Evaluating Machine-Learning-Based Methods for Modeling a Digital Twin of Battery Systems Providing Frequency Regulation

Abstract

Battery energy storage systems (BESSs) have the potential to become providers of frequency regulation services in future energy systems. In this case, the reliable operation of BESS has a significant impact on system stability. One of the tools that ensures the secure BESS operation is applying a digital twin to forecast the BESS state of charge (SOC). It allows to predict BESS behavior and identify potential failures or cyberattacks in BESS. In this article, we compare multiple data-driven approaches to forecast the SOC of a BESS based on realistic dataset and real battery operation. Recurrent neural networks, e.g., feedforward neural network, gated recurrent unit, long short-term memory, support vector regression, random forest, and AdaBoost methods are applied to evaluate each method's performance. We compare the methods based on the maximum and average forecast errors, the required computational capacity, as well as expected training speed. To validate the results, in addition to the data gathered from simulations, we used experimental setup with a BESS of 79 kWh providing containment reserve for normal operation, which is a BESS service with a high economical potential in Nordic Region. Furthermore, we provide recommendations for the methods that are suitable to be applied for modeling the digital twin of a utility-scale battery digital twin.