Real-time data driven model predictive control for efficient energy consumption in smart buildings

Abstract

Model Predictive Control (MPC) has received a lot of attention in recent years mainly in the field of Heating, Ventilation, and Air Conditioning (HVAC) control in smart buildings. MPC is an optimal control that improves the energy efficiency of HVAC systems. This is achieved by using a model-based control approach that integrates a mathematical representation of the building with the most important factors that affect the building dynamics. However designing an appropriate controller that accurately models the dynamics of the physical system is a challenging task in real applications, especially for multi-zone building and for different HVAC system types. Moreover, the non linearity of the buildings thermal dynamics makes the Indoor Air Temperature (IAT) prediction more challenging since it is affected by complex factors such as controlled and uncontrolled points, outside weather conditions and occupancy schedule. Modern smart buildings are equipped with multiple sensors that collect data, which is then used by optimal control techniques to improve energy efficiency while maintaining comfort levels. The availability of historical data opens the opportunity to develop data-driven control solutions based on artificial intelligence algorithms. Data-driven control reduces the cost and time consuming tasks caused by MPC that requires an accurate and complex modeling processes. Thus, the goal of this dissertation is to provide an efficient and scalable data-driven HVAC control framework that minimizes energy consumption, carbon emission, peak demand and discomfort during occupied hours under self-tuned setpoint, temperature ramp and equipment cycling constraints which integrates a multi-step temperature prediction model that consider control sensitivities. In order to meet this goal, four key issues are required to be addressed in our framework and are summarized as follows: i) how to model IAT in a multi-zone smart building and for different types of HVAC systems without decreasing the prediction accuracy?, ii) how to accurately model a multi-step IAT prediction in a data-driven MPC framework without bias on the optimization decision for the control outputs?, iii) how to design and deploy an efficient real-time data-driven MPC optimization problem suitable for a real-time HVAC system application?, and iv) how to model a more scalable data-driven control system for HVAC system while reducing energy consumption and carbon footprint? As part of our contributions to address the first issue highlighted above and to accurately model an IAT prediction model especially for multi-zone building and for different HVAC system types, we propose a new IAT prediction model based on Long Short Term Memory (LSTM) model. LSTM-MISO and LSTM-MIMO strategies are built for multi-input single-output and multi-input multi-output, respectively. A direct prediction with sequence-to-sequence (S2S) approach has been developed to predict multi-step ahead. Furthermore, a feature selection analysis has been performed to obtain optimal model structure for both variable air volume (VAV) and constant air volume (CAV) systems. Since the temperature behavior depends on the time of the action taken by control variables in the HVAC system, it is found that the consideration of these control variables as input increases the prediction accuracy performance. The performance of different strategies has been evaluated based on two case studies on real smart buildings operational data using VAV and CAV systems. For both buildings, experimental results showed that the proposed models outperform Multilayer Perceptrons models by reducing the mean absolute percentage error by 50%. To address the second issue, we extend the first research objective and propose a new multi-step IAT prediction model based on a context-aware multivariate LSTM (CAM-LSTM) to be used in data-driven MPC framework without bias on the optimization decision for the control outputs. CAM-LSTM is based on high-level and low-level interaction between input features and considers the sensitive relationship between temperature and control parameters. Moreover, CAM-LSTM uses a dual-stream neural networks based on multivariate time series of controlled and uncontrolled inputs. In addition, an attention mechanism is applied on controlled parameters to give them more weight to better predict the zone temperature. To address the third issue, we propose an efficient real-time data-driven control framework named Model Predictive Control via Genetic algorithm (MPC-GA) allowing the optimal operation of HVAC system and has been experimentally validated in a multi-zone retail building. The MPC-GA combines CAM-LSTM model with a MPC framework. The prediction model is used in the optimization model which minimizes: energy consumption, peak demand and discomfort during occupied hours under self-tuned setpoint, temperature ramp and equipment cycling constraints. A heuristic search algorithm using a genetic algorithm is used to solve the real-time data-driven MPC-GA models and obtain the future optimal combination settings of all controls for all the zones over a prediction horizon. The benchmark results showed that the MPC-GA outperforms RBC control systems with more than 50% and 80% reduction in energy consumption and discomfort respectively. Finally, we introduce a scalable multi-agent based distributed approach for optimized control of a multi-zone smart building based on a set of local agents which represent individual zones in the building, coordinated by a central agent. For each control horizon, the coordinator minimizes the overall carbon emissions and assigns an individual energy budget to each local agent. Each local agent minimizes the discomfort in its zone while respecting the energy budget assigned by the coordinator. We propose a heuristic search based on a genetic algorithm to find the optimized control sequences in each zone, and formulate an integer linear programming (ILP) model for the coordinator problem which can be solved using an ILP solver. For a representative winter test day, the proposed methodology gave an energy savings of 8.8% and reduced the carbon footprint by 23.4%.

Date	27 Feb 2023
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Mohamed Cheriet (Supervisor) & Kim Khoa Nguyen (Co-supervisor)