Where Solar Surges Next: Distribution-Planning Forecasts from Zero-Inflated Neural Models

Distribution utilities face a planning paradox: residential rooftop PV arrives in spatially clustered, temporally volatile surges that strain local infrastructure, yet standard forecasting methods provide only point estimates without quantifying plausible ranges or identifying high-uncertainty neighborhoods. At census block-group resolution, annual new installations are severely zero-inflated and overdispersed—roughly half of California block-groups added no systems in 2016—so MSE-trained regressors systematically smooth over the long-tail surges that actually drive feeder constraints. We present SolarCast, a zero-inflated negative binomial (ZINB) spatiotemporal neural network that jointly models the zero-generation process and count intensity in a single likelihood, producing one-year-ahead distributional forecasts with calibrated prediction intervals at neighborhood scale. SolarCast encodes multi-year adoption histories with neighbor peer effects, socioeconomic and satellite-derived built-environment features, and learned year embeddings that capture common statewide temporal effects observed in the training period. Evaluated on 2016 across 23,133 California block-groups, SolarCast attains 95% empirical coverage at the 90% nominal level, MAE of 5 installations among non-zero block-groups, and explains substantial variation at both neighborhood (R2 = 0.50) and county scales (R2 = 0.84). Unlike deterministic baselines, SolarCast lets utilities size equipment for plausible adoption ranges, stress-test plans under percentile scenarios, and flag high-uncertainty neighborhoods—capabilities useful for risk-aware infrastructure decisions.