A novel net load prediction approach using multi-scale deep learning network based on Trend-Multi-Period encoder and segment imbalance regression

With the increase of the proportion of new energy penetration, the source-load balance and stable operation of the new power system depend on more accurate prediction. In this study, a triple-stage multi-step net load forecasting approach is proposed by applying Trend-Multi-Period encoder (TMPE) with multi-scale convolutional neural network (MSCNN) and segment imbalance regression (SIR) with crisscross optimization (CSO). In the information extracting stage, TMPE is proposed to decouple the complex net load characteristics to obtain trend information and multi-period information. In the prediction stage, MSCNN, which can map the time series into a picture-like multi-channel structure, is utilized to mine the global and local information of upstream features. In the third stage, CSO-based SIR is applied to alleviate the negative impact caused by data imbalance in view of the high prediction error of net load in high and low power segments, where CSO maximizes the performance of SIR. The model proposed in the one-step prediction, compared with other baseline models, the maximum reduction in RMSE and MAE is 60.06 % and 66.58 %, respectively, and the R2 is as high as 98.85 %. Extensive experiments demonstrate the superiority of the proposed method over other state-of-the-art methods.