SkyGPT: Probabilistic Short-term Solar Forecasting Using Synthetic Sky Videos from Physics-constrained VideoGPT (2306.11682v1)

Abstract: In recent years, deep learning-based solar forecasting using all-sky images has emerged as a promising approach for alleviating uncertainty in PV power generation. However, the stochastic nature of cloud movement remains a major challenge for accurate and reliable solar forecasting. With the recent advances in generative artificial intelligence, the synthesis of visually plausible yet diversified sky videos has potential for aiding in forecasts. In this study, we introduce \emph{SkyGPT}, a physics-informed stochastic video prediction model that is able to generate multiple possible future images of the sky with diverse cloud motion patterns, by using past sky image sequences as input. Extensive experiments and comparison with benchmark video prediction models demonstrate the effectiveness of the proposed model in capturing cloud dynamics and generating future sky images with high realism and diversity. Furthermore, we feed the generated future sky images from the video prediction models for 15-minute-ahead probabilistic solar forecasting for a 30-kW roof-top PV system, and compare it with an end-to-end deep learning baseline model SUNSET and a smart persistence model. Better PV output prediction reliability and sharpness is observed by using the predicted sky images generated with SkyGPT compared with other benchmark models, achieving a continuous ranked probability score (CRPS) of 2.81 (13\% better than SUNSET and 23\% better than smart persistence) and a Winkler score of 26.70 for the test set. Although an arbitrary number of futures can be generated from a historical sky image sequence, the results suggest that 10 future scenarios is a good choice that balances probabilistic solar forecasting performance and computational cost.

• The paper introduces SkyGPT, a two-stage framework that integrates VideoGPT with physics-based constraints to generate realistic cloud dynamics.
• It leverages a vector quantized variational auto-encoder and a conditional transformer to produce diverse, plausible future sky scenarios.
• SkyGPT outperforms benchmarks like SUNSET in metrics such as CRPS and Winkler Score, leading to more reliable short-term solar forecasting.

Probabilistic Short-term Solar Forecasting with SkyGPT

The research paper titled "SkyGPT: Probabilistic Short-term Solar Forecasting Using Synthetic Sky Videos from Physics-constrained VideoGPT" presents an innovative advancement in short-term solar power prediction through the development of the SkyGPT forecasting system. This two-stage deep learning framework leverages generative models to predict cloud dynamics with particular attention towards generating plausible sky videos, providing a novel approach to photovoltaic (PV) power forecasting.

Key Contributions and Methodology

The primary contribution of this work is the development of SkyGPT, which combines the strengths of VideoGPT with the inclusion of physics-informed constraints from the PhyDNet model architecture. This setup allows for generating diverse and realistic future sky videos by modeling cloud dynamics effectively using past sky image sequences. SkyGPT utilizes a vector quantized variational auto-encoder (VQ-VAE) to compress high dimensional video data, and a conditional transformer is used to predict future tokens, facilitating the generation of numerous possible future scenarios. The incorporation of physical laws via PhyCell in the transformer allows it to retain the dynamics of cloud motion which is crucial for accurate forecasting.

In the subsequent stage, these predicted sky images are input into a subsequent convolutional neural network (CNN)-based framework, specifically a modified U-Net architecture, to estimate future PV output. This setup contrasts with traditional deterministic models by offering a probabilistic forecasting approach that attempts to capture uncertainty through the stochastic generation of cloud scenarios.

Results and Performance

Experimental evaluations, as discussed, demonstrate the effectiveness of SkyGPT in generating realistic sky imagery and enhancing solar forecast accuracy. When benchmarked against existing methods such as SUNSET and smart persistence models, the proposed methodology displayed improved performance in metrics such as Continuous Ranked Probability Score (CRPS) and Winkler Score (WS).

Particularly, the stochastic nature of SkyGPT shines through in cases showing highly variable weather conditions, where deterministic models typically fail to accurately track sudden changes in cloud cover. The generation of multiple future scenarios in SkyGPT contributes to superior probabilistic prediction by covering a wider range of potential future states, resulting in increased reliability and sharper predictions.

Implications and Future Directions

The implications of this research are manifold, both for practical applications and theoretical advancements in AI-driven forecasting. Practically, improved short-term solar forecast accuracy can aid in the stability and operational efficiency of power grids heavily reliant on solar sources, reducing reliance on supplementary dispatchable resources. Theoretically, the integrations of physics-informed constraints into generative models open new pathways in reliably modeling stochastic processes in environmental settings, suggesting similar approaches may be beneficial across various domains involving fluid dynamics or other meteorological phenomena.

Future explorations could emphasize refining the physics-constrained components for enhanced model competence in complex situations. Investigating the integration of additional meteorological data sources and improved transfer learning techniques to accommodate varying environmental conditions across geographic regions is another promising direction to amplify this framework’s generalizability. Furthermore, enhancing the PV output prediction model sophistication could advance the frontiers of accuracy, particularly through leveraging novel architectures like Vision Transformers.

In conclusion, SkyGPT presents significant progress in probabilistic solar forecasting, showcasing the potential of generative AI techniques when augmented with domain-specific knowledge constraints. This paper exemplifies the convergence of AI and physical science, reinforcing the notion that collaborative integration can spearhead cutting-edge developments in forecasting technologies.