A Hybrid Machine Learning Approach: Analyzing Energy Potential and Designing Solar Fault Detection for an AIoT-Based Solar–Hydrogen System in a University Setting

Abstract

This research aims to optimize the solar–hydrogen energy system at Kangwon National University’s Samcheok campus by leveraging the integration of artificial intelligence (AI), the Internet of Things (IoT), and machine learning. The primary objective is to enhance the efficiency and reliability of the renewable energy system through predictive modeling and advanced fault detection techniques. Key elements of the methodology include data collection from solar energy production and fault detection systems, energy potential analysis using Transformer models, and fault identification in solar panels using CNN and ResNet-50 architectures. The Transformer model was evaluated using metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), and an additional variation of MAE (MAE2). Known for its ability to detect intricate time series patterns, the Transformer model exhibited solid predictive performance, with the MAE and MAE2 results reflecting consistent average errors, while the MSE pointed to areas with larger deviations requiring improvement. In fault detection, the ResNet-50 model outperformed VGG-16, achieving 85% accuracy and a 42% loss, as opposed to VGG-16’s 80% accuracy and 78% loss. This indicates that ResNet-50 is more adept at detecting and classifying complex faults in solar panels, although further refinement is needed to reduce error rates. This study demonstrates the potential for AI and IoT integration in renewable energy systems, particularly within academic institutions, to improve energy management and system reliability. Results suggest that the ResNet-50 model enhances fault detection accuracy, while the Transformer model provides valuable insights for strategic energy output forecasting. Future research could focus on incorporating real-time environmental data to improve prediction accuracy and developing automated AIoT-based monitoring systems to reduce the need for human intervention. This study provides critical insights into advancing the efficiency and sustainability of solar–hydrogen systems, supporting the growth of AI-driven renewable energy solutions in university settings.

1. Introduction

The integration of renewable energy systems, especially solar–hydrogen systems, is becoming increasingly critical as universities seek sustainable solutions to meet their rising energy needs. Recent studies have focused on optimizing the design and implementation of these systems, with a particular emphasis on their architecture, seamless integration, and the substantial benefits they offer in promoting clean energy usage. These systems typically combine photovoltaic arrays, fuel cells, electrolyzers, and hydrogen storage, integrated with energy management systems tailored to the university’s needs. Research has shown that integrating renewable energy sources with existing grid networks can significantly enhance sustainability and reliability, reducing CO2 emissions and grid dependency while providing cost savings. The optimization of these integrated systems is crucial for meeting varying energy demands, especially in rural areas. While challenges in implementation exist, the potential for sustainable energy harvesting in academic environments is promising, with benefits including increased renewable fraction, reduced emissions, and improved energy management. Solar–hydrogen systems are emerging as promising sustainable energy solutions for universities. These systems integrate solar photovoltaic technology with water electrolysis to produce hydrogen for energy storage and use. The design and implementation of such systems in university settings involve considerations of energy management, infrastructure integration, and economic viability [1,2,3]. While the technology shows potential for reducing campus energy consumption and achieving sustainability goals, challenges remain in terms of scalability and cost-effectiveness. According to economic calculations, the components that absorb light have a greater influence on the cost of producing hydrogen than the materials used in electrolysis. Optimized designs have the ability to produce hydrogen for less than USD 2.90 per kilogram after distribution and compression. In order to solve implementation issues and investigate the long-term effects of solar–hydrogen systems on the economy and education in university settings, more study is required [4,5].

Solar energy is used in solar–hydrogen systems to create hydrogen, a clean, storable energy source. Photovoltaic (PV) cells are commonly utilized in these systems to provide electricity, which is subsequently utilized in electrolyzers to separate water into hydrogen and oxygen [6]. By allowing for long-term energy storage, this strategy tackles the unpredictability of renewable energy sources. These integrated solar PV–hydrogen systems can meet diverse energy needs, powering homes and vehicles while emitting only water vapor [7,8]. Various methods for solar hydrogen production exist, including photochemical, semiconductor, photobiological, and hybrid systems. Research focuses on developing more efficient electrolyzers and advanced storage media to enhance overall system effectiveness. Solar–hydrogen technology offers potential for preserving major energy system options and supporting the transition from fossil fuels to sustainable energy sources [9].

Fault detection in solar panels is crucial for maintaining system efficiency and longevity. Recent research has focused on developing intelligent and automated methods for identifying faults in photovoltaic (PV) systems. IoT-based approaches using smart monitoring devices can detect faults in real-time, improving overall system performance [10]. Advanced techniques such as Bayesian belief networks [11] and fuzzy inference systems combined with multi-resolution signal decomposition [12] have been proposed for more accurate fault detection and diagnostics. These methods can identify various types of faults, including DC-side short-circuits, which are particularly challenging to detect under low-irradiance conditions. Developing efficient fault detection and diagnostic procedures requires a thorough understanding of physical, electrical, and environmental defects in PV systems [13]. Implementing reliable fault detection methods is critical for optimizing energy production, preventing damage to PV panels, and mitigating potential fire hazards. IoT-based approaches using wireless sensor nodes and machine learning algorithms enable real-time fault detection and diagnosis, reducing downtime and maintenance costs. A comprehensive understanding of physical, electrical, and environmental faults in PV systems is crucial for developing effective fault detection techniques. Various methods have been suggested for PV fault diagnostics, focusing on the DC side of the system. Implementing reliable fault detection is essential for optimizing energy production, preventing damage to PV panels, mitigating fire hazards, and extending the operational lifespan of solar energy systems [14,15].

The purpose of this research is to analyze the energy potential of solar–hydrogen systems within university settings, specifically at the Kangwon National University’s Samcheok campus, using advanced machine learning models. Additionally, this study aims to design an efficient fault detection system for solar panels by comparing the performance of Convolutional Neural Networks (CNNs) and ResNet-50. Through this case study, the research seeks to optimize renewable energy systems in academic environments, enhancing their reliability, sustainability, and economic viability.

2. Energy Prediction and Fault Detection in Solar–Hydrogen Systems

2.1. Solar–Hydrogen Systems

Solar–hydrogen systems are emerging as a promising sustainable energy solution, integrating solar power with hydrogen production to address growing energy demands and environmental concerns. These systems can produce clean hydrogen from water using various methods, including electrolysis powered by photovoltaics, thermochemical water splitting, and photoelectrochemical cells. The integration of solar energy and hydrogen production offers multiple benefits, such as electricity generation, heating, cooling, and freshwater production [16,17]. Solar–hydrogen systems can be implemented in distributed residential settings, providing a comprehensive energy solution that combines photovoltaics for immediate electricity needs and hydrogen production for energy storage and transportation. While challenges remain, including cost and efficiency improvements, solar–hydrogen technologies show promise for a sustainable future. Life cycle assessments indicate that photoelectrochemical cells may offer the most sustainable option for hydrogen production when considering environmental impact, resource use, and efficiency [18]. Solar electricity technologies, particularly crystalline silicon photovoltaics and wind turbines, have shown remarkable growth, but producing solar fuels remains more complex [19]. Hydrogen, as a versatile energy carrier, plays a crucial role in the transition to a clean energy economy [20].

Solar-driven water electrolysis, which combines photovoltaic (PV) technology with water electrolyzers, is a promising technique for creating green hydrogen. This device splits water into hydrogen and oxygen via electrolysis, which is powered by PV-generated electricity. The generated hydrogen has the ability to store energy for a long time and be utilized as a clean energy carrier. There are several different setups, such as photoelectrochemical water-splitting systems, PV–water electrolyzer systems, and PV–rechargeable energy storage–water electrolyzer systems. Although wind turbines and crystalline silicon photovoltaics are currently at the forefront of renewable energy deployment, research on alternate photovoltaic technologies and synthetic fuel production methods is still ongoing. Notwithstanding several obstacles, solar-powered water electrolysis exhibits potential for expansion, ranging from domestic to commercial uses, thereby promoting a more sustainable energy environment [21,22].

The development of sustainable energy can benefit greatly from the merging of solar energy with hydrogen generation. Hydrogen is a very efficient means of storing solar energy, mitigating intermittent power outages, and guaranteeing a consistent flow of electricity. These systems help mitigate climate change by reducing greenhouse gas emissions by substituting solar-produced hydrogen for fossil fuels. Because of its adaptability, hydrogen may be used for a wide range of processes in industry, transportation, and the production of energy. Energy security is improved by solar–hydrogen systems because they lessen dependency on imported fuels and centralized infrastructures. These integrated systems have payback times between 2.8 and 6.7 years, demonstrating their economic viability. Although solar hydrogen generation methods have advanced significantly, more development is necessary before they can be widely used [22,23,24] All things considered, integrating solar and hydrogen energy offers a viable way to achieve sustainability and energy independence.

A potential approach to the generation and storage of sustainable energy is the combination of solar–hydrogen systems. Infrastructure for hydrogen storage, electrolyzers, and solar panels are important parts. The goal of research is to increase the scalability and efficiency of these systems, with a focus on photovoltaic–electrochemical, photocatalytic, and photoelectrochemical water-splitting techniques. Combining solar electricity with hydrogen energy storage solves the intermittent nature of renewable energy sources and is in line with carbon reduction objectives. Nonetheless, there are still difficulties in translating basic research into useful applications, such as scaling up and increasing efficiency. This subject is developing as a result of ongoing developments in flexible battery technologies and comparison studies between lithium-ion and hydrogen battery storage [25,26].

Table 1. Key components and benefits of solar–hydrogen integration.

No	Component	Description	Benefits
1	Photovoltaic (PV) Panels	Converts sunlight into electricity.	Renewable energy generation, zero emissions.
2	Electrolyzers	Uses electricity to split water into hydrogen and oxygen.	Produces clean hydrogen, enables energy storage.
3	Hydrogen Storage	Stores hydrogen for later use in fuel cells or other applications.	Long-term energy storage, energy security.
4	Fuel Cells	Converts stored hydrogen back into electricity.	Reliable power generation, supports grid stability.
5	Distribution Infrastructure	Networks for transporting hydrogen to various applications.	Flexibility in energy use, supports diverse energy needs.
6	Energy Management Systems	Optimizes the operation of the integrated solar–hydrogen system.	Enhances system efficiency, reduces operational costs.

lists the key elements of solar–hydrogen systems and the advantages they provide, showing how the combination of solar energy and hydrogen generation promotes the development of sustainable energy sources. This technological integration offers a viable path toward a future in energy that is cleaner, more resilient, and more flexible.

A number of techniques are being investigated for solar hydrogen synthesis, which is a promising way to convert sunlight into sustainable energy [27]. Technologies related to photocatalysis, photoelectrochemistry, solar thermochemistry, photovoltaic–electrolysis, photothermal catalysis, and photobiology are among them [28]. Water is the primary source of hydrogen, which has the ability to mitigate environmental emissions, promote sustainability, and ensure energy security [29]. However, there are still many issues with infrastructure, safety, pricing, manufacture, distribution, and storage. To overcome these challenges, researchers, legislators, business executives, and the general public must work together. Improving hydrogen energy systems’ economy, sustainability, safety, and efficiency are major areas of attention for next research. To fully utilize hydrogen energy, new storage systems, better infrastructure designs, and seamless integration technologies must be developed [30].

2.2. Energy Potential Analysis Using Transformer Models

Transformer models are useful for energy forecasting, especially when it comes to renewable energy sources, as evidenced by recent studies. In a variety of situations, these models do a better job than more conventional techniques like ARIMA and LSTM at capturing long-term dependencies in time series data. Transformers have shown success in predicting electricity consumption, multi-stream load forecasting, and power output for multiple wind farms. Their ability to handle complex correlations and dependencies among multiple data sources makes them particularly valuable for renewable energy forecasting [31,32]. By leveraging contextual information and seasonality trends, Transformer-based models have achieved significant improvements in forecasting accuracy for both wind and solar power generation, reducing prediction errors compared to LSTM and RNN approaches [33]. These advancements highlight the potential of Transformer models in enhancing energy management and grid stability. Recent research has focused on improving long-term forecasting for multi-energy systems using Transformer-based models. These models excel at capturing long-range dependencies, crucial for accurate predictions [34]. Several innovations have been proposed to address limitations in existing approaches. The PWDformer introduces a Deformable-Local aggregation mechanism to enhance time information aggregation and adapt to complex patterns [35]. Patchformer employs patch embedding to better capture local and global semantic dependencies in multivariate time series [36]. DTformer utilizes temporal top windowed attention and dual variable attention to handle extended temporal and intervariable dependencies while regulating computational complexity [37]. These models have demonstrated significant improvements in forecasting accuracy across various datasets, including energy, traffic, and weather. The advancements in Transformer-based models for time series analysis show promise in addressing the challenges of long-term multi-energy load forecasting in integrated energy systems.

Transformer-based models have shown potential; for onshore wind and solar power output, the Informer model has demonstrated better prediction capabilities. Long Short-Term Memory (LSTM) models, however, have occasionally performed better than alternative strategies [38]. Higher accuracy in predicting energy generation rates and demand has been demonstrated by ensemble models utilizing neural networks and support vector machines [39]. In comparison to single models and other hybrid techniques, the H-Transformer hybrid system, which combines SARIMA with a Transformer neural network, has demonstrated better performance [40]. Furthermore, better performance and resilience in wind power prediction have been shown by the Powerformer model, a modified Transformer-based method [41]. Improvements in forecasting accuracy are critical for maintaining grid stability and integrating renewable energy sources. Transformer-based model use for renewable energy forecasting has been investigated recently. Transformers have demonstrated their efficacy in solar irradiance prediction, with a maximum worst-case mean absolute percentage error of 3.45% achieved for forecasts made one day in advance [42]. A hybrid CNN–LSTM–Transformer model was suggested for forecasting solar energy production, and it performed more accurately than previous models [43]. SolarFormer, a multi-scale Transformer model that meets or outperforms cutting-edge models in the field of solar panel identification from aerial data, was introduced by [44]. Expanding the use to wind energy, [45] created a ChatGPT-inspired Transformer-based model that surpassed conventional forecasting models in accuracy. Together, these studies show how Transformer-based models can enhance forecasting and mapping of renewable energy, thereby advancing the field of sustainable energy.

Table 2. Applications and benefits of transformer models in energy forecasting.

No	Application	Description	Benefits
1	Solar Power Generation Forecasting	Predicts future solar energy output based on historical weather and irradiance data.	Improved grid integration, optimized energy storage and usage.
2	Wind Energy Output Prediction	Forecasts wind power production by analyzing patterns in wind speed and direction.	Enhanced reliability of wind energy, better resource allocation.
3	Energy Demand Forecasting	Estimates future energy consumption by considering factors like temperature and usage trends.	Balanced energy supply and demand, reduced risk of outages.
4	Seasonal Trend Analysis	Captures long-term patterns in energy production and consumption.	Supports long-term planning and system optimization.
5	Regional Energy Grid Forecasting	Scales predictions to cover larger geographical areas, such as cities or regions.	Enables efficient energy distribution and management.

lists the numerous ways Transformer models are used in energy forecasting, especially in relation to renewable energy sources [46,47,48,49,50]. Using Transformers can lead to increased integration of renewable energy sources into the grid, better resource management, and more accuracy in estimating energy outputs. These models present a viable method for maximizing renewable energy systems and guaranteeing a steady supply of electricity in an environmentally friendly way.

When compared to more conventional models like Autoregressive Integrated Moving Average (ARIMA) models and Long Short-Term Memory (LSTM) networks, Transformer models have shown to be a potent tool for energy prediction. Since each of these models has advantages and disadvantages, they can be used for various energy forecasting jobs. In order to anticipate energy usage, recent research has compared deep learning techniques—specifically, LSTM networks—with conventional time series forecasting techniques like ARIMA. Although ARIMA has proven to be a strong statistical analytic tool, LSTM models have shown to be more accurate in predicting energy use [51]. According to empirical research, LSTM performs better than ARIMA, reducing error rates on average by 84–87% [52]. Strong seasonal patterns in the data, however, may be necessary for these models to function well [53]. Other deep learning architectures, such as RNN, GRU, and Time Series Transformers, have been studied more recently for forecasting home energy usage. Despite possible overfitting problems, results imply that Transformer designs function better with fewer samples. Additionally, a 23% improvement in hourly granularity forecasting has been demonstrated by optimizing a voting ensemble using the Simulated Annealing metaheuristic [54].

2.3. Machine Learning Techniques for Fault Detection in Solar Panels

Because Convolutional Neural Networks (CNNs) are so good at processing images and recognizing patterns, they have become a popular method for identifying defects in solar panels. Imaging techniques such as electroluminescence imaging and infrared thermography are widely used in solar panel monitoring to identify anomalies and faults that may not be readily evident. Because CNNs can automatically learn and extract features from the data, they are especially well-suited for analyzing these images and enabling precise defect identification. With their great accuracy and efficiency, Convolutional Neural Networks (CNNs) have become a potent tool for identifying solar panel defects. Research has indicated that Convolutional Neural Networks (CNNs) are a useful tool for identifying flaws including fractures, shading, and possible induced degradation in a variety of image forms, including true color and electroluminescence images. With accuracy rates ranging from 91.1% for binary classification to 98.07% for multi-class defect detection, CNN-based models have demonstrated remarkable performance. These models perform better than manual feature detection techniques and are more robust against changes in input data. Additionally, to improve defect detection capabilities even further, CNNs can be coupled with other machine learning methods like Decision Trees and Support Vector Machines. The solar sector stands to benefit greatly from the revolutionary quality control that CNN-based fault detection systems promise to bring about. This will enhance module dependability and facilitate the rise of renewable energy sources [55,56,57,58].

Automated fault identification of solar panel faults with the application of deep learning algorithms. Convolutional Neural Networks (CNNs) have demonstrated encouraging outcomes in the identification of different fault types, including as dust, bird marks, and microcracks. High accuracy in defect classification has been shown using transfer learning algorithms that use pre-trained models such as VGG16, VGG19, and AlexNet; some studies have reported accuracy rates above 90%. The application of Fully Convolutional Neural Networks (F-CNNs) has also been utilized to attain pixel-level prediction of photovoltaic backsheet degradation mechanisms. These automated inspection methods are much more efficient in maintaining solar panel performance than manual ones, and they also reduce subjectivity and human error. The accuracy of flaw identification in monocrystalline and polycrystalline solar panels has been significantly improved by the use of ensemble learning algorithms [59,60,61,62]. Deep learning approaches have been applied in recent research to build effective solutions for defect identification and classification in solar photovoltaic (PV) systems. In categorizing different types of faults using thermographic images and recognizing PV arrays in aerial photos, Convolutional Neural Networks (CNNs) have demonstrated encouraging outcomes. These methods have proven to be highly accurate in identifying a variety of fault types, such as offline modules, hot spots, and cracking. In several research, the use of RGB images for fault classification has been investigated, and cracks, shadows, and dust have been identified with a moderate degree of accuracy. Transfer learning and multi-scale CNN architectures have been employed to improve classification performance and overcome challenges such as imbalanced datasets [63,64,65]. Additionally, researchers have proposed methods combining meta-heuristic algorithms for parameter extraction with CNN-based fault classification, showing promising results in various atmospheric conditions [66].

Table 3. CNN applications in solar panel fault detection.

No	Aspect	Description	Strengths
1	Image Processing	Analyzes solar panel images to detect defects like cracks or hotspots.	High accuracy in fault identification and classification.
2	Pattern Recognition	Learns hierarchical features to recognize subtle degradation patterns.	Effective in early detection of potential issues.
3	Scalability	Handles large volumes of image data efficiently.	Ideal for large-scale solar installations with automated monitoring.
4	Versatility	Can be combined with other data sources for enhanced fault detection.	Allows for a comprehensive assessment of solar panel health.
5	Computational Requirements	Requires substantial computational resources for training.	Delivers rapid and accurate fault detection once trained.

details the key aspects of CNNs in detecting faults in solar panels, emphasizing their advantages in image processing, pattern recognition, and scalability [67,68,69,70,71]. CNNs excel at detecting and classifying faults, making them a vital component of contemporary solar monitoring systems. Their capacity to handle large datasets and integrate with other technologies further boosts their effectiveness in ensuring the efficiency and reliability of solar energy systems.

A deep convolutional neural network architecture called ResNet-50 has shown success in a number of image categorization applications. Deeper networks can be trained by utilizing residual connections, which preserve information flow and stop gradient dissipation [72]. ResNet-50 has been successfully applied to breast cancer classification using MRI scans, achieving 92.01% accuracy [73]. In emotion recognition, ResNet-50’s deep architecture allows for capturing complex patterns, offering advantages in performance and computational efficiency [74]. For poultry disease recognition, combining SURF feature analysis with K-means clustering to preprocess images before feeding them into ResNet-50 improved accuracy by 20%, reaching 93.56% [75]. These studies highlight ResNet-50’s versatility and potential in various domains, including medical imaging, biometrics, and computer vision applications, demonstrating its ability to extract meaningful features and achieve high classification accuracy.

By solving the vanishing gradient issue in extremely deep networks, ResNet architectures have completely changed the field of deep learning. The main invention is the residual block, which uses skip connections to make information flow easier. ResNet variants, such as ResNet-50 and ResNet-152, differ in depth but share the same fundamental building blocks [76,77]. In an effort to improve feature map exploration, recent research has looked into ways to improve ResNet performance. For example, [78] proposed ResNetX, which adds a “fold depth” dimension and increases network disorder. Additionally, [79] showed that by using contemporary training techniques on the original ResNet-50 architecture, remarkable results can be achieved, including 80.4% top-1 accuracy on ImageNet without the need for additional data or distillation.

ResNet-50’s architecture demonstrates scalability across various applications and hardware implementations. It can be efficiently accelerated on FPGAs, achieving high throughput and low latency for complex image recognition tasks [80]. The model’s scalability extends to large-scale distributed training on supercomputers, where it achieves over 90% scaling efficiency using up to 104 K ×86 cores, reducing training time to just 28 min. Additionally, the introduction of the Collapsed Ensemble technique allows ResNet-50 to achieve accuracy comparable to ResNet-152 while maintaining its original topology, further highlighting its scalability and potential for improved performance [81,82]. ResNet-50’s versatility is further demonstrated by its application in Wi-Fi fingerprint-based indoor localization systems for several buildings and floors, which exhibit near-state-of-the-art performance with lower complexity and energy consumption [83].

Table 4. ResNet-50 in solar panel fault detection.

No	Aspect	Description	Strengths
1	Deeper Architecture	Comprises 50 layers, enabling it to capture complex patterns and features.	Enhanced ability to detect subtle and complex faults in solar panels.
2	Residual Learning	Utilizes shortcut connections to prevent vanishing gradients.	Improves learning efficiency in deep networks.
3	Feature Extraction	Excels at identifying minute differences in image data.	High accuracy in fault detection across varied fault types.
4	Scalability	Suitable for large-scale solar installations with diverse fault detection needs.	Ensures reliable performance in extensive solar monitoring systems.
5	Computational Requirements	Requires significant computational resources and longer training times.	Delivers superior accuracy and reduces false positives.

outlines the key elements of ResNet-50 in solar panel fault detection, highlighting its deep architecture, advanced feature extraction abilities, and scalability. The model’s capacity to accurately detect and classify intricate faults makes it a vital component of sophisticated solar monitoring systems. Although ResNet-50 requires more computational resources and longer training periods, its enhancements in fault detection accuracy and system reliability make it an invaluable tool for sustaining the performance and efficiency of solar energy systems [84,85,86,87,88].

2.4. Emerging Trends in Renewable Energy Monitoring

Monitoring and fault detection for renewable energy is being revolutionized by the merging of AI and IoT (AIoT) technologies. Real-time data collecting from several renewable energy sources is made possible by IoT devices, and AI algorithms can examine this data to identify anomalies and schedule maintenance. The Isolation Forest algorithm, for instance, has shown effectiveness in identifying faults in renewable energy systems. AI-driven predictive maintenance minimizes downtime, extends equipment lifespan, and optimizes energy output by adapting to changing environmental conditions. These developments help to increase the effectiveness, dependability, and affordability of renewable energy installations. As artificial intelligence (AI) technology develops, it is anticipated that its influence on renewable energy will intensify, hastening the shift to greener energy sources [89,90,91]. Insights, anomaly identification, and fault diagnosis for hybrid energy systems are provided by real-time monitoring systems that integrate AI and IoT, supporting sustainability and effective energy management [92]. The combination of artificial intelligence (AI) with the Internet of Things (IoT) for intelligent monitoring and predictive maintenance across multiple domains, including smart grids, has been the focus of recent study. IoT sensor data abnormalities are being found using machine learning techniques like autoencoders, one-class SVM, and isolation forests in order to develop proactive maintenance plans [93]. With temperature detection accuracy reaching 98.7%, real-time monitoring paired with AI-based predictive analytics has demonstrated notable gains in the early detection of problematic trends [94]. Large-scale IoT data processing requires artificial intelligence (AI) in order to support both post-event and real-time analysis [95]. Machine learning algorithms are being utilized in renewable energy systems to optimize smart grids, estimate energy supply and demand, and increase energy efficiency. These developments are expected to improve grid sustainability, lower operating costs, and improve asset management [96].

The integration of AI and IoT enhances fault detection capabilities in industrial systems. The precision and accuracy of fault identification and localization can be increased by training deep learning models to identify certain fault signs in IoT sensor data [97]. Machine learning techniques, combined with IIoT architectures, enable real-time data collection and processing for fault detection, prediction, and prevention [98]. Energy-aware intelligent fault detection schemes using optimized deep learning mechanisms can significantly improve fault detection accuracy while reducing false alarm rates and energy consumption in IoT-enabled wireless sensor networks [99]. Ensemble deep learning techniques, such as PropensityNet, DNN, and CNN-LSTM, can be applied to fused sensor data from IIoT devices for efficient fault diagnosis [100]. These developments in IoT and AI integration lead to safer and more effective industrial operations, lowering maintenance costs and downtime. The integration of AI and IoT in renewable energy systems is summed up in

Table 5. AI and IoT (AIoT) integration in renewable energy monitoring.

No	Aspect	Description	Benefits
1	Real-Time Monitoring	Continuous data collection by IoT devices across renewable energy systems.	Enables timely detection of anomalies and optimizes system performance.
2	Predictive Maintenance	AI analyzes historical and real-time data to forecast equipment failures.	Reduces downtime and maintenance costs, extends equipment lifespan.
3	Fault Detection	AI models identify specific fault patterns from IoT sensor data.	High accuracy in detecting faults, leading to prompt corrective actions.
4	Energy Optimization	AI adjusts operations based on real-time conditions, like weather or demand.	Maximizes energy production efficiency and reliability.
5	Challenges	Issues include data security, network reliability, and computational demands.	Overcoming these challenges is crucial for widespread adoption.

, which also highlights the advantages that come with it, including improved problem detection, real-time monitoring, and predictive maintenance [101,102,103,104,105,106,107]. AIoT, or artificial intelligence and the Internet of Things, not only makes renewable energy systems more reliable and efficient but also comes with problems that must be overcome in order to reach its full potential. This tendency looks to be very important for managing renewable energy in the future as it develops.

3. Research Methodology

3.1. Data Collection

In order to evaluate energy potential and find solar panel defects, gathering data is an essential part of this study process. The energy data are sourced from Kangwon National University’s Samcheok campus, with a particular focus on the Green Energy Building as the sample site. These data encompass metrics related to solar energy production, including historical weather patterns, solar irradiance levels, temperature, and humidity. These variables are vital for predicting energy output using advanced machine learning models such as the Transformer model. Historical weather data offer insights into seasonal trends while solar irradiance data reflect the available sunlight, both essential for accurate energy forecasting.

For solar panel fault detection, the study collects data in the form of images and sensor readings. This includes thermographic images that reveal temperature variations on the solar panel surfaces, which can indicate potential issues like hotspots or degradation. Furthermore, sensor data—such as voltage and current measurements—assist in locating abnormalities in performance. Convolutional Neural Networks (CNNs) and ResNet-50 are two machine learning models that are trained with these photos and sensor data in order to automatically identify and categorize solar panel defects. The university’s solar–hydrogen system analysis and problem detection procedures depend critically on the integration of these datasets.

3.2. Energy Potential Analysis Using Transformer Model

This section provides a detailed explanation of the Transformer model architecture implemented in Python, used to predict energy output in solar–hydrogen systems. The Transformer model was first created for natural language processing, but because of its capacity to identify long-term dependencies in data, it has grown in popularity for time series forecasting. Because of the self-attention mechanisms built into its architecture, the model is able to assess the importance of different time steps in the input data, which makes it a powerful tool for studying complex, multivariate datasets, such as those found in the energy production industry.

A dataset of historical energy production data, meteorological data, and solar irradiance levels gathered from the university’s solar energy installations is used to train the Transformer model. Subsets of the dataset are used for testing, validation, and training to make sure the model learns well and can be applied to fresh, untested data. To maximize model performance, important training parameters including learning rate, batch size, and number of epochs are carefully adjusted. Using patterns from past data, the model is trained to predict future energy output. Overfitting is avoided by using validation procedures to monitor the model’s development.

Several metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), and Mean Absolute Error Squared (MAE2), are used to assess the Transformer model’s performance. An easy-to-understand indicator of prediction accuracy, MAE yields an average of the absolute disparities between values that were anticipated and those that were observed. By accounting for the square of the differences, MSE helps to discover notable outliers by assigning greater weight to larger errors. The MAE2 variance, which squares the MAE, highlights bigger disparities even more, enabling a more thorough assessment of the model’s forecasting performance. When combined, these measures offer a thorough evaluation of how well the Transformer model predicts energy output in a solar–hydrogen system.

3.3. Solar Panel Fault Detection Approach

This section outlines the methodology for solar panel fault detection, with a focus on the architectures of ResNet-50 and Convolutional Neural Networks (CNNs), along with the metrics used for performance evaluation and training, implemented in Python.

CNNs are frequently utilized in image processing and are very good at locating solar panel defects. Convolutional layers, one of the several layers that make up CNN architecture, are capable of automatically identifying elements in images, such as edges, textures, and patterns that point to defects like fractures or hotspots. The network’s efficiency is increased by further processing these features through pooling layers, which lower the dimensionality of the data. The retrieved features are then interpreted by the fully connected layers at the end of the network to determine whether a problem is present in the solar panel photos.

By adding residual connections, ResNet-50, a more sophisticated deep learning model, expands upon the fundamental CNN architecture. The network can go deeper thanks to these connections without running into the vanishing gradient issue. ResNet-50, with its 50 layers, can extract more complicated characteristics and improve the accuracy of complex defect detection. Compared to a standard CNN, the residual connections allow the model to learn more abstract and detailed features while maintaining its efficiency as it goes deeper.

Important measures like data augmentation are done to the training dataset to increase the variety of input images and boost the models’ capacity to generalize to new data in the training process for both models. Color corrections, flips, and rotations of images may be part of this procedure. The models are trained across a number of epochs using optimization methods to minimize the loss function, such as the Adam optimizer or stochastic gradient descent. To prevent overfitting and guarantee that the models can correctly identify errors in fresh, unviewed photos, the training process is closely watched.

Several measures are utilized to assess and compare CNN’s and ResNet-50’s performance. Precision shows if the flaws that have been recognized are accurate, whereas accuracy evaluates the percentage of faults that have been identified correctly. The model’s recall evaluates its capacity to find all potential errors, and the F1-score offers a trade-off between recall and precision. Since faster models are preferred for real-time monitoring, inference time—the amount of time needed for the model to analyze an image and create a prediction—is also taken into account. When combined, these measures offer a comprehensive assessment of the two architectures’ efficacy and efficiency in solar panel fault detection.

4. Results and Discussion

4.1. Energy Potential Analysis

The three main metrics used in this section to assess the Transformer model’s prediction of the university’s solar hydrogen energy potential are Mean Absolute Error (MAE), Mean Squared Error (MSE), and Mean Absolute Error 2 (MAE2). Forecasts of energy production that are accurate are necessary to maximize energy storage, resource management, and the overall effectiveness of renewable energy systems. The university’s solar hydrogen output was predicted using the Transformer model, a potent machine learning method renowned for its capacity to spot intricate patterns in time series data.

The average difference between the model’s predictions and the actual energy output, as indicated by the MAE result (Figure 1) of 1685.04, is around 1685 units. The average prediction error (MAPE), which may be simply interpreted, is a simple metric that gives information about the overall accuracy of the model.

The average of the squared discrepancies between the expected and actual values is represented by the MSE result (Figure 2), which is 4,361,833.37. Larger errors are given more weight in this metric, since it squares the errors before averaging. Even if the majority of predictions are quite near to the actual values, there are certain cases with sizable mistakes, which raise the overall MSE, as indicated by the relatively high MSE.

The MAE2 result (Figure 3) of 1685.038, which is a variation of the MAE, underscores the importance of minimizing errors. This result, closely aligning with the original MAE, indicates a consistent average deviation of around 1685 units, even when the error is squared.

These metrics form a comprehensive evaluation of the Transformer model’s accuracy. The relatively low MAE and MAE2 suggest the model is generally effective, while the higher MSE points to areas where larger errors might need to be addressed, especially when extreme values are involved. This analysis feeds directly into the broader narrative of optimizing renewable energy systems by improving prediction accuracy, leading to more efficient energy management. The detailed evaluation using these metrics directly informs both the Model Performance and Error Analysis sections. By understanding how the Transformer model performs in predicting solar hydrogen energy production, we can identify its strengths and weaknesses, ultimately guiding future improvements and refinements to enhance its real-world application.

The Transformer model’s ability to forecast the university’s solar hydrogen energy potential is evaluated visually in the graph in Figure 4 (Model Performance). The model’s predictions (represented by the orange line) are compared against the actual measured values (blue line) over a substantial period, revealing key insights into its efficacy. Overall, the model demonstrates a commendable ability to track the fluctuations and trends in solar hydrogen production. The close alignment of the predicted and real values in many instances suggests that the Transformer model is effective at learning and generalizing the patterns of energy generation, which are influenced by factors such as sunlight availability, seasonal variations, and perhaps the operational efficiency of the solar–hydrogen systems. This alignment implies that the model can serve as a reliable tool for forecasting the university’s energy potential, aiding in the strategic planning of energy use and storage.

Despite the Transformer model’s generally strong performance, the graph also uncovers several areas where its predictions fall short, highlighting some limitations and sources of error (Error Analysis). Notably, there are significant discrepancies between the predicted and actual values during certain periods, particularly around the 1800 to 2100 mark on the X-axis. These deviations suggest that the model occasionally struggles with accurately predicting the solar hydrogen output, leading to instances of both overprediction and underprediction. These errors might stem from the model’s inability to fully account for sudden or extreme changes in environmental conditions, such as unexpected weather events that alter solar irradiance. Additionally, the model may have been trained on data that did not sufficiently capture the full range of possible conditions, leading to less accurate predictions in outlier scenarios. These issues highlight the importance of ongoing model refinement.

To mitigate these errors and improve the model’s predictive accuracy, several approaches could be considered. First, incorporating more comprehensive environmental data, such as real-time weather information or atmospheric conditions, might help the model better anticipate fluctuations in solar energy production. Secondly, enhancing the model’s architecture or employing techniques like ensemble learning could improve its robustness and ability to generalize from diverse datasets. Lastly, a thorough analysis of the periods of significant deviation could provide insights into the specific factors that the model currently overlooks, guiding targeted improvements. By addressing these limitations, the Transformer model could be further optimized to provide more reliable and accurate forecasts of the university’s solar hydrogen energy potential.

4.2. Fault Detection Performance

Reduced energy production results from trash accumulation on solar panels, including dust, snow, bird droppings, and other debris. This reduces the panels’ capacity to convert sunlight into energy. To ensure that solar panels remain efficient, regular monitoring and cleaning are crucial. Implementing a systematic monitoring and cleaning procedure is essential for optimizing resource use, lowering maintenance costs, and enhancing the efficiency of the panels. By establishing a well-planned routine for monitoring and cleaning, solar panel owners can maximize energy production, extend the lifespan of their panels, and contribute to broader sustainability goals. The goal of this dataset is to assess how well different machine-learning classifiers detect dust, snow, bird droppings, and mechanical and electrical problems on solar panel surfaces. Six different class folders are included in the dataset for classification: mechanical damage, electrical damage, snow, bird droppings, garbage, and dirt (Figure 5). There is a small imbalance in the quantity of photos gathered, because the data are taken from the internet.

To ensure the integrity and quality of the dataset used for training a machine learning model, several steps are involved in the data verification process. Preparing the data for training requires performing pretreatment tasks, such as cleaning, normalization, and feature engineering. It is important to review the preprocessed data for any anomalies, inconsistencies, or missing variables that could affect the model’s performance. This may include visualizing the data using plots or graphs to identify patterns or outliers. Additionally, addressing class imbalances where certain classes are overrepresented or underrepresented is critical to mitigate potential biases in the model.

It is critical to divide the dataset into training and validation sets in order to evaluate the model’s performance on untested data. This guarantees the robustness and generalizability of the model by enabling the use of cross-validation procedures. Maintaining the model’s accuracy and efficacy over time also depends on constant observation and updating of the training data as new information becomes available. Machine learning practitioners can create more accurate and dependable models for a variety of applications, such as solar panel fault detection, by carefully analyzing the training data.

The results of defect detection models applied to solar panel images reveal differing levels of performance between two architectures, VGG-16 and ResNet-50, as presented in the “Prediction Results” section (Figure 6). The model based on the VGG-16 architecture achieved an accuracy of 80% with a loss value of 78%. This indicates that while the model correctly identified 80% of instances, there is still a significant gap between the predicted and actual labels, reflected in the higher loss value. Conversely, the ResNet-50-based model demonstrated improved performance, with an accuracy of 85% and a reduced loss value of 42%. This suggests that the ResNet-50 model is better at identifying patterns and features in solar panel images, leading to more accurate predictions and a smaller prediction error. Despite these advances, there remains room for further optimization to minimize errors and enhance the precision of fault detection in solar panels. Overall, both models show promising results in identifying defects, with ResNet-50 outperforming VGG-16 in accuracy and loss metrics.

Examining flaw detection models on solar panel photos reveals the relative advantages and drawbacks of two well-known architectures: VGG-16 and ResNet-50 (Figure 7). With an accuracy of 80%, the VGG-16 model demonstrated a respectable capacity to accurately detect solar panel flaws. The high loss value of 78%, on the other hand, indicates a significant difference between the actual and predicted labels, indicating that the model may not be completely accurate in capturing the intricate patterns present in the data. In contrast, the ResNet-50 model demonstrated superior performance with an accuracy of 85% and a significantly lower loss value of 42%. This improvement suggests that ResNet-50’s deeper architecture and advanced feature extraction capabilities allow it to more effectively recognize intricate patterns and features in the solar panel images, resulting in more accurate predictions and a lower prediction error. The reduced loss in the ResNet-50 model indicates that it is better at minimizing errors in its predictions, making it more reliable for fault detection in solar panels. However, even with this enhanced performance, the presence of a 42% loss value reveals that there is still room for further refinement to enhance accuracy and reduce errors. This could involve fine-tuning the model parameters, improving the quality of the training data, or integrating additional data preprocessing steps. Overall, the analysis underscores the potential of ResNet-50 as a more effective model for solar panel defect detection compared to VGG-16, particularly in terms of accuracy and loss. Nonetheless, both models exhibit promising capabilities, and with further optimization, they could provide even more reliable tools for maintaining and monitoring the efficiency of solar energy systems.

5. Future Prospects

5.1. Analysis and Future Directions

This study assesses how well machine learning models perform in two crucial domains: energy potential analysis, using the Transformer model, and solar panel failure detection. There are significant distinctions between the ResNet-50 and VGG-16 architectures when defect detection models are analyzed. With a loss value of 78%, the VGG-16 model had a comparatively high accuracy of 80%. This suggests that although the model was able to accurately detect most errors, it had difficulty with more intricate or subtle problems, which led to a substantial discrepancy between expected and actual results. Conversely, the ResNet-50 model outperformed the others, with an accuracy of 85% and a significantly lower loss value of 42%. ResNet-50’s deeper design, which enables it to extract more complex characteristics and more accurately recognize patterns in solar panel photos, is responsible for this improvement. ResNet-50 appears to be a more dependable model for identifying solar panel defects, as seen by its lower loss value, which also shows that it is more successful at reducing prediction errors. Even with these improvements, there is still opportunity for additional optimization to raise the model’s precision and lower its prediction error.

Three important metrics were utilized to assess the Transformer model’s performance in the context of energy potential analysis: Mean Absolute Error (MAE), Mean Squared Error (MSE), and a variation known as MAE2. The Transformer model was used to estimate the university’s solar hydrogen energy output. An accurate indicator of the model’s accuracy is the MAE result of 1685.04, which shows that the model’s predictions and the actual energy production differ by an average of 1685 units. The MSE result of 4,361,833.37 suggests that while the majority of predictions are close to the actual values, there are some significant outliers that increase the overall error. This is because MSE places more weight on larger errors by squaring them. The MAE2 result of 1685.038, which is very close to the original MAE, shows a consistent average error, further validating the model’s performance. Despite these generally strong results, the Transformer model does show some limitations, particularly in scenarios where the energy output varies significantly from the norm, suggesting that further refinement is necessary.

Moving forward, several strategies could be implemented to improve both the fault detection models and the energy potential prediction model. For fault detection, incorporating more diverse training data, such as additional environmental factors, could help models like ResNet-50 further improve their accuracy. Additionally, hybrid approaches that combine different architectures might leverage the strengths of each to better handle complex or subtle faults. For the Transformer model used in energy potential analysis, integrating more comprehensive environmental data, such as real-time weather conditions, could help the model anticipate and account for sudden changes in solar energy output. Further, exploring advanced model architectures or ensemble techniques could enhance the model’s ability to generalize from the data and reduce significant prediction errors. In both areas, ongoing model refinement, real-time monitoring, and continuous updates to training data will be crucial for maintaining the accuracy and reliability of these machine-learning applications. By addressing these areas, the models can be better optimized to support the efficient and sustainable operation of the solar–hydrogen energy systems at the university and potentially in broader applications.

5.2. AIoT-Based Solar–Hydrogen System in the University

Energy management and sustainability can be significantly improved by integrating artificial intelligence (AI) and Internet of Things (IoT) technology into a university’s solar–hydrogen system (

Table 6. Key aspects of the AIoT-based solar–hydrogen system in the university.

No	Aspect	Details
1	Real-Time Monitoring	Continuous tracking of solar irradiance, hydrogen production, weather conditions, and equipment status.
2	AI-Driven Optimization	AI algorithms optimize energy production, predict maintenance needs, and enhance system efficiency.
3	Predictive Maintenance	Anticipates equipment failures to reduce downtime and maintenance costs.
4	Sustainability Impact	Maximizes renewable energy efficiency, reduces reliance on traditional energy, and lowers carbon footprint.
5	Educational Value	Provides students and faculty with hands-on experience in advanced energy technologies.
6	Operational Efficiency	Enhances system reliability and longevity by responding to real-time data and optimizing operations.
7	Cost Savings	Reduces operational costs through efficient energy management and predictive maintenance.

). Using IoT sensors and AI-driven analytics, this AIoT-based system improves predictive maintenance, real-time monitoring, and overall system efficiency. Critical parameters including solar irradiance, hydrogen generation, weather, and equipment performance are continuously monitored via IoT sensors. Artificial intelligence algorithms analyze the gathered data to forecast maintenance requirements, enhance operational effectiveness, and maximize energy generation and storage. For example, AI can automatically modify operations to maintain constant hydrogen production levels if the system senses a drop in solar irradiation.

The university gains a number of advantages from this combination. Quick responses to changes in system performance are made possible by real-time monitoring and control, and AI-driven predictive analytics helps to detect equipment failures, which lowers maintenance costs and downtime. Furthermore, by maximizing the efficiency of renewable energy sources and lowering the carbon imprint, the system helps the university achieve its sustainability goals. The AIoT-based solar–hydrogen system also functions as a research and teaching tool, giving academics and students hands-on exposure with cutting-edge energy technologies. In addition to enhancing energy management, this arrangement supports the university’s overarching goals of advancing sustainability and cutting-edge energy solutions.

6. Conclusions

This study focused on integrating advanced technologies such as artificial intelligence (AI), machine learning, and the Internet of Things (IoT) into a solar–hydrogen system at Kangwon National University’s Samcheok campus. The primary goal was to enhance the efficiency, reliability, and sustainability of renewable energy systems, particularly in an academic setting. The findings offer valuable insights into the application of data collection, predictive modeling, and fault detection techniques to improve solar energy output and storage while ensuring the long-term performance of solar panels. The integration of solar–hydrogen systems represents a significant advancement in the pursuit of sustainable energy solutions. These systems combine solar energy generation with hydrogen production and storage to address the intermittent nature of solar power, resulting in a flexible and reliable energy source. The Green Energy Building at the university’s Samcheok campus serves as an effective demonstration of how such technologies can be implemented in a real-world setting. This solar–hydrogen system increases the campus’s energy independence, offering a cleaner alternative to fossil fuels. Even during periods of low solar irradiance, the system provides a stable and reliable energy supply by storing excess energy as hydrogen. This setup reduces the university’s carbon footprint, decreases reliance on external energy sources, and offers a model for other institutions seeking to adopt similar technologies.

Accurate data collection is crucial for optimizing any energy management system. This study gathered solar energy data from various sources, including sensor data from solar panels, solar irradiance measurements, and historical weather data. This information is essential for understanding the performance of the solar–hydrogen system and for developing predictive models that improve system operations. Additionally, thermographic images and sensor data were used for fault detection in solar panels. This enabled the identification of issues like dust accumulation, mechanical damage, and electrical defects, which could otherwise reduce energy output. The use of machine learning algorithms further enhanced the reliability and efficiency of the solar–hydrogen system. One of this study’s main objectives was to assess the Transformer model’s ability to forecast energy output from the university’s solar–hydrogen system. Transformer models are well-regarded for their capacity to detect complex patterns in time-series data. The model’s performance was evaluated using three key metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), and a second variation of MAE (MAE2). The Transformer model performed well in predicting energy production, with an MAE of 1685.04, an MSE of 4,361,833.37, and an MAE2 of 1685.038. These results indicate that the model’s predictions were generally accurate, with errors averaging around 1685 units. However, the relatively high MSE suggests that there are cases where the model’s predictions deviate significantly from actual values, highlighting areas where further refinements are needed to improve accuracy.

The research also investigated the use of Convolutional Neural Networks (CNNs) and ResNet-50 for fault detection in solar panels. These deep learning models are highly effective in processing and analyzing thermographic images of solar panels. The CNN model, based on the VGG-16 architecture, achieved an accuracy of 80% but with a high loss value of 78%, indicating some challenges in fault detection. In comparison, the ResNet-50 model demonstrated superior performance, achieving 85% accuracy with a significantly reduced loss value of 42%. This highlights ResNet-50’s better ability to capture complex patterns in the data, leading to more accurate detection of solar panel defects. Despite this success, further optimization is required to reduce prediction errors and enhance the model’s capacity to identify subtle defects. Several key discoveries emerged from the analysis of these models. The Transformer model’s ability to predict energy output shows its potential for optimizing solar–hydrogen systems. However, the model’s relatively high error rates in certain cases suggest that additional improvements are necessary, particularly to increase its accuracy in more challenging scenarios. The comparison of ResNet-50 with VGG-16 demonstrated that ResNet-50 is more suitable for fault detection tasks, given its higher accuracy and lower loss values. Nonetheless, further development is needed to improve the model’s performance in detecting minor defects that may not be immediately apparent.

Building on the results of this study, several avenues for future research can further improve the effectiveness and reliability of solar–hydrogen hybrid systems. First, integrating more comprehensive environmental data—such as real-time weather reports or air quality measurements—could enhance the Transformer model’s accuracy by reducing prediction errors and improving its ability to account for fluctuations in solar energy production. Second, fault detection models could be improved by exploring alternative deep learning architectures or incorporating additional data sources, such as vibration sensors or acoustic emissions. Using ensemble learning techniques, which combine multiple models to increase overall performance, could also strengthen fault detection. Finally, expanding the use of AI and IoT technologies could lead to the development of a fully automated monitoring and maintenance system. Such a system would enable real-time issue detection and correction, minimizing downtime and maximizing energy production. Predictive analytics could also be used to optimize energy usage and storage, ensuring that the university’s energy infrastructure operates at maximum efficiency at all times. This study highlights the potential of integrating AI, IoT, and machine learning into solar–hydrogen systems to improve renewable energy management in academic institutions. The Transformer model demonstrated potential in energy output prediction, while ResNet-50 showed superior performance in fault detection. However, there remains room for improvement in both models. Future research should focus on refining these models, incorporating more data sources, and exploring the scalability of solar–hydrogen systems in diverse settings. Through continued advancements in technology, institutions can contribute to the global transition to sustainable energy sources.