Synergistic Optimization of Agrivoltaic Systems: A Review of Intelligent Equipment Control for Agricultural Production Adaptability

Abstract

Agrivoltaic systems configure photovoltaic power generation and agricultural production within the same land space, providing a new pathway for alleviating the conflict between energy development and farmland conservation; however, array shading and structural constraints also alter crop growth and agricultural equipment operating environments. Following the PRISMA process, this review searched studies published from 2010 to 2026 in the Web of Science Core Collection, Scopus, and China National Knowledge Infrastructure, and ultimately included 206 publications. The review focuses on array-induced environmental reconfiguration, equipment adaptation, light–thermal sensing and prediction, and coordinated agrivoltaic operation. Existing evidence indicates that the environmental effects of agrivoltaic systems are clearly influenced by climate and array configuration; in representative vertical or tracking systems, annual-scale photosynthetically active radiation decreased by approximately 11% to 34%. Mismatch among module height, row spacing, and implement width reduces field-operation efficiency, which fell to approximately 45% under severe mismatch in some experiments; elevated arrays also cause GNSS signal attenuation and increase the difficulty of continuous positioning beneath the panels. Because existing studies differ considerably in site conditions, evaluation indicators, and validation periods, the above results mainly reflect representative performance ranges. Matching criteria between photovoltaic arrays and agricultural machinery have not yet been established, long-term field-measured data in complex field environments are insufficient, model adaptability across regions is limited, and coordinated scheduling of agricultural production and photovoltaic operation and maintenance remains inadequate. Overall, coordinated design of arrays and agricultural machinery, multi-sensor fusion navigation, environmental prediction corrected for array structure, and hierarchical scheduling under safety constraints are the main development directions for intelligent coordinated operation of agrivoltaic systems.

1. Introduction

Population growth and climate change are jointly intensifying pressure on global food, energy, and land systems. By 2050, the world population is expected to approach 10 billion, and global food production will need to increase substantially to meet the corresponding demand [1,2]. At the same time, human-induced global mean warming over 2013–2022 has reached 1.14 °C, and the frequency and intensity of extreme heat, drought, and heavy precipitation events have further increased [3,4,5]. Therefore, simultaneously strengthening the resilience of food production and energy supply under conditions of limited land resources and rising climate risks has become a core challenge for sustainable agriculture [6].

Against this background, photovoltaic power generation, with advantages including low carbon emissions during operation, continuously declining construction costs, and flexible deployment, has become a major source of newly installed renewable-energy capacity worldwide [7]. Global cumulative photovoltaic capacity increased from approximately 100 GW in 2012 to more than 1 TW in 2022 [8]. However, the continued expansion of large-scale ground-mounted photovoltaic power plants may occupy farmland and intensify competition for land between energy development and food production [9]. Agrivoltaics combines photovoltaic power generation and agricultural production within the same land unit, providing a new technical pathway for alleviating this conflict and increasing total land output.

Early studies mainly focused on the effects of array density, module tilt, layout, and shading ratio on crop growth, attempting to find a reasonable balance between photovoltaic power generation and crop production [10]. These studies demonstrated that module geometry not only determines photovoltaic output, but also changes the total amount, uniformity, and temporal rhythm of radiation received by the crop canopy [11]; even with similar installed capacities, different configurations may produce different growth responses because of differences in shadow trajectories and crop-level radiation distribution [12].

In recent years, the core scientific issue of agrivoltaic systems has no longer been simply electricity generation above the panels and planting beneath them, but the integrated regulation of crop light environments, thermal and moisture environments, water use, and physiological and metabolic processes by photovoltaic modules [13]. On the one hand, new materials such as semitransparent modules, spectrum-selective modules, and bifacial photovoltaic modules have been used to improve the light environment beneath the panels and to alleviate competition for light between photovoltaic modules and crops; on the other hand, adjustable photovoltaic arrays, tracking supports, and dynamic shading-control technologies have gradually been applied to agrivoltaic systems, transforming photovoltaic modules from fixed shading structures into active environmental-regulation units [14]. Related studies show that module transmittance, changes in light quality, and light uniformity directly affect crop photosynthesis and quality formation; therefore, establishing a matching relationship between photovoltaic structural parameters and crop growth requirements is essential for improving the overall benefits of agrivoltaic systems. Colored semitransparent photovoltaic panels can coordinate crop production and photovoltaic power generation through selective utilization of different wavelength bands [15]; meanwhile, optimized tracking agrivoltaic strategies for field crops such as maize can reduce crop-yield losses and improve total land productivity [16]. Recent studies further indicate that the design of agrivoltaic systems must also comprehensively consider the relationships among spectral composition, crop type, and facility energy demand. Modules with different transparency and spectral characteristics can induce markedly different crop responses; some semitransparent modules can alleviate lettuce-yield reductions under simulated future-climate conditions, while integrating semitransparent photovoltaics with heat pumps extends array optimization from light allocation to coordinated greenhouse energy supply and carbon-emission reduction [17,18,19].

In addition, the presence of photovoltaic arrays also imposes new engineering requirements on agricultural-equipment operations. Compared with conventional farmland, agrivoltaic systems contain support columns, row-spacing constraints, low-clearance spaces, and GNSS signal interference, making it difficult for conventional agricultural machinery and robots to adapt directly to under-panel operating environments [20]. The core objective of agrivoltaic systems is to simultaneously realize agricultural production and photovoltaic power generation within the same land space, thereby improving land-use efficiency. However, achieving this objective depends not only on light-energy allocation between crops and photovoltaic modules, but also on whether the system structure can meet the normal operating requirements of agricultural machinery. Because mechanized agriculture involves multiple operations such as tillage, monitoring, and harvesting, different crop field layouts and equipment types impose differentiated requirements on support height, passage width, and positioning methods in APV systems [21]. The development of intelligent agricultural equipment should enable sensing results to genuinely enter prediction, decision-making, and control processes, and digital twins and edge computing provide new implementation pathways for this purpose [22,23,24]. For agrivoltaic systems, the next step should strengthen the linkage among environmental, crop, and equipment information so that task arrangement and operational control become more coordinated.

Existing domestic and international studies have systematically discussed crop responses, microclimatic effects, array configurations, and land-use benefits of agrivoltaic systems [9]. International research expanded relatively early from combined land use toward semitransparent modules, dynamic tracking arrays, and under-array equipment positioning [10,15,16,25], whereas domestic research has focused more on regional light–thermal environment simulation, array-parameter optimization, and UAV remote-sensing monitoring [26,27,28]. Overall, existing studies are mostly limited to single sites, specific arrays, or local production links, and systematic connections among structural design, environmental prediction, and equipment operation are still lacking.

Against this background, this review focuses on the chain effects of photovoltaic-array-induced environmental reconfiguration on agricultural production and intelligent-equipment operation, with particular attention to the coupling among array structure, crop requirements, environmental sensing and prediction, and equipment operations. Unlike existing reviews that mainly discuss crop yield, microclimate, and land-use benefits, this review further analyzes, from the agricultural-equipment execution level, how environmental information is converted into bases for passage, navigation, variable-rate operations, and task scheduling, so as to clarify the technical pathway by which agrivoltaic systems develop from spatial co-location toward coordinated operation.

2. Literature Review and Research Gaps

2.1. Existing Agrivoltaic Reviews

Existing agrivoltaic reviews mainly focus on land co-use, array configuration, crop response, microclimatic change, and overall system benefits. As agrivoltaic systems have gradually developed from experimental cultivation toward large-scale engineering applications, mechanization compatibility has begun to receive attention. Existing reviews have summarized agrivoltaic research from the perspectives of agronomic effects, array design, light–thermal environments, system modeling, and mechanization compatibility, but systematic linkage among these aspects toward field execution is still lacking. The marginal contribution of this review is not to again list the general benefits of agrivoltaics, but to use agricultural equipment as the execution carrier and place array structure, environmental state, crop requirements, and operation tasks within a unified analytical framework, with emphasis on the conversion mechanism from sensing and prediction results to operation instructions, the applicability boundaries of different technologies, and coordinated optimization of agricultural operations and photovoltaic operation and maintenance. The research boundaries of related reviews and their differences from this review are shown in

Table 1. Comparison of the research scope and contributions of this review and existing agrivoltaic reviews.

Representative Review	Research Perspective and Main Content	Representative Conclusions	Coverage Boundary Relevant to this Review
Weselek et al. [9]	Agrivoltaic applications, shading effects, crop responses, and land use	System benefits are jointly influenced by crop type, shading degree, and climatic conditions, and require site-specific configuration	Focuses mainly on agronomic effects and application potential, with limited coverage of machinery operations and coordinated control
Zainol Abidin et al. [29]	Photovoltaic structures, light environment, crop selection, and irrigation management	Photovoltaic structures and agronomic management should be coordinately designed around crop light and water requirements	Does not systematically discuss machinery passage, navigation and positioning, or operation decision-making
Toledo and Scognamiglio [30]	Array spatial design, performance evaluation, and landscape-ecological integration	Agrivoltaic-system design should simultaneously consider energy efficiency, agricultural use, and landscape-ecological impacts	Focuses on planning and design and does not extend to equipment operation or dynamic task coordination
Omer et al. [31]	Systematic evaluation of microclimate, water-use efficiency, and crop yield	Agrivoltaic systems can improve water–thermal environments and water-use efficiency, but excessive shading may reduce the yield of light-demanding crops	Mainly explains environmental and crop responses and does not discuss how sensing results are converted into operation instructions
Zainali et al. [32]	Modeling, simulation, and optimization of radiation, microclimate, crops, and photovoltaic systems	Multi-model coupling is a basis for coordinated optimization, but standardized data and cross-configuration validation remain insufficient	Focuses on planning and simulation optimization, with limited discussion of real-time task execution and multi-equipment coordination
Bellone et al. [20]	Compatibility between photovoltaic arrays and agricultural mechanization	Mismatch among array clearance, row spacing, headland space, and machinery scale reduces field-operation efficiency; both need coordinated design	Focuses on structural–machinery physical adaptation and has not formed a sensing–prediction–scheduling–execution closed loop
This review	Integrated analysis of array configuration, light–thermal environment, equipment adaptation, navigation sensing, prediction and decision-making, and task scheduling	Agrivoltaic systems need to shift from spatial co-existence toward coordinated operation jointly constrained by structure, environment, crop, and equipment states	Uses agricultural equipment as the execution carrier and constructs an agrivoltaic technical chain from sensing and decision-making to execution

.

2.2. Evolution Toward Intelligent Agrivoltaic Systems

From the perspective of research evolution, early agrivoltaic technologies mainly addressed the spatial compatibility of photovoltaic power generation and agricultural production within the same land unit, with research focusing on the effects of array density, module height, tilt angle, and shading ratio on crop yield and land-use efficiency. With the development of semitransparent modules, bifacial modules, and dynamic tracking technologies, research has gradually expanded from fixed-structure design to active regulation of light, thermal, and water resources, transforming photovoltaic modules from passive shading structures into adjustable environmental elements.

At the same time, large-scale operation of agrivoltaic systems has further expanded the research object from the “array–crop” relationship to the “array–environment–equipment” relationship. Photovoltaic supports, limited clearance, and regular passages alter the passage and turning conditions of agricultural machinery, while module obstruction and metal structures also affect positioning and sensing information from GNSS, vision, and LiDAR. Therefore, the role of environmental monitoring and light–thermal prediction is no longer limited to explaining crop responses, but must further serve specific operations such as irrigation, crop protection, inspection, navigation, and module adjustment. The extension of existing research from array design and environmental evaluation toward multi-source sensing, prediction and decision-making, and coordinated equipment operation indicates that agrivoltaic systems are gradually evolving from simple combined land-use modes toward intelligent production systems in which structure, environment, crops, and equipment jointly participate.

2.3. Research Gaps and Scope of This Review

For large-scale and long-term operation, three interrelated key gaps remain in existing research. First, in equipment adaptation, although existing studies have separately addressed array spacing, minimum clearance, machinery dimensions, and navigation and positioning methods, systematic quantitative design standards for machinery passage, navigation and positioning, and operational safety in photovoltaic-farm environments remain incomplete; in particular, safety constraints and applicability boundaries under module obstruction, GNSS multipath effects, and alternating bright and dark conditions still lack sufficient validation. Second, in environmental prediction, heterogeneous array shading causes pronounced spatiotemporal heterogeneity in light, thermal, and moisture environments. Existing models are mostly developed for a single site, a specific array, or a specific crop, and high-accuracy prediction models with cross-regional and cross-configuration applicability remain insufficient. Third, in operational management, agricultural operations, module regulation, and photovoltaic operation and maintenance are usually studied separately, and agricultural production and photovoltaic operation still lack a multi-objective coordinated-scheduling mechanism and engineering implementation paradigm that can jointly consider crop requirements, power-generation benefits, operating costs, and operational safety. Around these three research gaps, Section 5 focuses on the quantitative basis for machinery passage, navigation and positioning, and operational safety; Section 6 analyzes environmental sensing, prediction, and cross-regional applicability under heterogeneous shading; and Section 7 further discusses multi-objective coordinated scheduling and engineering implementation mechanisms between agricultural production and photovoltaic operation.

3. Materials and Methods

This study adopted a literature review method combining systematic retrieval and thematic analysis. The literature was primarily collected from the Web of Science Core Collection, Scopus, and China National Knowledge Infrastructure (CNKI). The search period covered publications from 2010 to 2026, with a particular focus on studies related to environmental sensing, light–thermal prediction, and intelligent equipment regulation in agrivoltaic systems.

3.1. Literature Search Strategy

Keyword groups were combined using the Boolean operators “AND” and “OR” to ensure the systematic coverage and comprehensiveness of the retrieved literature. Synonymous terms within each keyword group were connected using “OR”, whereas different thematic groups were connected using “AND”. English keywords were used for searches in the Web of Science Core Collection and Scopus, while the corresponding Chinese keywords were used in CNKI.

The first group (research objects) included agrivoltaics, photovoltaic agriculture, photovoltaic greenhouses, agriPV, transparent PV, semi-transparent PV and other related systems.

The second group (environmental sensing and light–thermal prediction) included satellite remote sensing, unmanned aerial vehicles, multispectral imaging, LiDAR, thermal infrared imaging, sensor monitoring, and light–thermal prediction.

The third group (agricultural equipment and control technologies) included agricultural machinery, agricultural robots, navigation and positioning, variable-rate operations, adaptive control, path planning, and intelligent regulation.

3.2. Inclusion Criteria

Research objects: Studies focusing on agrivoltaic systems, photovoltaic greenhouses, photovoltaic–livestock systems, and other related agricultural systems.

Research content: Studies involving the reconstruction of the environment beneath photovoltaic arrays, environmental information acquisition, light–thermal resource prediction, spatial adaptation of agricultural equipment, navigation and positioning control, and related intelligent regulation technologies.

Publication type: Peer-reviewed journal articles, high-quality international conference papers, or doctoral dissertations.

3.3. Exclusion Criteria

Non-academic publications: Patent specifications, news reports, product manuals, technical white papers, and online resources that had not undergone peer review.

Studies unrelated to the review topic: Studies mainly focusing on photovoltaic power generation, economics, general crop physiology, or energy statistics without addressing the core technologies of environmental sensing, light–thermal prediction, or equipment regulation.

Duplicate publications: When both conference and journal versions of the same study were available, only the version containing the most complete content and detailed information was retained.

Studies with insufficient information: Studies lacking full-text availability, providing unclear descriptions of key technologies, or failing to support qualitative or quantitative analysis.

3.4. Literature Screening Process

Title-and-abstract screening and full-text eligibility assessment were independently completed by two researchers. The two researchers separately recorded inclusion, exclusion, and uncertain decisions according to predetermined inclusion and exclusion criteria, and did not exchange screening opinions before completing the judgment at each stage. For studies on which their judgments differed, consensus was first reached through discussion; disagreements that still could not be resolved were adjudicated by a third researcher. Reasons for full-text exclusion were also recorded during screening to ensure traceability of the literature-selection process.

The literature search was conducted using the Web of Science Core Collection, Scopus, and CNKI, and all retrieved records were exported to EndNote 21 for management. After the results from the three databases were merged, duplicate records were identified and removed using EndNote-assisted automatic deduplication combined with manual verification. The remaining records were then subjected to further screening.

A total of 1856 publications were initially retrieved, including 1422 records from the Web of Science Core Collection, 283 from CNKI, and 151 from Scopus. After 227 duplicate records were removed, 1629 publications remained for title-and-abstract screening, during which 1041 studies unrelated to the research topic were excluded, leaving 588 publications for full-text assessment. Further screening was then conducted according to the predefined inclusion and exclusion criteria: 72 publications were excluded because of insufficient key technical information, 31 because the full text was unavailable, 45 because the research findings were duplicated, and 234 because they did not focus on the core topics of environmental sensing, light–thermal prediction, and equipment regulation in agrivoltaic systems. A total of 382 publications were excluded during the full-text assessment, and 206 publications were ultimately included in the subsequent qualitative review analysis. Figure 1 shows the literature-screening process.

3.5. Data Extraction and Analysis Framework

After eligibility screening of the included studies was completed, a structured evidence matrix was used to code the 206 included publications. The extracted information was divided into five categories: (1) literature and research-background information: publication year, country in which the study was conducted, climatic conditions, crop or livestock type, and experimental scale; (2) photovoltaic-system characteristics: system configuration, module technology and transmittance, array height, row spacing, tilt angle, and tracking mode; (3) technical methods and data: sensing platform, sensor type, prediction or control algorithm, model inputs and outputs, and validation scheme; (4) reported results: radiation distribution, air and soil temperature, soil moisture, crop yield and quality, photovoltaic power generation, land equivalent ratio, prediction or positioning accuracy, and operating efficiency; and (5) implementation conditions and limitations: machinery passability, system robustness, cost, technical complexity, technology maturity, and scalability. Information not reported in the original studies was uniformly marked as “not reported” and was not subjectively inferred.

Cross-study comparisons were conducted across six dimensions: agronomy, energy, environment, technology, operation, and implementation. Quantitative results were compared only when indicator definitions, reference conditions, and units of measurement were highly consistent; otherwise, narrative thematic synthesis was used. Conflicting research conclusions were interpreted in combination with photovoltaic-array geometric parameters, shading intensity, measurement location and duration, crop variety and growth stage, climatic conditions, and experimental-validation scale, so as to distinguish generalizable conclusions from findings applicable only to specific sites or specific systems.

As shown in Figure 2, existing studies can be broadly grouped into four interrelated thematic clusters: (1) crop–environment coupling studies centered on agrivoltaic systems, crop yield, microclimate, and water management; (2) environmental sensing and prediction studies centered on greenhouse systems, evapotranspiration, and prediction models; (3) equipment-adaptation studies involving agricultural machinery, passage navigation, and economic evaluation; and (4) site operation-and-maintenance studies centered on photovoltaic modules and operation and maintenance. Overall, crop response and microenvironmental effects remain the research basis of this field, whereas environmental prediction, machinery passage and navigation, and photovoltaic operation and maintenance show a stronger engineering-application orientation.

Figure 3 shows the annual scientific output of agrivoltaic research. The overall number of studies was small during the early stage (2010–2017), when research mainly focused on the basic concept of combining photovoltaic power generation with agricultural production. Early research concentrated on land-use efficiency, photovoltaic layout, and photovoltaic–agriculture compatibility. Dupraz et al. proposed the concept of agrivoltaics and pointed out that combining photovoltaic power generation with agricultural activities could improve total land productivity [10]. However, studies during this period were mainly based on model analysis and small-scale experiments, and validation of applicability across different climate zones, crop types, and long-term production conditions remained limited. From 2018 to 2021, agrivoltaic research began to shift from proof of concept toward environmental regulation and system optimization. During this stage, research attention gradually moved toward spatial heterogeneity of the light environment and interactions between photovoltaic structures and the agricultural microclimate [33]. At the same time, modeling studies further demonstrated that module height, spacing, and orientation significantly affect under-panel radiation distribution and crop-production capacity [27]. However, substantial differences among studies in array structure, climatic conditions, and crop-growth stages made direct comparison of reported light–thermal changes and crop responses difficult, and the field still lacked unified experimental designs and evaluation indicators. Since 2021, agrivoltaic research has entered a period of rapid development, with a marked increase in annual publication output. Semitransparent modules, dynamic tracking systems, and multi-source monitoring technologies have gradually been used to improve coordination between energy production and agricultural production [28,34]. Since 2021, agrivoltaic research has shown a clear trend toward engineering application and intelligent operation. Compared with the previous year, publication output in 2023 and 2025 increased by approximately 40.1% and 55.7%, respectively. This growth reflects increasing attention to environmental remote sensing, machine-learning prediction, intelligent agricultural equipment, and related directions.

4. Reconfiguration of Agricultural Production and Operating Environments by Photovoltaic Arrays

4.1. Types and Application Scenarios of Agrivoltaic Systems

According to agricultural-production scenarios, agrivoltaic systems can be divided into open-field, protected-horticulture, livestock-production, and aquaculture types. Among them, open-field systems mainly include elevated and inter-row configurations; protected-horticulture systems include rooftop, semitransparent, and spaced configurations; livestock-production systems include canopy-frame and livestock-building rooftop configurations; and aquaculture systems include fixed-support and floating configurations, as shown in Figure 4.

The main trade-off in open-field systems is between support cost and machinery passage; the economic return of protected-horticulture systems depends on whether light-transmission losses can be compensated by crop value, cooling effects, and facility-energy benefits; agricultural benefits in livestock-production systems are derived more from animal thermal comfort and forage water conservation, requiring additional animal-safety design; aquaculture systems can achieve large-scale combined use of water surfaces but face uncertainty in aquatic-environment responses, equipment corrosion protection, and aquaculture-operation access.

4.1.1. Open-Field Systems

Open-field agrivoltaic systems are based on maintaining continuity of field agricultural production and are suitable for grain crops, forage, open-field vegetables, and some cash crops. According to array configuration, they can be further divided into elevated fixed, inter-row fixed, vertical bifacial, and tracking systems. Elevated fixed systems can retain under-panel space for crops and agricultural machinery but have relatively high support-structure costs; inter-row fixed systems have relatively simple structures but can divide farmland into narrow operating units; vertical bifacial systems have a small ground projection, which is favorable for rainfall entry and machinery passage, but impose higher requirements on row-spacing and orientation design; tracking arrays can adjust module orientation according to solar position, crop light demand, and weather conditions, but have greater control and operation-and-maintenance complexity. The core issue for this type of system is coordinating shading distribution, crop requirements, and machinery operating space.

4.1.2. Protected-Horticulture Systems

Protected-horticulture systems mainly include photovoltaic greenhouses and other protected-production systems, and can use rooftop opaque modules, spaced modules, semitransparent or spectrum-selective modules, and adjustable rooftop arrays. Their advantage is that photovoltaic power generation can directly serve ventilation, irrigation, supplemental lighting, and temperature-regulation equipment, forming a local energy cycle. The main limitation is that rooftop modules can easily cause insufficient canopy radiation and uneven spatial distribution. Therefore, the design focus of this system type is coordination among module coverage ratio, light-transmission characteristics, and facility environmental-control energy consumption.

4.1.3. Livestock-Production Systems

Livestock-production systems include elevated pasture systems, shading-canopy systems, and photovoltaic systems installed on the roofs of livestock buildings. Shading produced by photovoltaic modules can improve livestock thermal comfort and may influence forage moisture conditions and regrowth processes by reducing evapotranspiration. However, engineering design must also account for fencing, drinking water, feeding, manure treatment, and animal-activity safety. Dust, ammonia, and corrosive gases in livestock environments may also accelerate module contamination and aging of metal structures; therefore, module cleaning, ventilation, and corrosion protection should be incorporated into system design.

4.1.4. Aquaculture Systems

Aquaculture systems mainly use fixed supports over water surfaces, pond-spanning structures, or floating photovoltaic structures. Array shading can reduce peak water temperature and surface evaporation under strong-radiation conditions, but excessive coverage may also suppress phytoplankton photosynthesis and alter dissolved oxygen and primary productivity in the water body. Compared with terrestrial scenarios, this system type is also constrained by wind and waves, floating-platform stability, feeding, harvesting, pond patrol, and water-quality monitoring; therefore, trade-offs are required among power-generation area, aquatic ecological requirements, and aquaculture-operation passages.

As shown in

Table 2. Typical structures and application characteristics of agrivoltaic systems under different agricultural production scenarios.

Agricultural Production Scenario	Representative Array Structure	Typical Application Regions and Representative Scale	Main Advantages	Key Constraints	Equipment-Operation Constraints
Open-field	Elevated fixed, inter-row, vertical bifacial, or tracking arrays	Mainly found in temperate agricultural regions of Europe and North America. The Heggelbach project in Germany has a capacity of 194 kWp, and the Borgo Virgilio project in Italy is approximately 2.15 MWp [39,40].	Improves overall land-use efficiency; alleviates heat and drought stress; retains space for mechanized operations	High support-structure cost; uneven light distribution; restricted passage, turning, and safety distances	Row spacing and clearance restrict passage of large machinery, and headland turning space is insufficient; module obstruction and metal supports interfere with GNSS positioning.
Protected horticulture	Rooftop opaque, spaced, semitransparent, or adjustable photovoltaic modules	Mainly found in Mediterranean protected-horticulture regions and high-radiation greenhouse production areas. The photovoltaic greenhouse in Italy has a capacity of 68 kWp, and the photovoltaic greenhouse in Türkiye has a capacity of 48.6 kWp [41,42].	Supplies facility energy on site; alleviates summer overheating; supports coordinated environmental regulation and power generation	Low light and light patchiness; trade-off between coverage ratio and crop yield; possible insufficient light and heat in winter	Narrow passages and low clearance make equipment turning difficult; indoor GNSS is constrained, and reflections and alternating bright–dark conditions affect visual positioning.
Livestock production	Elevated pasture, shade-canopy, or livestock-building rooftop photovoltaic systems	Mainly found in the United States, Australia, and temperate pasture regions of Europe. The photovoltaic grazing project in Oregon, United States, is approximately 1.4 MW [43].	Improves animal shade and thermal comfort; maintains forage moisture; reduces manual vegetation maintenance	Requirements for fencing and drinking-water facilities; risks of animal gnawing and collisions; manure, dust, and corrosion problems	Fences, drinking-water facilities, and animal activity obstruct passage and turning; canopy obstruction and dynamic obstacles reduce positioning and obstacle-avoidance reliability.
Aquaculture	Fixed supports above fishponds, pond-spanning systems, or floating photovoltaics	Mainly concentrated in pond-aquaculture regions along the eastern coast of China and coastal areas of Taiwan. The Jiangsu fishery–photovoltaic project is 50 MW, and the Cixi project in Zhejiang is 200 MW [44].	Combined use of aquaculture water surfaces; reduces water-surface evaporation; supplies power to aeration, pumping, and monitoring equipment	May alter water temperature, dissolved oxygen, and plankton; restrict feeding, harvesting, and pond patrol; floating-structure and electrical-safety risks	Pond banks, supports, and cables restrict equipment passage, and turning space on the water surface is limited; reflections and module obstruction affect vision and GNSS positioning.

, different agrivoltaic systems differ markedly in application region and engineering scale. Open-field systems are mainly distributed in temperate agricultural regions of Europe and North America, with representative project capacities ranging from the hundred-kilowatt level to the megawatt level; their design focus is to ensure array clearance, machinery passages, and headland turning space while improving overall land-use efficiency. Protected-horticulture systems are mostly applied in Mediterranean and high-radiation greenhouse-production regions. Because of limitations imposed by greenhouse area and rooftop coverage, representative systems are usually at the tens-of-kilowatts scale; therefore, greater coordination is needed among module transmittance, canopy photosynthetically active radiation, and energy consumption for facility environmental regulation. Livestock-production systems are mainly found in the United States, Australia, and temperate pasture regions of Europe. Constrained by local climatic conditions, power-generation capacity per unit area is generally relatively low, although projects can reach the megawatt scale. Photovoltaic shading helps improve animal thermal comfort and forage moisture conditions, but dust, ammonia, manure, and animal collisions increase requirements for module cleaning, structural corrosion protection, and electrical protection. Aquaculture systems are concentrated in pond-farming regions along the eastern coast of China and in Taiwan; representative projects can reach tens to hundreds of megawatts, clearly larger than those of other scenarios, which is related to continuous water-surface space and conditions favorable for centralized engineering development. However, larger installed capacity also increases the risk of impacts on water temperature, dissolved oxygen, plankton, and feeding, harvesting, and inspection passages. Overall, installed capacity reflects the engineering application scale of different scenarios but cannot independently represent agricultural suitability. Each system type still requires comprehensive evaluation of energy output, agricultural returns, and system safety in combination with effective agricultural area, array structure, environmental-regulation requirements, equipment-operating conditions, and operation-and-maintenance risks.

4.2. Reconfiguration of Light, Thermal, and Water Environments by Photovoltaic Arrays

4.2.1. Reconfiguration of the Light Environment

Investigating the mechanisms by which photovoltaic arrays affect the internal environment of agricultural systems is the basis for evaluating crop adaptability and carrying out engineering optimization. After photovoltaic modules are introduced into agricultural space, the first change occurs in the radiation distribution within the facility. For example, Cossu et al. compared solar-radiation distributions in typical European photovoltaic greenhouses and pointed out that pronounced light gradients may still form among different greenhouse regions in winter [33]. Kujawa et al. further compared tomato production under different shading levels and found that increased shading alters the greenhouse microclimate and increases the risk of yield reduction, although the effects of moderate shading still depend on regional climate and greenhouse structure [45].

For open-field photovoltaic farmland, modules intercept part of the direct radiation and change the proportion of diffuse radiation, creating light gradients of different intensities beneath panels, between panels, and at array edges. The crop response is manifested not only as a reduction in total radiation, but also as changes in the temporal pattern and spatial distribution of received light. Fixed-tilt and tracking bifacial arrays produce different spatiotemporal shading trajectories; partial shading beneath photovoltaic panels significantly changes the radiation conditions received by the crop canopy, and the magnitude and spatial uniformity of photosynthetically active radiation reduction at crop level differ substantially among array configurations [46,47]. As shown in Figure 5, the daily mean shading rate beneath OAVS modules and between adjacent modules exhibits a clear spatial gradient in both summer and winter.

Figure 5 shows that the OAVS array does not produce uniform shading, but rather a radiation pattern with clear spatial zoning and seasonal variation. In summer, shading within the array ranges from approximately 3% to 100%, with the southern region generally showing higher shading and the central region showing the largest variation range; in winter, the overall shading rate increases to 63.9–100%, the northern region remains above 93.0%, and the southern region is relatively lower [26]. This seasonal difference is mainly related to changes in solar elevation and azimuth: lower solar elevation in winter increases shadow projection distance and expands highly shaded areas, while the change in shading intensity among different orientations reflects the combined effect of array orientation and the solar trajectory.

In facility-based agrivoltaics, photovoltaic modules affect the light environment more directly. Photovoltaic greenhouses usually place modules on the roof or on part of the roof surface, which can readily cause uneven spatial light distribution inside the greenhouse. Rooftop photovoltaic modules alter the spatial distribution of solar radiation inside the greenhouse, but under certain coverage ratios and checkerboard layouts, their effects on crop yield and some microclimatic parameters are not significant [48]. Semitransparent modules and light-diffusing materials can improve crop light conditions. Integrating semitransparent photovoltaic modules into greenhouse systems can reduce evapotranspiration water consumption while meeting a certain power-generation demand, without causing significant adverse effects on lettuce growth [49]; meanwhile, introducing light-diffusing films beneath rooftop photovoltaic modules allows diffuse light to weaken local shadow fluctuations caused by the modules [50]. Further optimization-design studies of agrivoltaic systems indicate that array height, spacing, and layout can be quantitatively screened through light-environment simulation [51]. Therefore, the optimization focus of facility-based systems is not only average transmittance, but the PAR actually received by the crop canopy and its uniformity.

Overall, fixed arrays mainly create relatively stable spatial light gradients, tracking arrays exhibit dynamic changes in shadow position and duration, and greenhouse light environments are additionally affected by module transmittance and roof-layout patterns.

4.2.2. Reconfiguration of the Thermal and Moisture Environment

In addition to the light environment, photovoltaic arrays also reconfigure under-panel thermal and moisture environments and soil–water processes through radiation obstruction, changes in ventilation conditions, and rainfall redistribution. During the day, modules block part of the shortwave radiation and can reduce canopy, surface, and soil temperatures; at night, they may suppress longwave-radiation loss from the surface, producing a certain heat-retention effect beneath the panels. Comparative facility studies found that the mean daytime air temperature in agrivoltaic systems can be lower than that in conventional agricultural facilities [52]; CFD simulations further show that array configuration and the resulting local airflow jointly affect module and ground-surface temperature distributions [53].

The effects of photovoltaic arrays on water processes are mainly reflected in evapotranspiration demand, soil evaporation, rainfall redistribution, and water-use efficiency. Moderate shading can reduce crop evapotranspiration demand and soil evaporation and may alleviate water stress under hot and dry conditions by increasing soil water content [54,55,56,57]. In semitransparent photovoltaic greenhouses in arid regions, transmitted total solar radiation and PAR decreased by approximately 40% and 37%, respectively, while canopy-temperature fluctuations during periods of strong radiation were also reduced [34]. Under fixed and dynamic arrays, the interaction between shading and deficit irrigation can also alter crop growth and water responses [58].

Because studies differ in regional climate, array structure, measurement location, and crop type, the reported PAR reduction rates and cooling magnitudes should not be compared without considering experimental conditions.

Table 3. Comparison of light–thermal–water environmental responses and applicability boundaries under different agrivoltaic array configurations.

Array Configuration and Climatic Conditions	Changes in the Light Environment	Changes in Temperature and Water Conditions	Main Influencing Factors and Applicability Boundaries	Complexity and Economics
Elevated fixed; temperate farmland in southwestern Germany	Daily mean PAR decreased by approximately 30% [59]	Daily mean air and soil temperatures decreased by approximately 1.1 °C and 1.2–1.4 °C, respectively, and relative humidity increased by 2% during some periods [59].	Array geometry shapes light–thermal gradients, and rainfall redistribution determines water conditions beneath the panels.	High structural investment; LCOE is approximately 0.0829 EUR·kWh−1, higher than 0.0603 EUR·kWh−1 for conventional ground-mounted photovoltaics [39]
Vertical bifacial; semi-arid field in Pakistan	Annual-scale PAR in similar vertical or tracking configurations decreased by approximately 11% to 34%, with a light-uniformity index of approximately 91% to 95% [60,61]	Air temperature in the vertical bifacial system was approximately 0.5 °C lower than the control, and soil moisture increased by approximately 26% [62].	Orientation and row spacing determine radiation uniformity, and semi-arid conditions strengthen the water-conservation effect.	Medium complexity; economics are sensitive to row spacing, foundations, and land loss
Semitransparent crystalline-silicon greenhouse; naturally ventilated conditions in Kunming, China	Solar radiation decreased by approximately 35% to 40% [49].	On sunny days, indoor temperature decreased by 1–3 °C, and evapotranspiration decreased from 1.72 to 0.80 mm·d−1 [49].	Transmittance determines heat load and evapotranspiration; water saving mainly results from radiation reduction rather than increased air humidity.	Complex system integration; economics depend on high-value crops, facility energy savings, and photovoltaic self-consumption
Single-axis tracking; tomato field under the Mediterranean climate of Israel	Seasonal cumulative photosynthetic irradiance in the middle of the array decreased by only 1.96%, that in transition rows decreased by approximately 10.2% to 24.9%, and that directly beneath modules decreased by 91.3–92.9% [63]	Daytime air temperature directly beneath the modules was 2.5–3.0 °C lower than in the middle of the array, while nighttime temperature was approximately 0.5 °C higher [63].	Tracking trajectory and lateral distance jointly shape light–thermal gradients.	High control, maintenance, and failure costs

selects representative configurations and summarizes their light, temperature, and water responses and applicability boundaries. Table 3 includes representative light, temperature, and water responses under elevated fixed, vertical bifacial, semitransparent greenhouse, and single-axis tracking arrays.

Overall, the PAR reduction rates and cooling magnitudes reported for different agrivoltaic systems are jointly affected by climatic background, array structure, observation scale, and crop type. High-radiation, arid, or semi-arid regions have higher background radiation and evapotranspiration demand, and moderate shading generally shows more pronounced cooling and water-conservation effects; in temperate or low-radiation regions, the same shading ratio may instead intensify crop light limitation. Array height, row spacing, module coverage ratio, transmittance, and tracking trajectory further determine shadow intensity, duration, and spatial uniformity. At the same time, PAR reduction directly beneath modules, in transition zones between panels, and in the middle of arrays may differ severalfold, and measurements above the canopy, within the canopy, and near the ground are not directly comparable. Crop height, canopy structure, shade tolerance, and growth stage determine whether environmental changes are translated into yield or quality effects.

Although the magnitude of change differs substantially among studies, several basic points of consensus have emerged: photovoltaic arrays generally reduce effective radiation beneath the panels, but this effect is clearly spatially heterogeneous and temporally dynamic; under high-temperature, strong-radiation, or water-limited conditions, moderate shading is generally beneficial for reducing evapotranspiration demand, maintaining soil moisture, and alleviating crop heat and drought stress. Therefore, the environmental effect of arrays is more appropriately understood as a redistribution of light, thermal, and water resources in time and space.

Current controversies mainly concern three aspects: first, no consistent conclusion has been reached on the direction and magnitude of the effect of arrays on air temperature; second, whether short-term cooling and water-conservation effects can be stably translated into whole-season yield and quality benefits remains uncertain; and third, sufficient evidence is still lacking as to whether transferable optimum shading thresholds exist across climate zones, array configurations, and crop types. An important reason for these controversies is that existing studies have not yet adopted unified standards for array parameters, measurement height, control settings, statistical time scales, and crop-growth stages. This also indicates that future studies need cross-regional and cross-configuration validation under unified observation criteria.

4.3. Matching Array Types with Crop Types

4.3.1. Crop Requirements and Spatial Adaptation

Existing studies show that array structural parameters are the basis for determining the actual light conditions received by crops. Simulation of maize production and photovoltaic power generation in elevated tracking agrivoltaic systems shows that reducing module density, adjusting module height, and optimizing array layout can alleviate insufficient crop light and achieve a balance between crop yield and energy output [64]. Shading modeling for typical agrivoltaic configurations such as fixed vertical, single-axis tracking, and dual-axis tracking shows clear differences in the reduction of crop-level photosynthetically active radiation and in light uniformity among different array designs, indicating that array-configuration optimization is a key means of regulating the under-panel light environment and reducing uncertainty in crop yield [60]. In addition, dynamic agrivoltaic systems use rotatable photovoltaic panels to redistribute daytime solar radiation between crops and modules, so that photovoltaic arrays are no longer merely fixed shading structures but become controllable light–thermal resource-allocation units, thereby increasing total productivity per unit land area [61]. Therefore, adjustment of module density, height, row spacing, and movement trajectory provides an engineering basis for crop spatial arrangement and growth-stage adaptation.

However, the ability of arrays to regulate the environment does not mean that shading, cooling, and water-conservation effects necessarily translate into crop benefits. Their actual value depends on whether environmental changes match the physiological requirements of crops at specific growth stages. The same shading level may alleviate heat and drought stress during vegetative growth but cause radiation deficits during flowering, grain filling, or maturity; the same array may also shift from a protective structure to a limiting factor as season, weather, and crop-development processes change.

Table 4. Agrivoltaic array adaptation strategies driven by crop growth requirements.

Crop-Adaptation Scenario	Core Physiological and Production Requirements	Main Risks	Array-Adaptation Strategy
Seedling and vegetative growth stages	Maintain appropriate leaf-area expansion and stand establishment, and avoid canopy overheating and water loss	Persistent low light causes seedling etiolation or insufficient biomass accumulation; strong light and high temperature aggravate water stress	Use zoned planting according to seedling light requirements; retain moderate shading during high-temperature periods and reduce prolonged obstruction during low-radiation seasons
Flowering and fruit-set stages	Ensure effective radiation and suitable temperature required for floral-organ development, pollination, and fruit set	Excessive shading restricts flower-bud differentiation and causes flower or fruit drop or reduced fruit-set rate	Place crops sensitive during reproductive stages in low-shading zones; reduce module obstruction during critical periods and avoid continuous low light
Grain-filling stage	Maintain photosynthate formation and transport to grains or fruits	Insufficient cumulative radiation restricts dry-matter formation and leads to smaller fruits or lower yield	Arrange critical growth stages to avoid high-shading seasons; increase effective light-receiving time through spatial zoning or orientation adjustment
Maturity and quality-formation stage	Ensure formation of sugars, color, dry matter, and market quality	Persistent low light at the late stage causes delayed maturity, reduced sugar content, or insufficient coloration	Reduce shading intensity during maturity; preferentially place quality-sensitive crops in areas with more uniform light
High-temperature or drought-stress period	Reduce canopy heat load and transpiration consumption and maintain plant water status	Strong radiation and high vapor-pressure deficit cause wilting, stomatal closure, and photosynthetic inhibition	Increase protective shading during periods of high temperature and strong radiation and coordinate it with irrigation timing

summarizes array-adaptation strategies and evaluation indicators for different growth stages and stress scenarios according to crop requirements.

As shown in Table 4, the benefits of moderate shading are concentrated mainly during periods of high temperature, strong radiation, or water stress, whereas its costs often become evident during reproductive growth and quality formation. Therefore, evaluating array suitability using whole-season mean radiation or final yield may conceal short-term light limitations at key stages and their irreversible effects. Fixed arrays can only reduce this stage-specific mismatch through crop selection, spatial zoning, and sowing-date adjustment, but have difficulty adapting simultaneously to seasonal radiation changes and extreme weather. Adjustable arrays have the potential for dynamic compensation according to phenological stage, increasing protective shading during hot and dry periods and reducing obstruction during light-sensitive stages such as flowering, grain filling, and maturity. However, these dynamic-regulation schemes require continuous monitoring of quality-formation processes, phenological dynamics, and stage-specific physiological responses, resulting in relatively high monitoring costs and operation-and-maintenance workloads.

4.3.2. Crop-Oriented Regulation Logic for Dynamic Arrays

Based on the above crop-adaptation principles, dynamic agrivoltaic systems need to translate stage-specific crop requirements into executable module-adjustment commands. Their regulation objectives can be divided into three levels: first, maintaining the photosynthetically active radiation, canopy temperature, and water status required during key crop-growth stages so as to reduce yield and quality losses; second, reducing photovoltaic power-generation losses while satisfying agronomic requirements; and third, ensuring system safety during extreme weather, agricultural-machinery passage, and module operation and maintenance. Because crop light acquisition, photovoltaic power generation, and operational safety are not always consistent, dynamic regulation is essentially the allocation of radiation resources under multi-objective constraints rather than simply pursuing maximum shading or maximum power output.

Specifically, when crops enter light-sensitive stages such as flowering, grain filling, or maturity and canopy radiation is below stage-specific requirements, shading should be reduced; when canopy temperature, vapor-pressure deficit, or water stress exceeds the set range, protective shading can be increased; under continuous rainy or overcast conditions, overlap between naturally low light and module shading should be avoided; when strong winds, heavy rainfall, agricultural-machinery passage, or module-maintenance tasks occur, priority should be given to switching to a safe or avoidance position. In this way, module movement changes from being driven solely by solar position to being jointly driven by crop status, weather changes, and operational tasks, as shown in

Table 5. Regulation strategies and applicability boundaries for dynamic agrivoltaic arrays.

Regulation Strategy	Main Trigger Basis	Regulation Objective and Adjustment Method	Main Advantages	Main Limitations
Fixed-schedule regulation	Preset operating periods, calendar-day progression, and key phenological nodes	Change module tilt or shading duration at preset times	Simple control, stable operation, and low implementation cost	Difficult to respond to abrupt weather changes and actual crop conditions
Solar-position tracking	Solar elevation angle and azimuth	Improve photovoltaic radiation capture through single-axis or dual-axis tracking	Mature technology with a clear power-generation objective	The maximum-power orientation may conflict with the crop light-demand objective
Weather-threshold regulation	Air temperature, wind speed, and precipitation status	Increase shading during high-temperature and strong-radiation periods and enter a safe orientation under adverse weather	Can respond to high-temperature and extreme-weather events	Thresholds depend on region and crop and are affected by forecast errors
Crop-state feedback	Canopy PAR, temperature, soil moisture, leaf water potential, and phenology	Dynamically adjust module orientation according to actual light–thermal–water stress	Can directly respond to crop needs and has strong agronomic targeting	Depends on sensor stability and has high monitoring costs

.

Table 5 shows that existing studies usually use radiation, temperature, or soil moisture directly as control-trigger variables. Because these variables are affected by sensor location, varietal differences, and phenological stage, fixed thresholds are difficult to transfer across regions. Even when crop requirements can be accurately identified, module adjustment is not a local response to a single variable but produces coupling effects across multiple system links. Changes in module orientation not only alter canopy light acquisition, but also simultaneously affect power output, rainfall redistribution, ventilation conditions, and machinery passage, so local optimization under a single objective may create new system mismatches. Accordingly, agrivoltaic systems can adopt a two-layer control framework: the first layer adjusts shading position and duration according to crop phenology and stage-specific light–thermal–water requirements; the second layer comprehensively considers photovoltaic power generation, agricultural operations, and module operation-and-maintenance tasks to determine system operating priorities.

5. Adaptation and System Evaluation of Agricultural Equipment Under Photovoltaic-Array Constraints

5.1. Matching Array Space with Agricultural-Machinery Dimensions

Within agricultural-production space, photovoltaic arrays mainly alter the operating boundaries of equipment. Agrivoltaic systems differ from conventional photovoltaic power stations in that their spatial design must also ensure that agricultural production can continue [30]. From the perspective of operating space, mechanization constraints beneath photovoltaic arrays mainly include horizontal-space constraints and vertical-space constraints [65]. Therefore, the operating boundaries of agrivoltaic equipment are determined not only by field shape and crop-row direction, but are also jointly constrained by module layout, support structure, and the safe operating range [66]. Particularly in inter-row, vertical, or low-profile photovoltaic arrays, machinery operating space is mainly determined by module row spacing, and machinery cannot simply operate continuously according to the original field width [29]. If column spacing, crop-planting spacing, and implement working width cannot be matched, the operable space is reduced, land-use losses increase, and machinery-passage efficiency declines.

This working-width matching problem differs among sowing, fertilization, spraying, and harvesting operations [67]. Seeders and fertilizer applicators can adapt to narrower passages by shutting off some metering or fertilizer units, and therefore have a certain degree of adjustability. In restricted facility spaces, dedicated rail platforms, air-assisted atomization, and autonomous control can improve crop-protection coverage and passage stability [68]. For sowing, fertilization, and crop-protection operations, section control can reduce repeated coverage near headlands and obstacles. In existing open-field studies, repeated sowing area caused by field shape accounts for approximately 0.1% to 15.5%, while boom section control based on positioning information can reduce repeated pesticide application by approximately 15.2% to 17.5% [69,70]. Harvesting machinery is more difficult to adapt because header width and turning radius are usually large; insufficient array row spacing or restricted headland space can limit continuous harvesting operations and increase turning losses [71]. Agrivoltaic systems should determine row spacing by comprehensively considering the relevant operating equipment required throughout the entire crop-rotation cycle.

In addition, headland turning is an important source of efficiency loss in mechanized operations beneath photovoltaic arrays. In conventional farmland, equipment can turn at the headland; in agrivoltaic systems, however, array orientation is often selected primarily for power-generation efficiency and may not be consistent with natural field boundaries, crop-row direction, or the optimal machinery-operating direction. When headland width is insufficient, array-entrance spacing is small, or supports obstruct turning paths, equipment must reduce speed, increase reversing maneuvers, or detour into the next operating passage, thereby increasing non-productive time. Design studies on agrivoltaic vineyards show that array layout must simultaneously consider crop-row orientation, equipment passage, and headland operating space [72]; headland optimization should therefore not be limited to improving a single turning trajectory, but should determine operating direction, entrance position, and passage-transition sequence simultaneously during array planning. After field-path optimization, the number of turns that can be reduced accounts for 14.1% of the total number of turns, and headland area is reduced by 7.6% [73]. For operator-driven agricultural machinery, this increases the difficulty of operation; for robots, it increases the complexity of path planning, trajectory tracking, and obstacle-avoidance control.

In addition to horizontal constraints, vertical-height constraints also determine whether equipment can operate smoothly. Vertical-space constraints mainly depend on the clearance relationship among the lowest point of the photovoltaic modules, the lower edge of the support structure, and the highest point of the agricultural machinery. Under-panel mechanized operation is feasible only when the clearance height satisfies machinery-passage requirements [74]. For elevated agrivoltaic systems, greater module installation height can retain under-panel passage space, allowing tractors, sprayers, and some harvesting equipment to perform conventional agricultural operations beneath the modules [75]; in low-profile or dense systems, large equipment is difficult to accommodate, and operations must rely on low-profile and autonomous equipment [76]. Insufficient clearance height restricts tractors, sprayers, harvesting platforms, and combine harvesters from entering under-panel space and also creates safety risks for operators or robots when approaching the modules. For elevated agrivoltaic systems, the height of the lowest module point should be determined in combination with the height of commonly used regional agricultural machinery, final crop height, and the height of operating mechanisms. If early design considers only planting or management stages while ignoring the height required during harvesting, key operations may later become impossible to mechanize.

Representative engineering projects further demonstrate that array parameters must be designed around the actual agricultural-machinery system. In a German elevated agrivoltaic system, ground clearance is 5 m, photovoltaic-module row spacing is 9.5 m, and module-row width is 3.4 m. The large clearance and support span allow combine harvesters to operate beneath the array; however, increasing row spacing also reduces photovoltaic installed capacity per unit area by approximately 25% compared with conventional ground-mounted photovoltaic power stations [39]. Vertical and single-axis tracking systems use 9 m row spacing and set a 0.5 m buffer on each side of the photovoltaic structure, resulting in an effective cultivation width of approximately 8 m and an approximately 11% loss of cultivable area [77]. Actual operations show that, in the absence of high-precision automatic navigation, a 0.5 m buffer is still insufficient to ensure safety, and operators need to further increase the distance on each side.

Table 6. Adaptation ranges of typical agricultural machinery and agrivoltaic array spatial parameters.

Typical Equipment	Main Dimensional Characteristics	Minimum Clearance Height	Effective Passage or Array Parameters	Safety Margin and Operating Limitations
Dryland tractors and seeding/tillage equipment	Maximum tractor width 2.5 m and maximum height 2.8 m; median working widths of seeders and cultivators are both 4.00 m, and that of disc harrows is 4.94 m [20]	Minimum clearance for dryland operations in the Korean elevated system is 3.0 m [20]	Minimum clear distance between supports in the Korean elevated system is 3.3 m [20]	The 3.3 m support clearance was determined by adding 10% to the maximum implement width of 3.0 m [20]
Rice transplanters and crawler combine harvesters	Maximum transplanter width 3.0 m; maximum combine-harvester width 2.4 m and maximum height 3.1 m [20]	Minimum clearance for paddy-field operations in the Korean elevated system is 3.4 m [20]	Minimum clear distance between supports is 3.3 m and is matched with rice planting row spacing [20]	Passage width is determined by adding 10% to the maximum implement width [20]
Combine harvesters and harvesting headers	Average harvesting-header working width 7.50 m and median width 7.00 m [20]	Minimum clearance height of the German Heggelbach elevated system is 5.0 m [59]	Module-row width is 3.4 m and clear distance between module rows is 6.3 m [59]	A 13.0 m support spacing is determined from two 6.0 m machinery working widths plus a total 1.0 m safety margin.
6 m boom sprayer (German Heggelbach elevated system)	The locally common boom working width used in project design is 6.0 m [20]	Clearance height is 5.0 m [39,59]	Nominal support spacing is 19 m transversely and 12 m longitudinally, corresponding to clear passage widths of approximately 18.4 and 11.75 m [39,59]	The 19 m transverse spacing was determined according to an integer-multiple relationship with the working width of locally common machinery

presents the adaptation ranges of typical agricultural machinery and agrivoltaic-array spatial parameters.

Different agrivoltaic structures have different effects on mechanization adaptability. Measurements in agrivoltaic test fields show that when the minimum module-edge height is approximately 3.2 m, tractor field-operation efficiency is approximately 72%, close to the open-field control; when the minimum edge height decreases to 0.91 m, operation efficiency falls to approximately 50% because of increased turning, reversing, and implement-adjustment time [78]. For the effect of array row spacing, scenario analysis using 6–10 m of effective operating space, implement widths of 1–9 m, and speed reductions of 5%, 15%, and 25% shows that field-operation efficiency may fall to approximately 45% when effective operating space and implement width are severely mismatched [20]. The restrictions imposed by photovoltaic arrays on equipment dimensions and configuration also align with the trends toward electrification and miniaturization of agricultural machinery. Existing studies show that electric-drive systems can reduce the structural constraints imposed by conventional mechanical transmission through flexible layout and refined energy management, creating conditions for narrow and low-profile chassis design [79,80,81]. Related results have extended from large hybrid harvesting equipment to small orchard tractors and dedicated harvesting machinery. Among these, a small leafy-vegetable combine harvester improves passage and maneuverability in restricted spaces by reducing the overall machine envelope, configuring dedicated operating components, and using drive-by-wire control [82]. Greenhouse crawler electric tractors also demonstrate that low-speed electric-drive chassis can meet continuous operating requirements in facility spaces [83].

Overall, the restrictions imposed by photovoltaic arrays on clearance, passages, and turning space make it difficult for conventional large agricultural machinery to adapt directly to under-panel operations. Equipment operating beneath arrays therefore needs to emphasize compact configuration, flexible drive, and task-switching capability; small electric platforms and replaceable operating modules are consequently more suitable for tasks such as inspection, crop protection, irrigation, and transport. The development of agrivoltaic equipment should shift from passive modification of large individual machines to dedicated platform design for array environments.

5.2. Positioning and Navigation in GNSS-Constrained Environments

Among technologies validated in photovoltaic-farm scenarios, the mean carrier-to-noise density ratio in photovoltaic-covered areas of an elevated agrivoltaic orchard in Germany decreased from 30.62 dB-Hz to 26.92 dB-Hz, and horizontal and vertical positioning accuracy were also affected [25]; however, PDOP, HDOP, VDOP, and time to first fix did not change significantly, indicating that positioning degradation is mainly associated with module obstruction, reflection from metal supports, and increased multipath observations. Therefore, photovoltaic farms should not simply be regarded as completely GNSS-denied environments: RTK-GNSS can provide global coordinates outside the array and in headland areas, whereas after entering areas beneath modules or close to supports, GNSS weight should be reduced according to carrier-to-noise density ratio, number of available satellites, and RTK fixed-solution status, and other sensors should maintain relative positioning. For tracking arrays, module tilt and operating time should also be included in GNSS-reliability evaluation.

A zoned scheme combining in-row LiDAR navigation beneath the panels with GPS correction in open headland areas can support inter-row movement and obstacle avoidance [84]. In agrivoltaic soybean fields, field deployment of RTK-GPS, 3D LiDAR, and robotic platforms has also been demonstrated [85]. However, these studies have not yet fully reported path-tracking error, cross-row misclassification rate, or localization-loss frequency; therefore, they should be regarded as direct evidence of signal characteristics, zoned navigation, and platform deployment rather than evidence that complete autonomous-navigation performance has already been validated.

For transferable technologies and targeted improvement experiments, studies in other structured agricultural scenarios further show that fusion of visual SLAM with inertial information, LiDAR row-center extraction, and LIO-SAM fused with artificial markers can respectively use visual, geometric, and prior features to maintain relative positioning under GNSS degradation and suppress accumulated drift during long-term operation [86,87,88,89]. Regularly arranged supports, beams, and lower module edges can provide stable array-direction and lateral-distance constraints for LiDAR positioning. Photovoltaic supports can be identified, their three-dimensional positions recovered, and the array direction fitted to generate navigation targets parallel to the module rows [90]. However, repeated support spacing may also produce similar point clouds in adjacent arrays, causing positioning drift along the array direction. Regularly arranged supports, beams, and module edges can provide lateral-distance and heading constraints for LiDAR, but the high similarity of point clouds among adjacent array rows can easily cause cross-row mismatching and longitudinal drift. Therefore, LiDAR is more suitable for under-panel lateral positioning and short-range obstacle avoidance. When multiple array rows have similar matching scores, the system should reduce speed and request absolute-position correction rather than continue using a low-confidence SLAM result.

In addition, soft soil, standing water, slope changes, and frequent turns in agrivoltaic fields can readily cause wheel slip, and implement vibration can accelerate inertial-error accumulation. Therefore, IMU and wheel-speed odometry are only suitable for maintaining positioning continuity during short-term degradation of GNSS or vision and should not be used as independent positioning sources for long-distance under-panel operations. Slip can be detected by comparing the consistency of displacement estimated from wheel speed, IMU, and LiDAR, and accumulated error should be corrected promptly after a reliable GNSS fixed solution or artificial-marker observation is regained.

Based on field-measured GNSS degradation in agrivoltaic scenarios, demonstrations of LiDAR–GPS zoned navigation, and multi-sensor fusion methods from other structured scenarios, this review proposes the conceptual fusion framework shown in Figure 6. The framework is intended to adjust the observation weights of different sensors according to GNSS signal quality, observability of array geometry, and visual-imaging quality, while using IMU and wheel-speed odometry to provide short-term motion constraints. Its overall performance still requires validation under real agrivoltaic operating conditions.

As shown in Figure 6, GNSS, LiDAR, and RGB-D respectively provide global position, array geometry, and visual-environment information, while the IMU provides attitude and short-term motion constraints. When GNSS is affected by module obstruction and multipath interference, LiDAR is affected by repetitive array structures, or vision suffers from underexposure and specular reflection, the onboard computing unit can dynamically adjust observation weights according to sensor quality, thereby reducing the effect of degradation in any single sensor on continuous positioning.

Visual navigation is affected by hard shadows from modules, glass reflections, and rapid light–dark transitions. When a robot enters a shadow, it may become briefly underexposed, and when it exits the shadow, it may become overexposed, resulting in fewer feature points and failed inter-frame matching. To address this, semantic features such as supports, lower module edges, and passage boundaries should be extracted preferentially, and exposure time and visual-observation weight should be adjusted in real time according to image entropy, brightness, gradient, and the number of tracked features [91]. When image quality decreases, positioning should temporarily rely on LiDAR and IMU; after exposure and feature tracking recover, visual constraints can be reintroduced. For low-illumination and strongly shaded areas that remain after exposure adjustment, low-light image enhancement can further be used to recover passage edges, and semantic-segmentation networks such as UNet can be used to extract traversable areas [92]. On this basis, short-term visual memory and recognition of bright and dark regions can support obstacle avoidance and motion decision-making, providing a methodological reference for bright–dark region recognition and continuous passage navigation beneath photovoltaic arrays [93]. Different schemes differ substantially in accuracy, robustness in array environments, hardware cost, and applicable regions, as compared in

Table 7. Performance and applicability boundaries of navigation and positioning schemes in agrivoltaic scenarios.

Navigation Scheme	Overall Robustness Beneath Photovoltaic Arrays	Relative Cost and Deployment Complexity	Current Engineering Positioning
Single RTK-GNSS	Sensitive to sky obstruction, module reflection, and multipath from metal supports; errors may vary periodically between array rows, resulting in low robustness.	Moderate cost. No additional environmental sensors are required, but a differential base station, network service, or reference station is needed.	Suitable as a source of global coordinates outside the array; not suitable as the sole positioning method for operations near supports beneath the modules.
GNSS–IMU–wheel-speed fusion	Can alleviate short-term obstruction but is difficult to use for continuous long-distance operation beneath the panels; robustness is moderate.	Moderate cost. Sensors are mature and computational load is moderate, but calibration and tire-slip handling are required.	Suitable for array entrances, exits, and transition areas with short-term GNSS degradation; it is an option between single GNSS and autonomous positioning.
LiDAR–IMU-coupled SLAM positioning	Essentially unaffected by bright–dark changes but affected by dust, rainfall, crop occlusion, and sparse point clouds; robustness is relatively high.	High cost. Three-dimensional LiDAR and real-time computing platforms are expensive, with high calibration and maintenance requirements.	Can serve as the main relative-positioning scheme for areas beneath modules where GNSS is constrained for long periods.
RGB-D semantic visual navigation	Can identify support and passage semantics but is susceptible to hard shadows, glass reflection, exposure lag, and support occlusion; robustness is moderate.	Moderate cost. Cameras are relatively inexpensive, but dedicated datasets, model training, and continuous updates are required.	Suitable for array-structure recognition, row-number determination, and local-path generation; it should not independently undertake safety-critical positioning.

.

After obtaining a reliable pose, the robot still needs to complete path planning, local replanning, and obstacle recognition. Existing reviews have systematically summarized obstacle-detection and avoidance methods and path-planning and tracking-control technologies in complex agricultural environments [94,95]; for repetitive row structures and dynamic-obstacle environments, combining object detection with visual SLAM has also been used to improve robot localization and environmental mapping [96]. In addition, region segmentation and coverage-path planning can optimize operating direction and coverage sequence in narrow operating areas, improved D* Lite can reconstruct paths promptly when local passages are blocked, and adaptive look-ahead distance and steering-state feedback can improve path-tracking stability [97,98,99]. Based on a visual navigation path, robot-center deviation and heading deviation can be calculated in real time and corrected through steering control [100]. Obstacle-recognition results should also be combined with positioning confidence and passage geometry rather than simply following the shortest path; deep-learning-based field obstacle detection can provide the basis for this local avoidance [101,102].

Overall, current direct evidence mainly verifies the effects of elevated photovoltaic structures on GNSS signals, the engineering feasibility of LiDAR in-row navigation with GPS headland correction, and the deployment capability of RTK-GPS and LiDAR platforms in agrivoltaic fields; visual–inertial fusion, SLAM, semantic passage recognition, and coverage-path planning remain mainly cross-scenario transferable technologies. Future studies need to test positioning error, array-row mismatching rate, and path-tracking deviation separately between array rows and at array edges before the practical boundaries of different navigation schemes can be further determined.

5.3. Comprehensive Benefit Evaluation of Agrivoltaic Systems

Benefit evaluation of agrivoltaic systems comprehensively considers factors such as agricultural production, photovoltaic power generation, resource utilization, and engineering cost. Agrivoltaics has substantial potential in food and energy production and can provide farmers and operators with dual income through crop production and electricity sales. Existing studies show that the economic advantages of agrivoltaic systems are strongly condition-dependent. The economic and environmental costs of agrivoltaics are comparable to those of other photovoltaic options to a certain extent, but its value also lies in reducing additional land occupation and maintaining agricultural-production functions [103]. Comparisons among different array configurations show that techno-economic performance is influenced not only by power generation, but also by array investment, agricultural returns, and land-use mode [104]. These results indicate that the economic rationality of agrivoltaics depends on land opportunity cost, changes in agricultural output, and policy conditions. Related surveys show that farmer income opportunities, local economic benefits, project siting, and fairness in benefit distribution remain key factors determining social acceptance [105]. When the evaluation boundary is extended from the power-generation side to crop revenue per unit land area, multiple agrivoltaic configurations may outperform stand-alone photovoltaics, although some scenarios still require additional electricity-price support to compete with stand-alone agriculture [104]. An East African case also shows that mounting structures can account for approximately 45% of the initial investment in systems without storage, extending the payback period by 1–3 years compared with ground-mounted photovoltaics, although additional crop income can partially offset this difference [106]. LCOE focuses on electricity-generation cost, land-revenue models simultaneously account for agricultural output, and regional project evaluations are additionally affected by financing cost, electricity price, and policy support.

The land equivalent ratio is a commonly used integrated productivity indicator for agrivoltaic-system evaluation. It compares the combined crop and energy output produced within the same land unit with separate production; when LER is greater than 1, combined use has an overall output advantage [21]. However, the benefits of photovoltaic greenhouses depend on module light-transmission characteristics, roof-coverage layout, and crop light requirements [107]. Lifecycle assessment further extends the evaluation boundary to the entire process of system construction, agricultural production, and end-of-life recycling. Existing LCA studies show that the environmental advantages of agrivoltaic systems do not increase in parallel with land-use efficiency. Elevated agrivoltaic systems increase support height and structural strength to retain agricultural-production space, thereby increasing construction and electricity-generation costs [39]. Vertical systems or systems with simpler structures require relatively less material input, but may be constrained in terms of crop suitability and energy output per unit land area [108]. In addition to photovoltaic modules, the production and transport of raw materials for supporting facilities such as supports, inverters, and energy storage must be included in accounting; on the agricultural side, production inputs and field operation and maintenance must also be considered.

Equipment and operational adaptation is an important evaluation dimension that distinguishes agrivoltaic systems from conventional photovoltaic power stations. Evaluation indicators for agricultural-machinery operations need to cover four dimensions: passage performance, operation quality, operational safety, and economic energy consumption, including key parameters such as passage efficiency, operation accuracy, obstacle-avoidance capability, energy consumption, and operation-and-maintenance cost. Because supports, module edges, and crop distribution jointly alter passage conditions and operating environments, the performance of the same equipment beneath panels, between panels, and at array edges is often inconsistent. Therefore, the evaluation focus should vary with the task: crop-protection and irrigation operations place greater emphasis on operation uniformity and input utilization, navigation operations focus on path deviation and obstacle-avoidance safety, and module operation and maintenance require evaluation of cleaning effectiveness, defect detection, and energy consumption. Compared with a single efficiency indicator, this type of evaluation based on actual operating outcomes can better reflect the suitability of equipment in agrivoltaic scenarios.

Economic advantages at the site level also do not mean that intelligent equipment is necessarily investment-feasible; its benefits should be calculated incrementally relative to a conventional agrivoltaic system. Savings in labor and inputs and reductions in soil-compaction losses may offset the cost of automated control, but the break-even point varies substantially with farm area, the number of suitable operating days, and the need for human supervision [109]. Field fragmentation, relocation distance, and passage conditions also reduce effective equipment utilization and increase cost per unit area [110]. In agrivoltaic scenarios with dense supports, restricted passages, and shared time windows for agricultural operations and module maintenance, losses caused by equipment waiting, route detours, and downtime may be even more pronounced. Therefore, expanding the service area can distribute fixed investments in sensors, communications, and software, but does not necessarily continue to reduce marginal cost. Long-term evaluation must also include battery and sensor replacement, software subscriptions, equipment renewal, and withdrawal of subsidies; intelligent investment is scalable and economically sustainable only when cost savings and additional income continuously cover maintenance and renewal expenditure over the full life cycle.

There are continuously changing trade-offs among crop growth, photovoltaic power generation, and machinery operations. Changes in weather, crop-growth stage, and shading position may all make an existing scheme no longer applicable. The linkage between environmental monitoring and operational decision-making should be further strengthened so that changes in light, thermal, and water conditions can be promptly translated into specific adjustment criteria for irrigation, crop protection, travel speed, and operating areas. Only in this way can array design, crop production, and equipment operation gradually shift from mutual constraint to coordinated operation.

6. Light–Thermal Sensing and Prediction for Adaptive Agricultural-Machinery Operations

6.1. Environmental Sensing Beneath Photovoltaic Arrays

Photovoltaic arrays transform the remote-sensing object of farmland from a relatively continuous crop canopy into a composite scene in which modules, shadows, crops, and inter-panel passages are interwoven. Solar elevation, module tilt, and array row spacing further cause land-cover composition and illumination state to change with time. Therefore, environmental sensing in photovoltaic farms should focus on resolving parameter-retrieval biases caused by mixed pixels, structured shadows, and array geometry, and converting corrected light–thermal information into agricultural-machinery operation zones and regulation bases. In general agricultural remote sensing, high-resolution satellites, Landsat/Sentinel, and UAV platforms have been used respectively for crop-yield estimation, vegetation gross primary productivity, and land-surface temperature estimation. Related studies show clear differences among platforms in spatial resolution, observation timing, and matching with ground scales [111,112,113,114]. These cross-scale differences are further affected in agrivoltaic scenarios by photovoltaic-module obstruction, structured shadows, and mixed pixels; therefore, satellite or UAV inversion results cannot be used directly to judge under-panel crop status without correction.

Satellite or low-resolution UAV imagery in photovoltaic arrays usually contains modules, shadows, vegetation, bare soil, and inter-panel passages within the same pixels, and direct calculation of vegetation indices or land-surface temperature can therefore produce clear mixed-pixel bias. High-resolution UAV imagery or semantic-segmentation methods should first be used to extract masks for modules, shadows, crops, and passages, after which pixel decomposition and scale correction should be carried out according to area fractions of different objects, spectral endmembers, and ground observations. By combining field surveys, high-resolution UAV imagery, and Sentinel-2 data, Zou et al. corrected vegetation monitoring around photovoltaic power plants and increased estimated NDVI from 0.248 to 0.298 after excluding photovoltaic panels and bare soil [28], indicating that array-structure information should be used explicitly in remote-sensing inversion rather than treating the photovoltaic site as homogeneous vegetation.

Only after modules, shadows, vegetation, bare soil, and inter-panel passages have been separated can crop growth and water–thermal stress beneath photovoltaic arrays be reliably retrieved. In this scenario, a decrease in vegetation index may indicate restricted crop growth, but may also be caused by module shading and exposure changes; abnormal canopy temperature may result from restricted transpiration, but may also be affected by hot modules and bare-soil mixing. Therefore, semantic segmentation should be used to establish masks for modules, sunlit vegetation, shaded vegetation, bare soil, and passages, and vegetation indices, canopy temperature, and water-stress indicators should be calculated separately within each region [115,116]. In existing ground-based photovoltaic scenes, module recognition has already achieved high precision through object detection and deep-learning segmentation [117,118]. These methods can provide a basis for identifying modules, but under agrivoltaic conditions, crops may overlap with supports and panel shadows, so direct transfer of photovoltaic-site segmentation models is still insufficient. For under-panel crops, model training should explicitly include sunlit leaves, shaded leaves, photovoltaic panels, bare soil, and inter-panel passages, and a data-adaptation layer should be introduced to reduce differences caused by array structure, illumination, and sensor type [13,24,119].

Array-structure information can further serve as a prior constraint for inversion of light–thermal parameters. Multi-temporal surveys should simultaneously record module height, tilt, azimuth, row spacing, and solar elevation and azimuth, calculate the theoretical shadow range from these data, and then correct actual shadow boundaries using image segmentation. For multispectral imagery, downwelling-light sensors and reflectance calibration panels can be used for radiometric normalization [120]; for thermal-infrared imagery, module and bare-soil pixels should first be excluded, and ground blackbodies or contact temperature measurements should be used to correct temperature bias under different illumination conditions. Flight altitude, flight direction, shooting time, sun position, and module posture should be recorded as mandatory metadata. For short-term comparison within the same site, multi-temporal imagery should be collected at similar solar positions; for cross-season or cross-site comparison, bidirectional reflectance and shadow-proportion correction should be applied before vegetation indices and temperature indicators are compared.

For heterogeneous photovoltaic scenes in agrivoltaics, a physics-guided hyperspectral unmixing model, PTV-Mamba, has been proposed [121]. Built on an extended linear mixing-model framework, the model introduces an illumination-scaling term at the pixel dimension and uses a physics-constrained decoder to decouple multiple interference factors, separately removing photovoltaic-module abundance, overall brightness shifts caused by illumination, and residual spectral distortion induced by scattered light and environmental reflection. Unlike conventional approaches that identify photovoltaic modules mainly from pixel-brightness characteristics, this method can effectively reduce radiometric errors associated with module tilt, local shadows, and complex underlying-surface conditions, which is particularly suitable for quantitative inversion of vegetation beneath photovoltaic arrays. Structured shadows produced by modules still need to be treated separately after mixed-pixel unmixing. Existing ground photovoltaic studies have proposed four shadow categories—module shadow, structure shadow, bush shadow, and cloud shadow—and corrected images according to illumination coefficients in different shadow zones [122]. The method can serve as a reference for identifying structured-shadow zones in agrivoltaic fields, but its direct applicability to crop canopies still requires validation. In addition, a triple multilinear unmixing model can use spatial and spectral information jointly and has been used for component extraction in complex ground photovoltaic scenes [123].

Environmental sensing in agrivoltaic systems has formed a multi-scale technical route combining continuous observation by fixed nodes, fine mapping by UAVs, and regional monitoring by satellites. However, existing studies still focus mainly on whether sensors can observe environmental differences, with insufficient discussion of whether data are comparable across time, arrays, and platforms. Current bottlenecks include the lack of unified standards for sensor installation height and measurement-point coordinates, incomplete recording of module posture and solar geometry, insufficient cross-scenario validation of shadow and mixed-pixel correction methods, and scale mismatch between remote-sensing results and continuous ground data. Future studies should synchronously incorporate module structure, shadow trajectory, and equipment state into sensing design and establish a complete data chain from raw observation and spatial correction to operation zones.

6.2. Near-Ground Multimodal Sensing for Agricultural Equipment

Unlike in-situ microclimate monitoring with fixed nodes, which emphasizes continuous recording of environmental factors, machine vision and robotic near-ground sensing focus more on detailed characterization of individual crops and canopy status. Fixed-node sensors are usually used to continuously record environmental variables such as PAR, wind speed, and module temperature, whereas near-ground platforms can carry RGB, depth, spectral, thermal-infrared, and LiDAR sensors to obtain structural and physiological information from the side of crops or from within the canopy [124,125,126,127]. In agrivoltaic scenarios, however, the technical focus should not remain at listing sensor types, but should analyze the effects of module shading, structural occlusion, and limited clearance on data quality and equipment deployment.

The primary problem faced by near-ground sensing in agrivoltaic systems is data incomparability caused by structured shading. In an actual agrivoltaic alfalfa system, time-series RGB imagery was combined with near-ground LiDAR, and shading zoning and color normalization were used to improve comparability of data across different times and spaces; however, joint occlusion by modules and vegetation still prevented effective imagery from being obtained in some under-panel areas [128]. This indicates that the core difficulties of under-panel sensing are not limited to insufficient spatial resolution, but also include structural occlusion and radiometric differences between bright and dark regions. For the latter issue, HDR imaging can expand the effective brightness range but cannot independently eliminate shadow misclassification; further combination with color-space transformation and multi-threshold shadow removal can improve vegetation-segmentation accuracy and correction accuracy [129]. In addition, sunlit areas, deep-shadow areas, and light–dark transition areas can be treated as different visual domains, with data augmentation and domain adaptation used to alleviate cross-region performance degradation of models [130].

At the hardware level, active illumination can improve the consistency of spectral and fine-phenotyping data. PhenoRob-F stabilizes hyperspectral and RGB-D acquisition conditions through an adjustable imaging module and artificial light source [131]. Taken together, existing studies indicate that the focus of multimodal optimization beneath photovoltaic arrays is not to increase the number of sensors, but to dynamically adjust the reliability of each modality according to illumination quality, degree of occlusion, and data completeness. When RGB images show overexposure, underexposure, or shadow mis-segmentation, the weight of visual results should be reduced and supplemented with depth, point-cloud, or multi-view information; the occlusion ratio, data confidence, and abnormal areas should also be output.

Airborne remote sensing mainly characterizes the top of the canopy from a top-down view and has limited ability to observe lower leaves, stem bases, and side structures of the canopy. Ground-based proximal platforms can supplement this information through side-view, oblique-view, and close-range observations. Existing comparative studies also show that ground-based LiDAR can retain more structural detail at the top and sides of the canopy than UAV photogrammetry [132,133]. Mobile platforms can additionally integrate RGB, thermal infrared, imaging spectroscopy, and depth sensors to continuously obtain crop structural and physiological information at relatively high spatiotemporal resolution [134,135]. For crops beneath photovoltaic arrays, this type of near-ground data can not only supplement under-panel blind areas in UAV and satellite images, but can also further identify local differences in crop development among different regions of the system.

Three-dimensional near-ground imaging can obtain structural parameters such as canopy-height distribution, leaf area, and leaf angle from side-view and close-range perspectives, compensating for the insufficient representation of canopy sides and organ-scale information by a single top-view image [136,137,138]. Among these methods, RGB-D combined with lightweight instance segmentation can extract plant height and canopy geometry at relatively low equipment and computational cost, and related three-dimensional reconstruction studies have also demonstrated that visual sensors can quantify plant height, leaf area, and canopy structure [127,139]. However, three-dimensional measurement accuracy obtained under conventional farmland conditions cannot directly represent performance beneath panels, because module occlusion, low light, and limited observation distance may reduce point-cloud completeness. Therefore, three-dimensional parameters still need to be calibrated through multi-view acquisition, marking of occluded areas, and under-panel field measurements.

The value of near-ground sensing lies not only in recording crop status, but also in converting phenotypic information into equipment-control parameters. In targeted spraying operations, spray decisions need to simultaneously consider crop, environmental, and equipment information and adjust spray volume, nozzle on/off status, and operating speed accordingly [140]. Deep learning and machine vision can identify and spatially locate weeds and select treatment methods such as mechanical removal or spot spraying accordingly, thereby reducing repeated operations and chemical input in non-target areas [141]. Therefore, near-ground vision can be used to identify canopy boundaries and density, and distance and attitude information can be combined to adjust nozzle position, reduce speed, or shut off local nozzles when approaching supports or modules. Taken together, different monitoring platforms show clear differences in measurement accuracy, observation scale, temporal continuity, and implementation cost.

Table 8. Performance, cost, and applicability of different environmental-monitoring schemes in agrivoltaic systems.

Monitoring Scheme	Accuracy and Spatiotemporal Scale	Implementation Cost and Components	Applicability to Agrivoltaics
Fixed sensor network	High single-point accuracy and temporal resolution, but limited spatial representativeness	Single-node purchase cost is relatively low; multi-zone deployment requires data loggers, power supply, communication, and protective facilities, while long-term calibration, fault replacement, and network maintenance increase lifecycle cost	Suitable for continuous monitoring and model calibration; under-panel, inter-panel, and edge zones need to be covered
UAV multispectral/thermal infrared	Centimeter-level spatial resolution, but data can only be acquired by flight mission	The flight-platform cost is moderate and multispectral and thermal-infrared payloads are relatively expensive; repeated monitoring also involves route planning, battery maintenance, image mosaicking, and data processing, resulting in relatively high integrated cost for high-frequency aerial surveys	Suitable for fine zoning in open arrays; blind areas exist beneath panels and inside greenhouses
Satellite remote sensing	Good regional coverage and long-term continuity, but intra-array accuracy is limited by mixed pixels	Open data such as Landsat and Sentinel have low acquisition costs, while commercial high-resolution imagery requires purchase or subscription; no field deployment is required, but cloud/shadow processing, mixed-pixel correction, and multi-temporal data analysis increase computing and personnel costs	Suitable for regional evaluation of large projects; not suitable for directly judging crop conditions beneath panels
Near-ground imaging and LiDAR	High local structural accuracy, but affected by occlusion, low light, and platform passage	RGB and depth cameras are relatively inexpensive, while thermal infrared, hyperspectral, and three-dimensional LiDAR are more expensive; costs also include mobile-platform integration, multi-sensor synchronization and calibration, supplemental lighting, dust/water protection, point-cloud processing, and data storage	Suitable for monitoring canopy structure, crop stress, and remaining clearance

compares the applicability boundaries of fixed sensors, airborne remote sensing, and near-ground sensing technologies in agrivoltaic scenarios.

The hierarchical monitoring shown in Table 8 has practical significance only when data at different scales can be mutually calibrated. Differences in vegetation indices, canopy temperature, and structural parameters obtained by different platforms do not necessarily represent real changes in crop status, because solar position, shadow proportion, mixed pixels, and ground-calibration methods can all affect inversion results. Satellite imagery may underestimate vegetation status because modules and bare soil are mixed into pixels, whereas shadow correction in UAV and near-ground imagery may weaken real stress signals. In addition, methods such as HDR, active supplemental illumination, and domain adaptation have mainly been validated in conventional farmland or general phenotyping scenarios, and their applicability beneath photovoltaic arrays still lacks same-site comparative evidence. The value of light–thermal sensing should be evaluated according to the degree to which it reduces environmental-prediction error and uncertainty in agricultural-machinery decision-making.

6.3. Prediction Methods for Light–Thermal–Water Environments

Light-resource prediction in agrivoltaic systems has certain methodological commonalities with conventional photovoltaic power prediction. The difference is that conventional photovoltaic prediction mainly focuses on module output power, whereas agrivoltaic systems also need to estimate under-panel radiation, effective photosynthetic light for crops, and its spatial uniformity, while further considering coordination between agricultural-production and power-generation objectives [142]. Existing methods can generally be divided into three categories: physical models, statistical models, and machine-learning methods.

Physical models have the advantage of explicitly introducing module height, tilt, row spacing, and solar position and are suitable for explaining spatial differences in light–thermal–water environments after array configuration changes. The radiation-distribution model for photovoltaic greenhouses developed by Cossu et al. was validated in a commercial greenhouse with 50% rooftop photovoltaic coverage; the mean R² over 12 d of observations was 0.88 ± 0.06, although error varied with measurement position, and its applicability was mainly limited to fixed photovoltaic greenhouses with known geometric parameters [143]. For open agrivoltaic systems, a coupled module-geometry–solar-radiation–CFD model predicted air and soil temperatures under three module heights, with relative errors between simulated and measured values below 8% [144]. Therefore, although physical models have strong configuration interpretability, their cross-regional application still depends on recalibration of optical parameters, soil thermal parameters, and boundary meteorological conditions. Cossu et al. study, Gong et al. study.

In general agricultural environments, the Stanghellini model, Shuttleworth–Wallace model, and crop-coefficient method have been used respectively for greenhouse and field evapotranspiration estimation, showing that canopy resistance, crop coefficients, and energy-partition parameters are important factors affecting evapotranspiration prediction [145,146,147,148,149,150]. However, these models are mostly based on unshaded or relatively uniform-radiation conditions, and their parameters cannot be transferred directly to agrivoltaic systems; recalibration is still required according to radiation and canopy state beneath panels, between panels, and in shadow-transition zones. For crop-response prediction, Moretta et al. coupled three-dimensional radiation, soil–water balance, and alfalfa-growth processes. Across open-field, fixed-array, and dual-axis tracking treatments, yield prediction achieved R² = 0.94 and RMSE = 110.34 g·m⁻², LAI prediction achieved R² = 0.96 and RMSE = 0.31, and canopy–air temperature-difference prediction achieved R² = 0.83 and RMSE = 0.88 °C [40]. This allows crop-canopy changes to feed back into subsequent radiation calculations and enables continuous simulation of shadow evolution, light–thermal–water environments, and crop growth.

Data-driven models use historical radiation, meteorological, and system-operation data to learn relationships between input variables and prediction targets, including methods such as regression analysis, support vector machines, neural networks, and fuzzy inference [151,152]. They have relatively high computational efficiency and can fit complex nonlinear relationships, but most studies use point radiation or photovoltaic output as prediction targets and do not include module height, row spacing, real-time orientation, and crop spatial position as inputs; therefore, they have difficulty directly describing spatial heterogeneity of the under-panel environment.

6.3.1. Conventional Machine-Learning Prediction

Solar-radiation prediction provides a methodological basis for estimating under-panel light resources in agrivoltaic systems. Nonlinear regression methods can establish prediction relationships using historical radiation and weather information and have good applicability when complete mechanistic parameters are unavailable [153]. In comparisons of NARX, MLP, ARMA, and persistence models for hourly global solar-irradiance prediction in New Zealand, NARX introduces historical outputs and external meteorological variables and can effectively describe the time-lag characteristics of radiation changes [154]. The predictive performance of conventional machine learning depends not only on algorithm type, but also on input features, data quality, and validation method [155]. To summarize the general process of light–thermal environment estimation beneath photovoltaic arrays, Figure 7 summarizes data preprocessing, model construction, and performance evaluation.

As shown in Figure 7, light–thermal environment estimation usually includes three stages: data preprocessing, model construction, and performance evaluation. Effective features are screened through correlation analysis, collinearity testing, and outlier processing; physical models, conventional machine-learning models, and deep temporal models are then established separately and validated using typical-weather and daily data; finally, the predictive performance of different models is compared using RMSE and R².

In conventional machine-learning research, nonlinear networks and combined prediction are important areas. For hourly solar-radiation prediction using multilayer feedforward neural networks, radial-basis-function networks, support-vector regression, and ANFIS, different models show different sensitivities to combinations of input variables and weather types [156]. Fidan et al. used harmonic analysis to extract daily-periodic characteristics of solar radiation and used morning observations to estimate radiation changes during subsequent periods, providing another approach to feature construction from historical sequences [157]. Combining fuzzy logic with neural networks uses fuzzy inference to process uncertainty in sky conditions and meteorological information and then uses neural networks to establish radiation-prediction relationships, improving adaptability under changing weather conditions [158]. Under-panel radiation prediction is closely related to model type, weather classification, historical time window, and input-feature selection.

Conventional machine-learning studies can provide candidate methods for under-panel radiation and transpiration prediction, but existing accuracy mainly comes from conventional solar-radiation or greenhouse-crop scenarios. In New Zealand hourly global solar-irradiance prediction, the NARX model had an RMSE of 0.243 MJ·m⁻², lower than those of MLP, ARMA, and persistence models at 0.484, 0.315, and 0.514 MJ·m⁻², respectively [154]. In greenhouse tomato studies, CARS-CatBoost achieved whole-season transpiration-rate prediction with R² = 0.917 and RMSE = 0.014 mm·h⁻¹ [159], while a random-forest model incorporating relative leaf-area index achieved R² values of 0.9472 and 0.9654 and RMSE values of 19.75 and 18.85 g in two test sets [160]. These results indicate that crop-state variables help improve prediction accuracy for water processes. However, such systems still cannot reveal the applicability boundaries of different algorithms in agrivoltaic scenarios, particularly under complex weather conditions such as cloudy and rainy periods, for which estimation stability remains unclear.

6.3.2. Deep Temporal Prediction

With the accumulation of continuous monitoring data, light–thermal environment prediction has gradually shifted from static nonlinear mapping to deep temporal modeling. Recurrent neural networks can use historical hidden states to describe continuous changes in environmental variables and are suitable for radiation, temperature, humidity, and other sequences with temporal dependence. By retaining historical hidden states, recurrent neural networks can describe the continuous temporal changes of greenhouse environments. In greenhouse microclimate prediction, introducing lagged time windows into a hybrid RNN model produces lower errors than radial-basis networks and ordinary feedforward networks, indicating that recurrent structures are more suitable for environmental sequences with temporal dependence [161]. As prediction time scales shorten and the number of input variables increases, research has further shifted toward deep temporal networks such as LSTM. Synthetic weather forecasts have been used as model inputs for short-term photovoltaic-power prediction with LSTM, showing that forecast meteorological information can supplement prediction based only on historical sequences [162]. Combining weather forecasts with LSTM has also enabled hourly day-ahead solar-irradiance prediction [163]. Although these studies mainly address conventional solar-energy systems, their time-series modeling approaches can be transferred to agrivoltaic systems for predicting intraday changes in under-panel radiation and PAR. In crop-process prediction, LSTM has been used to estimate greenhouse-tomato transpiration. Compared with nonlinear autoregressive networks, Elman networks, and ordinary RNNs, LSTM uses gating structures to reduce information decay during long-sequence training and shows better predictive performance in indicators such as coefficient of determination and mean absolute error [164]. Deep temporal models can therefore predict not only thermal and moisture environments, but also the propagation of environmental changes into crop-transpiration responses.

To enhance extraction of time-series features, CNN and LSTM can be combined, with convolutional layers identifying local change patterns in photovoltaic-power and related meteorological sequences and LSTM then describing dependencies between successive time points, thereby improving hourly photovoltaic-power prediction [165]. Combining CNN–LSTM with feature pyramids can also fuse environmental features at different scales [166,167]. Here, the CNN mainly processes continuous numerical sequences rather than directly extracting spatial features from sky images. To address rapid irradiance fluctuations caused by cloud movement, sky images have further been introduced into short-term photovoltaic-output prediction, using deep learning to extract visual information such as cloud cover and obstruction status so that models can more directly reflect the effects of cloud changes on photovoltaic output [168].

On the basis of deep temporal models, attention mechanisms can allocate weights according to the contribution of different time steps and input variables to prediction results. When LSTM is combined with an attention mechanism, key periods such as abrupt cloud-cover changes, rapid irradiance decreases, and temperature changes can be emphasized; this approach has been used for short-term photovoltaic-power prediction [169]. In agrivoltaic scenarios, solar radiation and air temperature beneath photovoltaic arrays have significant spatiotemporal heterogeneity, and temporal models such as LSTM-Attention can effectively capture dynamic light–thermal changes under different coverage densities and complex weather conditions, providing a data basis for crop management and array regulation [170]. This can improve prediction stability under cloudy conditions and rapid light–dark transitions. Combinations of generative adversarial networks and convolutional neural networks are mainly used for weather classification and photovoltaic-power prediction; by generating or expanding samples of different weather types, they alleviate insufficient training data for rare operating conditions such as rain and strong convection [171]. Therefore, GAN is more suitable as a data-augmentation method than as a temporal-prediction model directly parallel to RNN or LSTM.

The dynamic shading effect of photovoltaic modules causes light–thermal parameters within the system to exhibit pronounced time-series characteristics and spatial heterogeneity, creating dual challenges for temporal-feature capture and nonlinear-mapping accuracy. LSTM has advantages in handling long-term sequence dependencies because of its gating mechanism. As shown in

Table 9. Comparison of light–thermal resource prediction models.

Model Type	Typical Models	Advantages	Limitations	Suitable Prediction Targets
Physical-mechanism models	Solar geometry, energy balance, CFD	Strong interpretability	Many parameters and relatively high zoning-calibration and computational costs	Shadow distribution and temperature–humidity environment
Traditional machine learning	SVM, RF, XGBoost, ANFIS	Suitable for nonlinear regression and relatively fast to train	Depends on feature selection and is insufficient for modeling long-term dependence	Radiation, temperature and humidity, transpiration, and irrigation demand
Time-series deep learning	LSTM, RNN, CNN–LSTM	Can extract local fluctuations and long-term temporal dependence	Large data requirements and limited cross-scenario generalization	Radiation, PAR, canopy-temperature, and transpiration sequences
Attention models	LSTM-Attention, Transformer	Can focus on critical times and spatial regions	Complex structure and relatively high training and computational costs	Light–thermal prediction under cloudy weather and alternating bright–dark conditions

, different model types have different applicability boundaries for environmental prediction in agrivoltaic systems.

Data-driven models may achieve high accuracy when datasets from the same site are randomly divided, but such results can easily overestimate their generalization capability across seasons, crops, and arrays. Physical models may have relatively large local fitting errors, but are more suitable for explaining the effects of array geometry and weather changes. At the same time, a low RMSE obtained for field-average radiation or temperature does not demonstrate that a model can correctly predict shadow edges, light–dark transition times, or local water gradients. Therefore, in addition to conventional error indicators, model evaluation should include cross-site validation, extreme-weather testing, and array-zonal errors.

Agrivoltaic-environment estimation models have evolved from high-accuracy but high-cost ray-tracing models, to energy-balance models coupling light–thermal–water processes, and then to deep hybrid models oriented toward multi-source data and real-time prediction. Future model development should place greater emphasis on multi-source data fusion, adaptive parameter updating, and real-time decision support, in order to improve prediction reliability of agrivoltaic systems under differences in region, weather, and array structure.

7. Coordinated Operation of Agricultural Equipment Beneath Photovoltaic Arrays

7.1. Coordinated-Optimization Framework and Operation-Task Generation

Coordinated optimization of agrivoltaic systems is not the separate optimization of module regulation, agricultural operations, and photovoltaic operation and maintenance, but the joint determination, within unified spatial and temporal boundaries, of module orientation, agricultural tasks, equipment allocation, operating paths, operation-and-maintenance periods, and energy-supply schemes. Let the decision vector $x$ include module tilt, zonal irrigation amount, task start time, equipment–task allocation relationships, travel paths, and charging/discharging power; coordinated agrivoltaic optimization can be expressed as follows:

$${\displaystyle (}\underset{\mathbf{x}}{\min }\mathrm{F}{\displaystyle (}x{\displaystyle )}=\left(1-{\stackrel{{\displaystyle ~}}{E}}_{PV},1-{\stackrel{{\displaystyle ~}}{Y}}_{crop},1-{\stackrel{{\displaystyle ~}}{\eta }}_{op},{\stackrel{{\displaystyle ~}}{C}}_{O\&M}\right){\displaystyle )}$$

(1)

${\stackrel{{\displaystyle ~}}{E}}_{PV}{\stackrel{{\displaystyle ~}}{Y}}_{crop}{\stackrel{{\displaystyle ~}}{\eta }}_{op},{\stackrel{{\displaystyle ~}}{C}}_{O\&M}$ In the equation, photovoltaic power generation, crop yield or production return, and effective operating amount per unit time or unit energy consumption are considered together with equipment energy consumption, labor, maintenance, task delay, and photovoltaic downtime losses. After normalization of different indicators, the trade-offs among power generation, crop production, operating efficiency, and operating cost can be described using a Pareto front, or dynamic weights can be set according to crop-growth stage, electricity price, task urgency, and equipment status. Because final crop yield cannot be obtained directly in short-term rolling control, minute-to-hour scales can use daily-light-integral deficit, canopy heat stress, water deficit, and expected yield loss as proxy indicators, and their cumulative effects on final yield can then be evaluated at the seasonal scale.

Existing studies provide a direct basis for the above multi-objective modeling. Mengi et al. constructed a crop-driven digital replica and used multi-objective genome optimization to search for trade-off solutions between crop production and photovoltaic power generation [172]. Zainali et al. further conducted multi-objective optimization of agrivoltaic systems from a water–energy–food perspective [173]. These studies mainly solve array–crop configuration problems; agricultural-equipment passage efficiency and operation-and-maintenance costs should also be fully incorporated into the unified objective function.

Coordinated optimization must also satisfy agricultural-demand, structural-safety, time-window, and resource-capacity constraints. Agricultural constraints include minimum photosynthetically active radiation or daily light integral at different growth stages, upper and lower soil–water-content limits, meteorological conditions for pesticide application, and irrigation and harvesting requirements; structural and safety constraints include module tilt and rotation-speed ranges, prohibition of conflicts between module motion and implement operation, minimum clearance, and safe distance from supports; time constraints include task duration and mutually exclusive passage occupation; and resource constraints include equipment operating capability, implement matching, battery state of charge, charging power, and site power balance.

Different solution methods should be used at different time scales. Seasonal- or daily-scale array–crop–equipment configuration is a multi-objective mixed optimization problem, for which NSGA-II, MOEA/D, ε-constraint methods, and mixed-integer or nonlinear programming can be used to obtain Pareto solution sets [172,173,174]. Module orientation, irrigation amount, and energy allocation are clearly dynamic and are suitable for model predictive control, which repeatedly optimizes according to updated irradiance, weather, and crop status and executes only the control amount for the current period. Research on dual-axis module tracking in agrivoltaic systems has used convex relaxation and shading approximation to transform rolling optimization of power generation and crop yield into a second-order cone programming problem [175]; combinations of standard tracking and backtracking can also form adjustable operating strategies among power generation, crop light receipt, and economic returns [176].

When multiple agricultural machines, inspection UAVs, and cleaning robots compete for tasks, passages, and energy-replenishment resources, mixed-integer programming, improved ant-colony algorithms, task auctions, contract-net protocols, cooperative-evolution algorithms, or coalition games can be used for task allocation and path scheduling. Improved ant-colony algorithms have been used for static and dynamic task allocation among multiple agricultural machines while simultaneously considering equipment capability and path cost [177], while cooperative discrete artificial-bee-colony algorithms can optimize multi-robot task allocation, operation sequence, and system completion time in intelligent farms [178]. Further studies have combined remote-sensing yield prediction with heterogeneous-machinery task allocation, harvesting-transport scheduling, and collision-free path planning, forming a technical chain from production-state prediction to coordinated multi-machine execution [179]. These multi-machine scheduling methods have not yet been directly validated beneath photovoltaic arrays; transfer will require additional scenario constraints such as module-motion envelopes, mutually exclusive array passages, minimum clearance, positioning confidence, and energy-replenishment time windows.

The site-planning layer generates seasonal- or daily-scale Pareto schemes and constraint boundaries according to planting plans, power-generation targets, and operation-and-maintenance budgets; the edge-coordination layer receives weather predictions, crop status, module status, and equipment status and uses rolling MPC and multi-agent scheduling to determine task priorities, operation time windows, and passage allocation; and the equipment-execution layer completes path tracking, variable-rate operations, and local obstacle avoidance and feeds back actual trajectories, energy consumption, and completed work. Figure 8 shows the operation and coordinated-regulation framework for intelligent agricultural equipment beneath photovoltaic arrays.

Operation instructions should contain at least task type, target area, implementation time, priority, travel path, and implementation parameters. Soil-water content, predicted evapotranspiration, and rainfall forecasts can be used to calculate zonal water deficit and generate irrigation areas, irrigation amounts, valve flow rates, and operation sequences; canopy temperature and water-stress indicators can be used to determine local irrigation or cooling tasks; pest, disease, and weed locations and canopy density can be used to adjust nozzle on/off status, application rate, and travel speed; and canopy height, support distance, and obstacle position can be used to set implement height, local path, vehicle speed, and safety boundaries.

When equipment approaches a support, positioning confidence decreases, or a module enters the implement operating envelope, travel speed should be reduced, the safety boundary enlarged, and operations paused if necessary. After task completion, actual trajectory, irrigation flow, application rate, energy consumption, and crop remeasurement results should be fed back immediately. Sensing and prediction results first need to be converted into crop or operating states, and then structured instructions should be generated in combination with agronomic thresholds, equipment capability, and array-safety constraints. The corresponding state judgments, trigger conditions, and agricultural-machinery operation instructions for different information are shown in

Table 10. Conversion relationships from agrivoltaic sensing information to agricultural machinery operation instructions.

Sensing and Prediction Information	Decision Variable	Generated Agricultural Machinery Instruction	Main Constraints and Feedback
Soil water content, predicted evapotranspiration, and rainfall forecast	Zonal water deficit and irrigation priority	Target area, irrigation amount, and operation sequence	Rainfall probability and soil infiltration capacity; feedback of actual flow and water content
Canopy temperature, VPD, and water-stress index	Heat-stress intensity and duration	Local irrigation and cooling operations	Crop growth stage and shading state; feedback of canopy temperature
Pest, disease, and weed locations and canopy density	Target confidence and damage level	Spraying location and application rate	Recognition confidence; feedback of pesticide application amount and reinspection results
Canopy height, support distance, and obstacle position	Clearance and collision risk	Implement height, local path, speed, and safety boundary	Module orientation and positioning error; feedback of trajectory deviation and minimum distance
PAR, shadow trajectory, and anomalous-growth areas	Anomaly duration and need for remeasurement	Inspection area, sampling location, and task priority	Instantaneous-shadow and mixed-pixel interference; feedback of remeasurement results

.

Table 10 shows that sensing variables and equipment actions do not have a simple one-to-one relationship; the same environmental state may generate different instructions depending on crop-growth stage, weather trend, equipment capability, and position within the array. For agrivoltaic systems, the most important issue is to reasonably coordinate the light-receiving relationship between modules and crops and to account for both energy output and crop-growth requirements by combining environmental and crop information with module operating status [32]. Therefore, when a model predicts that a certain under-panel area will remain under low PAR or poor light uniformity for a long period, that area can be marked as a key low-light-response zone, and robots can be given priority to enter the area to collect crop-growth information and soil-moisture data, in order to determine whether crops exhibit photosynthetic limitation or growth compensation [180]. Research on intelligent agricultural greenhouses shows that environmental models should be combined with control strategies and actuating equipment so that prediction information is incorporated into production regulation [181].

7.2. Precision Field Operations Beneath Photovoltaic Arrays

Within the above unified optimization model, precision irrigation, variable-rate application, and crop monitoring are agricultural-task subproblems. Their main objectives are to reduce crop stress and task delay and improve equipment operating efficiency, while changes in photovoltaic power generation, module orientation, and array clearance enter the optimization process as energy and safety constraints. Intelligent-irrigation studies show that irrigation decisions need to consider crop water demand, weather, and system operating state, and fixed thresholds are difficult to adapt to continuous changes in water demand [182]. Because photovoltaic arrays alter soil evaporation and crop transpiration demand, water deficit differs among zones, and uniform fixed irrigation can easily cause local over-wetting or water shortage. Mobile irrigation and fertigation equipment can replenish water and fertilizer or inspect pipelines in specific zones according to prediction results and measured data, thereby improving water and fertilizer use efficiency. In addition to spatiotemporal irrigation decisions, implementation-end parameters also affect operating stability under photovoltaic-energy conditions. Existing photovoltaic-irrigation studies have optimized system operation from nozzle structure and number configuration, indicating that irrigation actuators, pumping demand, and photovoltaic supply capability need to be matched coordinately [183,184].

Module shading changes canopy radiation, temperature, evapotranspiration, and soil water, thereby affecting the zones and periods for irrigation, fertilization, and crop-protection operations; occupation of passages and operating time by agricultural tasks in turn limits adjustable module orientations. During high-temperature and strong-radiation periods, protective shading can reduce evapotranspiration load and be coordinated with local irrigation; during light-sensitive stages such as flowering and grain filling, shading duration should be controlled, and irrigation should be prioritized to alleviate water–thermal stress. Deficit-irrigation experiments under agrivoltaic conditions likewise show that although shading can improve water productivity, excessive reduction of irrigation can still cause clear yield loss [185].

Array orientation plays both light–thermal regulation and operation-coordination roles in this process. Implementing backtracking during key maize growth stages and restoring solar tracking during other stages allows module orientation to be adjusted according to crop light-demand windows and can maximize crop returns [16]. Recent predictive control and active–reactive control further incorporate weather changes, crop light targets, and their accumulated deviations into tilt-angle decisions [175,186]. During field operations, module orientation should also obey task occupancy: before equipment enters an operating zone, relevant modules can be temporarily moved to an orientation suitable for operation; local tracking can be paused or restricted during equipment operation; and crop-oriented or power-generation-oriented operation can resume after task completion. The duration of orientation adjustment should be controlled so that resulting crop-light losses and lost power-generation opportunities remain within acceptable ranges.

The degree of coupling between different agricultural tasks and array orientation also differs. Irrigation can be implemented by zone and time and is suitable for rolling adjustment with shadow movement and water deficit; crop-protection operations are constrained by wind speed, temperature, and pesticide-application windows and therefore cannot be scheduled only according to passage availability; harvesting and transport have stronger farming-time constraints and require the array to remain stable during operation. When tasks conflict, non-delayable crop requirements and strict meteorological windows should first be identified, after which passage-occupancy conflicts can be resolved through task splitting, zone rotation, or short-term orientation adjustment.

Module orientation also changes the implementation conditions of field tasks, and there is no single coordination method between different tasks and module orientation. For delayable or divisible tasks, periods with more suitable module orientation and light–thermal conditions can be prioritized; for operations with strict farming-time requirements, module tilt can be adjusted temporarily, or local array tracking paused to obtain stable operating space. Power-generation losses, operation delays, and changes in crop light receipt caused by module-orientation changes should remain within acceptable ranges.

Overall, existing studies have separately demonstrated the feasibility of dynamic irrigation prescriptions and photovoltaic water-supply systems, but direct evidence for joint regulation of “light–thermal change module orientation–field task” remains limited. Future studies should compare, under different array configurations and crop-growth stages, the effects of adjusting irrigation alone, adjusting module orientation alone, and jointly adjusting both, while simultaneously evaluating irrigation uniformity, water-use efficiency, canopy temperature, effective radiation, operation delay, and power-generation loss. Only by establishing quantitative relationships among these indicators can it be determined under which circumstances module yielding, task delay, or zoned operations are respectively appropriate.

7.3. Coordination of Photovoltaic-Module Maintenance and Agricultural Tasks

In photovoltaic-module operation and maintenance, information on module temperature, hot spots, abnormal power-generation efficiency, and surface dust can be further converted into inspection and cleaning priorities. Deposits on module surfaces and local shading reduce power-generation efficiency, while dust generated by agricultural operations may further intensify contamination [187]. Coordinated scheduling of module inspection and under-panel crop monitoring can use UAVs to rapidly detect thermal anomalies and visible defects, followed by ground robots for close-range inspection and targeted maintenance. The post-operation power-recovery rate and fault-handling results should also be fed back to sensing and prediction models. Figure 9 shows a specific photovoltaic-maintenance process.

As shown in Figure 9, the system integrates UAV inspection, a ground cleaning robot, an edge-computing controller, and an energy-storage device into the photovoltaic-maintenance process. The UAV collects visible-light and thermal-infrared information along a preset route to identify module-temperature anomalies, while the ground robot plans cleaning paths according to dust distribution and inspection results. The red trajectory in the figure denotes the UAV inspection route, the blue dashed line denotes the cleaning robot operating path, circular markers denote thermal-anomaly regions, and the background color reflects dust-concentration distribution.

In the field of coordinated operation and maintenance for conventional photovoltaic sites, preliminary multi-machine coordinated-scheduling approaches have been developed. An integrated air–ground collaborative inspection system for photovoltaic sites has been implemented based on a data-interaction mechanism between aerial and ground robots [189]. De Benedetti et al. further constructed a two-level scheduling framework combining site-level task auctioning and robot local path planning, assigning inspection tasks to heterogeneous aerial and ground robots and incorporating multidimensional costs such as equipment energy consumption, battery aging, and mechanical wear into task decision-making [190]. For multi-robot collaboration in photovoltaic cleaning, overlapping-coalition game theory has been introduced to dynamically form collaborative units according to panel-surface contamination, robot operating capability, and task priority, balancing multiple objectives of operation-and-maintenance cost control and operating timeliness [191]. From the perspective of power-generation benefits, photovoltaic operation-and-maintenance tasks need to be dynamically scheduled according to temporal power-generation characteristics to avoid opportunity losses in power generation caused by shutdown operations during peak-output periods [192]. Overall, existing research on photovoltaic multi-machine coordination mainly focuses on conventional photovoltaic-site scenarios, and dedicated research on dynamic multi-robot coordination and adaptive task scheduling under the combined operating conditions of agrivoltaic systems remains lacking.

When the above methods are transferred to agrivoltaic systems, jointly solving task sequence, array-passage occupancy, energy/liquid replenishment nodes, and vehicle routes can help balance robot workloads and reduce operating costs [193]. Agricultural operations and module operation and maintenance should be arranged in a unified manner according to crop requirements, weather, equipment status, and passage occupancy. When daylight conditions are suitable, crop-phenotyping monitoring or module inspection can be conducted; crop-protection and irrigation operations should still be scheduled according to wind speed, temperature, and crop water demand. Module cleaning can be arranged during periods of low irradiance or outside peak agricultural-operation periods, and autonomous cleaning can be performed at night to reduce effects on normal power generation and other operations [194]. By uniformly coordinating agricultural requirements, photovoltaic operation and maintenance, passage resources, and equipment status, multiple types of tasks can be comprehensively optimized in terms of time, space, and energy use.

7.4. Photovoltaic-Powered Energy Supply for Agricultural Equipment

Rising fuel prices or falling battery costs will become important drivers of the commercialization of solar-powered equipment. Replacing internal-combustion equipment with solar-powered equipment can substantially reduce environmental external costs. The energy supply of photovoltaic-powered agricultural equipment should be designed according to equipment power, operation duration, and the energy-supply conditions of the agrivoltaic system; photovoltaic technology can replace part of conventional grid electricity and supply power to electric agricultural vehicles and autonomous robots on farms [195]. Comparison of multiple photovoltaic–electric-vehicle integration modes for oil-palm mechanization shows that onboard photovoltaic systems have good applicability, with actual tests showing that solar energy contributes approximately 20% to 30% of vehicle energy demand [196]. Redpath et al. further deployed battery electric vehicles charged by a 10 kWp photovoltaic array in remote agricultural areas, demonstrating that fixed photovoltaic arrays can provide a replenishment source independent of diesel fuel and weak power grids for light agricultural transport and field-operating equipment [197].

Onboard photovoltaics are more suitable as a supplementary energy source for low-power equipment, while batteries remain the core energy-storage unit for maintaining continuous equipment operation. Batteries can store surplus electricity from photovoltaic arrays and supply loads at night or on cloudy days; therefore, their performance directly affects the endurance of photovoltaic-powered agricultural machinery [198]. A photovoltaic-powered multifunctional agricultural robot can drive operating mechanisms using onboard photovoltaic and battery systems and maintain field operations according to solar output and battery charging/discharging status [199]. In addition, Ghobadpour et al. constructed an agricultural mobile robot with a hybrid photovoltaic–battery energy supply. Adding only the onboard photovoltaic system increased vehicle range by up to 5%. Although onboard photovoltaics cannot independently supply high-power traction loads, they can delay the decline in battery state of charge and reduce the frequency of returning to charging stations [200]. A photovoltaic agricultural transport vehicle stores solar electricity in lithium-ion batteries for transporting people, tools, and agricultural products, showing that photovoltaic-energy architectures have good feasibility in light agricultural equipment [201].

For high-power machinery such as electric tractors and harvesting equipment, onboard modules alone cannot meet long-duration operating requirements; therefore, fast charging and battery-swapping facilities need to be constructed based on agrivoltaic fields. Electricity generated by photovoltaic arrays can be delivered to field equipment through fixed power-supply networks or stored first and then used for mobile operations [202]. Wallander et al. compared field charging and field battery swapping in fully electric agricultural-machinery systems and found that the operating effectiveness of battery swapping was higher than that of direct field charging, reducing equipment waiting time and improving continuous operating capability during peak farming periods [203]. The agricultural photovoltaic charging facility proposed by Goh et al. further integrates agricultural production, photovoltaic power generation, and electric-vehicle charging within the same land unit and can provide a centralized equipment-energy replenishment node for agrivoltaic fields [204]. In addition to mobile equipment, electrified environmental-control devices such as greenhouse thermoelectric dehumidifiers also constitute adjustable loads in agrivoltaic systems [205] and can be included in coordinated scheduling together with irrigation pumps, supplemental lighting, and equipment-charging tasks. Studies of stand-alone multi-energy complementary agricultural irrigation systems show that water-pump power demand and renewable-energy capacities such as photovoltaics need to be jointly configured [206]; energy configuration in agrivoltaic systems should simultaneously consider periodic agricultural loads such as irrigation.

The system should continuously obtain equipment task-power demand and operation deadlines and use this information to determine battery-swapping timing and task allocation. For low-power tasks, onboard photovoltaic auxiliary charging or automatic recharging during task intervals can be prioritized; for high-power tasks such as harvesting and heavy-load transport, centralized charging stations, replaceable batteries, or mobile energy-supply equipment should be configured. Charging and high-energy-consumption operations should be arranged when photovoltaic power generation is sufficient, while energy storage and the grid should jointly supply power when photovoltaic generation is insufficient or agricultural operations are concentrated, thereby reducing charging waiting time, increasing local consumption of photovoltaic electricity, and improving the energy autonomy of agricultural equipment. Although solar-powered equipment and agricultural robots have advantages in energy saving and emission reduction, high initial costs limit commercialization. As the prices of sensors, control units, and photovoltaic modules continue to decline, more products are expected to enter large-scale agricultural production.

8. Discussion

This paper extends the research boundary from traditional quantification of environmental characteristics and prediction of system performance to the levels of task decision-making and field-operation execution, focusing on analysis of the internal mechanism by which zoned environmental states are transformed into implementable operation instructions under the multiple coupling of crop-growth requirements, photovoltaic-module orientation, agricultural-machinery operating capability, and photovoltaic operation-and-maintenance constraints.

At present, this transformation process is continuously constrained by fragmentation among multiple spatiotemporal scales. Fixed nodes usually output continuous data at the measurement-point scale, remote-sensing results reflect canopy differences at flight or pixel scale, radiation models can generate minute-scale gridded states, and crop models often update biomass at a daily scale, while agricultural robots need to complete positioning, obstacle avoidance, and variable-rate operations at a second-scale. Even when each type of information is individually accurate, it cannot be used directly for coordinated control without spatial correspondence, temporal aggregation, and uncertainty descriptions. Therefore, scale conversion itself should be regarded as an independent model layer connecting prediction and control, reconstructing raw observations into system states with unified spatial boundaries, validity periods, change processes, and confidence levels. The decision layer should first use these states to determine the feasible domain jointly allowed by crops and equipment and then trade off energy and operating objectives within that domain. Prediction accuracy has practical control value only when the integrity of information across scales can be maintained.

Based on the above understanding, the core scientific gap identified in this paper is the current lack of a state system capable of uniformly describing array status, zoned environments, crop requirements, equipment capabilities, and operation-and-maintenance tasks, together with the lack of a verifiable conversion mechanism from prediction results to operation instructions. This gap makes experimental results difficult to transfer, models difficult to couple, and technical performance difficult to directly transform into agricultural and economic benefits at the system level. It should be noted that the studies included in this paper show clear heterogeneity in experimental scale, indicator definitions, and validation periods, and some under-panel sensing, navigation, and multi-machine coordination schemes are still mainly borrowed from conventional farmland, greenhouse, or ground-mounted photovoltaic scenarios. Therefore, the conclusions of this paper are mainly intended to extract technical consensus, applicability boundaries, and research gaps, rather than to provide a unified performance ranking of specific configurations or algorithms.

9. Conclusions and Development Outlook

Agrivoltaics couples crop production and photovoltaic power generation within the same land unit, providing an important pathway for alleviating conflicts among food production, energy development, and land occupation. A relatively clear technical chain can be identified from the existing research. First, array configuration is a fundamental constraint determining agricultural-equipment adaptability. Machinery dimensions, operating width, and turning requirements should be considered simultaneously during array planning. Second, when array structures interfere with satellite signals and the visual environment, multi-sensor fusion is the main pathway for achieving continuous navigation under GNSS-constrained conditions. GNSS, LiDAR, vision, and IMU provide complementary information, but the positioning accuracy and long-term stability of related fusion methods in real agrivoltaic scenarios still require validation. On this basis, stable equipment operation also depends on accurate sensing and prediction of the heterogeneous environment beneath arrays. The key to sensing and prediction in photovoltaic scenarios is correction of systematic biases introduced by array structures; mixed pixels, structured shadows, module orientation, and related factors need to be incorporated into parameter inversion and model correction, while preserving spatial zoning and uncertainty information, so as to further support operation decisions such as variable-rate irrigation, crop protection, and inspection. Ultimately, array design, positioning and navigation, and environmental sensing all need to enter a unified operational-decision framework. Coordinated agrivoltaic operation is essentially a multi-objective optimization problem under safety and agricultural-demand constraints. The system should prioritize personnel and equipment safety and irreversible crop requirements and use hierarchical control to coordinate power-generation benefits, crop yield, operating efficiency, energy consumption, and operation-and-maintenance cost.

The most prominent current bottleneck is that the different links still lack a unified coordination mechanism: long-term real-scene observation data are insufficient, cross-scenario generalization capability of models and algorithms is limited, there are no general adaptation standards for array deployment and implementation operations, and production, module adjustment, site passage, energy supply, and equipment maintenance also lack consistent state interfaces and constraint systems. To provide clearer guidance for future research, the following core research priorities are proposed:

(1) Existing research has not yet established matching criteria among array row spacing, minimum clearance, equipment width and height, operating-mechanism overhang, positioning error, and safety distance. Future research should establish calculation methods for effective operating width and safety buffer zones, quantify operating losses caused by support foundations, module movement envelopes, and headland space, and formulate array-design requirements adapted to agricultural machinery.

(2) Most existing prediction studies are still based on data from a single site or specific climate region, and their generalization capability across regions and configurations remains to be verified. Future research can use transfer-learning frameworks to integrate agrivoltaic data from different geographical regions, such as arid and high-altitude areas, and develop dynamic estimation models with climate adaptability. At the same time, reinforcement-learning algorithms can be combined to construct a photovoltaic-module-layout–environment-response feedback mechanism and achieve closed-loop optimization between estimation models and agricultural decision-making systems.

(3) Positioning and sensing beneath photovoltaic arrays still lack scenario-specific evidence. Most existing multi-sensor fusion, SLAM, and visual-recognition methods come from conventional farmland, orchards, or greenhouses, and the effects of metal-support multipath, repetitive array structures, and alternating bright and dark conditions on GNSS, LiDAR, and vision have not yet been sufficiently quantified. Future research should conduct zoned tests beneath arrays, output positioning confidence according to observations, and automatically reduce speed when positioning degrades to ensure the safety of continuous operations.