AI-Based Rooftop Solar Access Values: Equivalence Testing Against a Drone-Acquired Reference

This study tested whether AI-generated solar access values (SAVs) may be statistically equivalent to drone-acquired SAV measurements for residential rooftop design and found equivalence at the strictest practical tolerance evaluated. Solar access value quantifies the fraction of unshaded annual solar resource available at a rooftop location after accounting for shading and sky obstruction. SAV is central to residential solar and roofing design, panel layout, and proposal quality. Remote, AI-based SAV workflows can reduce field effort and accelerate quoting relative to drone-acquired surveys, but their operational use depends on whether they produce SAV estimates that are close enough to reference measurements to support equivalent design decisions. This report evaluates whether SAVs generated by an AI-workflow are practically equivalent to SAVs generated by a drone-acquired reference workflow using a matched dataset of seven buildings, 100 rooftop locations, and monthly, seasonal, and annual reporting slices across California and Texas. Because the dataset contains multiple rooftop locations per building and repeated observations through time, equivalence is assessed using a dependence-aware two one-sided tests (TOST) workflow rather than independence-based uncertainty estimates. In the global monthly analysis, the estimated mean paired difference is 0.765 SAV for drone-acquired SAVs relative to AI-derived SAVs. The dependence-aware spatiotemporal confidence interval is [0.644, 0.904] and the bootstrap sensitivity interval is [0.612, 0.938]. Both intervals lie fully within the strictest equivalence region (-1, +1). Equivalence is uniformly supported at margins three, five, and 10 SAVs across monthly, seasonal, annual, and building-level analyses. At margin 1 SAV, conclusions become sensitive to slice- and roof-level uncertainty. The study demonstrates credible SAV equivalence assessment requires uncertainty models that account for clustered, spatial, and spatiotemporal dependence.