Short-term Solar Power Forecasting: An Accuracy Comparison of Machine Learning Methods versus Typical Daily Curve

In this study, we perform short-term power generation prediction for the fluctuation of seasonal new energy power generation, and firstly, we explore the shortcomings of the traditional K-means clustering method in the prediction of solar power generation. Then, in order to improve the prediction accuracy and model robustness, we systematically evaluate a variety of advanced machine learning models, such as Random Forest(RF), Back Propagation Neural Network (BPNN), Stacked Regression(STAC), Linear Regression (LiR), Random Forest(RF) and XGBoost(XGB), BPNN+STAC, Support Vector Machine (SVM), Gradient Boosted Decision Tree (GBDT), BPNN+STAC, Long-short-term memory network (LSTM), Gated Recurrent Unit (GRU), etc., and ultimately select the best-performing BPNN+STAC combined model. Based on the results, we compare the performance of BPNN+STAC model with typical daily curve by K-means method, by quantifying the key indexes of MAE, RMSE, RMAPE and $\mathrm{R}^{2}$, the results show that the BPNN+STAC model has a significant advantage in terms of prediction accuracy and stability, and the MAE and RMSE are reduced by 44.55% and 50.12%, respectively, and the $R^{2}$ is improved by $49.23 \% .49 .23 \%$. This study not only provides a new theoretical framework for the accurate prediction of wind power generation, but also demonstrates its strong potential for practical application in virtual power systems and its wide spread value.