Research On Real-time Electricity Price Forecasting Based On Deep Learning

In recent years,with the continuous development of the global power market reform,the pattern of monopoly and government control of the power sector has been gradually broken.The efficiency of power industry has been improved,and economic and social progress has been promoted.In the context of marketization,electricity has become a commodity that can be traded freely,and its price varies in real-time.Therefore,electricity prices have become the most concerned issue.Due to the special properties of electricity,electricity prices are affected by many factors and show high volatility and uncertainty,which brings great challenges to the accurate prediction of electricity prices.For market participants,the acquisition of electricity price forecasts in advance makes it possible to earn more profits in power trading.For electricity consumers,reasonable electricity consumption plans can be made based on predicted electricity prices to reduce electricity costs.Additionally,power suppliers can rely on accurate electricity price forecasts to deal with power generation scheduling of power shortage or surplus events during certain time periods,which is conducive to improving system load rate,reducing system operation cost,and ensuring the safety and stability of power systems.Based on the above background,this thesis conducts an in-depth research on the forecasting of dayahead spot price in power market,and implements a novel forecasting model.The main contributions of this thesis are as follows:(1)Considering the integration of power market and the signal effect of neighboring markets,a bidirectional recurrent neural network and integrated market based electricity spot price forecasting model,BRIM is proposed.The experimental results in the EPEX-FR show that the accuracy of our proposed model for day-ahead spot price forecasting is significantly higher than that of the comparison schemes,and the day-ahead data of neighboring markets is conducive to improving predictive accuracy.(2)Considering the sudden occurrence of a few price spikes in the electricity spot market will negatively affect the accuracy of real time electricity price forecasting,a spike price occurrence prediction scheme is proposed in this thesis.Specifically,to alleviate the impact of few spike price samples in historical data,borderline-SMOTE is utilized to synthesize some spike data to increase the number of price spikes at the data level,and a loss function with a misclassification penalty to increase the cost of missing price spikes is designed at the algorithm level.The experimental results in the EPEX-FR prove that our scheme improves the accuracy and recall of price spike occurrence prediction compared with the conventional models.(3)By analyzing the forecasting performance of BRIM in different price intervals,it is found that the price spike prediction needs to be further optimized.Based on the research on preprocessing methods of electricity price data,a two-stage electricity price forecasting scheme based on multi-source data is proposed.This scheme chooses variance stabilizing transformations suitable for electricity prices in different intervals,respectively,and combines a price spike calibration model based on deep neural networks to further improve the accuracy of day-ahead electricity price forecasting.The experimental results show that the two-stage forecasting scheme significantly improves the accuracy of spike electricity price forecasting,without affecting the accuracy of normal electricity price forecasting.