PV-MMCD: Multimodal Conditional Diffusion for Probabilistic Photovoltaic Power Forecasting Based on Multi-Source Data
Year: 2026
Type: article
Abstract: Accurate photovoltaic (PV) power forecasting is essential for enhancing grid stability and optimizing economic operation. Existing methods often fail to adequately capture the intrinsic connection between cloud-driven variability and the resulting uncertainty in PV power generation. To address this limitation, this paper proposes PV-MMCD, a multimodal conditional diffusion model that improves probabilistic PV power forecasting by leveraging cross-modal correlations and cloud-to-power uncertainty propagation. A neural SDE–based module further extrapolates latent PV power and cloud features as dynamic conditions to strengthen the temporal uncertainty modeling. In the reverse process, a multimodal network progressively denoises the target sequences by integrating satellite imagery and NWP data through cross attention and feature-wise linear modulation, enabling more realistic PV power reconstruction. Once trained, PV-MMCD generates ensemble forecasts from multiple noise initializations, which are calibrated to produce reliable probabilistic forecasts. Experimental results across three geographically distinct PV sites show that PV-MMCD achieves a probabilistic forecast skill score of 61.7% over the 0–4 hour horizons, outperforming other multimodal forecasting methods and demonstrating strong robustness under diverse sky conditions.
Source: IEEE Transactions on Smart Grid
Language: English
Field: Energy
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7214737859