Photovoltaic Power Prediction Based On The Fusion Of Multiple Models

Energy is an important material basis for the development of human society.With the progress of society,the problems of energy consumption and environmental pollution have become increasingly prominent.As a renewable and clean energy,photovoltaic power generation has strong intermittent and volatility compared with traditional fossil fuel power generation.Accurate photovoltaic output power prediction can provide technical guidance for the smooth grid connection of photovoltaic power stations and avoid grid power quality problems caused by huge photovoltaic power fluctuations.Therefore,it is essential to obtain more accurate photovoltaic power prediction data.Traditional photovoltaic power prediction methods generally only consider regression prediction of historical series or error correction based on similar days.In order to further improve the accuracy of photovoltaic power prediction,the article considers both time series input data and similar daily input data,and uses different model structures for training.The trained model adopts a variety of model fusion strategies to improve the accuracy and stability of the model.The main research contents include:(1)Aiming at the problem of short-term prediction using historical photovoltaic power data,a W-CNN-BiLSTM model is proposed.The model utilizes the ability of the LSTM model(Long Short-Term Memory)to propagate predictions on the time axis,and improves the prediction accuracy through multiple improvements to feature extraction.In view of the large fluctuation and randomness of photovoltaic power output,the improved model uses wavelet decomposition to decompose the photovoltaic power curve into high-frequency and low-frequency signals,and uses CNN structure(Convolutional Neural Network)to extract the spatial structure of the signal Features,using bidirectional long short-term memory neural network for prediction.Finally,various improved models are compared to verify the effectiveness of the proposed model.(2)Aiming at the problems of the inability to parallel training of the recurrent structure of long short-term memory neural networks and the weak feature extraction capabilities of convolutional neural networks in long-term sequences,an improved Transformer model is proposed.The Transformer model that has achieved good performance in natural language processing tasks is improved into a parallel training structure with similar days as input.The model is completely based on the attention mechanism to achieve prediction.First,the similar days are used as the input of the model encoder,and the self-attention mechanism of the encoder is used to extract the features of the similar days.Then,the predicted time series features of the day to be predicted are input to the decoder and the attention mechanism is used to extract the features of the similar day at the corresponding time to obtain the predicted data at the next time.Finally,compared with other prediction models,it is found that this model can have strong prediction stability and high prediction accuracy under weather with large fluctuations in photovoltaic power.(3)In view of the fact that a single model can no longer meet the requirements of prediction accuracy,in order to further improve the prediction accuracy and increase the robustness of the model,two models with large differences and high prediction accuracy are fused,and a photovoltaic power prediction method with multiple models is proposed.The historical photovoltaic power sequence is used as the input of the W-CNN-BiLSTM model and the similarity date is used as the input of the improved Transformer model.Their prediction results are used as a new dataset,and the Xgboost model(Extreme Gradient Boosting)is trained using this new data set,and finally realize the output prediction result.Experimental results show that compared with other single models,the fused model has higher accuracy and better stability.