Energy consumption forecasting in energy management systems

Abstract

Energy Management System (EMS) is emerging as a solution for the power grid management, which is indicated by the profit, stability, and reliability to its end users. Load forecasting is a component of an EMS, and it is vital to foresee the future demand to generate a proper schedule for a microgrid. A load forecasting study should cover all types of loads: commercial buildings and households, which has been incompletely studied in other papers. Since 2018, with the improvement in computing power, Deep Learning (DL) models have caught attention in many studies, and shown their potential over the traditional method Auto Regression Moving Average (ARIMA) in solving forecasting problems. In the DL branch, Recurrent Neural Network (RNN) is the type of model having memory, which is designed to estimate future values on history data. This thesis studies the application of RNN in forecasting energy consumption of both load types and implements a service that forecasts demand focusing on commercial buildings. We experiment with six different RNN-based models on two sets of data, commercial buildings in 2004 in San Francisco and households between 2011 to 2014 in London. The performance of the RNN family significantly exceeds ARIMA in accuracy and processing time. In two study cases, the reported error of the best performer in RNN-based models is 48% and 23% lower than ARIMA. The Forecasting service for commercial buildings is then implemented as a restful Application Programming Interface (API) for further deployment in an EMS.