Research And Application Of Power Time Series Data Forecasting Technology

With the rapid development of smart grid,the power industry can generate hundreds of millions of power time series data every day.The analysis and forecasting of power time series data has important practical significance for the safe and stable operation of power system and the rational allocation of resources.In recent years,more and more scholars began to study power time series data forecasting.Most of the existing methods forecast the future trend by studying the static relationship or historical sequence characteristics between multiple power time series data.However,the dynamic change of multiple power time series data has not been fully considered.This will lead to the model can not accurately capture the relationship between power time series data,which limits the accuracy of power time series forecasting.Therefore,this paper not only deeply considers the historical sequence characteristics of power time series data,but also fully explores the dependence between power time series data,and designs two power time series forecasting models and develop a power time series data processing platform.The main work and contributions of this paper are as follows:(1)In this paper,a two-view deep neural network model is proposed to forecast power time series data.The model includes a view based on the decomposition of shallow time series features and a view based on the mining of deep time series features.From the view of shallow time series feature decomposition,the statistical model is designed to decompose the trend,periodicity and holidays feature in power time series data.From the view of deep time series feature mining,a deep learning model is designed to mine the potential deep time series features in power time series data.Finally,the complementarity of the two view time series prediction model is explored by organically integrating the time series characteristics.The experimental results show that the model is better than the comparison method in the accuracy and interpretation of induction motor proportional data prediction.(2)In this paper,a dynamic and static graph attention network model is proposed to forecast power time series data.The model designs a dynamic and static graph learning layer to mine the short-term pattern and long-term pattern in the time series data.The dynamic graph is used to capture the short-term change trend of the time series data,and the static graph is used to mine the long-term evolutionary model of it.In addition,the model also designs a temporal convolutional layer for power time series data,which can effectively learn the time series feature of different time scales.Finally,graph attention network is adopted to aggregate topological information of dynamic and static graphs,respectively,to interact with short-term and long-term patterns,thereby enhancing the feature representation of time series data.The experimental results on photovoltaic energy and household power consumption datasets show that the model is better than the comparison method.(3)Based on the research of power time series forecasting,this paper designs and implements a power time series data processing platform.The platform consists of missing value processing module,original data analysis module,correlation analysis module and power time series prediction module.This platform provides a set of power time series processing scheme,and also provides a significant reference for the decision-making of relevant power departments.