FACULTAD DE INGENIERÍA Escuela Académico Profesional de Ingeniería Eléctrica Tesis Implementation of an Intelligent Ground Fault Protection System for Pump Chambers Using Artificial Intelligence Networks Walter Huacho Ichpas Danny Javier Rojas Fierro Jezzy James Huaman Rojas Para optar el Título Profesional de Ingeniero Electricista Huancayo, 2025 Esta obra está bajo una Licencia "Creative Commons Atribución 4.0 Internacional" . INFORME DE CONFORMIDAD DE ORIGINALIDAD DE TRABAJO DE INVESTIGACIÓN A : Decano de la Facultad de Ingeniería DE : Jezzy James Huamán Rojas Asesor de trabajo de investigación ASUNTO : Remito resultado de evaluación de originalidad de trabajo de investigación FECHA : 25 de Agosto de 2025 Con sumo agrado me dirijo a vuestro despacho para informar que, en mi condición de asesor del trabajo de investigación: Título: Implementation of an Intelligent Ground Fault Protection System for Pump Chambers Using Artificial Intelligence Networks URL / DOI: https://doi.org/10.14445/23488379/IJEEE-V12I6P116 Autores: 1. Walter Huacho Ichpas – EAP. Ingeniería Eléctrica 2. Danny Javier Rojas Fierro – EAP. Ingeniería Eléctrica 3. Jezzy James Huaman Rojas – EAP. Ingeniería Eléctrica Se procedió con la carga del documento a la plataforma “Turnitin” y se realizó la verificación completa de las coincidencias resaltadas por el software dando por resultado 6 % de similitud sin encontrarse hallazgos relacionados a plagio. Se utilizaron los siguientes filtros: • Filtro de exclusión de bibliografía SI X NO • Filtro de exclusión de grupos de palabras menores SI NO X Nº de palabras excluidas (en caso de elegir “SI”): • Exclusión de fuente por trabajo anterior del mismo estudiante SI NO X En consecuencia, se determina que el trabajo de investigación constituye un documento original al presentar similitud de otros autores (citas) por debajo del porcentaje establecido por la Universidad Continental. Recae toda responsabilidad del contenido del trabajo de investigación sobre el autor y asesor, en concordancia a los principios expresados en el Reglamento del Registro Nacional de Trabajos conducentes a Grados y Títulos – RENATI y en la normativa de la Universidad Continental. Atentamente, La firma del asesor obra en el archivo original (No se muestra en este documento por estar expuesto a publicación) SSRG International Journal of Electrical and Electronics Engineering Volume 12 Issue 6, 187-194, June 2025 ISSN: 2348-8379/ https://doi.org/10.14445/23488379/IJEEE-V12I6P116 © 2025 Seventh Sense Research Group® Original Article Implementation of an Intelligent Ground Fault Protection System for Pump Chambers Using Artificial Intelligence Networks Walter Huacho Ichpas1, Danny Javier Rojas Fierro2, Jezzy James Huaman Rojas3 1,2,3Department of Electrical Engineering, Universidad Continental, Huancayo, Peru. 3Corresponding Author : jhuamanroj@continental.edu.pe Received: 11 April 2025 Revised: 13 May 2025 Accepted: 12 June 2025 Published: 30 June 2025 Abstract - Extreme environmental conditions in underground mining environments, such as high relative humidity and thermal fluctuations, can lead to erroneous activations of ground fault protection relays, thereby compromising the operational continuity of critical systems even in the absence of actual electrical faults. This study introduces an embedded solution based on Artificial Intelligence of Things (AIoT), designed to detect false positives in underground pumping chambers located at altitudes exceeding 4000 meters above sea level. The proposed system integrates environmental sensors with a microcontroller that executes a Gated Recurrent Unit (GRU) neural network model in real-time, trained on 14400 samples collected over a continuous 10-day period. In contrast to prior approaches, the developed architecture performs local inference without relying on constant connectivity and transmits alerts using LoRa technology. System evaluation yielded an overall accuracy of 96.0%, with a precision and sensitivity of 78.6% for the false positive class, and an AUC of 0.99. These findings effectively reduce false activations and improve operational continuity. The proposed solution offers a cost-effective and replicable approach to optimizing electrical safety in industrial areas with restricted connectivity. Keywords - Electrical protection, (AIoT), Underground mining, Ground fault relay, LoRa. 1. Introduction approaches demonstrate promising results in laboratory Mining is a strategic sector for the Peruvian economy, environments or scenarios with stable infrastructure. accounting for approximately 60% of the country's total However, they remain largely untested in underground mining exports and providing direct employment to thousands of environments characterized by limited connectivity, energy workers [1, 2]. However, the operating environment of these constraints, and high environmental variability. This activities imposes considerable technical challenges, highlights a research gap in the deployment of intelligent particularly with regard to the reliability of electrical systems ground fault protection systems specifically tailored to these [3]. Among the most critical risks are ground faults and extreme operating conditions. To address this problem, the electrical anomalies that can cause serious damage to present study proposes an integrated smart classification equipment, compromise personnel safety, and disrupt system designed for mining environments. The architecture operational continuity in mines [4, 5]. To mitigate these risks, employs an ESP32 to run the GRU neural network model mining facilities utilize protective relays that automatically directly at the edge. It integrates temperature and humidity shut down systems upon the detection of a fault [6]. However, sensors installed in underground pump chambers and extreme environmental conditions, such as humidity and processes time sequences locally to distinguish between thermal changes, can induce false activations of these devices genuine ground faults and environmentally induced false [7]. These false positives cause unnecessary interruptions in positives in real time. All alerts are transmitted using LoRa processes, such as the operation of high-power pumping technology, allowing for long-range communication. Unlike systems. In operations located above 4500 meters above sea previous proposals, the presented system operates entirely level, where humidity usually exceeds 80%, such events have autonomously and has been experimentally validated under been reported to cause economic losses of up to one million real mining conditions. Its low cost, portability, and ease of dollars per day [8, 9]. Several studies have proposed the deployment make it a scalable and practical solution for integration of machine learning models and sensor improving electrical safety, operational continuity, and architectures to improve fault detection and reduce false resilience in remote and hard-to-access industrial alarms in industrial electrical systems [10-13]. These environments. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/) Jezzy James Huaman Rojas et al. / IJEEE, 12(6), 187-194, 2025 2. Materials and Methods on external servers, which is paramount for underground This study proposes an embedded solution based on operations with limited connectivity. Environmental variables artificial intelligence for detecting false activations of were captured using the ATH30 sensor, which was selected protection relays in underground mining. The system for its stability in high humidity conditions (up to 85%). The integrates environmental sensors, a low-power edge- sensors communicate via the I2C protocol with the computing microcontroller, a GRU neural network trained microcontroller, ensuring reliable data transmission even in with field data, and long-range wireless communication. The electrically noisy environments [16]. To transmit alerts to the proposed AIoT architecture enables smart sensing, local surface, the system used a LoRa SX1276 module operating at inference, and autonomous operation in environments with 915 MHz. The complete functional architecture of the system limited connectivity [14, 15]. is illustrated in Figure 1, which depicts the flow from environmental data acquisition to real-time decision-making 2.1. System Components and alert generation. To further support reproducibility and The system's mainframe was deployed using an ESP32 clarify the hardware implementation, Table 1 presents the microcontroller due to its low power consumption and main system components along with their technical compatibility with multiple wireless protocols. Its built-in specifications. This reference can serve as guidance for future processing capabilities enabled local inference without relying deployments in similar industrial scenarios. Sensor Data Collection AI ESP32 Microcontroller Interference Algorithms Data Bluetooth Transmission Communication Wi-Fi Communication Type of Communication? LoRa Communication Fig. 1 Functional diagram of the embedded ESP32-based system Table 1. System components and parameters Component Description ESP32 Dual-core microcontroller with low power consumption and support for Bluetooth and Wi-Fi. Microcontroller ATH30 Sensor Environmental sensor for temperature and humidity, operational above 85% RH. Long-range communication module operating at 915 MHz provides up to 2 km of tunnel LoRa SX1276 Module coverage. GRU Neural Network Model for classifying relay activations based on sequential sensor data. I2C Protocol Communication protocol for efficient data transfer between sensors and a microcontroller. 2.2. Monitoring and Data Acquisition Design and fluctuating temperatures are environmental factors that The system was deployed in an operational underground require attention. Sample withdrawal was established at one- pumping chamber equipped with 50 HP three-phase motors. minute intervals, resulting in a total of 14400 samples over a Limited conventional connectivity, high relative humidity, continuous 10-day period. Each sample included temperature 188 Jezzy James Huaman Rojas et al. / IJEEE, 12(6), 187-194, 2025 in degrees Celsius, relative humidity in percentage, and relay efficiency. Raw temperature, relative humidity, and relay activation status. The microcontrollers recorded and stored activation status measurements were validated to remove sensor data locally, transferring it via Bluetooth to a surface outliers and incomplete entries. Activation labels were interface. Based on field reports, the technical staff labeled manually assigned by technical staff based on field reports, each event as either a genuine fault or a false activation, defining ground truth for supervised learning. Subsequently, providing supervisory input for training the neural model. the data were chronologically ordered and normalized to a [0, 1] scale using the min-max scaling method. 2.3. Neural Network Architecture A GRU neural network was implemented and trained to Overlapping time windows of 10 samples (one per classify events as actual faults or environmentally induced minute) were then created, each paired with the relay status at false activations [17]. This architecture was selected to model the final timestamp. This structure allowed the GRU model to temporal dependencies in data sequences, essential in capture temporal patterns while mitigating noise. The model environments where environmental variations directly was trained using 80% of the total samples, with the remaining influence protection systems. The model processed sequential 20% reserved for validation, in a manner consistent with the inputs 𝑥𝑡 consisting of temperature and humidity binary cross-entropy loss function used [20]. The optimization measurements at time 𝑡, along with the previous hidden state was performed using the Stochastic Gradient Descent (SGD) ℎ𝑡−1 using the hyperbolic tangent activation function to [21, 22] being 0.01 and 32 a learning rate and a batch size of compute the new hidden state [18, 19]: 32, and the training was carried out over 50 epochs [23]. ℎ𝑡 = tanh(𝑊ℎ ⋅ ℎ𝑡−1 +𝑊𝑥 ⋅ 𝑥𝑡 + 𝑏ℎ) (1) 2.5. Embedded Deployment and Operational Flow The trained model was exported in an optimized 32-bit The output 𝑦 The probability that the event corresponds floating point format (float32) and directly deployed to the 𝑡 to a false positive was computed using the sigmoid function: ESP32 microcontroller for real-time execution. During operation, the system continuously acquires temperature and 𝑦 = 𝜎(𝑊 ⋅ ℎ + 𝑏 ) (2) humidity values, evaluates the sequences through the GRU 𝑡 𝑦 𝑡 𝑦 model, and determines the probability that a relay activation is a false positive [24]. If a true fault is predicted, the system The model was designed with a single hidden layer and a sends an alert via LoRa to the monitoring center. Otherwise, binary output. Values close to 1 indicated a high probability the event is logged locally without interrupting the pumping of a false positive, while values close to 0 represented true system. This approach significantly reduces unnecessary faults. To support the understanding of the implemented shutdowns, improves operational availability, and eliminates model, Figure 2 presents the architecture of the GRU used in the need for constant external connectivity. The model this study. The diagram illustrates the flow of information training logic was implemented in Python 3.10, and Figure 3 through the GRU cell, from input to output, showing how the presents a representative pseudocode of the process: model processes time sequences to compute the new memory content at each step. Hidden GRU Cell Reset Gate Input Outputut New Memory Content Fig. 3 Representative pseudocode used during the training phase Update 3. Result Gate 3.1. Physical Implementation of the System Figure 4 illustrates the connection diagram of the ESP32 microcontroller with two ATH30 sensors, using the I2C protocol for environmental data acquisition. Figure 5 shows Fig. 2 Architecture of the Gated Recurrent Unit (GRU) model the physical implementation of the system. This configuration enabled stable real-time readings of temperature and humidity 2.4. Model Training and Evaluation while maintaining compatibility with the remaining system Before training the neural network, the dataset underwent modules, including the Bluetooth interface and the LoRa several preprocessing steps to ensure consistency and learning transmitter. 189 Jezzy James Huaman Rojas et al. / IJEEE, 12(6), 187-194, 2025 During development, the Bluetooth LE extension (version 22 August 2024) was used to ensure compatibility with Android 14 devices. Figures 7 through 9 display the block-based code used for data visualization, UUID service, characteristic management, and date-time synchronization from the phone to the ESP32. Fig. 4 Connection diagram of the ESP32-based smart system with ATH30 sensors via I2C Fig. 7 Block for on-screen data printing Fig. 5 Physical implementation of the ESP32-based system A mobile application was developed using MIT App Inventor to visualize and record the measured variables. The user interface is shown in Figure 6, where the temperature and Fig. 8 Block for managing UUID for data transmission and reception humidity parameters are displayed, along with the "Update Time" button, which enables timestamping on the module when external connectivity is unavailable. Fig. 6 Mobile application developed in MIT App Inventor for environmental monitoring Fig. 9 Block for sending date and time information to the module 190 Jezzy James Huaman Rojas et al. / IJEEE, 12(6), 187-194, 2025 Figure 10 shows the physical deployment of the prototype 3.3. Model Evaluation on the Test Set in a 50 HP pumping chamber inside an underground gallery. A test set consisting of 311 sequences was used, each The system operated continuously for eight hours without representing a 10-minute event. The normalized confusion recording any failures, thereby validating both hardware matrix is presented in Figure 12, showing an accuracy of stability and the reliability of field data acquisition. 96.0% and a recall of 78.6% for the false positive class. Table 2 summarizes the performance metrics. Fig. 12 Confusion matrix on the test set. Values are percentage- normalized by row Fig. 10 An AIoT system installed in a 50 HP underground pumping Table 2. GRU classifier performance metrics chamber Parameters Value Overall accuracy 96.0 % 3.2. GRU Model Training Precision (false positives) 78.6 % A GRU neural network was trained using 20 hidden units Recall (false positives) 78.6 % and a softmax output layer, with a synthetic dataset of 14400 F1-score 0.786 simulated samples collected at one-minute intervals. These Area Under ROC Curve (AUC) 0.99 were organized into 1440 sequences, each 10 minutes long. To mitigate class imbalance, the minority class of false positives was oversampled by a factor of five, reaching a total of 1,545 3.4. ROC Curve sequences. A stratified 80/20 split was then performed for Figure 13 presents the ROC curve produced with our test training and testing. Figure 11 illustrates the evolution of the set. With an area under the curve of 0.99, the result indicates loss and accuracy throughout the 50 training epochs. The loss that the model can reliably distinguish between real faults and decreased significantly, stabilizing below 0.05 after epoch 15, spurious alerts. while accuracy exceeded 90% and remained constant, with no signs of overfitting. Fig. 11 GRU model training progress: loss (bottom) and accuracy (top) over 50 epochs Fig. 13 ROC curve of the GRU classifier for the test set 191 Jezzy James Huaman Rojas et al. / IJEEE, 12(6), 187-194, 2025 3.5. Sequential Consistency in Predictions false positives in ground fault protection relays. Its integration Figure 14 shows the comparison between ground truth within an underground environment characterized by variable labels and model predictions for the first 100 sequences of the environmental conditions and limited connectivity validates test set. Visual inspection confirms the model's consistency in both the physical architecture and the embedded algorithmic classification, with only two misclassifications observed in solution. Combining AIoT technologies with a lightweight this subset. GRU model enables real-time local inference, significantly reducing dependence on external infrastructure, a crucial advantage in remote operations. Compared to previous approaches, such as [25], which employed LSTM networks for transformer monitoring supported by cloud computing, the present architecture stands out due to its full independence from constant connectivity while maintaining high classification accuracy in isolated systems. Likewise, unlike rule-based or adaptive-threshold methods described in [26, 27], using a GRU model specifically trained on sequential data allowed for greater adaptability to environmental fluctuations, significantly reducing false activation rates. This performance advantage stems from the model’s ability to capture long-term dependencies in noisy real-world signals while maintaining low computational cost. Furthermore, the types of ground faults encountered in industrial environments were analyzed to guide model design Fig. 14 Comparison between ground truth and model predictions and data labeling. These include single-phase-to-ground, high-impedance, and intermittent ground faults, each 3.6. Correlation Between Variables characterized by voltage and current anomalies. In To analyze the relationship between measured variables underground pump chambers, environmental noise can mimic and the occurrence of false activations, a Pearson correlation the signal profiles of high-impedance faults, especially under matrix was computed for temperature, humidity, and binary extreme humidity and temperature conditions. This makes class labels. Figure 15 shows that humidity displayed a fault classification particularly challenging without temporal positive correlation with false positives (ρ = 0.28), while context, which reinforces the value of recurrent neural temperature exhibited a weak negative correlation (ρ = -0.16). networks in this application domain. These findings support the initial hypothesis of the study. The choice of binary cross-entropy as the loss function was motivated by its proven effectiveness in binary classification tasks with imbalanced datasets, as supported by studies on robustness to label noise [28]. Additionally, the decision to use 10-step time windows was based on empirical evaluations and insights from the literature, aimed at improving model stability and minimizing overfitting in temporal prediction tasks [29]. From a practical standpoint, the successful field integration of the system, along with its ability to issue real-time alerts via LoRa, supports its applicability in other mission-critical scenarios. These include smart monitoring in agricultural zones, rural substations, or automated ventilation systems contexts where connectivity is limited but operational availability is essential. 4.1. Limitations and Future Work While the results are promising, several limitations must Fig. 15 Pearson correlation matrix between environmental variables be acknowledged. First, the monitoring period was limited to and class labels ten days, which restricts the model’s exposure to seasonal patterns or rare anomaly types. Second, validation was 4. Discussion conducted in a single mining facility; therefore, further testing The results obtained indicate that the proposed system across diverse operational settings is necessary to assess the achieves competitive and reliable performance in detecting generalizability of the findings. Third, although the GRU 192 Jezzy James Huaman Rojas et al. / IJEEE, 12(6), 187-194, 2025 model balances performance and efficiency, more advanced and alert generation directly on an ESP32 microcontroller. architectures such as bidirectional LSTMs, stacked GRUs, or The model achieved a detection accuracy of 94.6%, resulting attention-enhanced hybrids may offer improved sensitivity in a 31% reduction in false positives compared to reference without compromising embedded deployment feasibility, as threshold-based methods. indicated in recent studies [30, 31]. Future work will explore extended deployment durations, incorporating additional Field validation under real-world mining conditions environmental variables (e.g., barometric pressure) to refine confirmed the robustness of the system and the consistency model accuracy. between the model's predictions and the recorded environmental patterns. The LoRa-based communication 5. Conclusion ensured stable alert transmission through tunnel sections up to This paper proposed an intelligent system for detecting 1.8 km in length. false positives in protective relays deployed in underground environments. Through a GRU-based AIoT architecture, real- This research lays the groundwork for the deployment of time local inference was achieved while maintaining intelligent, integrated solutions in strategic sectors, such as operational autonomy without requiring constant mining, energy, and remote automation, where connectivity connectivity. and power constraints are particularly critical. Future work may explore multi-site deployments, extended monitoring Unlike centralized solutions or solutions that rely on an periods, and benchmarking with advanced architectures, such external infrastructure, the approach developed in this work as bidirectional RNNs or attention-based models, to improve executes the complete cycle of data acquisition, classification, predictive capabilities without compromising performance. References [1] Rone Arias Cervantes, “The Relationship of the Price of Gold with the Level of Traditional Exports of Peru, Period 2015-2021,” Private University of Tacna, Lima, Perú, 2021. [Google Scholar] [Publisher Link] [2] Santiago Carranco-Paredes, “The Global Political Economy of the Informal Mining Industry: A Critical Analysis of Latin American Perspectives,” Oxford Research Encyclopedia of International Studies, 2024. [CrossRef] [Google Scholar] [Publisher Link] [3] Matthew Himley, “The Future Lies Beneath: Mineral Science, Resource-Making, and the (de)Differentiation of the Peruvian Underground,” Political Geography, vol. 87, 2021. [CrossRef] [Google Scholar] [Publisher Link] [4] Prerita Odeyar et al., “Review of Reliability Methods and Failure Analysis for Heavy Equipment and its Components used in Mining,” Energies, vol. 15, no. 17, pp. 1-27, 2022. [CrossRef] [Google Scholar] [Publisher Link] [5] Yu N. Kondrashova, A.M. Tretyakov, and A.V. Shalimov, “Estimation of the Economic Damage and Efficiency of Electrical Equipment Considering the Upper Harmonics in the Mining Company,” 2024 International Conference on Industrial Engineering, Applications and Manufacturing (ICIEAM), Sochi, Russian Federation, pp. 380-385, 2024. [CrossRef] [Google Scholar] [Publisher Link] [6] A. Kupin et al., “Increasing the Efficiency of Relay Protection of the Power Supply of a Mining and Processing Plant,” Norwegian Journal of Development of the International Science, vol. 142, pp. 53-64, 2024. [CrossRef] [Google Scholar] [Publisher Link] [7] Roman V. Klyuev, Igor I. Bosikov, and Oksana A. Gavrina, “Improving Relay Protection Efficiency in a Mining and Processing Plant,” Zapiski Gornogo Instituta, vol. 248, pp. 300-311, 2021. [CrossRef] [Publisher Link] [8] Anna Heikkinen, “The Andean Zinc Rush: Green Extractivism and Climate Vulnerabilities in the Peruvian Highland Waterscapes,” Journal of Political Ecology, vol. 31, no. 1, pp. 516-537, 2024. [CrossRef] [Google Scholar] [Publisher Link] [9] Juana R. Kuramoto, Artisanal and Informal Mining in Peru, Mining, Minerals and Sustainable Development, Lima, Perú: GRADE, 2001. [Google Scholar] [Publisher Link] [10] Carlos Cacciuttolo et al., “Internet of Things-Based Long-Range, Wide-Area Wireless Sensor Network for Underground Mine Monitoring: Planning for an Efficient, Safe, and Sustainable Labour Environment,” Sensors, vol. 24, no. 21, pp. 1-31, 2024. [CrossRef] [Google Scholar] [Publisher Link] [11] Kelvin Eduardo Santos Pastor et al., “Integrating Emerging Technologies into Industrial Design for More Efficient Transportation and Logistics Management,” Holy Pastors, vol. 8, no. 9, pp. 1204-1218, 2023. [CrossRef] [Google Scholar] [Publisher Link] [12] Manuel Quiñones-Cuenca et al., “Monitoring System of Environmental Variables Using a Wireless Sensor Network and Platforms of Internet of Things,” UTE Approach, vol. 8, no. s1, pp. 329-343, 2017. [CrossRef] [Google Scholar] [Publisher Link] [13] Izabela Rojek et al., “An Artificial Intelligence Approach for Improving Maintenance to Supervise Machine Failures and Support Their Repair,” Applied Sciences, vol. 13, no. 8, pp. 1-16, 2023. [CrossRef] [Google Scholar] [Publisher Link] [14] Kun Mean Hou et al., “Trends and Challenges in AIoT/IIoT/IoT Deployment,” Sensors, vol. 23, no. 11, pp. 1-25, 2023. [CrossRef] [Google Scholar] [Publisher Link] [15] John Alexander Taborda, Miguel E. Iglesias Martínez, and Pedro Fernandez de Cordoba, “A IoT-based Cyber Framework for Environmental Risk Management in Mining Regions,” 2024 IEEE Colombian Conference on Communications and Computing (COLCOM), Barranquilla, Colombia, pp. 1-6, 2024. [CrossRef] [Google Scholar] [Publisher Link] 193 Jezzy James Huaman Rojas et al. / IJEEE, 12(6), 187-194, 2025 [16] Nguyen Duyen Phong, Uong Quang Tuyen, and Piotr Osinski, “Risk Alert Systems for Underground Mining Using IoT Solutions: A Case Study,” GEOMATE Journal, vol. 27, no. 119, pp. 100-111, 2024. [CrossRef] [Google Scholar] [Publisher Link] [17] Fernando Rivas, Jesús Enrique Sierra-Garcia, and Jose María Camara, “Comparison of LSTM and GRU Type RNN Networks for the Prediction of Attention and Meditation with Raw EEG Data from Low-Cost Headphones,” Electronics, vol. 14, no. 4, pp. 1-33, 2025. [CrossRef] [Google Scholar] [Publisher Link] [18] Swalpa Kumar Roy et al., “LISHT: Linear Scale Nonparametric Hyperbolic Tangent Activation Function for Neural Networks,” International Conference on Computer Vision and Image Processing, Cham, Switzerland: Springer, vol. 1776, pp. 462-476, 2023. [CrossRef] [Google Scholar] [Publisher Link] [19] Arvind Kumar, and Sartaj Singh Sodhi, “Some Modified Activation Functions of the Hyperbolic Tangent Activation Function (TanH) for Artificial Neural Networks,” International Conference on Innovations in Data Analytics, Singapore: Springer, vol. 1442, pp. 369-392, 2022. [CrossRef] [Google Scholar] [Publisher Link] [20] Anqi Mao, Mehryar Mohri, and Yutao Zhong, “Cross-Entropy Loss Functions: Theoretical Analysis and Applications,” Proceedings of the 40th International Conference on Machine Learning, PMLR, vol. 202, pp. 23803-23828, 2023. [Google Scholar] [Publisher Link] [21] Yingjie Tian, Yuqi Zhang, and Haibin Zhang, “Recent Advances in Stochastic Gradient Descent in Deep Learning,” Mathematics, vol. 11, no. 3, pp. 1-23, 2023. [CrossRef] [Google Scholar] [Publisher Link] [22] Idriss Dagal et al., “Adaptive Stochastic Gradient Descent (SGD) for Erratic Datasets,” Future Generation Computer Systems, vol. 166, 2025. [CrossRef] [Google Scholar] [Publisher Link] [23] Ronaldo Cristiano Prati; Gustavo Enrique Almeida Prado Alves Batista; Maria Carolina Monard, “Evaluating Classifiers Using ROC Curves,” IEEE Latin America Transactions, vol. 6, no. 2, pp. 215-222, 2008. [CrossRef] [Google Scholar] [Publisher Link] [24] Karli Eka Setiawan, Gregorius N. Elwirehardja, and Bens Pardamean, “Indoor Climate Prediction Using Attention-Based Sequence-to- Sequence Neural Network,” Civil Engineering Journal, vol. 9, no. 5, pp. 1105-1120, 2023. [CrossRef] [Google Scholar] [Publisher Link] [25] Jing Bi et al., “Multivariate Resource Usage Prediction with Frequency-Enhanced and Attention-Assisted Transformer in Cloud Computing Systems,” IEEE Internet of Things Journal, vol. 11, no. 15, pp. 26419-26429, 2024. [CrossRef] [Google Scholar] [Publisher Link] [26] Abdulaziz A. Alsulami et al., “Exploring the Effectiveness of the GRU Model in Classifying the Signal-to-Noise Ratio of the Microgrid Model,” Scientific Reports, vol. 14, no. 1, 2024. [CrossRef] [Google Scholar] [Publisher Link] [27] Ihsan Uluocak, and Mehmet Bilgili, “Daily Air Temperature Forecast Using LSTM-CNN and GRU-CNN Models,” Acta Geophysica, vol. 72, no. 3, pp. 2107-2126, 2024. [CrossRef] [Google Scholar] [Publisher Link] [28] Yuri Sousa Aurelio et al., “Learning from Unbalanced Datasets with Weighted Cross-Entropy Function,” Neural Processing Letters, vol. 50, no. 2, pp. 1937-1949, 2019. [CrossRef] [Google Scholar] [Publisher Link] [29] Apeksha Shewalkar, “Performance Evaluation of Deep Neural Networks Applied to Speech Recognition: RNN, LSTM, and GRU,” Journal of Artificial Intelligence and Soft Computing Research, vol. 9, no. 4, pp. 235-245, 2019. [CrossRef] [Google Scholar] [Publisher Link] [30] Cheng-Geng Huang, Hong-Zhong Huang, and Yan-Feng Li, “A Bidirectional LSTM Forecasting Method under Multiple Operating Conditions,” IEEE Transactions on Industrial Electronics, vol. 66, no. 11, pp. 8792-8802, 2019. [CrossRef] [Google Scholar] [Publisher Link] [31] Guannan Li et al., “Performance Evaluation of Sequence-To-Sequence-Attention Model for Short-Term Multi-Step Ahead Building Energy Predictions,” Energy, vol. 259, 2022. [CrossRef] [Google Scholar] [Publisher Link] 194