Improved NaS Battery State of Charge Estimation by Means of Temporal Fusion Transformer

Abstract

The stability of the grid is being challenged by the increasing penetration of renewable energy resources for green and diversified power generation, as well as for the reduction of CO2 emissions. The installation of battery energy storage systems (BESS) to support intermittent and variable renewable energy generation is a promising solution and Sodium sulfur (NaS) batteries have shown an outstanding performance for energy-intensive and high utilization BESS solutions due to their low cost and long lifecycles. The ability to accurately estimate the battery’s state of charge (SOC) at all potential utilization scenarios is a critical element of the effective BESS operation. Despite the considerable number of studies available on the SOC estimation of lithium-ion BESS for improved battery management systems (BMS), there is scarcely any literature about the challenges and methods for the SOC estimation of NaS batteries. This work highlights the challenge of reliable SOC assessment in NaS BESS by means of a pulse charge/discharge test, introduces a methodology to refine the SOC values collected from an associated BMS / SCADA system and proposes a data-driven approach for the corresponding SOC estimation using a Temporal Fusion Transformer model. After the application of hyper-parameter tuning, this state-of-the-art deep learning (DL) model demonstrates an R-square (R2) value of 0.997, which is superior to the R2 of 0.987 achieved by a recurrent neural network / long short-term memory (RNN/LSTM) DL architecture.