Research On Short Term Power System Load Forecasting Based On Deep Learning

The production of electric power is closely related to the national economy.Over-production or under-production will stably operation of the power system.Therefore,accurate load forecasting technique has practical significance in a power grid environment where the scale is continuously expanding.Traditional load forecasting methods are difficult to meet the needs of power grid regulation due to problems like insufficient feature extraction capabilities.In view of the shortcomings of existing load forecasting methods,this paper using deep learning to design short-term load forecasting models,and compares the prediction result with various load forecasting models,which shows the superiority of the proposed load prediction model in prediction accuracy and training speed.Firstly,the problem of power system load forecasting is analyzed.Basic principles of power load forecasting concepts,classification,characteristics,accuracy and the factors affected by the forecasting model is explored,various domestic and foreign load forecasting models are analyzed,the steps of load forecasting problem resolution are explained by using the load forecasting model metrics,and the forecasting model to implement the overall process of forecasting is designed.Secondly,the application of deep learning in short-term load forecasting is studied.Basic theories of commonly used models of neural networks and deep learning is introduced,steps of load forecasting,data preprocessing method and division of data sets used is explained,and the advantages and disadvantages of LSTM,GRU,SVR,DBN,and XGBoost load forecasting models is analyzed.Finally,the problem of short-term load forecasting is studied.By establishing the CNN-LSTM model,the active feature richness under the condition of limited input feature number is realized.By establishing a parallel multi-model fusion model,multi-dimensional feature extraction is achieved when the features of input load data are rich,the prediction model and XGBoost algorithm are combined with the reciprocal error method to construct a combined prediction model to achieve further improvement in error accuracy.Combined with the above researches,the Python language is used to perform prediction operations in two environments,PyCharm and Jupyter Notebook.The results of the calculation example show that the MAPE of the CNN-LSTM short-term load forecasting model is 0.608%,and the prediction accuracy and training speed are improved when only the load data is available;the MAPE of the parallel multi-model fusion short-term load prediction method is 2.677%,and the short-term load forecasting accuracy is improved when multi-dimensional features are available.Both models solved the problem of insufficient short-term load forecasting accuracy,which helps to reduce the operating cost of the power grid and improve the power grid's control capability.