Short-term Multi-step Load Forecasting In High Latitude And Cold Regions Based On Recurrent Neural Network

Electricity load forecasting refers to the statistics and analysis of historical electricity data,investigation to find out the internal correlation and development law between social,economic,natural and other related factors,and according to the prediction results of future economic and social development,to make estimates and forecasts of future electricity negative to meet economic and social development and power system demand.However,with the increase of data volume,the applicability of traditional load forecasting methods has certain limitations,therefore,the god will network method gradually becomes a popular method for power load forecasting.This paper first introduces the research background of load forecasting and the significance of load forecasting in high latitude cold regions of China,and describes the current status of research on electric load forecasting and deep neural networks.The classification of load forecasting is listed,the factors influencing electric load are clarified,the principles and differences between single-step and multi-step electric load forecasting are explained,and the theoretical basis of machine learning-based time series forecasting methods and neural networks is illustrated.The data used in this paper is the electricity load data of a region in Northeast China from 2017 to 2019,which has a time granularity of 15 minutes and a time span of 3 years,with a total of 105,120 electricity load data.Two full years of data from 2017 and 2018 were used as the training set,and the full year of 2019was used as the training set.The data were normalised,normalised and windowed,and then the model was built.In this paper,a total of 14 multi-step prediction models of power load based on machine learning and artificial intelligence methods are established,and most of the models in the overall experiment can basically chase the trend of time series changes,and have acceptable test set results.Among the 7 machine learning methods,the best prediction effect is the KNN model,MAPE\MAE\R~2 is 98.72%\133\97.69%.Further,a GRU-LSTM deep learning method was selected for prediction,and the sample indicators of the prediction result,MAPE\MAE\R~2 were99.02%\102\98.71%,respectively,which was aimed at the machine learning model,and all the prediction indicators were improved.Through comparison,it is found that the LSTM-GRU deep convolutional neural network model proposed in this paper has the best r^2 and mape indicators,and the model has the best prediction effect and is most suitable for load prediction in high-latitude cold regions...