Research On Short-Term Electric Load Forecasting Based On Feature Set And Temporal Convolutional Network

Highly accurate short-term power load forecasting is conducive to the stable operation and real-time dispatch of power systems and the maintenance of high power quality,and also provides some guidance for China to strive to achieve the "double carbon" goal of carbon peak by 2030 and carbon neutrality by 2060.However,due to the randomness and non-linearity of power load and the influence of various environmental factors,the structure and characteristics of the obtained raw power load data are unclear,which will greatly increase the difficulty of power load forecasting and make it difficult to obtain more accurate forecasting results.Therefore,in order to achieve higher accuracy in short-term power load forecasting,this paper proposes a short-term power load forecasting model based on feature set and temporal convolutional network by integrating various deep learning theories and methods.The main work of the full paper is as follows:1.This paper firstly outlines the concept and significance of electric load forecasting,and describes the current status of domestic and foreign research in related fields.Then the types and characteristics of electric load and the classification,characteristics and basic process of electric load forecasting are analyzed and detailed,and the influencing factors of electric load are explored and analyzed.Then the theory and methods of some basic modules in deep learning are elaborated,and the theoretical knowledge of the modules used in the constructed short-term electric load forecasting model based on feature set and temporal convolutional network is analyzed and described in detail,and the role played by each module in the model construction is analyzed.Then,we analyze the fluctuation and nonlinear characteristics of the power load series in a regional public power load dataset as a sample,and describe the pre-processing methods and processes for the data,such as normalizing the missing values and outliers in the dataset and normalizing the scale and magnitude of the data for uniformity.2.propose and build the framework of short-term electric load forecasting model based on feature set and temporal convolutional network.Firstly,based on the original load,the load change rate sequence with the highest correlation with the original load is obtained by calculating the load change rate at different time intervals through Pearson coefficient analysis.Then the original electric load series are optimally reconstructed using empirical modal decomposition(EMD)and principal component analysis(PCA),and the optimized feature series are used together with the load change rate series and other factor series to construct the multi-load feature set(MLFS).Then the MLFS is used as the input sequence,and the data features are mined and learned using a temporal convolutional network(TCN)and a bidirectional long short-term memory network(BILSTM),and the important feature information is further strengthened and highlighted by dynamically adjusting the weights through an attention mechanism,while the warmup and cosine annealing decay strategy(COSA)are fused in the neural network training process as WUCA optimizes the learning rate parameters to further accelerate the model convergence speed and improve the training accuracy of the model.3.Under the premise of the same experimental environment and model parameters settings,three case groups are set up to build the basic models with CNN,LSTM,BILSTM and TCN and their combined models with each other and with the attention mechanism respectively for experiments,and the comparative analysis of the models is carried out by three evaluation indexes,RMSE,MAE and MAPE,which are used to verify the effectiveness of the proposed model,MLFS and WUCA,and then the robustness of the proposed model is verified by forecasting the electric load at different stages.Then,the effects of different sliding window widths,different activation functions and different optimization algorithms on the model are investigated.Finally,the effectiveness of the proposed model is further verified by cross-sectional experimental comparison.The experimental results show that the RMSE,MAE and MAPE of the proposed model are 100.601 MW,74.107 MW and1.153%,respectively,which show good results in short-term power load forecasting.