LightGBM-Integrated PV Power Prediction Based on Multi-Resolution Similarity

Abstract

Improving the accuracy of PV power prediction is conducive to PV participation in economic dispatch and power market transactions in the distribution network, as well as safe dispatch and operation of the grid. Considering that the selection of highly correlated historical data can improve the accuracy of PV power prediction, this study proposes an integrated PV power prediction method based on a multi-resolution similarity consideration that considers both trend similarity and detail similarity. Firstly, using irradiance as the similarity variable, similar-days were selected using grey correlation analysis to form a set of similar data to control the similarity, with the overall trend of the day to be predicted at a macro level. Using irradiance to calculate the similarity at each specific point in time via Euclidean distance, similar-times were identified to form another set of similar data to consider the degree of similarity in detail. The above approach enables the selection of similarity data for both resolutions. Then, a 1DCNN-LSTM prediction model that considers the feature correlation of different variables and the temporal dependence of a single variable was proposed. Three important features were selected by a random forest model as inputs to the prediction model, and two similar data training models with different resolutions were used to generate a photovoltaic power prediction model based on similar-days and similar-times. Ultimately, the learning of the two predictions integrated with LightGBM compensate for each other, generating highly accurate predictions that combine the advantages of multi-resolution similarity considerations. Actual operation data of a PV power station was used for verification. The simulation results show that the prediction effect of ensemble learning was better than that of the single 1DCNN-LSTM model. The proposed method was compared with other commonly used PV power prediction models. In the data case of this study, it was found that the proposed method reduced the prediction error rate by 1.48%, 11.4%, and 6.45%, compared to the LSTM, CNN, and BP, respectively. Experiments show that model prediction results considering the selection of similar data at multiple resolutions can provide more extensive information to an ensemble learner and reduce the deviation in model predictions. Therefore, the proposed method can provide a reference for PV integration into the grid and participation in market-based electricity trading, which is of great significance.