AI learns how buildings shade themselves to predict facade solar power

As cities run out of roof space, the next great solar frontier is vertical. Building-integrated photovoltaics, or BIPV, turn walls into power plants, but predicting how much electricity a facade will actually produce has always been fiendishly difficult. Unlike a rooftop array with a single tilt and orientation, a facade experiences wildly different solar conditions from floor to floor: lower storeys may sit in the shadow of neighbouring buildings while upper floors bask in open sky. A new study published in Results in Engineering tackles this problem head-on with an occlusion-aware machine learning framework that predicts annual facade energy yield across climates, and its results suggest designers may soon be able to screen thousands of facade configurations in seconds.

The research team, led by Xinsen Wang and colleagues, assembled a hybrid dataset of 12,080 BIPV facade cases spanning 62 cities in 41 countries and four major Köppen climate groups. The data combine weather-driven physical simulations using the pvlib solar engineering library, converted records from PVsyst and other photovoltaic references, and an independent compilation of real building projects. After rigorous quality control that enforced monthly-to-annual energy conservation and excluded saturated shading cases, the final dataset covers three module families—polycrystalline, monocrystalline and thin-film—and installation heights ranging from roughly one metre to nearly 45 metres. The prediction target is deliberately practical: annual electricity generation per square metre of active PV area, expressed in kilowatt-hours per square metre per year, which places configurations of different sizes and technologies on a common productivity footing.

The real innovation lies in how the framework represents occlusion geometry. Each facade configuration is described not just by its orientation and climate, but by a set of nine height-and-shading descriptors: installation height, floor index, obstruction height, horizontal separation, obstruction elevation angle, average shading fraction, signed clearance, positive clearance and relative height. The geometry works by comparing the sun’s apparent elevation against the angle subtended by a neighbouring obstruction, but only when the sun lies within the facade’s front-facing sector. Solar positions are computed hourly using the Reda-Andreas algorithm, and direct beam irradiance is screened out whenever the sun dips below the obstruction angle. A separate time-weighted shading fraction captures how much of the front-facing daylight hours are geometrically blocked, and engineered features multiply annual irradiation by the unshaded time fraction to approximate the resource that actually reaches the modules.

With the feature matrix in place, the team benchmarked nine machine learning models under a common protocol, ranging from a simple Ridge regression baseline to modern gradient-boosted ensembles including XGBoost, LightGBM and CatBoost. The clear winner was ExtraTrees, an extremely randomized tree ensemble, which achieved a test coefficient of determination of 0.856 and a mean absolute percentage error of 9.57 percent under a city-grouped holdout split—making it the only candidate to break the ten percent error barrier. In an additional five-fold city-grouped evaluation covering every record, ExtraTrees performed even better, with a pooled error of just 7.10 percent and a coefficient of determination of 0.907. The linear and AdaBoost baselines trailed far behind, confirming that the nonlinear interactions among climate, azimuth, module type and shading geometry demand tree-ensemble machinery.

Accuracy on held-out data is one thing; transferring to unseen cities and climates is another. The validation ladder therefore included leave-one-city-out tests, in which the model must predict for an entire city it has never seen, and even leave-one-climate tests that withhold whole climate groups. Under the ExtraTrees protocol, leave-one-city-out prediction achieved a pooled error of 6.95 percent. Withholding complete climate groups revealed a telling gradient: tropical climates were easiest to predict at 7.18 percent error, while continental climates proved hardest at just over 15 percent. The authors note that much of this apparent difficulty stems from a small-denominator effect—low-yield configurations show larger percentage errors but smaller absolute errors—so the continental penalty is less alarming than it first appears. Removing latitude from the feature set, meanwhile, nearly doubled the pooled error, underscoring how much composite geographic and solar-path information that single variable carries.

To disentangle what each feature block actually contributes, the researchers ran paired ablation experiments on identical data folds. Climate-only models failed spectacularly, with errors above 33 percent. Adding facade and module information cut the error roughly in half, and introducing average shading slashed it again to around 7.5 percent. Interestingly, adding the full explicit obstruction geometry on top of average shading yielded almost no further gain—a difference of just 0.016 percentage points, with a confidence interval spanning zero. The full engineered set, however, did deliver statistically significant improvements of roughly 0.4 to 0.5 percentage points over reduced variants. The lesson is subtle: average shading already encodes much of the height information, but the combined system of geometry, effective irradiation and module characteristics is what truly drives accuracy.

Interpretability came through Shapley additive explanations, or SHAP, which decompose each prediction into per-feature contributions. Across the nine models, latitude, facade azimuth, module type and average shading repeatedly emerged as the dominant drivers, with the ExtraTrees analysis ranking thin-film module encoding, latitude and azimuth as its top three features. Spline-smoothed dependence curves located zero crossings—for example, average shading contributions flip sign near a shading fraction of about 0.31, and obstruction-angle crossings appear near 20.5 and 58.1 degrees. Crucially, the team tested whether these attribution rankings were stable: refitting the model across five outer city folds produced importance-rank correlations between 0.973 and 0.995, suggesting the identified design priorities are robust rather than artifacts of a single fit.

The framework was then confronted with reality in Hohhot, China, on a real south-facing facade instrumented across five floors with three equal-power module technologies: polycrystalline silicon, cadmium telluride and monocrystalline silicon. Against a pvlib physical reference for the annual case, the model achieved 10.38 percent error with a coefficient of determination of 0.872. For April and May field measurements, seasonally allocated predictions matched observations with 7.28 percent error, and all floor-module comparisons fell within a ten percent envelope. Measured yield climbed steadily with height, from about 2.49 kilowatt-hours per kilowatt-peak per day on the first floor to 3.22 on the fifth, and the predictions tracked this vertical gradient closely. Notably, the model preserved floor ordering far more reliably than module ordering, hinting that in real buildings, choosing the right floor matters more than agonizing over equal-power module choices.

The engineering payoff was quantified through candidate-screening and timing experiments. Across twelve city-excluded selection tasks combining four weather contexts with three obstruction scenarios, the surrogate achieved a mean reference-based selection loss of just 2.34 percent, compared with 7.81 percent for a coarse resource-shading score, and it picked the reference-optimal alternative in seven of twelve tasks. Timing tests showed that evaluating 1,000 candidates takes under two seconds even in a cold-start workflow—roughly five times faster than the physical reference—and over twenty times faster once weather data and models are cached. Single queries, however, remain slower than direct physical calculation, so the speed advantage is fundamentally a batch-screening benefit.

The authors are candid about limitations: the current geometry uses single-point hourly shading without obstacle width, multiple overlapping buildings or partial module shading, and a sensitivity experiment showed that simply changing the geometric representation can shift reference yields by up to 65 percent for narrow obstructions. Complex urban canyons will still need LiDAR, GIS or full building models for detailed evaluation. Yet the framework’s core message stands: by folding climate, floor-level obstruction geometry and module technology into one interpretable, transfer-tested model, facade solar assessment moves from aggregate resource estimation to genuine configuration comparison. For architects weighing which floors of a renovation deserve PV cladding, that could mean the difference between a facade that merely looks green and one that actually earns its keep.

Subject of Research: Occlusion-aware machine learning for climate-adaptive prediction of building-integrated photovoltaic facade energy yield

Article Title: Occlusion-aware machine learning for climate-adaptive BIPV facade yield prediction

Article References: Wang, X., Aoqier, B., Liu, Y., Li, G., & Xu, G. (2026). Occlusion-aware machine learning for climate-adaptive BIPV facade yield prediction. Results in Engineering, 32, Article 113363. https://doi.org/10.1016/j.rineng.2026.113363

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113363

Keywords: building-integrated photovoltaics, BIPV, machine learning, facade solar energy, shading geometry, ExtraTrees, SHAP, cross-city transfer, pvlib, urban solar potential, energy yield prediction, Köppen climate groups