Gulai Shen
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Gulai Shen has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
Ali Malkawi
6 shared publicationsNa Li
5 shared publications- Elence Xinzhu Chen3 shared publications
- Ahmed S. Alahmed
Electrical Engineering Department, King Fahd University of Petroleum and Minerals, KSA
2 shared publications Ali Mehmani
2 shared publications- Chuchu Fan
Massachusetts Institute of Technology
2 shared publications
Latest Publications
- 19 November 2026
- Published by Elsevier in Building and Environment
A Gaussian process-based method for interactive individual occupant comfort modeling
- 19 November 2026
- Published by Elsevier in Building and Environment
A Gaussian process-based method for interactive individual occupant comfort modeling
- Open Access
- 24 September 2026
- Published by Elsevier in SSRN Electronic Journal
Experimental Validation of an Ensembled Deep-learning ModelPredictive Control Framework for Natural Ventilation
Naturally ventilated buildings can substantially reduce energy use, but maintaining reliable indoor air quality (IAQ) and thermal comfort in winter remains challenging, particularly under high occupancy and single-sided ventilation. This study presents the full-scale deployment and winter validation of an Ensembled Multi-time scale deep-learning-based Adaptive Model Predictive Control (EMA-MPC) framework for automated window control. EMA-MPC integrates an ensemble of Long Short-Term Memory (LSTM) models that predict indoor temperature and CO2 at fast and slow time scales with data-driven disturbance models that localize weather forecasts and adjacent-zone temperatures. These predictions feed a multi-objective MPC formulation that minimizes CO2 exceedance, thermal comfort violations, and window actuation while enforcing operational constraints.Field experiments were conducted in two single-sided ventilation zones: a small office ”Live Lab” with 1–2 occupants and a basement Conference Room with 11–21 occupants. In the Live Lab, EMA-MPC increased the fraction of occupied time with CO2 below 800 ppm from 72.9% to 100% for single occupancy and from 60.6% to 90.1% and 95.9% for two-person occupancy, relative to a rule-based baseline, while maintaining indoor temperatures above 20°C and comparable heating loads despite colder outdoor conditions. In the Conference Room, under near-capacity occupancy, EMA-MPC reduced peak CO2 by approximately 45% compared with a pulse-ventilation baseline that exceeded 3,000 ppm, while maintaining temperatures near 22–24°C. These results demonstrate that adaptive deep-learning-based MPC can substantially improve winter IAQ without consistently increasing observed heating load while explicitly managing IAQ–comfort trade-offs under real ventilation constraints.
- 24 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessExperimental Validation of an Ensembled Deep-learning ModelPredictive Control Framework for Natural Ventilation
Naturally ventilated buildings can substantially reduce energy use, but maintaining reliable indoor air quality (IAQ) and thermal comfort in winter remains challenging, particularly under high occupancy and single-sided ventilation. This study presents the full-scale deployment and winter validation of an Ensembled Multi-time scale deep-learning-based Adaptive Model Predictive Control (EMA-MPC) framework for automated window control. EMA-MPC integrates an ensemble of Long Short-Term Memory (LSTM) models that predict indoor temperature and CO2 at fast and slow time scales with data-driven disturbance models that localize weather forecasts and adjacent-zone temperatures. These predictions feed a multi-objective MPC formulation that minimizes CO2 exceedance, thermal comfort violations, and window actuation while enforcing operational constraints.Field experiments were conducted in two single-sided ventilation zones: a small office ”Live Lab” with 1–2 occupants and a basement Conference Room with 11–21 occupants. In the Live Lab, EMA-MPC increased the fraction of occupied time with CO2 below 800 ppm from 72.9% to 100% for single occupancy and from 60.6% to 90.1% and 95.9% for two-person occupancy, relative to a rule-based baseline, while maintaining indoor temperatures above 20°C and comparable heating loads despite colder outdoor conditions. In the Conference Room, under near-capacity occupancy, EMA-MPC reduced peak CO2 by approximately 45% compared with a pulse-ventilation baseline that exceeded 3,000 ppm, while maintaining temperatures near 22–24°C. These results demonstrate that adaptive deep-learning-based MPC can substantially improve winter IAQ without consistently increasing observed heating load while explicitly managing IAQ–comfort trade-offs under real ventilation constraints.
- Open Access
- 3 July 2026
- Published by Elsevier in SSRN Electronic Journal
A Gaussian Process-based method for interactive individual occupant comfort modeling
Occupant-centric building control relies on accurate occupant comfort models to balance energy efficiency and occupant satisfaction, with increasing interest in transitioning from group-based to personalized modeling. However, modeling individual comfort remains challenging due to the difficulty of collecting personal data and the high variability in individual behavior. This paper presents a mobile-based human building interface for collecting and interacting with natural language feedback from occupants. Building on this, a novel Gaussian Process-based method is developed to efficiently and accurately model individual occupant comfort using occupants' natural language feedback. Using synthetic feedback generated from three simulated occupants with distinct comfort parameters and feedback reporting behaviors, the proposed method is able to recover true comfort with MAEs between 0.0458 and 0.0659 on a normalized scale. The results demonstrate that, with a reasonable amount of data, computational resources, and occupant feedback, the approach can adaptively and continuously learn individual comfort from sparse and natural language inputs. Furthermore, the proposed framework provides a pathway for integrating uncertainty-aware individual comfort modeling into occupant-centric control systems for smart and sustainable buildings.
- 3 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Gaussian Process-based method for interactive individual occupant comfort modeling
Occupant-centric building control relies on accurate occupant comfort models to balance energy efficiency and occupant satisfaction, with increasing interest in transitioning from group-based to personalized modeling. However, modeling individual comfort remains challenging due to the difficulty of collecting personal data and the high variability in individual behavior. This paper presents a mobile-based human building interface for collecting and interacting with natural language feedback from occupants. Building on this, a novel Gaussian Process-based method is developed to efficiently and accurately model individual occupant comfort using occupants' natural language feedback. Using synthetic feedback generated from three simulated occupants with distinct comfort parameters and feedback reporting behaviors, the proposed method is able to recover true comfort with MAEs between 0.0458 and 0.0659 on a normalized scale. The results demonstrate that, with a reasonable amount of data, computational resources, and occupant feedback, the approach can adaptively and continuously learn individual comfort from sparse and natural language inputs. Furthermore, the proposed framework provides a pathway for integrating uncertainty-aware individual comfort modeling into occupant-centric control systems for smart and sustainable buildings.
- Open Access
- 3 July 2026
- Published by Elsevier in SSRN Electronic Journal
A Markov chain-based method for interactive individual occupancy status forecasting
Occupant-centric building controls require accurate individual occupancy forecasts to improve energy efficiency and occupant comfort, but this is challenging because personal occupancy data are difficult to collect and individual behavior is highly variable. This paper presents a mobile human-building interface that enables interactive natural-language feedback to efficiently collect individual occupancy data. We also develop an adaptive Markov chain-based forecasting method that incorporates arrival information smoothing, duration-of-stay-based transition adjustment, and feedback-based schedule updates. Prediction performance is evaluated using standard classification problem metrics including accuracy, precision, and recall, where higher precision reduces false occupancy predictions and higher recall captures a greater fraction of actual occupancy periods. Using actual occupancy data collected, the proposed method achieves 0.97 accuracy, 0.91 precision, and 0.87 recall during the stable summer period, and 0.95 accuracy, 0.92 precision, and 0.78 recall during the more variable fall period. It outperforms a baseline using average historical data based on time and day type as well as a supervised-learning benchmark based on a time-inhomogeneous Markov chain model. Additional tests using a public occupancy dataset without interactive feedback also show consistent performance for the developed model, achieving the highest accuracy up to 0.97 and precision up to 0.95, while a slight drop in recall. These results demonstrate a practical approach for uncertainty-aware and interactive individual-level occupancy prediction in occupant-centric building control.
- 3 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Markov chain-based method for interactive individual occupancy status forecasting
Occupant-centric building controls require accurate individual occupancy forecasts to improve energy efficiency and occupant comfort, but this is challenging because personal occupancy data are difficult to collect and individual behavior is highly variable. This paper presents a mobile human-building interface that enables interactive natural-language feedback to efficiently collect individual occupancy data. We also develop an adaptive Markov chain-based forecasting method that incorporates arrival information smoothing, duration-of-stay-based transition adjustment, and feedback-based schedule updates. Prediction performance is evaluated using standard classification problem metrics including accuracy, precision, and recall, where higher precision reduces false occupancy predictions and higher recall captures a greater fraction of actual occupancy periods. Using actual occupancy data collected, the proposed method achieves 0.97 accuracy, 0.91 precision, and 0.87 recall during the stable summer period, and 0.95 accuracy, 0.92 precision, and 0.78 recall during the more variable fall period. It outperforms a baseline using average historical data based on time and day type as well as a supervised-learning benchmark based on a time-inhomogeneous Markov chain model. Additional tests using a public occupancy dataset without interactive feedback also show consistent performance for the developed model, achieving the highest accuracy up to 0.97 and precision up to 0.95, while a slight drop in recall. These results demonstrate a practical approach for uncertainty-aware and interactive individual-level occupancy prediction in occupant-centric building control.
- Open Access
- 15 April 2026
- Published by Elsevier in SSRN Electronic Journal
Experimental Validation of an Ensembled Deep-learning Model Predictive Control Framework for Natural Ventilation
Naturally ventilated buildings can substantially reduce energy use, but maintaining reliable indoor air quality (IAQ) and thermal comfort in winter, especially under high occupancy and single-sided ventilation, remains challenging. This study reports the full-scale experimental deployment and winter-season validation of an Ensembled Multi-time scale deep-learning-based Adaptive Model Predictive Control (EMA-MPC) framework for automated window control. The controller integrates an ensemble of Long Short-Term Memory (LSTM) models that predict indoor temperature and CO2 at fast and slow time scales, combined with data-driven disturbance models that localize weather forecasts and adjacent-zone temperatures. These predictions feed into a multi-objective MPC formulation that minimizes CO2 exceedance, thermal comfort violations, and window actuation while enforcing operational and comfort constraints. Field experiments were conducted in two single-sided ventilation zones: a small office ”Live Lab” with 1–2 occupants and a basement Conference Room with 11–21 occupants. In the ”Live Lab”, EMA-MPC increased the fraction of occupied time with CO2 below 800 ppm from 61.1% to 100% for single occupancy and from 41.5% to 81.4% for two-person occupancy, relative to a rule-based control baseline, while keeping indoor temperatures above 20 °C and heating loads comparable despite colder outdoor conditions. In the Conference Room, EMA-MPC acheived CO2 level 45% lower than a heuristic pulse-ventilation that exceeded 3,000 ppm—while maintaining temperatures near 22–24°C. These results demonstrate that an adaptive deep-learning based MPC can be robustly implemented in occupied naturally ventilated buildings to substantially improve winter IAQ without additional energy use, while explicitly managing IAQ–comfort trade offs under real ventilation constraints.
- 15 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessExperimental Validation of an Ensembled Deep-learning Model Predictive Control Framework for Natural Ventilation
Naturally ventilated buildings can substantially reduce energy use, but maintaining reliable indoor air quality (IAQ) and thermal comfort in winter, especially under high occupancy and single-sided ventilation, remains challenging. This study reports the full-scale experimental deployment and winter-season validation of an Ensembled Multi-time scale deep-learning-based Adaptive Model Predictive Control (EMA-MPC) framework for automated window control. The controller integrates an ensemble of Long Short-Term Memory (LSTM) models that predict indoor temperature and CO2 at fast and slow time scales, combined with data-driven disturbance models that localize weather forecasts and adjacent-zone temperatures. These predictions feed into a multi-objective MPC formulation that minimizes CO2 exceedance, thermal comfort violations, and window actuation while enforcing operational and comfort constraints. Field experiments were conducted in two single-sided ventilation zones: a small office ”Live Lab” with 1–2 occupants and a basement Conference Room with 11–21 occupants. In the ”Live Lab”, EMA-MPC increased the fraction of occupied time with CO2 below 800 ppm from 61.1% to 100% for single occupancy and from 41.5% to 81.4% for two-person occupancy, relative to a rule-based control baseline, while keeping indoor temperatures above 20 °C and heating loads comparable despite colder outdoor conditions. In the Conference Room, EMA-MPC acheived CO2 level 45% lower than a heuristic pulse-ventilation that exceeded 3,000 ppm—while maintaining temperatures near 22–24°C. These results demonstrate that an adaptive deep-learning based MPC can be robustly implemented in occupied naturally ventilated buildings to substantially improve winter IAQ without additional energy use, while explicitly managing IAQ–comfort trade offs under real ventilation constraints.
- 20 March 2026
- Published by Elsevier in Energy and Buildings
- Vol. 360
Asynchronous multi-agent reinforcement learning for coordinated control of natural ventilation and radiant cooling
- 20 March 2026
- Published by Elsevier in Energy and Buildings
- Vol. 360
Asynchronous multi-agent reinforcement learning for coordinated control of natural ventilation and radiant cooling
- 29 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Smart Grid
- Vol. 17 (3) , 1914-1927
Joint Scheduling of DER Under Demand Charges: Structure and Approximation
We study the joint scheduling of behind-the-meter distributed energy resources (DERs), including flexible loads, renewable generation, and battery energy storage systems, under net energy metering tariffs with demand charges. The problem is formulated as a stochastic dynamic program aimed at maximizing expected operational surplus while accounting for renewable generation uncertainty. We analytically characterize the optimal control policy and show that it admits a threshold-based structure. However, due to the strong temporal coupling of the storage and demand charge constraints, the number of conditional branches in the policy scales combinatorially with the scheduling horizon, as it requires a look-ahead over future states. To overcome the high computational complexity in the general formulation, an efficient approximation algorithm is proposed, which searches for the peak demand under a mildly relaxed problem. We show that the algorithm scales linearly with the scheduling horizon. Extensive simulations using two open-source datasets validate the proposed algorithm and compare its performance against different DER control strategies, including a reinforcement learning-based one. Under varying storage and tariff parameters, the results show that the proposed algorithm outperforms various benchmarks in achieving a relatively small solution gap compared to a theoretical upper bound.
- 29 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Smart Grid
- Vol. 17 (3) , 1914-1927
Joint Scheduling of DER Under Demand Charges: Structure and Approximation
We study the joint scheduling of behind-the-meter distributed energy resources (DERs), including flexible loads, renewable generation, and battery energy storage systems, under net energy metering tariffs with demand charges. The problem is formulated as a stochastic dynamic program aimed at maximizing expected operational surplus while accounting for renewable generation uncertainty. We analytically characterize the optimal control policy and show that it admits a threshold-based structure. However, due to the strong temporal coupling of the storage and demand charge constraints, the number of conditional branches in the policy scales combinatorially with the scheduling horizon, as it requires a look-ahead over future states. To overcome the high computational complexity in the general formulation, an efficient approximation algorithm is proposed, which searches for the peak demand under a mildly relaxed problem. We show that the algorithm scales linearly with the scheduling horizon. Extensive simulations using two open-source datasets validate the proposed algorithm and compare its performance against different DER control strategies, including a reinforcement learning-based one. Under varying storage and tariff parameters, the results show that the proposed algorithm outperforms various benchmarks in achieving a relatively small solution gap compared to a theoretical upper bound.
- Open Access
Co-Optimizing Distributed Energy Resources under Demand Charges and Bi-Directional Power Flow
We address the co-optimization of behind-the-meter (BTM) distributed energy resources (DER), including flexible demands, renewable distributed generation (DG), and battery energy storage systems (BESS) under net energy metering (NEM) frameworks with demand charges. We formulate the problem as a stochastic dynamic program that accounts for renewable generation uncertainty and operational surplus maximization. Our theoretical analysis reveals that the optimal policy follows a threshold structure. Finally, we show that even a simple algorithm leveraging this threshold structure performs well in simulation, emphasizing its importance in developing near-optimal algorithms. These findings provide crucial insights for implementing prosumer energy management systems under complex tariff structures.
Open AccessCo-Optimizing Distributed Energy Resources under Demand Charges and Bi-Directional Power Flow
We address the co-optimization of behind-the-meter (BTM) distributed energy resources (DER), including flexible demands, renewable distributed generation (DG), and battery energy storage systems (BESS) under net energy metering (NEM) frameworks with demand charges. We formulate the problem as a stochastic dynamic program that accounts for renewable generation uncertainty and operational surplus maximization. Our theoretical analysis reveals that the optimal policy follows a threshold structure. Finally, we show that even a simple algorithm leveraging this threshold structure performs well in simulation, emphasizing its importance in developing near-optimal algorithms. These findings provide crucial insights for implementing prosumer energy management systems under complex tariff structures.
- 7 December 2024
- Published by Elsevier in Building and Environment
- Vol. 269
Using PeopleHour for occupant-centric office building performance assessment
- 7 December 2024
- Published by Elsevier in Building and Environment
- Vol. 269
Using PeopleHour for occupant-centric office building performance assessment
- 15 November 2023
- Published by Elsevier in Building and Environment
- Vol. 247
Indoor environmental wellness index (IEW-Index): Towards intelligent building systems automation and optimization
- 15 November 2023
- Published by Elsevier in Building and Environment
- Vol. 247
Indoor environmental wellness index (IEW-Index): Towards intelligent building systems automation and optimization