Design And Application Of Electricity Sales Forecasting Algorithm Based On Multi-scale Convolutional Neural Network

As an important part of the power marketing management system,electricity sales are the main product of power companies and an important indicator to measure the management level of power companies.Based on the accurate estimation of electricity sales in a specific period of time in the future,predicting the operating status of power companies and adjusting operating strategies in a timely manner are of great significance for improving the power companies’ ability to avoid risks,maximizing the potential of the current power grid,and improving the stability of the power grid.The neural network modeling method can effectively process the nonlinear data in the electricity sales through the end-to-end powerful mapping ability,and improve the expressive ability of the model.Therefore,this method has been introduced by some scholars into the forecast of electricity sales.Up to now,the most popular application is electricity sales prediction based on BP neural network.This method extracts input data features through multi-layer neurons and can learn certain data rules.However,there are often many factors that affect electricity sales,and there is a certain degree of coupling between different factors.However,the electricity sales forecasting method based on BP neural network cannot effectively process data with strong coupling.In response to this problem,a multi-scale convolutional neural network（Attention-based Multi-scales CNN,AMSCNN）integrated attention mechanism is studied for the electricity sales prediction model,accuracy test,and specific implementation.First,the data characteristics of electricity sales are studied.Different influencing factors of electricity sales data have different span-period attributes,and electricity sales coupled by multiple factors have multi-scale characteristics;secondly,due to the lack of multi-scale feature extraction capabilities in traditional methods,a multi-time scale division method is studied.This method divides the input data into multiple scales,and the subsequent model extracts the time features implied by the data from multiple time scales;third,because different time scale features have different contributions to the accuracy of electricity sales forecasting,they need to be targeted at different times.Span features are assigned weights.To solve this problem,a multi-channel data fusion model based on the attention mechanism is proposed to realize the self-adaptive extraction of multi-channel features.The algorithm was tested with actual electricity sales data,and the results showed that compared with the Conventional deep learning methods,the performance of the electricity sales prediction model based on AMSCNN increased by 3.68%on average.Finally,based on the above model,the electricity sales forecasting module of the power marketing system is developed.Through the PyQt5 framework,the login and registration interface,the main interface and the electricity sales forecast module interface of this part of the module are built,and the detailed design is carried out around the parts of the electricity sales forecast module,such as data analysis,data visualization and different models.