Long Short-Term Memory Approach to Predict Battery SOC

Abstract

Estimating the ‘State of Charge’ (SOC) is a complex endeavour. Data-driven techniques for SOC estimation tend to offer higher prediction  accuracy compared to traditional methods. With the progression of Artificial Intelligence (AI), machine learning has found  extensive applications across various fields such as infotainment, driver assistance systems, and autonomous vehicles. This paper  categorizes the machine learning techniques utilized in Battery Management System (BMS) applications and employs a modern  supervised neural network approach to predict SOC. Accurate SOC estimation is crucial to prevent battery failures in critical situations,  such as during heavy traffic or when traveling with limited access to charging stations. Long Short-Term Memory (LSTM) networks are  particularly adept at classifying, processing, and predicting based on time series data. These models are capable of capturing and retaining features over time, making them suitable for this study. The model's predicted SOC closely matches the true SOC, and the SOC  prediction error remains nearly zero even with a large sample of input data.