Kalman filter anomaly values processing meta-model ensemble learning framework for Lithium-ion battery capacity prediction

To address the low prediction accuracy caused by single-view in Lithium-ion battery capacity prediction, a Kalman filter anomaly values processing-based meta-model ensemble learning (KF-AVP-M2EL) framework is proposed using block resampling and multi-activation extreme learning machine (BR-MA-ELM). First, BR-MA-ELM addresses the feature forgetting problems caused by time-series interference of online sequential extreme learning machine (OS-ELM) through block resampling. Second, BR-MA-ELM introduces a nonlinear activation function set (NLAFS), overcoming the limited nonlinear expression capability of OS-ELM. Third, during meta-dataset construction, the Kalman filter (KF) initializes the model with the capacity degradation data, records anomaly values and their indexes from health indicators (HIs) predictions, and filters corresponding results from another view, effectively reducing interference. Finally, BR-MA-ELM is employed as the meta-model, the anomaly values saved from the HIs view are then reinserted according to their corresponding indexes, preserving HIs’ ability to capture local changes while integrating the overall degradation trend from the capacity degradation data view, thereby comprehensively capturing the multidimensional characteristics of battery aging and ensuring more accurate predictions. In the comparison experiment, KF-AVP-M2EL achieved an RMSE of 0.0034, an MAE of 0.0026, and an MAPE of 0.02, outperforming four state-of-the-art algorithms. Ablation experiments further validate the superiority of the proposed algorithm.