Neuroplastic Fuzzy GCN-LSTM: A Hybrid Graph-based Deep Learning Model for Accurate Battery SoC Prediction

Abstract

Accurate prediction of the State of Charge (SoC) in batteries is critical for the efficient and safe operation of Electric Vehicles (EVs). In this work, a novel hybrid neural architecture called Neuroplastic Fuzzy GCN-LSTM is proposed for accurate SoC prediction. It involves fuzzy logic, dynamic graph modelling and graph convolutional networks (GCNs) to effectively handle the complex patterns of battery data. Initially, time-series sensor inputs of voltage, current, and temperature are normalised and transformed into fuzzy linguistic representations to address uncertainty and improve interpretability. Then, a dynamic graph is constructed to capture inter-feature dependencies via a Gaussian similarity kernel. These graphs are processed through spectral GCN layers to extract spatial correlations. Finally, the neuroplastic LSTM (NP-LSTM) is applied for SoC prediction. Unlike conventional LSTMs, the NP-LSTM adaptively modulates its memory cell updates using error-based synaptic plasticity, allowing the model to emphasise learning from past prediction errors. The performance of the Neuroplastic Fuzzy GCN-LSTM model is verified using NASA’s Prognostics Center of Excellence data sets in terms of Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE) and Coefficient of Determination (R²) compared to models. The Neuroplastic Fuzzy GCN-LSTM model achieved the best performance among all competing models, with the lowest MSE (0.009378), MAE (0.086234), and RMSE (0.096839).