Deep Learning Multi-Horizon Irradiance Nowcasting: A Comparative Evaluation of Three Methods for Leveraging Sky Images
In the realm of renewable energy, particularly solar energy, accurate predictions of global horizontal irradiance (GHI) are crucial for optimizing energy generation and management. Recent advancements in deep learning (DL) have opened new avenues for improving irradiance nowcasting, particularly through the use of all-sky imager (ASI) images. A recent study, documented in arXiv:2603.26704v1, investigates three distinct methodologies for integrating ASI images into DL models aimed at enhancing GHI forecasts.
Methodologies Evaluated
The study explores three innovative approaches to leverage ASI images in the context of irradiance nowcasting:
- Method 1: Direct Feature Extraction with CNN
This approach utilizes a convolutional neural network (CNN) to directly extract features from raw RGB images obtained from ASIs. By processing the images in their native form, the model aims to capture essential visual information that can correlate with irradiance levels.
- Method 2: Engineered 2D Feature Maps
The second method employs advanced algorithms to create 2D feature maps that incorporate domain knowledge. Features such as cloud segmentation, cloud motion vectors, solar positions, and cloud base heights are utilized to inform these maps. The CNN then processes these engineered inputs to extract compound features that are more relevant for irradiance predictions.
- Method 3: Time-Series Input from Engineered Features
The final methodology aggregates the engineered 2D feature maps into a time-series format. This temporal aspect allows the model to analyze changes over time, potentially enhancing the accuracy of multi-horizon forecasts.
Evaluation and Results
All three methods were employed within a deep learning framework trained on a high-frequency dataset spanning 29 days. The primary objective was to generate multi-horizon forecasts of GHI, with predictions extending up to 15 minutes ahead. To assess the performance of each model, the researchers utilized root mean squared error (RMSE) and skill score metrics over seven carefully selected days of data.
Key Findings
The results revealed significant insights into the effectiveness of the different methodologies:
- Aggregated engineered ASI features as model input outperformed the direct CNN approach, indicating that the integration of domain knowledge into the feature engineering process can lead to more accurate forecasts.
- The study highlighted that complex spatially-ordered DL architectures are not strictly necessary for effective irradiance nowcasting. Instead, simpler alternative image processing methods can yield comparable, if not superior, results.
- By emphasizing the potential of engineered features, the research opens new pathways for improving spatial DL feature processing methods, thereby enhancing the predictive capabilities of solar irradiance models.
Conclusion
This comparative evaluation underscores the importance of leveraging ASI images in deep learning frameworks for solar irradiance nowcasting. The findings advocate for further exploration of feature engineering and alternative processing methods, reinforcing the notion that innovation in data utilization can lead to significant advancements in renewable energy forecasting.