A Forecasting Method of Distributed Load Based on Residual Enhanced U-Net Structure

To overcome the inadequate prediction accuracy caused by the inherent nonlinearity and stochasticity of distributed loads (DL), a forecasting method of DL is proposed based on residual-enhanced U-Net structure. Firstly, the initial data undergoes loading and cleaning procedures, followed by temporal feature engineering and standardized preprocessing, to generate the formatted input dataset for ResU-Net. Subsequently, symmetrically structured encoder-decoder blocks with residual skip connections generate long-term forecasts, enhancing stability by mitigating feature degradation in deep layers. Finally, comparative experiments were conducted on a continuous three-year distributed load dataset, benchmarking against LSTM, Transformer, and TCN models. Experimental results demonstrate that the proposed method achieves a reduction of 18.39% in MAE, 16.96% in RMSE, and an improvement of 17.53% in R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, respectively. Additionally, it reduces forecasting bias by 15.17% during peak load periods, validating the effectiveness of the proposed methodology for distribution load forecasting tasks.