3D energy simulation: terrain, backtracking, and the numbers behind it

On sites with significant slope variation, standard flat-terrain energy models can overestimate yield because they cannot account for terrain-driven shading, row-level backtracking differences, or elevation-dependent irradiance distribution. The result is a gap between predicted and actual production that only becomes visible after construction, when correcting it costs far more than it would have during design.

The assumption that has been costing projects yield

Most utility-scale solar sites built in the 2010s were flat. That is not a coincidence. Flat ground made every part of the process simpler: grading, tracker installation, cable runs, and energy yield estimation. When the terrain is essentially a single plane, a plant-wide average irradiance value is a reasonable approximation. Errors are small because the geometry is simple.

That era of easy sites has largely passed. Land scarcity, environmental restrictions, and the push into new markets have moved projects onto terrain that would have been ruled out a decade ago: hillsides in Spain and Portugal, rolling landscapes in India and Japan, complex parcels in the Balkans. The geometry is no longer simple. But in many workflows, the modeling approach has not changed to match.

The problem is not that engineers are being careless. Standard energy yield models were built for simple terrain and they perform well under those conditions. The issue is what they cannot see. On uneven ground, rows at different elevations receive different irradiance. Trackers on a north-facing slope shadow neighboring rows at angles that a flat algorithm might not predict.

Where flat-terrain models may introduce error

The errors do not appear in isolation. They interact, and the interactions can mask each other in aggregate. A model that underestimates shading losses might simultaneously overestimate temperature losses, producing an annual total that looks accurate while the intermediate calculations are substantially wrong. That is why comparing only annual production figures between a model and a real plant can be misleading.

Four specific mechanisms drive most of the divergence on complex terrain.

First, terrain-amplified shading on fixed-tilt structures. On sites with slope variation, the effective shading geometry between fixed-tilt rows changes depending on where those rows sit on the terrain. A flat-terrain model applies a single inter-row shading assumption across the whole layout, which means it cannot distinguish between rows in favorable orientations and rows where the terrain increases shading exposure. The result is an averaged shading estimate that can look acceptable at the plant level while masking meaningful production differences across the site.

Second, backtracking. Single-axis trackers use a backtracking algorithm to reduce row-to-row shading at low sun angles by tilting back toward horizontal. A standard flat algorithm calculates the tracking angle assuming all rows are at the same elevation. On sloping ground, rows at different heights shadow each other at different angles and at different times. The flat algorithm either over-corrects or under-corrects depending on the row position relative to the slope, and the losses accumulate across the full year.

Third, bifacial rear-side irradiance. Rear-side generation depends on what the module sees behind it: ground albedo, reflected light from nearby structures, and the angle of incoming diffuse radiation. All three of these change with terrain elevation and slope. A fixed albedo value applied uniformly across a plant with gradient variation will produce rear-side estimates that are wrong in a directional sense, not randomly wrong, which means they do not average out.

Fourth, electrical mismatch. When rows at different positions on a slope receive different irradiance due to terrain-driven shading, the DC current in partially shaded strings is constrained by the weakest module. A plant-wide shading average cannot capture this because it loses the row-level variation that produces mismatch. RatedPower's 3D Energy simulation engine works module by module, calculating both the front-face and rear-face shading for each structure individually and carrying those results through the electrical model.

What a validated model actually requires

Correcting these errors requires modeling the physical reality of each site rather than approximating it. RatedPower's 3D Energy simulation engine was built around three technical decisions that address the mechanisms described above.

The irradiance calculation uses Radiance, a ray-tracing engine originally developed for architectural lighting simulation and now widely used in photovoltaic (PV) research. Unlike transposition models that calculate a single plane-of-array value, Radiance traces individual light rays from the sky and ground to each module surface, accounting for the actual geometry of the layout, including inter-row shading, near-shading objects, terrain slope, and rear-side reflected irradiance. The result is a module-by-module irradiance map rather than a plant average.

The terrain-aware backtracking algorithm replaces the flat-terrain stow angle with row-specific angles calculated from each tracker's actual position on the slope, which directly affects tracking angle optimization across the layout. Internal simulations comparing terrain-aware backtracking against standard flat backtracking on complex terrain showed energy yield differences of 1 to 3% in favor of terrain-aware backtracking. That range is not trivial: on a 100 MW plant, 1% of annual production represents a measurable difference in revenue that feeds back into financial models and debt service calculations.

The electrical model integrates with the irradiance calculation so that shading effects propagate through strings and inverters rather than being applied as a blanket derating. Partial shading of one string affects the DC bus in a way that depends on the current mismatch across that string, not on the average shading level across the plant.

What the validation results show

Claiming accuracy is straightforward. Demonstrating it against real metered production data from commissioned plants is considerably harder. Getting that data requires careful data cleaning to remove curtailment and maintenance intervals, and a methodology rigorous enough to distinguish model error from data error.

RatedPower's preliminary validation report, published in October 2025 and authored by Félix Ignacio Pérez Cicala, Diego Botija Slavicek, and Ignacio Álvarez Iberlucea, documents three stages of validation: internal testing of individual model components, internal yearly simulations, and, most importantly, comparison against metered production data from two commissioned plants. Both case studies are anonymized.

The first case study was a monofacial single-axis tracker (SAT) plant in central Spain, operating on mostly flat terrain. The model was built to match the real plant as closely as possible, using Spain's IGN terrain data and the plant's actual equipment specifications. Sanitized annual error, after removing curtailment and maintenance intervals from the dataset, was -0.09% in year A and -0.25% in year B. The normalized root mean squared error (NRMSE), which measures hourly accuracy rather than annual balance, was 1.91% and 2.21% respectively. The report also tested the sensitivity of these results to the sanitization threshold: raising the threshold from 20% to 60% still produced errors under 0.8%, confirming the accuracy is not an artifact of data cleaning.

The second case study was a bifacial fixed-structure plant in the Philippines. The data presented more challenges: only a few months were available, there were periods of missing irradiance data, and ground-measured albedo could not be obtained. The sanitized error across the available months was 1.79%, with an NRMSE of 3.20%. The monthly breakdown showed that 4 out of 5 months produced errors well under 2%, with the outlier month showing data quality issues rather than a systematic modeling error.

The report describes itself as preliminary, with further case studies planned. That is an accurate characterization. Two plants, even well-validated ones, do not constitute a comprehensive validation across all technology configurations and site types. The authors are transparent about data quality limitations and the assumptions required to compare real plant production against a simulation. That transparency is the reason the results are credible: a model that only reports its best outcomes is not one you can use for bankable yield analysis.

One finding from the internal validation stage is worth noting. When comparing the 3D model against RatedPower's previous energy yield model on complex terrain, the 3D model consistently found substantially more back-face irradiance, with differences of around 30% in some scenarios. Simpler models that assume flat terrain can systematically underestimate bifacial rear-side gains because they cannot account for how ground albedo and reflected irradiance change with slope. On bifacial projects in terrain with any meaningful gradient, that gap feeds directly into financial projections.

What this means for design decisions

The downstream consequences of yield error depend on how far into the development process the error is discovered.

At the feasibility stage, an overestimated yield inflates land value assessments and initial financial returns. Ground coverage ratio (GCR) decisions made on the basis of a plant-wide average shading assumption may not hold when row-level variation is accounted for.

On a site with slope variation, the shading geometry differs significantly from row to row, meaning a GCR that looks acceptable as a site average can produce unacceptable losses in unfavorable terrain sections. Layout choices, including tracker pitch and row orientation on sloped sections, are more reliably optimized when the model can show where production is higher or lower across the specific site geometry rather than returning a single average.

At the financing stage, lenders and independent technical advisors are asking increasingly specific questions about yield methodology. A bankable yield report needs to document its modeling assumptions clearly enough for the technical advisor to assess whether they are appropriate for the site. A model that runs module-by-module and produces a documented methodology, including validation results against real plants, is a different conversation from one that applies a plant-wide average with a conservative derating.

At the operating stage, if production falls short of the P50 estimate in an energy yield assessment (EYA), the question is always whether the gap is weather-related or model-related. A model validated against real metered data with documented error bounds makes that analysis more tractable.

For context on how 3D Energy fits into the broader RatedPower design workflow, the introductory piece on 3D Energy simulation covers the terrain modeling foundations in more detail.

3D Energy runs within RatedPower's platform, on the same layout used for the rest of the design process. There is no separate export, no rebuilding the plant in a different tool. A full terrain-aware simulation on a utility-scale layout typically completes in 10 to 20 minutes depending on plant size. The output includes a structured report with production figures, loss breakdowns, performance ratio by area, and simulation inputs, in PDF or other formats suitable for lender review.

See 3D Energy in your next project

The preliminary validation report, including the full methodology and case study results, is available to download. To see 3D Energy applied to a real project layout, request a demo and ask an engineer to walk through a terrain-aware simulation on a complex site.