Modelling, simulation and optimisation of agrivoltaic systems: a summary of a comprehensive review

Omer Jerusalem | 11/2026

In this essay I give my selective reading of the comprehensive review:

Zainali, S. et al. Modelling, simulation, and optimisation of agrivoltaic systems: a comprehensive review. Appl. Energy (2025).

The review divides the modelling of an agrivoltaic field into 5 headings:

1. Radiation
2. Photovoltaic
3. Microclimate
4. Crop
5. Optimisation

(On the subway map, where two lines meet, the output of the first model becomes the input of the next.)

In this essay I go into depth only on the first stage, radiation, and the five components the review describes for it. After that I expand a little on the integration model LER and point out, of all things, the drawbacks of using it on its own. I hope this reading encourages you to dig into the rest of the text. The interactive thinking map covers every part of the review and helps steer you to the topic you want to explore.

Radiation

Solar radiation is the energy source that drives (almost) everything. From photosynthetic assimilation to the coal-fired power station, by way of wind turbines and ocean waves, it all started with the sun. Look at it closely and you see that radiation is not a generic phenomenon that travels from point A to point B, but a component whose properties can be catalogued and meaningfully separated. It is the base of the pyramid for a system that relies on this inexhaustible source of energy.

Decomposition: Total radiation (GHI) is made up of direct radiation (BHI or DNI), the beam that travels in a straight line from the sun's disc, through the atmosphere, to the pyranometer; and diffuse radiation (DHI), the radiation scattered toward the pyranometer by particles in the upper layers of the atmosphere, dust particles in the lower layers, and clouds. For the decomposition stage alone, the review mentions 9 different models (in reality there are more than 140). They differ in accuracy, in complexity and in computational "weight". We will see later why this stage is so critical.

Transposition: Once we have split the sunlight into its two components, it hits the panels and the crop. Models at this stage calculate which part of the light reaches where: the direct light and the diffuse light. Direct radiation hits the panels and is absorbed; diffuse radiation also reaches the shaded crops. There are two families of models here, isotropic (2 models in the review) and anisotropic (4 models in the review), which simulate either a uniform sky or radiation whose distribution changes with the angular distance from the sun.

Shading: Perhaps the most painful topic when talking about PV in general and dual use in particular. Models that come from ground-mounted PV plants calculate the mutual shading between panels, which can significantly reduce the units' output. Here too there are two modelling approaches with the same familiar trade-off: simplicity and efficiency versus accuracy and reliability. 2D/3D view-factor models have a low computational load and can run thousands of simulations very quickly. To do so they rely on simplifying assumptions such as infinite rows (ignoring the end-of-row effect), uniform reflection of radiation from the ground (there is shading there too) and simplified plant structures. The second approach is called ray-tracing. Here the model builds a detailed virtual environment of the system's structure, the ground albedo and the topography, including the absorption, transmission and reflection coefficients of the system's different components. It then launches light rays and calculates their paths from the source (the sun and the sky) to the target surface. For direct radiation on the panels, the two approaches give fairly similar results. The significant difference for electricity production is in bifacial panel systems: on the underside of the panel, which receives radiation reflected from the ground and the crops, ray-tracing is more accurate by over 8% compared with the view-factor approach.

PAR: Photosynthetically active radiation, at wavelengths of 400–700 nm. We learned that in first year. So why does it matter here? Remember decomposition, above? Direct light and diffuse light do not have the same composition in terms of wavelength. The panel layout also affects and changes the composition of the light reaching the crops under the panels. And that will affect how much yield we get from the plot at the end of the season. The review cites 8 models that deal with this specific topic alone.

Validation: We have outlined dozens of models that deal only with the distribution of radiation as it passes through the atmosphere, scatters and hits the panels and the plants. Validation is the field test: how well does the model match the measurements taken by sensors in the field? Here too, the picture is complex. Models that focus on the amount of energy reaching the panels (decomposition, transposition and shading) usually achieve good results. Predicting the amount and composition of light reaching the plants is less promising. Calibration and comparison with field measurements are needed to reach satisfactory results. Under changing weather, the split between direct and reflected radiation is very hard to predict and requires calibration in the field.

That concludes a short summary covering only the first part of the review. From here we can start to derive the distribution of radiation reaching the panels and the crop, and from there the microclimate in the space between the crop and the panels, which in turn feeds back on electricity output and crop yield.

And one more "small" detail: all this mess means nothing to anyone unless all the numbers are tied to money at the end. So here I jump quickly to the fifth and final part of the review: optimisation and integration.

In practice, this means chaining the models of the different stages vertically to get a final simulation that estimates the amount of energy and the amount of yield. Only then can we look at the most important bottom line: profit.

LER: what it is and why it's a problem

When trying to assess the land-use efficiency of a combined plant, the most common and best-known KPI is LER, the Land Equivalent Ratio. LER compares the output of the combined solar system with that of a ground-mounted solar plant with the same characteristics, and the output of the shaded field with that of a regular agricultural field. It then adds the two parts together. When the sum is above 1, our solar field is supposedly more land-efficient than the conventional alternatives.

• Land Reduction Factor is a factor that corrects for the share of land taken up by the PV structure's posts.

The review cites an earlier study by P.E. Campana et al. that describes the problem with relying on this index alone. Because of how the equation is built, when a model is asked to design a combined system by LER, the result is a field with very narrow row spacing, because the PV term of the equation jumps dramatically compared with the drop in the agricultural term. LER on its own will prefer to maximise electricity production over agricultural produce.

Beyond that, dense panels cause mutual shading and lower electricity production relative to installed capacity (PVw). The further apart the rows, the larger the agricultural term of the equation: more radiation for the crops. The closer the rows, the larger the mutual shading between panels. Electricity production per unit of field area will increase, and with it LER, but production per square metre of panel will decrease.

To sum up, this essay does not cover even a small part of the extensive review carried out by the authors.

The review describes 5 stages in modelling an agrivoltaic operation. Each has dozens of possible models, and all of them fall into one of two categories: empirical models and mechanistic models. The trade-off always looks similar. Empirical models are simple to understand, cheap and quick to compute, but compromise on accuracy. Mechanistic models are remarkably accurate, but heavy, expensive and complex. As with any complex system, there is no silver bullet that answers every case, and the golden path probably runs through the middle: integrating the different types according to the characteristics of the site and the crops.

Finally, remember that this chain of models, with all its physical and agronomic depth, will not become a decision-support tool unless we pass its insights through the LCOE filter, put the electricity tariff into Excel and derive the business insights we need. But that is a topic for another article.

So what does it mean in practice? Building an agrivoltaic plant is a complex, intricate and expensive operation.

At the planning level:

• Decisions for 25 years. When setting up a project, we would rather invest in more dynamic and heavier models, to reduce the error as much as possible.
• A plant in the western Negev is not like a plant in the Galilee or on the Golan. Copy-paste won't work here. The design has to fit the climate, the topography, the soil, the crop type and so on. Errors that carry through will cost dearly.
• A model will never be 100% accurate, but it can give a sense of direction and explain which factors are critical in the specific plot we are designing. Define for yourselves the deviation you are willing to tolerate. It will show up in the ROI. This is all the more true as the price per kWh falls.
• Yield: no doubt it plays a marginal role in the bottom line, until you reach the threshold set by the Ministry of Agriculture. We won't want to gamble a plant whose construction cost runs to tens or hundreds of millions for a few more megawatts a year. Again we are back to the question of tolerable deviation.

At the operational level:

• The trackers run automatically, but the brain behind that automation will always be an integrative model that combines sub-models fed by the sensors in the field. And the output will be an angle. Control will always be automatic, except for maintenance events and farming operations.
• The better we get at collecting the right agricultural, meteorological and microclimate data, the more we will save on input costs. Just make sure we don't drown ourselves in sensors, detectors and graphs that end up hiding the forest.
• Integrated interfaces (dashboards) that we can adapt to the preferences of the grower and the operations team will play a decisive role here. A graph buried in a folder won't volunteer to show itself. Dynamism and adaptability matter a great deal.

In the end, the question to ask is: how much is the model a black box that spits out finished decisions, and how much do we really understand what it bases them on, and why it makes the decisions it makes.

Because we are the ones who pay the price, or take the profit, not the model.

References:

1. Zainali S, Ma Lu S, Fernández-Solas Á, Cruz-Escabias A, Fernández EF, Zidane TEK, et al. Modelling, simulation, and optimisation of agrivoltaic systems: a comprehensive review. Appl Energy 2025;386:125558. https://doi.org/10.1016/j.apenergy.2025.125558.
2. Campana PE, Stridh B, Amaducci S, Colauzzi M. Optimisation of vertically mounted agrivoltaic systems. J Clean Prod 2021;325:129091. https://doi.org/10.1016/j.jclepro.2021.129091.