Multi-scale Dilated Convolutional Residual Network based Model for Short-Term Load Forecasting

Short electricity load forecasting has a significant impact on both the interests of power generation enterprises and the stability of regional power grid supply. Due to the significant nonlinear feature of load data and the combined effect of multiple factors, existing deep learning methods have difficulty capturing the changes in power load, thereby limiting the forecasting performance of the method. To overcome these limitations, the paper proposes the CEEMDAN-SE-MDCRN prediction model. Firstly, CEEMDAN-SE-MDCRN uses Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose the original load to enhance the extraction capacity of the intrinsic characteristics of the load data. Next, Sample Entropy (SE) is adopted to calculate the complexity of different components and group the components with similar complexity. Then, the Multi-scale Dilated Convolutional Residual Network (MDCRN) is adopted to extract the spatial information of the load after grouping to produce the final output. The simulations demonstrate that the CEEMDAN-SE-MDCRN method can achieve better generalization ability compared with existing deep learning methods on the New England dataset.