Research

Description:
Heavy solar photovoltaic (PV) penetration in distribution networks can cause overvoltage during low-demand, high-irradiance periods and requires active power curtailment for voltage control. Existing curtailment-based voltage control strategies often impose disproportionate curtailment on electrically distant prosumers. Fairness-aware curtailment methods can mitigate this unfairness, but typically imposes fairness over short horizons and increase curtailed energy whereas some rely on topology reconfiguration, requiring additional switching and protection infrastructure. This paper proposes a memory-weighted online rolling linear programming (LP) framework for cumulative fair PV curtailment over a monthly billing horizon. The proposed framework operates at 15-minute resolution and tracks each prosumer’s cumulative curtailed energy and cumulative available PV energy to calculate a normalized curtailment burden. A time-varying fairness penalty shifts future curtailment toward lower-burdened prosumers while satisfying system constraints. The framework is validated on a modified CIGRE LV benchmark network and a modified IEEE-37 bus feeder under realistic varying-days scenario and synthetic similar days irradiance scenario for testing stress. Compared with a day-ahead LP baseline, the proposed method attains perfect month-end fairness with a lower price of fairness in both scenarios. On the CIGRE LV network, the price of fairness decreases from 7.53 to 4.51 percentage points in the varying-days scenario and from 12.47 to 7.83 percentage points in the similar-days scenario. On the IEEE-37 bus feeder, the corresponding reductions are from 20.73 to 14.37 and from 21.87 to 15.66 percentage points. Monte-Carlo based sensitivity studies demonstrate robustness of the framework against PV availability estimation errors. A priority weighted extension further demonstrates the framework’s ability for preferential treatment of a commercial prosumer while preserving fairness within the residential cohort. .  

First Author
Under Review: Applied Energy (Elsevier) 

Supervisor: Dr. Md. Ziaur Rahman Khan, Professor, Dept. of EEE, BUET

Based on Postgraduate Thesis 

Description:
High photovoltaic (PV) penetration in low-voltage (LV) distribution networks can cause overvoltage during high generation, low-demand periods, requiring active power curtail ments. Fairness-aware curtailment methods aim to distribute the curtailment burden equitably among prosumers. However, practical networks may include heterogeneous prosumer classes where a commercial prosumer receives preferential treatment under contractual obligation. This paper investigates how the feeder location of a priority PV prosumer affects the curtailment burden on residential users. A modified CIGRE LV network is used where along with the 16 residential PV users, a commercial PV unit of 63.25 kWp is connected at each of the PV-capable buses. Solving a day-ahead linear programming (LP) formulation with priority weighted curtailment ratio fairness under high PV, low-load scenario under six commercial priority weights show that for most buses stronger commercial priority reduces commercial curtailment while increasing residential curtailment. Results also show that, as the commercial prosumer is placed at a more remote location, the burden of priority-weighted fairness on the residential prosumers increases, exhibiting crucial topology dependence, exceptions being buses 16 and 18 due to their lateral placement and local voltage sensitivity. The results indicate that priority-based curtailment is strongly topology-dependent and must be evaluated using network-aware voltage constraints.  

First Author
Accepted in: 2026 IEEE INTERNATIONAL CONFERENCE ON POWER AND ENERGY (PECON 2026)

Supervisor: Dr. Md. Ziaur Rahman Khan, Professor, Dept. of EEE, BUET

Written as an extension of Postgraduate Thesis

Description:
The rapid growth of artificial intelligence workloads creates large and rapidly varying power fluctuations on distribution feeders, creating voltage violations which cannot be suppressed reliably by battery energy storage system alone, especially when storage headroom is limited. This paper proposes a cascaded primal-dual online feedback optimization controller which jointly dispatches a grid connected battery storage and GPU batch-size schedules for real-time voltage regulation. A cascade priority function, created from logistic state-of-charge indicators, continuously transfers regulatory authority from the battery to the compute side as storage approaches its operational limits. The proposed controller, therefore, replaces abrupt switches with a smooth transition between actuators. On the IEEE 13 bus distribution test feeder under an 1800 second heterogeneous workload, the proposed controller reduces mean voltage deviation by 76.3% against the uncontrolled baseline and by 23.6% against battery-only online feedback optimization with identical hardware. It also cuts the mean inter-token latency by 38.9% relative to the uncontrolled baseline, while retaining 77% of aggregate throughput. Monte Carlo uncertainty trials confirm voltage deviation variance to be 1.4 times tighter than battery-only control.  

First and Only Author
Under Review: IEEE Transactions on Smart Grid

Description:
To meet the rapidly increasing load demand, the Northern Grid of Bangladesh suffers from frequent under-voltage problems owing to weak infrastructure and lack of generation capacity. This study proposes a real-life data-driven approach using deep-learning for improving the voltage profile of the Northern Grid of Bangladesh. The real-world Northern Grid of Bangladesh has been simulated in DIgSILENT PowerFactory using actual load, generation and transmission data. Actual voltage data of 132 kV, 230 kV and 400 kV buses in the Northern Grid has been collected and analyzed using Long Short-Term Memory (LSTM) Network to predict voltages and rank the buses from weakest to strongest based on the forecasted voltages. The simulations demonstrate that placement of three SVCs at Rangpur, Panchagarh and Joypurhat 132 kV buses completely fix the undervoltage problem. The same SVC placements have also been applied to a different date’s simulation to validate the findings and the solution worked successfully for that date as well. Integrating deep learning for identifying weak bus, coupled with FACTS device placement, offers a scalable and intelligent framework for grid reinforcement in a growing power system like that of Bangladesh.

First Author
Under Review: IEEE International Conference on Power Electronics, Drives and Energy Systems (PEDES 2026) 

Description:
This project introduces a deep learning framework designed for month-ahead electricity price forecasting, a critical but often overlooked area in deregulated power systems like ISO New England (ISO-NE). By combining Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks, the proposed model effectively captures complex trends, residuals, and seasonal patterns in historical market data to predict the final locational marginal price (LMP). The simulation results indicate high accuracy, with forecasted prices closely matching real-world datasets, offering market participants a reliable tool for generation scheduling, bidding strategies, and risk management.

Supervised by: Dr. Sheikh Anowarul Fattah, Professor, Dept. of EEE, BUET

Part of Machine Learning and Pattern Recognition Course of M.Sc. 

Description:
This project details the development of an autonomous firefighting and rescue robot designed to serve as a first line of defense in high-rise buildings and industrial settings in Bangladesh. Utilizing a combination of IR and MQ2 gas sensors, the robot detects fire and smoke, moves toward the source, and mitigates the hazard using an onboard water tank and pump. A distinctive feature of this prototype is its Bangla voice recognition capability, which allows it to identify a victim's call for help and send a notification with location details to emergency services via a SIM 900a GSM module. Built with an Arduino UNO and powered by rechargeable lithium-ion batteries—designed for solar charging to align with Sustainable Development Goals (SDG)—the project offers a scalable, budget-friendly solution for improving fire response times and victim rescue efforts 

Supervised by: Dr. Celia Shehnaz, Professor, Dept. of EEE, BUET

Based on Control Systems Laboratory Project in Undergraduate 

Description:
This project details the development of a Bangla voice-based attendance system designed to replace manual, proxy-prone methods at BUET with a seamless biometric solution. Utilizing MATLAB, the system employs a subspace K-Nearest Neighbors (KNN) classifier and Mahalanobis distance to identify students based on unique vocal characteristics extracted via Mel Frequency Cepstral Coefficients (MFCC) and pitch. The researchers compiled a robust database of 4,740 voice samples from 237 individuals, implementing a 5th-order low-pass filter to mitigate classroom noise. Experimental results demonstrated high effectiveness, achieving a final individual detection accuracy of 95.62% for IDs and 95.26% for names, significantly outperforming previous models. Future implementation plans involve a central database and a dedicated mobile application, with potential upgrades to deep learning frameworks for enhanced precision. 

Supervised by: Dr. Celia Shehnaz, Professor, Dept. of EEE, BUET

Based on Digital Signal Processing Laboratory Project in Undergraduate