A multi-source feature engineering-enhanced framework for mid-to-long-term EV charging load forecasting: Integrating self-adaptive optimization and BiLSTM-iTransformer predictor

The unordered and spatiotemporally uneven charging behavior of Electric Vehicles (EVs) poses a significant threat to the stable operation of urban power grids, especially with large-scale EV fleet charging. To address this challenge, a multi-source feature engineering-enhanced framework is developed for mid-to-long-term EV charging load forecasting, integrating self-adaptive optimization and a BiLSTM-iTransformer predictor. The framework's feature engineering module leverages Robust Principal Component Analysis (RPCA) to separate underlying data patterns from noise and sparse anomalies, combined with the Mean Impact Value (MIV) method to robustly extract key features and quantify the contributions of multi-source inputs like user behavior and weather. At the core of the framework is a hybrid predictive model that integrates a Bidirectional Long Short-Term Memory (BiLSTM) network, adept at capturing local short-term dependencies, with an iTransformer architecture, renowned for its efficiency in modeling long-term global correlations. Furthermore, the Tuning-Enhanced Rime Optimization Algorithm (TE-RIME) is employed as the self-adaptive optimization component to dynamically tune the hyperparameters of the BiLSTM-iTransformer model, ensuring optimal performance through its enhanced exploration and exploitation mechanisms. Experimental results across multiple EV charging station datasets from different cities demonstrate that the framework effectively extracts key features, achieving superior prediction accuracy and generalization ability compared to benchmark methods. This provides a reliable tool for power grid planning and operation, facilitating a low-carbon energy transition.