Papers by Keyword: Machine Learning

Abstract: The foundation of power system reliability is voltage stability, which is required to promise a secure and stable supply of electricity. Insufficient generating capacity, a timeworn transmission facility, inadequate reactive power compensation, and the increasing integration of renewable energy sources are the main foundations of the existing voltage instability of Nigeria's power grid, specifically in the northern regions. Through evaluating contemporary disturbances such as smart grid technology and intelligent machine learning (ML) drives up for real-time voltage security evaluation and predictive analytics, this study provides a technically enhanced examination of these effects. Contrasting machine learning models, such as deep learning (DL), supervised learning, unsupervised learning, and reinforcement learning (RL), are explored for their capabilities in time-varying voltage prediction, robotic grid control, and anomaly detection. Also, it highlights the transformative impact of machine learning in improving voltage stability management and outlines strategic recommendations relating to guiding principle reforms and infrastructure transformation. The article intends to provide a forward-looking structure for deploying adaptive Machine stakeholder engagement e-learning-powered solutions to achieve resilient and voltage security in the Nigerian power system that structures long-term economic sustainability.

Abstract: The dynamic nature of cyber threats and the growing intricacy of network traffic frequently provide challenges for traditional approaches to network security and traffic regulation. In this study, we suggest using deep learning systems as a potentially effective way to improve traffic control and network security. Three deep learning architectures the Long Short-Term Memory Convolutional Neural Network (LSTM-CNN), the Convolutional Neural Network (CNN), and the Deep Neural Network (DNN) are thoroughly analysed here. The capacity of these systems tо categorise network traffic and identify unusual activity is assessed. Our tests carried out оn actual network datasets show how well these deep learning models perform іn precisely categorising network traffic and spotting any security risks. Additionally, we look into how well these models work in scenarios involving the source domain and the target domain. While the target domain assessment measures the models' ability tо generalise tо new data, the source domain evaluation evaluates the models' performance оn the training set. Our findings show that, in both domains, the LSTM-CNN design outperforms the CNN and DNN structures, achieving maximum accuracy and resilience. Our research indicates that deep learning systems in particular, LSTM-CNN architecture have a lot оf potential to enhance traffic control and network security. Network managers and cybersecurity experts may strengthen their networks' defences against online attacks and guarantee the uninterrupted operation оf vital network infrastructure by utilising deep learning.

Abstract: Machine Learning (ML) approach seeks to open new frontiers in the search for novel thermoelectric materials that convert heat waste into useful electrical energy. Five regression-type ML algorithms Linear Regression, Random Forest, XGBoost, Bagging Regressor, and Gradient Boosting Regressor were employed in this study to forecast the thermoelectric figure of merit (ZT) of doped chalcogenide compounds. Gradient Boosting Regressor achieved the best baseline performance (R² = 94.5%, MAE = 0.073, RMSE = 0.128), further improved with hyperparameter tuning to R² = 95.8%, MAE = 0.065, and RMSE = 0.112. Compared to the baseline, tuning reduced RMSE by 12.6% and MAE by 10.8%. The optimized model reliably reproduced experimental ZT trends in doped Bi₂Te₂Se and Ag₂Te, validating its predictive capacity. Our findings show that hyperparameter tuning is greatly recommended for high-fidelity predictions in thermoelectrics.

Abstract: CO₂ conversion to methanol via thermocatalytic hydrogenation is one of the viable alternatives to address climate change problem while producing a valuable industrial product. However, this comes with a challenge, i.e., predicting the performance of catalytic systems. In this work, we present a data-driven study to predict the performance of Cu-based catalyst based on a compiled dataset consisting of 15 features obtained from experiment data. Furthermore, we implement feature selection techniques such as univariate, RFE, and XGBoost to investigate how the performance of the prediction model changes with varied number of features. The results show that features selected by RFE method yields the best performance with 7 number of features, capable of even outperforms the baseline model in terms of accuracy and feasibilty. This suggests that feature selection technique is relevant in terms of constructing a machine learning model for predicting methanol production via CO₂ thermocatalytic hydrogenation.

Abstract: This research integrates experimental testing and Machine Learning (ML) techniques to predict the weld quality of Tungsten Inert Gas (TIG) and Shielded Metal Arc Welding (SMAW). A balanced dataset comprising weld parameters and mechanical test results including tensile strength, impact energy, and bend test outcomes was compiled for mild steel and stainless steel specimens with thickness ranging from 6mm to 10mm. Experimental results revealed that TIG welding produced superior tensile strength (up to 572 MPa) and impact energy (up to 58J) compared to SMAW. A Random Forest classifier achieved 100% accuracy in classifying weld quality as Good or Defect, while linear Regression produced tensile strength with an R2 of 0.68, Mean Absolute Error (MAE) of 17.5 MPa, and Root Mean Squared Error (RMSE) of 20.27 MPa. These results confirm the viability of ML techniques as non-destructive tools for weld quality prediction and mechanical property estimation. The framework developed in this research contributes to intelligent welding process control and supports the transition toward efficient, data driven manufacturing.

Abstract: This study explores the required Level of Detail (LOD) in 3D urban models to elicit observation responses similar to those in real spaces. Through experiments involving 30 participants, both real-world and 3D-modeled streetscapes were evaluated using psychological surveys and webcam-based eye-tracking. Results showed that higher model precision generally produced responses closer to those from real environments. However, inconsistencies appeared at higher LODs, likely due to fatigue or equipment limitations. Open horizontal spaces attracted greater attention, suggesting the need for detailed modeling in such zones. While a clear threshold of sufficient detail was not found, the findings highlight the potential of 3D models as substitutes for field observation and the necessity for standardized LOD criteria in urban simulations.

Abstract: The sintering behavior of spark plasma sintering was analyzed by extracting features from process data obtained during the fabrication of aluminum matrix composites and machine learning using the obtained features were performed to predict relative density of composites. Seventy-five samples were sintered with different types of reinforcement, and different temperature and pressure conditions. Regression methods include linear regression such as Ridge, Lasso and Elastic Net, and nonlinear regression such as random forest, gradient boosting and XGBoost were tested. XGBoost had the highest prediction accuracy and the trained model was used for Shapley additive explanations value analysis and inverse analysis.

Abstract: In the present work artificial neural networks (ANN) models have been implemented and trained as surrogate models to replicate two physics-based microstructure models for Al-alloys, i.e. the ALFLOW model, which predicts the sub-structure evolution and associated flow stress during plastic deformation and the softening model ALSOFT, which predicts the softening behavior after hot/cold deformation, in view of the combined effect of recovery and recrystallization. Input for both ANN models was limited to variables such as strain, strain rate, time, temperature and solute concentration, and the flow stress as the output. Accuracy and efficiency were tested for different ANN architectures. It is demonstrated that fully connected feed-forward neural network architectures with ∼3 hidden layers are suitable as surrogate models for both ALFLOW and ALSOFT, with a potential speed-up of ∼100x for ALFLOW and ∼10x for ALSOFT.

Abstract: This work numerically studies the water-in-oil (W/O) droplet formation inside a flow focusing on the micro junction formed by rectangular channels with dimensions of 390 × 190 μm² using OpenFoam. An automatic algorithm was developed to assess the effect of key parameters such as water viscosity, restriction ratio and water mass flow rate ratio on the droplet size. A total of 96 simulations, with different parameter combinations, were conducted to train a Machine Learning (ML) algorithm capable of predicting the droplet dimensions based on the key parameters mentioned. The ML algorithm was also compared to a Newtonian-based optimization method, where the geometry is iteratively adjusted to produce droplets of a fixed size. Results reveal that both methods appear valid in the prediction of droplet dimensions.

Abstract: Accurate prediction of tunnel water inflow (WI) is critical for preventing construction hazards and supporting engineering decision-making. This study proposes an enhanced stacking machine learning framework to achieve reliable WI prediction. SMOGN technology was used to address the issue of imbalanced data distribution. On this basis, the Bayesian optimization tool Optuna was utilized for automated hyperparameter tuning of various base learners, including elastic net (Enet), multilayer perceptron (MLP), support vector regression (SVR), and extreme gradient boosting (XGBoost). Base learner outputs were combined with original features to form augmented inputs, with ridge regression as the meta-learner. Gaussian process regression (GPR) modeled residuals for uncertainty quantification, while the Sobol method assessed feature importance. Results show that feature augmentation and residual modeling improve prediction performance. The proposed stacking model outperforms individual base models and achieves state-of-the-art results. From an engineering perspective, the framework can be used to forecast potential high-inflow zones, guide advance grouting or drainage measures.