The user interface is developed using Tkinter, the standard graphical user interface toolkit for Python. The web map contains multiple interrogation functionalities, allowing users to ‘click’ on a point to identify a location. Upon repositioning the marker to the preferred place, the user selects the “Predict” button to initiate the prediction. This retrieves solar potential data based on the chosen coordinates and then offers location-specific information, including estimated solar irradiance, optimal panel tilt, and energy yield calculations. The design is intended to be transparent to the end user, providing a natural experience while guaranteeing real-time feedback irrespective of the underlying processing complexity, as shown in Fig. 6. Furthermore, solar irradiance is calculated for each location on the world map using future meteorological data by creating scenarios (SSP1-2.6: sustainability scenario (1.5 °C warming), SSP2-4.5: middle road scenario (2.7 °C warming), SSP5-8.5: fossil-fuel development (4.4 °C warming)). Upon user selection of a coordinate (selected coordinates 21.42 N and 78.50 E), the system extracts solar potential data from the predictive map and displays it to the users as illustrated in Fig. 7. However, both the figure generated using Custom Tkinter (Modern GUI framework, Version-5.2.2) and Tkinter Map View (Interactive map widget, Version-1.29) software in Python. The Matplotlib library is used to generate a heat map (Fig. 7). In addition, diverse ML techniques have been employed to forecast solar potential and produce the solar feasibility report. A thorough comparative study was conducted across all implemented models, assessing performance across several indicators and differing geographical and meteorological variables. Table 1 shows the training evolution metrics for XGBoost.
Masked solar potential map (https://matplotlib.org/).
It can be observed from Table 1 that early stopping analysis indicated an optimal termination point at epoch 450, where validation loss stopped improving. Significantly, there were no signs of overfitting, as both the training and validation losses stabilised, and test set performance remained consistent after epoch 400. The patterns of change are evident in the learning curves: initially (epochs 0–150), the RMSE declines rapidly from 2.43 to 1.52; subsequently, during epochs 150–400, the model converges gradually with a low RMSE reduction; and in the saturation phase (epochs 400–500), the RMSE improvement is negligible, indicating that the model has reached its capacity. The tight alignment of validation loss, test loss, and training loss reflects the model’s effective generalisation. Furthermore, the comparative analysis (Fig. 8a) of solar potential prediction using machine learning models indicated a notable enhancement after implementing various algorithmic approaches. LR had moderate predictive capabilities, evidenced by an R² value of 0.71 and an RMSE of 2.14 kWh/m², indicating its effectiveness in early-stage forecasting while highlighting deficiencies in high-accuracy predictions. In contrast, LASSO demonstrated reduced accuracy (R², 0.74; RMSE, 1.98 kWh/m²), showing marginal enhancement of LR overall, although exhibiting superior concordance in high-dimensional meteorological contexts. The little enhancements in performance underscore the limitations of linear models in depicting atmospheric interactions. The enhancement resulted from the transition to ensemble methods. Five hundred estimators of optimised depth-constrained RF achieved an R² of 0.86 and an RMSE of 1.52 kWh/m², demonstrating the RF’s potential capability to capture the entire impacts of cloud cover and atmospheric conditions on solar irradiance heterogeneity. The GBM implementation demonstrated further enhancements, achieving an R² of 0.88 and an RMSE of 1.31 kWh/m², particularly excelling under diverse climatic circumstances. The observed outcomes align with theoretical insights regarding successive ensemble benefits in intricate meteorological interactions. Compared to all tested models, XGBoost demonstrated exceptional results across most measures, achieving an R² of 0.93 and an RMSE of 0.97 kWh/m², indicating much higher predictive performance across various geographical locations and atmospheric conditions. This observation corroborates prior research regarding the algorithm’s efficacy in structured prediction challenges, including non-linear feature interactions. The neural methodologies yielded competitive, if not superior, outcomes. The Multi-Layer Perceptron (MLP) model 9 achieved an R² of 0.85 and an RMSE of 1.55 kWh/m², requiring extensive hyperparameter adjustment to mitigate overfitting related to the temporal characteristics of meteorological data. The LSTM demonstrates potential with R² = 0.87 and RMSE = 1.41 kWh/m², particularly in capturing seasonal and daily radiation patterns, while it remains a computational burden for practical application. It is also observed from Fig. 8b that XGBoost attains the lower values of MSE (kWh/m2) and RAE (%) compared to other algorithms. This study highlights the varying performances of machine learning models in predicting solar potential, underscoring the importance of selecting an appropriate model for a certain forecasting task and feature condition.
Comparative analysis of different ML algorithms.
The comprehensive comparative analysis ultimately reveals XGBoost consistent superior performance across all evaluation metrics, with pronounced advantages in both predictive accuracy and computational efficiency relative to other high-performing models. These findings align with performance characteristics documented in comparative studies, which similarly identified XGBoost exceptional performance for solar radiation prediction across diverse meteorological conditions, establishing it as the optimal modelling approach for the solar potential prediction application. XGBoost employs an iterative gradient boosting technique to compute model weights. The process begins by initialising predictions with the mean GHI, subsequently followed by many boosting iterations during which gradients are computed. Decision trees are constructed based on these gradients, and the forecast is revised according to the formula F(m) = F(m-1) + η × tree(m). The ultimate forecast is achieved by aggregating the results of all the trees. XGBoost autonomously derives features by establishing non-linear decision boundaries via tree splits, approximating interactions through tree depth, and ultimately ranking feature relevance based on gain, which indicates the enhancement in the loss function. These advantages can be attributed to XGBoost excelling with tabular data, such as weather, where sequential dependencies are less significant, unlike time series, where LSTM is superior. Furthermore, XGBoost exhibits greater tolerance for missing or noisy data, offers enhanced interpretability for feature importance, and has superior computing efficiency, training 17 times quicker and predicting 7 times faster. The feature importance derived from the XGBoost model reveals that clear sky irradiance (31.2%), cloud amount (24.8%), day of year (12.3%), temperature at 2 m (8.7%), latitude (7.9%), clearness index (5.4%), surface albedo (3.2%), and relative humidity (2.8%) are the primary variables influencing the prediction of solar global horizontal irradiance (GHI), while other variables collectively account for 3.7% of the contribution. To validate methodological robustness, additional cross-regional analysis across diverse geographical contexts representing varied climatic conditions (Fig. 9), following established protocols. XGBoost consistently demonstrated superior generalization capabilities across all tested regions, with particularly notable performance advantages in regions characterized by high climate variability. This robust cross-regional performance, combined with computational efficiency and exceptional accuracy metrics, establishes XGBoost as the optimal model for integration into the solar potential prediction application.
Regional performance comparison of top models across diverse climatic zones.
The performance analysis of XGBoost is further examined concerning three categories of uncertainty, specifically tropical uncertainty (Mindanao, Philippines), which encompasses aspects such as monsoon fluctuation, typhoon frequency, and ENSO effects. Nevertheless, continental uncertainty (Gobi-Altai, Mongolia) encompasses the aspects of Dust storms, temperature extremes, and snow cover. Mediterranean uncertainty characterizes the Peloponnese region of Greece. Saharan dust, thermal waves, and alterations in precipitation. Table 2 shows the different SSP scenarios that significantly influence solar irradiance estimates under normal conditions and extreme events, with uncertainty range.
An uncertainty analysis produces several significant findings. XGBoost continuously excels (Table 3), even under more extreme conditions, demonstrating tolerance to climatic fluctuations. However, uncertainty rises with extended timeframes and more severe climate scenarios due to the increased pressures of environmental dynamics. Spatial data reveal that the most unreliable places are exclusively tropical, mostly influenced by convective processes, while it also highlights that continental regions possess significant solar energy potential alongside considerable seasonal variability.
Beyond solar potential prediction, the application offers a comprehensive project feasibility assessment report (Fig. 10). This report considers a wide range of factors that influence the viability of solar energy projects, including economic, regulatory, and site-specific constraints. The economic analysis component incorporates current market data on solar panel costs, installation expenses, and projected electricity prices. It also accounts for various incentive schemes and policy frameworks across different regions, allowing for tailored feasibility assessments based on local contexts. Regulatory considerations are integrated through a regularly updated database of solar energy policies and building codes for different jurisdictions. This ensures that feasibility calculations align with current legal and administrative requirements, reducing the risk of project delays or complications. The importance of such regulatory alignment has been highlighted in recent studies on barriers to solar energy deployment. Site-specific constraints are evaluated using geographical information system (GIS) data, including topography, land use patterns, and shading analysis. The application employs advanced algorithms to simulate shading effects from nearby structures and natural features, providing a more accurate assessment of the usable area for solar panel installation. This approach builds on recent advancements in solar resource assessment methodologies. The integration of climate data into the site assessment process represents a significant advancement in project feasibility evaluation. By incorporating high-resolution climate models, the application can account for potential future changes in solar radiation patterns, enhancing the accuracy of long-term performance projections. Furthermore, the application considers social and environmental factors in its feasibility assessments. Recent research has emphasized the importance of community acceptance and environmental impact in the success of renewable energy projects. By incorporating these elements, the application provides a more holistic approach to project evaluation. This comprehensive approach to feasibility assessment, combining technical, economic, regulatory, and socio-environmental factors, aligns with current best practices in sustainable energy development. By providing a multifaceted analysis, the application aims to enhance the accuracy of project viability predictions and contribute to more successful and sustainable solar energy deployment.
A solar project feasibility report for the coordinates 21.42 oN and 78.50 ° E.
Explainable AI (SHAP analysis)
The global feature importance analysis (Table 4) based on SHAP data indicates that clear sky irradiance (31.2%) and cloud amount (24.8%) are the primary determinants of solar production, with their impacts ranging from − 2.8 to + 3.1 and − 2.3 to + 0.8, respectively. Seasonal and spatial factors, such as the day of the year (12.3%) and latitude (7.9%), exert a large influence, indicative of solar geometry, whereas meteorological conditions (6.0% to 8.7% for temperature at 2 m; 5.4% for clearness index) have a minor effect. Additional secondary variables, such as surface albedo, humidity, and precipitation, accounted for merely 5.1% of the overall significance. Interaction effects enhance model complexity, exhibiting the greatest synergy for cloud quantity × clear sky irradiance (8.3%), succeeded by temperature–humidity (3.7%) and day of year–latitude (3.2%). Regional characteristics highlight the dominant climatic factors in the grids: the tropics are shaped by clouds, precipitation, and humidity associated with convection; deserts are affected by irradiance, temperature, and albedo/dust under clear-sky conditions; temperate zones are influenced by the day of the year, clouds, and irradiance, which reflect seasonality; and polar regions are determined by latitude, day of the year, and albedo, as production reacts to extreme solar angles. Explainability tests critically assess whether the model accurately learns physical relationships (e.g., the adverse impact of clouds on irradiance), effectively encodes seasonal cycles, and appropriately weights geographical factors, while ensuring the absence of spurious correlations as indicated by partial dependence plots. This further enhances confidence in the model’s physical consistency and its prediction dependability across various climate regimes.
Additionally, to demonstrate the practical application and versatility of the solar potential forecast tool, three case studies including varied geographical regions, climate zones, and socioeconomic circumstances are provided. These examples highlight how the application’s holistic approach to solar evaluation may guide decision-making across multiple scales and contexts.
Case study 1
Solar Potential Assessment for a Tropical Island Region (Mindanao, Philippines − 7.12°N, 125.48°E).
Mindanao, the southernmost major island of the Philippines, represents an ideal test case for the application, given its tropical climate, increasing energy demands, and government renewable energy targets. Located at 7.12°N, 125.48°E, the region experiences high solar irradiance year-round but faces challenges including frequent cloud cover during monsoon seasons, limited land availability, and vulnerability to typhoons.
Methodology
The XGBoost model was applied to predict long-term solar potential for Mindanao under three CMIP6 climate scenarios (SSP1-2.6, SSP2-4.5, and SSP5-8.5) through 2099. The technical and economic feasibility of a hypothetical 50 MW utility-scale solar installation was then analysed under each scenario, incorporating local policy frameworks and infrastructure constraints.
Results
Climate impact assessment
Under all scenarios, Mindanao is projected to experience a 1–3% decrease in annual solar irradiance by 2050, primarily driven by increased cloud cover and precipitation intensity. The most significant decreases occur during the July-September monsoon period, with potential irradiance reductions of up to 7% under SSP5-8.5.
Technological optimization
The application identified significant advantages for weather-resistant module technologies and advanced drainage systems in this region. The optimum tilt angle calculations suggested a 6.5° tilt facing south, with single-axis tracking systems providing only a 14% energy yield improvement compared to fixed-tilt systems, substantially lower than the global average advantage of 18% due to the region’s proximity to the equator.
Economic feasibility
The levelized cost of electricity (LCOE) analysis demonstrated moderate economic viability across all scenarios, with projected costs declining from $0.041/kWh in 2025 to $0.033–0.037/kWh by 2050 (in 2025 USD), depending on the climate scenario. The application identified potential challenges in the policy environment due to existing regulatory frameworks requiring updates to accommodate large-scale renewable integration.
Case study 2
Solar Potential Assessment for a Semi-Arid Steppe Region (Gobi-Altai, Mongolia − 45.08°N, 96.25°E).
This case study examines the potential for utility-scale solar installations in the Gobi-Altai region of Mongolia (45.08°N, 96.25°E). The region faces challenges of extreme temperature variations, sparse population, limited electrical infrastructure, but benefits from exceptionally high solar radiation and vast available land areas.
Methodology
The application was utilized to assess solar potential across a 200 km² region, evaluating the feasibility of large-scale solar installations under different climate scenarios. The analysis incorporated extreme temperature impacts on panel efficiency, dust accumulation effects, and integration with the national grid expansion plans.
Results
Climate impact assessment
The model projected an increase in annual solar radiation by 2050 under mid-range scenarios, with winter months showing the most significant improvements. The annual direct normal irradiance at this site is 1389 kWh/m²/year.
Temperature impact analysis
The application identified significant seasonal efficiency variations due to temperature extremes, with summer module temperatures potentially reaching 51.4 °C, reducing efficiency.
Technological optimization
For utility-scale applications, the optimum system configuration featured robust bifacial panels with anti-soiling coatings. The application recommended fixed-tilt systems at 38° facing south, if moving systems, then with Winter Optimum Tilt 48° facing south and Summer Optimum Tilt 28° facing south, as tracking systems showed disproportionately higher maintenance requirements due to dust and extreme temperature cycling. Module spacing calculations indicated optimal row spacing of 2.8 times the module height to balance land use efficiency with snow shedding and reduced soiling.
Economic feasibility
The economic analysis has demonstrated projected LCOE values of $0.051/kWh by 2030, and the application also identified significant grid integration challenges, recommending phased development coordinated with transmission infrastructure expansion, potentially incorporating battery storage to address grid stability concerns during periods of curtailment.
Case study 3
Solar Potential Assessment for a Coastal Mediterranean Location (Peloponnese, Greece − 37.36°N, 22.17°E).
This case study explores the application of the tool for sustainable tourism development in the Peloponnese region of Greece (37.36°N, 22.17°E), a Mediterranean coastal area with strong seasonal tourism patterns, heritage protection requirements, and ambitious national renewable energy targets.
Methodology
The application was employed to assess distributed solar potential across a 30 km² area, integrating building data, seasonal energy consumption patterns, and heritage preservation requirements. The analysis included both technical potential assessment and visual impact evaluations, considering the region’s cultural and aesthetic significance.
Results
Climate impact assessment
The model projected stable solar production patterns through 2050 under moderate scenarios, with a slight increase in summer output (2–3%) and negligible winter changes. Notably, the analysis revealed an 87% correlation between peak tourism energy demand and solar generation patterns, creating favourable conditions for high self-consumption rates without extensive storage requirements.
Technological optimization
The application identified 12.4 MW of viable rooftop solar potential across 247 buildings within the study area, with an additional 8.7 MW possible through carport and shade structure integration. The spatial analysis revealed that focusing on large hotel rooftops (> 500 m²) could achieve 65% of the total potential while involving only 18% of the buildings, significantly reducing administrative complexity.
Economic feasibility
The levelized cost of electricity (LCOE) analysis demonstrated moderate economic viability across all scenarios, at about $0.051/kWh. The feasibility assessment demonstrated that achieving 50% solar penetration in the tourism sector by 2030 would require streamlined permitting processes specifically designed for heritage-sensitive regions. The application identified optimal phasing of installations, calculating that focusing first on larger, less visually sensitive properties would achieve the most favourable cost-benefit ratio while establishing implementation protocols for subsequent phases.
