Optimization framework for multi-objective energy management strategy in hybrid electric vehicles integrating explainable artificial intelligence

This study presents a multi-objective optimization (MOO) and explainable artificial intelligence (XAI) integrated framework for establishing MOO strategies in hybrid electric vehicle (HEV) and analyzing the decision-making patterns of the system. Specifically, the study first establishes a power-split HEV model and incorporates fuel consumption, electricity consumption, battery degradation, and the number of engine start-stops into the system evaluation metrics. Subsequently, the dynamic programming (DP) algorithm is employed to develop an offline globally optimal multi-objective strategy based on the multi-indicator vehicle model, and the non-dominated sorting genetic algorithm-II (NSGA-II) is used to perform MOO of the strategy's cost function. By combining the cognition-driven analytical hierarchy process (AHP) decision-making method with the data-driven technique for order preference by similarity to ideal solution (TOPSIS) method, the AHP-TOPSIS decision-making method is used to select the optimal solution from the Pareto frontier. Tree-based XAI methods are introduced, employing mean decrease impurity (MDI) and partial dependence plots (PDP) to analyze the interaction mechanisms among the four objectives during the decision-making process. A double-layer random forest (RF) energy management strategy is constructed, combining a five-fold cross-validated RF pattern recognition model with an engine power prediction model based on the optimal solution dataset. The results demonstrate that the multi-objective strategy exhibits better overall performance compared to single-objective strategies. The proposed double-layer RF strategy reduces fuel consumption by 1.8 %, maintains similar electricity consumption, decreases battery degradation by 7.1 %, and reduces the number of engine start-stops by 82.1 % compared to a rule-based (RB) strategy, with minimal deviation from the DP strategy. This validates the superior performance of the strategy in multi-objective control processes.