Research On Short-term Load Forecasting Of Power System Based On Convolutional Neural Network

The safe and stable operation of the power system is closely related to the development of the national economy and social personal safety.Accurate short-term load forecasting results are an important basis for the grid corporations to guide the power system in formulating power generation plans,coordinating unit operation,dispatching load distribution and formulating maintenance plans.Smart grids are developing quickly,and more distributed power sources are connected to power systems.These make the factors that affect the short-term load forecasting results more complex and make the load more nonlinear.The severe environment also puts forward higher requirements for the accuracy and adaptability of short-term load forecasting.The load of power system has the characteristics of randomness,non-linearity and time sequence because of the large amount of data and the influence of many factors.Convolutional neural networks(CNN)is an accurate and efficient feature extraction method,which can deeply find the multiple characteristics of information and obtain useful feature description.Combined with the characteristics of load series,this paper uses CNN to predict the short-term load of power system,as follows:Combined with the actual characteristics of load forecasting,Convolutional neural networks suitable for the field of load series data forecasting is determined.To solve the problem that the traditional back propagation algorithm is easy to bring slow convergence or even non convergence,particle swarm optimization(PSO)algorithm is used to optimize the connection parameters of convolution neural network to improve the comprehensive optimization ability and convergence speed.Simulation results show that CNN can effectively extract sample information through its convolution layer and pooling layer.After particle swarm optimization,it has achieved good results in prediction accuracy and prediction speed.Considering that there will be some noise data in the process of sample data collection due to external factors or impact load,an improved prediction model of convolutional neural network based on the combination of autoencoder and PSOA-CNN is proposed to further improve the prediction accuracy and stability of convolutional neural network.The autoencoder can reduce the noise variates of the required data by data reconstruction.Then,the weight and threshold of CNN are optimized by using the comprehensive optimization ability of particle swarm optimization,so that the prediction performance of the model is improved effectively.This paper combines the discontinuous feature extraction ability of the convolutionalneural network with the time-series learning ability of the long short-term memory neural network,and proposes two short-term load forecasting models based on AE-PSOA-CNN and LSTM.The short-term load forecasting model based on the combination of AE-PSOA-CNN and LSTM takes the feature vectors extracted by AE-PSOA-CNN in sequence to replace the original data as the input of LSTM.Then use particle swarm optimization and Adam algorithm to optimize the two model parameters respectively.Based on the AE-PSOA-CNN and LSTM integrated network short-term load forecasting model,the two models are used to learn different characteristics of the data.The BP neural network outputs the prediction result after taking the optimal weight.The simulation verifies that the combined idea can not only perform regression prediction,but also consider the correlation between data.Among them,the short-term load forecasting model combining AE-PSOA-CNN and LSTM has better forecasting performance.