Leveraging Generative Models for ASI-based Solar Nowcasting

Short-term variations in PV power are an increasingly important challenge for solar energy integration. By anticipating sudden changes in irradiance caused by passing clouds, all-sky imager-based solar nowcasting can help address this challenge. However, the utility of nowcasting systems is highly dependent on the quality of the forecast. While recent data-driven models have shown great potential in standard forecast metrics such as root-mean-square error (RMSE) and forecast skill, they tend to produce smoothed forecast curves and may not be well suited to detect ramps. An alternative data-driven approach lies in generative modeling. Instead of forecasting solar irradiance directly from available data, like radiometer measurements or sky images, we propose a two-step method to predict cloud dynamics and irradiance separately.

Using novel denoising diffusion models [1], we show that realistic sequences of sky images can be generated. By conditioning video prediction on the latest acquired sky images, plausible future sky conditions are produced. In contrast to traditional methods that only predict cloud motion, changes in cloud shape can also be represented. Another advantage of diffusion-based video prediction is the versatility of possible outcomes. By introducing samples of random noise during inference, the model generates different outputs that vary depending on the conditioned input.

In the second step, we apply an irradiance model to the generated synthetic sky images. Each image is processed independently and returns a corresponding irradiance value. Thus, an irradiance distribution can be obtained from the samples of synthetic sky images for each lead time. As a result, the uncertainty of the forecast can be estimated, since a larger variation of synthetic sky images will lead to a larger distribution of corresponding irradiance.

We evaluate our novel generative nowcasting approach not only on standard forecast metrics, but especially on its ability to detect ramp events. Preliminary results already indicate that such a generative video prediction on sky images in combination with an irradiance model can overcome the problem of smoothed forecast curves [2]. Furthermore, the intermediate results of synthetic sky images enhance interpretability, and the generation of varying scenarios enables probabilistic forecasting.

 

[1] Ho, Jonathan, Ajay Jain, and Pieter Abbeel. "Denoising diffusion probabilistic models." Advances in neural information processing systems 33 (2020): 6840-6851.

[2] Paletta, Quentin, Guillaume Arbod, and Joan Lasenby. "Benchmarking of deep learning irradiance forecasting models from sky images–An in-depth analysis." Solar Energy 224 (2021): 855-867.