Research On Short-term Load Forecasting Of Power System Based On Improved Temporal Convolutional Neural Network With Multi-scale Feature Enhancemen

The rapid development of the power industry has brought great challenges to the safety and stability of the power grid due to the integration of renewable energy and the addition of various new loads such as electric vehicles.The short-term load forecasting results of the power system are an important data foundation for guiding power dispatch personnel to formulate basic power generation and transmission plans,unit maintenance and repair,load dispatch and distribution,as well as participating in spot market transactions and dayahead market quotations by power selling companies in the existing power market environment;It is the key basis for ensuring the safe and stable operation of the system and maintaining the supply-demand balance of the power system.High precision short-term load forecasting of power systems is of great significance for the scientific scheduling of power grid resources and the efficient,safe,and stable operation of the power grid.From the perspective of improving model performance and achieving high-precision short-term load forecasting,this article fully combines the characteristics of load data,and studies short-term load forecasting in power systems based on the Temporal Convolutional Network,which has strong data mining capabilities and efficient data processing capabilities that can be parallelized.The details are as follows:To fully explore the effective temporal information contained in power load data and conduct scientific and accurate short-term load forecasting for power systems,a short-term load forecasting model based on hybrid expansion convolution improved time convolution neural network is proposed.Learning power load data features based on parallel computable Time Convolutional Neural Networks(TCNs);To address the grid effect issue in the expansion convolution structure of TCN,a hybrid expansion convolution layer is constructed to improve the basic TCN residual block and avoid information loss.Through actual power grid data simulation,the results show that the model can effectively improve model performance and complete load forecasting tasks.Taking into account the multi time scale characteristics of power load data and the problems of information discontinuity and distance information irrelevance in traditional TCN models,an improved time convolutional neural network(MS-DHTCN)model based on multi-scale feature extraction is proposed.Firstly,causal convolution with four different sizes of convolution kernels is used to extract load data features and obtain multi-scale load features;Then,the basic TCN residual block structure is improved using double mixed expansion convolutional layers,taking into account the shallow details of load characteristics and deep connections.Finally,the constructed multi-scale feature extraction layer is combined with the improved TCN model to build the MS-DHTCN load prediction framework.The simulation results of actual power grid load data show that the MSDHTCN model proposed in this paper can effectively improve prediction accuracy.To further enhance the ability of the prediction model to extract key load features and improve prediction accuracy,a multi-scale feature enhanced DHTCN short-term load prediction model is proposed.This model introduces three attention modules to construct a multi-scale feature enhancement unit based on the previously constructed multi-scale feature extraction unit.Among them,the SE module trains weight parameters by setting squeezing and excitation layers to enhance the proportion of useful features;The Efficient Channel Attention Network module utilizes fast one-dimensional convolution to generate channel weights and achieve local cross channel interaction without dimensionality reduction,improving model performance while adding minimal parameters;The Convolutional Block Attention Module calculates the weights of load data from both channel and spatial dimensions,and then multiplies the obtained attention weights with the input features for adaptive learning of features,further highlighting key information.Firstly,input the load data into a multi-scale feature extraction unit for information extraction;Then,the extracted multi-scale load features with channel or spatial attention are input into the DHTCN model for short-term load prediction.The simulation results show that the introduction of attention module can effectively enhance the model’s ability to capture key information and improve the accuracy of load forecasting.