A Domain-Knowledge Guided Grey Wolf Optimized Random Forest framework for state of charge estimation under dynamic thermal and loading conditions

Accurate state of charge (SOC) estimation is fundamental to safe lithium-ion battery operation in electric vehicles and challenging under varying temperature and load. This study presents a Domain-Knowledge Guided Grey Wolf Optimized Random Forest (DGWO-RF) framework in which temporal and signal features engineered exclusively from voltage, current, and temperature, including filtered and differential voltage, power, and squared current, drive a bagged regression-tree ensemble tuned by Grey Wolf Optimization. Evaluation uses a public drive-cycle dataset spanning 5–45 °C. Under condition-specific calibration within one cycle–temperature condition, average root mean square error (RMSE) is 1.14% on the US06 cycle, rising to 1.53% at 5 °C; this accuracy does not extend to untrained temperatures, and a blocked-split analysis shows the sample-wise protocol to be optimistic by up to a factor of two. Under a leakage-free cycle-wise partition withholding complete drive cycles measured on separate cells, augmenting the inputs with trailing moving averages over four timescales reduces transfer error by 73% and systematic offset by an order of magnitude, yielding 0.62% RMSE and 0.46% mean absolute error on the withheld FTP-75 cycle at 25 °C. At equal budget over five seeds, the choice among Grey Wolf, particle swarm and Bayesian optimization has little effect on accuracy; random search is less consistent than all three. The results indicate that, within the calibrated operating range, temporal feature support and search-space design have a substantially greater influence on estimation accuracy than the specific metaheuristic used for hyperparameter optimization.