Minute-scale forecasting of wind power using long-range lidar data A thesis accepted by the Faculty of Aerospace Engineering and Geodesy of the University of Stuttgart in partial fulfillment of the requirements for the degree of Doctor of Engineering Sciences (Dr.-Ing.) by Ines Würth born in Stuttgart, Germany Main referee: Prof. Dr. Po Wen Cheng Co-referee: Prof. Dr. Charlotte Bay Hasager Date of defense: 06.12.2021 Institute of Aircraft Design University of Stuttgart 2022 Contents Abbreviations xiii List of Symbols xv Abstract xvii Kurzfassung xix 1 Introduction 1 1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Research areas in this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.3 Research objectives, methodology and organization . . . . . . . . . . . . . . . . 4 2 Background 7 2.1 Intra-hourly variability of wind power generation . . . . . . . . . . . . . . . . . 7 2.2 Application areas for minute-scale forecasting . . . . . . . . . . . . . . . . . . . 9 2.3 Lidar in wind energy applications . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.4 How is wind power forecast today? . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.4.1 A note on terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.4.2 Numerical weather prediction . . . . . . . . . . . . . . . . . . . . . . . . 13 2.4.3 Statistical time series models . . . . . . . . . . . . . . . . . . . . . . . . 15 2.4.4 Where is the gap that needs to be closed? . . . . . . . . . . . . . . . . . 15 2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 3 The forecasting chain 19 4 First step in the forecasting chain: from radial velocity to wind field 23 4.1 Measurement setup at the onshore and offshore site . . . . . . . . . . . . . . . . 24 4.1.1 Definition of scientific objectives . . . . . . . . . . . . . . . . . . . . . . . 24 4.1.2 Site selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 4.1.3 Site characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 4.1.4 Experiment layout design . . . . . . . . . . . . . . . . . . . . . . . . . . 29 4.1.5 Infrastructure planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 4.1.6 Deployment and calibration procedures . . . . . . . . . . . . . . . . . . . 32 4.1.7 Scanning modes design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 iv Contents 4.1.8 Execution and data collection . . . . . . . . . . . . . . . . . . . . . . . . 33 4.1.9 Decommissioning and post-calibration procedures . . . . . . . . . . . . . 35 4.1.10 Data availability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 4.2 Getting a useful wind speed out of a lidar . . . . . . . . . . . . . . . . . . . . . 37 4.2.1 Data filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 4.2.2 Wind field reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . 43 4.3 Lessons learned . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 5 Along the forecasting chain: wind field evolution 59 5.1 Using Taylor’s hypothesis in wind energy applications . . . . . . . . . . . . . . . 60 5.2 Using Taylor’s hypothesis for minute-scale forecasting . . . . . . . . . . . . . . . 61 5.3 Lessons learned . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 6 Reaching the end of the chain: minute-scale forecasts of wind speed and power 65 6.1 Methodology of probabilistic minute-scale forecasts . . . . . . . . . . . . . . . . 66 6.1.1 Converting wind speed to power . . . . . . . . . . . . . . . . . . . . . . . 69 6.1.2 Choosing a forecast horizon . . . . . . . . . . . . . . . . . . . . . . . . . 70 6.1.3 Quantifying the uncertainty with probability density function and cumu- lative distribution function . . . . . . . . . . . . . . . . . . . . . . . . . . 70 6.1.4 Evaluating probabilistic forecasts . . . . . . . . . . . . . . . . . . . . . . 75 6.1.5 Comparing to benchmark forecasting method persistence . . . . . . . . . 79 6.2 Results of minute-scale forecasts from the onshore campaign . . . . . . . . . . . 82 6.2.1 Availability of forecasts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 6.2.2 Forecasts in the time domain . . . . . . . . . . . . . . . . . . . . . . . . 85 6.2.3 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 6.2.4 Sharpness and skill . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 6.3 Results of minute-scale forecasts from the offshore campaign . . . . . . . . . . . 93 6.3.1 Availability of forecasts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 6.3.2 Forecasts in the time domain . . . . . . . . . . . . . . . . . . . . . . . . 95 6.3.3 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 6.3.4 Sharpness and skill . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 6.4 Impact of wind ramps on the forecast accuracy . . . . . . . . . . . . . . . . . . . 113 6.4.1 Detecting wind ramps . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 6.4.2 Assessing the ramp impact on the forecast accuracy . . . . . . . . . . . . 118 6.4.3 Case study of a failed ramp forecast in the onshore campaign . . . . . . . 118 6.5 Lessons learned . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 7 Challenges for the implementation of lidar-based minute-scale forecasting 127 7.1 How far do we see? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 7.1.1 Correlating the measurement range and the forecasting horizon . . . . . 128 7.1.2 Analysing the measurement range . . . . . . . . . . . . . . . . . . . . . . 128 7.2 Overcoming barriers to adoption with the help of the community . . . . . . . . . 140 7.2.1 Availability of measurements . . . . . . . . . . . . . . . . . . . . . . . . . 141 7.2.2 Reliability and pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 7.2.3 Need for standards and common tools . . . . . . . . . . . . . . . . . . . . 142 Contents v 7.3 Lessons learned . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 8 Summary and conclusions 145 8.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 8.1.1 How should the lidar data information be processed to gain a power forecast for a wind turbine? . . . . . . . . . . . . . . . . . . . . . . . . . 145 8.1.2 How does the measurement setup and measurement site influence the forecast? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 8.1.3 What is the forecast horizon of lidar-based forecasts and what influences the forecast horizon? How does the lidar-based forecast perform in com- parison to state-of-the-art statistical methods and what are its benefits? . 147 8.2 Discussion and future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 8.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 A Appendix 151 A.1 Lidar Data Sheet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 A.2 Met mast sensor equipment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152 Bibliography 155 List of Figures 2.1 Example of wind ramps in wind speed and generated power. . . . . . . . . . . . 9 2.2 Forecasting horizons of diferent wind energy applications. . . . . . . . . . . . . . 10 2.3 State-of-the-art forecasting approaches. . . . . . . . . . . . . . . . . . . . . . . . 13 2.4 Different classes of weather models. . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.5 Forecast error for different forecasting techniques. . . . . . . . . . . . . . . . . . 16 3.1 The lidar forecasting chain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 4.1 Measurement sites marked on map of Germany. . . . . . . . . . . . . . . . . . . 25 4.2 Wind characteristic in Stötten measured at the met mast. . . . . . . . . . . . . 26 4.3 Map of the measurement site in Stötten. . . . . . . . . . . . . . . . . . . . . . . 27 4.4 Wind characteristic in alpha ventus. . . . . . . . . . . . . . . . . . . . . . . . . . 28 4.5 StreamLine XR lidar on the top level platform of the radio tower. . . . . . . . . 29 4.6 StreamLine XR lidar on the nacelle of the AV04 turbine in alpha ventus. . . . . 30 4.7 Schematic drawing of the lidar measurement setup. . . . . . . . . . . . . . . . . 33 4.8 Typical time series of LOS lidar data unfiltered and filtered. . . . . . . . . . . . 39 4.9 Radial wind speed over CNR sorted by the measurement range. . . . . . . . . . 40 4.10 Image of the Stuttgart TV tower with edge detection filter. . . . . . . . . . . . . 41 4.11 Demonstration of edge filter for a generic wind speed time series. . . . . . . . . . 42 4.12 Percentage of good points after filtering with edge filter. . . . . . . . . . . . . . 42 4.13 Radial velocity over CNR filtered with the edge filter. . . . . . . . . . . . . . . 43 4.14 Overview of the coordinate systems used for the wind field reconstruction. . . . 46 4.15 Wind direction and wind speed using global-local reconstruction method. . . . . 47 4.16 Regression of wind data using the global-local reconstruction method. . . . . . . 48 4.17 Sketch of the local moving window reconstruction method. . . . . . . . . . . . . 49 4.18 Regression using the moving window reconstruction method. . . . . . . . . . . . 50 4.19 Regression parameters using the moving window reconstruction method. . . . . 51 4.20 Scan plots of wind direction and wind speed from different reconstruction methods. 53 4.21 Overview of the coordinate systems used in alpha ventus. . . . . . . . . . . . . . 54 4.22 Reconstructed horizontal wind speed in alpha ventus. . . . . . . . . . . . . . . . 56 4.23 Correlation of the yaw angle and the wind direction. . . . . . . . . . . . . . . . . 56 5.1 Sketch of a turbulent eddy passing a wind turbine. . . . . . . . . . . . . . . . . 60 5.2 Scan plots of wind speed propagation. . . . . . . . . . . . . . . . . . . . . . . . . 64 6.1 Timeline of measured and forecasted wind speed. . . . . . . . . . . . . . . . . . 66 viii List of Figures 6.2 Probabilistic forecast process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 6.3 Power curves of the onshore and offshore site. . . . . . . . . . . . . . . . . . . . 71 6.4 Example of uncertainty information for forecasted power data. . . . . . . . . . . 73 6.5 Examples of PIT histograms of different forecasts. . . . . . . . . . . . . . . . . . 76 6.6 Predictive PDF and CDF with low sharpness and high sharpness. . . . . . . . . 78 6.7 Comparison of the wind speed measurements at the met mast and turbine. . . . 79 6.8 Correlation of the wind speed measured at turbine and met mast. . . . . . . . . 80 6.9 Example of a probabilistic persistence power forecast. . . . . . . . . . . . . . . . 82 6.10 Availability of power data and forecasts onshore. . . . . . . . . . . . . . . . . . . 84 6.11 Example of the probabilistic wind speed and power forecast. . . . . . . . . . . . 86 6.12 PIT histograms for onshore wind speed forecasts. . . . . . . . . . . . . . . . . . 87 6.13 PIT histograms for onshore power forecasts. . . . . . . . . . . . . . . . . . . . . 88 6.14 CRPS of onshore forecast. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 6.15 Histograms of wind speed CRPS onshore. . . . . . . . . . . . . . . . . . . . . . . 92 6.16 Histograms of power CRPS onshore. . . . . . . . . . . . . . . . . . . . . . . . . 92 6.17 Availability of power data and forecasts offshore. . . . . . . . . . . . . . . . . . . 94 6.18 Example of the probabilistic forecast offshore. . . . . . . . . . . . . . . . . . . . 96 6.19 PIT diagrams wind speed forecasts offshore. . . . . . . . . . . . . . . . . . . . . 98 6.20 PIT diagrams power forecasts offshore. . . . . . . . . . . . . . . . . . . . . . . . 99 6.21 CRPS of offshore forecasts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 6.22 Histograms of CRPS of wind speed forecast offshore. . . . . . . . . . . . . . . . 102 6.23 Histograms of CRPS of power forecast offshore. . . . . . . . . . . . . . . . . . . 102 6.24 Example CDFs of wind speed and power forecast. . . . . . . . . . . . . . . . . . 103 6.25 Histograms of error of deterministic forecast. . . . . . . . . . . . . . . . . . . . . 104 6.26 Polar plots of CRPS depending on wind direction. . . . . . . . . . . . . . . . . . 105 6.27 CRPS depending on the turbine wind speed. . . . . . . . . . . . . . . . . . . . . 106 6.28 Increase of measurement height against the measurement range. . . . . . . . . . 107 6.29 Deviation of wind speed between lidar and hub height. . . . . . . . . . . . . . . 108 6.30 CRPS for different measurement ranges. . . . . . . . . . . . . . . . . . . . . . . 110 6.31 CRPS depending on the number of measured wind speeds vectors. . . . . . . . . 112 6.32 CRPS depending on stability classes. . . . . . . . . . . . . . . . . . . . . . . . . 113 6.33 Schematic diagram of a ramp matrix. . . . . . . . . . . . . . . . . . . . . . . . . 115 6.34 Number of upward and downward ramps. . . . . . . . . . . . . . . . . . . . . . . 116 6.35 Timeline of power generation with ramps marked. . . . . . . . . . . . . . . . . . 117 6.36 CRPS for periods with and without ramp event. . . . . . . . . . . . . . . . . . . 119 6.37 Timeline of a ramp event at the onshore site. . . . . . . . . . . . . . . . . . . . . 120 6.38 Wind speed propagation during a ramp event. . . . . . . . . . . . . . . . . . . . 122 7.1 Forecast horizon calculated based on Taylor. . . . . . . . . . . . . . . . . . . . . 129 7.2 Schematic visualization of two methods to determine the maximum range. . . . 130 7.3 Comparison of the Sum Range and Weighted Range method. . . . . . . . . . . . 131 7.4 Maximum measurement range over time. . . . . . . . . . . . . . . . . . . . . . . 131 7.5 Comparison of of the valid measurement points over the measurement range. . . 133 7.6 Boxplot of maximum measurement range for different number of pulses. . . . . . 135 7.7 Histograms of maximum range for different number of pulses. . . . . . . . . . . . 135 7.8 Maximum measurement range versus different environmental conditions. . . . . 137 7.9 Correlation of humidity and rain sensor data. . . . . . . . . . . . . . . . . . . . 137 List of Figures ix 7.10 MIC between maximum range and environmental variables. . . . . . . . . . . . . 138 7.11 Maximum measurement range on June, 8th 2016. . . . . . . . . . . . . . . . . . 138 7.12 Webcam pictures of the measurement site on June 8th, 2016. . . . . . . . . . . . 139 A.1 Sensor equipment of the met mast in Stötten, effective 13.10.2015 . . . . . . . . 152 List of Tables 4.1 Overview of lidar scans at the onshore campaign in Stötten. . . . . . . . . . . . 34 4.2 Overview of lidar scans at the offshore campaign in alpha ventus. . . . . . . . . 34 4.3 Overview of available data for the onshore campaign. . . . . . . . . . . . . . . . 36 4.4 Overview of available data for offshore measurement campaign. . . . . . . . . . . 36 6.1 Filter criteria for onshore power curve. . . . . . . . . . . . . . . . . . . . . . . . 69 6.2 Example timeline of turbine power data. . . . . . . . . . . . . . . . . . . . . . . 81 6.3 Example of probabilistic persistence power data sets. . . . . . . . . . . . . . . . 81 6.4 Total number of available forecasts onshore. . . . . . . . . . . . . . . . . . . . . 83 6.5 Number of available forecasts offshore. . . . . . . . . . . . . . . . . . . . . . . . 95 6.6 Median of wind speed CRPS offshore. . . . . . . . . . . . . . . . . . . . . . . . . 100 6.7 Stability classes using RiS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 7.1 Overview of measurement periods with different number of pulses. . . . . . . . . 134 A.1 Data sheet for Stream Line XR lidar. . . . . . . . . . . . . . . . . . . . . . . . . 151 A.2 Sensors on the met mast of the offshore research platform FINO1. . . . . . . . . 153 Abbreviations BSH Bundesamt für Seeschifffahrt und Hydrographie (Federal Maritime and Hydrographic Agency of Germany) CDF Cumulative Distribution Function CFD Computational Fluid Dynamics CNR Carrier-to-Noise Ratio CRPS Continuous Ranked Probability Score DBS Doppler Beam Swinging DOF Degree Of Freedom FFT Fast Fourier Transform FWHM Full Width at Half Maximum IFB Institut für Flugzeugbau (Institute of Aircraft Design) LES Large Eddy Simulation lidar Light detection and ranging LOS Line-Of-Sight MAE Mean Average Error MIC Maximum Information Coefficient NWP Numerical Weather Prediction PDF Probability Density Function PIT Probability Integral Transform PO Project Organizer PPI Plan Position Indicator PV Photovoltaics RAVE Researach At Alpha Ventus RHI Range Height Indicator RMSE Root Mean Square Error SCADA Supervisory Control and Data Acquisition SWE Stuttgart Wind Energy UTC Universal Time Coordinated VAD Velocity Azimuth Display List of Symbols Greek letters α wind shear coefficient ∆ range [m] γ̄ turbine tilt angle [◦] λ diameter of eddy [m] Θ̄v average potential temperature [K] Roman letters A position matrix of measurement points [-] b offset of regression d distance to measurement point [m] f probability density function F cumulative density function g acceleration due to gravity [m s−2] h smoothing parameter K Kernel function m slope of regression n counter O observation p pressure P power [% of rated power] q percentile Q quantile r vector of measured radial wind speeds [m s−1] xvi List of Symbols R2 coefficient of determination Ris Richardson number speed s vector of unknown wind components [m s−1] t time u longitudinal wind component [m s−1] v lateral wind component [m s−1] V EL radial velocity [m s−1] w vertical wind component [m s−1] x longitudinal coordinate [m] X random variable y lateral coordinate [m] z vertical coordinate [m] Subscripts (·)i referring to a specific measurement point (·)I referring to the inertial coordinate system (·)L referring to the lidar coordinate system (·)r referring to a wind ramp definition (·)W referring to the wind coordinate system Operators (·)-1 inverted ∼ proportional Abstract With the introduction of renewable energies, the power grid has transformed from a centralised to a decentralised system. To balance the supply and demand of power in the energy grid at all times in spite of the volatile nature of wind and solar power, grid operators have to rely on accurate forecasts. However, state of the art wind power forecasting methods are not able to forecast changes of power in the minute-scale accurately. Therefore new methods are needed. This thesis investigates the use of a long-range lidar to forecast wind power on the minute-scale. To that aim, two measurement campaigns were carried out. One was an onshore campaign, where the lidar was installed fixed on a radio tower next to a turbine that a forecast was made for. The second was an offshore campaign where the lidar was installed on top of the nacelle of a wind turbine. Both campaigns lasted over several months and the wind speed was measured in several kilometers in front of the turbine. During this time the turbine‘s own data system also recorded the 10-minute average power from the turbine. In this thesis, a wind power forecast process is established. Lidar data is transformed from radial velocity to filtered horizontal wind speed and wind direction. The wind field information is then propagated to the wind turbine with an advection model based on Taylor’s hypothesis. The forecasted wind speed at the turbine is then transformed into a forecasted power with the help of the power curve of the turbine. To account for the uncertainty in the wind speed and power forecast, probabilistic forecast methods are applied. The results show that lidar-based forecasts at the offshore site are accu- rate in a forecast horizon up to ten minutes and outperform the benchmark forecast method persistence. Longer forecast horizons are biased because only small wind speeds measured fur- ther away from the wind turbine arrive with a delay of more than ten minutes. At the onshore site, persistence outperforms the lidar-based method in all forecast horizons, includinig the forecast horizon up to 10 minutes. The reason is that the Taylor based advection model does not model the actual propagation at the complex onshore site well enough. During ramp events, the lidar-based forecast demonstrates its strength: information from the xviii Abstract wind speed measured a few kilometers in front of the turbine allows us to forecast changes of power. In comparison, persistence only uses old information and therefore cannot forecast any future changes. It is concluded that the added value of using a lidar for minute-scale forecasts lies in forecasting changes of power. As wind ramps are potentially critical to the grid stability, or can affect the cost of balancing the power system if they are not forecast well, using lidars at wind farms to improve the power forecast is advised. However, challenges to the implementation of lidar-based forecasts remain. Lidar measure- ments depend on the aerosol content in the air and therefore the availability of the measurements for a forecast is not guaranteed. A fallback solution is needed such as statistical models or nu- merical weather prediction. To achieve forecast horizons of more than 10 minutes, the lidar measurement range needs to be extended beyond 10 kilometers. And to establish lidars as a state-of-the-art forecasting tool, standards are needed, which could be enabled by groups such as the IEA Wind community. Wind lidar data coupled with propagation models and power curves has fundamental ad- vantages for minute-scale wind power forecasting. Although this thesis has shown that current approaches may not be perfect, the rapid pace of wind lidar technology development, the in- creasing number of users, and the growing network of third party service providers, suggests that wind lidar is the future of minute-scale wind power forecasting. Kurzfassung Mit der Einführung der erneuerbaren Energien hat sich das Stromnetz von einem zentralen zu einem dezentralen System gewandelt. Um trotz der volatilen Natur von Wind- und Solaren- ergie das Angebot und die Nachfrage von Strom im Energienetz jederzeit auszugleichen, sind Netzbetreiber auf genaue Vorhersagen angewiesen. Die aktuellen Methoden zur Vorhersage der Windkraft sind jedoch nicht in der Lage, Leistungsänderungen im Minutenbereich genau vorherzusagen. Daher werden neue Methoden benötigt. In dieser Arbeit wird der Einsatz eines long-range Lidars zur Vorhersage der Windleistung im Minutenbereich untersucht. Zu diesem Zweck wurden zwei Messkampagnen durchgeführt. Die erste war eine onshore Kampagne, bei der das Lidar fest auf einem Funkturm neben einer Windenerrgieanlage in- stalliert wurde, für die eine Vorhersage gemacht werden sollte. Die zweite war eine offshore Kampagne, bei der das Lidar oben auf der Gondel einer Windkraftanlage installiert wurde. Beide Kampagnen dauerten mehrere Monate und die Windgeschwindigkeit wurde in mehreren Kilometern vor der Anlage gemessen. Während dieser Zeit zeichnete das anlageneigene Daten- erfassungssystem auch die 10-minütige Durchschnittsleistung der Anlage auf. In dieser Arbeit wird ein Verfahren zur Windleistungsvorhersage entwickelt. Die Lidardaten werden von der Radialgeschwindigkeit in die gefilterte horizontale Windgeschwindigkeit und Windrichtung transformiert. Die Windfeldinformationen werden dann mit einem Advektion- smodell, das auf der Taylor-Hypothese basiert, auf die Windenergieanlage übertragen. Die prognostizierte Windgeschwindigkeit an der Anlage wird dann mit Hilfe der Leistungskurve der Anlage in eine prognostizierte Leistung umgewandelt. Um die Unsicherheit in der Windgeschwindigkeits- und Leistungsvorhersage zu berücksichti- gen, werden probabilistische Vorhersagemethoden angewendet. Die Ergebnisse zeigen, dass lidarbasierte Vorhersagen am offshore Standort in einem Vorhersagehorizont von bis zu zehn Minuten genau sind und Persistenz als Benchmark-Vorhersagemethode übertreffen. Längere Vorhersagehorizonte sind fehlerbehaftet, da nur kleine Windgeschwindigkeiten, die weiter ent- fernt von der Windkraftanlage gemessen werden, mit einer Verzögerung von mehr als zehn xx Kurzfassung Minuten eintreffen. Am onshore Standort übertrifft die Persistenz die lidarbasierte Methode in allen Vorhersagehorizonten, einschließlich des Vorhersagehorizonts bis zu 10 Minuten. Der Grund dafür ist, dass das Taylor-basierte Advektionsmodell die tatsächliche Ausbreitung an dem komplexen onshore Standort nicht gut genug abbildet. Bei Rampenereignissen spielt die lidarbasierte Vorhersage ihre Stärke aus: Die Information aus der wenige Kilometer vor der Anlage gemessenen Windgeschwindigkeit erlaubt es, Leis- tungsänderungen zu prognostizieren. Im Vergleich dazu nutzt die Persistenz nur alte Informa- tionen und kann daher keine zukünftigen änderungen vorhersagen. Es wird gefolgert, dass der Mehrwert der Verwendung eines Lidars für Prognosen im Minutenbereich in der Vorhersage von Leistungsänderungen liegt. Da Windrampen potenziell kritisch für die Netzstabilität sind oder die Kosten für den Ausgleich des Stromsystems beeinflussen können, wenn sie nicht gut vorhergesagt werden, ist der Einsatz von Lidaren in Windparks zur Verbesserung der Leis- tungsvorhersage ratsam. Allerdings bleiben Herausforderungen bei der Implementierung von lidarbasierten Vorher- sagen bestehen. Lidarmessungen sind abhängig vom Aerosolgehalt in der Luft und daher ist die Verfügbarkeit der Messungen für eine Vorhersage nicht garantiert. Es wird eine Auswe- ichlösung benötigt, wie statistische Modelle oder numerische Wettervorhersagen. Um Vorher- sagehorizonte von mehr als 10 Minuten zu erreichen, muss der Lidarmessbereich außerdem auf mehr als 10 Kilometer erweitert werden. Und um Lidare als modernes Vorhersageinstrument zu etablieren, werden Standards benötigt, die durch Gruppen wie die IEA Wind Community ermöglicht werden könnten. Windlidardaten, die mit Ausbreitungsmodellen und Leistungskurven gekoppelt sind, haben fundamentale Vorteile für die Vorhersage von Windenergie im Minutenbereich. Obwohl diese Arbeit gezeigt hat, dass die derzeitigen Ansätze nicht perfekt sind, legen die rasante Entwick- lung der Wind-Lidar-Technologie, die steigende Anzahl von Nutzern und das wachsende Net- zwerk von Drittanbietern nahe, dass Windlidar die Zukunft der Windleistungsvorhersage im Minutenbereich ist. 1 Introduction 1.1 Motivation The use of renewable energies is an essential part of reducing climate-changing emissions. Dur- ing the last years, the energy grid transitioned from centralized and manageable energy produc- tion with conventional sources such as coal, nuclear and gas, to decentralised energy production from solar and wind power. Both technologies work with energy sources that nature provides for free, but they come with a catch. Solar and wind power are fluctuating energy sources and are governed by local changes. Without intermediate storage, they do not provide a steady energy output. However, in order to keep the energy grid stable, the supply from power plants and the demand from consumers need to be balanced at all times. As a result of the energy transition, weather forecasting became a crucial tool to tackle the challenge of grid balancing, because it helps to manage the variable energy supply. If one looks up weather forecasting in Wikipedia, it says that “Weather forecasting is the application of science and technology to predict the conditions of the atmosphere for a given location and time.” In fact, people have tried to forecast the weather for millennia and since the 19th century national weather services successfully forecast meteorological conditions [1]. With the introduction of renewable energies into the energy grid, weather forecasts became important to forecast wind and solar power. 2 1 Introduction Wind power generation forecasts hours or days ahead of real time are based on weather forecasts which are calculated with Numerical Weather Prediction (NWP) models. NWP mod- els are accepted as baseline forecasts and use input data such as temperature, humidity and pressure from weather stations and simulate future weather conditions by solving physical and mathematical equations [2]. NWP models are computationally expensive and so tend to be run by large governmental organisations, such as the Deutscher Wetterdienst (German Meteo- rological Service), or commercial providers. However, for wind power forecasts up to 60 minutes, different forecast methods are needed because due to their spatial resolution of several kilometers, NWP models are not accurate for short time scales. The state-of-the-art is to use statistical methods that are based on historic measurement data of wind and power at the target wind turbine or wind farm. These forecasts are reliable, cheap and do not need much computing power [3], but have a disadvantage. Because they use old measurement information to predict the future performance, they are not able to predict changes in power generation. The successful integration of wind energy into the grid therefore requires a different ap- proach to wind and power forecasting. Ideally this would be an accurate forecast that provides temporally-resolved data at the point of interest and can be updated within a few seconds or minutes. Wind lidar may be an ideal tool for this application. Long-range lidars can measure the wind speed several kilometers upstream of a wind turbine or wind farm. This preview information of the wind speed that will affect the turbine’s power generation, can be used to generate power forecasts for the next minutes. The lidar is able to measure the variation in wind speed and should therefore be able to predict changes in generated power more accurately than conventional time series based methods. This thesis will therefore investigate the use of a long-range lidar to predict the power output of a wind turbine in the minute-scale. 1.2 Research areas in this thesis Lidars are used for many different applications in wind energy such as site assessment, turbine or wind farm control and power curve measurements [4]. All applications have one thing in common: the lidar data needs to be processed to be usable. This is the case for all data from measurements, and not specific to lidar data. The goal of lidar data processing for wind power forecasting is to make sure that only accurate wind measurements are used to generate the forecast. 1.2 Research areas in this thesis 3 The first step is to check for outliers in the data and filter the data accordingly. Lidars measure radial wind speed, so the second step is to apply wind field reconstruction methods to retrieve wind parameters such as horizontal wind speed and wind direction from the mea- surements [5]. When the work for this thesis was started, long-rage lidars were still new on the market and there was not much experience in terms of data processing. Therefore, the process- ing chain first had to be developed in this thesis based on existing methods for ground-based and short-range nacelle-based lidar measurements. Having chosen wind lidar as the tool to deliver the wind data, it is necessary to consider how to convert this into a forecast. The forecasting method that is applied in this thesis is probabilistic forecasting. Probabilistic forecasting is also called uncertainty forecasting and aims to provide a forecast of a value and quantifies the uncertainty of the forecast [6] at the same time. The goal is to help and facilitate decision making processes. For wind power forecasting, this means that not only the power output of a wind turbine is forecast for a specific time in the future, but also a probability that this power will be generated is given. Probabilistic power forecasting is able to quantify the uncertainty resulting from the volatile nature of the wind, which is responsible for the power generation. This thesis builds on existing methods for probabilistic forecasting and adapts them for the application of lidar-measurements. In parallel to the work carried out for this thesis, similar research was carried out by col- leagues at other institutes. In 2019, Elliot Simon finished his PhD thesis on ”Minute-Scale Wind Forecasting Using Lidar Inflow Measurements” [7]. He focused on using scanning-lidar to examine space-time correlations of wind patterns measured onshore, upstream of a lidar. In the same year Laura Valldecabres defended her PhD thesis. She showed in her work that lidars can forecast wind speeds at near-coastal conditions [8] and then focused on using radars to carry out minute-scale probabilistic power forecast [9], [10] and forecast ramp events [11]. The author of this thesis worked together with Elliot Simon and Laura Valldecabres and to- gether they organised a collaborative workshop between the IEA Wind Tasks 32 and 36 on ”Very short-term forecasting of wind power”, which took place in 2018. The outcome of this workshop was an overview paper of methods for minute-scale forecasting, led by the author of this thesis [12] and reported elsewhere in this thesis. 4 1 Introduction 1.3 Research objectives, methodology and organization This thesis aims to improve minute-scale forecasts of wind power using wind speed data from a long-range lidar and power data from wind turbines. The following research questions arise: • How should the lidar data information be processed to gain a power forecast for a wind turbine? • How does the measurement setup and measurement site influence the forecast? • What is the forecast horizon of lidar-based forecasts and what influences the forecast horizon? • How does the lidar-based forecast perform in comparison to state-of-the-art statistical methods and what are its benefits? The data needed to answer these questions were collected in two measurement campaigns. In one campaign a wind lidar was installed next to an onshore wind turbine. In another, a wind lidar was installed on top of an offshore wind turbine. Both campaigns were carried out over a period of several months. Simultaneous to the lidar data, turbine data was made available by the operator and used in the forecast process to generate the power curve of the turbines and validate the forecasts. At both sites, data from a meteorological mast was available which recorded meteorological conditions. The methods used in this thesis to process the data and to generate the forecasts were previously established for other applications. This thesis introduces each method, discusses its context and previous application, and explains how it was adapted to be used for lidar-based minute-scale forecasting in this work. Assumptions that were made for simplification are stated clearly. The methodology applied in this thesis is valid for the power forecast of a single wind turbine. To forecast the power output of a wind farm, the methods applied here need to be extended and wind farm effects need to be considered. The thesis starts by explaining background information in Chapter 2. An overview of the use of lidar in wind energy applications is given and the need for minute-scale forecasts is explained by discussing the variability of intra-hourly wind power generation. By describing the state of the art forecasting methods, the gap that lidar-based forecasts can close is described. The forecast chain that is established in this thesis is explained in Chapter 3. The chain is an overview of which steps are taken in order to process lidar data and to gain a wind power forecast of a wind turbine. The thesis structure then follows the chain links step by step and explains the processing steps. 1.3 Research objectives, methodology and organization 5 • In the fist link of the forecasting chain, wind speed measurements from the lidar are converted into horizontal wind speeds, which are the basis for the forecast. The methods for filtering lidar data, the wind field reconstruction, as well as the measurement setup are explained in Chapter 4. • In order to forecast the power of a wind turbine, these distant measurements need to be propagated through time and space to the wind turbine’s location. This is the second link in the forecasting chain and the use and implementation of a propagation model are explained in Chapter 5. • The final link in the forecast chain is to transform the predicted wind speeds into pre- dicted power, and to calculate the forecasts. The methods applied are explained in 6 and the results of the forecasts from an onshore and offshore campaign are compared and discussed. The challenges for the implementation of lidar-based minute-scale forecasting are discussed in Chapter 7. To that end first the correlation of the measurement range and the forecasting horizon is analysed, and then solutions to overcome barriers are suggested and discussed. The thesis concludes with Chapter 8, where conclusions and recommendations for further research are given. 2 Background This chapter explains the background that is needed to understand the need for minute-scale forecasting and the reason why a lidar was chosen as a tool. First the reason for the intra-hourly variability of wind power is explained in Section 2.1 and the need for minute-scale forecasting in different application areas that arises from the power variability is discussed in Section 2.2. The use of lidar in different wind energy applications is explained in Section 2.3. In Section 2.4 the state-of-the-art forecasting methods are described and the need for lidar-based forecasting is derived. Section 2.5 summarises the key information needed to understand the background of this thesis. 2.1 Intra-hourly variability of wind power generation The need to forecast wind power on timescales in the order of minutes arises from the variability of the wind speed in these time scales. The cause for the variability can be found in atmospheric phenomena with spatial scales from one up to tens of kilometers that introduce the fluctuations in the wind. There are four dominant phenomena [13]: • Open cellular convection is an offshore phenomenon that occurs when cold air is convected over warm sea water. Clouds are formed that have a honeycomb-like structure and rising air in the cloudy cell walls leads to a horizontal wind speed variability. When such cells move over an offshore wind farm, the farm experiences the spatial wind speed variability as a temporal variability. 8 2 Background • For coastal wind farms the land and sea breeze can lead to power fluctuations in the minute scale. The breeze is caused by the surface temperature difference of mainland and water causing sea breezes during the day and land breezes at night. Although the sea breeze circulation is predominantly a diurnal effect, it can also influence smaller scales when, for example, convergence of the sea breeze front with the background flow results in squall lines or showers near the coast. • Gravity waves in the atmosphere are a phenomenon that occurs over land and over sea. Gravity waves are waves generated in a fluid medium or at the interface between two media when the force of gravity or buoyancy tries to restore equilibrium. They cause wave-like cloud structures in the atmosphere and similar wave structures in the wind speed pattern with amplitudes of several meters per second. • Low level jets are regions of increased wind speed in the lower hundred meters of the atmospheric boundary layer. They occur onshore and offshore and are of special interest to wind energy as the extreme wind shear over the rotor causes extreme loading on the turbine and unexpectedly high power generation. As the mechanism for low-level jet formation is not fully understood, they are difficult to forecast. All these phenomena cause fluctuations in the wind speed in the minute scale that are aug- mented to the third power when transferred to the power generated by a wind turbine or wind farm. In Figure 2.1 an example time series of one day of wind speed and generated power of a single wind turbine is given. Small fluctuations in wind speed result in large fluctuations in generated power. The figure also shows phenomena referred to as ramp events. These rapid and strong changes in generated power are caused by extreme changes in wind speed or direc- tion and are associated with the passage of weather fronts. Ramp events can also have other causes than extreme changes in wind speed. Ramp events are especially critical around cut-out wind speed. If the cut-out wind speed of a wind farm is reached the whole wind farm shuts down. In this case a small fluctuation in wind speed can lead to a strong decrease in the power production. Ramp events are critical for the grid balancing as they are difficult to forecast and thus often occur unexpectedly. In [14] an overview of recent ramp forecasting techniques is presented. The article also points out that despite being critical, there is no common definition of ramp events. Instead, the definition usually depends on location as well as the size of the considered wind turbine or wind farm. However, the basic parameters that need to be forecast are clear: ∆t is the duration of the ramp and ∆P is the minimum change of power within ∆t. Positive ∆P implies an upward ramp, negative ∆P a downward ramp. Downward ramps tend to be more critical for grid stability than upward ramps. A downward ramp causes an energy shortage that needs to be compensated using available power generations, while an upward ramp can 2.2 Application areas for minute-scale forecasting 9 Figure 2.1: Example time series of wind speed and generated power of a single wind turbine with wind ramps marked for a time window of 60min and a change of power of 40%. Each data point in the time series corresponds to a 10-minute average. Reproduced without modifications from Würth et al. [16] with permission. be managed by curtailing the wind farm [15]. Wind ramp forecasting errors are specified as level errors and phase errors. Level error describe the magnitude of the power change that was forecast inaccurately, while phase errors describe the deviation in time when the ramp occurred. 2.2 Application areas for minute-scale forecasting Minute-scale forecasts of wind speed and power are needed in three application areas in the wind energy sector [12]: • Wind farm control: the controller of the wind farm uses preview information of the wind speed and direction to optimize the power output and reduce loads. • Power grid balancing: the Transmission System Operators (TSO) need information about changes of the produced power in their grid to balance power production and load at all times and to manage power reserves. • Energy and ancillary services markets: wind power energy trade takes place in intra-day markets and forecasts are needed to reduce imbalance costs and increase revenue. Each application area is associated with a different minute-scale forecast horizon in the minute scale (Figure 2.2). Wind turbine and wind farm control uses forecast information of the wind parameters of 10 2 Background Forecast horizon 1s 1min 10min 30min 1hour Wind turbine and wind plant control Power grid balancing Scanning lidar-based models Statistical methods NWP models + measurements Methods Applications 5min Energy and ancillary services market Radar-based models Figure 2.2: Overview of forecast horizons of different wind energy applications in the second and minute scale. Reproduced with modifications from Würth et al. [12] with permission. up to 10 minutes before they arrive at the turbine or wind farm. These forecasts can be used to control wind turbines by modifying turbine alignment or blade pitch. For instance for yaw control, an estimation of the wind direction is used to align the rotor with the wind direction. The preview wind information can also be used for wake steering which uses the misalignment effect to steer the wake of a wind turbine away from a downstream turbine. TSOs have the task to balance the power supply and power demand at all times to keep the frequency of the grid stable at 50 Hz. If the power production differs from the demand, balancing actions are needed. A shortage of power supply (or a rise in consumption) would otherwise lead to a frequency drop, a rise in power supply (or a drop in consumption) would lead to a frequency increase. To balance the supply and demand of power all times TSOs therefore need a control reserve, which is also called balancing power. There are three different control mechanisms established to activate the balancing power and they are categorized according to the activation time: primary control (within 30 sec), secondary control (within 5 min) and tertiary control (directly activated or supplied in schedules of up to 4 times 15 min). Forecasts are therefore useful in time ranges of up to 60 minutes to reduce the need for balancing actions and to lower the amount of balancing power that a TSO needs to keep on hold. Short-term wind energy trading takes place in intra-day markets. The regulations for the market depend on the country and also the lead times are country specific. In Germany and Australia for example the lead time for the trade is 5 min and the trading block of power is 15 min. This means that if a TSO or wind farm operator wants to sell their produced power on such a market, they need to know the amount of power they can sell as accurately as possible in advance. Forecasts are therefore necessary in the range of 5 min to 60 min. Minute-scale forecasts help to reduce the risk for penalties, that arise when the actual produced power differs from the power that was traded. This can happen if the produced power differs significantly from the forecast. Apart from the wind energy sector, wind speed forecasts are also relevant for other sectors, e.g. 2.3 Lidar in wind energy applications 11 the construction sector. Wind speed forecasts could help increase the safety during commission and operations on construction sites. If it is known in advance that a gust is coming, accidents can be avoided if heavy objects are lifted by a crane. 2.3 Lidar in wind energy applications Doppler wind lidars use the Doppler frequency shift of light to measure the speed of airborne particles. The measuring principle is based on the backscattering of the emitted laser light to a receiver [17]. The backscattering takes place at particles in the air, so-called aerosols, which scatter back the light frequency shifted by their own speed. This frequency shift is based on the Doppler effect. Due to the fact that the Doppler shift only affects particles moving in the direction of the laser beam, only this directional component of the wind speed can be measured. The radial wind speed is known as the line-of-sight wind speed and is denoted vlos. The aerosols that the lidar measuring principle is based on are [18]: • sea salt aerosols, especially near the sea. • dust aerosols, mineral origin • secondary aerosols, nitrates and sulfates • biological aerosols, e.g. fungal spores and pollen • smoke aerosols, from forest fires or anthropogenic caused • volcanic aerosols, especially in higher air layers These aerosols with a size in the order of the laser wavelength of 1.5 µm are particles that have very low settling velocities and thus are suspended by the wind. The quantity of the individual aerosols in the air however is strongly dependent on the location. The optical properties of the particles are subject to strong fluctuations. By accumulation of moisture from the air, for example, the diameter of the particles can fluctuate strongly and thus change the backscattering properties [18]. As a result, the lidar measurement properties are subjected to fluctuations and depend on the existence and the properties of the aerosols. Lidars are used for different applications in wind energy such as site assessment, turbine or wind farm control and power curve measurements [4]. For each application, a different lidar type is used, because the requirements for the wind measurements vary. For site assessment, ground- based vertical measuring lidars are ideal, which measure the wind speed and wind direction over several heights [19]. For turbine control, nacelle-based lidar systems are used, which use the preview information from the incoming wind field several hundred meters in front of the rotor, to determine control actions at the wind turbine to optimize the power performance and 12 2 Background reduce loads [20]. For wind farm control, forecasting of the wind is necessary, but not on a minute-scale, but on a second-scale [12]. Ground-based, and nacelle-based lidars are also used for power curve assessment to correlate the inflow wind speed over the whole rotor area to the generated power of the wind turbine [21]. For this application long-range lidars are used and mounted e.g. on the transition piece of offshore wind turbines, to measured the inflow wind speed outside of the induction zone of the turbine over several heights [22]. Minute-scale forecasting is a new application for wind lidars, and so far it is not established as common practice. All investigations conducted so far used long-range scanning lidars to measure the upstream wind velocity and forecast future power generation [12]. 2.4 How is wind power forecast today? The classic method to generate a wind power forecast comprises one or several of the following ingredients (Figure 2.3): • A numerical weather prediction (NWP) model which computes the state and evolution of the atmosphere e.g. weather forecast, • Observational data, e.g. Supervisory Control and Data Acquisition (SCADA) data from the wind turbine or wind farm that is either used to compute a forecast using statistical time series methods or serves as input for the forecast model, • Terrain information and information about the wind farm layout, These ingredients are combined using a forecast model that uses the information from NWP, observational data, and terrain information to generate a power forecast for a wind turbine or wind farm [23, 2]. In the following the ingredients are explained in more detail and their limitations for minute- scale forecasts are discussed. The goal is to derive the need for the lidar-based forecasting method. But first, the forecast terminology is clarified in the next section. 2.4.1 A note on terminology When talking about forecasting of wind power, the concept of the forecast horizon is often mentioned. The forecast horizon describes the time period or point in time in the future, for which the forecast is generated. When it comes to forecast horizons from seconds to a few hours, the terminology used for the description varies between nowcasting, very short-term, and short-term forecasting with no clear definition of the length of the time period. Therefore, for the sake of clarity, it was decided in the collaborative IEA Wind Task 32 and 36 Workshop on 2.4 How is wind power forecast today? 13 Figure 2.3: Elements in the forecasting approaches. Reproduced without modifications from Giebel et al. [2] with permission. “Very short-term forecasting of wind power”, to use the exact forecast horizon as a description. Therefore, this thesis only uses the term minute-scale forecast which describes a forecast horizon from one minute up to one hour. Longer forecast horizons would then be called hour-ahead or day-ahead forecasts. 2.4.2 Numerical weather prediction Numerical weather prediction models divide the atmosphere up into cells. Physical models are used to describe the state of the atmosphere in each of these cells. The parameters describing the atmosphere can then change over space and time. Depending on the size of the cells and the domain that is covered, the models are roughly divided into different classes (Figure 2.4). More accurate physics or bigger domains require more computational effort. As a result, forecasting the weather over large areas and longer periods of time is usually only possible for large organisations such as national weather services [23]. It is possible to provide high resolution time series forecasts by decreasing the size of the domain, or reduce the details of physics (but this might lead to increased uncertainty). Minute- scale forecasting requires a combination of high spatial resolution and domain sizes in the order 14 2 Background Domain size [km] H o ri zo n ta l r es o lu ti o n [m ] 0.001 0.01 1 10 0.01 0.1 100 1,000 0.1 100 1 10 1,000 10,000 100,000 CFD model Microscale / local model Mesoscale / regional model 10,000 Global model Figure 2.4: Different classes of weather models. The borders of the boxes are fuzzy, because the size specifications are to be understood as approximate values. of 10 km to 50 km. This makes numerical weather forecasting challenging, because it requires a combination of high resolution and comprehensive physics that is computationally expensive. NWP models also require boundary and initial condition data to deliver accurate minute- scale forecasts. Unfortunately these data are often not available for wind farms which makes it difficult to provide accurate forecasts. NWP forecasts are usally time consuming, meaning that it may be unable to deliver a minute-scale forecast in time for the operators to take actions. For these reasons, minute-scale forecasting is more typically based on simple algorithms based on the available on-site data. 2.4 How is wind power forecast today? 15 2.4.3 Statistical time series models Statistical approaches to forecasting mainly rely on deducing patterns from past observational data and extrapolating these relationships to predict future values over a desired time step [12]. Forecasts in wind energy are carried out for one dimensional time series signals such as a wind speed measurement, or SCADA data such as wind turbine or wind farm active power signal. The chosen forecast horizon should relate to the time resolution of available input data, and at a minimum be one sample (time step) ahead to avoid errors introduced by interpolation. Minute-scale statistical forecasting methods are largely identical to techniques employed for longer horizons. The main differences are the temporal resolution of the data and the variability of the physical process being predicted [12]. Benchmark statistical time series models are persistence and climatology. Persistence is a very simple forecast method and assumes that the forecasted conditions are the same as the present conditions. This means the most recent power measurement of a wind turbine is used for the power forecast. Climatology uses statistics from historic measurements, e.g. an average of the last n hours of measured power generation, to create a forecast. 2.4.4 Where is the gap that needs to be closed? NWP models are optimized to produce forecasts in the hour- and day-scale. They can produce weather forecasts for up to 15 days ahead. However, they are not very accurate in the minute- scale, where statistical time series models perform better (Figure (2.5). In fact, persistence frequently outperforms hour-ahead or day-ahead forecasts in the time range up to 60 min. This means new methods for minute-ahead forecasts need to be more accurate than persis- tence and persistence is therefore the benchmark. Persistence however has one disadvantage: it uses historic measurements to forecast future events. This approach produces large errors if the future event deviates significantly, such as a wind ramp. Considering that these changes in power are crucial information for TSOs who need to balance the grid, or wind farm operators selling their produced power, a better forecast method than persistence is needed. It is therefore of great interest to investigate if lidar-based forecasting can close the gap of forecasting power changes in the minute-scale and outperform the persistence model. 16 2 Background Figure 2.5: Qualitative visualization of the forecast error development over the first hours of a forecast for different temporal forecast techniques. Reproduced without modifications from Würth et al. [12] with permission. 2.5 Summary 17 2.5 Summary Minute-scale forecasts of wind speed or power are important for TSOs to keep the grid stable and to reduce balancing power, or for wind farm operators to optimize wind farm control or energy trading if they sell power on rolling markets. State-of-the-art NWP models are optimized to forecast wind conditions in the hour and day- scale. Time-series based forecasts produce more accurate minute-scale forecasts than NWP models, but rely on historical measurement data from the wind farms to forecast the power. Therefore they cannot forecast large changes in future power output, e.g. from wind ramps. Long-range lidars measure the wind speed remotely and can be used to measure the upstream wind speed of a wind turbine or wind farm. Lidar-based measurements therefore contain pre- view information of future wind speed changes. Therefore it is of great interest to investigate if lidar-based minute-scale forecasts are able to outperform conventional methods such as per- sistence. 3 The forecasting chain: from radial velocity to minute-scale wind power forecasts In order to gain minute-scale forecasts of wind power of a wind turbine from long-range lidar data, a forecasting chain had to be established in this thesis. The thesis will follow the chain link by link in the next chapters and explain the steps in detail. This chapter presents an overview of the steps and serves as orientation. In principle, the chain shows that data exists in one form and needs to be processed in order to reach a new form of data (Figure 3.1). On overview of the data forms and the steps of processing is given in the following. • LOS. Lidars measure the wind speed in Line-Of-Sight (LOS) direction along the laser beam and therefore measure the radial component of wind field in this direction. • Wind field reconstruction. As the lidar measures the radial component of the wind field at each measurement point, the three wind vector components u, v, w need to be deduced from this measurement at each point. This process is called wind field recon- struction. In order to reconstruct the wind field components from the LOS measurements, assumptions have to be made and algorithms have to be applied to the data, see Chapter 4. 20 3 The forecasting chain • Wind field. Long-range lidars are able to measure the wind field simultaneously in several measurement points up to a range of several kilometers. After wind field recon- struction, the wind field components u, v, w and the wind direction in those points are known. This data of the wind field is the basis for the minute-scale forecasting, because the measured wind field contains the preview information of the wind conditions the tur- bine will experience in the following minutes, if the lidar measures the inflow of the wind turbine. • Propagation model. The wind vectors measured in the distance are transported through space and time. This is known as wind field propagation. A model needs to be applied to the measured wind field data that determines how the propagation is hap- pening. With the help of the propagation model, the goal for minute-scale forecasting is to determine which wind vector measured in the inflow of the turbine will reach the turbine and at which future point in time this will happen, see Chapter 5. • Predicted wind speed at turbine. With the help of the propagation model, the measured wind field is transported though time and space and the predicted wind speed at turbine level is determined. This means, the wind conditions at the turbine location for the minutes after the forecast is issued are known. This wind speed forecast contains information from the lidar measurements several minutes prior to when the forecast is issued. • Power curve. In order to gain the power forecast of the turbine, the predicted wind speed at the turbine needs to be converted to power. This is achieved through means of the turbine’s power curve. The power curve is a unique property of each turbine type and sets the inflow wind speed in relation to the power the turbine produces for this wind speed. The power curve of a turbine is determined in a measurement campaign using free stream wind speed measurements and measured power values. • Predicted power. The predicted wind speed of the turbine is converted into the pre- dicted power using the turbine’s power curve. Similar to the predicted wind speed, this means the power output of the turbine for the minutes after the forecast is issued is known. The forecast can be generated for different forecast horizons, depending on the desired horizons and depending on for how many minutes in the future the forecasted power is actually available, see Chapter 6. This chain represents a very high level overview. To get from link to link, many more steps need to be considered, which is the task of the next chapters. It should also be noted that in this thesis the forecast of only one wind turbine is investigated. However, if the forecast is extended to wind farm level in the future, the chain can be extended with considerations of taking into account wake effects to aggregate the predicted power of several wind turbines. 21 Wind field reconstruction LOS Propagation model Wind field Power curve Predicted wind speed at turbine Predicted power Figure 3.1: The lidar forecasting chain. Data exists in one form (circle) and needs to be processed (arrow) in order to each a new form of data. Reproduced without modifications from Würth et al. [16]. 4 First step in the forecasting chain: from radial velocity to wind field Wind field reconstruction LOS Wind speed Wind field In the first step in the lidar forecasting chain, the radial wind speed measurements from the lidar are converted in several processing steps into horizontal wind speeds, which are the basis for the forecast. This chapter explains the processing steps but starts with an overview of the measurement setup in Section 4.1. Section 4.2 then explains the methods applied for filtering and wind field reconstruction in order to obtain a useful wind speed and wind direction signal from the lidar. 24 4 First step in the forecasting chain: from radial velocity to wind field 4.1 Measurement setup at the onshore and offshore site The data used in this thesis were generated in the projects VORKAST and ParkCast. VORKAST was a German national funded research project to optimise the design and operational man- agement of hybrid power plants and energy storage technologies by means of wind and Photo- voltaics (PV) power minute-scale forecasting [24]. The focus of the project was on onshore sites for wind and PV power forecasting. It was running from September 1st, 2014 to October 31st, 2017. ParkCast is the follow-up project to VORKAST and started in November 2018 with the goal to develop, optimize and evaluate new methods for minute-scale forecasts of offshore wind farms. ParkCast will end in October 2021. The author of this thesis was Project Organizer (PO) for the University of Stuttgart’s contributions for both projects. To describe the measurement setup that was used in the projects and that resulted in data that are used in this thesis, a 10-step methodology has been applied that was introduced by Vasiljevic et al. in 2017 [25]. This methodology provides guidance on how to carry out a lidar measurement campaign in order to ensure its success. The next sections will following these steps and explaining how they apply to the present campaigns, in order to give the reader a comprehensive understanding of the measurement setup. 4.1.1 Definition of scientific objectives Three scientific goals drove the planning of the VORKAST measurement campaign. First, the idea for the project was born in a time where commercial lidar measurement systems were first brought on the market that had an extended range and were able to measure the wind speed in distances of several kilometers. Therefore the first goal of the project (and thus the measurement campaign) was to test one of the new systems for functionality, and its applicability for minute-scale forecasting. Second, during the campaign the lidar should be able to measure as far as possible. The goal was to find out how to set up the lidar and carry out the measurements in order to obtain the best wind speed measurement for forecasting. This included the development of laser scan strategies to extract the wind field components from the radial wind speed, testing the data analysis process, and the maximum measurement range that could be reached. Third, the goal was to carry out minute-scale forecasts of wind speed and also wind power of an onshore wind turbine using the data from the long-range lidar. To that end, new forecasting methods using the lidar data should be developed, which had to take into account the dynamic variability of the wind. Thus the forecasts should be able to capture the variability of the power fluctuation. 4.1 Measurement setup at the onshore and offshore site 25 Figure 4.1: Measurement sites marked on map of Germany. Reproduced with modifications from [27] with permission. In ParkCast, the results from VORKAST should be used and transferred to an offshore wind farm. The goal was to investigated how the lidar-based minute-scale forecast methods developed for one turbine onshore could be transferred to a wind farm with several turbines offshore and how the forecasts perform under offshore conditions. It should be pointed out, that this thesis compares the methods developed for minutes-scale forecasts for one turbine onshore and offshore and does not take into account the forecasts of the whole offshore wind farm. 4.1.2 Site selection The site that was selected for the onshore measurement campaign was near Stötten in the south of Germany (Figure 4.1 bottom marker). The site was chosen due to its proximity to Stuttgart and prior use in related studies [26]. The proximity to the institute was important, because for this campaign a new lidar system had to be tested. When testing a new measurement system, it has proven beneficial to have easy access to be able to adjust its settings or repair it in case of failure. Due to related studies, the site also offered easy access to local wind turbine data and access to meteorological data from the institute’s meteorological (met) mast. As offshore site the alpha ventus wind farm in the Northern Bight 45 km north of the German island of Borkum was chosen (Figure 4.1 top marker). For ParkCast is was important to find a site that offered the possibility to install the lidar and get access to turbine data and also meteorological data. In 2003 the research platform FINO 1 (Forschungsplattformen in Nord- und Ostsee) was erected at the site, measuring the meteorological conditions with a met mast. In 2007 the German Federal Ministry for the Environment, Nature Conservation and 26 4 First step in the forecasting chain: from radial velocity to wind field 0◦ 15◦ 30◦ 45◦ 60◦ 75◦ 90◦ 105◦ 120◦ 135◦ 150◦ 165◦180◦195◦ 210◦ 225◦ 240◦ 255◦ 270◦ 285◦ 300◦ 315◦ 330◦ 345◦ 2.0 4.0 6.0 8.0 10.0 12.0 [0.0−2.0[ [2.0−4.0[ [4.0−6.0[ [6.0−8.0[ [8.0−10.0[ [10.0−12.0[ [12.0−14.0[ [14.0−16.0[ [16.0−18.0[ [18.0−20.0[ [20.0−22.0[ [22.0−24.0[ [24.0−∞] m s (a) Wind rose 0 5 10 15 20 25 30 v [ m s ] 0.00 0.02 0.04 0.06 0.08 0.10 0.12 F re qu en cy [− ] (b) Weibull distribution Figure 4.2: Wind characteristic in Stötten measured at the met mast. (a) Reproduced with modifi- cations from Hofsäß et al. [26] with permission. . Nuclear Safety (BMU) launched the RAVE (Research At Alpha Ventus) research initiative. The initiative’s goal was to support and facilitate research projects at the offshore wind farm. In 2009 alpha ventus started operating with 12 turbines and became Germany’s first offshore wind farm [28]. University of Stuttgart is member of the RAVE initiative and has conducted several research projects at the site. Access to the wind farm data and the FINO 1 data is obtained through data portals hosted by the Federal Maritime and Hydrographic Agency of Germany (BSH). Alpha ventus was chosen as a site for the ParkCast project, because as a RAVE member, easy access to the turbines to install and maintain a lidar and access to turbine and meteorological data was guaranteed. 4.1.3 Site characterization Stötten is in the Swabian Alps; the location is a very hilly area consisting of high plateaus surrounded by a pronounced 100 – 150 m tall wooded escarpment known in the region as the Albtrauf. A detailed study of the local meteorology [26] shows that the main wind direction is west to north-west and the most frequent wind speed is around 5 m s−1 (Figure 4.2). The measurements described in [26] are centred on a 100 m high met mast in relatively flat land less than 1 km easterly from a section of Albtrauf (Figure 4.3). On the plateau several wind turbines and a radio tower are located. alpha ventus is an offshore wind farm consisting of 12 turbines with two different turbine types set out in a 4-by-3 grid (Figure 4.4 (c)). The two northern rows are 5 MW turbines of type REpower 5M, with a rated power of 5 MW, a hub height of 92 m, and a rotor diameter of 126 m. The two southern rows are 5 MW turbines of the type Adwen AD 5-116, with a rated power of 5 MW, a hub height of 90 m, and a rotor diameter of 116 m. The turbines are enumerated row 4.1 Measurement setup at the onshore and offshore site 27 Radio tower Reference turbine Met mast 1350 m Figure 4.3: Map of the measurement site in Stötten. Map data: ©OpenStreetMap-Mitwirkende, SRTM | map display ©OpenTopoMap (CC-BY-SA) wise to identify them, starting with the AV01 which is the turbine at the top left corner, and AV12 which is the turbine at the bottom right corner. FINO 1 is located directly west of the AV04 turbine in a distance of 405 m. alpha ventus was the first wind farm installed in the area, but not the last. Around the site, several other wind farms were commissioned after alpha ventus (Figure 4.4 right) and they influence the inflow conditions of alpha ventus. With 60 MW total capacity alpha ventus is small in comparison to the wind farms around it. Direct neighbors of alpha ventus are Borkum Riffgrund I with 312 MW capacity and Borkum Riffgrund II with 448 MW capacity in the south-west, and Merkur with 396 MW capacity in the north-west. The closest distance between alpha ventus and the surrounding wind farms is around 2 km. The main wind direction at alpha ventus is south-west (Figure 4.4a), the mean wind speed at 91 m is 8.52 m s−1 (Figure 4.4b). The wakes of the surrounding wind farms in the main wind direction affect the power production of alpha ventus and lead to an increased turbulence intensity at the site [29]. https://www.openstreetmap.org/copyright https://opentopomap.org/#map=6/52.052/9.789 https://creativecommons.org/licenses/by-sa/3.0/ 28 4 First step in the forecasting chain: from radial velocity to wind field 2% 4% 6% WEST EAST SOUTH NORTH 0 - 2 2 - 4 4 - 6 6 - 8 8 - 10 10 - 12 12 - 14 14 - 16 16 - 18 >=18 Wind speed [m/s] (a) Wind rose 0 5 10 15 20 25 30 Wind speed [m/s] 0 0.02 0.04 0.06 0.08 0.1 0.12 F re qu en cy [- ] (b) Weibull distribution (c) Layout of alpha ventus. Reproduced with modifications from [30]. (d) alpha ventus sourrounded by other wind farms. Repro- duced with modifications from [31]. Figure 4.4: Wind characteristic in alpha ventus measured at FINO1 at 91 m (top) and layout of alpha ventus situated in the North Sea (bottom). 4.1 Measurement setup at the onshore and offshore site 29 Figure 4.5: StreamLine XR lidar on the top level platform of the radio tower. 4.1.4 Experiment layout design The lidar system that was used for the measurement campaign in VORKAST is a StreamLine XR pulsed doppler scanning-wind-lidar from the company Halo Photonics. The lidar was chosen because of its measurement range of 10 km and because of its light weight and compact form. The lidar was mounted on the top level platform of the radio tower (Figure 4.5) at a height of 736 m above sea level. The unobstructed view towards the main wind wind direction west/north–west was the reason for installing the lidar on the 91 m tower (Figure 4.3). Together with the lidar, a webcam was installed on the platform in order to take pictures of the view in westerly direction. The 102 m met mast was located in 1350 m distance westerly from the tower at 652 m above sea level and is fully equipped with meteorological sensors. Adjacent to the tower a reference turbine is located for which the minute-scale power forecasts are carried out. The hub height of the turbine is at the same level with the lidar mounting. Due to a confidentiality agreement with the turbine owner, data of this turbine will be displayed normalized by relevant parameters such as rated power. The met mast was operated by University of Stuttgart as part of the project LidarComplex to record the environmental conditions at the site. It is equipped with sensors at 5 m, 50 m, 75 m, and 98 m that record high resolution meteorological data. Wind speed and wind direction data from the mast are used in this thesis to verify the reconstructed wind speed from the lidar data (cf. Section 4.2.2). Other meteorological data from the mast such as temperature, relative humidity, and precipitation are used to assess environmental conditions at the site. This will 30 4 First step in the forecasting chain: from radial velocity to wind field Figure 4.6: StreamLine XR lidar on the nacelle of the AV04 turbine in alpha ventus. be relevant for the assessment of the analysis of the lidar measurement range and the possible forecast horizon (cf. Section 7.1). Details of the met mast’s sensor equipment are given in the appendix. The met mast was dismantled in August 2016, which means only a few months of concurrent lidar and met mast data were available. In alpha ventus, the same lidar system (StreamLine XR) was installed on the nacelle of the AV04 turbine (marked in Figure 4.4c). It was installed behind the rotor in a corner of the service platform of the turbine (Figure 4.6). The lidar was raised up on a 2.5 m high frame, to measure above the railing of the service platform and to have more clearance before the lidar beam hits the nacelle when measuring in a vertical pattern, e.g. with an Range Height Indicator (RHI) scan. It was decided to mount the lidar on top of a turbine - and not for example on the transition piece - as the lidar then rotates with the yaw angle of the turbine and faces the inflow direction. Another benefit of installing the lidar on the nacelle is that a horizontal measurement is automatically at the hub height and therefore can be directly used to assess the power production of the turbine. The turbine AV04 was chosen, as it is a turbine at the outside of the wind farm layout facing the main wind direction, and located directly opposite of the met tower FINO1. This configuration ensured that the lidar was able to measure directly the inflow of the wind farm and at the same time this inflow could be further characterised using the meteorological data from the met mast. FINO1 is a platform equipped with a met mast that measures wind speed, wind direction, air temperature, air pressure, precipitation and relative humidity at several heights up to a 4.1 Measurement setup at the onshore and offshore site 31 100 m [32]. Wind speed and wind direction are measured with cup anemometers and sonic anemometers. The wind speed data from the cup measurements are available with a mast correction, which account for lateral speed-up effects, upwind flow retardation and downwind wake effects due to the mast construction itself [33]. A table with all sensors and their respective measurement heights can be found in Table A.2 in the Appendix. The minute-scale forecasts at alpha ventus in this thesis are performed for the turbine AV04. Sensitive data of the turbine, such as the power curve are displayed normalized for confiden- tiality reasons. 4.1.5 Infrastructure planning Two factors are important to consider when measuring at remote sites: the power supply of the measurement device and the remote access to the device to change the settings and transfer data. During the campaign at Stötten the lidar was plugged into the power supply of the radio tower and remote data access was ensured via a modem. Measurement data from the lidar and met mast was automatically downloaded every night. New trajectories to test different configurations and methods for the forecasting were set via the remote access. The met mast was plugged into a nearby wind turbine and had access to the turbine’s internet connection. The data from the reference turbine was supplied by the owner. At alpha ventus, the lidar was connected to the turbine’s power supply and the remote access to the device was established by connecting the lidar directly to the network of the wind farm. Data was downloaded directly to the institute server every night. Access to the turbine data was established via the data portal hosted by BSH. A data user agremeement regulated the conditions between BSH and University of Stuttgart. It is important to note that the turbine data of the wind farms is not automatically uploaded to the data portal, but has to be handed over by the wind farm operator. This causes delays in access to the data of several months. The FINO1 data is available through another data portal hosted by the BSH, where just a registration but no user agreement is necessary. 32 4 First step in the forecasting chain: from radial velocity to wind field 4.1.6 Deployment and calibration procedures After initial tests, the lidar was mounted for the first campaign on the top level platform of the radio tower in Stötten in October 2015 and measured there until August 2017. To find out the exact device alignment, a step-by-step procedure was established in the beginning of the campaign to ensure that the exact position of the laser beam is well known. The procedure is described in [16] in detail and was not part of this thesis. The wind speed calibration of this device was part of this thesis in so far, as they are needed to reconstruct the horizontal wind speed using measurements from the radio tower. To test these methods, the results were compared to the horizontal wind speeds measured at the met mast. The details are explained in Section 4.2.2. For the second offshore campaign in alpha ventus, the lidar used in the VORKAST project was installed on the nacelle of the AV4 turbine in March 2019. Unfortunately, the device had a malfunction and did not measure more than around 1 km. Therefore it had to be decommissioned and sent to repair to the manufacturer. In the meantime an identical unit of a StreamLine XR lidar could be installed in the same spot and started measuring in October 2019. At the time of writing of this thesis, the measurement is still ongoing and planned until summer 2021. These data are available for the ParkCast project and this thesis. 4.1.7 Scanning modes design The StreamLine XR lidar is able to measure radial wind speeds along the line-of-sight direction of the emitted laser beam. The beam can be steered to any direction in the hemishere above the device and in an angle of 15° below. The steering angles are defined as elevation for vertical movements and azimuth for horizontal movements (cf. Figure 4.7). The measurement distance of the lidar is divided into range gates and as it is a pulsed device, the measurements are carried out simultaneously. The length of the range gates can vary from 18 to 60 m. A radial velocity is measured for each of the range gates. A sequence of emitted beams along a predefined azimuth and elevation angle is called a scan. For more technical device parameter, see Table A.1 in the appendix. The scanning modes carried out during the first onshore campaign had the goal to measure the horizontal wind speed component which defines a wind turbine’s power conversion and not the vertical component. Therefore only horizontal scans, with varying azimuth and zero degree elevation angle were carried out. The so-called Plan Position Indicatior (PPI) scans were directed westerly into the main wind direction. An overview of the different scans is given in Table 4.1. Each scan in the table is tagged with a scan ID and the time period, the number 4.1 Measurement setup at the onshore and offshore site 33 Horizontal plane Elevation Lidar xL zL (a) Side view Azimuth Lidar Range gate length Range gate xL yL (b) Top view Figure 4.7: Schematic drawing of the lidar measurement setup. of rays per horizontal scan, the number of range gates for each ray and its number of pulses, the azimuth angle of the scan and the time it takes to perform the scan is given. In the first few months of the campaign, only the number of pulses was varied, to test the influence on the measurement range (Section 7.1). It should be noted that along with the number of pulses, also the scan time varied, as these two parameters are directly linked. Later, also the range of azimuth angle was broadened. All scans were carried out in a step-stare mode, which means that the scan motors stop, and only move on after the measurement is carried out. For the offshore campaign a similar approach for the scan modes was chosen. The goal was to measure the horizontal wind component, therefore horizontal PPI scans were carried out. The range of the azimuth angle was set broad, in order to capture the inflow of the whole wind farm. The number of pulses was only changed once. An overview of the different scans in given in Table 4.2. 4.1.8 Execution and data collection During the onshore campaign, the lidar measured reliably on the radio tower platform from October 2015 to August 2017. Data from lidar and webcam was downloaded every few days via the modem connection or on regular maintenance checks on site. New scan modes were tested every few weeks. Data from the met mast was collected automatically every night. During the offshore campaign, the lidar measured from March 2019 to August 2019 with a malfunction and was decommissioned. The replacement lidar was installed in October 2019 and measured reliably from then on. The data was collected automatically every night. Only in June 2020, the lidar was shut down due to a wind farm shut down and could only be started 34 4 First step in the forecasting chain: from radial velocity to wind field ID Period No. No. range No. Azimuth angle Scan rays gates pulses time 1 28.04.2016 00:00 41 167 10,000 250°– 290° 1′12′′26.05.2016 23:40 2 07.06.2016 00:00 41 167 30,000 250° – 290° 2′30′′02.09.2016 12:20 3 02.09.2016 14:00 41 167 40,000 250° – 290° 3′10′′26.09.2016 08:20 4 26.09.2016 08:30 41 167 60,000 250° – 290° 4′27′′05.10.2016 08:20 5 05.10.2016 08:10 41 111 60,000 250° – 290° 4′27′′03.11.2016 10:00 6 19.01.2017 10:27 16 111 60,000 252° – 282° 1′42′′26.04.2017 14:10 7 26.04.2017 14:20 11 111 45,000 282°– 302° 0′52′′01.08.2017 00:00 Table 4.1: Overview of lidar scans at the onshore campaign in Stötten. Azimuth angle given in geographical coordinate system. Maximum measurement range always set to 10 km. ID Period No. No. range No. Azimuth angle Scan rays gates pulses time 1 14.01.2020 00:00 91 7980 20,000 240° – 60° 4′21′′27.02.2020 09:30 2 11.01.2021 15:42 71 400 40,000 260° – 40° 2′47′′11.01.2021 15:30 Table 4.2: Overview of lidar scans at the offshore campaign in alpha ventus. Azimuth angle given in geographical coordinate system. In the first scan, gate overlapping was turned on. Maximum measurement range always set to 12 km. 4.1 Measurement setup at the onshore and offshore site 35 again a month later. This was the only gap in the measurements. 4.1.9 Decommissioning and post-calibration procedures For the onshore campaign, the lidar was removed from the radio tower in August 2017 as the VORKAST project is finalized. Post-calibration procedures were not performed, as no met mast was available at that time. The met mast was decommissioned a year before in August 2016, as the project Lidar Complex finished. The offshore campaign is still ongoing at the time of this writing. It is planned to send the lidar for maintenance to the manufacturer, after decommissioning at alpha ventus. 4.1.10 Data availability An overview of available data used from the onshore and offshore campaigns is given in Table 4.3 and Table 4.4 respectively. Data is available from different sources: lidar, webcam, met mast, and turbine. The lidar data is available for both campaigns with a time resolution that depends on the scan configuration. For the onshore campaign, a webcam was available which recorded a picture every minute. For the onshore campaign, met mast data is available only until August 2016 but with a high time resolution which depends on the sensor. Offshore, FINO1 met mast data is only available as 10-minute mean values. The turbine data at Stötten are available as 10-minute averages for a period from July 2016 to December 2017. Available data are the turbine power, nacelle anemometer wind speed and the corresponding time stamp. For the AV04 offshore turbine, besides the above mentioned data the yaw angle of the nacelle is also available. For the alpha ventus wind turbine, also only 10-minute averaged data is available. 36 4 First step in the forecasting chain: from radial velocity to wind field Device Recorded signal Time resolution Available period Lidar Radial wind speed Scan dependent Oct. 2015 – Aug. 2017 Webcam Pictures in westerly direction 1-min Apr. 2016 – Aug. 2017 Met mast Wind speed 50 Hz Oct. 2015 – Aug. 2016 Wind direction 50 Hz Temperature 1 Hz Relative humidity 1 Hz Precipitation 1 Hz Turbine Power 10-min mean Oct. 2015 – Aug. 2016 /Nacelle wind speed 10-min mean Table 4.3: Overview of available data for the onshore campaign. More details on the met mast instrumentation is given in Appendix A.2. Device Recorded signal Time resolution Available period Lidar Radial wind speed Scan dependent Jan. 2020 – Jan. 2021 Met mast Wind speed 10-min mean Jan. 2020 – Sep. 2020 Wind direction 10-min mean Temperature 10-min mean Relative humidity 10-min mean Precipitation 10-min mean Turbine Power 10-min mean Jan. 2020 – Apr. 2021Nacelle wind speed 10-min mean Nacelle azimuth angle 10-min mean Table 4.4: Overview of available data for offshore measurement campaign. More details on the met mast instrumentation is given in Appendix A.2. 4.2 Getting a useful wind speed out of a lidar 37 4.2 Getting a useful wind speed out of a lidar Conventional ground-based, profiling lidar devices are prepared by the manufacturer and are ready for costumers to use. For applications such as wind resource assessment the relevant data are 10 minute averages of horizontal wind speed wind direction and turbulence intensity. In these cases, the lidar is treated as a black box and algorithms that are applied for the data processing are not modified. However, the StreamLine XR lidar that is used in this thesis requires customized data pro- cessing before the data can be used for power forecasting. 4.2.1 Data filtering The lidar measurement principle is based on the reflection of laser pulses on particles in the air which backscatter the light with a frequency shift due to the airspeed of the particles (cf. Section 2.3). The measurement depends on the existence of these aerosols. If the concentration in the air is too high or too low, or the laser energy of the scan is set too low, the device may measure an incorrect signal. Especially in far range gates, the backscattered signal intensity is often low. As a consequence lidar data needs to be filtered and the measurement range may deviate from the maximum range given in the product data sheet. When designing a filter for lidar data, filter requirements should be defined first, as the requirements vary from application to application. In this thesis the following requirements apply: 1. Conservative filtering with least possible data loss 2. Adaptability to varying environmental conditions 3. Highly efficient processing for real time capability. The requirements are specified for the application of minute-scale forecasting. Therefore as many corrupted data as possible are to be filtered out (conservative filtering), but at the same time the least possible amount of data should be lost in order to reach the maximum measurement distance. The filter should also be robust and work in varying environmental conditions, and in order to be able to use it for real time application of the forecast, the processing speed needs to be accordingly high. 38 4 First step in the forecasting chain: from radial velocity to wind field Carrier-to-Noise Ratio (CNR) filter The standard approach for detecting outliers and noise in the lidar raw data is to use the CNR, which is an indicator for the signal quality. The CNR is an output signal of the lidar device. It will be used in this thesis with its normalized unit in decibel [dB]. The CNR is saved for to every radial velocity measurement for every range gate. Very high CNR is the result of the beam hitting a hard target. A low CNR value can be the result of either low aerosol concentration at a certain gate and therefore not enough backscattered signal or the result of too low signal intensity reaching the gate. A typical timeline of unfiltered measured radial velocity V EL over the range of a typical measurement period is plotted in Figure 4.8a . The measured wind speed data in the first kilometers is coherent. After a transition range where outliers enter the homogeneous data, it becomes very noisy in the far measurement ranges. In the 30-hour time series that is shown, the sign of the wind speed changes from negative (wind speed towards LOS) to positive (wind speed away from LOS). From this, a change in wind direction from westerly to easterly can be deduced. The radial wind speed is plotted over the corresponding CNR sorted by the measurement range in Figure 4.9. The data in the first 1000 m start off as a cloud in a CNR range from -12 to 0 dB. This cloud then moves gradually towards lower CNR levels with increasing measurement distance. Between 3000 to 4000 m range the first wind speed outliers occur. These outliers increase with increasing distance and spread from -20 to 20 m s−1 until the former data cloud vanishes in the outlier spread. There is a CNR threshold at around −22 dB below which wind speed outliers occur. In CNR filter algorithms this threshold is used as a filter parameter. One can be sure that radial wind speed measurements that are tagged with a CNR above this threshold, are valid measurements. The only exception are data with CNR values above the threshold and a wind speed around 0 m s−1. These data can result from hard targets, i.e. solid objects. A standard approach to filter lidar data using the CNR values is to set the device specific CNR threshold and mark all wind speed data below the threshold as invalid. This approach is used for example for applications such as lidar-assisted control for wind turbines where short-range wind lidars are used, and the use of any incorrect lidar data must be avoided [20]. However, the disadvantage is that potentially valid wind speed data with CNR below the threshold (and therefore marked as invalid) are lost and the measurement range therefore is cut short unnecessarily (Figure 4.8b). The challenge for the application of minute-scale forcasting is therefore to implement a filter algorithm that removes only the incorrect wind speed data, and includes valid wind speed measurements although they might have low CNR. 4.2 Getting a useful wind speed out of a lidar 39 06:00 09:00 12:00 15:00 18:00 21:00 00:00 03:00 06:00 09:00 12:00 time [HH:MM] 0 2000 4000 6000 8000 10000 ra ng e [m ] -15 -10 -5 0 5 10 15 V E L [m /s ] (a) Unfiltered 06:00 09:00 12:00 15:00 18:00 21:00 00:00 03:00 06:00 09:00 12:00 time [HH:MM] 0 2000 4000 6000 8000 10000 ra ng e [m ] -15 -10 -5 0 5 10 15 V E L [m /s ] (b) CNR filtered; CNR threshold of −22 dB 06:00 09:00 12:00 15:00 18:00 21:00 00:00 03:00 06:00 09:00 12:00 time [HH:MM] 0 2000 4000 6000 8000 10000 ra ng e [m ] -15 -10 -5 0 5 10 15 V E L [m /s ] (c) Edge filtered; window size 3, ∆V EL limit > 3 m s−1 Figure 4.8: Typical time series of LOS lidar data unfiltered and filtered. 40 4 First step in the forecasting chain: from radial velocity to wind field Figure 4.9: Radial wind speed over CNR sorted by the measurement range. Edge detection filter The solution to filter lidar wind speed data comes from the area of image processing.1 When processing any image, chart or photo, often there is an edge detection algorithm involved. An edge within an image is classified as a significant local change in the image intensity, which typically occurs right on the boundary of two adjacent areas within the image [34]. For this reason “edge detection is frequently the first step in recovering information from images” [34]. Figure 4.10 shows a photograph of the Stuttgart TV tower on the left and the result of the edge detection filter on the right. Local changes in the color data are detected with the filter. Transferred to the application of wind speed filtering, an edge detection algorithm detects local changes in the wind speed by calculating the difference of maximum and minimum of radial wind speed ∆VEL within a predefined window. A window size of [1 3] for the edge filter means that three velocity values are included in the window from the same range gate of three neighbouring beams. The window then moves over all range gates and beams. Thus, a matrix with differences for each measurement point is stored. Afterwards it is checked whether the differences ∆VEL exceed a predefined ∆VEL threshold. If this is the case for both adjacent values of a measurement point, the corresponding measurement is marked as invalid. This results in a logical matrix which contains an entry for valid or invalid measurement points for all range gates and beams. It is important to note that by checking ∆VEL of adjacent measurement points, only the outliers are marked as invalid, and not necessarily the neighbours. To demonstrate the functionality of the filter, an example of a generic wind speed time series 1This section is based on the Bachelor thesis "Adaptive filtering of long range lidar data" carried out by Malte Justus Niemeier, which was handed in at SWE in 2016 and supervised by the author of this PhD thesis and her colleague Maayen Wigger. 4.2 Getting a useful wind speed out of a lidar 41 Figure 4.10: Image of the Stuttgart TV tower (left) and edge detection filtered (right). Left image reproduced with modifications from [35] with permission. is given in Figure 4.11. The Figure also gives the wind speed difference ∆VEL for a window size of [1 3]. For the first and last beam in the scan, the left and right values next to the center point of the moving window are ignored. For a ∆VEL threshold of >2 m s−1, the measurement points from beam 1, 5 and 6 would be marked as invalid in this example. For a wind threshold of >3 m s−1 only beam 1 would be excluded. The filter parameters that are relevant for the performance of the edge filter are the window size and the wind speed limit ∆VEL. To test filter requirement 1, a parameter study is carried out, to find out how many data remain after filtering with the edge filter for different window sizes and wind speed limits (Figure 4.12). This study was carried out for the time series in Figure 4.8a. The results show that for a window size of 3 adjacent data points (smallest possible window), the least amount of data are filtered. By increasing the window size to 5, around 10 % more data are filtered. Increasing the window more, leads to more data loss. The wind speed limit ∆VEL also has a significant influence on the data availability after filtering. Setting the limit to 0.5 m s−1, thus allowing only very small wind speed fluctuations, leads to a rigorous filtering of the data. When increasing the limit gradually, the amount of available data increases rapidly at first, and then evens out. Figure 4.8c shows the result of the edge filter for a window size [1 3] and a wind speed limit ∆VEL of >3 m s−1. Compared to the CNR filter (26 % data availability), much less data is discarded with the edge filter (65 % data availability). Hence, the measurement range is increased significantly. Wind speed values with a CNR below the threshold of −22 dB are then 42 4 First step in the forecasting chain: from radial velocity to wind field moving window ΔVEL 4 4 2 3 6 6 3 1 1 Figure 4.11: Demonstration of edge filter for a generic wind speed time series for window size [1 3] and ∆VEL threshold of > 2 m s−1. Filtered values marked grey. 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 VEL limit [m/s] 0 10 20 30 40 50 60 70 P er ce nt ag e of v al id p oi nt s af te r fil te rin g 3 5 7 9 11 window size Figure 4.12: Percentage of good points after filtering with edge filter with different ∆VEL limits (thresholds) and window sizes. Study carried out for time series in Figure 4.8 4.2 Getting a useful wind speed out of a lidar 43 Figure 4.13: Radial velocity over CNR filtered with the edge filter. “rescued” when applying the edge detection filter with those filter parameters (Figure 4.13). Although some outliers remain in the data, it is decided to apply the edge detection filter with a window size [1 3] and a wind speed limit ∆VEL of >3 m s−1, for the sake of the increased measurement range for both the onshore and offshore data. The edge filter can adapt to varying environmental conditions such as the wind direction change, and is very simple so that it is computationally highly efficient. Thus all the filter criteria are fulfilled. 4.2.2 Wind field reconstruction A lidar measures the radial component V EL of the wind vector in the laser beam direction LOS, in contrast to classical wind measuring systems such as a cup anemometer that measure horizontal wind speed. The challenge is to derive the horizontal wind velocity from the measured radial wind velocities of the discrete measuring points of the lidar scan; this is known as wind field reconstruction. The basic equation to reconstruct the three wind components u, v, w from the radial compo- nent is V EL = xi di ui + yi di vi + zi di wi. (4.1) where the lidar measures at the coordinate point [xi yi zi] in a distance di the unknown wind vector [ui viwi] [5]. At least three measurement points are necessary to solve this equation. 44 4 First step in the forecasting chain: from radial velocity to wind field Therefore assumptions must be made. For example assuming a constant and homogeneous wind field during the