ShadowSense: Edge Optimized Self-Supervised Learning for Dynamic Shadow Mapping of Solar Panels

We introduce ShadowSense, a self-supervised, edge-deployable framework for dynamic shadow mapping and short-term PV power forecasting. Unlike prior methods that require labor-intensive manual annotations or rely on overly complex physical models, ShadowSense uses transient power dips as weak supervision signals, enabling fully unsupervised learning from synchronized sky imagery and power measurements. The architecture integrates a masked spatiotemporal autoencoder to learn cloud motion dynamics, a differentiable geometric projection module to adapt to diverse panel orientations, and a causal alignment system to maintain temporal correspondence between cloud shadows and power responses. Evaluated over 122,400 observations across 92 days in Kaunas, Lithuania, the system achieved a 32.7% reduction in mean absolute error (MAE) compared to baselines, captured 92.3% of rapid power ramps, and operated with an average inference latency of 66 ms and energy cost of 0.52 J per prediction on a solar-powered NVIDIA Jetson Nano, demonstrating scalable, real-time PV forecasting in resource-constrained environments.