ASHRAE OR-16-C077-2016 Smart Windows Control Strategies for Building Energy Savings in Summer Conditions A Comparison between Optimal and Model Predictive Controllers.pdf

1、 Jean-Michel Dussault is a PhD candidate in the Dpartement de gnie mcanique, Universit Laval, Qubec, Qubec, Canada. Maarten Sourbron is a Post doctoral research affiliate in the Dpartement de gnie mcanique, Universit Laval, Qubec, Qubec, Canada. Louis Gosselin is a professor in the Dpartement de gni

2、e mcanique, Universit Laval, Qubec, Qubec, Canada. Smart Windows Control Strategies for Building Energy Savings in Summer Conditions: A Comparison between Optimal and Model Predictive Controllers Jean-Michel Dussault, Eng. Maarten Sourbron, PhD Louis Gosselin, Eng., PhD Student Member ASHRAE Member

3、ASHRAE ABSTRACT Advanced control strategies for smart windows (SW) are discussed in this paper. Since smart windows are used both to reduce energy consumption and to improve thermal and visual comfort, the optimal solar flux passing throught the window is the result of a complex trade-off between da

4、ylighting and heat flow balance. A typical office building zone is modeled in TRNSYS with an integrated electrochromic smart window. Two types of advanced SW controllers, i.e. (i) a genetic algorithm based controller and (ii) a model predictive control based controller, are studied and compared to a

5、 base case scenario. The advanced controllers evaluate the hour-by-hour state of the smart window required to minimize the overall energy consumption (heating, cooling, lighting) while respecting constraints related to thermal and visual comfort. Results have shown that the two controllers, while pr

6、esenting different control strategies, offer very similar and promising results in terms of energy savings and peak load reductions. Finally, opportunities resulting from the present work are discussed. INTRODUCTION Smart windows (SW) (Jelle et al. 2012) retained the attention of many researchers ov

7、er the years since the early 90s. Research was initially oriented toward the development of potential technologies and the evaluation of the thermal and optical performance of idealized SW (Reilly et al. 1991). Then, simple control strategies applied to real SW technologies have been simulated and i

8、t was showed that SW could reduce the peak loads and the overall energy consumption compared to conventional “passive” glazings (Selkowitz et al. 1994). Later on, the notion of visual comfort was added in the control strategies in order to increase the market acceptance (Lee et al. 2006). Nowadays,

9、althought there are still many research projects on the development of materials offering enhanced performances for smart windows (Llordes et al. 2013), the focus on smart windows is moving toward a deeper understanding of the optimal use of the existing SW technologies. With the increasing amount o

10、f available data (current weather and/or weather forcasts) and the highly sofisticated building energy management systems available (Rocha et al. 2015), it is now possible to think about the optimal implementation of control systems for active faades such as SW. The main purpose of this paper is to

11、investigate the performance of two different SW controllers, i.e. a controller based on a genetic algorithm (GA) strategy and a controller based on a model predictive control (MPC) strategy. Since literature reveals that SW present higher energy savings in hot climates (Bahaj et al. 2008), this stud

12、y is focused on the analysis of these control strategies during summer conditions. BUILDING MODEL The building model considered in the present work represents a typical single zone office space developed in TRNSYS (TRANSSOLAR Energietechnik GmbH, 2012) with Type 56. All simulations were performed wi

13、th a time step of one hour. Since the focus of this work is about control strategies in summer conditions, the simulation period covered the months of June and July. While the hours simulated during the month of June were used as a warm up period as suggested by (Delcroix and Kummert, 2012), every h

14、ours of the month of July were used for the analysis presented in the next sections. Building Location, Geometry and Construction The building was located in the city of Montreal (Quebec), Canada, with EnergyPlus weather data for Montreal (.epw file) being used as the weather file. The building geom

15、etry consists of a 10 m (32.8 ft) 10 m (32.8 ft) 3 m (9.8 ft) (width (L) depth (P) height (H) building zone with a south facing exterior wall (see Figure 1). The exterior wall considers a 10 m (32.8 ft) wide (L) by 2 m (6.6 ft) high (Hsw) double insulated unit electrochromic SW (Jelle et al. 2012) o

16、n the upper part of the wall and an opaque wall (U-Value = 0.45 W/m2K (0.08 Btu/hr-ft2-F) at the botton (the wall construction being, from outside to inside: concrete siding, lightweight frame and gypsum indoor finishing). The thermal mass is accounted for in the envelope and in the floor (0.10 m (4

17、 in) concerete floor slab). All other surfaces were defined as adiabatic interior surfaces. For the sake of illustration, Table 1 provides the SW center of glazing properties at normal incidence. However, the model uses the complete and detailed properties varying with the angle of incidence obtaine

18、d in the IGDB. Figure 1 : Building zone dimensions Table 1. Smart Window Center-of-Glazing Properties Smart window states U-Value SHGC Tvis Tsol W/m2K Btu/h-ft2-F - % % State 1 (S1) (bleached) 1.63 0.287 0.47 62.1 38.1 State 2 (S2) 1.63 0.287 0.17 21.2 8.6 State 3 (S3) 1.63 0.287 0.11 5.9 2.4 State

19、4 (S4) (fully tinted) 1.63 0.287 0.09 1.5 1.0 Gains and Schedules Internal gains include artificial lighting, people and equipment. Table 2 presents the building zone heat gains with their radiative and convective fractions. Only sensible heat has been considered in the model. Table 2. Building Zone

20、 Heat Gains Gain Types Max. Heat Gains Convective Fraction Radiative Fraction W Btu/hr % % Occupants (10) 730 2491 30 70 Equipment 800 2730 30 70 Light 352 1201 41 59 As presented in Table 2, the zone could accept up to 10 occupants (73W/occupant (249 Btu/hr/occupant) and contains a power density of

21、 8 W/m2 (2.54 Btu/hr-ft2) (floor area) for office equipment. The building lighting model calculates the illuminance distribution on interior surfaces of the zone considering daylight and artificial light from the lighting system (eight T8 lamps of 55 W (188 Btu/hr) each, i.e.: 440 W (1501 Btu/hr) to

22、tals). The artificial lighting system considers a 20 % heat to return. Illuminance levels from daylight were calculated using Daysim simulation software (Reinhart and Breton, 2009). On the other hand, illuminance levels from the artificial lighting system were directly calculated from Radiance (Ward

23、, 1994) and considered lamps uniformly distributed over the ceiling area. Since this paper was designed to compare SW control strategies, a simplified representation of the lighting system was considered, i.e.: a linear relation between lighting and power (no ballast factor and standby power loss co

24、nsidered). In order to offer proper luminosity on the work plane, a sensor has been set in the middle of the rooms width and at two thirds of the rooms depth (from the glazed wall). The minimal required luminosity at the sensor is labeled WPreq and has been set to 500 lux (46.5 fc) during occupancy

25、hours (Dubois and Blomsterberg, 2011). The installed artificial lighting system is assumed to provide only the necessary amount of artificial light to meet the illuminance requirements during occupancy hours (dimmable system to prevent overlighting). Occupancy and work plane lighting requirements ar

26、e presented in Figure 2. Figure 2 : Week days schedule for occupancy (left axis) and work plane lighting requirement (right axis) HVAC&R System The HVAC&R system is simplified to a convective heating load Qheat and a cooling load Qcool that act directly on the air node of the building model. The ind

27、oor minimal and maximal temperature set points are 21C (69.8 F) and 25C (77 F) which are respectively lowered and raised by 3C (5.4F) outside the occupation hours. The system sizing considered the maximum heating and cooling loads for the building with the smart window in its clearer state (33 W/m2

28、(10.5 Btu/hr-ft2) and 83 W/m2 (26.3 Btu/hr-ft2) (floor area), respectively). A constant air volume ventilation system was installed. The volume flow rate was roughly estimated at 340 m3/h (200 cfm). A heat exchanger was used to transfer heat between the air exhaust and make up air with 60 % efficien

29、cy. SMART WINDOW CONTROL STRATEGIES Base Case Scenario The SW base case controller consideders a SW set at its clearest state (S1) at all time. This passive approach is used to define the reference energy consumption of the building zone. It represents the energy consumption of an office building zo

30、ne whose envelope is composed of a typical passive low-E window (low-E on surface #2). No shading system is installed in the base case. The control strategies presented in the following subsections are all compared to this base case scenario in order to assess the benefits of advanced control strate

31、gies. Genetic Algorithm The purpose of the genetic algorithm controller (optimal control) is to establish the possible minimum overvall energy consumption of the zone (heating, cooling and artificial lighting) (Dussault and Gosselin, 2014). Results of other control strategies, such as MPC, could the

32、n be compared to the results achieved with the GA in order to assess their performance. To achieve such optimal control for active SW, a controller based on a genetic algorithm (Gosselin et al. 2009) assuming a perfect building model representation and perfect weather forecast (i.e. one perfectly kn

33、ows the future weather parameters) have been used. The algorithm minimizes an objective function (in this case, the overall energy consumption) by evaluating a certain number (population) of different combinations (phenotypes) of the design variables (SW state at each time step). The initial populat

34、ion evolves generation by generation by keeping the phenotypes of a generation that give the best results and by creating the following generation from those phenotypes and newly created ones (children) by crossovers and mutations. The design variables involved are thus the SW states at each simulat

35、ed hour where sunlight is available. To minimize computational time, hourly artificial lighting variables have not been considered as design variables, but rather as values adjusted to meet requirements depending on the SW state, the lighting control strategy and the lighting set point. The objectiv

36、e function to minimize is the overall energy consumption (QTot), in Wh/m2 (Btu/ft2) of floor area, defined as: (1) where QHeat is the total energy consumed for heating in Wh/m2 (Btu/ft2), QCool is the total heat removal required for cooling in Wh/m2 (Btu/ft2), COP is the coefficient performance of t

37、he cooling system (COP = 3 in this case) and QLight is the total lighting energy consumption in Wh/m2 (Btu/ft2). Constraints of minimal light requirements were defined. Optimization parameters and convergence criteria used for the optimization runs are presented in Table 3. It is important to unders

38、tand that the GA optimization is not intended for real-time control of SW since the required computational time is too long for such applications. It serves here only to establish the optimal achievable level of performance with an “optimal control” against which real-time strategies could be compar

39、ed. Table 3. Parameters of the Genetic Algorithm Parameters Value Units Number of phenotypes per generation 40 - Maximum number of generations 75 - Number of generations with unchanged QTot value before convergence 30 - Proportion of children per generation 80 % Children mutation probability 5 % Num

40、ber of chromosomal crossovers 2 - Model Predictive Control (MPC) In the MPC approach (see Figure 3), a linearized internal model (based on a simplified space-state resistance-capacitance thermal model of the building developed in Matlab (Dussault et al. 2012) runs as a closed-loop observer of the bu

41、ilding states ( ) in parallel to the “real” process (TRNSYS model). These states are then used to calculate the predictions of the building outputs T over the prediction horizon Hp. These predictions are then optimized considering the possible design variables (SW states) in order to minimize Qtot (

42、eq. 1) over the entire prediction horizon. Uc represents the inputs to supply to the building energy management system (heating/cooling loads, lighting load and SW state). The optimization environement used in this work was YALMIP with the GUROBI Solver. The optimization constraints were the minimum

43、 lighting requirements (hard constraints), the HVAC&R system maximal capacities (hard constraints) and the temperature setpoints (soft constraints). The MPC controller includes the observer, the predictor and the optimization procedure. In this work, the prediction horizon considered in the MPC cont

44、roller was 24 hours. Figure 3: Model based predictive control architecture Besides the optimization solvers, the MPC controller is different from the GA controller by the fact that it contains an internal model that is different from the “real” process and introduces model errors in the predictions.

45、 For this reason, the states of the internal model are updated at each time step with the measured outputs of the TRNSYS model in order to maintain convergence. As presented in Figure 3, the architecture of the MPC controller is developed for real-time control applications. However, as for the GA co

46、ntroller, prefect weather forecast is used in the MPC controller in order to assess its maximal performances. RESULTS A comparison of the base case scenario, the GA and the MPC controllers is illustrated in the following subsections on a monthly and an hourly basis. From the results, one could clear

47、ly see that the two advanced controllers show considerable savings. Monthly Results Table 4 presents results of energy consumption (cooling, lighting and total) as well as the maximum peak load for the month of July. Fan power was not considered in the results analysis presented in this section. Tab

48、le 4 illustrates the fact that GA and MPC controllers reduce the overall energy consumption by 33% and 32 % respectively, compared to the base case scenario. For both the GA and the MPC controllers, this overall energy reduction is the result of a trade-off between cooling energy consumption (36% an

49、d 35% decrease, respectively, compared to the base case) and lighting energy consumption (approximatively 0.22 kWh/m2 (70 Btu/ft2) increase in both cases). Total energy savings are mainly explained by the fact that the darker states of the SW reduce dramatically the solar heat gains entering the building while forcing the artificial lighting system to operate a little more to respect illuminance requirements. Table 4 also illustrates cooling peak load reductions of 26% and 25% for the GA and MPC, respectively,