Power Time Series Forecasting Based On Deep Learning

In contemporary society,the combination of electricity and human life and production has become increasingly close,and electricity is an important energy source for social development.With the booming development of big data technology.there is also more variability in the intelligent allocation of electricity.To achieve efficient and precise distribution of electricity resources,it is necessary to accurately predict electricity data.Accurate load forecasting can help the power system achieve supplydemand balance and has high economic value and social significance.Currently,the methods for predicting power load have evolved from traditional time series methods to deep learning methods.Existing models can achieve accurate results in the short-term forecasting field,but few models can maintain high accuracy in long-term forecasting.Therefore,this article proposes the CNN-Informer model.which combines Convolutional Neural Network(CNN)and Informer model,and demonstrates through actual power data experiments that the model has good performance in medium-to-long-term power load forecasting.Since power data conforms to the characteristics of time series data,this study explores relevant papers on time series forecasting by domestic and foreign scholars.Due to the influence of factors such as holidays and seasons,it is necessary to introduce relevant covariates and perform variational mode decomposition on the load data.The decomposed mode components are more stable,and the original target data is reconstructed using the decomposed modes and covariates.The reconstructed data is then subjected to feature extraction using CNN.and then input into the encoder of the Informer model,allowing the model to extract local features while extracting correlations for long sequences of load.After preprocessing the IESO power company market demand load data.Ontario demand data,and hourly load data of a factory in Suzhou,the proposed CNN-Informer model and commonly used machine learning models such as Xgboost.neural network models such as LSTM,Transformer,etc.are used to predict the data sets and compare the results.Finally,public power data sets are used to verify the generalization ability of the models.Multiple experiments demonstrate that the proposed model has higher accuracy than other existing models in medium-to-long-term forecasting.