Modeling Of Thermal Power Unit Load Prediction Based On Phase Space Reconstruction And Bidirectional Long Short-term Memory Network

Unit load prediction is one of the key technologies of power grid operation and dispatching and distribution.In the process of its research,it is necessary to not only explore the variation rule of unit load itself and the correlation between each influencing factor and unit load,but also to make a legitimate short-term forecast of the value of future unit load.Strict and effective unit load forecast is of great consequence to the work of smart grid.Accurate unit load prediction can provide relevant technical and scientific basis for power patching,economic load distribution,load control and the overall economic operation of power grid.On behalf of exactly express the features of unit load time series,expand the amount of information in unit load time series and enhance the precisely of unit load model,this article center around the internal laws and features of unit load data.Phase Space Reconstruction(PSR)and Bidirectional Long-Short Term Memory(BILSTM),which are capable of fully representing time series,were selected.Design load prediction model at the same time by using PSR and BILSTM,using both advantages based load forecasting model,and according to the load is vulnerable to the feed water flow,main steam pressure,the impact of various limitations for instance main steam temperature characteristics,find the main limitations influencing the final forecast precision of the model,In this paper,the 600MW unit of a power plant in Hunan province is taken as the study object,and the load timing information of power plant unit is deeply studied and predicted.The core study indices are as follows:Firstly,the historical unit load data of a power plant and the data of various influencing factors are taken as samples.After comprehensive pre-processing of the collected unit data,the noise reduction of unit load is carried out by Empirical Mode Decomposition(EMD).After noise reduction,the chaotic characteristics of the unit load sequence are analyzed,and the time sequence is updated through time delay,time window and implanting dimension to increase the characteristic dimension of the unit load sequence,fully display the internal information of the sequence,and obtain the multi-dimensional unit load sequence.A Least Absolute Shrinkage and Selection Operator(LASSO)algorithm is used to chuse variables from the collected multiple impacting limitations,and the limitations with important impacting limitations are chused as characteristic variables for further research.Secondly,the Long Short Term Memory network model(LSTM)is constructed,and the prediction results of the PSR-LSTM model and the newPSR-LSTM model are compared to verify the validity of the LASSO algorithm as the variable selection.On behalf of further the forecast precisely of unit load,BILSTM model was built,PSR and BILSTM were combined,and the model performance was improved by training and improving various parameters in the model.A number of influencing factors after LASSO and the historical load sequence of the unit were simultaneously input as model variables.After case study,it was discovered that,The method has high precision.Finally,after comparing the PSR-BILSTM method with newPSR-LSTM method and standard BILSTM method,it is found that the prediction ability of the model is crucially modified.In the same prediction time range,linked to other models,the prediction curve of the PSR-BILSTM model is closer to the original data.And the error can be kept to the lowest in the case of sharp fluctuation of unit load.By comparison,it is found that RMSE decreases from 5.197 to 2.054,MAPE decreases from 1.056 to 0.482,RMSE decreases from 1.843 to 1.11,MAPE decreases from 0.359 to 0.221 in the prediction within 12h,MAPE decreases from 1.056 to 0.482,MAPE decreases from 0.359 to 0.221 in the prediction within 5min.The error of THE PSR-BILSTM model is always lower than that of the other two models,whether from the average value of MAPE and RMSE or the maximum and minimum value of MAPE and RMSE.It is proved that the research method used in this article has preferable prediction effect and higher prediction precision.