Remaining useful life prediction of lithium battery based on deep reinforcement learning fusion network

Lithium batteries are the most commonly used energy storage devices, and predicting their remaining useful life (RUL) is crucial. But in existing methods, single-model structures generally struggle to parallelly process the spatio-temporal correlations of features. Although multi-model fusion can enhance prediction accuracy, accurately allocating contributions from different models remains challenging. Therefore, this paper proposes a Deep Reinforcement Learning Fusion Network (DRLFN) for predicting the RUL of lithium batteries. This method initially integrates LSTM and Transformer for preliminary RUL predictions of lithium batteries. The prediction results are then fused and optimized using a Deep Q-Network (DQN). The process distinguishes regular and irregular prediction regions based on the spatial distribution of prediction errors in the validation set. For regular regions, common prediction bias patterns are identified and divided into sub-intervals, with a DQN exploring optimal model weight combinations within each sub-interval. For irregular regions, credibility scores based on individual model prediction errors are assigned, and the model with the higher score is selected for each prediction. Experiments utilize the lithium battery dataset from the MIT-Stanford-Toyota Research Center. By comparing the performance with mainstream models, the effectiveness and superiority of DRLFN for predicting the RUL of lithium batteries is validated.