| With the continuous advancement of China’s electricity market reform,the electricity sales side has gradually become a key part of the entire power system.While the number and scale of electricity sales companies continue to rise,there are also many losses and withdrawal from the market.It can be seen that the entity of the electricity sales company still has some room for improvement in its own profit acquisition in the process of operation and trading,and the profit acquisition is usually obtained from the difference in the process of purchasing and selling electricity.This article will focus on the power purchase process,through the electricity price forecast,based on the deviation of the electricity assessment to propose the electricity sales company’s daily declaration strategy.The specific research content is as follows:(1)For the problem of accurate prediction of electricity prices,based on the consideration of the periodicity and volatility of electricity prices,a short-term electricity price prediction model that is more suitable for the spot market of electricity is constructed.The model decomposes the original dataset by empirical mode decomposition method,obtains the local feature information of different time scales of the original data signal,divides the frequency by sample information entropy for different decomposed signals,uses the support vector regression method to train the data for the low-frequency modal component,and uses the convolution-bidirectional long-term short-term memory neural network training data based on the attention mechanism for the high-frequency modal component,and finally reconstructs and summs,which effectively improves the accuracy of the model prediction and improves the training rate of the model to a certain extent.(2)For the improvement of the accuracy of the prediction model,given that electricity price and charge are largely correlated,multi-task learning can be used to train multiple sets of models at the same time.The neural network has a large amount of computation,and multiple single-task models repeatedly train multiple sets of data,which is not only inefficient,but also does not have strong generalization performance.Electricity sales companies need to declare their electricity purchases in the spot market every day,and it is important to optimize the performance of the model.The data to be predicted in this paper includes day-ahead electricity prices,real-time electricity prices and real-time electricity consumption,there is a strong correlation between the three,suitable for using multi-task learning for simultaneous training and prediction,on the one hand,parameter sharing can be achieved between different tasks,so that the total number of parameters is reduced,the amount of memory occupied is correspondingly reduced,and the overall training speed is improved;On the other hand,related tasks can share information to a certain extent,complement each other’s performance,and thus improve the comprehensive prediction effect.(3)For the trading strategy of electricity sales companies,based on accurate forecasting of electricity prices,this paper proposes to formulate a day-ahead reporting strategy of electricity sales companies with the goal of minimizing the cost of electricity purchase.After obtaining the predicted values of day-ahead electricity price and real-time electricity price through the constructed model,compare the size of the two,and adopt the decision of over-reporting the electricity before the day or under-reporting the amount of electricity declared before the day.By predicting real-time electricity consumption,combined with the assessment range of deviations in each region,the value of the specific amount of electricity declared before the specific date is determined based on real-time electricity consumption within the scope of the price difference income processing mechanism,so as to obtain the maximum price difference benefit and achieve the goal of reducing the cost of electricity purchase. |
PDF Full Text Request