Seq2Seq regression learning-based multivariate and multistep SOC forecasting of BESS in frequency regulation service

Abstract

Battery Energy Storage Systems (BESS) has a fast response time and high ramping capability, making it suitable for smart grid frequency regulation service. However, providing the required regulation capacity without a problem at any point in time is bound with the BESS state of charge (SOC) development periodically. Moreover, knowing SOC variation in advance will help the BESS producers keep SOC at the desired level. The BESS SOC can be maintained in the admissible operation area to avoid the mismatch penalties while following regulation signals. Therefore, the maximum service reward can also be achieved. Unfortunately, forecasting the SOC of BESS in frequency regulation service is not a straightforward problem. The forecasting method should deal with multiple dependent variables that periodically determine SOC’s development. Moreover, developing a one-time manner multistep SOC forecasting model is also a challenge. To solve both problems, a sequence-to-sequence (seq2seq) regression learning architecture that has been proven to deal with sequentially interdependent data is adopted in our proposed forecasting framework. Various state-of-the-art memory cells in deep regression learning, i.e., long short-term memory (LSTM), gated recurrent unit (GRU), bi-directional (bi)-LSTM, and bi-GRU, were utilized and evaluated. The evaluation result shows that the developed models outperform the existing machine learning-based forecasting methods. The Bi-GRU cells provide the best performance in root mean square error (RMSE) and mean absolute error (MAE) evaluation metrics.