Research Topics

Overview (Click to enlarge)

Testbed Illustration (Click to enlarge)

Background of Existing Air-Conditioning Systems

In the centralized Variable Air Volume (VAV) air-conditioning system, the water chiller produces chilled water at a fixed temperature (typically 4°C to 7°C in Singapore) which flows at a fixed flow rate. While receiving the chilled water, the Air Handling Unit (AHU) receives a mix of outside air and recirculated air at the same time. Outside air is used because it keeps carbon dioxide within mandatory levels; recirculated air is used because it has lower humidity and is already cooled. A heat exchanger uses the chilled water to cool the air supplied to the AHU to a preset temperature setpoint (typically 12°C to 14°C in Singapore). The cool air output is then forced by a supply fan into the duct network in the building. Fans are responsible for creating sufficient pressure differences to ensure that enough cool air is supplied to the cooling zones.

Fig. 1. A cooling zone in a building HVAC system.

A zone is an area inside a building that is controlled by a single thermostat. It could be part of a large room or might comprise of several small rooms. The user sets the temperature he/she requires on the thermostat in the room. The VAV box (i.e. controller) receives this information from the thermostat and changes the damper position by sending an electrical signal to the damper actuator.

Cool air comes in from the supply duct and mixes with the existing warmer air to cool the room. It then flows back to the AHU through the return duct. The damper actually controls the mass flow rate of the supply air into the zone by altering the cross-sectional area of the duct. When a user sets a lower temperature, the damper will open more to allow more cool air into the room; when a user sets a higher temperature, the damper will turn the other direction to close or to limit the amount of cool air into the room.

Problem Statement

(1) The power drawn by the fans is huge and is approximately a quadratic function of the mass flow rate of air. Thus, large peaks in the air mass flow rate are undesirable.

(2) Conservative pre-cooling strategies fail to flatten the air mass flow rate sufficiently. Also, precooling well in advance of what is required increases energy consumption because the cumulative cooling load is larger than necessary.

(3) Currently, many commercial buildings in Singapore use a simple pre-cooling strategy. The cooling system is switched on at a fixed time (say 1 hour) before the work day begins (9 am). All zones are cooled to the same setpoint, irrespective of their cooling load and without regard to occupancy.

(4) Many HVAC control methods are centralized and involve sophisticated optimal control methods which aim to minimize the total energy consumption across all zones.

Token-Based Scheduling

Our interest is in minimizing the overall operational power consumption (fan power and chiller power) in the AC system through smart scheduling and distributed control. We propose a novel distributed architecture (called the token-based architecture) for controlling HVAC systems in commercial buildings.

Fig. 2. A flowchart of our token-based scheduling architecture.

The zone module receives temperature, humidity, and carbon dioxide measurements transmitted from general room sensors. Then it performs forecasting of temperature, cooling load, and occupancy. This information is used to compute request for cooling in the form of mass flow rate. We called this a token request. The mass flow rate token request translates into damper opening (in the form of percentage) to be mechanically executed by the damper actuator. The entry of supply air into the zone (or room) is therefore adjusted.

The token requests from various zones may be competing with each other and eventually overloading the system, leading to energy inefficiency. Hence, a centralized scheduler is proposed to balance requests by allocating tokens to each zone efficiently by demand. In this approach, we minimize total energy use while respecting operational constraints.

In this model predictive control framework, zone modules update their local models based on the measured thermal response under allocated tokens. Then they re-compute and forward fresh token requests for subsequent time slots.

Our objective is to minimize the total energy consumption over the window k = 1,...,W, where

is subject to zone thermal dynamics, constraints on the acceptable temperature ranges, and limits on the cooling air mass flow rates. Hp is hour; Pf is fan power; Pc is chiller power.

This proposed token-based architecture offers several advantages:

(a) The architecture is scalable in actual building implementation and robust to environmental changes.

(b) Computational burden on both the zone modules and the centralized scheduler is modest.

(c) The zone module only processes and handles its own zone, thus cutting down computation time.

(d) The hardware infrastructure will be modular. This implies that deployment cost will be reduced.

Learning-based Token Scheduling

This is an enhancement of the earlier token-based scheduling method. Learning-based token scheduling is a hierarchical distributed algorithm that aims at achieving energy savings in HVAC operations without compromising human comfort. It captures the influences of occupancy and user interactions. The algorithm involves these 5 steps:

1) Learning: each zone is controlled using numerous zone modules, which first runs a learning algorithm to correct (if needed) the zone thermal model using an moving horizon estimation approach.

2) Token Requests: the main role of the zone module is to run a model predictive controller (MPC) using forecast information on weather, occupancy, and cooling demands plus sensor readings (temperature, thermostat, and occupancy sensors) to compute the minimal energy required without breaching user-defined comfort margins. The minimum cooling energy computed is called cooling energy token.

3) Coefficient of Performance (COP) Constraints: the centralized controller checks whether the increase in cooling energy to be supplied using tokens as the lower bound will benefit the COP of the chiller system, thereby increasing the energy efficiency and reducing maintenance cost. The COP constraints are computed in the central scheduler and provided as input to the the quadratically constrained quadratic program (QCQP) model based on token requests from zone modules.

4) Mass Flow Rate Constraints: the tokens are converted into mass flow rates and the corresponding constraints are calculated. The minimum number of tokens required by each zone is computed based on the mass flow rate constraints in the central scheduler using zone temperature profiles and token requests from zone modules.

5) Token Allocation: the algorithm modifies the supply mass flow rate allocation to all zones, such that the duct pressure constraints are satisfied. It also ensures that the minimum damper position constraints are not violated. The computation is reduced to QCQP.

Optimal Scheduling with Non-preemptive Air Distribution for Precooling

The in-building section of a HVAC system is a major contributor to its overall energy consumption. This work looks into energy savings at the supply fan (which takes up to 30% of HVAC in-building energy consumption) while energy consumptions at the chilled water pump and the chiller are also taken into account. By elongating the cooling process, the energy consumption can be significantly reduced due to the nonlinear feature of the fan power consumption. To ensure that a computed assignment of mass flow rates is practically feasible, we developed an algorithm to check whether there exist a fan supply pressure and a damper opening to realize these mass flow rates.

Fig. 3. Analysis of air distribution in every duct and tracking the duct for a certain amount of air flow.

We want to determine the minimum supply pressure at the entrance of the supply duct. After choosing any one supply pressure (which should be more than the minimum supply pressure), we can determine the pressure value at each node of the duct network. We are then able to determine the damper opening in each room based on their assigned mass flow rates. Since we are dealing with a scheduling problem, that means we are interested in steady-state mass flow rates or average mass flow rates in the very slow cooling process. By locking onto the values of damper opening and fan supply pressure, steady-state mass flow rates will appear in the duct network.

An Internet of Things Compliant Model Identification Methodology for Smart Buildings

The token-based scheduling system (TBSS) provides significant scalability and reasonable performance. However, during implementation, it requires solving the MPC in a distributed fashion. The MPC assumes a reasonable model of the zone, estimate on the disturbances, and uses an optimization routine to compute the future control moves that optimize the energy consumption. As building models are complex and nonlinear, it depends on numerous variables intrinsic and external to a building that are time-varying, the model needs to be learned online.

In a TBSS, the network bandwidth emerges as a key concern as sensor readings have to be transmitted at regular intervals. As the scale of the building increases, techniques to reduce network bandwidth without affecting model fidelity are required. This requires optimal utilization of the network bandwidth and allocation of sensors. The problem is compounded by the presence of data losses. Consequently, model identification methods that can reduce network bandwidth and are robust to packet loss are required for HVAC systems. This can be achieved by showing that system identification can be done with more sparse measurements. Therefore, the user does not need to transmit intense data sets through the network to identify the zone model. At the same time, since model identification is done in real time, an efficient system identification algorithm is needed to identify many different sparse zone model at the same time with the available data.