Artificial neural network architectures for state estimation in lithium-ion batteries

Abstract

This dissertation investigates the application of artificial neural networks for predicting the state of charge, state of health, and temperature of lithium-ion battery cells. The study evaluates several model architectures, including encoder-based models, informer-based models, and transformer-based approaches (collectively referred to as attention-based models), as well as long short-term memory networks. The evaluation considers key aspects such as model size, complexity, and reproducibility. For state of charge and temperature predictions, the training data consists of charge cycles and worldwide harmonized light-duty vehicle test procedure cycles. A novel training algorithm was developed and consistently applied across all models to ensure comparability. Encoder-based models are systematically analyzed, focusing on architectural features such as positional encodings, normalization layers, and input scaling strategies, as well as autoregressive methods like artificial feature extraction and artificial recurrence. Similarly, both stateful and stateless long short-term memory models are evaluated for their robustness and predictive power. Among the attention-based models, the encoder-only architecture, incorporating positional encodings, post-normalization layers, and zero-importance scaling for current data, achieved notable results. However, long short-term memory models, particularly in stateless configurations, consistently outperformed encoder-based models in state of charge predictions, with the best-performing long short-term memory model achieving a root mean square error as low as 0.744 percent. This demonstrates that long short-term memory networks are more robust and effective for time series forecasting in the context of state estimation for lithium-ion battery cells. The challenges of temperature prediction are highlighted by limitations in the quality of the training data, which significantly impacted model performance. While increasing the number of trainable parameters improved model accuracy to some extent, these improvements eventually plateaued, emphasizing that data quality plays a more critical role than model complexity. Long short-term memory models exhibited more consistent performance, whereas encoder models were more variable, underscoring the importance of high-quality data in real-world applications. For state of health predictions, a novel classification model was developed by filtering out current pulses during complex operations, such as drive cycles, which were then analyzed by the artificial neural network to estimate the state of health. Long short-term memory models once again demonstrated superior performance and stability compared to encoder-based models. The findings reveal that long short-term memory networks remain a highly effective approach for state estimation, outperforming attention-based models in both state of charge and state of health predictions. These results underscore the potential of artificial neural networks in battery management systems while identifying key factors, such as data quality and architectural decisions, that significantly influence model performance.