Improved NaS Battery State of Charge and State of Health Estimation: A Novel Integration of Temporal Fusion Transformer, Isolation Forest, and Support Vector Regression

Abstract

Precise State of Charge (SOC) and State of Health (SOH) are crucial for the effective operation and longevity of Sodium-Sulfur (NaS) Battery Energy Storage Systems (BESS). Real-time knowledge of SOC allows for optimal discharge planning and prevents over-discharging, while a reliable SOH estimate facilitates preventive maintenance and diminishes unexpected system failures. This paper proposes a data-driven approach to address the need for robust SOC and SOH estimation in NaS BESS. The proposed framework utilizes machine learning techniques for precise SOC and SOH estimation in NaS BESS, essential for integrating renewable energy sources into the electrical grid and deriving valuable insights into battery health and capacity. Unique challenges associated with NaS batteries, such as the significant hysteresis effect, demand sophisticated estimation techniques. To address this, a Deep Neural Network (DNN)-based Temporal Fusion Transformer is employed for SOC estimation, yielding an exemplary R-square value of 0.997, thereby surpassing the performance metrics of conventional Recurrent Neural Network/Long Short-Term Memory (RNN/LSTM) and Gated Recurrent Units (GRU) architectures. For the estimation of SOH, a dual-strategy method is implemented, using support vector regression (SVR) coupled with an Isolation Forest model to facilitate the prediction of various operational cycles and enhance anomaly detection capabilities. The proposed approaches not only demonstrate superior accuracy in SOC and SOH estimation but also establish a robust framework for comprehensive assessment in NaS BESS. The findings of this study contribute to the advancement of battery management systems, which support the sustainability and reliability of renewable energy-rich power grids.