Research On Short-Term Photovoltaic Power Prediction Method Based On Improved Similar Day Algorithm And LSTM Neural Network

With the increasingly prominent environmental issues,photovoltaic(PV)power generation,as an important clean and renewable energy generation technology,has developed rapidly.However,due to the randomness,intermittency and volatility of PV power generation,the large-scale grid connection of PV power will seriously affect the operational stability of the power grid.In order to ensure the quality of the power grid and assist the power department in formulating scheduling plans,PV power prediction is an effective solution.On the basis of existing research,this thesis explored an innovative short-term PV power prediction method that combines deep learning frameworks and multiple algorithms.The main research content and focus are summarized as follows:This thesis firstly utilizes qualitative and quantitative analysis methods to select the influencing factors of PV power generation with strong correlation;then,considering the accuracy of the dataset,a method combining physical recognition and statistical recognition is proposed to identify abnormal data,and to delete and correct abnormal data to ensure the effectiveness of the data used in establishing the model.Secondly,considering the significant differences in PV power generation laws under different weather types,a weather type classification method combining K-means++and Fuzzy C-means lustering(FCM)is proposed.Firstly,eight statistical features of historical power series are calculated as input to the clustering algorithm,and K-means++ is adopted to explore the optimal classification number.FCM algorithm is used to correct the edge samples of strong feature sets,and the power curve clusters under four weather types have obviously regular characteristics.Then,considering the limited accuracy of the weather forecast information on the second day,it is determined to use the combination of historical data samples and weather forecast data as the input of the model.For the effective utilization of historical daily samples,this thesis proposes an improved similar day selection algorithm that combines weighted Gray Relation Analysis(GRA)and Cosine similarity(Cosine).And this method is compared with similar date selection algorithms in other literature.The mean square error of the forecast day and the actual day(SRMSE)is proposed as an indicator.The comparison results verify the advantages of the improved similar day algorithm.Then,based on the weather type classification method and the improved similar day algorithm,this thesis establishes an innovative hybrid short-term PV power prediction model(H-LSTM)combined with the Long Short-Term Memory neural network(LSTM).The model structure and model hyperparameter settings are determined through K-fold cross validation and control variates optimization methods.Then,the processed dataset is used for example analysis of the hybrid model.The calculation results illustrate that the average Root Mean Square Error(RMSE)and average Mean Absolute Deviation(MAD)of the H-LSTM model are 0.3139MW and 2.33%,respectively;Compared with the unclassified weather prediction model,the training time of the H-LSTM model has been shortened by 5.66%,and the MAD has been cut by 62.30%.Compared with other improved similar day algorithm models,the RMSE of the H-LSTM model is reduced by at least 31.00%.Compared with other neural network hybrid models,the RMSE of the H-LSTM hybrid model has increased by a maximum of 24.13%.The comparative results validate the superiority of the hybrid model proposed in this thesis,which combines weather type classification,improved similar day algorithm,and LSTM neural network.Finally,considering the issue of large power fluctuation leading to significant prediction bias under non-ideal weather conditions,an optimized hybrid model(H-WPD-LSTM)combining Wavelet Packet Decomposition(WPD)is proposed.Each feature parameter is decomposed into four sub-sequences based on WPD,and each subsequence model is established.The final optimization result is obtained by adding the output results.Example analysis is conducted under ideal and non-ideal weather conditions.The results indicate that the H-WPD-LSTM model has a significant improvement in fitting and prediction ability under non-ideal weather conditions,and the mean MAD has increased by 8.55%compared to the original mixed model.The innovative hybrid prediction model proposed in this thesis has high prediction accuracy and can effectively predict the short-term power output of PV power plants.Meanwhile,the analysis of the optimized model provides a reference basis for improving prediction accuracy in non-ideal weather conditions.