Research On Daily Load Forecasting Of M Thermal Power Plant Based On Deep Learning Algorithm

With the proposal of China’s "carbon neutral" strategy,new energy power generation methods such as wind energy and solar energy have been largely installed and connected to the grid in recent years.As the supply of new energy power generation is unstable,it has a certain impact on the stable operation of China’s power grid.This leads to frequent startup and shutdown of units and load lifting and lowering of thermal power plants in order to maintain grid stability.Therefore,the research on load forecasting of thermal power plant becomes more and more important.In the process of data processing for daily load forecasting research of M thermal power plant with two 660 MW units,in view of the irregular distribution of the data set,the data of daily power generation of M thermal power plant is optimized by introducing the parameter "average daily power generation hours of units" to the data set,and the two dimensions of "number of generating units" and "daily power generation hours" are introduced.Therefore,this optimization method can also be extended to the model of four units and above,The model of double(multiple)units is simplified to single unit model.This optimization method transforms the data set from a bimodal distribution(multimodal distribution)to an approximate normal distribution.Through comparative tests,it can be seen that the average absolute percentage error MAPE of the load forecasting results of the optimized data set using BP neural network is reduced from 18.7% to 15.3%compared with the results of the load forecasting of the original data set using BP neural network.The accuracy of prediction results of RNN,LSTM and GRU models has also been improved.In the process of establishing the load forecasting model,this paper proposes VMD-Adam-LSTM algorithm model.Because the data set in this paper is aperiodic and nonlinear,this paper selects the variational empirical mode decomposition(VMD)method which is more suitable than the wavelet decomposition method to process the data.The optimization algorithm uses Adam optimization algorithm.Because the mean value of the first order moment and the mean value of the second order moment are used,Adam optimization algorithm has the advantages of inertia retention and comfort perception in the calculation of learning rate.Adam optimization algorithm and LSTM algorithm use the same time scale t,so it has a high degree of agreement with LSTM network algorithm.The LSTM algorithm updates the data stream and the modified data separately with logic gate operation.On the basis of the advantages of RNN network in processing data with sequence characteristics,it solves the problem of gradient explosion and gradient disappearance in RNN network,and also enhances the sensitivity of data stream to the original data.The comparative test results of VMD-Adam-LSTM algorithm model using the optimized dataset and BP neural network model using the pre optimized dataset show that this model reduces the average absolute percentage error of load forecasting from 18.7% to 12%.In the process of exploring the optimal model,ten models are proposed in this paper: BP neural network,RNN,LSTM,GRU,SGD-GRU,SGD-LSTM,Adam-LSTM,Adam-GRU,VMD – Adam-LSTM,VMD – Adam-GRU.Finally,the comparison test between VMD-Adam-LSTM algorithm model using optimized data set and BP neural network model using optimized data set shows that this model reduces the average absolute percentage error of load forecasting from 18.7% to 12%.The research in this paper improves the accuracy of daily load forecasting of M thermal power plant,which is of great significance to daily load forecasting of thermal power plant.