Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Reports volume 16, Article number: 3249 (2026)
887
Metrics details
Highway slope photovoltaic (HSPV) systems represent a significant approach to achieving transportation-energy integration and reducing carbon emissions, as well as a crucial utilization method for renewable energy. Given the complexity of assessing the installable area of photovoltaic (PV) panels on highway slopes across large spatial scales, a regression assessment analysis was conducted using Random Forest. The results demonstrated that as of 2024, the total installable PV area on highway slopes across 14 cities in Guangxi is 0.989 km2. Furthermore, accounting for regional variations in PV system efficiency, the power generation efficiency of HSPV systems in different areas was evaluated by integration of the Analytic Hierarchy Process (AHP) and Entropy Weight Method. Based on these assumptions, the estimated installed capacity of HSPV in Guangxi reaches 155.93 MW, with an annual power generation potential of 169.03 GW · h. After conducting an economic analysis, the project is economically feasible, as evidenced by a net present value > 0, an internal rate of return > 8%, a benefit-cost ratio > 1, and a payback period < 10 years. The proposed integrated framework—combining random forest-based slope-area extrapolation with AHP-entropy efficiency evaluation—remains reliable even without high-resolution imagery, offering replicability and scalability for highway-slope PV assessments. The findings provide actionable guidance for PV siting and economic appraisal in other provinces or linear infrastructures (e.g., railways, transmission corridors).
Globally, distributed photovoltaic (DPV) technology is widely recognized as a core means of achieving sustainable energy development and reducing carbon emissions. The international community has increasingly recognized the importance of assessing PV potential, particularly in scenarios where land resources are limited. The exploration of PV applications in unconventional spaces, such as highway slopes, has emerged as a prominent area of research interest1. As an auxiliary space in highway construction, the traditional functions are limited to slope protection. The deployment of PV systems has transformed them into a renewable energy production base, realizing the “dual utilization” of land resources. China’s highway network is rapidly expanding, with a total mileage of over 160,000 km, providing a unique opportunity for road slope PV integration.
The assessment of large-scale distributed PV potential has been extensively investigated by various researchers. For instance, Qiu et al.2 developed a GIS-MCDM based evaluation framework that integrates land availability, solar radiation and slope to quantitatively assess regional PV potential. Zhu et al.3 combined satellite remote sensing with deep learning to accurately identify rooftop areas suitable for PV deployment at the city scale. Kim et al.4 proposed a two-stage “low-resolution screening–high-resolution actuarial” approach to balance computational efficiency and accuracy when selecting suitable road-related sites. Zhang et al.5 applied machine-learning regression to estimate rooftop areas in multiple Chinese cities and then evaluated the overall rooftop PV potential. In parallel, a number of studies2,3,4,5,10,33,34 have shown that digital elevation models (DEMs), together with derived slope and aspect maps from contour or topographic data, can be combined with empirical equations to perform high-resolution solar resource assessments that explicitly account for geographical effects. More recently, Yifan et al.33 and others3,9,29,30,31,32 have extended these DEM- and MCDM-based approaches to urban-scale PV potential evaluation and multi-criteria siting analysis, providing an important methodological foundation for the present work.
Previous studies obtained detailed surveying and road slope mapping data through high-resolution satellite imaging technology, which was costly in terms of information processing. Thus, it was difficult for most Chinese cities to obtain this data. Most existing urban studies use easily accessible national statistics, such as highway mileage and slope gradient, to indirectly determine slope area. However, the reliability of the assessment depends on the quality of the initial data. More precise, detailed geospatial data is essential for evaluating the potential of HSPV. Therefore, new methods must be developed to address the lack of data on the paved area of highway slopes across urban areas. At the same time, there is no quantitative impact of geographical environment differences on the power generation efficiency of PV systems in the existing large-scale distributed PV power generation potential assessment.
Currently, China’s highway construction is in a period of rapid advancement. Research on high-speed information collection is constantly evolving with the support of new technologies, such as artificial intelligence. These technologies provide new opportunities for precise measurement of highway slopes. In this context, there is a need for methods that can extrapolate the installable area of highway slopes from limited detailed samples to entire regional networks without exhaustive high-resolution mapping, and (ii) can evaluate the power generation efficiency of PV systems in a way that explicitly reflects regional environmental differences. The main contributions of this paper are as follows: (1) A Random Forest model based on machine learning is proposed to evaluate and extrapolate the paved area of HSPV along highway slopes in 14 cities in Guangxi; (2) A comprehensive evaluation model is proposed for the purpose of assessing the power generation efficiency of PV systems in various regions, which integrates the Analytic Hierarchy Process (AHP) and Entropy Weight Method to account for both expert knowledge and data-driven information; (3) Based on the above two models, the power generation potential and installed capacity of HSPV in Guangxi are evaluated, and the economic feasibility under different scenarios is analyzed at the regional scale.
In response to the problem of high accuracy when calculating the laid area of large-scale DPV, the problem of high accuracy is lower than the cost. The proposed study utilizes a random forest model, a machine learning algorithm, to assess the extent of PV pavement implementation on highway slopes. Considering the large-scale potential assessment of existing research, due to the difficulty in quantifying external environmental factors, the numerical values of system efficiency are usually obtained by taking empirical values, and a comprehensive evaluation model is proposed to evaluate the PV power generation efficiency in various regions using AHP and entropy methods. Figure 1 is the technical route of this article.
Research route for the evaluation of power generation potential of HSPV.
To calculate the slope area of the existing paved photovoltaics on the expressway in Guangxi, refer to Fig. 2, which presents a highway distribution map with a total mileage of 8500 km. Establish a random forest regression model based on machine learning method. Select the elevation, slope, direction, and type of the slope as the characteristic variables of the model, and the slope area as the target variable.
Highway distribution map and sampling locations for highway-slope PV evaluation in Guangxi; major cities (e.g., Nanning, Liuzhou) are labeled for reference.
Slope elevation
Elevation is a basic indicator that reflects the topography and topography, which directly affects the mechanical equilibrium conditions of the slope6. In mountainous expressway scenarios, slopes of different elevation sections face different risks of geological disasters7. Elevation directly affects solar radiation flux. According to 8. Jogunuri et al.8, in the evaluation of PV system efficiency, the average atmospheric transmittance increases by 1.2–2.5% for every 100 m of altitude increase. Therefore, elevation is a key variable in predicting the generation potential per unit area.
Slope gradient
The slope directly affects the inclination angle of the PV panel, thereby changing the solar radiation reception efficiency. Research in 9. Balaska et al.9 shows that the performance differences in PV modules with different inclinations at the same geographical location are significant. Steep slopes may increase the difficulty and cost of PV array installation, and may easily cause the risk of soil erosion or landslide. The analysis of GIS-AHP site selection model10 shows that the gradient of more than 15 ° will significantly increase the construction difficulty and maintenance cost, and may cause the risk of water and soil loss. An empirical study in Ontario, Canada, shows that11, when the slope is less than 2°, the rainwater accumulation leads to a 40% increase in the surface pollution rate of PV panels, and a 19% increase in the snow cover time; A minimum slope of 2° is recommended to ensure natural drainage. Therefore, the slope gradient of 2°~15° is the appropriate slope for the laying of PV on the highway slope.
Slope orientation
Solar irradiance exerts a direct influence on the power generation efficiency of PV modules, with the direction of the slope determining the incident angle of sunlight. Traditional PV systems in the northern hemisphere are optimally oriented with the southern position (180°). Barbón et al.12 research shows that when the azimuth is from southeast to southwest (135°–225°), the energy loss is usually less than 5%. That is, the suitable azimuth range of PV module for Expressway Slope in the northern hemisphere is 135°–225°.
Slope type
The geologic structure and stability mechanisms of different types of slopes vary significantly. Figure 3 presents a schematic representation of the three types of slopes. Fill-type slopes are formed by the accumulation of external fills, exhibiting spatial variability in soil compactness, grain size distribution, and drainage properties. These slopes are susceptible to uneven settlement13. Cut-type slopes are formed through the excavation of natural strata, and their stability is contingent upon the shear strength of the native rock and soil mass. Conversely, fill-cut slopes are distinguished by the simultaneous execution of fill and excavation processes. Cut-type slopes are typically more precipitous than fill-type slopes, and the installation of PV modules on cut slopes can result in glare issues that may compromise vehicle safety. Consequently, the optimal locations for installing PV modules are the filled portions of filled and cut slopes.
Types of highway slopes. (a) fill-type; (b) Cut-type; (c) fill-cut.
The sample data collected in 14 cities in Guangxi were normalized and divided into two data sets: 80% as the training set and 20% as the test set. Subsequently, the random forest model, which is predicated on the machine learning method, is employed for regression analysis. The fundamental steps involved in this process are as follows:
Sampling14: The training dataset S is utilized to generate K datasets through a process of sampling and replacement. This process is repeated to construct a decision tree for each dataset;
Expanding: Within each sub-node, m features are selected at random from the set of M attributes, with the selection based on the Gini coefficient;
The hyperparameters of the model were tuned using grid search15 and cross-validation methods to obtain optimal model performance;
Prediction: For a novel dataset, the mean of all decision tree predictions16 is taken as the final output;
The new dataset is predicted using the trained model, and the average of all the decision tree predictions is taken as the final output.
The performance of random forests is highly sensitive to the selection of hyperparameters, including the number of trees, tree depth, and the minimum number of samples that are required for node splitting, among other factors. The manual adjustment of these parameters is both inefficient and prone to yielding suboptimal results. Consequently, this paper employs grid search to optimize the random forest parameters. Grid search is performed by defining a hyperparameter space that contains candidate values for multiple hyperparameters and then evaluating all possible combinations of these hyperparameters. A systematic approach is employed to identify the most effective model configuration by searching for predefined combinations of hyperparameters. This process enhances the model’s performance.
This paper organizes the data into 14 sample areas. Appendix B Fig. B1 shows the administrative divisions of Guangxi, using Baise, China as an example. Figure 4 illustrates the correlation between characteristic variables and slope area. The coefficient of determination of the training set of the model is calculated to be 0.89, and the average ab-solute error is 684.5 m2. The coefficient of determination of the test set is 0.84, and the average absolute error is 742.7 m2. It is evident that the correlation between the feature variables that are selected by the model and the slope area is enhanced, and the pre-diction misprediction is reduced.
Relationship between features, error distribution and extrapolation model performance. (a) Calculated importance (MDI) by summing the impurity reduction of statistical features across all split nodes; (b) training set error distribution; (c) test set error distribution.
In order to verify whether the random forest model is the optimal choice for this study, this paper utilizes the four models of decision tree, linear regression, K-nearest neighbor, and random forest to conduct regression analysis on the sample areas of 14 cities. The specific results are shown in Table 1. A thorough examination of the data presented in Table 1 reveals that the random forest model exhibits superior performance in comparison to the other models examined in this study. The superior performance of the model in this experiment can be attributed to its tuning to the hyperparameters through grid search and 10-fold cross-validation.
Most existing studies estimate power generation using fixed efficiency coefficients, ignoring the dynamic impact of environmental factors during operation17,18. The present study developed a comprehensive assessment model for PV system efficiency, with the objective of evaluating the potential of large-scale distributed PV power generation. The model integrates tomography and entropy analyses to systematically evaluate five environmental factors: solar radiation, temperature, relative humidity, wind speed, and rainfall. However, this paper does not address the impact of shadowing on power generation efficiency.
Solar radiation
Solar radiation is the primary driving force behind the power generation of PV systems, and its intensity directly determines the efficiency with which pho-tons excite electron-hole pairs. Numerous research experiments have proven that solar radiation is significantly and positively correlated with power generation output. The study referenced in Nakamoto et al.19 used monthly input–output data to confirm that solar radiation is a core input factor. Lim et al.20 further emphasizes that the unpredictability of solar radiation leads to intermittent power generation and directly affects grid stability.
Temperature
The operating temperature of PV cells directly impacts their electro-chemical performance. Temperature effects are the primary limiting factor in intrinsic physical processes. Efficiency losses are more pronounced in monocrystalline silicon and polycrystalline silicon PV panels at high temperatures21. According to the findings of the experimental data set, a 1 °C rise in temperature has been shown to result in a decrease of 0.4% to 0.5% in efficiency. This decline can be attributed to two primary factors: increased charge recombination and decreased carrier mobility22. Additionally, high temperatures accelerate the aging of backplane materials, which affects long-term reliability23.
Relative humidity
In high-humidity environments, water molecules penetrate the interior of PV modules, leading to electrode corrosion, aging of encapsulation materials, and in-creased surface leakage current. Preliminary findings suggest a positive correlation between humidity levels and the annual efficiency reduction rate of silicon-based modules. Specifically, it has been observed that when humidity exceeds 60%, there is an increase in the efficiency reduction rate of approximately 0.5% to 1.2% per year24. The adsorption of water molecules on the PV surface results in the formation of an insulating layer, thereby reducing the efficiency of heat dissipation. In environments characterized by high humidity, the temperature rise effect is known to be amplified. Research indicates that when relative humidity surpasses 80%, the operational temperature of the module increases by 3–5°C25. High humidity has been demonstrated to reduce atmospheric transmittance and enhance solar radiation scattering. In tropical regions, where relative humidity frequently exceeds 70%, direct radiation is known to decrease by 15–20%, resulting in a power generation loss of 10–15%26.
Wind speed
Increased wind speed has been demonstrated to exert a considerable influence on the temperature of PV panels, which can consequently result in a diminution in the efficiency of the constituent components. Preliminary experiments have demonstrated a positive correlation between wind speed and component temperature, with a documented decrease of approximately 2–3 °C for every 1 m/s increase in wind speed27. When wind speeds exceed 5 m/s, dust deposition is known to decrease. Research has demonstrated that in regions with high wind speeds, dust deposition is reduced by 37%, and efficiency losses are decreased by 9.3%28.
Rainfall
Rainfall has been shown to remove dust and pollutants accumulated on the sur-face of PV panels, thereby improving light transmittance and enhancing power generation efficiency. According to experimental studies, the intensity of rain-fall must reach a certain level to effectively remove dust. As indicated by Yao et al.29, when the rainfall exceeds 8 mm in a single event, it has been demonstrated to remove over 90% of the dust from the component surface. Furthermore, frequent light rain has been demonstrated to be more conducive to maintaining component cleanliness than occasional heavy rain30. However, the presence of rainwater on the panel surface results in the formation of a water film, which has been observed to increase light scattering and reflection losses, particularly in high-latitude regions where this effect is more pronounced31. In regions characterized by low levels of pollution and moderate latitudes, where annual rainfall exceeds 500 mm and the frequency of rainfall exceeding 8 mm per event surpasses 10 times per year, natural cleaning methods can effectively substitute for manual cleaning, thereby achieving a net positive benefit32. The Guangxi is situated within a subtropical climate zone, characterized by an annual precipitation average of 1500 mm and more than 30 precipitation occurrences per year.
There exists a multitude of methodologies for the calculation of evaluation indicators’ weights; the most widely utilized is the AHP. The AHP33 is a decision-making methodology that involves the systematic decomposition of complex problems into a series of factors. The factors are grouped into different levels based on their subordinate relationships, and a multi-level structural model is established. This model comprises target layers, criterion layers, and factor layers. Therefore, the AHP method was selected to determine the weights of evaluation indicators during the zoning process.
Construct a judgment matrix
To this end, a 1–9 scale was employed to make pair-wise comparisons between the significance of two factors at the same level to the higher level34, thereby constructing the judgment matrix A = (aij)m×n.
The calculation of the weighting coefficient
The judgment matrix A is to be normalized by columns;
The sum vector (bar{w}_{i}) is obtained by adding A row by row;
The next step is to normalize vectors (w_{i}) and (bar{w}_{i}) to obtain weight coefficients.
Consistency check
In order to ascertain the validity of the weighting results, it is imperative to assess the convergence of the constructed judgment matrix35. The test formula is as follows:
It is imperative to calculate the maximum Eigenvalue, λmax, of the judgment matrix, A;
The consistency index (CI) is determined by the following formula;
The consistency ratio (CR) can be calculated in the following way.
Where RI a quantitative metric that quantifies the randomness of the data. The specific values of the RI are provided in Table 2. In instances where CR is lower than 0.1, Preliminary analysis indicates that the judgment matrix meets the required consistency standards. Conversely, in cases where CR is greater than 0.1, the judgment matrix must undergo reconstruction until the consistency requirements are met.
Entropy, a physical quantity originating from thermodynamics, is defined as the conversion of heat into work. Subsequently, this concept was incorporated into the field of information theory, where it was employed as a metric to quantify the uncertainty associated with information. The degree of chaos in information distribution within a system is directly proportional to entropy. That is to say, the smaller the entropy, the more uniform the information distribution. Conversely, the larger the entropy, the more chaotic the distribution36. The entropy weight method has been demonstrated to objectively ascertain the extent to which each indicator affects system security, thereby37 rendering the indicator weight values more objective.
The m experts score n factors to form a m×n-order raw evaluation matrix R;
Standardized processing of indicators, the calculation formula for positive indicators is as follows:
The formula for calculating negative indicators is as follows:
The characteristic weight of the factor j in the section i must be calculated.
Calculate the information entropy ej of the evaluation indicator j.
Calculate entropy weight.
The assessment of the power generation potential of HSPV can provide a direct guide for the optimization of power supply structures for transportation facilities. It has been demonstrated that loads such as highway tunnel lighting and service area power consumption exhibit temporal and spatial stability. This property enables HSPV system to achieve “self-generation and self-consumption.” The result of this phenomenon is an increase in the penetration rate of green energy, as well as the promotion of sustainable resource development and the reduction of grid transmission losses and expansion pressures.
In this study, the utilization rate of the slope area that is available for installation is assumed to be 80%, and the conversion efficiency of PV modules is set at 21.3%. The utilization rate of the slope area available for installation is contingent up-on several factors, including vegetation coverage requirements, the frequency of extreme weather events, construction and maintenance costs, monitoring system density, land ownership boundaries, and shadows and obstacles.
The potential installation capacity, Pi, of the HSPV system is expressed as follows:
where Pr is the rated power per unit area of PV panels; S is the area of PV modules that can be installed on the highway; η is the actual utilization rate of the slope area available for installing PV arrays.
The annual power generation Pw of the HSPV system is expressed as:
where Si represents the area of the i-th slope, Gi denotes the annual surface solar radiation received by the i-th slope, Ci is the overall efficiency of the PV system corresponding to the i-th slope, K represents the conversion efficiency of the PV panels, and n is the total number of slope grids.
The highway network layout and sampling framework have been illustrated in Fig. 2; therefore, the following provincial-scale maps (Figs. 5, 6, 7 and 8) focus on the spatial distribution of installable area, system efficiency, installed capacity and annual power generation across Guangxi, rather than repeating the detailed road geometry.
The model obtained from the training of 14 sample areas suggests that the highway slope is divided into several grids to predict the non-sample areas. The technically feasible pavement area of HSPV in Guangxi has been calculated to be 0.989 km2 under the baseline assumption that 80% of the available slope area can be effectively used for PV installation. To better illustrate the impact of different utilization assumptions, Fig. 5 presents a set of provincial-scale multi-panel maps showing the spatial distribution of effective installable area under three utilization-rate scenarios (60%, 70% and 80%). Highway corridors and candidate slopes are highlighted according to the available feasibility constraints (slope, aspect and land-use), while the full provincial background remains visible, which may still include non-installable areas and is treated as a limitation in the Discussion. Major cities such as Nanning and Liuzhou are labeled to aid interpretation.
Provincial-scale distribution of effective installable area on highway-adjacent slopes in Guangxi under different utilization-rate scenarios: (a) 60%, (b) 70% and (c) 80%; highway corridors and candidate slopes are highlighted based on available feasibility criteria, while the full provincial background remains visible and may include non-installable areas; major cities are labeled to aid interpretation.
To guarantee the scientific integrity of the assessment results, the indicator weights were determined using a combination of the AHP and entropy weight method38. The subsequent section delineates the calculation process for the combination weight.
The calculation is expressed as follows:
where wi denotes the subjective weight of each indicator in the analytic hierarchy process, wj signifies objectivity of the indicator’s weight within the entropy weight method, α represents the coefficient of variation, 0 ≤ α ≤ 1, and δi is used to denote the comprehensive weight of the indicator.
The comprehensive weight is subject to variation in accordance with changes in α. In consideration of extant research findings and the prevailing circumstances of the evaluation indicator system in this paper, α = 0.5 is adopted as the baseline setting, which assigns equal importance to the AHP and entropy weights when computing the comprehensive weights shown in Table 3. In the Guangxi, the factors influencing PV power generation efficiency are multifaceted. Among these factors, the annual average solar radiation has the greatest influence, accounting for 30.96% of the total variance in efficiency. The annual average temperature also has a relatively high weight in the model, accounting for 20.77% of the total variance in efficiency. This finding suggests that the direct influencing factors are the most significant factors affecting PV power generation efficiency. A simple sensitivity check is conducted by varying α from 0.2 to 0.8 in increments of 0.1, and the resulting comprehensive weights change only slightly while the ranking of indicators remains stable; these results are summarized in Table 4.
According to the comprehensive assessment model previously referenced, the distribution of power generation system efficiency on HSPV in the Guangxi can be determined. The baseline efficiency values in this region are concentrated between 78% and 80.5%, suggesting that the overall efficiency of PV systems on highway slopes in Guangxi is relatively high, thereby rendering them suitable for PV system operations. To further account for the influence of shading on system performance, several uniform shadow-loss scenarios (10%, 20% and 30% additional losses) are introduced by applying the corresponding loss factors to the baseline efficiency index. Figure 6 presents a set of multi-panel maps that visualize the resulting efficiency distributions under these shadow-loss assumptions, thereby providing a more conservative view of effective system performance in environments such as urban sections with overpasses and sound barriers, mountainous terrain and vegetation-rich segments. In addition, a sensitivity analysis of the combined AHP–entropy weights was conducted by varying the parameter α and perturbing individual indicator weights, and the relative ranking of cities was found to be robust to these changes.
Provincial-scale distribution of PV power generation system efficiency for highway-adjacent slopes in Guangxi under different shadow-loss scenarios: (a) 10% additional losses, (b) 20% additional losses and (c) 30% additional losses; highway corridors and candidate slopes are highlighted based on available feasibility criteria, while the full provincial background remains visible and may include non-installable areas; major cities are labeled to aid interpretation.
According to the relevant parameters of PV modules in Appendix A Table A1, initial computations suggest that, under the baseline utilization assumption (80%), the installed capacity of PV power generation on highway slopes in Guangxi is 155.93 MW, with an estimated power generation potential of 169.03 GW·h. Building on the utilization-rate framework introduced in Sect. 2.3, we further examine how the spatial distribution of installed capacity and annual power generation changes under different utilization scenarios (60%, 70% and 80%) while keeping the technically feasible slope area unchanged.
Provincial-scale distribution of PV installed capacity on highway-adjacent slopes in Guangxi under different utilization-rate scenarios: (a) 60%, (b) 70% and (c) 80%; highway corridors and candidate slopes are highlighted based on available feasibility criteria, while the full provincial background remains visible and may include non-installable areas; major cities are labeled to aid interpretation.
Figure 7 shows multi-panel maps of the distribution of PV installed capacity on highway-adjacent feasible slopes under the three utilization-rate scenarios. In each panel, the same feasible slope grids as in Fig. 5 are used, but the installed capacity is scaled according to the assumed utilization rate (60%, 70% or 80%). This visualization highlights how conservative design choices that lower the utilization rate reduce the apparent concentration of installed capacity, especially in cities with extensive highway mileage such as Nanning and Liuzhou.
Provincial-scale distribution of annual PV power generation potential on highway-adjacent slopes in Guangxi under different utilization-rate scenarios: (a) 60%, (b) 70% and (c) 80%; highway corridors and candidate slopes are highlighted based on available feasibility criteria, while the full provincial background remains visible and may include non-installable areas; major cities are labeled to aid interpretation.
Figure 8 presents analogous multi-panel maps for the annual power generation potential under the 60%, 70% and 80% utilization-rate scenarios. These maps are derived by combining the scenario-specific installed capacities with the efficiency distributions discussed in Sect. 3.2, while treating shadow losses through the uniform loss factors illustrated in Fig. 6. Together, Figs. 5, 7 and 8 provide a consistent spatial picture of how the effective installable area, installed capacity and annual power generation respond to different utilization assumptions, and how these interact with the shadow-loss scenarios considered in the subsequent economic analysis.
Conducting an economic feasibility analysis for a PV project typically requires the use of a series of standard economic indicators, including net present value (NPV), internal rate of return (IRR), payback period, and benefit-cost ratio (BCR). These indicators can comprehensively assess the project’s profitability, investment efficiency, and risk tolerance. A compendium of related parameters is enumerated in Appendix A Table A1.
Initial investment:
where Ct is the initial total investment; Pi is the installed capacity; uc is the unit price of PV modules; Cs is other costs such as mounting brackets, inverters, and installation.
Annual maintenance cost:
where CO is the annual operation and maintenance cost of the entire PV system. Pi is the installed capacity. Pu is the power of a single PV panel. uo is the annual operation and maintenance cost of a single PV panel.
Power generation and attenuation:
where Pt is the electricity generation in the tenth year. Pw is the electricity generation in the first year. µa is the average annual degradation coefficient of PV panels from the second year onward.
Net present value (NPV):
where CFt is the net cash flow in year t. r is the discount rate. n is the project life. Ct is the initial total investment.
The internal rate of return (IRR) is the discount rate that results in the project’s NPV being zero. The IRR for the project is 11.3%, which exceeds the 8% threshold for capital cost, thereby indicating that the project is financially viable.
The dynamic payback period is a crucial metric in evaluating the financial viability of a project, as it provides a quantitative assessment of the time required to recoup the initial investment. The point in time at which the cumulative discounted net cash flow reaches zero. The following is the year in which the cumulative net cash flow becomes positive: The eighth year of the program has been reached.
Benefit-cost ratio (BCR):
where Pt is the electricity generation in year t; ue is the unit price of PV grid connection; n is the project life; Ct is the initial total investment; CO is the annual operation and maintenance cost of the entire PV system.
The calculation yielded the following result: The BCR is 1.72, which is greater than 1. This indicates that the project has significant economic viability.
In addition to the baseline case, which assumes that 80% of the technically feasible slope area can be effectively equipped with PV modules (as visualized in Figs. 5 and 7(c) and 8(c)), a simple scenario analysis was carried out by varying the assumed utilization rate (e.g., 60%, 70% and 80%) while keeping the technically feasible area itself and other parameters unchanged. In practice, these utilization-rate scenarios are implemented by proportionally scaling the installed capacity and annual generation relative to the 80% baseline case, rather than by modifying the underlying spatial distribution of feasible slope grids. The results show that both NPV and IRR decrease as the utilization rate is reduced; however, under moderate utilization levels (around 70%), the project still remains economically feasible with NPV > 0, IRR above the capital cost threshold and BCR greater than 1. This indicates that the economic conclusions are robust to reasonable variations in the utilization rate assumption. Furthermore, when the shadow-loss scenarios considered in Sect. 3.2 (Fig. 6) are combined with conservative utilization-rate assumptions, the economic indicators decrease but remain acceptable under moderate loss conditions, while under extreme loss conditions the feasibility becomes marginal in the most severely shaded segments.
It should also be noted that, in the economic calculation, the revenue from PV generation is based on the assumption that most of the electricity can be either self-consumed by nearby highway facilities or exported to the grid under current tariff schemes, without explicitly modeling curtailment or additional grid connection infrastructure costs. In practice, temporal mismatches between local demand and PV output, curtailment risks under high PV penetration and the need for grid reinforcement may reduce the effectively monetized energy and economic benefits. These aspects will be investigated in more detail in future techno-economic studies.
This study addresses the issues of low accuracy and high cost in calculating the potential area for large-scale distributed PV installations. The study puts forth a novel proposition of employing a random forest model, a machine learning algorithm, to evaluate the prospective areas suitable for PV installations on highway slopes. The findings of the simulation experiments demonstrate that the random forest model exhibits a marked superiority in terms of performance when compared to decision trees, linear regression, and k-nearest neighbor models, particularly in the context of regression analysis. This observation is based on the analysis of sample data from 14 cities.
This study proposes an integrated assessment model to address the limitations of existing studies that rely on fixed efficiency coefficients in large-scale potential assessments. This model integrates tomography analysis and entropy methods to systematically assess the impact of five environmental factors—solar radiation, temperature, relative humidity, wind speed, and rainfall—on the efficiency of PV systems. In comparison with conventional empirical methodologies, this model has been demonstrated to enhance the precision of assessments concerning power generation potential, while concomitantly addressing the challenge of quantifying environmental factors.
According to the aforementioned model, the calculated PV installed capacity for highway slopes in the Guangxi is 155.93 MW, with a power generation potential of 169.03 GW ·h. This assessment result provides a foundation for achieving energy self-sufficiency on highways. Following a thorough economic analysis, all key economic indicators met the established criteria, thereby affirming the economic feasibility of the project. These metrics included an NPV greater than zero, an IRR exceeding 8%, a BCR greater than one, and a payback period of less than ten years. It is also worth noting that the Random Forest model used to extrapolate the installable slope area was validated using an 80/20 split of samples drawn from the same set of 14 cities. Because nearby slopes along the same highway corridor may share similar characteristics, spatial autocorrelation cannot be fully eliminated, and the reported test performance may be optimistic when the model is transferred to completely new regions. Therefore, the current model is mainly intended for application within Guangxi under conditions similar to the training data, and further external validation using samples from deliberately excluded cities or other provinces will be required in future work to better assess generalizability.
However, this study is subject to two significant limitations:
Firstly, it should be noted that shadow effects are not included. The impact of shadow losses, attributable to nearby vegetation or terrain, has not been systematically assessed, a factor that may result in an overestimation of actual power generation.
Secondly, there is a paucity of system efficiency validation. The accuracy of the comprehensive assessment model has not been validated using actual power plant operational data.
Subsequent research endeavors should employ high-resolution remote sensing and 3D shadow simulation in conjunction with real-world operation data feedback to optimize the model within a closed-loop system. Furthermore, these efforts should investigate the potential of wind power, PV, and energy storage systems to enhance the reliability of transportation energy systems through a synergistic approach.
In this study, a machine learning-based evaluation framework was developed to assess the power generation potential and economic feasibility of highway slope photovoltaic (HSPV) systems in Guangxi, China. The framework integrates a Random Forest regression model for extrapolating the installable slope area, an AHP–entropy-weight model for evaluating the efficiency of PV systems under different environmental conditions, and an economic analysis based on standard financial indicators.
First, the Random Forest model was trained using detailed samples from 14 cities and then applied to the entire highway network to estimate the paved slope area that can host PV installations. The model achieved high accuracy on both the training and test sets, demonstrating clear advantages over decision tree, linear regression and K-nearest neighbor models for this regression task.
Second, the AHP–entropy-based comprehensive efficiency model was constructed to quantify the influence of solar radiation, temperature, relative humidity, wind speed and rainfall on PV system performance. The results indicate that annual average solar radiation and temperature are the most influential factors, and a sensitivity analysis of the indicator weights confirmed that the relative ranking of cities is robust to moderate changes in the weighting scheme.
Third, combining the estimated installable area and system efficiency, the total installed capacity of HSPV in Guangxi was estimated to be 155.93 MW, with an annual power generation potential of 169.03 GW·h under baseline assumptions. The subsequent economic assessment, including NPV, IRR, payback period and BCR, shows that HSPV projects on highway slopes are economically feasible under a range of scenarios, thereby supporting the concept of “self-generation and self-consumption” for transportation facilities.
Finally, this work highlights several directions for future research. In particular, more detailed modeling of shadow effects and validation using real operational data from highway-slope PV plants are needed to further improve the accuracy and applicability of the proposed framework. In addition, the methodology can be extended to other regions and combined with multi-energy systems, such as wind power and energy storage, to support the coordinated planning of low-carbon transportation infrastructure.
The datasets analysed during the current study are available in the zenodo repository, https://zenodo.org/records/16948274. All data generated during this study are included in this published article and its supplementary information files.
Rahdan, P., Zeyen, E. & Victoria, M. Strategic deployment of solar photovoltaics for achieving self-sufficiency in Europe throughout the energy transition. Nat. Commun. 16, 6259. https://doi.org/10.1038/s41467-025-61492-9 (2025).
Article CAS PubMed PubMed Central Google Scholar
Qiu, T., Wang, L., Lu, Y. & hang, M. in, W., ang, S., ang, L. Potential assessment of photovoltaic power generation in China. Renew. Sustain. Energy Rev. 154, 111900 (2022). https://doi.org/10.1016/j.rser.2021.111900
Zhu, R. et al. Multi-sourced data modelling of spatially heterogenous life-cycle carbon mitigation from installed rooftop photovoltaics: A case study in Singapore. Appl. Energy 362, 122957. https://doi.org/10.1016/j.apenergy.2024.122957 (2024).
Article CAS Google Scholar
Kim, B., Han, S., Heo, J. & Jung, J. Proof-of-concept of a two-stage approach for selecting suitable slopes on a highway network for solar photovoltaic systems: A case study in South Korea. Renew. Energy 151, 366–377. https://doi.org/10.1016/j.renene.2019.11.021 (2020).
Article Google Scholar
Zhang, Z. et al. Carbon mitigation potential afforded by rooftop photovoltaic in China. Nat. Commun. 14, 1. https://doi.org/10.1038/s41467-023-38079-3 (2023).
Article CAS ADS Google Scholar
Xie, H., Dong, J., Deng, Y., Dai, Y. & Sun, G. Prediction model of the slope angle of rocky slope stability based on random forest algorithm. Math. Probl. Eng.. 1–10(2022). (2022). https://doi.org/10.1155/2022/9441411
Huang, F. et al. Slope stability prediction based on a long short-term memory neural network: Comparisons with convolutional neural networks, support vector machines and random forest models. Int. J. Coal Sci. Technol. 10(1), 18–14. https://doi.org/10.1007/s40789-023-00579-4 (2023).
Article CAS Google Scholar
Jogunuri, S. F. T., Stonier, J., Peter, A. A., Jayaraj, G. & Ganji, J. S. J. Random forest machine learning algorithm based seasonal multi-step ahead short‐term solar photovoltaic power output forecasting. IET Renew. Power Gener. 19(1). https://doi.org/10.1049/rpg2.12921 (2025).
Balaska, A., Tahri, A., Tahri, F. & Stambouli, A. B. Performance assessment of five different photovoltaic module technologies under outdoor conditions in Algeria. Renew. Energy 107, 53–60. https://doi.org/10.1016/j.renene.2017.01.057 (2017).
Article CAS Google Scholar
Jiang, W. et al. Evaluation method for the availability of solar energy resources in road areas before route corridor planning. Appl. Energy 356, 122260. https://doi.org/10.1016/j.apenergy.2023.122260 (2024).
Article Google Scholar
Hailu, G. & Fung, A. S. Optimum Tilt angle and orientation of photovoltaic thermal system for application in greater Toronto Area, Canada. Sustainability 1(22), 6443. https://doi.org/10.3390/su11226443 (2019).
Article ADS Google Scholar
Barbón, A., Bayón-Cueli, C., Bayón, L. & Rodríguez-Suanzes, C. Analysis of the Tilt and azimuth angles of photovoltaic systems in non-ideal positions for urban applications. Appl. Energy 305, 117802. https://doi.org/10.1016/j.apenergy.2021.117802 (2022).
Article Google Scholar
Paudyal, P., Dahal, P., Bhandari, P. & Dahal, B. K. Sustainable rural infrastructure: Guidelines for roadside slope excavation. Geoenviron. Disasters 10(1), 11–14. https://doi.org/10.1186/s40677-023-00240-x (2023).
Article Google Scholar
Khan, K., Ahmad, W., Amin, M. N. & Deifalla, A. F. Investigating the feasibility of using waste eggshells in cement-based materials for sustainable construction. J. Mater. Res. Technol. 23, 4059–4074. https://doi.org/10.1016/j.jmrt.2023.02.057 (2023).
Article Google Scholar
Ogunsanya, M., Isichei, J. & Desai, S. Grid search hyperparameter tuning in additive manufacturing processes. Manuf. Lett. 35, 1031–1042. https://doi.org/10.1016/j.mfglet.2023.08.056 (2023).
Article Google Scholar
Wang, Z., Zhang, J., Ashish, T. & Asutosh He Zhang. Sustainable development study of biomass new energy in rural buildings based on LCA-emergy-carbon footprint and machine learning methods. Energy Sustain. Dev., 88, 101811 (2025). https://doi.org/10.1016/j.esd.2025.101811
Nakamoto, Y., Eguchi, S. & Takayabu, H. Efficiency and benchmarks for photovoltaic power generation amid uncertain conditions. Socio Econ. Plan. Sci. 94, 101971. https://doi.org/10.1016/j.seps.2024.101971 (2024).
Article Google Scholar
Yao, W. et al. Power generation evaluation of solar photovoltaic systems using radiation frequency distribution. J. Building Eng. 98 https://doi.org/10.1016/j.jobe.2024.110981 (2024).
Nakamoto, Y. & Eguchi, S. How do seasonal and technical factors affect generation efficiency of photovoltaic power plants? Renew. Sustain. Energy Rev. 199 https://doi.org/10.1016/j.rser.2024.114441 (2024).
Lim, S. C., Huh, J. H., Hong, S. H., Park, C. Y. & Kim, J. C. Solar power forecasting using CNN-LSTM hybrid model. Energies 15(21), 8233 (2022). https://doi.org/10.3390/en15218233
Sun, C., Zou, Y., Qin, C., Zhang, B. & Wu, X. Temperature effect of photovoltaic cells: A review. Adv. Compos. Hybrid. Mater. 5(4), 2675–2699. https://doi.org/10.1007/s42114-022-00533-z (2022).
Article Google Scholar
Kaldellis, J. K., Kapsali, M. & Kavadias, K. A. Temperature and wind speed impact on the efficiency of PV installations. Experience obtained from outdoor measurements in Greece. Renew. Energy 66, 612–624. https://doi.org/10.1016/j.renene.2013.12.041 (2014).
Article Google Scholar
Hudișteanu, V. S., Cherecheș, N. C., Țurcanu, F. E., Hudișteanu, I. & Romila, C. Impact of temperature on the efficiency of monocrystalline and polycrystalline photovoltaic panels: A comprehensive experimental analysis for sustainable energy solutions. Sustainability 16(23), 10566. https://doi.org/10.3390/su162310566 (2024).
Article ADS Google Scholar
Santhakumari, M. & Sagar, N. A review of the environmental factors degrading the performance of silicon wafer-based photovoltaic modules: Failure detection methods and essential mitigation techniques. Renew. Sustain. Energy Rev. 110, 83–100. https://doi.org/10.1016/j.rser.2019.04.024 (2019).
Article CAS Google Scholar
Jo, H. H., Kim, J. & Kim, S. Enhancing the power generation performance of photovoltaic system: Impact of environmental and system factors. Appl. Therm. Eng. 240 https://doi.org/10.1016/j.applthermaleng.2023.122221 (2024).
Ramli, M. A. M. et al. On the investigation of photovoltaic output power reduction due to dust accumulation and weather conditions. Renew. Energy 99, 836–844. https://doi.org/10.1016/j.renene.2016.07.063 (2016).
Article Google Scholar
Kaushik, S. C., Rawat, R. & Manikandan, S. An innovative thermodynamic model for performance evaluation of photovoltaic systems: Effect of wind speed and cell temperature. Energy. Conv. Manag. 136, 152–160. https://doi.org/10.1016/j.enconman.2017.01.011 (2017).
Article ADS Google Scholar
Salimi, H., Mirabdolah Lavasani, A., Ahmadi-Danesh-Ashtiani, H. & Fazaeli, R. Effect of dust concentration, wind speed, and relative humidity on the performance of photovoltaic panels in Tehran. Energy Sources Part A Recov, Utili. Environ. Effects. 45(3), 7867–7877 (2019). https://doi.org/10.1080/15567036.2019.1677811
Yao, W. et al. New models for the influence of rainwater on the performance of photovoltaic modules under different rainfall conditions. Renew. Sustain. Energy Rev. 173 https://doi.org/10.1016/j.rser.2022.113119 (2023).
Jo, W., Ham, N., Kim, J. & Kim, J. The cleaning effect of photovoltaic modules according to precipitation in the operation stage of a large-scale solar power plant. Energies 16(17), 6180. https://doi.org/10.3390/en16176180 (2023).
Article Google Scholar
Yao, W. et al. The impact of spectral distribution on photovoltaic power generation and its quantitative evaluation model. Appl. Energy 358, 122581. https://doi.org/10.1016/j.apenergy.2023.122581 (2024).
Article Google Scholar
Zhao, B., Zhang, S., Cao, S. & Zhao, Q. Cleaning cycle optimization and cost evaluation of module dust for photovoltaic power plants in China. Clean Technol. Environ. Policy 21(8), 1645–1654. https://doi.org/10.1007/s10098-019-01731-y (2019).
Article Google Scholar
Yifan, J., Li, Z., Jianghua, W. & Meng, W. Assessment of the photovoltaic potential at urban level based on parameterization and multi criteria decision-making (MCDM): A case study and new methodological approach. Energy. Sustain. Dev. 83, 101585. https://doi.org/10.1016/j.esd.2024.101585 (2024).
Article Google Scholar
Ashour, M. & Mahdiyar, A. A comprehensive state-of-the-art survey on the recent modified and hybrid analytic hierarchy process approaches. Appl. Soft Comput. 150, 111014. https://doi.org/10.1016/j.asoc.2023.111014 (2024).
Article Google Scholar
Can Yıldırım, B., Karakaya, G. & Gönül, M. S. Fibonacci Series-Based pairwise comparison scale for analytic hierarchy process. Int. J. Inform. Technol. Decis. Mak. 20(03), 959–986. https://doi.org/10.1142/S0219622021500243 (2021).
Article Google Scholar
Tianlin, D., Jianzhong, G., Fang, W. & Renjian, Z. Application of entropy-based multi-attribute decision-making method to structured selection of settlement. J. Vis. Commun. Image Represent. 58, 220–232. https://doi.org/10.1016/j.jvcir.2018.11.026 (2019).
Article Google Scholar
Wang, L., Huang, W., Zhao, C., Hu, Y. & Cui, S. Exploring the environment-nutrition-obesity effects associated with food consumption in different groups in China. J. Environ. Manag. 317, 115287. https://doi.org/10.1016/j.jenvman.2022.115287 (2022).
Article CAS Google Scholar
Wen, Z. Evaluation of heterogeneity in tectonically deformed coal reservoirs based on the analytic hierarchy process-entropy weight method coupling model: A case study. ACS Omega 8(40), 36700–36709. https://doi.org/10.1021/acsomega.3c02764 (2023).
Article CAS PubMed PubMed Central Google Scholar
Download references
The authors sincerely thank the editors and reviewers for their comments and guidance.
Not applicable.
School of Electrical Engineering, Guangxi University, Nanning, 530004, China
Jiyong Li, Yide Peng, Xiaoping Xiong, Zhiliang Cheng, Chen Ye, Hao Huang & Kaiyue Wang
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
J. L. and Y. P. responsible for developing the core theoretical framework of the algorithm and planning and managing the entire research project. Wrote the first draft of the manuscript and conducted a comprehensive review and edit to ensure scientific rigor and logical coherence.X. X. and Z. C. made key contributions to the design and development of research methods.H. H. and K. W. mainly responsible for software development and data analysis.Z. C. and C.Y. participated in the validation and interpretation of the results of the study.
The authors declare no competing interests.
Not applicable.
Not applicable.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Below is the link to the electronic supplementary material.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Li, J., Peng, Y., Xiong, X. et al. Assessment and economic analysis of photovoltaic power generation potential on highway slope: a case study of Guangxi, China. Sci Rep 16, 3249 (2026). https://doi.org/10.1038/s41598-025-33194-1
Download citation
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s41598-025-33194-1
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.
Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Scientific Reports (Sci Rep)
ISSN 2045-2322 (online)
© 2026 Springer Nature Limited
Sign up for the Nature Briefing: Anthropocene newsletter — what matters in anthropocene research, free to your inbox weekly.
