Short-term Electrical Load Forecasting Based On Deep Residual Neural Networks

With the highly penetrated flexible loads in distributed level,the diversified and stochastic energy consumption patterns brought a huge uncertainty and risks of system operation,which poses an additional challenge to the Short-Term Load Forecasting（STLF）.Accurate STLF can direct power operators to develop generation plans,reduce operating costs and achieve a supply-demand balance.To address the problems of accuracy,reliability and computational efficiency that are inherent in traditional STLF techniques,thesis focuses on STLF techniques for several scenarios based on deep residual networks.（1）A relatively smooth multivariate time series is obtained by decomposing the nonstationary original load series based on the Redundant Haar Wavelet Transform（RHWT）.Then an iResNet is proposed to learn the decomposed RHWT sequences.The residual connections solve the problems of gradient disappearance and model degradation during the training process,a Prelu activation function and Gaussian noise are introduced to improve generalizability and robustness.（2）Considering the uncertainty of load forecasting,thesis proposes two probabilistic forecasting strategies:ⅰ)A Quantile Score（QS）-based quantile regression strategy.Compared with the traditional quantile regression neural network,this method leverages training strategies such as batch training,early stopping and learning rate decay,which effectively reduces the training cost and constructs a reliable prediction interval;ⅱ)A probability density estimation strategy based on composite kernel function.A sparse Gaussian Process with a composite kernel function is designed to fit the residuals between predicted and actual values based on the RHWT-iResNet model.This approach can further improve the point load forecasting and quantify the uncertainties of future loads.（3）To address the STLF problem with small sample sets,thesis proposes a transfer learning based STLF method with limited training samples.Firstly,iResNet is used to learn knowledge in the source domain and embedded them to a series of parametric models,then with limited training samples of the target domains,the Bayesian Weighted Probability Averaging（BPWA）method is proposed to adaptively combine the transfer models to search the optimal forecasting models.Besides,the proposed pre-trainedadaptive combination framework also provides alternative methods for the application of transfer learning in regression tasks.（4）In the case study,public data sets from the low voltage side and households are selected for validation.Several evaluation indicators are developed to verify the effectiveness of the proposed method.