Eliminating backpropagation in electric vehicle battery management: A non-iterative liquid gated echo state network for SOC prediction

While deep learning models exhibit high accuracy in Electric Vehicle (EV) Battery State of Charge (SOC) forecasting, their reliance on iterative Backpropagation Through Time (BPTT) incurs massive computational costs, hindering real-time deployment on embedded Battery Management Systems (BMS). To bridge this gap, a novel backpropagation-free architecture, termed the Liquid Gated Echo State Network (LG-ESN), is proposed with innovations in a three-step manner. First, the architecture decouples temporal dynamics from learning via a dual-reservoir initialization, which calculates dynamic input-dependent leak rates without requiring weight updates. Second, by isolating these complex liquid time-constant equations within a randomly generated frozen reservoir, the model eliminates BPTT and the associated vanishing gradient problem. Third, the actual learning phase is reduced to a single Ridge regression readout layer, ensuring faster convergence to a mathematical optimum. This novelty grants the LG-ESN the rich temporal memory of deep recurrent networks while guaranteeing deterministic global convergence. Experimental validation over 15 independent random initializations demonstrates that the LG-ESN achieves an R 2 of 0.9315 ± 0.0041, an RMSE of 2.9923 ± 0.0878 and an MAE of 2.3770 ± 0.0395, outranking five competing architectures in every single run. A Friedman test followed by a Holm-corrected post hoc analysis confirms that these margins are statistically significant for all three metrics (adjusted p < 0.01). This indicates that the LG-ESN is a strong candidate for BMS implementation.