Research On Photovoltaic Power Prediction Method Based On Improved Multivariate Combination Algorithm

With the continuous popularization of photovoltaic energy,it is of great significance to improve the accuracy and reliability of photovoltaic power prediction for ensuring the safety and economy of the electrical power system.At present,most of the existing PV power forecasts are ultra short-term forecasts,while short and medium term forecasts are relatively few.In addition,the existing short and medium term forecasts are mostly based on numerical weather forecast data,and there are fewer studies that use historical data for long time series prediction and output all the forecast results at one time.Due to the advantages and disadvantages of a single model,most of the existing ultra-short-term forecasting is realized based on combination models,one is based on data decomposition,the other is based on weight coefficient combination model.However,the combination model based on data decomposition is still a single model in essence.The combination model based on weight coefficient assigns certain weight coefficients to a variety of single models for combination,while the traditional weight coefficient determination method has the problem of poor linear fitting.In this thesis,a short term prediction method for photovoltaic power generation based on BiLSTM-Informer algorithm is proposed firstly.Aiming at the problem that Informer algorithm is difficult to learn all time series information,the feature coding of time series data is improved.Bidirectional long short term memory(BiLSTM)algorithm is used to extract hidden features of time series data,and then input it into the encoder for encoding.After calculating the encoder layer and decoder layer of Informer algorithm,the short-term prediction results of photovoltaic power generation are output.In order to explore the prediction performance of LSTM model,Informer model and BiLSTMInformer model under different prediction sequence lengths,the prediction errors of the three models were simulated and compared,and the validity of the short-term prediction effect of the BiLSTM-Informer model was verified.Secondly,based on XGBoost algorithm,LSTM algorithm,Bayesian hyperparameter optimization and linear adaptive weights,a nonlinear adaptive weighted Bayesian optimization XGBoost-Bayesian optimization LSTM model considering error correction is proposed for ultra short term prediction of photovoltaic power generation.Firstly,Bayesian hyperparameter optimization is used to optimize the hyperparameter of the model trained by XGBoost algorithm and LSTM algorithm.Then,a new training data set was composed of the preliminary predicted values and the real values of the two single models.Neural network algorithm was used to train the proposed models,and adaptive weight coefficients were assigned to the preliminary predicted values of the two single models.Finally,according to the size of the predicted value of the model proposed in the training,the distribution of the prediction error was calculated piecewise.In the prediction,according to the predicted value of the proposed model,the average error value was accumulated on the basis of the prediction results for error correction,so as to further improve the prediction accuracy of the proposed model.The photovoltaic power prediction performance of the proposed model,traditional combined model and two traditional single models on sunny,cloudy and rainy days was simulated by Python language.The results show that: Compared with XGBoost-LSTM model,XGBoost model and LSTM model,the square mean root error of the proposed model is reduced by28.57%,39.39% and 49.79%,and the mean absolute error is reduced by 44.25%,53.33%and 64.8%,respectively.The coefficient of determination increases by 1.43%,2.38% and3.34%,respectively.The proposed model can more effectively reduce the photovoltaic power prediction error of the traditional single model,and optimize the weight coefficient of the traditional combined model.The prediction error of PV power of the proposed model under three weather conditions is relatively minimum and the robustness is strongest,which verifies the validity of the proposed model.