Research On Short-term Power Demand Combination Optimization Forecasting Of Power Supply Chain

Power industry is a major basic industry of the country,power supply and demand are more closely related to economic growth,which makes power demand forecasting become a hot research topic in many fields such as management science,electric power science and computer science,then high accuracy forecasting results can better support power demand-side management.This paper investigates the power demand forecasting methods in the power supply chain from the perspectives of model combination optimization,data feature selection,and optimal combination model selection,respectively.Based on the more accurate prediction results,we can analyze the supply and demand situation of the power industry,support the government and power enterprises to reasonably formulate policy measures,optimize power consumption and relieve the pressure of power outages,then realize the long-term benefits of energy conservation and environmental protection.Firstly,the Bi-directional long short-term memory(Bi-LSTM)model incorporating the Attention mechanism to improve the effectiveness and accuracy of prediction,which is used as a benchmark,and a newly developed e Xtreme gradient boosting(XGBoost)prediction model with good performance is introduced to optimize it in a weighted combination.The “Attention-Bi-LSTM+XGBoost” combined prediction model is then constructed.In the data pre-processing,the weighted gray correlation projection algorithm is used to eliminate the adverse effects of holiday data.After evaluating the forecasting method with the Singapore and Norway power market datasets,the results of the case studies show that the weighted gray correlation projection algorithm can effectively optimize the input data,and the “Attention-Bi-LSTM+XGBoost” combined model is closer to the real data with less error than the single model,which improves the accuracy of power demand forecasting.Secondly,considering that the demand forecasting process is limited by time,weather,and economic factors,the large number of influencing factors makes the forecasting process more complicated and inhibits the accuracy of the forecasting method.Therefore,the number of influencing features is increased to 27,and the model with the best performance is selected among multiple feature selection methods.The results of the case study show that the prediction results of the lasso regression method are closer to the real data with less error than the other feature selection methods,indicating that feature selection is a key element to achieve accurate prediction and make the sample data more realistic.Finally,on the basis of Attention-Bi-LSTM(ABL)and XGBoost models,single support vector machine(SVM),e Xtreme learning machine(ELM),gated recurrent unit(GRU)and temporal convolutional network(TCN)models are added to comprehensively evaluate the performance of six single models using weighted information criteria to arrive at the optimal combination model.The results of the case studies show that it is not found that the more the number of single models to be combinatorial optimization is,the better the prediction results are,which is determined by the forecasting performance of the single models constituting the combined model.The short-term power demand combination optimization forecasting method based on power supply chain proposed in this paper is effective in terms of combination optimization,feature selection and model selection,which can overcome the shortcomings of single model,improve the forecasting process,raise the accuracy and feasibility of model forecasting,reduce power supply losses to ensure the efficiency of the power supply chain.