Enhanced Solar Forecasting with Contrastive Learning Model: A 15-Minute Prediction
Part Of
Conference Record - IAS Annual Meeting (IEEE Industry Applications Society)
Start Page
1
End Page
7
ISBN
9781665457767
Date Issued
2025-06-15
Author(s)
Abstract
Accurate short-term solar photovoltaic (PV) forecasting remains challenging due to rapid fluctuations in irradiance caused by cloud dynamics. To address this, we propose CL-SUNSET, a novel 15-minute-ahead forecasting framework that integrates self-supervised contrastive learning with a CNN-based regression model, effective learning from unlabeled ground-based sky images. The model first extracts features using a contrastive learning module encoder, which are then combined with CNN-extracted spatial features and historical PV logs for regression. Empirical evaluations on two datasets - a subtropical climate dataset, Taipei-based, and a warm Mediterranean dataset, Stanford. The results demonstrate the superior performance of CL-SUNSET. On the Taipei-based dataset, it achieves an Root Mean Squared Error (RMSE) of 14.266 W and a forecast skill (FS) of +5.17%, outperforming the baseline CNN with -14.51% FS. On the Stanford dataset, CL-SUNSET attains an RMSE of 3.081 kW and an FS of +30.48%, compared to -14.20% from the CNN. The model maintains robustness across weather conditions and forecast horizons, with FS increasing from +1.85% (5 min) to +10.37% (30 min) in Taipei, and from +15.88% to +49.83% in Stanford. Furthermore, t-SNE visualizations of contrastive representations reveal coherent temporal clustering, indicating that the model learns temporal continuity without providing any time series. These results highlight the promise of contrastive self-supervised learning for enhancing forecast accuracy, data efficiency, and interpretability in PV power prediction.
Event(s)
2025 IEEE Industry Applications Society Annual Meeting, IAS 2025. 15 June 2025 - 20 June 2025, Taipei.
Subjects
Deep Learning
Energy Systems Integration
Interpretable AI
Self-Supervised Learning
Solar Power Forecasting
Publisher
IEEE
Type
conference paper