Long Short-Term Memory Recurrent Neural Network for Estimating State of Charge of Energy Storage System for Grid Services

Existing methods of state of charge (SOC) estimation have limitations such as requiring an accurate battery model or frequent calibration, making them unsuitable for energy storage system (ESS) applications. These limitations can be overcome using machine learning (ML) techniques. Among ML-based techniques, long short-term memory (LSTM) has a feedback-loop characteristic, which considers current and historical SOC. At present, LSTM has not been used for batteries in ESS during peak demand reductions. Therefore, this paper studies the use of LSTM for battery SOC estimation in ESS during peak demand reductions. LSTM is found to be effective even though the network begins with an inaccurate SOC, which makes it a suitable technique for batteries in ESS for grid services. The estimation accuracy of the LSTM is also higher than conventional SOC estimation techniques such as coulomb counting and empirical model, and other ML techniques such as the feedforward neural network.