Research On Short-term Load Forecasting Method Of Elec Tric Power Based On Ensemble Deep Learning

With the development of the economy,the energy structure has undergone great changes.The increase of distributed power generation,the rise of new energy power generation industry and the diversification of users’ power consumption activities make the load forecasting of the power system present nonlinear,random and uncertain characteristics.Accurate load forecasting can affect the rationality of power grid planning,the safety of operation and the economy of supply and demand balance.In order to quantify the uncertainty of the load and improve the prediction accuracy,this paper carries out the research on probability density load forecasting and multi step forecasting.The research contents are as follows:First of all,this paper introduces the characteristics of power load and the characteristics of the load.In this paper,Pearson’s correlation coefficient analysis is carried out on the data used later to obtain data features with high correlation,which is convenient for data feature learning.At the same time,an appropriate method is selected for the subsequent load data decomposition,and the prediction accuracy is improved through decomposition and noise reduction.Then,the research on probability density prediction is carried out.According to the existing problems in probability density prediction,a mixed ensemble prediction model based on Gaussian kernel function probability density estimation is proposed.A single prediction model has certain advantages and disadvantages,and the disadvantages will have a certain impact on the stability and accuracy of the model,and the integrated prediction model can solve this problem well,not only can it combine the advantages of multiple models,but also can be increased the robustness and accuracy of the model.Considering the advantages of feature learning in machine learning and the advantages of deep learning in processing large amounts of data,in order to better improve the accuracy of model prediction,the hybrid ensemble model proposed in this paper is composed of four basic models of Gaussian process regression,gradient boosting regression tree,gated neural network,and timedomain convolutional neural network.The original loss function in the basic model is replaced by the Pinball loss function,so that the output quantile can obtain the value of probability prediction.In order to obtain the probability density curve,a method is proposed to transform the quantile value into the probability density curve through the Gaussian kernel function,so as to realize the probability density prediction.At the same time,a method of transforming the probability density prediction problem into a quadratic programming problem is also proposed,and the final prediction model is formed by the superposition of weight values.Finally,through comparative experiments,the interval coverage of the prediction model is 23% higher than that of the benchmark model,and the interval width is 47.9%,which verifies the accuracy and feasibility of the model proposed in this paper.Later,the research on multi step prediction was carried out.In order to improve the accuracy of multi step prediction,a multi-output prediction model based on multi-channel multi-scale convolutional neural network and Gated Recurrent Unit（GRU）was proposed.First,the basic principles of the composition model are introduced.Then extract the time features in the load data,input the data information and the time feature information through two neural network models,and then perform feature fusion,so as to output the point prediction result of multi-step prediction.Finally,an example is given to verify that the multi-step prediction model improves the accuracy by 24.3% compared with the comparison model,which is feasible.