Research On Short-Term Power Load Forecasting Based On Deep Learning

Short-term load forecasting is the basic work of power system operation and control,and an important link to ensure social security and stability.High precision short-term load forecasting has always been the goal of power enterprises and researchers.With the construction and development of electric power market,the continuous progress of various measurement systems and communication technology brings both challenges and opportunities to short-term load forecasting research.On the one hand,with the grid connection of various mobile loads such as distributed renewable energy and electric vehicles,and the implementation of demand side management,the factors affecting the load become more complex than before,and the uncertainty of load change is greater.On the other hand,the popularity of smart electricity meters provides massive data support for load forecasting research,making it possible to carry out load forecasting research from the user side.In view of the short-term load forecasting problem in the current complex environment,this paper mainly carries out the following research:(1)The periodical law and influencing factors of load data are intuitively analyzed,and the maximum information coefficient algorithm is used to quantify the influence degree of various influencing factors on load data,and the input characteristics of load prediction model are determined according to the correlation analysis results.At the same time,the method of data preprocessing is introduced,which lays a foundation for the establishment of subsequent prediction model.(2)A short term load forecasting method based on deep clustering is proposed.Firstly,the user-side load data obtained by smart electricity meters are preprocessed to construct the user load characteristic curve to reduce the influence of the uncertainty of electricity consumption behavior on the load.Then,the depth clustering algorithm is used to cluster the user characteristic curves to get different clusters,which solves the problem of key information loss in the process of dimensionality reduction of highdimensional load data.According to the characteristics of different clusters,the bidirectional short and long time memory neural network is used to predict the load data,and the positive and negative laws are fully mined.Finally,the results are summarized to obtain the final prediction results.The experimental results show that this method has higher prediction accuracy when the sample data is large and the dimension is high.(3)A short term load probability prediction method based on quantile regression neural network is proposed.Firstly,the selected dataset was divided into four subsets according to the seasonal effects,and the data were preprocessed.The quantile regression time memory neural network was used to get the prediction values at different loci,and then the prediction intervals at different confidence levels were obtained,which overcame the problem that the traditional quantile regression model could not deal with nonlinear regression.The feature attention mechanism is integrated into the prediction model,and the model can adjust the attention weight in real time,so as to enhance the load of strong correlation factors and weaken the influence of weak correlation factors on the prediction results.Finally,according to the predicted values of different subpoints,the probability density function of load is obtained by kernel density estimation method,which reflects the probability distribution of load in the future comprehensively and intuitively.The reliability and accuracy of the probabilistic prediction model are proved by experiments.