Modelling, Simulation and Performance Analysis of Floating Photovoltaic Systems—A Systematic Review and Meta-Analysis

Abstract

Research into floating photovoltaics (FPV) has seen a significant increase in recent years. Still, the observed outputs are poorly quantified, isolated, and occasionally contradictory, with reported cooling-induced efficiency increases varying widely across sources. To address the need for consensus in the field, a systematic literature review (SLR) and meta-analysis were conducted using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework to provide a comprehensive overview of the current state-of-the-art in FPV systems. 3751 articles were identified through Boolean queries on three databases (Scopus, Web of Science, and Google Scholar). Using Python programming to ensure objectivity and replicability, the dataset was screened to 109 publications (subject to a manual, full-text review) relating strictly to modelling, simulation, and performance analysis of FPV systems with regard to the observed effect of reduced operating temperature. Focusing on these areas, this study provides a fundamental understanding of the temperature-based performance, as well as insights into the operation and simulation of FPV systems. Consistent temperature reductions were observed between ground-mounted and floating systems. Experimental data on FPV temperature were subject to a regression analysis, and the resulting equation was found to correspond well to a reported relation in the literature. The article concludes with a set of informed research directions to underpin the further development and implementation of FPV technology.

1. Introduction

1.1. Floating Photovoltaics: A Brief Overview

As global energy demands continue to rise [1], there remains a need for the consistent supply and generation of clean energy. Solar energy is a readily abundant source of energy [2], and the appropriate utilisation of this resource is key to meeting these demands [3,4]. Photovoltaic (PV) technology is the most dominant means of harnessing this energy from the sun and has been in mainstream commercial application for well over a decade [5]. Adoption has risen over the years, primarily due to relatively low production costs and high resource availability, as the Levelized Cost of Energy of solar PV electricity was noted to drop by 90% between 2010 and 2023 to as low as 0.044 USD/kWh [6,7].

Floating Photovoltaics (FPV) are a relatively new iteration of PV technology [8] which involves the installation of PV modules on bodies of water; indeed, with the increasing competition for land resources [9,10], FPV systems present themselves as a key component of the global energy landscape in the years to come [11]. The first operational FPV plant was an experimental 20 kW system designed and implemented in Aichi, Japan, by researchers at the National Institute of Advanced Industrial Science and Technology in 2007 [8,12]. The goal of the experiment was to assess the cooling effect observed between a ground-mounted and water-cooled system. Since then, several FPV plants have come into operation around the world, and global installed capacity reached 7.7 GW as of 2023 [13]. FPV application and implementation come in various forms, examples being the plants in Far Nientes and Gundlach Bunchu Wineries in California, which were deployed in 2008, shortly after the development of the plant in Aichi, to provide electricity for the farm [14,15]. Another contextual application exists in aquavoltaics, which is the combination of photovoltaic implementation with aquacultural activities [16,17]. Also prominent in FPV application is the integration with other renewable energy sources, with particular interest in hydropower [18], offshore wind [19], pumped hydroelectric storage plants [20], among others.

A generic FPV plant comprises PV modules and a support system, a floating structure, a mooring system, and an anchor system [21]. The PV modules are at the top of the configuration, held together by a frame (usually aluminium), which is located above water by floats or pontoons—usually made of HDPE (high-density polyethylene) or other lightweight materials. The above-water structure is kept in place by a mooring and anchoring system, which usually involves a system of metallic ropes to keep the floating structure at a stable mean distance from an anchor fixed into the seabed. Finally, electrical cabling systems are required to convey the generated electricity to ashore (see Figure 1). The actual design and configuration of these components differ widely across installations, but this general structure is usually consistent across different FPV systems.

1.2. Study Background

As the temperature of a photovoltaic cell rises, a slight increase in the short-circuit current is observed due to a narrowing of the band gap; conversely, however, the open-circuit voltage experiences a drop due to changes in the intrinsic carrier concentration within the semiconductor [22,23]. The net effect of this is a reduction in power output/efficiency. This negative correlation between temperature and cell efficiency is well-established and parameterised by the temperature coefficient (β) of a PV module, which denotes the responsiveness of a module’s performance to changes in temperature [24,25]. In the case of silicon cells, a temperature coefficient between 0.4 and 0.5%/K implies that for every unit degree change in module temperature, the module efficiency changes by about 0.5% [25]. By virtue of this relationship, the temperature-specific performance of a PV module can be estimated via the relation in Equation (1) where $\eta$ is the efficiency at a given temperature $T_c$, ${\eta }_{stc}$ is the module efficiency at standard test conditions, and $T_{ref}$ is the reference/base module temperature [26].

$$\eta ={\eta }_{stc}\left\{1-{\beta }_{ref}\left(T_c-T_{ref}\right)\right\}$$

(1)

As the microclimate around aquatic environments is typically cooler than terrestrial areas, a consequence of FPV deployment is a reduction in temperature compared to GPV (ground-mounted PV) systems, which then translates into increased performance for the FPVs relative to GPVs [27]. While this is theoretically expected and indeed justified, the extent to which this temperature reduction and expected performance increase are observed is still uncertain [13]. Attempts to experimentally monitor and accurately quantify the effects of temperature changes on FPV performance usually involve documenting ambient environmental conditions around the test-site, as well as module performance and temperature, to arrive at a reasonable correlation between these variables [28,29]. Several studies have confirmed the expected characteristics of lower temperature and increased performance for FPVs, but the reported results vary considerably across the research landscape [27].

Regarding the simulated performance of FPV systems, several approaches have been developed over the years (see Table 1). Initially, models developed for GPV systems were applied to FPV systems, but they fell short in that they generally failed to account for the complexity of the aquatic microclimate, thus resulting in limited accuracy [30]. The Sandia model developed by King et al. [31] (Table 1, row 1), and the Faiman model [32] (Table 1, row 3) are two of the most commonly used models and are fundamentally applied in popular modelling software, SAM [33] and PVSyst [34], respectively. Notably, the Faiman Model introduced the concept of heat-loss coefficients, now commonly referred to as U-values, quantifying how effectively heat is exchanged between a PV system and the surrounding environment [8,32]. U-values can be represented as a single term $U\left(\mathrm{W}/{\mathrm{m}}^2\mathrm{K}\right)$, or, for increased parameterisation, it can be split into two terms (Equation (2)): the wind-independent (or constant) term ${‘U}_c\left(\mathrm{W}/{\mathrm{m}}^2\mathrm{K}\right)’$ and the wind-dependent term ‘$U_w\left(\mathrm{W}\mathrm{s}/{\mathrm{m}}^3\mathrm{K}\right)$’, which scales with wind velocity $V_w$. The U-value of a system is a numerical representation of its cooling rate, and simply put, the higher a U-value, the more effective a system is at losing heat [32].

$$U=U_c+U_w·V_{air}$$

(2)

Liu et al. [35] attempted to organise research outputs in relation to U-values in a study conducted by researchers at the Solar Energy Research Institute of Singapore (SERIS) of the National University of Singapore, thereby classifying FPV structures at the SERIS testbed according to the sizes of the floating structure with respect to the amount of water coverage the structure has; the concept being that a higher water coverage or “footprint” would lead to reduced thermal contact between the PV module and the water surface, and also reduce the air-flow beneath the modules, factors which both significantly affect the heat transfer characteristics of an FPV system [36]. Based on this study, FPV structures were categorised into free-standing (modules high above the water), small (modules close to the water with minor water surface coverage), and large (modules close to the water with larger water surface coverage) footprint configurations. These classifications have since been largely adopted by the research community.

Furthermore, regarding module temperature modelling, a clearer picture was provided in the study by Kamuyu et al. [37], where the researchers derived a regression-based model from experimental empirical data. The model factored variables including radiation, ambient temperature, wind speed, and water temperature, as inputs to predict the module temperature (Table 1, row 11). This approach was largely successful and prompted the application of similar methodologies to other studies [36]. However, a direct consequence of this regression-based approach is that the models are heavily dependent on the conditions of the experiment, and applying them to other experimental datasets does not always yield the required results [38]. Recent studies are modifying traditional FPV modelling approaches by exploring various advanced methods, including transient analysis, solid state models, dynamic simulations, thermal resistance techniques, and CFD (computational fluid dynamics) [39].

Table 1. FPV Temperature modelling equations from the literature.

n	Equation	Notes	Ref
1	$T_{\mathrm{m}}=\mathrm{G}·e^{a+b·V_{ws}}+T_a$		Sandia [31]
2	$T_{\mathrm{c}}=T_m+{\displaystyle \frac{G_{POA}}{{\mathrm{G}}_{\mathrm{r}\mathrm{e}\mathrm{f}}}}\left(T_c-T_{ref}\right)$	Intra-Module temperature as a function of back-of-module temperature as given above. $T_c-T_{ref}$ is suggested to be 2 K–3 K for open rack modules.	Sandia [31]
3	$T_c=T_a+{\displaystyle \frac{G_{POA}}{U_c+U_w{V}_w}}\equiv T_a+{\displaystyle \frac{{\alpha G}_{POA}(1-\eta )}{U_c+U_w{V}_w}}$		Faiman [32]
4	$T_c=T_a+{\displaystyle \frac{G_{POA}}{{\mathrm{G}}_{\mathrm{N}\mathrm{O}\mathrm{C}\mathrm{T}}}}\left(T_{NOCT}-T_a\right)\left({\displaystyle \frac{9.5}{5.7+\left(3.8\times V_w\right)}}\right)\left(1-{\displaystyle \frac{{\eta }_{NOCT}}{\tau \alpha }}\right)$		Duffie & Beckman [40]
5	$T_m=T_a+\left({\displaystyle \frac{0.32}{8.91+\left(2.0\times V_w\right)}}\right)·G$		Skoplaki [41]
6	$T_m={\displaystyle \frac{T_a.U_{pv}+G·\left(0.81-{\eta }_{stc}·(1-{\beta }_{STC}·T_{STC}\right))}{U_{pv}+G·{\beta }_{STC}·T_{STC}}}$	$U_{pv}=26.6+2.3V_w or 24.1+2.9U_{pv}=26.6+2.3V_w$	Mattei [42]
7	$T_m=T_a+\mathrm{k}·\mathrm{G}$	A common value for k is 0.025 K·m2/W	Ross Model [43]
8	$T_m=T_a+\mathrm{G}·e^{\left({3.473-0.0594·V}_w\right)}$	Similar to the Sandia model [31] but assigns static, system-independent coefficients.	Kurtz Model [44]
9	$T_m=3.81+\left(1.31T_a\right)+\left(0.0282G\right)-\left(1.65V_w\right)$		Risser and Fuentes [45]
10	$T_m=\mathrm{a}{T}_a+\mathrm{b}\mathrm{G}{e}^{{cV}_w}=\mathrm{a}{T}_a+\mathrm{G}{e}^{{\ln \left(b\right)-cV}_w}$		Keddouda [46]
11	$T_{fpv}=2.0458+0.9458T_a+0.021G-1.2376V_w$		Kamuyu [37]
12	$T_{fpv}=1.8081+0.9282T_a+0.021G-1.2210V_w+0.0246T_w$	Kamuyu model [37] with the inclusion of water temperature.	Kamuyu [37]
13	$T_{fpv}=0.943T_a+0.0028\mathrm{G}-1.528V_w+4.3$		Tamizhmani [47]
14	$T_{fpv}=0.943T_a+0.0195\mathrm{G}-1.528V_w+0.3529$		Muzathik [48]
15	$T_{fpv}=1.2645T_a+0.0128\mathrm{G}-0.0873T_w+13.2554$		Hayibo [49]

T_c = cell temperature, T_m = GPV module temperature, T_fpv = FPV cell temperature, T_a = ambient temperature, T_w = water temperature, G = solar irradiance, V_w = wind speed, U_c = constant/wind-independent heat-loss coefficient, U_w = wind-dependent heat-loss coefficient, POA = Plane of array, NOCT = Nominal operating condition temperature, STC = standard test conditions.

Table 1. FPV Temperature modelling equations from the literature.

n	Equation	Notes	Ref
1	$T_{\mathrm{m}}=\mathrm{G}·e^{a+b·V_{ws}}+T_a$		Sandia [31]
2	$T_{\mathrm{c}}=T_m+{\displaystyle \frac{G_{POA}}{{\mathrm{G}}_{\mathrm{r}\mathrm{e}\mathrm{f}}}}\left(T_c-T_{ref}\right)$	Intra-Module temperature as a function of back-of-module temperature as given above. $T_c-T_{ref}$ is suggested to be 2 K–3 K for open rack modules.	Sandia [31]
3	$T_c=T_a+{\displaystyle \frac{G_{POA}}{U_c+U_w{V}_w}}\equiv T_a+{\displaystyle \frac{{\alpha G}_{POA}(1-\eta )}{U_c+U_w{V}_w}}$		Faiman [32]
4	$T_c=T_a+{\displaystyle \frac{G_{POA}}{{\mathrm{G}}_{\mathrm{N}\mathrm{O}\mathrm{C}\mathrm{T}}}}\left(T_{NOCT}-T_a\right)\left({\displaystyle \frac{9.5}{5.7+\left(3.8\times V_w\right)}}\right)\left(1-{\displaystyle \frac{{\eta }_{NOCT}}{\tau \alpha }}\right)$		Duffie & Beckman [40]
5	$T_m=T_a+\left({\displaystyle \frac{0.32}{8.91+\left(2.0\times V_w\right)}}\right)·G$		Skoplaki [41]
6	$T_m={\displaystyle \frac{T_a.U_{pv}+G·\left(0.81-{\eta }_{stc}·(1-{\beta }_{STC}·T_{STC}\right))}{U_{pv}+G·{\beta }_{STC}·T_{STC}}}$	$U_{pv}=26.6+2.3V_w or 24.1+2.9U_{pv}=26.6+2.3V_w$	Mattei [42]
7	$T_m=T_a+\mathrm{k}·\mathrm{G}$	A common value for k is 0.025 K·m2/W	Ross Model [43]
8	$T_m=T_a+\mathrm{G}·e^{\left({3.473-0.0594·V}_w\right)}$	Similar to the Sandia model [31] but assigns static, system-independent coefficients.	Kurtz Model [44]
9	$T_m=3.81+\left(1.31T_a\right)+\left(0.0282G\right)-\left(1.65V_w\right)$		Risser and Fuentes [45]
10	$T_m=\mathrm{a}{T}_a+\mathrm{b}\mathrm{G}{e}^{{cV}_w}=\mathrm{a}{T}_a+\mathrm{G}{e}^{{\ln \left(b\right)-cV}_w}$		Keddouda [46]
11	$T_{fpv}=2.0458+0.9458T_a+0.021G-1.2376V_w$		Kamuyu [37]
12	$T_{fpv}=1.8081+0.9282T_a+0.021G-1.2210V_w+0.0246T_w$	Kamuyu model [37] with the inclusion of water temperature.	Kamuyu [37]
13	$T_{fpv}=0.943T_a+0.0028\mathrm{G}-1.528V_w+4.3$		Tamizhmani [47]
14	$T_{fpv}=0.943T_a+0.0195\mathrm{G}-1.528V_w+0.3529$		Muzathik [48]
15	$T_{fpv}=1.2645T_a+0.0128\mathrm{G}-0.0873T_w+13.2554$		Hayibo [49]

T_c = cell temperature, T_m = GPV module temperature, T_fpv = FPV cell temperature, T_a = ambient temperature, T_w = water temperature, G = solar irradiance, V_w = wind speed, U_c = constant/wind-independent heat-loss coefficient, U_w = wind-dependent heat-loss coefficient, POA = Plane of array, NOCT = Nominal operating condition temperature, STC = standard test conditions.

1.3. The Need for a Systematic Literature Review

It has been suggested that the execution of a systematic literature review (SLR) should mandatorily precede any new research activity [50]. An SLR, if conducted properly, not only effectively explores a domain, but also enables increased levels of objectivity that are inherently difficult to achieve in regular narrative reviews. SLRs are becoming commonplace in academic research but are yet to be fully integrated into most areas [51]. At the time of writing (mid-2025), only three SLRs were found to exist in the field of FPVs, all of which were published within the last year. Two of the three were specifically investigating the considerations and criteria applied during site selection for prospective FPV plants. Forester et al. [52] investigated the nature and application of the exclusionary criteria for FPV site selection, while Ali et al. [53] conducted a similar assessment, but with a focus on co-located Offshore Wind Farms and FPV plants. The third publication, by Goncalves et al. [54], involved an SLR of the reported environmental impacts on water bodies caused by FPV deployment.

Despite the relative nascence of FPV technology, the constant evolution of the research landscape creates a need to stay abreast of the state-of-the-art. To achieve this objective, a systematic literature review has been conducted with a specific focus on the modelling, simulation, and performance analysis of FPV systems. Moreover, FPVs represent an emerging technology associated with greater levels of renewable energy adoption; however, there is a distinct lack of clarity on the planned and realised performance benefits of these systems. Therefore, there is an urgent need to capture data on the status of FPV performance and related engineering considerations to inform future research endeavours as well as the industrial adoption of FPV systems. In this context, modelling and simulation studies provide an appropriate lens to enable the capture of this data according to related performance parameters of FPVs; conducting an SLR, but crucially also incorporating meta-analysis of identified data, directly addresses this requirement.

The current study differs from existing reviews, firstly, by employing a formalised and systematic approach to assessing the corpus, which results in a higher degree of objectivity than is typically seen in narrative reviews. Secondly, the study is bolstered by the execution of a meta-analysis on quantitative data regarding the thermal performance of FPV systems, which also constitutes a novelty in the field of FPVs and provides much-needed synthesis to the research outputs in the literature. The findings, therefore, from the study can be applied as a platform on which future research is based.

The remaining sections of the study are organised as follows: Section 2 provides a detailed account of the methodological approach employed in the study. Section 3 presents the results of the Systematic Review, synthesising the outputs in relation to the study’s objectives. The key findings, limitations, and future research areas based on the results of the study are discussed in Section 4. Finally, the study is concluded in Section 5.

2. Methodology

There are several viable approaches to an SLR, as described by Xiao and Watson [55]. The method of choice is a function of the suitability for the dataset and applicability to its contents. In the current work, a textual narrative synthesis and meta-analysis have been selected for the study. Detailed information and guidance on executing these forms of analyses are provided in other studies [56,57] and will not be covered in this review. Adopting a meta-analysis does have its drawbacks, as an attempt at the joint analysis of the outputs from a wide array of studies inevitably carries the risk of incompatible data as different types of datasets are synthesised [58]. However, measures have been taken to homogenise and categorise the dataset as much as possible (details provided in Section 3.3) to reduce the likelihood of this occurrence.

The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework, originally developed for use in the clinical sciences [59], is widely documented as the general standard for conducting SLRs across. The framework details a 27-item checklist (See Appendix A.2) and a four-phase flowchart to aid researchers in the organisation and presentation of systematic reviews. Furthermore, the research methodology, in line with the PRISMA framework, is clarified and highlighted in the coming subsections. Additionally, the work of Kitchenham et al. provided the necessary guidelines within which the review was conducted [60,61].

2.1. Research Protocol

Before commencement of the review, attempts were made to draft the intended protocol to be taken for the SLR, in line with Kitchenham [60]. As this was an iterative process, the proposed protocol underwent refinements throughout the review, and the finalised version is provided in

Table 2. Summary of SLR protocol.

Item	Description
Background (Rationale)	Provide a clearer picture of the state of the art in the temperature-based performance, modelling, and simulation of FPVs
Research Questions	Technical performance of FPVs Comparison between FPVs and GPVs FPV modelling techniques
Search Strategy	DB queries (Scopus, Web of Science, Google Scholar)
Selection Criteria	Python (version 3.13.3)-aided text-based screening Manual Screening
Q.A. Procedures	Manual review
Data Extraction Strategy	Quantitative and textual data across multiple data categories (Textual data alone when quantitative data is not available)
Synthesis	Meta-Analysis, Thematic Analysis
Timeline/Timeframe	1 May 2025 to 31 July 2025

. This protocol was not formally registered in any repository and exists only within the context of this study. After a cursory review of the background literature and with the authors’ existing knowledge in the field, the following research questions were formulated:

• What are the technical performance characteristics of floating photovoltaic systems with respect to temperature?
• What observable differences exist between the performance of floating photovoltaic systems and conventional ground-mounted or roof-mounted systems?
• What is the state of the art in floating photovoltaic system performance modelling?
• What are the future research areas for floating photovoltaic systems?

2.2. Record Identification

The literature identification process involved an initial query of three databases (namely Scopus, Web of Science, and Google Scholar) for publications in the field of floating photovoltaics. Record extraction from Scopus and Web of Science was executed via manual visits to the respective webpages, while Google Scholar records were extracted using Harzing’s “Publish or Perish” (version 8.17.4863) [62], which is a third-party software. The data from the three sources underwent individual ‘cleaning’ processes and were combined into a single file with the use of Python (version 3.13.3) programming. As a result, 3751 records were extracted in total (see

Table 3. Database queries and output.

Database	Query	Records	Date
Scopus	TITLE (“FLOATOVOLTAIC” OR “FLOATOVOLTAICS” OR “FLOATING SOLAR” OR “FLOATING PHOTOVOLTAIC” OR “FLOATING PHOTOVOLTAICS” OR “PHOTOVOLTAIC FLOATING” OR “FLOATING PV”) OR KEY (“FLOATOVOLTAIC” OR “FLOATOVOLTAICS” OR “FLOATING SOLAR” OR “FLOATING PHOTOVOLTAIC” OR “FLOATING PHOTOVOLTAICS” OR “PHOTOVOLTAIC FLOATING” OR “FLOATING PV”)	1183	30 May 2025
Google Scholar 1	FLOATOVOLTAIC OR FLOATOVOLTAICS OR “FLOATING SOLAR” OR “FLOATING PHOTOVOLTAIC” OR “FLOATING PHOTOVOLTAICS” OR “PHOTOVOLTAIC FLOATING” OR “FLOATING PV”	1632	30 May 2025
Web of Science	((((((TI = (“FLOATOVOLTAIC”)) OR TI = (“FLOATOVOLTAICS”)) OR TI = (“FLOATING SOLAR”)) OR TI = (“FLOATING PHOTOVOLTAIC”)) OR TI = (“FLOATING PHOTOVOLTAICS”)) OR TI = (“PHOTOVOLTAIC FLOATING”)) OR TI = (“FLOATING PV”) OR ((((((AK = (“FLOATOVOLTAIC”)) OR AK = (“FLOATOVOLTAICS”)) OR AK = (“FLOATING SOLAR”)) OR AK = (“FLOATING PHOTOVOLTAIC”)) OR AK = (“FLOATING PHOTOVOLTAICS”)) OR AK = (“PHOTOVOLTAIC FLOATING”)) OR AK = (“FLOATING PV”) OR ((((((KP = (“FLOATOVOLTAIC”)) OR KP = (“FLOATOVOLTAICS”)) OR KP = (“FLOATING SOLAR”)) OR KP = (“FLOATING PHOTOVOLTAIC”)) OR KP = (“FLOATING PHOTOVOLTAICS”)) OR KP = (“PHOTOVOLTAIC FLOATING”)) OR KP = (“FLOATING PV”)	936	30 May 2025
TOTAL		3751

¹ Google Scholar records were extracted using Publish-or-Perish version 8.17.4863 [62].

).

2.3. Record Screening

To maintain maximum objectivity and minimise the risk of bias in the study, the majority of the screening and exclusion processes were executed directly using record metadata and via the Python programming language. The first stage involved record deduplication, which reduced the dataset to 1905 unique, individual records.

With the dataset ready for use, the next step was to define the screening criteria. This was a multi-step process that initially involved record screening based on record type, then a secondary screening process based on titular categorisation. The types and volume of records screened out in the first stage are detailed in. The objective of this stage was to limit the dataset to recently published peer-reviewed journal articles. To achieve this, with 2020 as the recency cut-off point, conference papers, preprints, book chapters, and publications solely detailing information on grants were understandably excluded. Publications that were found exclusively in private repositories (these included student theses), as well as reviews, were also excluded on this basis. Finally, records that were not published in English were also removed from the dataset. This reduced the volume of records in the dataset by about 50% to 962 (see

Table 4. Preliminary screening results.

Record Type	Non-English	Pre-2020	Repository Papers	Preprints	Review Papers	Book Chapters	Conference Papers	Chapters	Grants	Total
n	32	318	54	26	79	38	365	14	17	943

).

The topic of excluding records from the dataset based on their ‘quality’ was considered during the screening process, but this ultimately was not applied, as it has been argued that quality-based exclusion (usually on the source of the publication) is not necessarily a justified approach [63]. Additionally, even if quality-based exclusion were a set criterion, it would be a difficult and borderline unfeasible task to come up with an objective set of criteria for this exercise. Furthermore, an overarching goal of the study is to provide a holistic view of the research landscape in the specified areas; therefore, it would be better served if the screening process were oriented towards inclusivity, rather than exclusivity, where reasonable.

The secondary stage of record screening was based on the category/study area of each record. A word-based textual occurrence program was run on the titles of all the publications in the dataset to identify the frequently occurring words in the title column. Using this information, categories were formulated to classify records based on the words in their titles, where each category corresponded to the occurrence of certain keywords in a record’s title. For example, the “water_purification” category corresponds to the keywords “evaporator”, “evaporators”, “desalination”, and “still”. The developed categories, and an example of the categorisation framework in action, are depicted in Figure 2. Following the categorisation, in accordance with the goal of the study, and the research questions, records under four categories were selected for further assessment; these were Comparative Assessment, Cooling and Temperature, Modelling and Simulation, and Performance Analysis, the resulting dataset of which had 258 records.

2.4. Record Exclusion/Inclusion

Before commencement of the full-text review, a manual review of the abstracts and conclusions of the remaining records was executed. This was firstly to see whether any records had falsely bypassed the primary screening stage, and an additional 50 records were screened out based on their validity, leaving 208 records in the dataset. The second purpose of the manual abstract review was to begin the exclusion of records based on relevance to the scope of the study. To do this with a respectable degree of objectivity, a set of exclusion/inclusion criteria was devised. The development of the criteria involved several cursory reviews of the contents of the dataset, in a bid to harmonise the available data with the research objectives. Ultimately, the output of this process was a set of six criteria (see

Table 5. Inclusion/exclusion criteria.

Group	Criteria
A	C1: Temperature-dependent FPV Performance C2: FPV heat-transfer systems
B	C3: Determine FPV characteristics C4: Compare FPV and GPV characteristics C5: Develop modelling and simulation methods C6: Validate modelling and simulation methods

), split into two groups (two in group A, four in group B), and the requirement for a record’s inclusion in the final dataset was that it met at least one criterion in each group (i.e., C1 OR C2 AND C2, C3, C4, OR C5).

For most of the 208 records, an abstract review was sufficient to correctly assess the study’s relevance, and for those with any ambiguity, a cursory review of the publication was required. After this stage, 70 full-length articles remained in the final dataset, and while the chosen approach was largely successful at fulfilling the intended purpose, there were a few special cases that warranted exception. These exceptions were largely made up of Case Studies and Feasibility Studies. By nature, they typically failed to meet any of the T1 criteria (typically due to a lack of focus/data on module temperature) as they were generally oriented towards general performance assessments. However, the decision was made to include them in the final dataset, as their output contributed to adjacent objectives of the study; to contribute to the understanding of FPV model application (in the case of feasibility studies), and larger-scale FPV system performance (for case studies). With these categories included, a final dataset of 109 records was put forward for full-text review.

2.5. Record Synthesis and Data Collection

A cornerstone of an SLR is a standard format for data extraction across multiple categories [55,64]. In a similar approach to [65], a codebook has been developed for this effect. As the dataset is inherently heterogeneous, despite the numerous filtering processes, record coding was necessary to facilitate both the textual narrative synthesis and the meta-analysis, as the outputs require some level of homogeneity for any comparisons or analyses to take place. The codebook employs the criteria detailed in Table 5, as well as other categorical divisions, and is provided in

Table 6. Study synthesis codebook.

Study Area (A)	Study Focus (F)	Study Method (M)	Category (C) *
A1: Temperature-dependent FPV performance	F3: Determine FPV characteristics	M1: Experimentation	CC_Exp
A2: FPV heat-transfer systems	F4: Compare FPV and GPV characteristics	M2: Modelling/simulation	CC_Mod
	F5: Develop modelling and simulation methods		FPV_Exp
	F6: Validate modelling and simulation methods		FPV_Modval
			CT
			U-values
			FC_Study

* Additional columns were added to the codebook for the FC_Study textual narrative synthesis and are discussed in the corresponding section.

.

The methodological approach employed for coding the category column was synchronised with the data collection process, such that records were assigned categories based on the data they reported. Studies assessing and characterising the observed cooling effects in FPV systems as compared with GPV systems via experimentation (CC_Exp) and modelling (CC_Mod), studies conducting experimental analyses of the thermal and electrical performance of FPV systems (FPV_Exp), those investigating the various means of further improving FPV temperature (CT), assessing the accuracy of FPV modelling and simulation approaches (FPV_ModVal), and those focused on determining and characterising the heat-transfer coefficients of FPV systems (U-values), were the six main categories found in the dataset, and the records in these categories were subject to the meta-analysis. Data items such as ambient temperature, water temperature, wind velocity, and solar irradiance were extracted (where available) across all records, alongside the required variables for each category (See

Table 7. Collected Data items.

Category (C)	Variables	Data Items *
CC_Exp	Observed cooling effect	∆TF_G, ∆P, ∆η
CC_Mod	Observed cooling effect	∆TF_G, ∆P, ∆η
FPV_Exp	FPV performance characterisation	TFPV, TA, TW, G, Vw
FPV_Modval	Model accuracy	RMSE, MAE
CT	Observed cooling effect	∆TC_R
U-values	Heat loss coefficient	U, Uc, Uv

* ∆T_{F_G}, ∆P, and ∆η are the change in temperature, power, and efficiency, respectively, between FPV and GPV systems, RMSE: Root Mean Squared Error, MAE: Mean Absolute Error.

).

The only outliers in this process were the case and feasibility studies (FC_Study), as discussed earlier, and they were sorted into a category of their own. As the meta-analysis discusses the quantitative trends and overarching qualitative narratives, the textual narrative synthesis was limited to only those studies in the FC_Study category, whilst all other records featured directly or otherwise in the meta-analysis. There were some studies in the other categories with quantitative data that did not quite fit into the datasheets, but as they are in the minority (n = 3), they are integrated into the discussion on the results. The PRISMA flowchart, which pictorially summarises the record selection process, is provided in Figure 3, and the categorisation matrix, containing all the articles included in the final dataset and their respective categories and references, is detailed in

Table 8. Categorisation matrix with all included records.

Ref	C1	C2	C3a	C3b	C4a	C4b	C5	C6	Datasheet
Lee [66]	X	-	-	-	X	-	-	-	CC_Exp
Jeong et al. [67]	X	-	-	-	X	-	-	-	CC_Exp
Mandavi and Tiwari [68]	X	-	-	-	X	-	-	-	CC_Exp
Shukla et al. [69]	X	-	-	-	X	-	-	-	CC_Exp
Dzamesi et al. [70]	X	-	-	-	X	-	-	-	CC_Exp
Rehman et al. [71]	X	-	-	-	X	-	-	-	CC_Exp
Peters and Nobre [72]	X	-	-	-	X	-	-	-	CC_Exp
Refaai et al. [73]	X	-	-	-	X	-	-	-	CC_Exp
Kumar and Kumar [74]	X	-	-	-	X	-	-	-	CC_Exp
Majumder et al. [75]	X	-	-	-	X	-	-	-	CC_Exp
Elminshawy et al. [76]	X	-	-	-	X	-	-	-	CC_Exp
Nisar et al. [77]	X	-	-	-	X	-	-	-	CC_Exp
Ramanan et al. [78]	X	-	-	-	X	-	-	-	CC_Exp
Elminshawy et al. [79]	X	-	-	-	X	-	-	-	CC_Exp
Anusuya and Vijayakumar [80]	X	-	-	-	-	X	-	-	CC_Mod
Sukarso and Kim [81]	X	-	-	-	-	X	-	-	CC_Mod
Ramanan et al. [82]	X	-	-	-	-	X	-	-	CC_Mod
Tina and Bontempo Scavo [83]	X	-	-	X	-	X	-	-	CC_Mod
Kichou et al. [84]	X	-	-	-	-	X	-	-	CC_Mod
Karami and Khameneh [85]	X	-	-	-	-	X	-	-	CC_Mod
Semeskandeh et al. [86]	X	-	-	-	-	X	-	-	CC_Mod
Silva et al. [87]	X	-	-	-	-	X	-	-	CC_Mod
Sheikh et al. [88]	X	-	-	X	-	-	X	-	CT
Elminshawy, N et al. [89]	X	-	X	-	-	-	-	-	CT
Elminshawy, N et al. [90]	X	-	X	-	-	-	-	-	CT
Sazali et al. [91]	X	-	X	-	-	-	-	-	CT
Sutanto et al. [92]	X	-	X	-	-	-	-	-	CT
Elminshawy, N et al. [93]	X	-	X	-	-	-	-	-	CT
Amin and Kocher [94]	X	-	-	-	-	X	-	-	CT
Elminshawy, N et al. [95]	X	-	X	-	-	-	-	-	CT
Araimi et al. [96]	X	-	X	-	-	-	-	-	FPV_Exp
Intwala and Ghosh [97]	X	-	X	-	-	-	-	-	FPV_Exp
Ma et al. [98]	X	-	X	-	-	-	-	-	FPV_Exp
Manoj Kumar et al. [99]	X	-	X	X	-	-	-	X	FPV_ModVal
Suh et al. [100]	X	-	X	X	-	-	-	X	FPV_ModVal
Niyaz et al. [101]	-	X	X	X	-	-	X	X	FPV_ModVal
Amiot et al. [102]	-	X	X	X	-	-	X	X	FPV_ModVal
Osama et al. [103]	X	-	X	X	-	-	-	X	FPV_ModVal
Osama et al. [104]	X	-	X	X	-	-	-	X	FPV_ModVal
Rahaman et al. [105]	X	X	X	X	-	-	X	X	FPV_ModVal
Nicola and Berwind [106]	-	X	X	X	-	-	-	X	FPV_ModVal
Makhija et al. [107]	X	-	X	X	-	-	-	X	FPV_ModVal
Dörenkämper et al. [108]	-	X	X	X	-	-	-	X	FPV_ModVal
Willemse et al. [109]	-	X	X	X	-	-	X	X	FPV_ModVal
Kaplanis et al. [110]	X	X	X	X	-	-	X	X	FPV_ModVal
Wu, R et al. [111]	X	X	X	X	-	-	X	X	FPV_ModVal
Elminshawy, N et al. [112]	X	-	X	X	-	-	-	X	FPV_ModVal
Amiot et al. [113]	-	X	X	X	-	-	X	X	U-values
Lindholm et al. [114]	-	X	-	X	-	-	-	-	U-values
Makhija et al. [115]	-	X	X	X	-	-	X	-	U-values
Nysted et al. [116]	-	X	X	X	-	-	X	X	U-values
Wu, R et al. [117]	-	X	X	-	-	-	-	-	U-values
Dörenkämper et al. [118]	-	X	X	-	-	-	-	-	U-values
Dörenkämper et al. [119]	-	X	X	-	-	-	-	-	U-values
Kjeldstad et al. [120]	-	X	X	-	-	-	-	-	U-values
Wu, H et al. [121]	X	X	X	X	X	-	X	X	Mixed (CC_Exp, CT, and FPV_Modval)
Tina and Bontempo Scavo [122]	X	X	X	X	X	X	X	X	Mixed (FPV_Modval, CC_Mod, and CT)
Tina et al. [123]	X	X	X	X	-	-	-	X	Mixed (U-Values, FPV_ModVal)
Kjeldstad et al. [124]	X	X	X	X	-	-	-	X	Mixed (FPV_Exp, U-Values/FPV_ModVal)
Indartono et al. [125]	X	-	X	-	X	-	-	-	Mixed (CT, CC_Exp)
Sutanto et al. [126]	X	-	X	-	X	-	-	-	Mixed (CT, CC_Exp)
Lindholm et al. [127]	-	X	X	X	-	-	X	X	Mixed (U-Values/FPV_ModVal)
Amrizal et al. [128]	X	-	-	X	-	X	-	-	Mixed (CC_Mod and CT)
Agrawal et al. [129]	X	-	X	X	X	-	-	X	Mixed (CC_Exp and FPV_ModVal)
Tina et al. [130]	X	-	X	X	-	-	-	X	Mixed (FPV_Exp, FPV_ModVal)
He et al. [131]	-	-	-	X	-	-	-	-	FC_Study
Manolache et al. [132]	-	-	-	X	-	-	-	-	FC_Study
Gao et al. [133]	X	-	-	X	-	X	-	-	FC_Study
Avasthi et al. [134]	-	-	-	X	-	-	-	-	FC_Study
Ravichandran et al. [135]	X	-	-	X	-	X	-	-	FC_Study
Yakubu et al. [136]	-	-	-	-	X	X	-	-	FC_Study
Aboshosha and Hamad [137]	-	-	-	X	-	-	-	-	FC_Study
Asgher and Iqbal [138]	-	-	-	X	-	-	-	-	FC_Study
Jamroen et al. [139]	-	-	X	-	-	-	-	-	FC_Study
Abd.Wahab and Mustafa [140]	-	-	-	-	-	X	-	-	FC_Study
Micheli [141]	X	-	-	X	-	-	-	-	FC_Study
Maraj et al. [142]	-	-	X	-	-	-	-	-	FC_Study
Choi et al. [143]	-	-	X	-	-	-	-	-	FC_Study
Maia et al. [144]	X	-	-	-	-	-	X	-	FC_Study
Karatas and Yilmaz [145]	-	-	X	-	-	-	-	-	FC_Study
Cáceres González et al. [146]	X	-	-	-	-	X	-	-	FC_Study
Ayyad et al. [147]	X	-	-	-	-	X	-	-	FC_Study
Kaymak and Şahin [148]	X	-	X	-	X	-	-	-	FC_Study
Eyring and Kittner [149]	-	-	-	X	-	-	-	-	FC_Study
Choi et al. [150]	-	-	-	X	-	-	-	-	FC_Study
Irshad et al. [151]	-	-	-	X	-	-	-	-	FC_Study
Cosgun and Demir [152]	-	-	-	X	-	-	-	-	FC_Study
Abdulhadi et al. [153]	X	-	-	-	-	X	-	-	FC_Study
Passos et al. [154]	-	-	-	X	-	-	-	-	FC_Study
Goh et al. [155]	-	-	-	X	-	-	-	-	FC_Study
Mehadi et al. [156]	-	-	-	X	-	-	-	-	FC_Study
Ilas and Islam [157]	-	-	X	-	-	-	-	-	FC_Study
Rahmat et al. [158]	-	-	X	-	-	-	-	-	FC_Study
Ravichandran et al. [159]	-	-	-	X	-	-	-	-	FC_Study
Anbarasu and Suresh [160]	-	-	X	-	-	-	-	-	FC_Study
Getie and Jember [161]	-	-	-	X	-	-	-	-	FC_Study
Zhou et al. [162]	-	-	-	X	-	-	-	-	FC_Study
Al Shammary et al. [163]	-	-	X	-	-	-	-	-	FC_Study
Mekonnen et al. [164]	-	-	-	-	-	X	-	-	FC_Study
Al-Smairan et al. [165]	-	-	-	-	-	X	-	-	FC_Study
Kaymak and Şahin [166]	-	-	X	-	X	-	-	-	FC_Study
Dixon et al. [167]	-	-	-	X	-	-	-	-	FC_Study
Liu et al. [168]	-	-	-	X	-	-	X	-	FC_Study
Minda et al. [169]	-	-	-	X	-	-	-	-	FC_Study
Razak and Nor [170]	X	-	-	X	-	-	-	-	FC_Study
Zayed et al. [171]	X	-	-	X	-	-	-	-	FC_Study
Sulaiman et al. [172]	X	-	-	X	-	-	-	-	FC_Study
Khortsriwong et al. [173]	X	-	-	X	-	-	-	-	FC_Study
Huang et al. [174]	X	-	-	X	-	-	-	-	FC_Study

For C3 and C4, a and b refer to methods (a = experiment, b = modelling).

.

3. Results

The systematic review is split twofold, as earlier indicated, into a meta-analysis and a textual narrative synthesis. Furthermore, to provide additional insights into the corresponding knowledge base, a bibliometric analysis of the resultant dataset is also discussed before presenting the main results.

3.1. Dataset Characterisation and Bibliometric Analysis

The records were first characterised based on annual publications. The results (see Figure 4) clearly indicate that the field is gaining traction, with increases in annual publications observed in every year apart from 2023. As at the cut-off time for the study (30 May 2025), the number of publications for 2025 was already approaching the total figures for the year before and is on course to significantly exceed them as well.

Furthermore, the 109 records in the dataset were analysed with VOSviewer software (1.6.20) [175], which is a science mapping and bibliometric analysis tool used to identify key relationships within the dataset through co-occurrence analysis. Two knowledge maps were generated from the software: a keyword co-occurrence map (Figure 5) and a text-data co-occurrence map based on the abstracts and titles of the records (Figure 6). Both maps yielded similar results; the words associated with the dataset were broadly split across two categories: The first, termed ‘General Characteristics’, comprises general informative terms surrounding FPV and PV generation, also hinting at technoeconomic assessments (TEA) and case studies with terms like ‘levelized cost’, ‘potential’, ‘scenario’, and ‘location’. The second category includes more technical terms regarding the temperature-based performance of PV systems. In the text-data co-occurrence map, there is a third cluster highlighted in yellow by the software that focuses more specifically on FPV systems, in contrast with the red cluster that deals specifically with PV systems. Both have been grouped under the category of ‘Temperature-based performance’ as, despite the colour grouping, the links indicate strong co-occurrence.

Finally, an assessment was executed on the specified geographical location in the records in the dataset. The locations have been plotted using their coordinate data via the folium package in Python. The results indicate a strong bias towards the Northern and Eastern hemispheres, with ten per cent of the records assessing locations in the Southern or Western hemispheres (see Figure 7). The effect this has on the reported data, or the forthcoming analysis, is uncertain, but from a scientific standpoint, heavily localised data bears the inherent risk of bias [176].

3.2. Textual Narrative Synthesis

This section aims to present a categorical synthesis of the studies that did not possess sufficient quantitative data to fit into the meta-analysis scheme. The studies in the dataset are coded based on the reported data across the categories in the codebook to present a synthesised representation of the information in the dataset. The categories in the codebook are either definite (study location, scale, type) or open-ended (temperature modelling method, findings) based on the range of variation in the data field. The complete codebook for the textual narrative synthesis and the results of the dataset coding are detailed in

Table 9. Record coding system for textual narrative synthesis.

Study Location (Continent)	Scale of Study	Study Type	Temp. Modelling Method	Findings
AF = Africa	<1 MW	TEA = Technoeconomic Assessment	Open-ended (Faiman, Kamuyu, etc.)	Open-ended
AS = Asia	1 MW–50 MW	GPE = General Performance Evaluation
EU = Europe	>50 MW	INT = FPV Integration
GL = Global	NL = National Scale
NA = North America	GL = Global Scale
SA = South America

and

Table 10. Coded textual narrative synthesis dataset.

Ref.	Continent	Scale	Type	Temp. Modelling Method	Findings
[131]	AS	1 MW–50 MW	TEA	PVSyst (Faiman)	The fixed pile-based system outperformed the FPV system
[132]	EU	<1 MW	GPE	Nil	Good Generation Potential
[133]	AS	>50 MW	INT	Kamuyu, Hayibo	Integration is feasible; Water-cooling is more effective than air-cooling
[134]	AS	1 MW–50 MW	TEA	Nil	Bifacials outperformed monofacials
[135]	AS	1 MW–50 MW	TEA	Helioscope	Thin-film outperforms GPVs and Pontoon-based FPVs
[136]	AF	1 MW–50 MW	GPE	SAM (Sandia)	Bifacials outperformed better than monofacials
[137]	AF	>50 MW	TEA	PVSyst (Faiman)	Good Generation Potential
[138]	NA	<1 MW	TEA	HOMER PRO (D&B)	FPV was economically superior to the diesel system
[139]	AS	<1 MW	TEA	Exp	Good Performance
[140]	AS	<1 MW	GPE	Exp	GPV outperformed FPV
[141]	EU	National Scale	TEA	Kamuyu	Economic Viability
[142]	EU	<1 MW	GPE	Exp	Good Performance
[143]	AS	1 MW–50 MW	GPE	Nil	Good performance; High safety/structural strength
[144]	SA	1 MW–50 MW	GPE	Energy balance model	Good Performance
[145]	EU	<1 MW	GPE	Nil	Irradiance was significant in determining output
[146]	SA	National Scale	GPE	Kamuyu	Good Generation Potential
[147]	GL	Global	GPE	Multi-layer	Good Generation Potential
[148]	EU	<1 MW	GPE	Exp	Similar FPV and GPV performance
[149]	EU	National Scale	TEA	Nil	Technical Feasibility; Economic Viability
[150]	AS	1 MW–50 MW	GPE	Menicucci	An increase in water level leads to shading reduction and power increase
[151]	AS	National Scale	INT	Cheng	Integration is feasible
[152]	EU	1 MW–50 MW	TEA	PVSyst (Faiman)	Bifacials outperformed better than monofacials
[153]	AS	National Scale	GPE	Standard	Good Generation Potential
[154]	SA	>50 MW	INT	PVSyst (Faiman)	Integration is Feasible; Good economic performance and environmental impact
[155]	AS	National Scale	INT	Nil	Integration is Feasible; Good economic performance and environmental impact
[156]	AS	1 MW–50 MW	INT	PVSyst (Faiman)	Integration is Feasible
[157]	AS	<1 MW	GPE	Exp	FPV outperformed GPV system
[158]	AS	<1 MW	GPE	Exp	Good Performance
[159]	AS	<1 MW	INT	Sandia model	Integration is Feasible; Good economic performance
[160]	AS	<1 MW	TEA	Nil	FPV outperformed GPV system
[161]	AF	1 MW–50 MW	TEA	PVSyst (Faiman)	Good Generation Potential
[162]	AS	1 MW–50 MW	INT	TRNSYS	Good Generation Potential
[163]	AS	1 MW–50 MW	GPE	HOMER PRO (D&B)	Integration is Feasible; Good environmental impact
[164]	AF	1 MW–50 MW	TEA	Triyana	FPV outperformed GPV system
[165]	AS	<1 MW	TEA	PVSyst (Faiman)	FPV outperformed GPV system
[166]	EU	<1 MW	GPE	Exp	Similar FPV and GPV performance
[167]	AS	National Scale	TEA	Nil	Good Generation Potential
[168]	AS	National Scale	GPE	Sandia model	Good Generation Potential
[169]	EU	1 MW–50 MW	INT	PVGIS	Integration is Feasible

, respectively. The textual synthesis is presented across the various categories in the coming sub-sections.

3.2.1. Study Location

Similarly to the observations made on the dataset as a whole, the studies in this subset are primarily focused on locations in the Eastern and Northern hemispheres. Furthermore, the studies were geographically coded in terms of the continent which serves as the point of focus of each study. Asia was the most well-represented (n = 21), with major representation from China (including Hong Kong) [131,133,155,162,168] and India [134,159,160], having 5 and 3 records, respectively. European locations came in second with 9 records. 4 of the 9 assessed sites in Turkey while three of the four were specifically centred on the pilot plant on Lake Buyukcekme [145,148,166]. The other European studies were spread across the other Balkan nations [132,142,169], Switzerland [149], and Spain [141]. Finally, the dataset is wrapped up with 4 records on Africa [136,137,161,164], 3 South-American [144,146,154], 1 North-American [138], and 1 publication assessing Global FPV energy generation potential [147].

3.2.2. Scale of Study

Records were classified according to the scales or sizes of the assessed FPV system in each study. Four classes were created: <1 MW, 1 MW–50 MW, >50 MW, National Scale.

The most populated categories are the small and medium scale modules in the <1 MW and 1 MW–50 MW categories. These are typically small-scale case studies or proof of concepts confirming the performance of FPV systems [142,144,158,162,164]. The larger-scale studies are usually more in line with feasibility assessments, investigating the broader potential of FPV systems [149,153,155,167,168]. On the absolute extreme of this are the studies which assess national (or global) FPV potential [141,146,147,151].

3.2.3. Study Type

As discussed earlier, the records in this category are almost entirely made up of feasibility and case studies. Nonetheless, characterising the data, there were 3 main categories identified in the dataset: 14 studies classed as Technoeconomic assessments, 17 studies assessing the general performance of FPV, and 8 studies investigating the potential of integrating FPVs with other technologies.

The TEAs assessed the economic output of FPV systems using the modelled power outputs and local or general electricity prices. Publications in this category included GPV to FPV comparisons [135,160,164,165], utility/national scale cost feasibility assessments [141,149,167], and some niche applications [131,134,138,152].

Most of the studies relating to FPV integration had a component of combined FPV-Hydropower stations [154,156,159,169], with tidal [162] and PHS/PHES [133,151,155] also discussed in some studies. In these cases, technoeconomic analyses were sometimes conducted to quantify the financial implications of these integrated systems.

The third category (performance evaluation) was fairly straightforward. In most cases, it involved applying common modelling tools and techniques to assess the feasibility of FPV plant installation, particularly with respect to the potential power output of the studied plant [132,153,158,163], although regular comparisons between GPVs and FPVs were still prevalent in this classification [140,148,157,166].

3.2.4. Temperature Modelling Approach

This category was defined as ‘open-ended’ in the codebook, as the model choices reported across the dataset were fairly diverse. Characterising this data is particularly useful as it is an indicator of the state of the art, or at least the popular choices when modelling FPV temperature.

PVSyst, based on the Faiman model, was the most employed tool, with [131,137,152,154,156,161,165] making use of it. As well, refs. [159,168] used the Sandia model, and [133,141,146] used Kamuyu’s empirical model. The other records were quite varied in terms of model application, employing notable modelling and simulation software such as HOMER PRO, TRNSYS, Helioscope, etc.

Refs. [139,140,142,148,157,158,166] derived their outputs via experimental means and therefore did not need any temperature models. Furthermore, multiple records did not directly indicate any temperature modelling approach.

3.2.5. Study Output

Study output was a direct function of the study objectives. Records that assessed the performance variations between FPV and GPV systems generally indicated improved performance for FPV systems [157,160,164,165]. There was one outlier, which indicated a slight increase for the GPV system, but this was attributed to wind/wave disturbances [140].

Studies assessing the potential integration of FPVs with other energy generation techniques all indicated feasibility [151,156,169], and in the cases where economic performance was also investigated, results yielded substantial economic viability and improved economic performance upon FPV inclusion [154,155,159]. Standalone TEAs were of the same nature, indicating satisfactory technical, economic, and, when measured, environmental performance indices. Similarly, good performance was indicated by studies solely assessing FPV performance.

There were a handful of studies which investigated the comparative performance differences between monofacial and bifacial systems in an FPV environment, and as expected, the bifacials performed better than the monofacial systems [134].

3.3. Quantitative Meta-Analysis

Before delving into the individual categories and their results, some key points of consideration regarding the quantitative meta-analysis are detailed below:

• During the data extraction process, where available, representative data points have been extracted to characterise entire studies. Depending on the complexity and manner of reporting employed by the authors of the reviewed publications, single representative datapoints are not always available. Therefore, point data from available tables or graphs is used to substitute in these cases.
• Again, owing to the wide variation across (and within) most of the datasets, there are multiple instances of a dataset having more than one datapoint. In most cases, these are applied as is to the analysis. However, in cases in the analysis where the goal is to determine a representative metric for the dataset (e.g., mean, median, deviation, etc.), multi-value studies are often limited to a single value, so as not to skew the results. Mean values are calculated to achieve this objective.
• The following meta-analysis was executed using Python 3.13. All graphs were produced using the matplotlib library, and the statistical analysis functions were performed using the scikit-learn library.
• Record IDs as depicted in Figure 8, Figure 9, and Figure 10 are numerically synchronised with their references to aid reader comprehension. Citations corresponding to the IDs are provided below the figures in all cases.

3.3.1. CC_Exp

18 studies met the inclusion criteria for the quantitative analysis in this category, and the total number of datapoints came to 40. As this category represents the reported temperature differences between FPV and GPV systems (∆T_{F_G}), these values are plotted in Figure 8a. Nearly all the included studies indicated a temperature drop between GPV and corresponding FPV systems, as demonstrated by the positive ∆T_{F_G} values. The sole outlier is from a study conducted by Peters and Nobre [72], where a 2.8 MW floating plant was compared to a 12.8 KW rooftop system in Kampot Province, Cambodia. The authors indicated increased temperatures of up to 9 K for the FPV system as compared to the GPV. This was attributed to the reduced wind speed at the FPV level, compared to the GPV system, which was 3 m above the ground. The authors also noted that the sizes of the respective systems could have contributed to the temperature difference. Nonetheless, the mean value of ∆T_{F_G} of the dataset was calculated to be 3.76 K based on the raw data points extracted, and 3.35 K when each publication was limited to a single data point (described in Section 3.3). Furthermore, the standard deviation of the dataset was calculated to be 3.42 K and 4.01 K for the raw data and the single-point mean, respectively.

The noted outlier, record [72], was removed from the dataset, and the remaining values are plotted in Figure 8b. The mean temperature rose, as expected, to 4.09 K, with an approximately equal per-record value of 4.08 K. The standard deviation also dropped significantly to 2.74 K and 2.71 K for each case. As well, with the removal of the outlier, most of the ∆T_{F_G} values ranged between +1 K and +10 K.

As covered in Section 1.2, the temperature coefficient (β) of a PV panel is a design-specific parameter that denotes the rate at which the electrical performance of a PV panel changes with changes in module temperature, with typical values ranging between 0.4 and 0.5%/K [177], which implies proportionality between the change in temperature and the change in module efficiency. The reported temperature changes (∆T_{F_G}) were plotted against both the changes in power (∆P) and the change in efficiency (∆η) in Figure 9a,b to examine the existence of any such relationship, with only records having a ∆T_{F_G} and either a ∆P or ∆η value being included in the plot.

The study by Rehman et al. [71] stands out as an outlier in terms of ∆P and ∆η values. This was an experimental assessment undertaken by researchers at the King Fahd University of Petroleum and Minerals, Saudi Arabia, in which a GPV plant 150 m inland from the shoreline was compared to an FPV plant 25 m off the coast. Despite a relatively regular average temperature reduction of about 5.8 K, an increase in power/efficiency of up to 200% was reported. The cause in this instance is not clearly demonstrated, but factors such as the solar radiation and varied tilt angles (5° for the FPV and 20° for the GPV) could potentially explain the exceptional differences in performance. Also, it should be noted that the FPV performance data from this experiment indicated that the floating system performed close to the nominal efficiency at STC (standard test conditions); therefore, the results of the experiment may be interpreted rather as an underperformance of the GPV system rather than an overperformance of the FPV system.

There were also a few edge cases where the ∆P or ∆η values were observed to be negative. This was the case in an iteration of the experiment carried out by Lee [66], where the negative reading is likely attributed to minor alterations in surrounding conditions like water-wave stability and wind-loading, altering the incident radiation and power output. A similar outcome was noticed in the study by Kumar and Kumar in [74], where the observed cooling effect was assessed across different module technologies (HIT, CdTe, P-Si). The slightly negative readings came from the P-Si and HIT modules (2.7% and 0.4%), while the CdTe module saw a 3% increase in performance. While these reductions are in disagreement with expected values, as the experiment was conducted over 1 year, they were attributed to module degradation in the water environment.

In order to homogenise the dataset, the noted outliers were excluded, and the datapoints were reduced to one per study. The resulting plots, Figure 9c,d show very little correlation between ∆T_{F_G} and ∆P or ∆η, despite the exclusion and simplifications made to the dataset. To quantify this, basic linear regression analyses were executed with ∆T_{F_G} as the independent variable and ∆P and ∆η each as the dependent variables. The results, unfortunately, yielded no correlation, with coefficients of determination (R²) of 0.0009 between ∆T_{F_G} and ∆η, and 0.2248 between ∆T_{F_G} and ∆P.

3.3.2. CC_Mod

Similar analyses were conducted for this dataset, as it contains identical datatypes to CC_Exp, only that, as the name entails, it focuses solely on studies that employed modelling/simulation to conduct their research. The resultant dataset consisted of 10 papers, containing 66 datapoints. Unlike the CC_Exp dataset, not all studies reported data on ∆T_{F_G}, with only 28 recorded datapoints for this parameter. Nonetheless, the plot of the ∆T_{F_G} values (Figure 10a) is similar to that of CC_Exp, with generally positive values. The only negative ∆T_{F_G} values are from a CFD-based study by Ramanan et al. [82], where wind speed, ambient temperature, and water temperature were varied, and the observed effects on FPV and GPV modules were recorded. It was observed that ∆T_{F_G} was slightly negative when the water temperature was higher than or equal to the ambient temperature. Otherwise, the ∆T_{F_G} values of the dataset ranged between +1 K and +12 K. The mean values are of limited significance compared to the CC_Exp dataset, primarily because the simulated parameters are not necessarily replicable in real-life conditions; nevertheless, the dataset had an individual and per-record mean ∆T_{F_G} of 4.02 K and 5.94 K, respectively.

In terms of simulation methods, the records were split between the Kamuyu [37] model (records [80,81,86]), PVSyst—using the Faiman [32] model ([83]; which uses the Faiman model), Muzathik [48] model (Karami and Khameneh [85]), and Computational Fluid Dynamics (CFD) modelling [82,87,128], while Kichou et al.’s study [84] used a combination of the Kamuyu and Muzathik equations in modelling FPV temperature.

In plotting the relationship between ∆T_{F_G} and ∆P (Figure 10b), despite only 4 records fulfilling the inclusion requirements, unlike the CC_Exp dataset, there does seem to be an appreciable level of correlation between ∆T_{F_G} and ∆P. To confirm this, a linear regression was executed similarly to that with the CC_Exp dataset, which yielded a correlation coefficient of 0.9942 between ∆T_{F_G} and ∆P. The correlation between ∆η and ∆T_{F_G} was not plotted separately, as only 2 datapoints met the requirements.

3.3.3. FPV_Exp

This category contains studies where the performance of FPV systems was characterised via experimentation. The general approach of the studies in this dataset was the collation of ambient environmental data, as well as module temperature and performance data. In total, 5 studies, yielding 27 datapoints, fit into this category and were analysed. As the goal was experimental characterisation, 4 of the 5 studies [96,97,124,130] reported data on the parameters typically used in empirical regression models (T_FPV, T_AMB, T_W, G, and V_W) as described in Table 1. Therefore, in order to characterise this dataset, a regression analysis was executed, with FPV module temperature as the dependent variable and various combinations of the other experimentally determined values set as the independent variables. The assessed combinations were chosen based on the format of commonly applied models observed in the literature (Table 1). The goal of this analysis is to identify the accuracy with which FPV module temperature can be predicted using environmental variables. The resulting regression equations for all the investigated models are given below, with

Table 11. Regression analysis results from experimental FPV data.

S/N	nr *	np *	Model equation	R2	MAE	MSE	RMSE	MAPE **
A	4	16	TFPV = 1.2234 ∙ TW + 11.2679	0.3569	7.1358	64.5303	8.0331	18.3378
B	4	16	TFPV = 1.4560 ∙ TW + 0.0216 ∙ G − 7.0910	0.5613	5.8503	44.0151	6.6344	14.9731
C	4	16	TFPV = 1.4127 ∙ TA + 1.3551	0.7184	4.3743	28.2585	5.3159	12.0812
D	4	16	TFPV = 1.6672 ∙ TA − 0.3957 ∙ TW + 3.7794	0.7324	4.0812	26.8527	5.1820	11.9773
E	3	8	TFPV = 2.4433 ∙ TW + 6.6274 ∙ Vw − 20.2014	0.7843	4.3359	32.9569	5.7408	10.3714
F	4	16	TFPV = 1.4971 ∙ TA + 0.0187 ∙ G − 12.0936	0.8800	3.0158	12.0382	3.4696	8.9764
G	4	16	TFPV = 1.5473 ∙ TA − 0.0799 ∙ TW + 0.0185 ∙ G − 11.4134	0.8806	2.9643	11.9841	3.4618	8.9704
H	3	8	TFPV = 1.9824 ∙ TA + 5.3291 ∙ Vw − 22.4768	0.9010	2.9953	15.1236	3.8889	7.4056
I	3	8	TFPV = 2.8688 ∙ TW + 0.0376 ∙ G − 3.2419 ∙ Vw − 39.6515	0.9290	2.8070	10.8542	3.2946	8.3971
J	3	8	TFPV = 4.9411 ∙ TA − 4.0251∙ TW + 2.9896 ∙ Vw − 16.8384	0.9703	1.9365	4.5323	2.1289	5.3822
K	3	8	TFPV = 2.1393 ∙ TA + 0.0274 ∙ G − 2.0688 ∙ Vw − 34.0268	0.9833	1.3677	2.5536	1.5980	3.9254
L	3	8	TFPV = 3.6362 ∙ TA − 2.1006 ∙ TW + 0.0191 ∙ G − 1.0659 ∙ Vw − 27.6122	0.9947	0.8298	0.8049	0.8972	1.9069

* nr and np refer to the number of records (publications) and datapoints, respectively, in each analysis. ** descriptions of the error metrics are provided in Appendix A.1.

characterising the models and error metrics, and Figure 11 displaying the linear plots.

The error metrics indicate that the four-parameter model incorporating T_A, T_W, V_W, and G (Model L) performed the best, with a correlation coefficient of 0.9947 and an MAE of 0.8298 K. However, as the dataset is relatively small (as Tina et al. [130] did not include wind speed in their reported results), the high values could be a product of the model being overfit to the data. A more representative model encompassing an additional study is Model G, which does not incorporate wind velocity in its analysis and has respectable error metrics with a correlation coefficient, MAE, and RMSE of 0.8806, 2.9643 K, and 3.4618 K. As well, the worst performing model (Model A) with a correlation coefficient of only 0.35 was based on only water temperature. Notably, Model C, which incorporates only air temperature, performs significantly better, with a correlation coefficient of 0.72, indicating that air temperature has a greater effect on FPV temperature than water temperature.

Comparing the results to the other relations found in the literature, a notable observation is that Model G bears a striking resemblance to Hayibo’s [49] model (Table 1, row 15). The equation was derived from a regression analysis on experimental data and was found to predict module temperatures to a higher degree of accuracy than the Kamuyu model. While there is a reasonable degree of variation between the two models, the similarity in the coefficients indicates that the foundational theoretical relationships exist in both Models [49].

$$\mathrm{Hayibo}’\mathrm{s} \mathrm{Model}: T_{FPV}=1.2645{·T}_A-0.0873·T_w+0.0128·\mathrm{G}+13.2554$$

(3)

$$\mathrm{Model} \mathrm{G}: T_{FPV}=1.5473{·T}_A-0.0799·T_w+0.0185·\mathrm{G}+11.4134$$

(4)

3.3.4. FPV_ModVal

The next category of records is from publications that focused on developing and/or validating FPV temperature modelling/simulation approaches. The methodological approach of the studies in this section involved the experimental measurement of FPV module temperature, and comparison with modelled/simulated temperature derived from the model of choice. 19 studies and 109 datapoints were identified as appropriate for this study and were fitted into the datasheet. The Root Mean-Squared Error (RMSE) and Mean Average Error (MAE) were selected as the variables for analysis within this category, as they were frequently included in the assessed studies.

The best performing models, as depicted by the RMSE plot (Figure 12), are those by Tina and Bontempo Scavo [122], Wu H et al. [121], Lindholm et al. [127], and Wu R et al. [111]. Incidentally, all four models differ from the conventional empirical and regression-based models detailed in Table 1 in that they all employ multi-layer resistive thermal modelling, discretely assessing the heat transfer between the layers within the module, and between the module and the environment. This is usually conducted via detailed evaluation of the conductive, free, and forced convective, and radiative heat transfer mechanisms in the system. This new approach consistently yields higher accuracies than other models. Similar simulation methods were applied in studies by Amiot et al. [102], where the developed model was used to assess the thermal boundaries, and the influence of multiple parameters on module temperature was assessed; Rahaman et al. [105], where in addition to the multi-layer thermal model, a ‘simplest’ steady-state thermal model was also developed and assessed by the authors; and Willemse et al. [109], where the thermal modelling was coupled with an assessment of the effects of pontoon design on module temperature. An adjacent study, by Niyaz et al. [101], considered a different approach, modelling using similar modelling methods as those above, but assessing the accuracy of the model at predicting the temperature of modules of different types. The RMSE for Heterojunction with Intrinsic Thin Layer (HIT), Monocrystalline Silicon, and Cadmium-Telluride modules was found to be 2.37 K, 2.2 K, and 2.76 K, respectively.

Studies were executed by three groups (Agrawal et al. [129], Nicola and Berwind [106], and Tina et al. [130]), thereby assessing the potential optimisation of the more ‘conventional’ PV temperature simulation models, and their accuracy in predicting FPV temperature. In their study, Agrawal et al. assessed the accuracy of multiple regression-based models, in one of which they modified the Sandia model by substituting the ambient temperature for water temperature in the original model equation. However, the modified model performed worse than the regular version (See Figure 12).

Both the studies by Tina et al. and Nicola and Berwind employed a similar methodological approach, in that to optimise the parameters of the conventional models, the models were exposed to a training dataset, with which their parameters were optimised. After the optimisation process, the trained models were then tested on another dataset and assessed for their accuracy. The results of this analysis, in particular, are shown in the Mean average error plot (Figure 13).

According to Nicola and Berwind’s analysis, the particularly poor performance of the Zenit and Risser and Fuentes models was attributed to the former’s lack of a separate focus on the effect of wind speed, and the latter suffering from overfitting, as it was developed as an empirical model based on a regression analysis. The standard model was also lacking, due to its parametric simplicity. Both Skoplaki models performed relatively well pre-optimisation, and maintained accuracy after the optimisation process, yielding final MAE values of 1.73 K and 1.87 K. The Kurtz and Risser and Fuentes models were notably poor at predicting FPV temperature in their basic states, but were particularly responsive to the optimisation process, with the MAE values dropping from 7.61 K to 1.82 K, and 12.71 K to 2.13 K.

In reference to Tina et al.’s study, the assessed models (Sandia, Faiman, and Keddouda) performed reasonably well before and after optimisation. The Faiman was further tuned, introducing parameters to account for the inclusion of water temperature (See Equation (5)), and this adjusted model performed best in comparison to the other conventional models, with a mean average error of only 1.05 K. It is worth mentioning that the base Faiman model appears more accurate than the optimised model. This is because the chart portrays the average of the model accuracy values as reported across multiple studies; while [130] reported an average RMSE of approximately 2.39 K across the assessed model configurations, the combined average RMSE of [103,123] yielded approximately 2.39 K, therefore leading to an overall average RMSE of 1.66 for the base Faiman model, which is slightly lower than the RMSE of the optimised Faiman model as computed by [130] to be 1.71 K.

$$T_c=T_a+{\displaystyle \frac{G_{POA}}{U_c+U_w{V}_{air}}}\equiv T_a+{\displaystyle \frac{{\alpha G}_{POA}\left(1-\eta \right)}{U_c+U_w{V}_w}}+U_2{T}_w$$

(5)

The studies that employed CFD-based modelling implemented similar parametric methods and approaches [105,112]. While CFD itself is not a modelling method per se, the records are grouped under this umbrella because it is the most appropriate means of categorising them. Furthermore, conducting a detailed analysis of the various methods used in these studies would be overly complex and beyond the scope of this work.

Not covered on the graph, but still of note, is the study by Kaplanis et al. [110], which was conducted similarly to the others that employed resistive thermal modelling, with a deviation of as little as 0.2 K between the measured and modelled temperature. Osama et al. [104] and Dörenkämper et al. [108] also had data on temperature difference, but not according to the captured statistical metrics. Both assessed the accuracy of the Faiman and Sandia models, with [104] generating the differences between modelled temperatures of 3 K and 6 K, and [108] obtaining values between 8 K and 15 K, and 4 K and 9 K, respectively.

3.3.5. U_Values

Heat transfer coefficients are another well-represented category in the dataset, with 11 records and 61 datapoints available for analysis. Datapoints were plotted according to their reported Overall U-values and FPV-footprint/structural classification (in accordance with Liu et al.’s system [35]). Records that were not explicitly classified by their authors were assigned a classification based on the provided system descriptions or images. As U-values are usually dependent on wind speed, extracted data in periods of high wind speeds could skew the results of the analysis. To curtail this, the dataset was limited to points with wind speeds between v ≅ 0m/s and v ≅ 3m/s, thus reducing the number of datapoints to 51. The resultant dataset is represented with a boxplot (Figure 14), with the study method colour-coded onto the figure.

As Figure 14 shows, the U-values in the dataset follow the expected pattern as laid out by Liu et al., as larger footprints tend to restrict heat transfer, thus yielding lower heat transfer coefficients [117,118,119]. Another interesting output from [119] is that with similar ‘free-standing’ configurations, FPV systems show higher heat transfer coefficients than GPV systems. The “membrane” systems cited are based on an FPV plant in Skafta, Norway, where the FPV modules are placed in quasi-direct contact with the water via a floating membrane. The system characteristics were investigated by Kjeldstad et al. [124] via experimentation, and Lindholm et al. [127] via a multi-layer thermal model as described in Section 3.3.4, to ascertain the difference in heat-transfer properties between an FPV system in direct thermal contact with the water (water-cooled [wc]), and a system raised slightly above the membrane layer (air-cooled [ac]). The water-cooled system performed significantly higher U-values across multiple configurations, with values as high as 79W/m²K, compared to the ac system, which ranged around 20 to 40 W/m²k depending on the wind velocity. There also appears to be good harmony between the values derived from the different study methods. CFD-based assessments conducted by Lindholm et al. [114] and Nysted et al. [116] largely correlate with experimental studies of systems of similar structures (Kjeldstad et al. [120], Dörenkämper et al. [119]) and a study by Makhija et al. [115], which applied the Faiman model.

The reported U_c and U_v values are also presented in Figure 15a,b. The values were calculated for datapoints where they were not presented, but which contained overall U-values across a range of wind velocities. The wind-independent U-values exhibit a similar pattern in the overall values in terms of structural correlation, with the water-cooled membrane structure expectedly having the highest value, as most of its observed cooling effect is as a result of water contact. The wind-dependent values also exhibit a similar pattern, only with slightly more uniformity across the different categories. A feature not directly captured in either plot is the effect of bifaciality, as was investigated by Tina et al. [123] using a mix of experimental and modelling processes; it was observed that bifacial modules have slightly higher wind-independent thermal coefficients than monofacial modules.

3.3.6. CT

The final category explored in the quantitative meta-analysis details the various supplementary cooling techniques explored by studies in the final dataset and comprises 13 publications with 45 datapoints. The focus of the study was initially intended to consider only conventional FPVs. However, the drive to produce more efficient systems (particularly in relatively arid regions [93] has led to various innovations aimed at further reducing the temperature of FPV systems and will undoubtedly form a key part of the FPV research landscape. A plot of the identified cooling techniques and their reported temperature reduction relative to regular FPV systems (∆T_{C_R}) is depicted in Figure 16.

Sixteen iterations of cooling techniques were identified in the plot, with a wide range of explored options. Water veils, which involve the active pumping or spraying of a thin layer of water on the surface of the modules, were investigated by Tina et al. [122], Sazali et al. [91], and Wu H et al. [121]. Tina et al. conducted a modelled assessment using a model developed by the researchers, and attained temperature reductions of up to 17 K at very low wind speeds, while Wu H et al. executed the analysis experimentally and attained ∆T_{C_R} values of up to 18 K. Sazali et al. also conducted an experiment, but theirs differed in that the cooling system was only activated when the module temperature exceeded 35 K; therefore, the instantaneous ∆T_{C_R} values were significantly lower than the other two. The studies by [91,121] also assessed the effect of water sprays at the back of FPV modules, indicating ∆T_{C_R} values of up to 12 K and 1 K (instantaneous), respectively.

Indartono et al. [125] and Sutanto et al. [126] experimentally investigated the cooling effects of a thermosiphon system. This is a passive cooling method, making use of buoyancy-driven fluid circulation as the heat-transfer mechanism with cooling pipes in thermal contact with the back of the modules and the water body acting as the cold reservoir. Both studies indicated superior performance for the cooled system, with ∆T_{C_R} values of 8.46 K and 3 K, respectively. Sutanto et al., in another publication, investigated the cooling effect of a similar cooling system, with the primary alteration being the placement of the cooling channel above the PV module. The channel was transparent, and the effect of different coolants (pure water and water infused with silver nanoparticles) was assessed via CFD modelling in ANSYS Fluent 17.0. The Ag-infused system had better thermal performance but reduced electrical performance than the pure water system due to the increased spectral absorption, reducing the available incident radiation [92].

Other cooling methods explored by different researchers included the assessment of a variety of ‘fins’ attached to the PV modules [89,90,93,128], Serpentine back-of-module cooling channels [88,94], and partial module submergence [90].

4. Discussion

4.1. Key Findings

The first point of note was the geographical distribution of the studies in the dataset. As was highlighted during the dataset characterisation, the majority of the studies emanated from the Eastern and Northern Hemispheres, which could lead to latent biases in the reported data. More specifically, however, there is a clear prevalence of studies originating from Asia and Europe within the dataset. In contrast to other parts of the globe, the high concentration of FPV-related studies from this region is largely driven by high demand for electricity, coupled with government policies incentivising research into renewable systems and the existing infrastructural systems available. Furthermore, the availability of FPV-test and deployment sites (again often backed by the relevant policies) significantly contributes to the progress being made in these areas.

The textual synthesis yielded relatively homogenous results. Most of the studies in this category assessed FPV performance across multiple metrics, including electric/power output, economic viability, and feasibility of integration. The narratives emerging from the studies indicated good synergy between FPV and other renewable energy technologies and supplemented the meta-analysis in quantifying the overall performance of FPV systems. Hydro-electric (or hydro-based) integration in particular is of notable interest, as installation on conventional plants has the added advantage of being able to leverage the existing infrastructure for new developments.

The meta-analysis of the datasheets had varying results with respect to prior expectations. Analysis of the CC_Exp and CC_Mod datasets confirmed the expected temperature drop experienced (∆T_{F_G}) by FPV systems as compared with GPVs, with the ∆T_{F_G} values largely sitting between the bounds of 1K and 10K across both modelled/simulated and experimental studies. Despite the apparent homogeneity, there are still unanswered questions on the causes of the identified variations and outliers (a notable one being the 9K temperature rise as reported by Peters and Nobre [72]). There are multiple potential factors influencing these disparities: the complex microclimate of the aquatic environment is yet to be fully understood in the context of FPV performance, and beyond that, the macroclimates of the locations of the experimental studies could also play a significant role, tying well into the observations made on the geographical distribution of the studies in the dataset. The focus of this study was primarily on the observed temperature variations in FPV systems, and as alluded to in the introduction, reduced temperatures are usually linked to better PV system performance.

The intricacies of the relationship between temperature change and power output/electrical performance were not explored in detail, but the executed analysis indicates wider levels of variation in the correlation between temperature differences (∆T_{F_G}) and the changes in module performance (measured via changes in power output {∆P} and efficiency {∆η}). The variations observed in this relationship are considerably more complex, as alongside the potential causes raised previously, the electrical performance of an FPV module is directly influenced by other factors such as the underlying module technology; the hydrodynamic characteristics of the interactions between the host body of water and the floating structure/system; environmental variations in the quality of solar radiation; and other general losses experienced by PV systems: shading, soiling, mismatch, and module degradation—the effects of which are still yet to be fully understood in the context of FPV systems. However, the analysis of the modelled datasets indicated significantly higher levels of correlation between ∆T_{F_G} and ∆P (or ∆η), likely due to the inherent control over parameterised numerical models. Nevertheless, the conclusions drawn on this outcome are limited due to the size of the dataset.

The reports on FPV experimentation elucidated the thermal characteristics of FPV systems. The regression analysis executed on the reported experimental data yielded promising results, with the output indicating that increased parameterisation leads to higher model accuracy. This is understandably due to the inherent complexity of the FPV environment and the need for the standardised data collection protocols to be applied to FPV experimentation. The models developed in this study performed quite well, with respectable correlation coefficients of up to 0.9833 and 0.8806 for the best-performing five and four-parameter models, respectively. As with the previous category on FPV-GPV cooling, the resulting model should be subject to further validation, as only five [96,98,124,129,130] studies were included in this dataset, and as few as three of those featured in the five-parameter regression model. Nonetheless, in comparing the developed models to those in the literature, an appreciable level of similarity was observed between one of the models (Model G) and Hayibo’s regression model [49], potentially indicating the prevalence of the underlying principles of FPV thermal performance in both models. Conversely, however, the observable differences between these two regression models and those with constants of considerably different orders (e.g., Kamuyu [37], Tamizhmani [47], Muzathik [48]) raise further questions on the inherent issues with localised regression models and their applicability across different experimental environments.

The FPV model validation datasheet was equally rich in informative content. Despite the agreed-upon inconsistency of FPV modelling techniques, more recent approaches based on multi-layer thermal resistance modelling techniques were found to perform the best across the dataset [111,121,122,127]. Conventional models like the Faiman and Sandia models still maintain respectable levels of accuracy with optimised parameters. However, for practical use, the parameter optimisation process makes them less viable than the newer models. This is further supported by the work of Manoj Kumar et al. [99] and Suh et al. [100], where commercial tools like PVSyst, System Advisory Model (SAM), TRNSYS, and Helioscope were found to be relatively inaccurate when compared to experimental measurements. Finally, there was a surprising underrepresentation of the better-known empirical models in the dataset, particularly with the Kamuyu model, which is typically regarded as a benchmark for empirical FPV models.

The U-value of an FPV system directly dictates the rate at which the system loses heat. As discussed earlier, a major means of classifying the U-Values of different systems is according to the system structure. Employing Liu’s structural classification system, the reported outputs correspond to the expected value ranges. Large footprint systems were generally in the range of 28–32 w/m²k [114,117,118], small footprints from 31 to 46 W/m²k [116,119,120], and the only free-standing FPV system was calculated to be 55 W/m²k, considerably higher than the values for free-standing GPV systems, which range between 29 and 35 W/m²k [119]. Also of significance are the values of the water-cooled membrane system, which range between 71 and 79 W/m²k, considerably higher than air-cooled systems on the same structure (28–39 W/m²k) [119,124,127] (all based on the upper and lower bounds of the box plots). The variation between the heat transfer coefficients of different system types underscores the importance of FPV plant design considerations. Specifically, plant designers and researchers alike must properly understand the trade-offs between different system designs in terms of cost, structural stability, and ultimately, thermal and electrical performance. Finally, on the subject of U-Values, the work of Niyaz et al. [101] merits mention, where a different approach was taken to determining U-values, as the heat-loss coefficients were modelled as front-of-module and back-of-module coefficients separately.

Cooling techniques are a welcome development, especially in areas of particularly high ambient temperatures, as the temperature reduction can yield a significant increase in module performance [93,125,126]. Passive cooling methods like module submergence [89,90,95] and the incorporation of fins [89,128] are generally preferred to active methods due to the additional energy consumption; however, in certain cases, the expended energy is recouped, with a net-positive energy balance [91,121,122]. It should be noted that submerged PVs (also referred to as underwater or undersea) represent an area of prolonged interest in the PV research community for similar reasons to conventional FPVs. Also, although not discussed in detail in this study, the difference in performance characteristics between bifacial and monofacial modules is another key study area within this dataset. The electrical output [96,97,123], thermal characteristics [122,123,130], and economic viability [134,152] in particular were extensively assessed, with superior bifacial performance reported in all instances.

Finally, there were five studies [170,171,172,173,174] among the 109 that were not assessed quantitatively or in the textual narrative synthesis. These were records that leveraged artificial intelligence and machine learning to compute or derive key variables and/or correlations as pertained to the specific study. They were excluded from the meta-analysis as they did not fit into any of the data sheets, despite meeting the objective inclusion criteria. Whilst the focus of this study is on physical modelling, advanced mathematical applications leveraging these techniques are undoubtedly a key part of this field, as they are for most others.

4.2. Limitations

The limitations of this study are primarily centred around data availability. Ideally, the dataset would consist of a more encompassing set to enable a broader series of analyses. Perhaps most notable, there is insufficient data reported on evaporative cooling and the effect of relative humidity, which is known to contribute significantly to FPV temperature [110]. Comprehensive data recorded on this effect would be key to each data category, but would be especially useful in the FPV_Exp datasheet, as further parameterisation would lead to a more robust set of regression models.

The lack of a standard for reporting and analysing FPV data is a significant hindrance to effectively conducting a review of this nature. The standards in place for regular PVs do not explicitly apply to FPVs, therefore making data collection especially tedious, as this review has made clear. The DNV-RP-0584 [178] recommended practices are the best example of such a standard, as they provide an extensive list of meteorological measurements to be recorded on an FPV site. However, the recommendations are yet to be standardised within the research community. It is reported that an IEC (International Electrotechnical Commission) standard is also under development and should be released shortly, which would greatly aid the work in this field [13].

Apart from the data availability, the sorting methodology could have been amended to expand the dataset and lead to a more inclusive and holistic study. The Python-aided screening and exclusion were successful at maintaining objectivity; however, the possibility persists that relevant records may have been omitted from the dataset without manual review due to this strict methodology.

4.3. Future Research Agenda

The SLR also enables the synthesis of a series of future research areas arising from the findings of this study, which are summarised as follows:

• Further studies are required to determine the relationship between temperature reduction and possible efficiency improvements for FPVs. Such studies require standard test conditions as well as the elimination of other variables that have the potential to influence efficiency. Therefore, a suitable hypothesis model is required, combined with multivariate analysis, to comprehensively determine the effect of temperature change on FPV efficiency and hence power generation.
• Experimental studies are recommended to address the current lack of studies originating in the Southern Hemisphere, and especially geographic-based studies that take into account different weather systems across the world.
• The meta-analysis conducted in the study herein identified a regression model having suitability for FPV temperature prediction, but this model requires further refinement. It is recommended to apply normalisation techniques to the dataset to improve the model’s consistency and accuracy across different conditions.
• The explicit consideration of the effects of evaporative cooling and humidity is underrepresented in the literature. Enhancing scientific knowledge on these areas will significantly improve the understanding of the total cooling effect experienced in the FPV microclimate.
• Research considering the comparative assessment of different FPV cooling techniques is essential to maximise the natural cooling benefits. Passive and active cooling techniques should be directly compared to identify the optimal systems under test conditions.
• Further studies are recommended to elucidate the differences between monofacial and bifacial FPV systems. The models developed for FPV temperature prediction are primarily oriented towards monofacial/conventional PV modules. Research efforts should be directed towards the development of bifacial module temperature prediction models or holistic approaches that incorporate both monofacial and bifacial module characteristics in a unified framework.
• Further research accounting for the specific physical and electronic properties of PV cell materials. These exist in the domain of general photovoltaic application, but there is a need to extend these studies to account for the unique nature of FPV operation.
• A key step going forward is to develop a holistic understanding of the role of artificial intelligence in temperature prediction models. These models have been assessed, but primarily in isolation; therefore, a comprehensive evaluation of their application to the field of FPV modelling is necessary.
• There is limited research on the identification of optimal applications for FPV systems. Particularly with a focus on integration with other renewable energy systems, studies should assess the synergy between FPV and other energy technologies, and seek to clarify the ideal approaches to FPV integration.
• It is recommended that an international collaborative network between major stakeholders in FPV research be established. These collaborative endeavours should focus on consolidating research on FPV systems, thereby addressing scientific, engineering, environmental, economic, and wider societal considerations as part of a broader sustainability-oriented agenda for FPVs. Additionally, collaborative studies can be undertaken to develop technology roadmaps underpinning the development of FPV systems for different applications, e.g., integration with other renewable forms of energy, such as hydropower or offshore wind power. Such collaboration would also increase research diversity, accelerate innovation, enable data sharing and unified testing, and promote standardisation of methods in the field.
• Furthermore, future research should focus specifically on developing FPV-specific standards. These should be developed in collaboration between researchers, laboratories, and manufacturers, and the International Organisation for Standardisation (ISO) or similar relevant bodies, and should address the unique challenges faced by FPV systems, including mechanical fatigue, dynamic cabling requirements, and system durability.
• Finally, techno-economic assessments are required to evaluate the economic viability of FPV systems alongside the technical feasibility of these systems. This should involve assessing the financial implications of FPV systems as well as working towards system optimisation in line with the various modelling and performance assessment approaches thus far explored, as studies of this nature are required to achieve industrial application and adoption.

5. Conclusions

A systematic literature review and meta-analysis were conducted to ascertain the state of the art with regard to FPV modelling, simulation, and performance analysis. Through a detailed analysis of 109 documents, the FPV research landscape was thoroughly explored, and this work has shed light on the overarching characteristics of FPV systems regarding system temperature. The range of the experimentally observed and simulated temperature differences between ground-mounted and Floating PV systems is largely positive, thereby confirming the expected cooling effect observed in FPVs relative to GPVs; however, it was observed that the effect of this temperature change on the electrical performance of the PV modules was not uniform. Furthermore, calculated and measured heat transfer coefficients were found to fit within expected values for the respective structural archetypes, and in general, water-cooled systems had lower temperatures than air-cooled systems. Moreover, cooling systems designed to further reduce the temperature and increase the performance of FPVs are of increasing significance, particularly in arid areas with high temperatures. Systems applied to modelling and simulation of FPV temperature are constantly improving, with recent developments indicating that state-of-the-art prediction models are consistently more accurate than other approaches. A regression analysis performed on the experimental data produced several mathematical relationships for estimating module temperature, with one of these equations aligning closely with an established empirical model reported in the literature.

It should be highlighted that, beyond the explicit categories highlighted in the results, there are wider inferences that can be made from this study. A primary point of discussion is the ever-increasing drive towards developing environmentally friendly energy solutions as part of the global energy mix. Whilst floating PVs currently make up only a small percentage of the renewable energy sector, the increased interest bodes well for the future of the sector as part of the drive towards net-zero. Therefore, this invariably ties into the future agenda to promote collaboration and identify a clear path forward for economically viable technical innovation.

As a concluding point, the answers to the research questions posed in Section 2.1 are explicitly given below:

• According to the analysis reported in this study, floating photovoltaic systems experience a temperature reduction typically between 1 K and 10 K when compared to ground-mounted PV systems.
• There is an expected, and indeed confirmed, increase in efficiency and power generated that accompanies the temperature reduction. However, the rate of power/efficiency increase with temperature change varies significantly with experimental conditions.
• Multi-layer heat transfer models are the most accurate at predicting FPV module temperature. CFD-based models are increasingly explored, but are not as accurate as the former. Conventional models developed for ground-mounted PV systems can be tuned to function at high levels of accuracy.
• Future research is suggested based on the research areas identified in this study (Section 4.3). Advancing FPV modelling techniques and standardising FPV data reporting are crucial for progress. Improved data consistency will enhance model accuracy and comparability. Studies examining the optimised application of FPV systems will also likely have a major impact on the development and implementation of FPV systems as part of the wider adoption of renewable energy sources.