Abstract
Integrating solar power into the power grid while maintaining stability is essential for the widespread adoption of clean energy. This necessitates accurate solar irradiance forecasting, which directly determines solar power output. This paper addresses the challenge of short-term solar irradiance prediction using historical sky image sequences. Unlike existing machine learning methods that directly regress irradiance values, we propose a novel deep neural network that leverages a physical clear-sky model. Our architecture utilizes a Vision Transformer to encode spatial and temporal information from sky videos. The network is then trained to predict the solar irradiance residual relative to the clear-sky model's output. Extensive evaluation on benchmark datasets like TSI880 and ASI16 demonstrates that our approach significantly outperforms existing algorithms in both nowcasting and forecasting tasks up to four hours ahead.
Introduction
The transition to renewable energy sources is hindered by the inherent variability of solar power. Rapid changes in cloud cover can cause sudden drops in irradiance, challenging grid stability. Consequently, grid operators require precise, short-term forecasts of Global Horizontal Irradiance (GHI).
Predicting solar irradiance is a complex task involving two distinct components: the deterministic astronomical movement of the sun and the stochastic dynamics of local weather. Traditional machine learning approaches often attempt to learn both simultaneously through direct regression. However, this is computationally inefficient as it forces the model to relearn well-established physical laws of solar position. Furthermore, standard sequence models like LSTMs often fail to capture the intricate spatial relationships between distant clouds and the sun's position within wide-angle sky images.
Methodology
We propose a hybrid physics-AI framework that combines the representational power of Vision Transformers (ViT) with the reliability of atmospheric physics. The core innovation lies in predicting the "residual"—the difference between the actual observed irradiance and the theoretical maximum under a clear sky.
Network Architecture
As illustrated in Figure 2, our model consists of several key components:
- Nowcasting Network: A ViT-based module extracts high-level feature representations from individual sky images. Transformers are particularly suited for this task because their self-attention mechanism can model long-range dependencies, allowing the network to relate cloud formations on the horizon to the sun's current position.
- Temporal Transformer Encoder: This module processes a sequence of feature representations (typically from the past hour) to capture the motion and evolution of cloud cover over time.
- Clear-Sky Integration: Instead of predicting GHI directly, the network outputs a residual value. This residual is added to the initial prediction provided by a deterministic clear-sky model. By "subtracting" the predictable solar cycle, the neural network can focus its learning capacity entirely on the complex, non-linear dynamics of cloud movement and atmospheric interference.
Experimental Results
The proposed model was evaluated on two benchmark datasets: TSI880, which uses a mechanical sun tracker, and ASI16, which provides a wider field of view without tracking.
Performance Comparison
Our method demonstrates superior performance across various time horizons. On the ASI16 dataset, the model achieved a normalized Mean Absolute Percentage (nMAP) error of 6.9% for nowcasting, a significant improvement over the 13.1% error reported for previous state-of-the-art methods. For the 4-hour-ahead forecasting task on the TSI880 dataset, our model consistently outperformed the baseline (Siddiqui et al.), as shown in the qualitative comparisons in Figure 6.
Key Insights
Ablation studies confirmed that the clear-sky residual approach is critical for accuracy. Models that attempted direct regression without the physical prior consistently showed higher error rates. Additionally, attention map visualizations revealed that the Transformer naturally learns to focus on the solar disk and approaching cloud fronts, validating that the model's predictions are grounded in physically relevant visual features.
Conclusion
This research demonstrates that integrating physical models into deep learning architectures can significantly enhance the accuracy and efficiency of solar irradiance forecasting. By leveraging Vision Transformers to handle complex spatial-temporal data and a clear-sky model to handle predictable solar cycles, we achieve state-of-the-art results. Future work will explore generative forecasting to predict future sky states and the integration of additional meteorological data, such as wind vectors and satellite imagery, to further refine long-term predictions.
Reference: Gao, H., & Liu, M. (2022). Short-term Solar Irradiance Prediction from Sky Images with a Clear Sky Model. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2475-2483.
Summary updated