Research On Monthly Electricity Consumption Forecast Approach For Customer Clusters Of Electricity Retailer

The advancement of the reform of the electricity market has promoted the gradual formation of a diversified pattern of electricity sellers.In the deregulate d market environment,accurate forecasting of users’ monthly electricity demand is an important basis for electricity retailers to carry out mid-and long-term electricity trades.Research on the monthly electricity consumption forecasting(MECF)method is an important means to ensure the profitability of electricity retailers and enhance their market competitiveness.At present,most MECF methods utilize low-resolution historical monthly electricity consumption data to construct a one-step forecasting model.The lack of training samples often leads to prominent over-fitting problems of the forecasting model,which seriously affects the generalization ability.To solve the above issues,this paper has carried out the following research:1)To better mine the information in the massive high-resolution historical electricity consumption data to overcome the lack of training samples in traditional MECF methods,this paper investigates how to utilize the multi-step forecasting strategy to make use of historical hourly electricity consumption data in MECF.Then,the main reasons that affect the accuracy of the multi-step forecasting strategy-based model are analyzed,and the MECF model based on the series decomposition of historical electricity consumption series is proposed.First,according to the similarity between the electricity consumption behaviors on the same weekdays or weekends in different weeks,the historical hourly electricity consumption series is divided into seven sub-series according to the weekly label.Then,the recursive multi-step forecasting strategy is used to construct a multi-step forecasting model for the decomposed seven sub-series.Finally,the seven-part forecasting results are accumulated to obtain the MECF result.2)On the one hand,increasing the resolution of historical data utilized in the construction of the model can enlarge the number of training samples to avoid model over-fitting,on the other hand,it can also lead to a sharp increase of the forecasting steps and reduce the stability of the multi-step forecasting model.To solve this problem,an optimization method for the resolution compression scale of historical power consumption series is proposed based on the series decomposition method.First,the resolution of the decomposed sub-series is compressed,and forecasting models are constructed for the compressed series.Thus,these forecasting results can be summed to form the MECF result.Then,a variety of MECF results at different sample resolutions can be obtained by varying the compression scale.Finally,the optimization modeling method is used to combine these MECF results to obtain the final forecasting result.At the same time,the compression scale optimization results are also given.3)The loss of information in the resolution compression process severely reduces the predictability of the compressed series.To solve this problem,an auto-encoder neural network(AENN)data compression method is proposed to compress historical electricity consumption series.First,the historical hourly electricity consumption data is used to train an AENN whose number of neurons in the hidden layer is less than the number of neurons in the input layer.Then,the encoder part of the AENN is used to compress the electricity consumption series into the more predictable coding series.Finally,forecasting models are constructed for the coding series,and then the forecasting results are decoded and accumulated t o obtain the final MECF result.