One-hour-ahead solar irradiance forecast based on real-time K-means++ clustering on the input side and CNN-LSTM

Abstract

With the increasing proportion of photovoltaic power generation in the power system, accurate solar irradiance forecasting is crucial for power grid scheduling. The paper proposes a approach of clustering modeling and joint forecasting for solar irradiance: firstly, by using the real-time K-means++ to cluster the GHI sequence on the input side of the model, the model input is divided into 4 clusters. The number of clusters and clustering methods are quickly determined through sensitivity experiments. Then 4 CNN-LSTM models, corresponding to the 4 clusters, are separately established and trained to fully extract multivariate and multi-temporal information from the input side. These 4 models are combined to achieve 1-h-ahead solar irradiance forecast with higher accuracy. To verify the effectiveness of the proposed method, Auto regressive (AR), Convolutional neural network (CNN), Long short-term memory (LSTM) and Gradient boosting decision tree (GBDT) are used as comparative models. The results indicate that the proposed K-means++_CNN-LSTM is a feasible method, with MAE decreasing by about 9.13%, RMSE decreasing by about 6.58%, and $\rho$ increasing by about 1.46%, compared to K-means++_AR. Finally, T-test is used to verify the positive impact of the proposed framework on model accuracy.