Neelam Mughees
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Neelam Mughees has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
Abdullah Mughees
4 shared publicationsAnam Mughees
4 shared publications- Krzysztof Ejsmont
Warsaw University of Technology
3 shared publications - Abdullah Mughees
Electrical Engineering Department, Information Technology University (ITU) of the Punjab, Lahore, Pakistan
2 shared publications - Anam Mughees
Department of Electrical Engineering, University of Engineering and Technology (UET), Lahore, Pakistan
2 shared publications Haojie Yu
Jixi Technology (Tianjin) Co. Ltd. Tianjin China
1 shared publication
Latest Publications
- Open Access
- 17 March 2026
- Published by Elsevier in SSRN Electronic Journal
Multi-Agent Reinforcement Learning and Deep Sequence-to-Sequence Bi-LSTM Neural Networks based Nonlinear Control of an Unmanned Aerial Vehicle
This research proposes a novel AI-based nonlinear control strategy for quadcopter trajectory tracking using the conditioned adaptive barrier function integral terminal sliding mode control (CABFIT-SMC) framework. By integrating deep learning and reinforcement learning, the model enhances control performance. A comparative analysis with standard SMC highlights the improved efficiency of CABFIT-SMC in handling nonlinearities, including gyroscopic moments and aerodynamic effects, within a MATLAB discrete-time environment. A Lyapunov stability analysis verifies the system’s asymptotic stability. To mitigate the computational complexity of the control law, a sequence-to-sequence Bidirectional Long Short-Term Memory (Bi-LSTM) neural network is introduced. Trained on CABFIT-SMC data, Bi-LSTM significantly reduces computational expenses while maintaining control accuracy. BI-LSTM achieves a simulation time of 0.0325 s and a computational cost of 22.1752 s, substantially outperforming RLCABFIT-SMC and the original CABFIT-SMC. These results highlight Bi-LSTM’s potential as a cost-effective solution for real-time aerial vehicle control. While this study is based on simulations, future work will focus on real-world testing to validate the method’s practical applicability in autonomous aerial systems.
- 17 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMulti-Agent Reinforcement Learning and Deep Sequence-to-Sequence Bi-LSTM Neural Networks based Nonlinear Control of an Unmanned Aerial Vehicle
This research proposes a novel AI-based nonlinear control strategy for quadcopter trajectory tracking using the conditioned adaptive barrier function integral terminal sliding mode control (CABFIT-SMC) framework. By integrating deep learning and reinforcement learning, the model enhances control performance. A comparative analysis with standard SMC highlights the improved efficiency of CABFIT-SMC in handling nonlinearities, including gyroscopic moments and aerodynamic effects, within a MATLAB discrete-time environment. A Lyapunov stability analysis verifies the system’s asymptotic stability. To mitigate the computational complexity of the control law, a sequence-to-sequence Bidirectional Long Short-Term Memory (Bi-LSTM) neural network is introduced. Trained on CABFIT-SMC data, Bi-LSTM significantly reduces computational expenses while maintaining control accuracy. BI-LSTM achieves a simulation time of 0.0325 s and a computational cost of 22.1752 s, substantially outperforming RLCABFIT-SMC and the original CABFIT-SMC. These results highlight Bi-LSTM’s potential as a cost-effective solution for real-time aerial vehicle control. While this study is based on simulations, future work will focus on real-world testing to validate the method’s practical applicability in autonomous aerial systems.
- 12 February 2026
- Published by Wiley in Computer Applications in Engineering Education
- Vol. 34 (2)
Adaptive Deep Reinforcement Learning for Optimizing Teacher Professional Development Path
Ensuring equitable access to cybersecurity expertise has become increasingly critical, considering the growing complexity of digital threats. As educators are tasked with delivering instruction in areas such as computer fraud prevention and network security, there is a pressing need for adaptive, data‐informed professional development systems that can support individualized learning paths. To address this challenge, this study proposes an adaptive deep reinforcement learning framework for optimizing personalized teacher development path planning in cybersecurity education. Two enhanced models are introduced: an improved Deep Q‐Network (DQN) that integrates a multi‐layer perceptron, a cubic dynamic reward function, and an adaptive exploration strategy; and a PER‐D3QN model that combines dueling double deep Q‐learning (D3QN) and prioritized experience replay (PER) to mitigate Q ‐value overestimation and accelerate convergence. Experimental evaluation using real‐world teacher data demonstrates that the improved DQN achieved average performance scores up to 0.36, compared to 0.054–0.068 for the traditional DQN. Moreover, the PER‐D3QN model outperformed the ERDQN baseline, attaining an average reward of 4.738 versus 2.021, and an average score of 2.799 after 6000 training rounds, compared to 1.946 for ERDQN, indicating that network update speed has also been significantly improved. This research not only helps to enhance teachers' professional knowledge and technical application ability in the field of network security, but also provides scientific methodological support for educational institutions to ensure that they are aligned with changing security threats. Furthermore, this study emphasizes the importance of interdisciplinary cooperation and encourages experts from computer science, education, and psychology to work.
- 12 February 2026
- Published by Wiley in Computer Applications in Engineering Education
- Vol. 34 (2)
Adaptive Deep Reinforcement Learning for Optimizing Teacher Professional Development Path
Ensuring equitable access to cybersecurity expertise has become increasingly critical, considering the growing complexity of digital threats. As educators are tasked with delivering instruction in areas such as computer fraud prevention and network security, there is a pressing need for adaptive, data‐informed professional development systems that can support individualized learning paths. To address this challenge, this study proposes an adaptive deep reinforcement learning framework for optimizing personalized teacher development path planning in cybersecurity education. Two enhanced models are introduced: an improved Deep Q‐Network (DQN) that integrates a multi‐layer perceptron, a cubic dynamic reward function, and an adaptive exploration strategy; and a PER‐D3QN model that combines dueling double deep Q‐learning (D3QN) and prioritized experience replay (PER) to mitigate Q ‐value overestimation and accelerate convergence. Experimental evaluation using real‐world teacher data demonstrates that the improved DQN achieved average performance scores up to 0.36, compared to 0.054–0.068 for the traditional DQN. Moreover, the PER‐D3QN model outperformed the ERDQN baseline, attaining an average reward of 4.738 versus 2.021, and an average score of 2.799 after 6000 training rounds, compared to 1.946 for ERDQN, indicating that network update speed has also been significantly improved. This research not only helps to enhance teachers' professional knowledge and technical application ability in the field of network security, but also provides scientific methodological support for educational institutions to ensure that they are aligned with changing security threats. Furthermore, this study emphasizes the importance of interdisciplinary cooperation and encourages experts from computer science, education, and psychology to work.
- 27 June 2025
- Published by Wiley in Asian Journal of Control
- Vol. 28 (1) , 431-450
Pitch angle control of squirrel‐cage induction generator‐based wind farm using fractional order PID and neural network controllers
Wind farms constantly inject varying power into the traditional power grid, affecting its operation and resulting in power quality and stability issues. Multivariable‐based control strategies for pitch angle control are most widely utilized to stabilize the fluctuations in the output wind power. In this research work, Fractional Order PID (FOPID) and Model Reference Neural Network (MR‐NN) controllers are proposed for controlling the pitch angle of a squirrel‐cage induction generator (SCIG)‐based wind turbine to stabilize the output active power under varying wind speeds. For comparison purposes, an Integer Order PI (IOPI) controller is also designed and implemented. Particle Swarm Optimization, Artificial Bee Colony (ABC), and Genetic Algorithm have been used to optimize the tuning parameters of both the IOPI and FOPID controllers. All the optimizing algorithms and controllers have been designed and realized in MATLAB. A comparative analysis of all the optimized variants of IOPI and FOPI controllers with MR‐NN controllers has been performed. The FOPID controller using the ABC algorithm outperforms all other control techniques in both steady‐state and transient‐state performance. The FOPID‐ABC controller achieved the fastest settling time (27.3 s) and the lowest steady‐state error (−0.0006), demonstrating superior stability and precision in power regulation.
- 27 June 2025
- Published by Wiley in Asian Journal of Control
- Vol. 28 (1) , 431-450
Pitch angle control of squirrel‐cage induction generator‐based wind farm using fractional order PID and neural network controllers
Wind farms constantly inject varying power into the traditional power grid, affecting its operation and resulting in power quality and stability issues. Multivariable‐based control strategies for pitch angle control are most widely utilized to stabilize the fluctuations in the output wind power. In this research work, Fractional Order PID (FOPID) and Model Reference Neural Network (MR‐NN) controllers are proposed for controlling the pitch angle of a squirrel‐cage induction generator (SCIG)‐based wind turbine to stabilize the output active power under varying wind speeds. For comparison purposes, an Integer Order PI (IOPI) controller is also designed and implemented. Particle Swarm Optimization, Artificial Bee Colony (ABC), and Genetic Algorithm have been used to optimize the tuning parameters of both the IOPI and FOPID controllers. All the optimizing algorithms and controllers have been designed and realized in MATLAB. A comparative analysis of all the optimized variants of IOPI and FOPI controllers with MR‐NN controllers has been performed. The FOPID controller using the ABC algorithm outperforms all other control techniques in both steady‐state and transient‐state performance. The FOPID‐ABC controller achieved the fastest settling time (27.3 s) and the lowest steady‐state error (−0.0006), demonstrating superior stability and precision in power regulation.
- 19 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (15) , 30941-30954
IoT-Enabled Modular Bird Repellent System: A Pattern-Based Approach for Enhanced Detection and Control
Effective bird repellent systems are crucial for mitigating the negative impacts of birds on agriculture, aviation safety, and urban infrastructure. This article presents an Internet of Things (IoT)-enabled modular bird repellent system that integrates real-time sensing, intelligent classification, and adaptive actuation to enhance efficiency and scalability. A pattern-based method is proposed for classifying birds and assessing the risk of bird populations. Experimental data showed that the detection accuracy of the perception module reached 95.2%, and the signal-to-noise ratio during daytime and nighttime was about 43.6 and 40.5 dB after stable image acquisition. The success rate of the driving module during the daytime within $500~{\mathrm { m}}^{2}$ was about 97.6%. The error rate of human-machine interaction module status detection was less than 2.6%, and the remote-control response time tended to be 157 ms. In communication testing, Ethernet, serial port, and USB communication had lower packet loss and error rates, and biomimetic bus’s performance was better than conventional bus. After 48 h of on-site testing, the bird repellent efficiency reached 94.6%, with only two malfunctions and a recovery time of less than 5 s. The integration of functional modules ensured efficient collaboration, while the inclusion of multiple communication methods and a biomimetic bus enhanced the system’s stability and adaptability to complex environments. This work demonstrates the potential of IoT-based intelligent bird repellent systems for precision wildlife management and environmental sustainability.
- 19 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (15) , 30941-30954
IoT-Enabled Modular Bird Repellent System: A Pattern-Based Approach for Enhanced Detection and Control
Effective bird repellent systems are crucial for mitigating the negative impacts of birds on agriculture, aviation safety, and urban infrastructure. This article presents an Internet of Things (IoT)-enabled modular bird repellent system that integrates real-time sensing, intelligent classification, and adaptive actuation to enhance efficiency and scalability. A pattern-based method is proposed for classifying birds and assessing the risk of bird populations. Experimental data showed that the detection accuracy of the perception module reached 95.2%, and the signal-to-noise ratio during daytime and nighttime was about 43.6 and 40.5 dB after stable image acquisition. The success rate of the driving module during the daytime within $500~{\mathrm { m}}^{2}$ was about 97.6%. The error rate of human-machine interaction module status detection was less than 2.6%, and the remote-control response time tended to be 157 ms. In communication testing, Ethernet, serial port, and USB communication had lower packet loss and error rates, and biomimetic bus’s performance was better than conventional bus. After 48 h of on-site testing, the bird repellent efficiency reached 94.6%, with only two malfunctions and a recovery time of less than 5 s. The integration of functional modules ensured efficient collaboration, while the inclusion of multiple communication methods and a biomimetic bus enhanced the system’s stability and adaptability to complex environments. This work demonstrates the potential of IoT-based intelligent bird repellent systems for precision wildlife management and environmental sustainability.
- Open Access
- 14 April 2025
- Published by Elsevier in Energy Reports
- Vol. 13, 4441-4458
Optimizing electrical and thermal energy storage systems for hour-ahead integrated demand response in industries
- 14 April 2025
- Published by Elsevier in Energy Reports
- Vol. 13, 4441-4458
Open AccessOptimizing electrical and thermal energy storage systems for hour-ahead integrated demand response in industries
- Open Access
- 7 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 108986-109000
New Appliance Signatures for NILM Based on Mono-Fractal Features and Multi-Fractal Formalism
Smart energy management demands better ways to understand the energy consumption of buildings. Nonintrusive Load Monitoring (NILM) is an emerging technique that disaggregates total building consumption to individual appliances. However, low-power appliances pose a challenge as they exhibit similar power consumption patterns that are difficult to identify and distinguish from background noise. This research aims to bridge this gap by developing feature extraction and classification techniques specifically for low-power NILM. This may facilitate the development of targeted energy-saving strategies and result in more precise energy-use monitoring. To effectively identify and differentiate low-power appliances within a NILM system, this work proposes a novel feature extraction approach that combines mono-fractal and multifractal analysis of appliance startup current transients. Mono-fractal features, including fractal dimension and Hurst exponent, are extracted alongside Lacunarity and Multifractal features comprising the singularity spectrum and Hölder exponents. These features collectively create comprehensive appliance signatures that capture the unique characteristics of each appliance during its turn-on event. Building upon these extracted features, the work investigates the performance of three machine learning classifiers for appliance classification: deep neural network, support vector machine, and K-nearest neighbours. These classifiers are optimized using Bayesian optimization to achieve optimum performance. The proposed method demonstrates significant improvement over existing feature extraction approaches across all optimized classifiers, achieving an accuracy of up to 98.3% in classifying low-power appliances.
- 7 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 108986-109000
Open AccessNew Appliance Signatures for NILM Based on Mono-Fractal Features and Multi-Fractal Formalism
Smart energy management demands better ways to understand the energy consumption of buildings. Nonintrusive Load Monitoring (NILM) is an emerging technique that disaggregates total building consumption to individual appliances. However, low-power appliances pose a challenge as they exhibit similar power consumption patterns that are difficult to identify and distinguish from background noise. This research aims to bridge this gap by developing feature extraction and classification techniques specifically for low-power NILM. This may facilitate the development of targeted energy-saving strategies and result in more precise energy-use monitoring. To effectively identify and differentiate low-power appliances within a NILM system, this work proposes a novel feature extraction approach that combines mono-fractal and multifractal analysis of appliance startup current transients. Mono-fractal features, including fractal dimension and Hurst exponent, are extracted alongside Lacunarity and Multifractal features comprising the singularity spectrum and Hölder exponents. These features collectively create comprehensive appliance signatures that capture the unique characteristics of each appliance during its turn-on event. Building upon these extracted features, the work investigates the performance of three machine learning classifiers for appliance classification: deep neural network, support vector machine, and K-nearest neighbours. These classifiers are optimized using Bayesian optimization to achieve optimum performance. The proposed method demonstrates significant improvement over existing feature extraction approaches across all optimized classifiers, achieving an accuracy of up to 98.3% in classifying low-power appliances.
- Top Cited Papers
- 1 August 2021
- Published by Elsevier in Expert Systems with Applications
- Vol. 175
Deep sequence to sequence Bi-LSTM neural networks for day-ahead peak load forecasting
- 1 August 2021
- Published by Elsevier in Expert Systems with Applications
- Vol. 175
Top Cited PapersDeep sequence to sequence Bi-LSTM neural networks for day-ahead peak load forecasting