Skip to Content
SolarSolar
  • Article
  • Open Access

1 October 2026

33 Pages

Performance Prediction Based on Solar Photovoltaic Siting Using Geographic Information Systems and Machine Learning

,
,
,
,
,
,
and
1
Computer Science Department, College of Applied Sciences, University of Samarra, Samarra 34010, Iraq
2
Electromechanical Engineering Department, University of Samarra, Samarra 34010, Iraq
3
Sustainable Energy Engineering Department, Tikrit University, Tikrit 34001, Iraq
4
Air-Conditioning and Refrigeration Department, Al-Imam University College, Tikrit 34001, Iraq

Abstract

Expanding solar power plants requires strategic planning that accounts for future climatic, environmental, and topographic conditions. This study develops an integrated Geographic Information Systems (GIS), multi-criteria decision analysis (MCDA), and machine-learning framework to identify priority locations for solar PV development in Iraq under the ACCESS-CM2 SSP2-4.5 climate projection for 2050. Support vector machine (SVM), random forest (RF), and gradient-boosted trees (GBT) classifiers were evaluated against MCDA-derived reference classes using five-fold spatial block cross-validation as the primary validation procedure; the fixed stratified 70:30 split was retained only for comparison. Their spatial outputs were integrated through GIS-based intersection to identify consensus suitability zones. Under the adopted temperature- and dust-correction model and its reference parameter assumptions, the modeled mean PV-efficiency indicator was about 14.3%, and the modeled mean daily energy-yield indicator was about 872 Wh/m2/day at a mean projected air temperature of 33.3 °C and a mean DUEXTTAU index of approximately 0.20. These quantities are scenario-dependent screening-model outputs, not observed or field-calibrated measurements. The mean spatial-validation accuracies were 91.46 ± 2.90% for SVM, 84.03 ± 2.14% for RF, and 82.22 ± 4.47% for GBT; these values quantify spatially separated agreement with the MCDA-derived reference classes and do not constitute independent validation of physical site suitability. Accordingly, the resulting maps should be interpreted as a spatial screening and prioritization framework whose quantitative outputs remain conditional on the adopted climate projection, MCDA weighting scheme, PV-performance assumptions, and the absence of independent field-based ground truth.

1. Introduction

The global trend toward using renewable energy as an alternative to fossil fuels is steadily increasing because of geopolitical conflicts in fossil fuel-producing regions. Recent geopolitical conflicts in the Middle East have created substantial economic pressures on oil-dependent economies. Iraq, in particular, has experienced significant disruptions in economic activity and losses in oil revenues due to regional conflict and disruptions to oil exports [1]. Solar energy has become a focus for countries with long daylight hours to reduce dependence on traditional energy sources. In 2025, annual additions to solar power capacity exceeded 600 gigawatts for the first time, accounting for 75% of the total new renewable energy capacity installed worldwide. As a result, cumulative global solar power generation reached approximately 2.8 terawatts; as shown in Figure 1, this makes it the world’s largest power generation technology in terms of installed capacity [2]. Electricity generation from solar photovoltaics reached approximately 2000 terawatt-hours in 2024, providing nearly 7% of global electricity, compared to 5% in 2023. Solar power covered around 40% of the increase in global electricity demand in 2024 and is projected to meet almost half of the new demand through 2027 [3]. The International Energy Agency predicts that global renewable energy capacity will grow by about 4600 gigawatts from 2025 to 2030, with nearly 80% of that coming from solar PV. PV module costs are expected to continue to fall; project development times are expected to decrease; technology is expected to improve; and global and local support is expected to improve, all of which should keep solar power as the leading source of renewable energy growth through the year 2030 [4]. Careful solar farm site selection is needed to maximize energy productivity while ensuring environmental sustainability, economic viability, and reliable integration with power networks to meet this increasing demand. Effective solar farm planning should consider factors such as solar radiation, air temperature, land use, topography, and proximity to environmental constraints, transport infrastructure, and power grids. Geographic information systems and machine learning techniques can support a more systematic and efficient site-selection process [5].
Figure 1. Renewable electricity capacity growth, 2013–2030 [2].
Iraq has promising natural conditions for solar energy, with high levels of solar radiation and wide areas suitable for development, offering significant opportunities to establish solar farms.
The application and available data largely dictate the choice of algorithm appropriate for a specific problem and the success of a machine learning application. Global climate change represents one of the biggest challenges for energy systems and sustainable development. The consistent rise in greenhouse gas concentrations has led to significant changes in temperature, precipitation patterns, solar radiation, and the frequency of extreme weather events [6]. Many studies have shown that these changes affect not only ecosystems but also the operation of renewable energy systems, especially solar power plants, by affecting the efficiency and annual productivity of PV cells, as well as the climatic characteristics of suitable investment sites [7]. Combining future climate data with GIS, AI, and machine learning technologies has therefore become a modern research trend for assessing site suitability and predicting the performance of solar energy projects under climate change scenarios [5]. This contributes to the enhancement of the accuracy of spatial planning and minimization of the risks of future climate change. In this work, the impact of climate change on global solar photovoltaic (PV) energy potential was assessed by analyzing historical climate data for 1985–2014 and future projections for 2015–2100 using CMIP6-SSP scenarios. The study assessed the impact of changes in solar radiation, air temperature, and wind speed on the performance of photovoltaic systems. The results showed that low-emission scenarios may maintain or improve solar energy potential in some regions.
In contrast, high-emission scenarios lead to a significant decline in this potential due to rising temperatures and changes in solar radiation. The study emphasized the importance of incorporating climate change projections into the planning process for selecting sites for future solar energy projects [8]. Furthermore, forecasts indicate that tropical regions may experience a decline in solar energy potential of between 10% and 15% by the end of this century due to increased cloud cover and rising temperatures, while regions at higher latitudes may see an improvement in productivity of between 5% and 10%. This prediction was based on using 32 CMIP6 climate models under different emissions scenarios, suggesting flexible planning strategies to accommodate future climate risks [9].
SVM, RF, and GBT were selected because they represent three complementary supervised-learning strategies: margin-based classification, bagging, and boosting, respectively. SVM offers strong generalization capabilities and can model nonlinear decision boundaries through kernel functions. RF reduces overfitting by aggregating multiple decision trees and can provide information on predictor importance. In contrast, GBT sequentially corrects previous prediction errors and effectively represents complex nonlinear interactions among environmental and spatial variables. In the present framework, these algorithms are compared based on their ability to reproduce the reference suitability classes generated by the GIS-MCDA stage, rather than being treated as independent validators of physical site suitability.
Previous solar-siting studies in Iraq have successfully applied GIS-based AHP and MCDA to identify suitable land at national and governorate scales [10,11,12]. In contrast, studies in comparable arid regions have extended this direction by combining GIS with machine learning and by examining future renewable-resource variability [5,13]. However, these contributions do not resolve a specific planning gap for Iraq: a national-scale framework is still needed that links projected mid-century climate conditions with spatial site suitability and with the direct effects of temperature and dust on PV efficiency and energy yield. Accordingly, the novelty of the present study is not the simple combination of GIS, MCDA, and machine learning, but the coupling of (i) 2050 climate projections under SSP2-4.5, (ii) a dust- and temperature-corrected PV performance model, (iii) a national GIS-MCDA suitability assessment, and (iv) a comparative SVM, RF, and GBT agreement analysis within one decision-support framework for Iraq. This study aims to integrate projected climatic, environmental, and topographic criteria for 2050 within a unified spatial assessment framework, generate a reproducible MCDA-based reference suitability classification, and evaluate the ability of SVM, RF, and GBT models to reproduce the resulting suitability classes. It also seeks to quantify the effects of temperature and dust on PV efficiency and daily energy production and, ultimately, to identify priority regions for future large-scale solar PV development in Iraq.

2. Previous Studies

Recent reviews of photovoltaic site-selection research show that GIS integrated with multi-criteria decision-making (MCDM) remains the dominant framework because it can combine climatic, topographic, environmental, land-use, and infrastructure criteria in a spatially explicit manner. However, the literature also shows substantial variation in criteria weighting, sensitivity analysis, and validation practice, while machine learning-based and hybrid GeoAI approaches remain less common than conventional GIS–AHP/MCDM workflows [14]. Therefore, the relevant question for the present study is not whether GIS, MCDM, or machine learning have previously been used, but which capabilities have already been demonstrated in Iraq and comparable arid regions and which capabilities remain unintegrated.
In Iraq, Khazael and Al-Bakri [10] applied a national GIS–AHP framework based on climatic, geomorphological, economic, and environmental criteria and reported that approximately 19% of Iraq was optimal for solar farm development. Nozad and Shareef [12] later used GIS–MCDA/AHP around Kirkuk City, combining exclusion constraints with weighted suitability criteria to identify highly suitable areas. At a finer spatial scale, Al-Abadi et al. [15] compared Entropy–TOPSIS and AHP–TOPSIS models in Basrah Governorate and validated the resulting suitability patterns against less-suitable land-use/land-cover categories, identifying about 38% of the study area as highly suitable. Nassif et al. [11] extended national GIS–AHP analysis by discussing climate-change implications and projected temperature increases toward 2100. More recently, Ibrahim et al. [16] linked GIS with unsupervised and supervised machine learning in the Nineveh Governorate using fourteen environmental, climatic, topographic, and socio-economic factors; the boosted decision tree achieved 94.2% accuracy on a 70:30 split of labels derived from the study’s clustering/classification procedure.
Comparable arid-region studies demonstrate additional capabilities. Al Garni and Awasthi [17] developed a GIS–AHP model for Saudi Arabia using technical, economic, climatic, infrastructure, and protected area constraints and found approximately 16% of the country promising for utility-scale PV development. Imam et al. [13] combined GIS-based multi-criteria land-suitability assessment with techno-socio-economic analysis and an SVM model for renewable-resource forecasting in Saudi Arabia; CMIP6 projections were used to assess future resource variability, and the SVM model predicted solar irradiance with R2 = 0.893. In the Cholistan Desert of Pakistan, Ashraf et al. [5] integrated GIS with random forest, XGBoost, and multilayer perceptron models using fourteen conditioning factors and field-derived reference points; random forest achieved an AUC above 0.92, and SHAP was used to interpret feature importance. These studies demonstrate that GIS–MCDM, machine learning, future-climate analysis, and model interpretability have each been used successfully but usually as separate or only partially connected components Table 1.
Table 1. Summarizes the methodological progression and the specific gap.
Existing Iraqi studies are dominated by static GIS–MCDM/AHP suitability mapping, whereas the recent GIS–ML study in Nineveh is locally scaled and does not incorporate a mid-century climate scenario. Conversely, the Saudi study that incorporates CMIP6 uses machine learning primarily to forecast renewable resource variables rather than to couple projected climate conditions with a national suitability classification and an explicit temperature-and-dust PV-performance correction. Because the comparative studies use different datasets, target variables, sample sizes, spatial resolutions, and validation procedures, their reported performance values are not directly comparable and are not used here to rank models or claim numerical superiority. Such numerical benchmarking would require the evaluation of all models on the same dataset using the same target definition and validation protocol. Accordingly, the table focuses primarily on methodological differences, data characteristics, validation design, strengths, and limitations. All comparative studies included in the table correspond to complete references in the reference list. For the present study, we include only the classification metrics defined and reported in the methodology and results (SVM, RF, and GBT accuracy); we do not report regression metrics such as R2, RMSE, and MAE because no separate regression task was performed.

3. Study Area

Iraq is located in the southwestern part of Asia, between latitudes 29° and 38° North and longitudes 38° and 49° East. As shown in Figure 2, the country is characterized by great climatic and geographical diversity, ranging from the mountainous regions in the north and northeast to the floodplains of the Tigris and Euphrates Rivers in the central region and extending to the desert and semi-desert areas in the west and south. Iraq is bordered by Turkey to the north, Iran to the east, Kuwait and Saudi Arabia to the south, Jordan to the southwest, and Syria to the west. Administratively, Iraq consists of 19 governorates, with a total area of approximately 438,317 km2. Iraq was selected as the study area because of its strong solar energy potential, particularly in the western, central, and southern regions, where solar radiation and sunshine duration support large-scale PV power generation [18,19]. Developing solar energy is a strategic response to the growing demand for electricity in Iraq and the need to diversify its energy sources. Vast open lands, especially in desert and semi-desert areas, offer further opportunities for large-scale solar power plants, provided environmental, topographic, climatic, and infrastructural factors are carefully considered [20].
Figure 2. Study area: location and administrative boundaries of Iraq.

4. Research Framework and Methodology

This section describes the methodological approach and materials used to assess the suitability of sites for solar power plants in Iraq. This methodology combines geoinformation systems with machine learning techniques to analyze the spatial variation of factors influencing PV system performance, considering both the direct and indirect effects of these factors on solar power production.
All spatial layers were processed and standardized in ArcGIS and organized into a unified spatial database. The environmental predictors were first transformed into standardized suitability scores and combined through the weighted MCDA procedure to calculate a continuous reference suitability score for each spatial sample. This score was then converted into three reference target classes (Very Suitable, Suitable, and Unsuitable). These MCDA-derived classes were subsequently used as the supervised target labels for SVM, RF, and GBT. Accordingly, the machine-learning stage evaluates how effectively the algorithms reproduce the MCDA-derived classification; it does not independently validate physical site suitability. A mathematical PV-performance model was additionally used to represent the combined effects of high temperature and dust on panel efficiency and energy output. Figure 3 illustrates the overall workflow.
Figure 3. Methodological framework showing GIS harmonization, MCDA reference-class construction, separate SVM/RF/GBT classification, spatial block performance evaluation, GIS-based model agreement/disagreement assessment, and screening-level PV-performance indicators.
These criteria were transformed into a spatial database suitable for suitability scoring and classification. The MCDA stage generated three reference categories—Very Suitable, Suitable, and Unsuitable—to support the identification of priority locations for solar PV development. The subsequent machine-learning stage used these categories as target labels for supervised classification. Therefore, the ML outputs are interpreted as an assessment of model agreement with the MCDA-derived reference pattern rather than as an independent source of ground-truth labels. Table 2 lists the criteria considered for solar-farm selection, together with their data sources and roles in the adopted framework.
Table 2. Criteria selection.

4.1. Data Collection and Preprocessing

The adopted criterion weights were defined as a study-specific, literature-informed weighting scenario. The literature was used to establish the direction and relative priority of the criteria rather than to copy a numerical weighting vector from a single published study. Previous GIS-based PV siting studies support the use and relative importance of solar-resource, climatic, topographic, and site-feasibility criteria [10,11,12,17]. In contrast, PV-performance studies document the strong negative effects of elevated temperature [26,27] and dust deposition [28,29] on photovoltaic efficiency and energy yield. On this basis, solar radiation was assigned the highest weight (35%); temperature and dust were assigned 20% each, slope 15%, and elevation and wind speed 5% each Table 3. The authors adopted these percentages as a structured weighting scenario and did not generate them through AHP, entropy weighting, statistical optimization, or formal expert elicitation. No expert panel was used; therefore, an expert-selection protocol and AHP consistency ratio are not applicable. The adopted percentages are not claimed to represent a unique or objectively optimal weighting solution. To evaluate the robustness of the adopted weighting scheme, a one-at-a-time sensitivity analysis was performed. Each of the six criterion weights was independently varied by −20%, −10%, +10%, and +20% relative to its baseline value. After each perturbation, the complete set of weights was renormalized so that the total remained equal to 100%. The MCDA suitability score was then recalculated for all 671 spatial locations using the same normalized criterion values, while the baseline class boundaries were held fixed. Robustness was quantified as the number and percentage of locations retaining their baseline suitability class under each scenario.
Table 3. Criterion-level suitability interpretation thresholds used for spatial screening.
Spatial and descriptive data for the study area were collected from different sources and processed in ArcGIS. The data included projected solar radiation and temperature for 2050, elevation, land slope, mean wind speed, and dust. All spatial layers were transformed to a unified coordinate reference system, and their values were extracted to a common point database. To construct the continuous reference suitability score used for ML labeling, each predictor was normalized to a 0–100 scale using min–max normalization. For benefit criteria, r = 100(x − xmin)/(xmax − xmin), whereas for cost criteria, r = 100(xmax − x)/(xmax − xmin). Within the observed dataset, solar radiation and mean wind speed were treated as benefit criteria, whereas temperature, dust, slope, and elevation were treated as cost criteria. The normalized criteria were combined using the study-specific weights informed by the cited literature: solar radiation 35%, temperature 20%, dust 20%, slope 15%, elevation 5%, and mean wind speed 5%. Thus, the reference score was calculated as S = 0.35Rsolar + 0.20Rtemperature + 0.20Rdust + 0.15Rslope + 0.05Relevation + 0.05Rwind. Figure 4 shows these weights.
Figure 4. Baseline MCDA criterion weights used in the suitability model.
A final suitability map was produced after spatially comparing the results to determine the extent of similarity or difference between the algorithms. Spatial representation also helped identify the area most suitable for establishing solar farms, particularly those characterized by high solar radiation, relatively lower temperatures, low dust levels, and suitable slopes.
Data related to topographic, climatic, geological, and hydrological features were collected from various sources in different formats, including climate forecasted data for the study area for the year 2050. Maps of the study area and land use maps were also used in both raster and vector data formats, which underwent pre-processing and conversion into a digital format compatible with the requirements of machine learning algorithms, enabling their efficient analysis and classification. Several criteria (physical, social, economic, and environmental) were selected based on local regulations and previously published studies.

4.1.1. Natural Factors

Natural characteristics are among the most important factors influencing the suitability of solar power plant sites because they affect both system performance and long-term site feasibility. These characteristics include land slope, dust conditions, solar irradiation, and elevation above sea level. Figure 5 presents the principal natural characteristics considered across Iraq.
Figure 5. Natural-factor maps used in the analysis: projected annual-average daily solar irradiation for 2050 from ACCESS-CM2 under SSP2-4.5 (NEX-GDDP-CMIP6; 0.25°; kWh/m2/day) and terrain slope derived from the SRTM DEM (30 m; degrees).

4.1.2. Socio-Economic Factors

Socio-economic criteria were used to represent accessibility and infrastructure constraints relevant to utility-scale PV development. Distances from roads and transport infrastructure were considered to support site access and reduce construction, operation, and maintenance requirements, while proximity to transmission networks and populated or urban areas was treated as a planning consideration. Because buffer thresholds vary among jurisdictions and project types, they are treated here as study-specific planning assumptions rather than universal regulatory limits. Figure 6 summarizes the road, railway/airport, and populated-area layers used in the screening analysis. These buffer distances were not included in the sensitivity analysis and should therefore be interpreted as study-specific planning assumptions, not as national regulatory setback distances.
Figure 6. Socio-economic screening layers for Iraq: roads, railways/airports, and populated areas (OpenStreetMap/settlement GIS vector layers; source-layer vintage, no temporal averaging; vector geometry, so raster cell size and physical measurement units are not applicable).

4.1.3. Environmental Factors

Environmental considerations are a key criterion in selecting locations for solar farms. Sites of environmental significance, including protected areas and wildlife migration routes, should be avoided to reduce negative impacts on ecosystems and biodiversity. Waterways and surface-water features were considered as hydrological spatial constraints in the site-selection process. Figure 7 presents the waterways network of Iraq, which was used as a spatial reference for identifying proximity to rivers and other surface-water features and for applying the distance-based water-body constraint reported in Table 4. The waterway layer is therefore treated only as a hydrological reference layer.
Figure 7. Waterways network in Iraq.
Table 4. Shows the exclusion criteria and ranges of buffer zones adopted in selecting solar farm sites.
Land cover and land use are important environmental criteria in choosing locations for solar energy farms because they play a role in reducing environmental impacts and enhancing the sustainability of land use. Therefore, solar projects should be directed towards locations that have the least possible impact on land use, excluding protected areas, wetlands, and forested areas of high ecological value. On the other hand, some countries are adopting policies that encourage integration between solar energy production and agricultural activities, in order to achieve optimal use of land and enhance the efficiency of its investment. Table 4. GIS planning constraints with accepted/preferred conditions and excluded/less-preferred conditions. These rules are separate from the criterion-level suitability thresholds in Table 3 and were not used as machine-learning predictors or target-class thresholds. The distance bands in Table 4 are interpreted as study-specific planning rules rather than uniform statutory setbacks. Proximity to roads and transmission lines represents an accessibility or grid-connection preference, whereas the urban-area, water-body, and protected-area distances are treated as exclusion/protection zones. Comparable Iraqi GIS-MCDA studies also apply distance-based infrastructure and environmental screening [12]. Where no single national Iraqi solar-farm setback standard was identified, the adopted numerical distance is therefore reported explicitly as a planning assumption rather than as a mandatory regulatory threshold.

4.1.4. Climate Data

Predictive climate data for 2050 were used to assess the future impacts of climate change on the productivity of solar photovoltaic systems and to anticipate the extent to which expected climate changes will affect their performance efficiency. The analysis was based on the Combined Socio-Economic Pathway SSP2-4.5, which represents an average greenhouse gas emissions scenario and is widely used in assessing future impacts of climate change. Climate variables were extracted from mid-century climate model outputs, with 2050 as the reference year for analysis.
We obtained climate projections from the NEX-GDDP-CMIP6 dataset [22] using the ACCESS-CM2 global climate model under the SSP2-4.5 scenario. We used the r1i1p1f1 ensemble member and the gn grid configuration. Climate data for the calendar year 2050 were extracted at a spatial resolution of 0.25°, resulting in 671 grid-cell centers located within the Iraqi national boundary. For surface downwelling shortwave radiation (rsds), daily values for the 365 days of 2050 were aggregated to monthly means and then summarized as an annual-average daily solar-irradiation indicator. Accordingly, values reported in kWh/m2/day represent daily solar irradiation averaged over the year rather than an annual irradiation total. We used the same climate-model realization and scenario consistently for the projected climatic variables. No multi-model climate ensemble was constructed; therefore, the present analysis represents the ACCESS-CM2 projection rather than the mean of several CMIP6 models. The analysis used the projected 2050 values directly and did not calculate climate anomalies relative to a separate historical baseline period. Consequently, inter-model climate uncertainty is not explicitly quantified and should be considered when interpreting the projected suitability patterns. The climate variables included surface downwelling shortwave radiation (RSDS; CMIP6 variable rsds), which represents the downward shortwave radiative flux reaching the land surface and provides the solar-resource input for PV generation. Mean near-surface air temperature (TEM), representing air temperature at approximately 2 m above the surface, was used as a national-scale climatic screening variable; maximum and minimum temperature fields were used to characterize heat extremes and thermal variability. These air-temperature variables are not treated as direct measurements of PV-cell or module operating temperature. Under SSP2-4.5, the ACCESS-CM2 projection for 2050 provides a consistent scenario-specific basis for examining spatial variation in solar resource and thermal conditions. Integrating RSDS with near-surface air-temperature information therefore supports comparative screening of future PV suitability while avoiding interpretation of the climate variables as site-specific engineering measurements. For spatial sampling density, the spatial sampling locations were not generated using random or stratified spatial sampling. Instead, they correspond to the centers of the 0.25° × 0.25° NEX-GDDP-CMIP6 grid cells located within the national boundary of Iraq, resulting in 671 spatial sampling points. The sampling density was therefore determined by the native spatial resolution of the climate dataset, and no additional spatial densification was applied. The remaining environmental and topographic predictor values were extracted to these common spatial locations within the GIS environment to construct the machine-learning dataset. GIS harmonization and alignment were performed in WGS 84 (EPSG:4326). The Iraqi national boundary defined the analysis extent, and the native 0.25° × 0.25° NEX-GDDP-CMIP6 grid was retained as the common spatial framework and alignment reference. No new interpolation-based raster grid was generated for the machine-learning table; instead, continuous environmental and topographic values were extracted at the same 671 climate-grid cell centers. Categorical land-cover and exclusion layers were used as spatial masks/constraints rather than interpolated continuous predictors. Accordingly, bilinear or nearest-neighbor resampling was not applied to generate the 671-point classifier dataset.
Accordingly, the present analysis should be interpreted as a 2050 scenario snapshot rather than as a climatological mid-century mean. Interannual variability was not examined, and climatic conditions in the calendar year 2050 may not represent the average conditions of the surrounding decades.

4.1.5. Target-Class Construction and Machine-Learning Dataset

The final MCDA reference score was calculated for 671 spatial samples. To obtain three balanced target classes for supervised learning, the score distribution was divided at the 33.33rd and 66.67th percentiles. The corresponding class boundaries in the revised 671-sample dataset were approximately 62.41 and 68.64 on the 0–100 scale. Samples below 62.41 were labeled Unsuitable; samples from 62.41 to below 68.64 were labeled Suitable; and samples at or above 68.64 were labeled Very Suitable. This procedure produced 223 Unsuitable samples, 224 Suitable samples, and 224 Very Suitable samples. These boundaries were applied deterministically to every sample; Table 3 provides the complementary criterion-level interpretation used for spatial screening. The 671 labeled samples were partitioned using a stratified 70:30 train-test split so that class proportions were preserved. The training set contained 470 samples, and the held-out test set contained 201 samples. Specifically, the training set contained 156 Unsuitable, 157 Suitable, and 157 Very Suitable samples, while the test set contained 67 samples from each class. Geographic coordinates, point identifiers, the continuous reference suitability score, and target-code fields were excluded from the predictor set. The six classifier inputs were the normalized criterion-specific suitability variables corresponding to solar radiation, projected annual temperature, dust, slope, elevation, and mean wind speed. The fixed stratified 70:30 split was retained only as a secondary comparison, whereas five-fold spatial block cross-validation was adopted as the primary model-validation procedure. Because the target classes were generated from the same six environmental criteria subsequently supplied to SVM, RF, and GBT, the ML task is intentionally interpreted as a surrogate classification/agreement analysis. Consequently, accuracy, precision, recall, F1-score, and confusion matrices measure agreement with the MCDA-derived reference classification and must not be interpreted as independent validation of real-world solar-farm suitability. Independent validation would require external ground-truth information such as operational solar-farm locations, field observations, or independently labeled suitable and unsuitable sites. Accordingly, the ML stage complements rather than replaces direct MCDA mapping: it is used to assess the spatial transferability and robustness of the MCDA-derived class structure across structurally different nonlinear classifiers. Agreement among SVM, RF, and GBT identifies more stable screening zones, whereas disagreement highlights locations that warrant additional site-specific investigation.

4.1.6. Machine-Learning Preprocessing and Reproducibility

The supervised classifiers were implemented in the KNIME Analytics Platform as support vector machine (SVM), random forest (RF), and gradient-boosted trees (GBT) models; XGBoost was not used as a classifier in this study. Before model training, the six predictor fields were checked for completeness, and no missing values were present in the 671-sample machine-learning table; therefore, no missing-value imputation or row deletion was required. The final classifier inputs consisted of six normalized suitability predictors scaled to the [0, 1] range, corresponding to solar radiation, temperature, dust, slope, elevation, and wind speed. No additional normalization was applied within the final KNIME workflows. Point_ID, longitude, latitude, the continuous MCDA reference score, spatial-block/fold identifiers, Split_70_30, and all target/code fields were excluded from the predictor matrix to prevent identifier, spatial, or target leakage. The fixed stratified 70:30 split was retained only as a secondary comparison, while the five-fold spatial block cross-validation described below was used as the primary validation procedure. For SVM, a radial basis function (RBF) kernel was used with an overlapping penalty parameter C = 100 and the KNIME RBF width parameter sigma = 0.10. Because the KNIME SVM Learner parameterizes the RBF width using sigma, this value is reported exactly as implemented rather than relabeled as gamma. SVM hyperparameters were selected using an internal stratified 80:20 split of the original 470-sample training subset (random seed = 42). Candidate C-sigma combinations were evaluated on this internal validation subset; C = 100 and sigma = 0.10 yielded the best validation performance and were then retained unchanged for the spatial cross-validation. For RF, 200 trees were used with the Information Gain Ratio split criterion and a fixed random seed of 42. Tree depth was not explicitly limited, while the minimum node-size and feature-sampling settings were left at their KNIME defaults; no separate RF hyperparameter search was performed. For GBT, 100 boosting models were used with a maximum tree depth of 4 and a learning rate of 0.10. All six predictors were available to each tree without attribute subsampling; row subsampling was disabled. Midpoint splitting for numeric predictors was enabled, and a fixed random seed of 42 was used Table 5. No separate GBT hyperparameter search was performed. The final RF and GBT configurations were also kept fixed across all five spatial folds. No supplementary point-level dataset, fixed-split file, or exact geographic coordinates are provided with this article. Instead, reproducibility is supported by reporting the six predictor definitions, the MCDA weighting and class-construction rules, the stratified 70:30 and five-fold spatial-block assignment procedures, the final model settings, the pooled confusion matrices, and the KNIME workflow structure used in the analysis. The spatial sampling framework can be regenerated by selecting the centers of the native 0.25° × 0.25° NEX-GDDP-CMIP6 grid cells that fall within the national boundary of Iraq in WGS 84 (EPSG:4326). The fixed stratified 70:30 split is retained only as a secondary benchmark, and no supplementary split file is distributed.
Table 5. Final KNIME model configuration and hyperparameter settings.
Spatial block cross-validation. To reduce the influence of spatial autocorrelation on model evaluation, a five-fold spatial block cross-validation procedure was implemented. The 671 sampling locations were grouped into geographically separated, non-overlapping 1° × 1° spatial blocks based on longitude and latitude. All observations belonging to the same spatial block were assigned to the same fold, preventing locations from the same block from being simultaneously included in the training and testing subsets. Longitude and latitude were used only to construct the spatial blocks and were not included as predictor variables. In each iteration, four folds were used for model training and the remaining fold was used for testing. The same predictor set and fixed model configurations were maintained across all five folds. Model performance was summarized using the mean and sample standard deviation across folds. The original stratified 70:30 split was retained only as a secondary comparison. After completing the five iterations, the out-of-fold predictions from the five spatial test folds were pooled so that each of the 671 locations contributed once to the overall spatial evaluation. These pooled predictions were used to calculate the confusion matrices, class-specific precision, recall, and F1-score, together with Macro-F1 and balanced accuracy. Because neighboring grid cells can share similar environmental conditions, a random train–test split can yield optimistic apparent accuracy when spatial autocorrelation is present; therefore, random-split performance is not treated as the primary validation result. The 671 locations formed 56 unique spatial blocks. Fold assignment was generated using a stratified group five-fold procedure (StratifiedGroupKFold; n_splits = 5, shuffle = True, random_state = 42), with Spatial_Block_ID as the grouping variable and the MCDA-derived reference class as the stratification label. This kept every 1° × 1° block entirely within a single fold while approximately balancing class composition across folds. The resulting fold labels were fixed and used identically for SVM, RF, and GBT.

4.1.7. Spatial Consensus, Intersection, and Model-Agreement Procedure

After completing the five-fold spatial block cross-validation, the pooled out-of-fold predictions from SVM, RF, and GBT were matched one-to-one by Point_ID across all 671 sampling locations and linked back to their corresponding spatial coordinates in GIS. This ensured that every location contributed exactly once to the three-model agreement assessment; no sampling location was removed because of GIS matching, exclusion masking, or missing model output. Full agreement was defined as all three classifiers assigning the same suitability class; partial agreement was defined as any two classifiers agreeing while the third differed; and complete disagreement was defined as all three classifiers assigning different classes. Only full-agreement locations were retained in the final consensus suitability map, whereas partial- and complete-disagreement locations were treated as non-consensus areas. The integrated map therefore represents GIS-based spatial consensus among the three pooled out-of-fold model predictions rather than majority voting, weighted voting, averaged class probabilities, stacking, or a fourth trained classifier. Accordingly, no separate classification accuracy is assigned to the consensus map, and the agreement categories are interpreted as model agreement/disagreement rather than probabilistic uncertainty.

4.2. Screening-Level Model of Temperature and Dust Effects on PV Performance

Environmental factors such as temperature and dust accumulation are critical for the efficiency and performance of PV solar panels. High solar radiation contributes to increasing the production of photovoltaic energy by providing a greater amount of energy falling on the panels. However, the high temperature of the solar cells limits their efficiency as a result of the decrease in open circuit voltage and the increase in internal losses, which leads to a decline in the overall performance of the system [26,27]. Temperature is one of the main factors affecting the performance of photovoltaic panels, as their efficiency decreases at a rate ranging between 0.3% and 0.5% for every rise of one degree Celsius above the reference temperature (25 °C), with this rate varying depending on the type of photovoltaic cells and the technology used [27,30]. Dust accumulation is a major factor affecting the performance of photovoltaic systems, especially in arid and semi-arid regions. The deposition of particles on the surface of solar panels reduces the transmittance of solar radiation to photovoltaic cells, which negatively affects the production of electrical energy [29,31]. The degree of performance degradation depends on a number of factors, including the density of dust, its physical and chemical properties, and the duration of its accumulation on the surface of the panels. In areas where dust storms are frequent, energy production losses may exceed 20% due to decreased energy conversion efficiency and increased need for cleaning and maintenance operations [28,29]. Dust deposition on the surface of photovoltaic panels contributes to an increase in their operating temperature due to the absorption of part of the incident solar radiation, which leads to aggravation of thermal losses and a decrease in energy conversion efficiency [28,31]. The combined effect of high temperatures and dust accumulation also has a cumulative effect on the performance of photovoltaic systems. Dust reduces the amount of solar radiation reaching cells, while high temperatures reduce the efficiency of converting solar energy into electrical energy [27,29]. Studying these factors is gaining increasing importance in evaluating the current and future performance of photovoltaic systems, especially in light of climate change scenarios that indicate rising temperatures and increased frequency of dust storms in many regions, including the Middle East and Iraq [26]. Understanding these effects contributes to the development of effective mitigation and adaptation strategies, which include adopting appropriate cooling technologies, establishing periodic panel cleaning programs, selecting photovoltaic materials that are more tolerant of harsh environmental conditions, as well as improving the design of systems to enhance their efficiency and operational reliability [30,32]. Based on this, this research aims to evaluate the individual and combined effects of dust accumulation and high temperatures on the efficiency and productivity of photovoltaic solar panels, which contributes to improving their performance and supporting future planning for solar energy projects in dry and semi-arid environments. These equations were applied to the climatic and spatial data for Iraq to estimate PV performance under the adopted modeling assumptions. The calculations used the annual-average daily solar-irradiation indicator (kWh/m2/day), the projected annual mean air temperature, and the MERRA-2 DUEXTTAU dust optical-depth index together with the reference constants listed in Table 6. DUEXTTAU was treated as a relative indicator of spatial atmospheric dust exposure rather than as a direct measurement of material deposited on the PV-module surface. Because the adopted coefficients were not calibrated against site-specific measured PV soiling or performance data, the resulting efficiency and energy-yield values are interpreted as illustrative, scenario-dependent modeled estimates rather than field-validated observations or precise soiling losses.
Table 6. Constants and reference parameters used in the screening-level PV-performance calculations.
The temperature coefficient was used to represent the sensitivity of the screening-level PV-efficiency indicator to the adopted temperature input relative to the 25 °C reference condition. In this national-scale analysis, projected annual mean near-surface air temperature is used as a climatic proxy rather than as a direct estimate of instantaneous PV-cell or module temperature. The temperature-response relation is expressed as follows, consistent with established PV temperature-performance literature [26,27].
Tf = 1 − β(Tproxy − Tref).
By using the values adopted in this study, the above equation becomes
Tf = 1 − 0.004 ∗ (Tproxy − 25),
where Tproxy denotes the projected annual mean near-surface air-temperature input used as the climatic proxy in the screening equations, rather than measured PV-cell or module temperature; Tref = 25 °C, and β = 0.004 °C−1.
In the above equations, the temperature factor represents the temperature-dependent component of the adopted performance model. The dust-response term was represented using DUEXTTAU as a relative atmospheric dust-exposure indicator rather than as a direct measure of deposited dust mass. The exponential response function was applied as shown in Equation (3) [29,33]. The baseline attenuation coefficient was set to k = 1.0 as a normalized scenario parameter, not as a universally calibrated PV-soiling coefficient. To evaluate the dependence of the modeled response on this assumption, k was varied from 0.5 to 1.5 while all other model inputs were held constant.
Dustfactor = e−kD.
After accounting for the adopted temperature and dust response terms, these factors were combined to obtain a modeled PV-efficiency indicator for screening purposes. The basic equation used is [26]
η = ηref[1 − β(Tproxy − Tref)] ∗ e−kD.
By substituting the value in Table 6 in Equation (4),
η = 0.18 ∗ [1 − 0.004 ∗ (Tproxy − 25)] ∗ e−D,
where η represents the modeled PV-efficiency indicator after applying the adopted temperature and dust response terms. It is a screening-level model output rather than a field-calibrated module efficiency. The annual-average daily solar-irradiation input was then converted to an equivalent mean irradiance for the subsequent power calculation:
G avg = solar − Annual − avg − kWh m 2 day ∗ 1000 24 .
Average electrical power per unit module area was calculated from the modeled efficiency indicator, module area, and equivalent mean irradiance using:
P = ηAGavg.
By substituting Equations (5) and (6) into Equation (7),
P = 0.18 [ 1 − 0.004 T proxy − 25 ] e − D solar − Annual − avg − kWh m 2 day ∗ 1000 24 ,
where P represents the modeled power indicator in W/m2 after applying the adopted temperature and dust response terms; therefore, this quantity is a screening-model indicator rather than a field-calibrated module-power prediction. The modeled daily energy-yield indicator per square meter was then calculated using the modeled PV-efficiency indicator and the daily solar-irradiation input. The general formula is
Edaily = η ∗ A ∗ solar − Annual − avg − kWhm2day ∗ 1000.
This value is interpreted as a modeled daily electrical-energy indicator (Wh/m2/day) under the adopted temperature and dust-response assumptions. It is used for relative spatial comparison among candidate locations and should not be interpreted as a field-calibrated estimate of module soiling loss or engineering-grade plant yield.
Edaily = 0.18[1 − 0.004(Tproxy − 25)] ∗ e−D ∗ solar − Annual − avg − kWhm2day ∗ 1000.
The above equations provide screening-level modeled PV-performance indicators for relative spatial comparison. Solar radiation, projected near-surface air temperature, and relative dust exposure are combined to derive illustrative PV-efficiency, power, and daily energy-yield indicators under the adopted assumptions. These quantities are not interpreted as the measured module efficiency, field-calibrated power output, or engineering-grade plant yield.

5. Results

The results of this study indicate that the integration of GIS and machine-learning methods provides a spatial screening framework for identifying candidate areas and estimating modeled PV-performance indicators in Iraq for 2050. The SVM, RF, and GBT outputs showed high agreement with the MCDA-derived suitability classes and were used to characterize the corresponding spatial suitability patterns. The intersection of the results from the three algorithms showed that the most suitable areas are clearly concentrated in the western and southwestern parts of Iraq, where large areas, relatively high solar radiation, and suitable spatial conditions are available for the construction of large-scale solar farms. In contrast, unsuitable locations were more clearly identified in parts of the central and southeastern regions, a finding often associated with high temperatures, increased dust impact, or less favorable topographic, environmental conditions or land use. The concentration of priority areas in western and southwestern Iraq is broadly consistent with earlier Iraqi GIS-based assessments, although the spatial patterns are not identical. At the national scale, Khazael and Al-Bakri [10] reported the strongest suitability mainly in southern, southeastern, and some western parts of Iraq, providing partial spatial agreement with the western and southern tendency identified here. At the governorate scale, Al-Abadi et al. [15] reported a substantial share of highly suitable land in Basrah, supporting the broader suitability of parts of southern Iraq. In contrast, Nozad and Shareef [12] identified highly suitable areas mainly in northern and northwestern Kirkuk, while Ibrahim et al. [16] identified highly suitable locations within Nineveh. These differences are expected because the studies were conducted on different spatial scales and used different criteria, weighting schemes, exclusion constraints, spatial resolutions, and climatic reference periods. The present study also incorporates projected 2050 climatic conditions under SSP2-4.5, whereas most earlier Iraqi assessments used current or static climatic conditions. Accordingly, differences in the exact locations of the highest-priority zones are interpreted as methodological and temporal differences rather than direct contradictions among the studies. The efficiency-correction model illustrates how the adopted temperature and dust inputs influence the calculated PV-performance indicators. The mean projected air temperature was approximately 33.3 °C and the mean DUEXTTAU index approximately 0.20; under the baseline assumptions these produced a modeled mean PV-efficiency indicator of about 14.3% and a mean daily energy-yield indicator of about 872 Wh/m2/day. These values are illustrative, scenario-dependent outputs of the deterministic screening model rather than field-calibrated performance measurements. Within these assumptions, solar irradiation, temperature, and relative dust exposure should be considered jointly when comparing candidate locations. In terms of model performance, five-fold spatial block cross-validation provided a more conservative assessment of geographic generalization than the original random split. SVM achieved the highest mean spatial accuracy (91.46 ± 2.90%), followed by RF (84.03 ± 2.14%) and GBT (82.22 ± 4.47%). The corresponding mean Cohen’s kappa values were 0.872 ± 0.044, 0.760 ± 0.032, and 0.733 ± 0.067, respectively Table 7. These spatially separated results are treated as the primary validation results, whereas the original stratified 70:30 results are retained only for comparison. The spatial validation metrics quantify agreement with the MCDA-derived reference classes and should not be interpreted as independent validation of physical site suitability because the reference classes were constructed from the same environmental criteria used as ML predictors.
Table 7. Five-fold spatial block cross-validation accuracy for SVM, RF, and GBT.
To complement overall accuracy, classification performance was also evaluated at the class level using the pooled spatial out-of-fold predictions. Because every location served as a test observation in exactly one spatial fold, the pooled evaluation contains all 671 samples without reusing any observation within the same test assessment. Table 8 summarizes the pooled overall metrics, Table 9 reports class-specific precision, recall, and F1-score, and Table 10 presents the corresponding pooled confusion matrices.
Table 8. Pooled five-fold spatial classification performance for SVM, RF, and GBT.
Table 9. Class-specific precision, recall, and F1-score from pooled spatial out-of-fold predictions.
Table 10. Pooled spatial out-of-fold confusion matrices for the three classifiers.
SVM provided the strongest and most balanced class-level performance, with F1-scores of 0.944, 0.878, and 0.922 for the Unsuitable, Suitable, and Very Suitable classes, respectively. RF and GBT also maintained useful class discrimination, although their lower Macro-F1 values indicate greater difficulty in reproducing the middle suitability class under geographically separated testing.
These results show that the spatial validation assesses performance across all three suitability classes rather than relying on overall accuracy alone. The class-level metrics and confusion matrices are therefore used as the primary classification-performance evidence together with the fold-wise spatial accuracies. The decision-support value of the machine-learning stage therefore lies in testing whether the MCDA-derived class structure can be reproduced by structurally different nonlinear learners in geographically separated areas. Cross-model agreement provides a robustness and consensus indicator for preliminary prioritization, whereas disagreement identifies less stable locations that may require additional site-specific investigation. This diagnostic role complements, rather than replaces, direct MCDA mapping and is not interpreted as external ground-truth validation. The MCDA weight-sensitivity analysis presented in Table 11 showed generally high classification stability under moderate perturbations. Across the ±10% scenarios, 87.33–99.25% of the 671 locations retained their baseline class. Wider ±20% perturbations produced a retention range of 73.03–98.96%. The largest response occurred for temperature, whereas elevation and wind-speed perturbations produced the smallest changes. Solar radiation, dust, and slope showed intermediate sensitivity. Overall, these results indicate that the adopted MCDA classification is reasonably robust to moderate weight changes while remaining most sensitive to the higher-impact climatic and topographic criteria.
Table 11. MCDA weight-sensitivity analysis: percentage of locations retaining the baseline suitability class.
Overall, the 2050 spatial analysis indicates that priority areas combine high solar radiation, low slope, lower dust levels, and lower modeled temperature penalties in the PV-performance screening indicator. Figure 8 reports signed Pearson correlation coefficients (r) for the principal spatial and environmental variables (No. = 671), preserving both the strength and direction of the relationships. Solar radiation showed a strong negative association with latitude (r = −0.84), indicating a pronounced latitudinal gradient in solar-resource distribution (Figure 9). Temperature and dust showed a strong positive relationship (r = 0.81), consistent with the co-occurrence of hot and dry conditions that can increase airborne dust levels (Figure 10). Dust and longitude were positively associated (r = 0.78), indicating a marked geographic gradient in dust conditions across Iraq (Figure 11). Temperature and elevation showed a negative association (r = −0.71), with lower-elevation areas tending to record higher temperatures (Figure 12). Wind speed and slope showed comparatively weak associations with most variables. These coefficients are interpreted descriptively rather than as independent inferential evidence. Ordinary correlation tests assume independent observations, whereas neighboring grid cells may exhibit spatial autocorrelation; therefore, conventional p-values and confidence intervals were not emphasized.
Figure 8. Signed Pearson correlation coefficients (r) among the principal spatial and environmental variables (N = 671).
Figure 9. Relationship between annual-average daily solar irradiation and latitude (N = 671).
Figure 10. Relationship between DUEXTTAU and projected annual mean air temperature (N = 671).
Figure 11. Relationship between DUEXTTAU and longitude (N = 671).
Figure 12. Relationship between elevation and projected annual mean air temperature (No. = 671).
The final consensus suitability map shown in Figure 13 was generated by matching the pooled out-of-fold SVM, RF, and GBT predictions by Point_ID for all 671 grid-cell centers and retaining only locations for which all three models assigned the same suitability class. The one-to-one matching was complete for all 671 locations; therefore, no cells were lost between the spatial evaluation and the model-agreement analysis. The final consensus map contains 540 full-agreement locations, whereas the remaining 131 locations with partial or complete disagreement are treated as non-consensus areas. The “Very Suitable” consensus class is concentrated mainly in western Iraq, the “Suitable” class extends through transitional areas of central and northern Iraq and parts of the south, and “Unsuitable” consensus locations are concentrated primarily in central and southeastern Iraq. The spatial consensus is not an additional trained classifier; therefore, no separate classification accuracy is reported for the consensus map. Quantitative predictive performance remains reported for SVM, RF, and GBT individually.
Figure 13. Final 2050 consensus suitability map derived from the GIS intersection of SVM, RF, and GBT class outputs on the common NEX-GDDP-CMIP6 0.25° spatial framework; unit: categorical suitability class.
Quantitative model-agreement analysis was performed on the complete set of 671 pooled out-of-fold predictions. Full three-model agreement was obtained for 540 locations (80.48%), while 130 locations (19.37%) showed partial agreement, with two classifiers assigning the same class and the third assigning a different class. One location (0.15%) showed complete three-way disagreement. Within the partially agreed locations, RF and GBT agreed in 65 cases, RF and SVM in 43 cases, and SVM and GBT in 22 cases. These percentages are calculated relative to all 671 sampling locations, and no location was excluded from the agreement analysis Table 12. The results indicate substantial spatial consistency among the classifiers while also identifying a limited set of non-consensus locations that should be interpreted more cautiously, particularly near suitability-class transitions. Accordingly, these categories are described as model agreement/disagreement rather than formal probabilistic uncertainty.
Table 12. Quantitative agreement among SVM, RF, and GBT spatial classifications.
The 671 MCDA-derived reference locations were approximately balanced by construction: 223 grid-cell centers were Unsuitable (33.23%), 224 were Suitable (33.38%), and 224 were Very Suitable (33.38%). Under the three-model out-of-fold consensus criterion, 540 of the 671 locations (80.48%) received a final consensus class: 186 were Unsuitable (27.72% of all grid-cell centers), 162 were Suitable (24.14%), and 192 were Very Suitable (28.61%). The remaining 131 locations (19.52%) were non-consensus because at least one classifier disagreed. These percentages are simple proportions of the 671 grid-cell centers and should not be interpreted as percentages of Iraq’s land area; no area-weighted GIS calculation was used for these proportions.
The spatial distribution of the modeled daily PV energy-yield indicator across Iraq is illustrated in Figure 14. Under the adopted screening model, the highest modeled values occur mainly in western and northwestern Iraq, reaching approximately 990 Wh/m2/day. This spatial pattern reflects the combined influence of solar resource, projected near-surface air temperature, relative dust exposure, and the other suitability criteria used in the screening framework. Lower modeled values occur mainly in northern and northeastern locations. These values are scenario-dependent screening indicators and should not be interpreted as field-calibrated or engineering-grade predictions of plant energy production. Under the spatially separated evaluation, SVM achieved the highest mean accuracy (91.46 ± 2.90%), followed by RF (84.03 ± 2.14%) and GBT (82.22 ± 4.47%). The lower performance relative to the original random split is expected under geographically separated testing and provides a more conservative estimate of spatial generalization. The three models nevertheless showed substantial agreement with the MCDA-derived reference classes; however, this agreement is not treated as independent ground-truth validation of site suitability.
Figure 14. Modeled daily PV energy-yield indicator for 2050 on the common 0.25° spatial framework; derived from the adopted solar-radiation, temperature-proxy, and dust-response model; units: Wh/m2/day.
Figure 15 illustrates the spatial distribution of the active PV module area required to meet a target daily electrical energy output of 9 MWh/day. Because the calculated PV output is expressed as daily energy yield per unit area (Wh/m2/day), the area calculation is reported as an energy-based screening metric rather than as installed DC or AC power capacity. The required active module area was calculated as follows: To improve readability, Figure 15 encodes the modeled area requirement using a continuous point-color scale and omits individual numeric point labels.
APV = Etarget/Edaily
where APV is the required active PV module area (m2), Etarget = 9,000,000 Wh/day (9 MWh/day), and Edaily is the modeled daily energy-yield indicator (Wh/m2/day). Therefore, locations with a higher modeled daily energy-yield indicator require a smaller active module area to achieve the same daily energy target Figure 16. This calculation represents only the active PV module area, not the total land footprint of a utility-scale PV farm.
Figure 15. Modeled active PV module area required to meet a daily energy target of 9 MWh/day for 2050 across the 671 grid-cell centers on the common 0.25° spatial framework; point color represents the modeled area requirement; units: m2.
Figure 16. Relationship between the modeled daily PV energy-yield indicator (Wh/m2/day) and the active PV module area required to meet a 9 MWh/day screening energy target.
The relationships between the environmental inputs and the modeled PV-performance indicators were interpreted as model-response patterns rather than as independent validation. Because the modeled PV-efficiency and daily energy-yield indicators are mathematically derived from solar radiation, the temperature proxy, and relative dust exposure, the observed associations with these variables are partly imposed by the adopted equations. Figure 17 therefore reports signed Pearson coefficients (r) rather than absolute |r| values (N = 671), so that the direction of each model-response relationship is explicit. Figure 18 dust shows a strong inverse association with the modeled PV-efficiency indicator (r ≈ −0.91) and the modeled daily energy-yield indicator (r ≈ −0.89). Temperature is also inversely associated with the calculated performance indicators, whereas the efficiency and energy-yield indicators are positively associated Figure 19. These coefficients describe the sensitivity of the adopted deterministic screening model to its inputs and should not be interpreted as independent empirical evidence of causality or predictive validity. As with the other spatial correlations, neighboring grid cells may be spatially autocorrelated, so ordinary independence-based significance tests would require caution Figure 20.
Figure 17. Signed Pearson model-response correlations (r) between environmental inputs and calculated PV-performance indicators (N = 671).
Figure 18. Screening-model response of calculated PV-efficiency indicator to DUEXTTAU (N = 671).
Figure 19. Screening-model response of calculated daily PV energy-yield indicator to DUEXTTAU (N = 671).
Figure 20. Association between projected annual mean air temperature and DUEXTTAU (N = 671).
A one-factor-at-a-time sensitivity analysis was conducted for the dust attenuation coefficient k using the 671-point dataset, with k = 0.5, 0.75, 1.0, 1.25, and 1.5 and all other inputs held constant. Relative to the k = 1.0 baseline, the mean modeled daily energy-yield indicator changed by +9.88%, +4.81%, 0.00%, −4.57%, and −8.90%, respectively. Thus, decreasing k by 25% and 50% increased the mean modeled energy response by 4.81% and 9.88%, whereas increasing k by 25% and 50% decreased it by 4.57% and 8.90%, respectively Table 13. The absolute PV-performance values are therefore sensitive to k and are interpreted as illustrative scenario outputs rather than experimentally calibrated soiling losses.
Table 13. Sensitivity of the modeled dust-response term to k (N = 671; baseline k = 1.0).

Study Limitations

  • The climate assessment utilizes a single CMIP6 model (ACCESS-CM2) under SSP2-4.5, which does not capture the full range of uncertainty present in a multi-model climate ensemble.
  • The climate assessment is based on a single calendar year (2050). Therefore, it represents a scenario-specific snapshot rather than a climatological average, and interannual variability was not evaluated. Conditions in 2050 may differ from the mean conditions of the surrounding decades.
  • The PV temperature-response calculation uses projected annual mean near-surface air temperature as a national-scale climatic proxy. Air temperature is not interchangeable with operating module or cell temperature; a Faiman- or NOCT-based thermal model would require temporally synchronized irradiance and wind-speed inputs together with module-specific thermal parameters that are not available in the present national-scale dataset.
  • Dust conditions were assumed to be stationary with respect to the 2050 climate projection, and MERRA-2 DUEXTTAU was used as a relative atmospheric dust-exposure proxy rather than a direct measure of PV-module soiling. The coefficient k = 1.0 was not experimentally calibrated for Iraq; the sensitivity analysis (k = 0.5–1.5) therefore quantifies the dependence of the illustrative PV-performance outputs on this normalized assumption.
  • The MCDA criterion weights are literature-informed and study-specific. The weight-sensitivity analysis showed generally high-class stability under ±10% perturbations, with greater changes for some ±20% perturbations; therefore, the adopted weights should not be interpreted as a unique optimal weighting scheme.
  • Buffer and exclusion thresholds were implemented as distinct GIS planning constraints and were excluded as predictor variables in the SVM, RF, and GBT models.
  • The 0.25° spatial resolution is appropriate for national-scale screening; however, it may not adequately capture local variations necessary for project-level site design.
  • The machine learning target classes were derived from the same environmental criteria used as predictors; consequently, the reported accuracies primarily reflect agreement with the MCDA-derived classes.
  • Independent field measurements, operational PV-farm records, and external ground-truth suitability labels were not available for validation.
  • Therefore, the final maps should be interpreted as outputs for national-scale screening and prioritization, rather than as definitive results for project-level site selection.

6. Conclusions

  • The integration of Geographic Information Systems (GIS), multi-criteria decision analysis (MCDA), and machine-learning algorithms established a consistent framework for mapping solar-farm suitability in Iraq under projected 2050 conditions. The machine-learning models demonstrated strong concordance with the MCDA-derived reference classes.
  • The adoption of a multi-criteria methodology enabled the site assessment to incorporate climatic, environmental, and topographical factors alongside solar irradiance.
  • Spatial intersection of the support vector machine (SVM), random forest (RF), and gradient-boosted trees (GBT) outputs identified consensus suitability zones and highlighted locations where all three algorithms produced consistent classifications. This cross-model consensus is interpreted as a robustness and prioritization indicator rather than as independent external validation of physical site suitability.
  • The adopted framework identified western and southwestern Iraq as priority areas for large-scale solar development by 2050, as these regions combine favorable modeled climatic and environmental conditions within the screening criteria.
  • Incorporating the near-surface air-temperature proxy and relative dust exposure into the screening framework provides illustrative, scenario-dependent PV-performance indicators rather than field-calibrated module-efficiency estimates or engineering predictions of plant output.
  • The model-response analysis showed that the modeled PV-efficiency and energy-yield indicators are sensitive to the adopted dust input and attenuation coefficient, particularly under higher temperature-proxy conditions; this sensitivity is interpreted within the screening model rather than as empirical causality.
This study provides a spatial screening framework to support preliminary planning and investment prioritization for solar energy projects in Iraq. The projected 2050 map should be considered a preliminary prioritization tool rather than independently validated ground truth. Site-level implementation should be preceded by field-based and engineering validation. The 2050 results should therefore be interpreted as a scenario snapshot and not as the climatological mean of the mid-century period.

Author Contributions

Conceptualization, O.F.K. and S.H.A.; methodology, A.A.G. and A.A.; software, O.F.K. and H.E.; formal analysis, A.A.G. and R.S.; investigation, S.H.A. and M.O.; resources, A.A.G. and R.S.; data curation, O.F.K.; writing—original draft preparation, S.H.A. and H.E.; writing—review and editing, G.F.A.-D. and M.O.; supervision, G.F.A.-D. and A.A. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the Deanship of Scientific Research, Vice Presidency for Graduate Studies and Scientific Research, King Faisal University, Saudi Arabia (Grant No. KFU265545).

Data Availability Statement

The point-level predictor table, exact geographic coordinates, MCDA point scores, and derived GIS layers are not publicly released, and no supplementary fixed-split dataset is provided. To support reproducibility, the manuscript reports the predictor definitions, MCDA weighting and class-construction rules, train–test and spatial-fold assignment procedures, model configurations, pooled confusion matrices, and the KNIME workflow structure. The 671-location sampling framework can be regenerated from the centers of the native 0.25° × 0.25° NEX-GDDP-CMIP6 grid cells falling within the national boundary of Iraq in WGS 84 (EPSG:4326). No supplementary point-level prediction file is distributed; instead, Section 4.1.6 reports the exact spatial-fold construction procedure (56 spatial blocks; five-fold StratifiedGroupKFold with shuffle = True and random_state = 42, grouped by Spatial_Block_ID and stratified by the MCDA-derived reference class).

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. World Bank. Macro Poverty Outlook for Iraq: April 2026 (English). Macro Poverty Outlook (MPO) Washington, D.C.: World Bank Group. Available online: https://documents.worldbank.org/en/publication/documents-reports/documentdetail/099614308062626168 (accessed on 10 June 2026).
  2. International Energy Agency. Global Energy Review 2026; IEA: Paris, France, 2026. [Google Scholar]
  3. International Energy Agency. Electricity 2025: Analysis and Forecast to 2027; IEA: Paris, France, 2025. [Google Scholar]
  4. International Energy Agency. Renewables 2025: Analysis and Forecasts to 2030; IEA: Paris, France, 2025. [Google Scholar]
  5. Ashraf, H.A.; Li, J.; Li, Z.; Sohail, A.; Ahmed, R.; Butt, M.H.; Ullah, H. Geographic information system and machine learning approach for solar photovoltaic site selection: A case study in pakistan. Processes 2025, 13, 981. [Google Scholar] [CrossRef] [Scilit]
  6. Peters, I.M.; Buonassisi, T. The impact of global warming on silicon PV energy yield in 2100. In 2019 IEEE 46th Photovoltaic Specialists Conference (PVSC); IEEE: New York, NY, USA, 2019; pp. 3179–3181. [Google Scholar] [CrossRef] [Scilit]
  7. Yin, J.; Molini, A.; Porporato, A. Impacts of solar intermittency on future photovoltaic reliability. Nat. Commun. 2020, 11, 4781. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  8. Liu, C.; Chen, L.; Li, K.; Jin, X.; Chen, X.; Qiao, W.; Liao, H. Historical and future climate changes impact global solar photovoltaic power potential: Role of key meteorological variables. Atmos. Ocean. Sci. Lett. 2025, 18, 100619. [Google Scholar] [CrossRef] [Scilit]
  9. Adigun, P.; Dairaku, K.; Ogunrinde, A.T.; Xue, X. Climate Change Influence on Solar Photovoltaic Energy Production and Its Associated Drivers in CMIP6 Ensemble Projections. J. Geophys. Res. Atmos. 2025, 130, e2024JD042971. [Google Scholar] [CrossRef] [Scilit]
  10. Khazael, S.M.; Al-Bakri, M. The optimum site selection for solar energy farms using AHP in GIS environment, a case study of Iraq. Iraqi J. Sci. 2021, 62, 4571–4587. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  11. Nassif, W.G.; Al-Timimi, Y.K.; Elhmaidi, D. GIS-AHP for Optimal Solar Site Selection: A Case Study of Iraq and Its Implications for Climate Change. Int. J. Sustain. Energy Plan. Manag. 2026, 47, 71–89. [Google Scholar] [CrossRef] [Scilit]
  12. Nozad, A.A.; Shareef, M.A. Solar Energy Farm Site Selection Using GIS Multi-Criteria Decision Analysis Surrounding Kirkuk City. Iraqi J. Sci. 2025, 2, 922–941. [Google Scholar] [CrossRef] [Scilit]
  13. Imam, A.A.; Abusorrah, A.; Marzband, M. Potentials and opportunities of solar PV and wind energy sources in Saudi Arabia: Land suitability, techno-socio-economic feasibility, and future variability. Results Eng. 2024, 21, 101785. [Google Scholar] [CrossRef] [Scilit]
  14. Spyridonidou, S.; Vagiona, D.G. A systematic review of site-selection procedures of PV and CSP technologies. Energy Rep. 2009, 9, 2947–2979. [Google Scholar] [CrossRef] [Scilit]
  15. Al-Abadi, A.M.; Handhal, A.M.; Abdulhasan, M.A.; Ali, W.L.; Hassan, J.J.; Al Aboodi, A.H. Optimal siting of large photovoltaic solar farms at Basrah governorate, Southern Iraq using hybrid GIS- based Entropy-TOPSIS and AHP-TOPSIS models. Renew. Energy 2025, 241, 122308. [Google Scholar] [CrossRef] [Scilit]
  16. Ibrahim, S.J.; Al-Ozeer, A.Z.; Al-Ahmady, K.K. Integrating GIS and AI for Solar Farm Site Selection: A Case Study of Nineveh Governorate–Iraq. J. Rafidain Environ. 2004, 4, 1–26. [Google Scholar] [CrossRef] [Scilit]
  17. Al Garni, H.Z.; Awasthi, A. Solar PV power plant site selection using a GIS-AHP based approach with application in Saudi Arabia. Appl. Energy 2017, 206, 1225–1240. [Google Scholar] [CrossRef] [Scilit]
  18. Al Garni, H.Z.; Awasthi, A. Chapter 2—Solar PV Power Plants Site Selection: A Review, in Advances in Renewable Energies and Power Technologies; Yahyaoui, I., Ed.; Elsevier: Amsterdam, The Netherlands, 2018; pp. 57–75. [Google Scholar]
  19. Hamza, A.A.G. Solar Power System Solution for Iraq; ScienceOpen Preprints: Berlin, Germany, 2024. [Google Scholar] [CrossRef] [Scilit]
  20. Visser, W. The World Guide to Sustainable Enterprise: Volume 1: Africa and Middle East; Routledge: London, UK, 2017. [Google Scholar]
  21. Global Solar Atlas. Solar Iraq Map; The World Bank Group: Washington, DC, USA, 2026. [Google Scholar]
  22. Thrasher, B.; Brosnan, I. NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP-CMIP6). 2025. Available online: https://nex-gddp-cmip6.s3.us-west-2.amazonaws.com/index.html (accessed on 10 June 2026).
  23. Catalog, E.E.D. MERRA-2 M2T1NXAER: Aerosol Diagnostics V5.12.4. 2026. Available online: https://developers.google.com/earth-engine/datasets/catalog/NASA_GSFC_MERRA_aer_2?hl=ar (accessed on 10 June 2026).
  24. U.S. Geological Survey (USGS). EarthExplorer; U.S. Geological Survey: Reston, VA, USA, 2026. Available online: https://earthexplorer.usgs.gov/ (accessed on 10 June 2026).
  25. jhowry. (n.d.). SRTM_DEM_Med [Tile layer]. ArcGIS Online, 2026. Available online: https://www.arcgis.com/home/item.html?id=cd7602eb6b91461d8fc0e34a84771d40 (accessed on 15 May 2026).
  26. Skoplaki, E.; Palyvos, J.A. On the temperature dependence of photovoltaic module electrical performance: A review of efficiency/power correlations. Sol. Energy 2009, 83, 614–624. [Google Scholar] [CrossRef] [Scilit]
  27. Dubey, S.; Sarvaiya, J.N.; Seshadri, B. Temperature dependent photovoltaic (PV) efficiency and its effect on PV production in the world—A review. Energy Procedia 2013, 33, 311–321. [Google Scholar] [CrossRef] [Scilit]
  28. Sarver, T.; Al-Qaraghuli, A.; Kazmerski, L.L. A comprehensive review of the impact of dust on the use of solar energy: History, investigations, results, literature, and mitigation approaches. Renew. Sustain. Energy Rev. 2013, 22, 698–733. [Google Scholar] [CrossRef] [Scilit]
  29. Jaszczur, M.; Teneta, J.; Styszko, K.; Hassan, Q.; Burzyńska, P.; Marcinek, E.; Łopian, N. The field experiments and model of the natural dust deposition effects on photovoltaic module efficiency. Environ. Sci. Pollut. Res. 2019, 26, 8402–8417. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  30. Kalogirou, S.A. Solar Energy Engineering: Processes and Systems; Elsevier: Amsterdam, The Netherlands, 2023. [Google Scholar]
  31. Mani, M.; Pillai, R. Impact of dust on solar photovoltaic (PV) performance: Research status, challenges and recommendations. Renew. Sustain. Energy Rev. 2010, 14, 3124–3131. [Google Scholar] [CrossRef] [Scilit]
  32. Chen, Y.; Wang, D.; Liu, Y.; Dong, Y.; Liu, J. Research on photovoltaic performance reduction due to dust deposition: Modelling and experimental approach. J. Therm. Sci. 2019, 28, 1186–1194. [Google Scholar] [CrossRef] [Scilit]
  33. Zhang, J.; Li, X.; Wang, J.; Qiao, L. CFD–DEM simulation of dust deposition on solar panels for desert railways. Appl. Sci. 2022, 13, 4. [Google Scholar] [CrossRef] [Scilit]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.