Research On Short-term Power Load Prediction Method Based On Deep Learning

Short-term load forecasting(STLF)plays a significant role in power system automation control,power security,market operation,and scheduling optimization.The forecast range is generally 1 day to 1 week in advance.Due to the nonlinear and non-stationary characteristics of the electric load sequence itself and the complex influence of various influencing factors on the load,it is difficult to achieve high-precision STLF.This paper proposes a hybrid algorithm that combines Similar Day Selection(SD),Variational Mode Decomposition(VMD),and Long Short-Term Memory(LSTM)neural networks,and builds a hybrid algorithm based on SD-VMD-LSTM short-term power load forecasting model.The main research contents of this paper are as follows:(1)This article introduces the basic theory of power load forecasting in detail and comprehensively.It mainly introduces the basic concepts,influencing factors and error analysis of power load forecasting,and summarizes the basic process of short-term power load forecasting.(2)Add key features: At present,factors such as temperature,humidity,day-ahead load and date type have been widely used as input features of STLF,but this article also realizes that STLF is very sensitive to day-ahead peak load through the analysis of Xgboost algorithm,so the peak value The load characteristics are added to the input characteristics of the model proposed in this paper.The calculation example shows that the forecast accuracy of the model that increases the peak load before the day is higher than that of the original model during the peak power consumption period.(3)Propose a similar day selection method based on Xgboost-K-means: This method first uses Xgboost to score each input feature and adds a weight coefficient to each feature,and finally uses an improved K-means clustering algorithm based on weighted coefficients A similar day selection is made for the electric load.Numerical examples show that the proposed Xgboost-k-means method can effectively merge similar days into a cluster.(4)Propose a short-term power load forecasting model based on SD-VMD-LSTM: use the VMD method to decompose the data processed by the SD method into several intrinsic modal functions(IMFs),and then use the LSTM neural network to analyze each IMF separately The sequence is predicted,and finally the predicted value of each LSTM model is reconstructed.The method proposed in this paper is used for load forecasting one day in advance and one week in advance with LSTM,SD-LSTM,VMD-LSTM,SD-ARIMA,SDBPNN and SD-SVR models on the real load data set of a certain area in the United States.The results of the calculation example show that the SD-EMD-LSTM model is better than most existing prediction models under a longer level load(1 day to 1 week).