Abstract Electric load forecasting serves as a crucial decision-making foundation for the operation of power systems, offering important guidance for system scheduling optimization, cost reduction, and capacity planning. With the increasing penetration of new power electronic devices, such as electric vehicle charging stations, load volatility and nonlinearity have become more pronounced. At the same time, the scarcity of historical data for newly added loads often leads to challenges in achieving high forecasting accuracy. This paper proposes a hybrid deep learning model for mid-term electric load forecasting based on BKA optimization and CWGAN-GP data augmentation. First, the original load dataset is expanded through data augmentation using CWGAN-GP. Then, CNN-BiTransformer-BiLSTM and BiTCN-BiGRU-Attention models are independently trained on similar semi-monthly (14-day) datasets, producing two sets of validation prediction vectors. Subsequently, these vectors are combined with the validation set and used to train the final model via XGBoost. Finally, the features of the target semi-month are fed into the final model to obtain the forecasted results. Experimental results across three groups demonstrate that the proposed model reduces RMSE by at least 3.02% and lowers MAPE by at least 17.7%, validating its superior accuracy and robustness.
Luo et al. (Wed,) studied this question.