Application Of Deep Learing In Power Spot Price Foreasting

Electricity cannot be stored on a large scale.The price discovery function of the market can make the fluctuation of electricity price not affected by human intervention and truly reflect the state of electricity supply and demand.Especially for the regions where new energy power generation accounts for a large proportion,the power price will fluctuate greatly with various weather factors and time factors.The power dispatching center needs to conduct corresponding dispatching and control of thermal power,hydropower and other controllable power sources according to the fluctuations of new energy,so as to ensure the real-time balance of power generation.New energy power generation to ensure the priority of the cashier,thermal power units also have a minimum generation load requirements.The prediction of power spot price is to ensure the real-time dynamic balance of power generation side and load side.Therefore,it is necessary to study the application of time series prediction in the field of power spot price.The main work of this paper is to forecast and analyze the power price Aiming at three different perspectives of time series prediction model structure selection,input feature optimization and data dimension processing,this paper studies the improvement of short-term electricity price prediction accuracy.This paper has done the following work:(1)The relevant technologies and theories of time series type data prediction are summarized.Statistical ARIMA model,machine learning support vector machine model,deep learning cyclic neural network and long and short memory network are respectively described.(2)For the price of electricity is influenced by many factors this feature,the data set increases the power generation side of photovoltaic power generation the characteristics of various forecasting model and compares the prediction effect,after experimental verification,both short-term and long-term memory network LSTM model prediction effect is significantly higher than contrast of ARIMA model and support vector machine(SVM)model,Compared with the GRU model with the same network structure,the fitting effect of LSTM is better,indicating that the prediction effect of LSTM is relatively best when new energy characteristic data accounts for a large number of data.(3)Considering that the addition of new energy features has a great negative impact on the fitting effect of price prediction,the paper proposes a high proportion of new energy price prediction scheme.Specifically,through the comparative analysis of various features and electricity prices,the ratio of new energy and load is added as a new input feature and PCA dimension reduction processing to improve the prediction accuracy.The experimental verification shows that the ratio of new energy to load,as a newly added input feature,is a key input parameter for the optimization of electricity price prediction.Compared with the direct input of original data into the LSTM model prediction,the fitting effect is optimized and the prediction accuracy in the low peak period is greatly improved.(4)In order to further improve the prediction effect,this paper makes an innovation on the basis of the LSTM model with optimized input features and adds the CNN model.LSTM-CNN hybrid model can optimize high-dimensional data with features,and further improve the accuracy of short-term electricity price prediction.The experimental results show that the LSTM-CNN hybrid model can significantly improve the accuracy of peak price prediction while maintaining the prediction accuracy of the whole time period.