Adaptive SOC estimation of grid-level BESS for multiple operational scenarios
摘要
Accurate estimation of the state of charge (SOC) is a key technical foundation for ensuring the efficient and safe operation of battery energy storage systems (BESS). Current research primarily focuses on electric vehicle power batteries, which may not be well-suited to the dynamic characteristics of grid-scale BESS, particularly in handling multi-scenario and multi-time-scale coupling in operations such as frequency regulation and peak shaving. This paper proposes a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM)-based adaptive SOC estimation framework that integrates scenario decoupling and multi-time-scale feature fusion. First, a dynamic threshold current decoupling algorithm, adapted to different operational scenarios, is designed to effectively separate the frequency regulation and peak shaving scenarios. Next, a multi-branch deep network structure is developed, where a dual-convolution kernel CNN module extracts multi-time-scale local features from short-term frequency regulation signals, while a multi-branch LSTM module captures long-term dependencies in the hourly to daily trends of peak shaving. Finally, the dynamic fusion of multi-time-scale features is achieved through an attention mechanism layer. Experimental results show that the current condition decoupling significantly improves the accuracy of SOC estimation. Compared to single-times-cale neural networks, the proposed multi-time-scale model reduces the mean absolute error, maximum error, and root mean square error by 26.15%, 23.56%, and 26.38%, respectively. The proposed model effectively accounts for both the long-term trends in peak shaving and short-term fluctuations in frequency regulation, enabling adaptive SOC estimation under different operational scenarios.
