| Keywords |
Dropout; PV; Machine learning; Confidence interval |
| Abstract |
This study investigates the effect of the Dropout technique on the generalizability of PV generation prediction models utilizing Artificial Neural Networks (ANNs). Using solar energy data from Daejeon, South Korea, covering 2017 to 2021, we developed three ANN-based predictive models with Dropout rates of 0%, 10%, and 20%. The Dropout method was implemented during the training phase to mitigate overfitting by randomly deactivating neurons in the network. The model with a 20% Dropout rate (Case 3) demonstrated the most consistent and reliable performance across various metrics. Although all models met the performance criteria based on CV(RMSE), those utilizing appropriate Dropout rates exhibited enhanced stability, particularly during specific timeframes. The results indicate that while Dropout effectively prevents overfitting and enhances prediction stability, variations in dropout rates do not necessarily lead to significant changes in overall model performance. This study highlights the importance of Dropout as a regularization technique within ANN frameworks and suggests further exploration of time-series analysis and other methods to improve solar power forecasting models. |