Smart Street Lighting Powered by Renewable Energy: A Multi-Criteria, Data-Driven Decision Framework

Abstract

Renewable energy sources, such as solar and wind power, are gaining increasing global attention. To facilitate their integration into transportation infrastructure, this paper proposes a multi-criteria assessment framework for identifying the most suitable renewable energy sources for street lighting at any given location. The framework evaluates three key metrics: cost–benefit, reliability, and power generation potential, using time-series weather data. To demonstrate its effectiveness, we apply the framework to data from Georgia, USA. The results show that the proposed approach effectively classifies locations into four categories: solar-recommended, wind-recommended, hybrid-recommended, and no recommendation. Specifically, wind energy is primarily recommended in the southeastern region near the coastline, while solar energy is favored in the northwestern region. A hybrid of both sources is mainly recommended along the coast and in transitional areas. In several isolated parts of the northwest, neither energy source is recommended due to unfavorable weather conditions influenced by the local terrain. Since processing long-term time-series data is computationally intensive and challenging during inference, we train machine learning models, including Multilayer Perceptron (MLP) and Extreme Gradient Boosting (XGBoost), using temporally aggregated features for efficient and rapid decision-making. The MLP model achieves an overall accuracy of 92.4%, while XGBoost further improves accuracy to 94.3%. This study provides a practical reference for regional energy infrastructure planning, promoting optimized renewable energy use in street lighting through a robust, data-driven evaluation framework.

1. Introduction

In light of the ongoing development of global energy structures, the transition from traditional fossil fuels to sustainable energy is becoming increasingly crucial [1]. Renewable energy sources, including solar, wind, and hydroelectric power, have played a pivotal role in shaping the energy supply chain [2]. The necessity for sustainable energy solutions is evidenced by international agreements such as the Paris Agreement, wherein signatories pledge to curtail carbon emissions to combat global warming [3]. To facilitate the advancement of sustainable energy solutions, governments have enacted legislation and established objectives to ensure the development of these commitments. This has resulted in accelerated progress in sustainable energy over the past decade [4].

Solar and wind power stand out among many potential options due to their sustainability, affordability, and wide accessibility [5]. For the past two decades, solar energy has experienced an escalation on a global scale, especially the photovoltaic (PV) system, which has gained wide attention due to the advantages that it can utilize both direct normal irradiance (DNI) and diffuse horizontal irradiance (DHI) at a relatively lower cost [6]. Following solar energy, wind power is the most widely used renewable energy source, producing 4% of global electricity and contributing 7% of the electricity supply in the United States [7]. In the transportation sector, solar and wind power play a crucial role in transitioning energy consumption toward cleaner alternatives, thereby reducing greenhouse gas emissions. In practical applications, hybrid renewable systems that combine both solar and wind sources are often used to enhance energy reliability and provide a more stable power supply under varying environmental conditions [8,9,10]. Particularly, these renewable energy sources can provide sustainable street lighting solutions in budget-constrained areas, particularly in remote or rural regions, where traditional grid-based infrastructure is impractical. Improved street lighting has been shown to significantly reduce traffic crashes [11]. Recognizing these benefits, researchers worldwide, particularly in countries such as the U.S. [12], China [13], and Italy [14], have increasingly focused on integrating renewable energy into transportation infrastructure.

Despite the numerous advantages of solar, wind, and hybrid energy, several challenges remain that require further consideration. One major constraint is the intermittency of these energy sources, particularly under extreme weather conditions [15]. Weather plays a decisive role in ensuring the stable operation of these renewable power generation systems. For PV systems, solar irradiance, measured by the global horizontal irradiance (GHI) and temperature, is the primary influencing factor. GHI can be further decomposed into diffuse horizontal irradiance (DHI) and direct normal irradiance (DNI). In the case of wind turbines, the generation potential is directly dependent on wind speed [16]. Insufficient solar radiation and inadequate wind speeds can compromise the reliability of traffic infrastructure powered by renewable energy sources. Moreover, the suitability of solar and wind energy is highly dependent on geographical conditions. For instance, regions where the maximum annual wind speed does not meet the start-up threshold of a wind turbine or areas with persistent cloud cover throughout the year may not be viable for implementing these renewable energy solutions. Meanwhile, in certain areas, such as mountainous, forested, or heavily shaded regions, the deployment of solar and wind energy may not be viable due to limited resource availability or unfavorable cost–benefit considerations. Addressing these location-specific constraints is essential for practically integrating solar and wind power into transportation infrastructure.

To ensure sustainable and optimal utilization of solar and wind power, it is essential to evaluate their power generation potential [17,18,19]. Advancements in artificial intelligence have provided powerful tools to address this challenge. Over the past decade, machine learning (ML) models have been widely applied in the solar and wind energy sectors [20,21,22,23,24]. One notable application is clustering analysis, which identifies patterns in weather data. Common clustering techniques include k-means clustering [25,26], fuzzy clustering [27], and hierarchical clustering [28], among others. Additionally, ML models have been extensively used for predicting solar [29,30,31] and wind power [32,33,34] generation potential across various timeframes, enhancing the reliability and efficiency of renewable energy integration.

Current research on wind and solar energy evaluation remains limited in scope, primarily focusing on large-scale power generation systems. Many of these studies employ traditional multi-criteria decision-making (MCDM) methods, including the Weighted Sum Method [35], analytical hierarchy process (AHP) [36], and Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) [37], to assess site suitability or system performance based on predefined criteria. However, these conventional methods encounter significant limitations when applied in practical settings. These systems exhibit limited scalability when confronted with substantial spatial datasets and are not designed to manage long-term, time-series weather data. Consequently, these models are less effective for dynamic, data-driven decision-making in renewable energy planning at regional scales. To address the aforementioned challenges, this paper introduces a multi-criteria assessment framework for determining the most suitable renewable energy sources for any given location using time-series weather data. To enable efficient and continuous decision-making in practice, we train machine learning models using temporally aggregated features, enhancing the framework’s practicality. By leveraging energy generation simulation tools and ML classification models, the framework supports scalable, automated, and robust energy planning at high spatial resolution. The remainder of the paper is structured as follows: Section 2 details the methodology of the proposed framework; Section 3 presents experimental results for the state of Georgia. Section 4 discusses limitations and potential adaptations of the framework in future research; Section 5 concludes the paper.

2. Methodology

2.1. Overview of Evaluation Framework

The evaluation framework is comprised of four steps: data collection, power generation modeling, energy source evaluation, and energy source recommendation. These are presented in Figure 1. The first step involves the collection of solar-related weather data, including GHI, DHI, DNI, etc. The National Solar Radiation Data Base (NSRDB) [38] is a publicly accessible repository that provides comprehensive solar and meteorological data with high temporal (30 min) and spatial (2 km) resolution. The Wind Integration National Dataset [39] offers a similar resource for wind-related weather data. Following the download, the data undergoes preprocessing to remove anomalous and null values. After that, a comprehensive database is built, which is then utilized in the subsequent modeling step. Next, the power generation modeling is based on the Renewable Energy Potential Model (reV) proposed by the National Renewable Energy Laboratory (NREL) [40]. The model is capable of calculating the annual power generation and the average generation cost for different renewable energy sources for each location. The third step is to evaluate each alternative renewable energy source at any given location from three perspectives: economic, reliability, and generation potential. The final weighted score for each energy source is determined by applying a specific weight to each aspect, resulting in a comprehensive assessment of each study location. The energy source with the highest weighted score is then recommended for that location. The final step entails training machine learning models for rapid and continuous energy source recommendation using temporally aggregated features (statistics) of time-series weather data, such as maximum, minimum, and mean values.

2.2. Power Generation Modeling

For power generation modeling, the reV model is leveraged, which is an open-source geospatial techno-economic tool for the comprehensive evaluation of solar and wind resources and their geospatial intersection with grid infrastructure and land use characteristics. The model employs a dynamic modeling approach to assess renewable generation, the levelized cost of energy (LCOE), spatial exclusions on developable land, and the renewable energy supply curve [40].

This section presents a method for calculating the annual power generated by renewable energy sources at a specified location. As illustrated in Figure 2, the input data are provided in a time-series format and encompass a range of meteorological variables, including solar radiation, temperature, wind speed, pressure, and so forth. The generation parameters for solar panels and wind turbines are summarized in

Table 1. Technical parameters for solar panel.

Item	Description
Adjust Constant	Adjustment parameter
Array Type	Array types usually include fixed array, single-axis tracking array, and double-axis tracking array
Azimuth	The direction in which the solar panel is pointing
DC/AC Ratio	The ratio between the DC output power of the PV array and the AC-rated power of the inverter
Inverter Efficiency	Conversion efficiency from DC to AC
Loss Ratio	System losses, including transmission losses, dust, temperature, etc.
Module Type	PV module types
System Capacity	Total installed capacity of the solar panel
Tilt	The tilt angle of the solar panel

and

Table 2. Technical parameters for wind turbine.

Item	Description
Adjust Constant	Adjustment parameter
System Capacity	Total installed capacity of the wind turbine
Generic Turbine Loss	Generic turbine losses due to factors such as electrical losses, drivetrain inefficiencies, and maintenance downtimes
Wind Resource Model	The wind resource data source model
Wind Resource Shear	Wind shear exponent
Wind Resource Turbulence	Turbulence coefficient for wind speed fluctuations
Wind Turbine Hub Height	The hub height of the wind turbine
Wind Turbine Power Curve	The turbine power output at different wind speeds
Rotor Diameter	The rotor diameter of the wind turbine

, respectively. The financial parameters include capital cost, fixed operating cost, variable operating cost, and fixed charge rate.

To calculate the total amount of solar power generated by a solar panel, it is necessary to first determine the total amount of radiation received by the panel, denoted by G_tilt, using Equation (1).

$${\mathrm{G}}_{\mathrm{tilt}}=DNI\times \cos \theta +DHI\times {\displaystyle \frac{1+\cos \beta }{2}}+GHI\times \rho \times {\displaystyle \frac{1-\cos \beta }{2}}$$

(1)

where θ is the angle between the zenith and the incident beam; β is the tilt angle of the solar panel; ρ is the land surface albedo.

The total DC power, P_DC, is calculated using Equation (2).

$$P_{\mathrm{DC}}=G_{\mathrm{tilt}}\times A\times {\eta }_{\mathrm{m}}(1-f)$$

(2)

where A is the total area of the solar panel; f is the loss ratio of the system; η_m is the module efficacy, which is decided by the material of the panel.

The total AC power, P_AC, is calculated by Equation (3).

$$P_{\mathrm{AC}}={\eta }_{\mathrm{inv}}\times P_{\mathrm{DC}}$$

(3)

where the η_inv is the inverter efficiency.

The annual solar power generation, E_ys, is computed by Equation (4).

$$E_{\mathrm{ys}}=\sum _{\mathrm{i}=1}^{\mathrm{N}}{P}_{\mathrm{AC},\mathrm{i}}$$

(4)

where N represents the total number of effective daylight hours in a year.

The wind power generation, E_yw, is determined by the wind turbine power curve for each hour and summed over a year by Equation (5).

$$E_{\mathrm{yw}}=\sum _{\mathrm{i}=1}^{\mathrm{N}}{P}_{\mathrm{ACW},\mathrm{i}}$$

(5)

where P_ACW is the output power corresponding to wind speed per hour; N = 8760.

The economy of renewable energy sources can be measured by LCOE, computed by Equation (6).

$$LCOE=\left({\displaystyle \frac{R_{\mathrm{char}}\times C_{\mathrm{cap}}+C_{\mathrm{oper}}}{E_{\mathrm{y}}}}+C_{\mathrm{var}}\right)\times 1000$$

(6)

where R_char is the fixed charge rate; C_cap is the capital cost; C_oper is the fixed operating cost; E_y is the annual energy generation; C_var is the variable operating cost.

2.3. Cost–Benefit

In a multi-criteria assessment framework for system performance, economic considerations are essential. Given the real-world budget constraints, this factor requires careful evaluation to ensure feasibility and effectiveness. Given the meteorological conditions, the aforementioned power generation model is capable of calculating the LCOE of disparate renewable energy sources for a given location. This is achieved through the calculation of power generation and the application of relevant financial parameters, including fixed charge rate (R_char), capital cost (C_cap), fixed operating cost (C_oper), and variable operating cost (C_var). In this study, the reciprocal of the LCOE is employed as the index, C_d, for quantifying the economic performance, consistent with the reliability and generation potential metrics. A higher value of 1/LCOE indicates an enhanced economic performance.

2.4. Reliability

In addition to economic considerations, the reliability of any renewable energy system is another crucial factor. As illustrated in Figure 3, the battery can be charged by a solar panel and/or wind turbine, thereby enabling the street light to be powered at night. If the power stored in the battery is sufficient to power the light over one night, the subject day is deemed satisfactory; otherwise, it is deemed unsatisfactory. The reliability score, R_d, is defined by Equation (7).

$$R_{\mathrm{d}}={\displaystyle \frac{D_{\mathrm{sat}}}{D_{\mathrm{y}}}}\times 10$$

(7)

where D_sat is the total number of satisfactory days in a year; D_y is the total number of days in a year.

2.5. Power Generation Potential

From the perspective of power generation, the power generation potential has a certain impact on the decision to use renewable energy sources for a location. The power generation potential, P_d, is defined as the ratio of total actual generated energy and maximum generation capacity of the system and computed by Equation (8).

$$P_{\mathrm{d}}={\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{N}}{P}_{\mathrm{AC},\mathrm{i}}}{\sum _{\mathrm{i}=1}^{\mathrm{N}}{P}_{\mathrm{rate},\mathrm{i}}}}$$

(8)

where P_rate is the system-rated power and N = 8760, which is the number of hours in a year.

2.6. Multi-Criteria Evaluation

The original scores for various renewable energy sources can be obtained based on the three aforementioned metrics for a given location, as shown in Figure 4. However, since the three evaluative metrics are on different scales, it is necessary to normalize them to a common scale [23]. For each metric, the scores of different energy sources are normalized together, as illustrated in the secondary column of the flow chart in Figure 4. Subsequently, the weighted scores are calculated, and the energy source with the highest weighted score is recommended for the location in question. In the event that the weighted scores of all considered renewable energy sources fall below a specified threshold, the location in question will be designated as a no-recommendation for renewable energy.

2.7. Energy Source Recommendation

The aforementioned multi-criteria assessment framework, relying on long-term fine-grained time-series weather data, is capable of recommending an energy source for any given location. However, to enhance practicality and reduce computational complexity, machine learning models can be trained for adaptive, efficient recommendations using temporally aggregated features. In our setting, the aggregated features are derived from the extraction of minimum, maximum, and mean values from time-series weather data over a year. The targets are the recommended energy categories obtained from the aforementioned multi-criteria evaluation framework.

For this purpose, two commonly used ML models were evaluated: MLP and tree ensemble. The MLP can be viewed as a generic function mapping, which is a feedforward neural network that propagates inputs through multiple layers of fully connected neurons to the final layer for either a regression or classification task. Each layer applies a linear transformation, followed by a nonlinear activation function, in order to capture feature interactions and inherent nonlinearity. For layer k, the output h^(k) is calculated as

$$h^{\left(\mathrm{k}\right)}=\sigma \left(W^{\left(\mathrm{k}\right)}{h}^{\left(\mathrm{k}-1\right)}+b^{\left(\mathrm{k}\right)}\right)$$

(9)

where W^(k) and b^(k) are the weight matrix and bias vector for layer k. h^(k−1) is the input from the previous layer, and σ is the activation function, such as ReLU.

For the classification, the outputs are fed to a softmax function to compute probability across target categories or classes. Cross-entropy loss is computed between the predicted distribution and the target outcome and used for model training.

Besides the traditional MLP model, we also evaluated another popular ML method: tree ensemble. Specifically, we train a gradient boosting tree model. The objective function can be generally written in Equation (10).

$$\mathrm{L}\left(\mathrm{X},\theta \right)=\sum _{i=1}^{n}l(y_{\mathrm{i}},{\widehat{y}}_{\mathrm{i}})+\sum _{k=1}^k\Omega (f_{\mathrm{k}}\left(X,{\theta }_{\mathrm{k}}\right))$$

(10)

where $l(y_{\mathrm{i}},{\widehat{y}}_{\mathrm{i}})$ is the loss function between the target value, $y_{\mathrm{i}}$, and the prediction value, ${\widehat{y}}_{\mathrm{i}}$; $\sum _{k=1}^k\Omega (f_{\mathrm{k}}(X,{\theta }_{\mathrm{k}}))$ is the regularization term of the tree.

The final output of the gradient boosting tree is the sum of the outputs of the sequential trees within the ensemble, as denoted in Equation (11).

$${\widehat{y}}_{\mathrm{i}}=\sum _{k=1}^k{f}_{\mathrm{k}}\left(X,{\theta }_k\right)$$

(11)

In this study, we employed XGBoost (version 2.1.2) [41], a second-order boosting tree model. For model evaluation, we utilized commonly used classification metrics: precision, recall, F1-score, and accuracy, as defined in Equations (12)–(15). Precision quantifies the proportion of correctly predicted samples among those classified as positive, reflecting the reliability of positive predictions. Recall measures the proportion of actual positive samples correctly identified, indicating the model’s ability to detect positive cases. The F1-score, the harmonic mean of precision and recall, provides a balanced assessment, particularly beneficial for imbalanced datasets. Accuracy, calculated as the proportion of correctly classified samples among all samples, serves as a general measure of overall model performance.

$$\mathit{Precision}={\displaystyle \frac{\mathit{TP}}{\mathit{TP}+\mathit{FP}}}$$

(12)

$$\mathit{Recall}={\displaystyle \frac{\mathit{TP}}{\mathit{TP}+\mathit{FN}}}$$

(13)

$$F1\text{-}\mathit{score}=2\times {\displaystyle \frac{\mathit{Percision}\times \mathit{Recall}}{\mathit{Percision}+\mathit{Recall}}}$$

(14)

$$\mathit{Accuracy}={\displaystyle \frac{\mathit{TP}+\mathit{TN}}{\mathit{TP}+\mathit{TN}+\mathit{FP}+\mathit{FN}}}$$

(15)

where TP and TN are the numbers of correctly predicted positive and negative samples, respectively; FP and FN are the numbers of falsely predicted positive and negative samples, respectively.

3. Experiments

3.1. Data Description

To demonstrate our framework, we utilize historical weather data for Georgia, USA, as a case study. Georgia benefits from relatively favorable levels of solar radiation and wind resources, rendering it a representative region for assessing the applicability of various renewable energy technology portfolios. The renewable energy sector primarily consists of three sources: solar energy, wind power, and hybrid energy, which combines both wind and solar resources.

To ensure comprehensive geographical coverage across Georgia, a total of 37,194 sites were sampled, each spaced 2 km apart. This spacing provides sufficient spatial resolution for accurately assessing the distribution and potential energy outputs of wind and solar resources. Once a study site is selected, accounting for the temporal variability of weather data is essential for evaluating the performance of renewable energy systems. Given the considerable inter-annual fluctuations in solar irradiance and wind speed, both of which directly impact the annual energy output of renewable energy systems, relying on weather data from a single year for economic analysis may lead to substantial overestimation or underestimation of cost-effectiveness [42,43].

To address this issue, two primary approaches are typically employed. The first approach involves simulating renewable energy systems using a multitude of Actual Meteorological Years (AMYs) [44]. While this method offers high accuracy and robustness, it requires considerable computational resources. The second approach models renewable energy systems using representative year data, which consists of 12 typical months. In this study, we utilize data from the National Renewable Energy Laboratory (NREL) [38], based on a Typical Meteorological Year (TMY) of 2022. This approach strikes a balance between computational efficiency and result reliability. The relevant weather variables, listed in

Table 3. The weather variables in the database.

Items	Elevation (Meters)	Items	Elevation (Meters)
Temperature	2	DNI	-
	10
Pressure	2	GHI	-
	10
Relative-humidity	2	Wind speed	2
			10
DHI	-	Wind direction	2
			10

Note: Elevation values refer to standard measurement heights of 2 m and 10 m for meteorological observations.

, are time-series data, recorded hourly, totaling 8760 h over the course of a year.

3.2. Power Generation Model

Solar panels and wind turbines were selected for energy generation.

Table 4. Technical parameters for solar panels and wind turbine.

Solar Panel		Wind Turbine
Item	Value	Item	Value
Adjust Constant	0	Adjust Constant	0
Array Type	0	System Capacity (kW)	0.3
Azimuth (Degree)	180	Generic Turbine Loss (%)	0.1
DC/AC Ratio	1.2	Wind Resource Model	0
Inverter Efficiency (%)	0.96	Wind Resource Shear	0.14
Loss Ratio (%)	14.07	Wind Resource Turbulence	0.1
Module Type	0	Wind Turbine Hub Height (m)	6
System Capacity (kW)	0.1	Wind Turbine Power Curve	-
Tilt (Degree)	32	Rotor Diameter (m)	2
Capital Cost (USD/kW)	1988	Capital Cost	1590
Fixed Charge Rate	0.098	Fixed Charge Rate	0.098
Fixed Operating Cost (USD/kW-year)	13	Fixed Operating Cost (USD/kW-year)	51
Variable Operating Cost	0	Variable Operating Cost	0

lists the key technical parameters for solar panels and wind turbines, which are based on recommended values from the official technical documentation of the NREL [40]. The parameters used in this study were chosen to reflect typical configurations of small-scale, distributed renewable systems suitable for street lighting applications, which ensures the accuracy of the simulation process and the representativeness of the results. To evaluate the robustness of the proposed framework, a sensitivity analysis was conducted by varying key parameters, such as capital costs and fixed operating costs, by ±20% within realistic bounds. Additionally, several inputs to the reV model, including solar panel efficiency and wind turbine rated power (each varied by ±10%), were adjusted to assess their influence on annual energy output. The results indicated that while these variations led to some changes in estimated power generation (typically within ±5–8%), the recommended energy source classifications across locations remained consistent. This confirms that the evaluation framework delivers reliable performance despite moderate changes in technical and economic assumptions.

Regarding the specific setup of the solar panels, the value of 0 for the array type indicates a fixed solar panel, which is appropriate for the intended small-scale application of street lighting in Georgia. An azimuth angle of 180 degrees was selected to optimize solar exposure in the northern hemisphere. The module type, indicated by the value of 0, represents a standard crystalline silicon module, which is known for its durability and efficiency. The material utilized is crystalline silicon, with an approximate nominal efficiency of 19%, making it well-suited for consistent and reliable energy generation. The tilt angle of the solar panel is set to the average latitude of Georgia, 32 degrees. For small-scale solar panels, the system capacity is assumed to be 0.1 kW, which is the common value in street lighting systems. These parameters ensure that the PV system can effectively capture solar radiation and convert it into electricity with high efficiency. The choices of DC/AC ratio, inverter efficiency, and loss ratio are based on average PV system performance, reflecting energy conversion efficiency under realistic conditions.

The technical parameters for the wind turbine were also selected from NREL technical documentation, which provided detailed information on different aspects, including the wind energy model, wind resource shear, wind resource turbulence, rotor diameter, and hub height. These parameters were chosen to accurately represent the operational conditions for small-scale street lighting systems. The wind resource shear and turbulence values were selected to align with typical conditions observed in urban and rural settings, thereby ensuring that the model is capable of capturing wind speed variations and turbulence impacts. The rotor diameter and hub height were chosen in accordance with the requirements for light pole applications, thereby ensuring effective power generation under low wind speed conditions while minimizing both physical and visual impacts. Figure 5 illustrates the power curve of the wind turbine utilized in this study. The wind turbine starts generating power after reaching a start-up wind speed of 1.5 m/s, and the output power demonstrates a gradual increase in concert with the wind speed. At a wind speed of 4 m/s, the turbine reaches its rated power of 0.3 kW and maintains this output until the wind speed reaches the cutoff wind speed of 25 m/s. Once the cutoff wind speed is exceeded, the turbine ceases operation to safeguard the equipment from damage due to high wind speeds. This power curve is based on manufacturer specifications for a representative small-scale turbine, as provided in the NREL System Advisor Model (SAM) database. These values reflect idealized conditions under standardized test environments and are commonly used in modeling distributed wind energy systems. Although real-world turbine performance may vary slightly due to local site effects, the selected curve provides a reliable and practical reference for simulation and planning purposes.

The hybrid system combines both solar panels and wind turbines, with a fixed rated power ratio of 1:3 between solar and wind components. All other technical parameters, such as efficiency, loss rate, and cost coefficients, are kept consistent with those used in the standalone solar and wind configurations.

Moreover, for the sake of economic analysis, the study incorporates parameters for both fixed costs and operating costs associated with each energy source. These include capital costs per unit system capacity (solar PV at USD 1988/kW, wind turbine at USD 1590/kW), fixed charge rates (0.098), and fixed operating costs (solar PV at USD 13/kW-year, wind system at USD 51/kW-year). These parameters were chosen in accordance with the standardized economic models provided by NREL to ensure that the research results are economically reasonable and broadly applicable.

3.3. Multi-Criteria Assessment

This study evaluates three renewable energy sources: solar, wind, and hybrid (a combination of both), where the total hybrid power generation is the sum of solar and wind power output. As illustrated in Figure 6, solar power generation ranges from 130 kWh to 160 kWh, while wind power generation varies significantly, ranging from 0 kWh to 2000 kWh. This indicates the greater variability of wind power compared to solar energy. The multi-criteria assessment framework considers three key factors: cost–benefit ratio, reliability, and power generation potential.

The cost–benefit ratio is calculated from the total power generation and the associated cost. The average LCOE for solar power is USD 1.38 per kilowatt-hour (kWh). Conversely, the LCOE for wind power exhibits significant variability. It is noteworthy that in certain locations, the maximum annual wind speed falls below the start-up speed of the wind turbine. In such cases, the LCOE cannot be calculated, resulting in zero cost–benefit scores for these locations. In the case of hybrid power, the total generation and cost are the sum of solar and wind power, and the LCOE ranges from 1.43 USD/kWh to 2.54 USD/kWh. From an economic perspective, solar power emerges as the most cost-effective option.

The reliability of power systems at each location is assessed based on energy demand. The street lighting system is assumed to have a rated power of 0.5 kW, operating for 10 h per night, resulting in a total energy requirement of 0.5 kWh per night. Figure 7 shows the relationship between power generation and satisfaction rates for solar, wind, and hybrid energy systems. As shown in Figure 7a, solar power generation exhibits a moderate positive correlation with satisfaction rates, reaching approximately 60% as generation increases. Figure 7b indicates that wind power achieves nearly 100% satisfaction even at relatively low generation levels, demonstrating its high reliability and efficiency. Additionally, as depicted in Figure 7c, the hybrid system reaches near-complete satisfaction more quickly than wind power alone, requiring less energy generation to achieve the same level of reliability for the purpose of street lighting at night.

Finally, the power generation potential is evaluated. Solar power generation is more evenly distributed across regions, with an average actual generation rate of 17%. In contrast, wind power exhibits significant variability across regions, with actual generation rates ranging from 0% to 71.2%. For hybrid energy systems, the actual generation rate falls within a range of 3.8% to 57.7%.

To ensure consistency across the three evaluation criteria, the original scores are normalized to a fixed range between 0 and 1. The weight coefficients can typically be determined using expert scoring methods [45] or the analytic hierarchy process [46]. In this study, the weight coefficients for cost–benefit ratio, reliability, and power generation potential are set at 0.2, 0.6, and 0.2, respectively. The weights were determined by prioritizing agencies’ preferences for the reliability of solar and wind energy sources, while also accounting for relevant economic and technical considerations. It is important to note that this weighting scheme may require adjustment in different geographic or policy settings to align with local priorities. The energy source with the highest weighted score is designated as the recommended energy option for each location. If all weighted scores fall below 4.5, the location is classified as not suitable for renewable energy implementation.

Figure 8 illustrates the distribution of renewable energy recommendations across the state of Georgia, categorizing sampled locations into solar, wind, hybrid, and no recommendation groups. The analysis reveals that wind energy holds the largest share, with 15,945 locations (42.87% of all sites), highlighting its substantial applicability and potential. Solar energy is recommended for 9020 locations, representing 24.25% of all sites. Hybrid energy, combining both solar and wind, accounts for 8004 locations (21.52%), suggesting that an integrated approach may offer greater reliability and economic benefits in certain areas. Conversely, 11.36% of locations are classified as not recommended for renewable energy, likely due to resource limitations or economic constraints.

To visualize the spatial distribution of the classification results, a renewable energy recommendation map was generated. This map depicts all 37,194 sampled sites across Georgia, with each site color-coded according to the framework’s recommendation: solar, wind, hybrid, or no recommendation. As shown in Figure 9, wind energy is predominantly recommended in the southeastern region of the state, where average wind speeds are relatively higher and more consistent. Solar energy is favored in the northwestern region, particularly around the Atlanta metropolitan area, where direct and diffuse solar irradiance levels are stronger and more stable. Hybrid systems are recommended in transitional zones that benefit from moderate levels of both wind and solar resources. Areas with limited solar irradiance and low wind speeds, such as forested or mountainous regions, are classified as not suitable for the deployment of solar or wind energy systems.

3.4. Classification Model

To facilitate implementation and adaptability to new data and new locations, machine learning models, including MLP and XGBoost, are trained to recommend energy sources using temporally aggregated features, framing the problem as a classification task. As shown in Figure 10, key features, such as minimum, maximum, and mean values, are extracted from the original meteorological time-series data. This approach effectively captures underlying trends and statistical distributions, enhancing the model’s ability to generalize across different locations.

The dataset is divided into two subsets: 80% is used for training, while the remaining 20% is reserved for testing. The MLP model is designed with three hidden layers, comprising 128, 64, and 32 neurons, respectively. To enable nonlinear mapping, a rectified linear unit (ReLU) activation function is employed. Additionally, a dropout layer with a rate of 0.3 is used to mitigate overfitting and improve generalization. The model is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and a total of 50 training epochs. These parameters were selected based on experimental results to ensure stable learning and smooth convergence.

Table 5. Evaluation metrics for MLP models.

Energy Source	Precision	Recall	F1-Score	Accuracy
Solar	0.91	0.86	0.88	0.924
Wind	0.98	0.99	0.98
Hybrid	0.95	0.91	0.93
No Recommendation	0.73	0.83	0.78

summarizes the evaluation metrics for the MLP model for renewable energy recommendation. The results indicate that the model performs exceptionally well in the wind recommendation category, attaining the highest F1-score of 0.98. This is followed by the hybrid and solar recommendation categories, with F1-scores of 0.93 and 0.88, respectively. Although the no recommendation category has a slightly lower F1-score of 0.78, it exhibits a strong capability in identifying relevant samples, as evidenced by its recall value of 0.83. The model’s overall accuracy of 92.4% demonstrates its effectiveness in classification using temporally aggregated features.

Besides the MLP model, the XGBoost model was also implemented to further enhance classification accuracy. The hyperparameters of the XGBoost model were optimized using a grid search method to achieve optimal performance for the renewable energy recommendation task. The search space included maximum tree depth ∈ {6, 9, 12}, learning rate ∈ {0.05, 0.10, 0.15}, iterations ∈ {80, 100, 120}, subsample ratio ∈ {0.4, 0.6, 0.8}, and column sampling ratio ∈ {0.2, 0.4, 0.6}. A 5-fold cross-validation strategy was applied. The macro-averaged F1-score was used as the evaluation metric to guide parameter selection. Based on the grid search results, the final configuration includes a learning rate of 0.15, a maximum tree depth of 9, 120 iterations, a subsample ratio of 0.6, and a column sampling ratio of 0.4. As shown in

Table 6. Evaluation metrics for XGBoost models.

Energy Source	Precision	Recall	F1-Score	Accuracy
Solar	0.90	0.91	0.91	0.943
Wind	0.99	0.99	0.99
Hybrid	0.97	0.95	0.96
No Recommendation	0.81	0.80	0.80

, the XGBoost model achieved an overall accuracy of 94.3%, surpassing the 92.4% accuracy attained by the MLP model. Furthermore, it demonstrates superior precision, recall, and F1-scores across most categories, with a particularly notable improvement in precision for the “no recommendation” category (0.81 vs. 0.73), reflecting better control over false positives.

4. Discussion

This section discusses the key aspects influencing the effectiveness and applicability of the proposed framework.

4.1. Model Generalization and Transferability

Once the classification models are trained for a specific region, their performance can be largely influenced by the similarity of weather patterns when applied to a different region. To mitigate this impact, the mean and variance of the weather data from the original training location can be used to standardize the weather data from the new region. This process aligns the feature distributions of the new region with those of the original training data, reducing discrepancies caused by regional variability. In the Georgia case study, the XGBoost model achieved an overall accuracy of 94.3%, while the MLP model reached 92.4%, demonstrating the effectiveness of aggregated weather features for reliable energy recommendation classification. These results suggest that the proposed framework generalizes well and, with appropriate normalization or transfer learning techniques, can be adapted for application in other regions. Moreover, if substantial differences in meteorological patterns persist between the original and new locations, leading to considerable prediction errors even after data normalization, model fine-tuning may be necessary.

4.2. Weight Coefficients in Multi-Objective Evaluation

In our multi-criteria evaluation framework, economy, reliability, and generation potential are three core indices. The weight coefficients play an important role in the final recommendation. In this study, the weights were assigned to reflect general preferences regarding the reliability of solar and wind energy sources. Empirical results indicate that wind energy was recommended in 42.87% of the sampled locations, primarily in the southeastern region near the coastline. Solar energy was favored in 24.25% of locations, mainly in the northwestern region. A hybrid of both sources was recommended in 21.52% of locations, mostly along the coast and in transitional regions, while 11.36% of the locations were categorized as not recommended for either source. The prioritization of these indicators may vary across regions. For instance, in rural regions, economics is typically the most important indicator. Conversely, in urban regions with robust policy support, greater emphasis may be placed on generation potential or reliability. Therefore, the values of weight coefficients need to be determined in response to the environment, policies, or socio-economic conditions.

On the other hand, the utilization of renewable energy may have a negative impact on the ecosystem. For example, the fan blades of wind turbines may pose a threat to bird activity. Therefore, environmental impacts could be considered as complementary indicators in future assessments to achieve a better balance between sustainable energy and ecological protection.

4.3. Limitations and Future Research

In some remote areas, such as mountainous or high-latitude regions, the recommended renewable energy sources based on meteorological conditions may encounter significant challenges in terms of deployment and maintenance. To address this, future frameworks could integrate infrastructure and terrain conditions as additional constraints to improve their practical feasibility. Furthermore, the social acceptance of different renewable energy types can vary by region. For instance, wind turbines may face public opposition in certain areas due to concerns over noise and visual impact. As such, the multi-objective evaluation framework should also consider local cultures, public preferences, and potential social impacts to offer more nuanced guidance for energy planning. In addition, local policy preferences, such as dedicated support for wind or solar, can influence the recommended outcomes. The framework could therefore be optimized to incorporate policy variables, further refining the analysis and recommendations.

5. Conclusions

This study proposes a multi-criteria evaluation framework for renewable energy recommendation, addressing economic analysis, reliability evaluation, and generation potential assessment. The framework effectively identifies the most suitable renewable energy source for each site. To facilitate implementation and adaptability to new data and new locations, machine learning models, such as MLP and XGBoost, were trained to recommend energy sources using temporally aggregated features, achieving robust performance. The main conclusions of the study are as follows:

(1)

Among the 37,194 evaluated locations across Georgia, wind power emerged as the most recommended source, accounting for 15,943 sites (42.87%). Solar and hybrid power were recommended for 9021 sites (24.25%) and 8010 sites (21.52%), respectively. The remaining 4220 sites (11.35%) were classified as not suitable for renewable energy deployment. These results demonstrate the effectiveness and practicality of the proposed multi-criteria evaluation framework in generating location-specific renewable energy recommendations based on resource potential, cost, and reliability.

(2)

The MLP model shows notable proficiency in the classification task, attaining an overall accuracy of 92.4% on the test dataset. The XGBoost model outperforms the MLP model, achieving an overall accuracy of 94.3% on the test dataset. The F1-scores for XGBoost across the four recommendation categories were 0.95 (Wind), 0.92 (Solar), 0.93 (Hybrid), and 0.89 (No recommendation). To improve transferability when applying these models to different regions, standardizing meteorological data or utilizing transfer learning can be beneficial. Additionally, customizing the weight coefficients in the multi-criteria evaluation framework is essential to account for regional variations in economic factors, reliability, and generation potential.

(3)

Considering environmental impacts, infrastructure constraints, and social acceptance factors can further enhance the operability of the framework, aligning it more closely with real-world scenarios and ensuring its practical relevance for energy planning. For example, 4220 locations (11.35% of total sites) were classified as not suitable for renewable energy deployment, many of which are situated in mountainous or forested areas with limited availability of solar and wind resources. In such regions, challenging terrain, installation difficulty, or public opposition to wind turbines may further hinder implementation, even if basic meteorological conditions are marginally acceptable. These findings underscore the importance of incorporating broader practical considerations into future extensions of the framework.

Overall, this study offers valuable insights into the integration of renewable energy sources for street lighting, offering a flexible and data-driven approach for effective energy planning and sustainable infrastructure development.