Energy Retail Sales Prediction Based On Customer Behavior

Sales prediction of energy industry can provide energy enterprise with the basis for efficient management,and help enterprises to improve the economic efficiency with macro-control.Most of the existing forecasting methods are based on historical sales data for time series analysis.Regression,neural network and other methods extract external data such as environmental factors to predict.However,sales are affected by macroeconomic factors,cyclical factors,surrounding environmental factors and other factors.They directly express as the customer's fine-grained consumer behavior.The existing works of the sales forecast are lack of deeply analyzing customer behavior.So it's necessary to dig factors related to sales and quantify them as the explanatory indexes for sales prediction.This paper introduces the fine granular customer consumption behavior as the important analysis factor.We understand the influence of the influencing factors on the sales volume from the customer behavior,and propose the sales forecasting framework based on the customer behavior analysis.By aggregating fragmented customer behavior data,we analyze the inherent correlation between purchasing behavior of representative groups and retail sales.We not only consider the difference of consumption habits among individuals,but also generalize the patterns consistency of group customers behavior,meanwhile provide managers with a basis for interpretation.According to the distribution rules of different types of customer groups,the behavior of group customer is forecasted and can further act as customer behavior feature of sales prediction.Three kinds of retail sales prediction targets are designed for energy retail management mode:single station short-term prediction,single station long-term prediction and multi-station joint prediction.The periodic features are obtained based on the time series of sales,and the customer behavior features are obtained by analyzing customer's consumption behavior.Single station prediction is based on single station correlation features.Multi-station joint prediction uses feature selection based on inter-station correlation to get other stations' features related to single station,which expands single-station prediction feature set,to achieve joint prediction,and then uses encapsulation feature selection with prediction models to predict sales.Prediction methods are validated in two aspects on two real oil trade data sets.On one hand,we compare the influence on the prediction accuracy of integrating different forms of customer behavior.On the other hand,we compare the influence on the prediction accuracy of using different prediction models,and evaluate the applicability of the prediction model.The experimental results show the effectiveness of this paper,and the sales prediction based on customer behavior can get better prediction results.