Publications

Reliable operation of grid-connected photovoltaic (PV) plants requires short-horizon forecasts that can track rapid generation changes caused by irradiance variation and plant operating conditions. This paper presents a 15-minute ahead DC power forecasting framework using real SCADA data from a grid-connected PV plant in Bangladesh. Conventional machine learning (ML) and deep learning (DL) baselines are evaluated using the same chronological data split and 15-minute-ahead forecasting target. To improve ultra-short-term prediction, a target-augmented residual BiGRU model is proposed. The model uses recent DC power information, including lag, ramp, and rolling statistical features, and learns a residual correction with respect to the last-known DC power value instead of estimating the future output directly. The proposed model achieves the best performance with an MAE of 57.15, RMSE of 146.59, WAPE of 17.76%, and R² of 0.93. Compared with the strongest DL baseline, BiGRU, it reduces MAE from 71.06 to 57.15 and WAPE from 22.08% to 17.76%. The results indicate that incorporating target-centric temporal features with residual learning improves ultra-short-term PV DC power forecasting while keeping the model architecture less complex than multi-stage hybrid approaches.

Precise photovoltaic (PV) output power forecasting is essential for ensuring the stability of renewable-rich grids. The inherent nonlinearity and weather-dependent variability of solar generation, compounded by distinct latent operating regimes, pose significant challenges to conventional models. This study introduces a novel regime-aware forecasting framework that leverages soft probabilistic memberships to capture transitional operating states using a Gaussian Mixture Model (GMM) to extract latent regimes. These probabilities are integrated as additional input features into a high-performance feature space, which also includes lag-window transformations and domain-specific feature engineering. The framework enhances four state-of-the-art baseline tree-based ensemble learners, Random Forest (RF), XGBoost, CatBoost, and LightGBM and validates using high-resolution real operational data from a 7.6 MWp PV real-world plant in Bangladesh. The results demonstrate that embedding soft regime probabilities consistently outperforms baseline models across the root mean square error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination. By effectively decoding the complexities of transitional operating states, this regime-aware framework establishes a robust and scalable standard for short-term PV output power forecasting in modern energy management systems.

A smart grid is the combination of networking equipment to capture parameters related to the cyber domain, as well as Advanced Metering Infrastructure (AMI) to obtain necessary consumption data and physical domain parameters. However, today’s cyber-physical grid systems are exposed to numerous threats, which have increased in frequency due to the integration of low-cost networking equipment and Internet of Things (IoT) devices. If we know that any attack has occurred in the grid, the operators can disconnect the utilities directly, and knowing the type of cyberattack will help the operators investigate the reason and solve it immediately. In this paper, we have done an analysis of machine learning classifiers to detect an attack case in the smart grid, as well as identify the type of cyberattack using the Smart Grid Cyber-Physical Attack dataset to help the operators disconnect the utilities and to investigate. It has been found that Random Forest achieves the highest accuracy of 98% while detecting the attack scenario, whereas LightGBM achieves an accuracy of 99% while identifying the type of attack. This result has been validated using 5-fold cross-validation, which proves the robustness of Random Forest and LightGBM in performing the task.

M. T. I. Aronno, M. S. H. Molla, M. T. R. Sahed, and M. N. M. Zishan, “Machine learning and explainable AI for robust smart grid security: A new dataset study,” in Proc. 2026 IEEE 2nd Int. Conf. Quantum Photonics, Artificial Intelligence, and Networking (QPAIN), Chittagong, Bangladesh, 2026, pp. 1–6, doi: 10.1109/QPAIN69676.2026.11545639.

@inproceedings{aronno2026machine,
  title        = {Machine learning and explainable AI for robust smart grid security: A new dataset study},
  author       = {Aronno, Md Tanjil Islam and Molla, Md Sabbir Hossain and Sahed, Md Tanvir Rahman and Zishan, Md Nasif Mahmud},
  booktitle    = {2026 IEEE 2nd International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN)},
  pages        = {1--6},
  year         = {2026},
  organization = {IEEE},
  doi          = {10.1109/QPAIN69676.2026.11545639}
}

This paper introduces a machine learning framework for binary dysarthria classification using acoustic and prosodic cues collected from speech recordings of dysarthric and non-dysarthric patients. Deep CNN14 embeddings are used to learn the audio signals' spectral and temporal patterns and to improve the classification accuracy, other prosodic descriptors such as pitch variability, energy dynamics, and spectral flatness are added. We train and combine several machine learning classifiers, such as KNN, Random Forest, XGBoost and CatBoost, and a soft-voting ensemble model. Adding prosody to the experiment made the overall accuracy go up from 97% to 98.1%, with big drops in both false positives and false negatives. SHAP-based explainable AI analysis shows that pitch-related and energy based parameters are two of the most important factors in detecting dysarthria. The proposed method presents a balanced framework that integrates deep acoustic representations with comprehensible features to attain reliable and clear dysarthric speech assessment.

M. T. I. Aronno, M. S. H. Molla, M. S. Rahman, and M. T. R. Sahed, “An explainable machine learning framework for dysarthria detection using deep acoustic embeddings and prosodic cues,” in Proc. 2026 IEEE 2nd Int. Conf. Quantum Photonics, Artificial Intelligence, and Networking (QPAIN), Chittagong, Bangladesh, 2026, pp. 1–6, doi: 10.1109/QPAIN69676.2026.11546156.

@inproceedings{aronno2026explainable,
  title        = {An Explainable Machine Learning Framework for Dysarthria Detection Using Deep Acoustic Embeddings and Prosodic Cues},
  author       = {Aronno, Md Tanjil Islam and Molla, Md Sabbir Hossain and Rahman, Md Shovon and Sahed, Md Tanvir Rahman},
  booktitle    = {2026 IEEE 2nd International Conference on Quantum Photonics, Artificial Intelligence \& Networking (QPAIN)},
  pages        = {1--6},
  year         = {2026},
  organization = {IEEE},
  doi          = {10.1109/QPAIN69676.2026.11546156}
}

Power theft is a serious concern for electricity providers all over the world. It makes the power supply less reliable, raises the load consumption, and costs more money. As the global energy sector moves towards using smart grid (SG) technologies, the use of advanced data analysis tools has created new ways to deal with these problems, but electricity theft detection (ETD) is still a major issue. Traditional methods often do not work well because of issues like imbalanced data distributions, overfitting, high-dimensional data, and too much dependency on individual detection techniques. This study's goal is to address these problems by developing a more effective and robust ETD model using modern prediction algorithms. This study addresses one of the voting ensemble models that employs some of the best-known machine learning boosting classifiers, like extreme gradient boosting (XGBoost), light gradient boosting (LightGBM), and categorical boosting (CatBoost) to enhance the prediction accuracy. We use a gated recurrent unit (GRU) to extract features, which makes it easier to work with time series data resulting in the improvement of model performance. There are three major tasks performed: synthetic minority oversampling technique with tomek link (SMOTETomek) was used for data pre-processing, a recurrent neural network (RNN) was used for the temporal feature extraction, and SecureGrid was proposed as a new voting-based ensemble model. The proposed model shows that it can achieve a test accuracy of 93%, an F1 score of 78%, and 97% ROC-AUC in the data set of the State Grid Corporation of China (SGCC). The proposed model outperforms traditional boosting models that have been used on similar datasets. We found that our proposed SecureGrid model has the ability to significantly enhance ETD in SGs. It offers a robust and near to accurate solution which is much better than current approaches.

M. S. H. Molla, M. T. I. Aronno, M. T. R. Sahed, M. N. M. Zishan, and A. K. M. N. Islam, “SecureGrid: A robust voting ensemble model for non-technical loss (NTL) detection in smart meter data,” in Proc. 2025 IEEE 2nd Int. Conf. Computing, Applications and Systems (COMPAS), Kushtia, Bangladesh, 2025, pp. 1–6, doi: 10.1109/COMPAS67506.2025.11381615.

@inproceedings{molla2025securegrid,
  title        = {SecureGrid: A robust voting ensemble model for non-technical loss (NTL) detection in smart meter data},
  author       = {Molla, Md Sabbir Hossain and Aronno, Md Tanjil Islam and Sahed, Md Tanvir Rahman and Zishan, Md Nasif Mahmud and Islam, AKM Nazrul},
  booktitle    = {2025 IEEE 2nd International Conference on Computing, Applications and Systems (COMPAS)},
  pages        = {1--6},
  year         = {2025},
  organization = {IEEE},
  doi          = {10.1109/COMPAS67506.2025.11381615}
}

Detecting and classifying violence is crucial for public safety and addressing societal violence. DL models have greatly improved the automation of violence detection systems by effectively capturing intricate visual patterns. However, the quality and diversity of the training data greatly impact the effectiveness of these models. Existing datasets may be biased towards specific situations, limiting their practical use. To address this limitation, we introduce the Movie Clip (MC) dataset to enhance the generalisability of Automated Violence Detection and Classification (AVDC) systems. The MC dataset encompasses a broad spectrum of near-real-world violent actions, incorporating diverse demographics, environmental circumstances, and cultural elements extracted from movies. Consequently, it accurately reflects the complexity and diversity of real-world violent scenarios. The potential of the new dataset is investigated against its existing counterparts, like Hockey Fight (HF) and AIRTLab datasets. These datasets are used to train the ConvLSTM models based on the VGG16 and VGG19 architectures. The proposed dataset significantly improves AVDC model generalisation, outperforming the generalisability of current datasets, thereby advancing violence detection and facilitating the development of more robust and efficient AVDC systems.

M. Haque et al., “Generalising violence detection with a new near-real-world violence dataset,” in Proc. 2024 3rd Int. Conf. Advancement in Electrical and Electronic Engineering (ICAEEE), Gazipur, Bangladesh, 2024, pp. 1–6, doi: 10.1109/ICAEEE62219.2024.10561682.

@inproceedings{haque2024generalising,
  title        = {Generalising violence detection with a new near-real-world violence dataset},
  author       = {Haque, Mahmudul and Nyeem, Hussain and Ovi, Tareque Bashar and Nahid, Al and Molla, Md Sabbir Hossain and Tuhin, Md Tanjim Mahmud and Shahab, Fardin and Alam, Ayat Subah and Alam, Saadia Binte},
  booktitle    = {2024 3rd International Conference on Advancement in Electrical and Electronic Engineering (ICAEEE)},
  pages        = {1--6},
  year         = {2024},
  organization = {IEEE},
  doi          = {10.1109/ICAEEE62219.2024.10561682}
}

Electricity theft poses a critical challenge among non-technical losses for electricity providers, resulting in severe consequences such as compromised supply quality, increased generation load, inflated bills for legitimate consumers, and broader economic repercussions. Smart grid technology offers a promising avenue to mitigate such losses through advanced data analytics. With the abundance of data generated by smart grid infrastructure, including individual power consumption patterns, machine learning and deep learning methodologies emerge as effective tools to accurately detect instances of electricity theft. However, existing schemes provide suboptimal performance in electricity theft detection (ETD) due to challenges such as imbalanced data distributions, overfitting issues, handling high-dimensional data, and reliance on individual schemes. In response to these limitations, this paper proposes a voting ensemble model based on machine learning boosting classifiers. In this model, we incorporate Convolutional Neural Networks (CNN), a widely recognized technique for feature extraction. The framework encompasses three key components, such as data preprocessing, data balancing, and classification, each addressing specific challenges in ETD. Notably, the proposed model leverages the strengths of extreme gradient boosting (XGBoost), categorical boosting (CatBoost), and light gradient boosting (LightGBM), enhancing its overall robustness and predictive accuracy. Rigorous validation using diverse metrics underscores the efficacy of the model, as evidenced by extensive simulations. Results indicate superior performance compared to individual models, with the proposed model achieving an impressive test accuracy of 99%.

M. S. H. Molla and M. N. M. Zishan, “Enhanced detection of electricity theft in smart grids using a machine learning boosting classifiers based voting ensemble model,” B.Sc. thesis, Dept. of EECE, Military Institute of Science and Technology (MIST), Dhaka, Bangladesh, 2024, doi: 10.13140/RG.2.2.23095.89767.

@thesis{molla2024enhanced,
  author  = {Molla, Md. Sabbir Hossain and Zishan, Md. Nasif Mahmud},
  title   = {Enhanced Detection of Electricity Theft in Smart Grids Using a Machine Learning Boosting Classifiers Based Voting Ensemble Model},
  school  = {Military Institute of Science and Technology (MIST)},
  address = {Dhaka, Bangladesh},
  type    = {B.Sc. thesis},
  month   = mar,
  year    = {2024},
  doi     = {10.13140/RG.2.2.23095.89767}
}