Enhancing Memory-Limited Feedforward Neural Networks for State of Charge Estimation through Temporal Feature Engineering

Abstract

State-of-charge (SOC) estimation is a key function of Battery Management Systems (BMS) in electric vehicles and battery energy storage systems. However, SOC is not directly measurable, making accurate estimation inherently challenging. Data-driven approaches offer a practical solution by leveraging measurable inputs such as voltage, current, and temperature. Feedforward Neural Networks (FNNs) are attractive due to their low computational complexity, but they lack inherent temporal memory, unlike recurrent architectures. This paper investigates three established casual smoothing techniques-moving average, Butterworth filtering, and exponential moving average-as temporal memory proxies for enhancing FNN-based SOC estimation. Their effectiveness is supported by frequency-domain analysis using the Fast Fourier Transform (FFT), which reveals that key signal dynamics occur at ultra-low frequencies (less than 0.1 mHz), justifying the use of smoothing as memory-preserving transformations. The main contribution of this work is a unified, frequency-informed evaluation framework that systematically benchmarks these techniques under consistent conditions and across varying temperatures. All models are trained and evaluated on LG 18650HG2 Lithium-ion battery data, with 20 repeated runs per model to ensure statistical robustness.