SolCAD-Net: Solar cloud advection dynamics for very-short-term photovoltaic power forecasting

Very-short-term photovoltaic (PV) power forecasting is essential for real-time grid operation, yet minute-ahead ramps are largely driven by transient cloud occlusion near the Sun. Satellite products and numerical weather prediction often lack the spatiotemporal resolution needed to capture such local cloud dynamics, motivating the use of ground-based all-sky images. However, global sky-texture modeling can dilute ramp-relevant circumsolar cues and increase sensitivity to background clouds and camera-dependent domain shifts, while strong power-history correlations may cause autoregressive shortcut learning. We therefore propose SolCAD-Net, a physics-guided multimodal framework that models Solar Cloud Advection Dynamics for ramp-aware PV nowcasting. First, a Sun-centered Polar48 encoder projects segmented cloud masks into Sun-referenced polar bins to represent circumsolar geometry. Second, wind-driven hard rotation and Advection-Guided Aerodynamic Cross-Attention canonicalize the upwind direction and apply a speed-conditioned radial bias to emphasize upstream cloud structures. Third, an upwind lookahead auxiliary objective promotes sensitivity to short-term changes in upwind cloud occupancy and suppresses power-history shortcuts. On the Stanford SKIPP’D benchmark, SolCAD-Net achieves an MAE of 0.3492 kW, an RMSE of 1.1490 kW, and a WAPE of 2.9967% at the 5-min horizon. On the top 10% high-magnitude ramp subset, it obtains an R 10 -MAE of 2.3405 kW and an R 10 -RMSE of 3.5180 kW, outperforming all compared methods. It also remains effective at 15 min with an MAE of 0.5311 kW. An independent-site reproduction on the Gatton dataset using site-specific downstream training demonstrates the applicability of the framework under different camera and operating conditions, with the best overall and ramp-focused performance across all metrics.