A machine learning approach to assess the climate change impacts on single and dual-axis tracking photovoltaic systems

Proposed work

The proposed research concentrates on creating an adaptive solar tracking system powered by AI that adjusts the orientation of solar panels according to the current climate conditions and seasonal changes over a period. The system evaluates the effect of climate change on the efficiency of solar tracking for static, single-axis, and dual-axis tracking systems using a 330 W solar panel setup. The system, by combining machine learning-based prediction and reinforcement learning-based optimization, dynamically selects the most efficient tracking mode to maximize energy output while reducing energy consumption.

The main goals of this study are:

Formulating an adaptive tracking algorithm that dynamically adjusts panel orientation based on real-time weather patterns and climate trends. To Evaluating the effects of seasonal climate changes, such as temperature, solar irradiance, cloud cover, wind speed, and humidity, on the performance of static, single-axis, and dual-axis tracking systems for a period of one year (January 2024–January 2025). To Apply a hybrid machine learning strategy that combines the convolutional neural network–long short-term memory (CNN–LSTM) for solar irradiance and climate forecasting. The XGBoost Regression for the estimation of energy yield based on various tracking modes.Deep Q-learning (DQL) was used to optimize real-time tracking modes based on energy efficiency forecasting and climate conditions. Tuning the performance of the ML-based adaptive tracking system compared with traditional tracking schemes to assess energy efficiency improvement and system resilience against climate change.

This study makes a novel contribution that is distinct from the classic solar tracking solutions. Climate predictions and adaptive reinforcement learning enable instant solar tracking maximization. Compared with classic trackers based on programmed algorithms, this system has an automatic adaptation capability that follows season-changing and instant-on weather updates. The use of the reinforcement model enables learning through self-practice and maximization, without requiring regular manual re-adaptation for optimal results and decision-making. Figure 1 depicts the well-defined workflow of the proposed AI-driven adaptive solar tracking system. The approach was initiated with the real-time acquisition of climate parameters, such as solar irradiance, temperature, humidity, wind speed, and cloud cover. Preprocessing of the collected data ensures normalization, elimination of outliers, and feature extraction to support accurate machine learning prediction. The CNN-LSTM model thereafter predicts climate patterns, which supports predictive insights regarding solar irradiance variability. Next, the Boost regression model predicts the energy yield in varying tracking modes to form a basis for decision making. The Deep Q-Learning model selects the best tracking mode (static, single-axis, or dual-axis) with the maximum efficiency in real-time climate input and forecasted power output. The system was continuously analyzed to enhance the model to fit seasonal patterns and maximize the long-term tracking efficiency. This method ensures that solar panels dynamically adjust to fluctuating weather patterns, leading to greater energy efficiency and reduced power usage, which aligns with the research objective of developing a climate-resilient solar-tracking optimization system.

The proposed AI-driven decision system integrates CNN-LSTM for climate prediction, XGBoost for power calculation, and Deep Q-Learning for tracking mode choice. This system guarantees greater adaptability and efficiency enhancement compared to traditional tracking systems. The proposed AI-powered tracking model scales with the ability to integrate it into large-scale solar farms and uses IoT-based edge computing for real-time decisions.

Data collection & preprocessing

The data applied in this research is one year (January 2024–January 2025) of climate and energy data of Sitapura, Jaipur, India. The data were derived from meteorological data, on-site sensor measurements, and historical data to present an extensive climate observation and accurate machine learning prediction.

Dataset description

The dataset of research includes the following key climate and energy parameters:

• Solar irradiance data: The Direct Normal Irradiance (DNI), Global Horizontal Irradiance (GHI), and Diffuse Horizontal Irradiance (DHI), which are recorded in every 15-minute intervals that allow assessment of the solar energy potential at different times of the day.

• Temperature & humidity: Measured hourly intervals to account for seasonal fluctuations that impact solar panel efficiency.

• Wind speed & Cloud cover: The Data obtained from weather stations and satellite images that evaluate the effect of environmental factors on the tracking performance.

• Energy output: The Collection from the three different tracking systems (a) static, (b) single-axis, and (c) dual-axis tracking, which logged at 30-minute intervals to compare their performance across varying climatic conditions.

The dataset includes solar irradiance, temperature, humidity, wind speed, cloud cover, and energy output for different tracking modes, recorded over a period of one year (January 2024–January 2025) in Sitapura, Jaipur, India.

The data pertain to solar radiation, climatic conditions, and power output for static, single-axis, and dual-axis trackers, to represent the impact of climate variation on solar panel efficiency. Tables 1 and 2 provide a clear presentation of the datasets used in this study. Table 1 summarizes the climate and energy parameters measured, including solar irradiance (DNI, GHI, and DHI), temperature, humidity, wind speed, cloud cover, and energy generation for different tracking modes to provide full data coverage for machine learning-based optimization and forecasting. Table 2 presents example data collected over different seasons (Winter, Summer, Monsoon, and Post-Monsoon) that reflect the variations in solar radiation, climatic conditions, and energy output of the static, single-axis, and dual-axis tracking mechanisms. This information is important for training and validating AI models to take climate-respondent tracking decisions in accordance with real-time conditions.

Justification for using Jaipur as the study site

Jaipur, India, is an ideal location for this research for the many reasons, the location has high solar irradiance zone, In Jaipur experiences an average DNI of ~ 5.5 kWh/m²/day, making it well suited for evaluating solar tracking efficiency. It has Distinct Seasonal Variations, in this region experiences extreme summers (~ 45 °C), cold winters (~ 5 °C), and monsoon seasons, providing diverse climate conditions to test the adaptive tracking performance. Existing Solar Installations in the Jaipur has several operational solar farms that facilitate real-world experimental validation and comparison with conventional tracking methods.

Data processing steps

For high-quality and consistent data, the following preprocessing methods are used:

• Handling missing values: Missing values caused by sensor faults or communication loss are treated using linear interpolation for short missing intervals and statistical modeling for long missing durations to ensure data consistency.

• Feature scaling & normalization: Min-Max normalization is used to scale all input variables to a common range (0 to 1) to avoid bias in machine learning models.

• Outlier detection & removal: Filtering based on Z-scores is employed to remove abnormal readings that can be caused by unusual weather conditions or faulty sensors.

• Data augmentation for machine learning: Synthetic data generation methods are utilized to improve model robustness to ensure that the ML model generalizes well to new climatic conditions.

Seasonal variability and cloud cover impact on solar tracking efficiency

The seasonal variation and cloud cover effect on the solar tracking efficiency are pivotal in the optimization of the adaptive positioning of solar panels. Seasonal changes in Direct Normal Irradiance (DNI), Global Horizontal Irradiance (GHI), and Diffuse Horizontal Irradiance (DHI) influence power output; hence, predictive climate simulation was used to optimize tracking efficiency. These variations are illustrated in Fig. 2(a) through a heatmap in which solar radiation on a monthly basis is plotted and labeled as varying with months and directly influencing the energy output of the single-axis, static, and dual-axis tracker systems. The boxplot representation of cloud cover impact in solar tracking mode is complemented by Fig. 2(b) and shows that static panels lose more efficiency in a cloudier period, whereas the dual-axis tracker possesses relatively consistent power output. These findings emphasize the need for machine-learning-based tracking decisions according to season and instantaneous weather conditions to ensure enhanced energy efficiency under dynamically changing environments.

Figure 2 illustrates the impact of seasonal variations on the available solar energy level, which is important for enhancing adaptive solar tracking methods. (b). Boxplot of the impact of cloud cover on solar panel efficiency for different tracking modes (static, single axis, and dual axis). The plot shows the variation in power generation due to changing cloud cover, which helps to optimize machine learning-based optimization techniques for decision tracking. Preprocessed data form the foundation for climate forecasting, estimation of energy yield, and real-time adaptive tracking, giving the proposed system a way of effectively responding to changing climate patterns and optimizing solar panel efficiency.

Climate impact prediction using CNN-LSTM

Solar irradiance variations and climate variability must be forecast to optimize adaptive solar tracking systems. Energy production is considerably affected by seasonal cycles, cloud cover, and temperature variations, making it crucial for an AI-based predictive system to be in place. A 10-day forecast of solar irradiance trends is provided by a hybrid CNN-LSTM model so that the solar panel orientation may be adaptively controlled in real time.

Model selection & justification

The proposed deep learning model integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) to capture spatiotemporal dependencies in climate patterns.

CNN for feature extraction:

• Input: Satellite cloud cover images of size 256 × 256 × 3 (RGB bands).

• Architecture: Three convolutional layers with kernel size 3 × 3 and ReLU activation.

• Pooling Layer: Max pooling (2 × 2) to reduce spatial dimensions.

• Feature Vector Output: Flattened to 512 features for LSTM integration.

LSTM for Time-Series Forecasting:

• Input: Climate feature matrix X_t​ with past 30 days of solar irradiance, temperature, and humidity.

• Architecture: Two LSTM layers with 128 hidden units each.

• Dropout: 20% (0.2) to prevent overfitting.

• Activation function: Tanh (for sequential dependencies).

• Output: 10-day forecast of DNI, GHI, and DHI values.

Mathematical model for solar irradiance prediction

The solar irradiance prediction model is mathematically expressed as in Eq. (1):

Where, I(t) = Predicted solar irradiance at time t̅, \(\:{\text{X}̅}_{\text{t}}\) = Climate feature matrix (Temperature, Cloud Cover, Humidity) and \(\:{\text{W}̅}_{\text{t}}\)​ = Past 30 days of irradiance values for LSTM-based forecasting.

The loss function used for the optimization is the Mean Squared Error (MSE), as shown in Eq. (2):

where I_actual​ is the true irradiance value, and I_predicted​ is the CNN-LSTM forecasted irradiance. The model was trained using the Adam optimizer with an initial learning rate of 0.001 and dynamically adjusted using a learning rate decay factor to prevent overfitting.

Expected output and uncertainty analysis

The new CNN-LSTM model is predicted to produce a high-precision 10-day ahead solar irradiance forecast, which can be utilized directly for energy yield estimation and adaptive tracking mode choice. However, owing to the natural unreliability of climate forecasting, confidence intervals of predictions must be evaluated to estimate potential variations from real solar irradiance levels.

Prediction confidence interval (PCI)

To guarantee the validity of the forecasts, the 95% Confidence Interval (CI) for DNI Prediction was obtained and calculated using Eq. (3):

where I_predicted​ is the forecasted solar irradiance of the CNN-LSTM model and Z_α/2​ is the critical value from the standard normal distribution (1.96 for a 95% confidence level). σ is the standard deviation of the prediction error of the model.

This confidence interval ensures that the true irradiance value lies within this range with 95% probability, providing uncertainty quantification for real-time decision-making.

Expected model performance

The forecasting accuracy is evaluated using the following performance metrics, as shown in Eqs. (4), (5), and (6):

To ensure high forecasting accuracy, the following evaluation metrics are used:

1. 1)

   Root mean squared error (RMSE) – Measures prediction deviation:

   $$\:\text{R}\text{M}\text{S}\text{E}=\sqrt{\frac{1}{n}}\:\:\sum\:{\:\:\left({\text{I}}_{\text{a}\text{c}\text{t}\text{u}\text{a}\text{l}}\text{}-{\text{I}}_{\text{p}\text{r}\text{e}\text{d}\text{i}\text{c}\text{t}\text{e}\text{d}}\right)}^{2}$$

   (4)

   Expected RMSE: ≤̅ 5% deviation.

2. 2)

   Mean absolute error (MAE) – Measures absolute differences using Eq. 5

   $$\:\text{M}\text{A}\text{E}=\frac{1}{n}\sum\:{\text{I}}_{\text{a}\text{c}\text{t}\text{u}\text{a}\text{l}}\text{}-{\text{I}}_{\text{p}\text{r}\text{e}\text{d}\text{i}\text{c}\text{t}\text{e}\text{d}}\text{}\mid\:\text{}$$

   (5)

   Expected MAE: ≤̅ 2.5%.

3. 3)

   R-Squared (R²) Score: measures model goodness-of-fit using Eq. 6

   $$\:\text{R}2=1-\:\frac{\sum\:{({\text{I}\text{}}_{\text{a}\text{c}\text{t}\text{u}\text{a}\text{l}}-{\text{I}}_{\text{p}\text{r}\text{e}\text{d}\text{i}\text{c}\text{t}\text{e}\text{d}\text{}})}^{2}}{{\sum\:(\text{I}\text{a}\text{c}\text{t}\text{u}\text{a}\text{l}\text{}-\text{I}\text{ˉ})}^{2}}\:\:\:\text{}$$

   (6)

   Target R² Score: ≥0.95 for high accuracy.

The use of the CNN-LSTM hybrid model in this study is critical because it incorporates sequential pattern recognition strength and spatial feature discovery capability, which are suitable for solar irradiance forecasting. CNN clearly extracts climate-related spatial patterns from satellite imagery, including humidity, atmospheric conditions, and cloud cover, which affect the solar radiation. During that time, LSTM computes past irradiance patterns and stores long-range relations; therefore, it can project future solar radiation variations with extreme precision. A mathematical formula for the expression guarantees the capture of temporal and spatial dependencies, resulting in effective and accurate forecasting. The precision of prediction was measured using the Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R² score, ensuring that the prediction was as similar as possible to the measurement. These forecasts play a vital role in adaptive solar tracking optimization, in the sense that they are employed directly for energy output estimation (through XGBoost) and tracking mode determination (through Deep Q-Learning). If the model predicts low irradiance from dense cloud cover, the tracking system can minimize movement and save energy, whereas high-irradiance predictions call for dual-axis tracking to ensure maximum energy capture. Therefore, the CNN-LSTM model is not only a prediction tool but also a major facilitator of real-time decision-making to ensure that solar tracking is performed efficiently under different climatic conditions.

Energy yield prediction using XGBoost

Accurate prediction of solar energy output under various tracking modes is important for maximizing energy production and enhancing system efficiency. Because solar power generation depends on climatic conditions, panel tilt angle, and past energy patterns, a predictive model based on machine learning is required to identify the best tracking mode for varying weather conditions. In this study, Extreme Gradient Boosting (XGBoost) was utilized for energy yield prediction based on its capability to deal with non-linear relationships, interactions among the features, and high computational efficiency.

Model input variables and feature selection

The XGBoost model accepts several inputs to provide an exhaustive analysis of the power generation factors:

Climate data:

• Solar Irradiance (DNI, GHI, DHI) – Modeled by the CNN-LSTM model.

• Temperature (°C) and humidity (%): seasonal influence on panel efficiency.

• Wind Speed (m/s) and Cloud Cover (%) influence real-time tracking performance.

Solar panel orientation & tracking mode:

Historical energy output data: The Historical records of power generation applied in supervised learning.

By incorporating the above features, XGBoost efficiently models subtle interactions between climate, panel azimuth, and power output to achieve a high-accuracy yield prediction. Multiple machine learning algorithms were compared for predicting the energy yield, and XGBoost was chosen because of its better performance in capturing nonlinear patterns, interactions between features, and computational performance. While linear regression requires a linear relationship between input and output variables, XGBoost is able to better model the complex nonlinear relationships between climate variables and power output, which is more appropriate for real-world scenarios where weather conditions vary dynamically. In addition, XGBoost is more computationally efficient than Random Forest because it uses gradient boosting with parallelization, which enables faster processing, and less memory usage compared to standard ensemble models. Figure 3 shows another important benefit of XGBoost because it has a lower RMSE than ANN-based methods for small- and medium-sized datasets. Although ANNs require plenty of data to generalize well, XGBoost also has high predictability with a moderate training data set and, therefore, was apt for the current research in which estimation of energy yield is based on one-year climate parameters and past solar power output data. Hence, XGBoost is a better compromise between predictability, computational cost, and interpretability; hence, it is an ideal model choice for predicting the energy yield in the current study.

The model combines forecasted climate data (temperature, humidity, and DNI) from the CNN-LSTM module and past energy output to forecast solar power generation in various tracking modes. Conditional validation guarantees the accuracy of the model with RMSE, MAE, and R̅²Score, along with an optimization refinement loop iterating over iterations. The results will help in choosing the optimal tracking mode for real-time energy maximization. By utilizing XGBoost to estimate the energy yield, this study presents a computationally effective and scalable framework that boosts real-time decision-making in solar tracking systems. The predictive capability of power generation over a range of weather conditions allows the tracking mode options to be dynamically optimized under the parameters of energy efficiency and environmental requirements.

Proposed algorithm for adaptive solar tracking (COMLAT)

Maximizing solar energy efficiency in dynamically changing climatic conditions requires an intelligent self-tuning tracking mechanism with real-time adaptation capability. For this purpose, we propose the Climate-Optimized Machine Learning Adaptive Tracking (COMLAT) Algorithm, an AI-based decision system for tracking decisions that best selects between Static, Single-Axis, or Dual-Axis solar tracking modes using reinforcement learning. The novelty of COMLAT is the integration of XGBoost-based energy yield prediction with Deep Q-Learning (DQL) for adaptive tracking control, wherein tracking decisions become data-driven, self-tuning, and scalable across various environmental conditions.

Algorithm framework and decision mechanism

The proposed COMLAT algorithm uses a machine-learning-based decision process that uses predictive modeling (XGBoost) and reinforcement learning (DQL) algorithms. To apply real-time adaptive tracking control.

Step 1

Input Climate Features & Predicted Irradiance.

The tracking system offers continuous monitoring of the real-time climate parameters. It integrates the multi-step solar irradiance forecasts from the CNN-LSTM model in Eq. (7):

Where the S_t​ represents a state space at time t̅, it is providing a climate-dependent decision factors.

Step 2

Energy Yield Estimation Using XGBoost Regression.

XGBoost predicts the expected power output for different tracking modes, enabling preemptive decision-making, as shown in Eq. (8):

where P_out​ is the predicted power output (kWh) for the different tracking configurations. The ability of XGBoost to handle nonlinear dependencies ensures highly accurate energy yield predictions.

Step 3

Reinforcement Learning-Based Tracking Mode Selection (DQL).

The Deep Q-Learning (DQL) algorithm optimally selects the most energy-efficient tracking mode and continuously refines its decisions through self-learning using Eq. 9.

Where,

Q(s, a) = Expected reward for selecting action aaa (tracking mode) in state ss (climate condition).

r̅ = Reward function based on energy yield improvement and tracking efficiency.

γ\gammaγ = discount factor (0.9), prioritizing long-term rewards.

Q(s′,a′) = Future Q-value estimate for the next state s̅′.

The reward function is defined as shown in Eq. (10):

where, \(\:{\text{P}}_{out}^{\text{s}\text{e}\text{l}\text{e}\text{c}\text{t}\text{e}\text{d}}\) = Predicted power output of the selected tracking mode, \(\:{\text{P}}_{\text{o}\text{u}\text{t}}^{\text{s}\text{t}\text{a}\text{t}\text{i}\text{c}}\)= Baseline energy yield for the static panel, Movement- Cost = Energy penalty for excessive tracking movements and α & β = Weighting factors that optimize the trade-off between energy gain and actuator efficiency.

Although the COMLAT algorithm presents a scientifically sound and AI-based method for adaptive solar tracking, various improvements can make it more computationally efficient, scalable, and applicable in real-world scenarios. These improvements address the issues of computational complexity, multi-agent cooperation, hyperparameter tuning, IoT integration, and adaptation to extreme weather conditions, making COMLAT a cutting-edge AI-based tracking system.

Strategic improvements and Long-Term evolution of COMLAT

Computational complexity optimization.

Reinforcement learning-based tracking incurs computational overhead, especially in real-time solar tracking applications, where decisions have to be made with low latency. In contrast to conventional rule- or PID-based tracking systems, COMLAT updates its policy functions at every iteration, which requires computational resources.

Optimization Approach:

• Computational complexity analysis (Big O analysis) was performed for state-action evaluations and Q-learning convergence.

• Apply dynamic learning rate adjustment to efficiently optimize Q-value updates.

• Benchmark COMLAT’s decision-making execution time compared with traditional tracking controllers to ensure real-time feasibility.

Multi-agent reinforcement learning (MARL) for large solar farms

The existing COMLAT was developed for single-agent decision-making by optimizing individual panel tracking. However, in large solar farms, multiple panels need to work together to avoid tracking conflicts and maximize energy distribution.

Proposed Enhancement:

• Integrate multiagent reinforcement learning (MARL), allowing multiple COMLAT agents to work together in tracking decisions.

• Design a cooperative reward-sharing scheme such that tracking optimization is performed at the system level and not for individual panels.

• Agent-to-agent communication should be implemented to avoid shadowing effects and opposing tilt angles in large-scale PV systems.

Adaptive hyperparameter optimization for improved convergence

The performance of Deep Q-Learning (DQL) in COMLAT relies on hyperparameters such as the learning rate, discount factor (γ), and exploration-exploitation trade-off. Inadequate tuning can result in a slow convergence or unstable tracking policies.

• Optimization strategy:

• Utilize Bayesian optimization or genetic algorithms for automatic hyperparameter optimization.

• Implement adaptive exploration methods via ε-greedy decay mechanisms, allowing the model to shift from exploration to exploitation more effectively.

• Tune the reward function scaling such that an ideal balance is maintained between maximizing the energy yield and actuator efficiency.

IoT integration and edge computing for real-time tracking

For real-world deployment, COMLAT must run on low-power embedded hardware, providing real-time tracking decisions without reliance on the cloud. Conventional cloud-based AI models incur communication latency, which renders real-time execution infeasible.

Proposed enhancement:

• Implement COMLAT on edge AI hardware (e.g., NVIDIA Jetson Nano, Raspberry Pi, or FPGA-based processors) to enable low-latency tracking control.

• The implementation of 5G-enabled IoT networks provides rapid communication among sensors, reinforcement learning models, and tracking actuators.

• Optimize the AI model with quantization and model compression to minimize the computational overhead while preserving the tracking accuracy.

Improved reward function for extreme weather adaptation

COMLAT optimizes tracking based on solar irradiance and cloud cover but does not specifically consider extreme weather conditions, such as high winds, storms, or sudden temperature changes, which may influence tracking stability.

Proposed enhancement:

• Modification of the reward function to penalize dangerous tracking angles during high-wind conditions to offer structural integrity to the panels.

• Add a hybrid mode in which COMLAT can alternate between AI-driven tracking and preconfigured safety modes for extreme weather conditions.

• Introduce long-term seasonal adaptation mechanisms in which the model adapts to previous seasonal fluctuations to refine tracking approaches for extended periods of cloudy or low-radiation months.

By integrating these enhancements, the COMLAT can become an extremely scalable, smart, and industry-grade AI solution for adaptive solar tracking. The addition of multi-agent reinforcement learning, edge computing deployment, real-time IoT integration, and extreme weather adaptation will keep COMLAT ahead of the next-generation AI-driven renewable energy solutions. These upgrades will not only enhance real-time adaptability but also support wider applications, such as autonomous solar grids, distributed PV networks, and self-optimized microgrids. With constant improvements, COMLAT has become a standard AI-based solar tracking technology to provide maximum energy efficiency under various climatic conditions.

Experimental setup

To prove the efficacy of the COMLAT algorithm, an appropriate experimental setup was established to infuse its scientific accuracy and practical use. The experiment included a testbed of a solar tracking setup installed in Sitapura, Jaipur, India, addressing three varying configurations of tracking types: Static, Single-Axis, and Dual-Axis tracking systems. The main goal of the experimental setup was to measure the energy yield, tracking efficiency, and computational load under varying climatic conditions by applying the AI-driven COMLAT tracking system.

Figure 4 depicts the experimental configuration of the AI-based dual-axis solar tracker system optimized in real-time for solar alignment. Full installation (A) illustrates the real-world application of Computational Machine Learning Adaptive Tracking (COMLAT), which guarantees the improved capture of solar irradiance through smart tracking choice. The system comprises motorized tracking hardware (B) powered by AI control logic to dynamically adjust solar panel tilts. The back structural view (C) displays the mechanical structure supporting the panel, whereas the electronic control module (D) incorporates sensors, microcontrollers, and motor drivers for effective operation.

Test location and climatic conditions

Experimental verification of the COMLAT algorithm was done at Sitapura, Jaipur, India, which was chosen on the basis of its highest solar irradiance and maximum climatic variability. Jaipur is a semi-arid climate with high seasonal swings; hence, it is a suitable test location to test the adaptive solar tracking efficiency under different environmental conditions. Geographical coordinates of the location are 26.85° N latitude and 75.80° E longitude, and the location falls under a high solar potential and highly fluctuating climatic area. Solar radiation in Jaipur is a suitable case study to track optimization analysis throughout a year. The average Direct Normal Irradiance (DNI) was approximately 5.5 kWh/m²/day, which is adequate for panel tracking testing. The weather is seasonal, with hot summers of up to ~ 45 °C, cold winters with a low of ~ 5 °C, and monsoon cloud cover reducing irradiance levels; hence, this location is ideal for testing the real-time adaptability of COMLAT under varying weather conditions. By choosing an area that has active climate change, the experiment seeks to confirm how accurately COMLAT adapts tracking angles from shifting irradiance levels, temperature changes, and cloud cover patterns. Seasonal and daily variation adaptation is a key measure for evaluating the efficiency and dependability of AI-powered solar tracking systems in actual field implementations.

Hardware and sensor setup

There are three solar tracking setups with automated monitoring systems for live data collection. For each setup, several environmental sensors were used to monitor the climatic factors that influence solar efficiency.

Solar panel details

The experimental setup utilized UTL 335 W Mono PERC photovoltaic modules, each comprising 72 monocrystalline cells with a rated power of 335 W. The panels have an open-circuit voltage of 45.0 V and short-circuit current of 9.10 A, with Vmpp and Impp values of 37.03 V and 8.79 A, respectively. With an efficiency of approximately 18.2%, the modules were installed across fixed-tilt, single-axis, dual-axis, and COMLAT-based tracking systems. All panels were mounted at 1.5 m facing true south, and energy output was recorded at 10-second intervals using calibrated inverters and IoT-based data loggers.

Total Panels Installed: 3 (One for Each Tracking Mode).

Setup:

• Static panel: Mounted in a fixed optimal tilt position for maximum yearly yield.

• Single-axis tracker: Turns in the East-West direction as the sun moves.

• Dual-axis tracker: The dual-axis Tracker: Dynamically adjusts the tilt in both azimuth and elevation angles for maximum exposure.

Sensor and data acquisition system

The following modules and sensors were utilized to measure the real-time environmental and electrical parameters:

• Solar irradiance sensors (DNI, GHI, DHI): Monitors real-time radiation intensity.

• Temperature & humidity sensor (DHT22): Monitors thermal variations that impact efficiency.

• Wind speed & direction sensor (Anemometer): Wind adaptation for panel safety.

• Current and voltage sensors (ACS712 and INA219): Keep track of power generation, voltage variation, and power consumption monitoring.

The integration of real-time sensors supports ongoing monitoring and adaptive tracking based on COMLAT’s Deep Q-Learning choices of the COMLAT.

Data logging and computational architecture

The system logs real-time sensor data at 5-second intervals and stores it in a cloud-based database for subsequent processing. The CNN-LSTM-based climate prediction model is executed on an edge computing platform, forecasting 10-day ahead irradiance levels, whereas the XGBoost-based energy yield estimation and Deep Q-Learning-based tracking decisions are executed on local embedded hardware.

• Data collection frequency: Every 5 s.

• Storage format: CSV and SQL database for batch processing & real-time analytics.

• Processing modules:

• CNN-LSTM Climate Prediction Model (Forecasting Solar Irradiance).

• XGBoost Energy Yield Estimator (Predicting Power Output).

• Deep Q-Learning COMLAT Tracker (Adaptive Tilt Angle Selection).

The experimental architecture provides low-latency decision making, enabling COMLAT to adjust the tracking angles in real time with minimal computational costs.

Performance metrics and evaluation criteria

Several performance metrics were used for benchmarking to assess the efficacy of the COMLAT algorithm.

Energy yield improvement (%)

Quantifies the percentage increase in power generation for Single-Axis & Dual-Axis tracking versus static configurations using Eq. 11.

Tracking adaptability score (TAS)

Using the Eq. 12 the measures how efficiently COMLAT adjusts tracking angles in response to climate variations.

Computational cost vs. energy savings

It evaluates whether the AI-driven tracking optimizations provide more energy savings than the computational cost required for real-time learning. Expected Outcome: 20–30% reduction in tracking power consumption.

The COMLAT system combines CNN-LSTM for climate forecasting, XGBoost for energy yield estimation, and Deep Q-Learning for adaptive tracking, making real-time solar panel optimization possible under different climatic conditions.