Reasearch On Forecasting Model Of Electricity Demand Under The Influence Of Spot Market

With the continuous advancement of the electricity system reform,the electricity market model suitable for each province has been initially completed to ensure a smooth transition of electricity from planned management to market competition,so that the electricity price in the market can reasonably demonstrate the relationship between the supply and demand.It is also helpful for reasonably scientifically guiding and investing and configuring related resources.Different types of electricity demand forecasting can provide decision supporting for the realization of electricity reform's goals.Different types of electricity have different prices and trading methods,etc.,so that the prediction results of different types are having reference value not only for the electricity grid but also for the optimal allocation of power generation enterprises.In the spot market environment,the RFE feature selection model selects suitable future matrix from a large number of indicators that affect electricity demand,such as macro,microeconomic,social population,market,weather,holidays,derivative indicators,historical electricity data characteristics and so on,as the model input for a specific type of electricity demand.Comparison of forecasting results between different models as ARIMA,SVR,PSO-SVR,SA-SVR and LSTM showed that the PSO algorithm can better optimized the forrcasting results of the SVR model on the train set,but LSTM model outperformed other models on the test set.The RMSE of industrial electricity,general commercial electricity,agricultural electricity and residential electricity in their optimal models are 0.076,0.088,0.094 and 0.091 respectively.The optimal model for electricity demand other than industrial power is PSO-SVR,but there is over-fitting problem which means the forecasting effect on the test set is not good.When the target variable seasonal decomposed and then the forecasting accuracy is greatly improved and the over-fitting problem is relieved.The RMSE of the four different electricity demand requirements on their optimal model are 0.046,0.025,0.047 and 009;and the optimal model is LSTM,the forecasting effects on the test set are 0.078,0.190,0.324,0.081.However,the agricultural electricity consumption is obviously different from other electricity demand's regular pattern,the SA-SVR model gains better forecasting results on the training set and the test set respectively of 0.096 and 0.205,the seasonal decomposition has worsened the over-fitting problem.The model method proposed in this paper can form a forecasting system in the process of continuous accumulation of forecasting experience.When faced with different forecasting conditions and target variables or industries,it can be analyzed by correlation analysis such as trend and expert systems to determine the appropriate range of forecasting models,and optimize model by the feedback through comparison between models.