Implementation of an adaptive occupancy and building learning temperature setback algorithm.
The choice of temperature setback schedule is one of the most influential decisions made by operators. It plays a substantial role over space heating and cooling loads and occupant comfort. If the nighttime setback period is set too long, occupants will more likely be present before the setpoints are met; if the setback period is set too short, there will be excessive energy use as a result of conditioning the space unnecessarily. In fact, both extremes are not uncommon in commercial buildings (Bordass et al. 2001a; Motegi et al. 2003).
An operator's decision for the temperature setback schedule is often restrained by two factors.
1. Diversity of a Building's Usage: Because more than a quarter of the North American workforce reported having flexible work hours (Golden 2001; McMenamin 2007; Zeytinoglu et al. 2009), job-specific restrictions and personal preferences translate into occupancy profiles. Consequently, in office buildings, the peak occupancy rates reduce and the occupancy spreads well before 09:00 and well after 17:00, as shown in Figure 1. In some cases, occupancy may even extend to weekends and holidays (Sunet al. 2014). Operators tend to choose conservatively long operating schedules in order to accommodate this diversity in building usage (Bordass et al. 2001b). As a result, energy use for heating and cooling during unoccupied hours can even exceed the occupied hours (Masoso and Grobler 2010).
2. Uncertainty in the Length of Nighttime Setback-to-setpoint Transition Time Period: The time needed to bring the indoor temperature from nighttime setback to the setpoint temperature depends on a number of factors including, but not limited to, (a) the characteristics of the terminal heating and cooling unit, (b) the thermal capacitance of the zone's fabric, (c) the thermal resistance and airtightness of the building envelope, (d) ambient environmental conditions (temperature and solar irradiation), (e) electrical appliance and lighting loads, and (f) shading position and window opening. These factors not only change in time (seasonally and daily) but also they can vary significantly between thermal zones.
Although optimal start algorithms--which dynamically determine the setpoint transition times--have been commercially available, they often require manual tuning and setup (e.g., operators may need to guess the amount of thermal mass) (Salsbury 2009) and they often employ inaccurate simplistic methods (Salsbury 2005). Consequently, operators need to conservatively guess a fixed setpoint transition duration, so that all offices can reach the setpoint temperature from the nighttime setback before occupants start arriving. Figure 1 illustrates how this decision can increase the effective operating hours in an office building.
In recognition of the challenges in employing the optimal temperature setback schedule in an office building, this paper puts forward an autonomous algorithm that learns the recurring occupancy patterns in tandem with the parameters of a simple model that describes the heat transfer process in a thermal zone from a small number of low-cost sensors. The algorithm was implemented in a local building controller actuating a radiant ceiling panel heater valve and the damper of a variable air volume (VAV) terminal unit serving a southwest-facing office space in Ottawa, Canada. Through this implementation, the algorithm's ability to identify parameters describing the occupancy patterns and the heat transfer process was assessed.
[FIGURE 1 OMITTED]
The algorithm was then implemented in the energy management system (EMS) application of the building performance simulation (BPS) tool EnergyPlus to assess annual energy and comfort performance. It was employed to dynamically adapt the temperature setback schedule for a terminal heating and cooling unit serving an office space modeled in EnergyPlus. Potential reductions in space heating and cooling loads were analyzed with respect to cases whereby different fixed setback schedules were selected by an operator.
This section first presents the controller algorithm and its following aspects: (a) how it learns the recurring occupancy patterns, (b) how it learns the parameters of the model describing the heat transfer process in a thermal zone, and (c) how it employs the recursively learned knowledge to apply dynamically changing setback schedules. Then, the characteristics of the office space where the controller algorithm was employed are presented. The results of this implementation were used to assess the predictive accuracy of the controller algorithm. Subsequently, the control algorithm was implemented in a BPS model of the monitored office to study the impacts of the algorithm in space heating and cooling loads. This section also presents the characteristics of this BPS model.
The controller algorithm has two recursive learning tasks: (a) to identify the recurring patterns of occupancy and (b) the parameters of a model describing the heat transfer process in a thermal zone. In order to execute these learning tasks, the algorithm inputs seven variables: the VAV terminal unit's discharge air pressure and temperature, indoor and outdoor temperatures, the heating unit's supply hot-water valve state, the readings from a light intensity sensor, and the movement detections from the motion sensor. By employing the recursively learned knowledge of the occupancy patterns and thermal zone's heat transfer characteristics, the controller algorithm outputs whether or not the temperature setback should be applied. The general structure of the controller algorithm is shown in Figure 2.
The Occupancy Learning Algorithm: Figure 3 presents the occupancy learning algorithm. The algorithm updates fives parameters describing the occupancy characteristics of a thermal zone based on three different types of occupancy events. The first movement detection of a day is an arrival event. The time of an arrival event is used to update the mean and the variance of the arrival times, as shown in Figure 3. The last movement detection of a day is a departure event. The time of a departure event is used to update the mean and the variance of the departure times, as shown in Figure 3. If no motion sensor detections are observed on a weekday, this is an event of absence. An event of absence is used to update the probability of having a weekday without occupancy.
The Building Learning Algorithm: Figure 4 presents a simple graybox model that describes the rate of change in indoor temperature in a perimeter office space. The model assumes that the rate of change in measured indoor temperature can be explained with the following five different loading conditions: (1) envelope losses and infiltration; (2) solar gains and electric lighting; (3) internal gains, plug loads, and heat losses due to occupant's use of doors; (4) heat output of the terminal heating unit (e.g., radiant panel heaters); and (5) heat output of the terminal ventilation/cooling units (e.g., VAV).
[FIGURE 2 OMITTED]
[FIGURE 3 OMITTED]
The model maps the five loading conditions onto the seven available sensors. The role of envelope and infiltration losses over the rate of change in the measured indoor temperature ([u.sub.1] x [x.sub.2] in Figure 4) was taken as a linear function of the difference between indoor and outdoor temperatures. The role of solar gains and electric lighting over the rate of change in the indoor temperature ([u.sub.2] x [x.sub.3] in Figure 4) was assumed as a linear function of the light intensity measurements from a photodiode on the ceiling between electric lighting and windows. Note that it is not argued that a photodiode sensor's readings can accurately represent the solar gains. We rather hypothesize that the rate of change in the indoor temperature due solar gains and electric lighting can be approximated as proportional to the photodiode sensor readings. Occupants, when present, were assumed to be associated with a recurring pattern of activities (e.g., keeping the door open, use of electric appliances, heat emitted during their recurring activities) resulting in a drift from the rate of change in temperature (see [x.sub.4] in Figure 4). The occupants' presence (see [u.sub.3] in Figure 4) was inferred from their movements detected by a passive-infrared (PIR) motion sensor. In a time frame of 30 minutes, if a movement was detected, the room was assumed occupied. This time delay value was selected in line with prior research (Page et al. 2008; Dong et al. 2010; O'Brien and Gunay 2014; Gunay et al. 2015). Note that here it is not argued that a motion sensor's readings can represent the occupancy-driven loads. We rather hypothesize that the occupancy-driven loads depend on occupants' presence and introduce an unknown time-varying bias to the natural temperature response of the office.
The role of the terminal heating system (e.g., radiant panel heaters) over the rate of change in the indoor temperature ([u.sub.4] x [x.sub.5] in Figure 4) was assumed linearly dependent on its actuation level [0,1] ([u.sub.4] in Figure 4). The role of the terminal ventilation/cooling system (e.g., VAV unit) over the rate of change in the indoor temperature was defined as [u.sub.5] x [x.sub.6] (see Figure 4). The model input [u.sub.5] was defined as the square root of the discharge air pressure ([square root of ([p.sub.da])] multiplied by the difference between the discharge ([T.sub.da]) and indoor air temperatures. Note the airflow rate based on the Bernoulli equation is linearly proportional to [square root of ([p.sub.da].
[FIGURE 4 OMITTED]
It is worth noting that, in a data-driven graybox model, one should not attribute explicit physical meanings to the nonphysical parameter weights (which also evolve in time). Whether or not these simplifications are reasonable or not can be tested by looking at the model's predictive accuracy.
In recognition of the uncertainties in model inputs, the unmodeled inputs and other modeling approximations, the heat transfer process defined in model (Figure 4) was transformed into a stochastic differential equation as follows:
d[x.sub.1] = ([u.sub.1][x.sub.2] + [u.sub.2][x.sub.3] + [u.sub.3][x.sub.4] + [u.sub.4][x.sub.5] + [u.sub.5][x.sub.6])dt + dw y = [x.sub.1] + e (1)
where w is a Wiener process, e is a Gaussian error in taking indoor temperature measurements, and y is the indoor air temperature measurements ([x.sub.1] in Figure 4). The measurable state ([x.sub.1] in Figure 4) and the parameters ([x.sub.2:6] in Figure 4) were recursively estimated using the extended Kalman filter (EKF). In brief, the EKF algorithm, which includes the parameters in an augmented matrix, makes a 15-minute-ahead prediction for the indoor temperature [x.sub.1]. When this time elapses, the prediction is compared with the indoor temperature measurement. This comparison, in tandem with the dynamics of the model, is used to improve knowledge about the unknown parameters ([x.sub.2] to [x.sub.6] in model 1). As a result, the model predictions ameliorate iteratively in time. The reader can refer to Grewal (2008) for further details about the EKF algorithm.
The Adapting Agent Algorithm: The adapting agent recursively inputs the updated parameters describing the occupancy patterns ([[micro].sub.aar], [[micro].sub.dpt], [[sigma].sub.aar], [[sigma].sub.dpt], [p.sub.abs]) and the parameters of the model describing the heat transfer process in the thermal zone ([x.sub.1] to [x.sub.6]) and outputs whether or not the thermal zone should be at temperature setback. On weekdays prior to the arrival of the occupants, the setback is terminated when the current time exceeds ([[micro].sub.aar] - 1.5(1 - [P.sub.abs]) [[sigma].sub.aar]) + [[DELTA]t.sub.warm-up]. The term [[DELTA]t.sub.warm-up] represents the time period needed by the terminal heating/cooling system to bring the indoor temperature from the nighttime setback to the setpoint temperature prior to 85% of the arrivals ([[micro].sub.aar] - 1.5(1 - [p.sub.abs]) [[sigma].sub.aar]). The adapting agent computes [[DELTA]t.sub.warm-up] as follows:
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (2)
where SP is the (heating or cooling) setpoint. Note that the [u.sub.3] is not included in Equation 2 because the office is unoccupied during the nighttime setback-to-setpoint transition period. On weekdays, if the time of day exceeds ([[micro].sub.aar] - 1.5(1 - [p.sub.abs]) [[sigma].sub.aar]) and an occupant has not arrived, meaning that there is less than 15% chance to observe occupancy for rest of the day, the temperature setback is reinstated. To ensure that occupants would find their offices thermally comfortable upon a long intermediate break (e.g., lunch break or an afternoon meeting), on occupied weekdays the offices remain conditioned between ([[micro].sub.aar] - 1.5(1 - [p.sub.abs])[[sigma].sub.aar]) and ([[micro].sub.dpt] + 1.5[[sigma].sub.dpt]). In the unlikely event that an occupant arrives during the temperature setback period, the setback is terminated.
[FIGURE 5 OMITTED]
Characteristics of the Office Space
The control algorithm was implemented in a southwest-facing shared corner office space in an academic building in Ottawa, Canada. Based on the results of this implementation, the control algorithm's ability in learning the parameters describing the recurring occupancy patterns and the parameters of the model describing the heat transfer process was assessed. Accuracy of the predictions made upon these recursively learned parameters was evaluated.
Figure 5 illustrates the main characteristics of this office space. The floor area of the office is 16 [m.sup.2] (172 [ft.sup.2]). The facade area is 28 [m.sup.2] (301 [ft.sup.2]); 16 [m.sup.2] (172 [ft.sup.2]) faces southwest and the other 12 [m.sup.2] (129 [ft.sup.2]) faces southeast. On the southwest facade, the area of the exterior window is 8 [m.sup.2] (86 [ft.sup.2]) (i.e., window-to-wall ratio is 0.5 on the southwest facade).
The steel-framed exterior wall assembly was designed to RSI 4.0 [m.sup.2o]Cx[W.sup.-1] (R 22 h[ft.sup.2]x[degrees]F/Btu) in compliance with ASHRAE (2013) for Climate Zone 6. The exterior window assembly consists of clear double-glazed windows and aluminum frames with thermal breaks--USI 3.1 Wx[m.sup.-2][degrees][C.sup.-1], (U 0.5 Btu/[h[degrees]F-[ft.sup.2]]), solar heat gain coefficient (SHGC) 0.4, and visible transmittance 0.6.
The office space was heated through the radiant ceiling panel heaters. The office was ventilated and/or cooled through the VAV terminal unit serving this office only. The light fixture included six recessed 32W (110 Btu/h) T8 florescent lamps.
The office space was intermittently used by four individuals. The occupants were able to control the motorized roller blinds and electric lighting. They were also able to specify their preferred indoor air temperature. Occupants could interact with these systems (e.g., lights on/off, blinds open/closed, increase/decrease temperature setpoint) through a wall-mounted control interface with four programmable buttons. The control algorithm was responsible for maintaining occupants' preferred indoor temperature during occupied periods. When the algorithm employs temperature setback based on the logic presented in the section title "Controller Algorithm," the heating setpoint was set back to 18[degrees]C (64.4[degrees]F) and the cooling setpoint was set back to 27[degrees]C (80.6[degrees]F).
The sensor locations are also noted in Figure 5. The office space was equipped with two PIR motion sensors and two thermistor-based temperature sensors. Two photodiode light intensity sensors were mounted on the ceiling. The temperature and the dynamic pressure of supply air were also monitored on the downstream side of the VAV unit's damper.
The BPS Model of the Monitored Office
In an effort to understand the space heating and cooling load implications of the control algorithm in contrast to a wide range of conventional fixed setback schedules, the BPS model of the monitored office was built in EnergyPlus v8.1.
The geometry and orientation of the office were set as shown in Figure 5. The two interior walls were assumed adiabatic, while the exterior surfaces were subject to the ambient conditions defined in the Canadian weather year for energy calculation for Ottawa. Two different floor assemblies were studied: 100 and 200 mm (4 and 8 in.) thick concrete slabs. The floor was covered by a carpet. The ceiling slab was covered with a plenum space that was enclosed by ceiling tiles.
The wall assembly was modeled with RSI-4.0 [m.sup.2]x[degrees]Cx[W.sup.-1] (R-22 h[ft.sup.2]x[degrees]F/Btu) insulation. The exterior window assembly was modeled as USI 3.1 W.[m.sup.-2][degrees]x[C.sup.-1] (U-0.5 Btu/[h[degrees]Fx[ft.sup.2]]), SHGC 0.4, and visible transmittance 0.6. When lights were switched on, they were assumed to consume 12 W/[m.sup.2] (3.8 Btu/[hx[ft.sup.2]]) (with a radiant fraction of 0.37 [IES 2010]). The infiltration rate was taken as 0.15 ach. The equipment power density was taken as 5.4 Wx[m.sup.-2] (1.7 Btu/[hx[ft.sup.2]]) (with a radiant fraction of 0.3) during the occupied hours (ASHRAE 2013). The manually controlled motorized roller blinds were assumed to have a transmittance of 0.05 and reflectance of 0.75.
The ventilation and cooling were delivered through a VAV unit with a chilled-water cooling coil (without a reheat coil). The cooling coil's temperature setpoint was 12.8[degrees]C (55[degrees]F). During the operating hours, outdoor air was introduced into the office at 40 L/s (85 cfm) (ASHRAE 2013). When outdoor air temperature was advantageous for sensible cooling, an economizer was set to increase the outdoor airflow rate. The maximum airflow rate was set at 100 L/s (212 cfm or 7.5 ach). The heating was supplied by a 1.5 kW (5100 Btu/h) electric radiant panel heater. Evidently, this does not represent the system efficiency of the radiant ceiling panel heater system served by a central boiler plant, but note that the BPS model was intended to compute the heating and cooling loads, not the HVAC equipment energy use. The heating and cooling systems were explicitly modeled solely to represent the temperature response of the building model.
By employing the observational dataset emerging from the implementation, the occupants' thermostat, lighting, and blinds use were modeled as discrete-time Markov logistic regression models. Further details on modeling and simulation of occupant behaviors as discrete-time Markov processes can be found elsewhere (Parys et al. 2011; Gunay et al. 2014b). The stochastic thermostat use model predicts the likelihood of increasing or decreasing the temperature setpoints in the next 30 minutes by looking at the indoor temperature (see Figure 6a). The lighting use model predicts the likelihood of a light switch-on action in the next five minutes by looking at the light intensity sensor readings on the ceiling (see Figure 6b). The blinds use model predicts the likelihood of opening or closing blinds by inputting the light intensity sensor readings (see Figure 6c). Similarly, drawing upon the observational dataset emerging from the implementation, occupants' arrivals and departures were modeled as a discrete-time Markov Gaussian mixture model. The occupancy model predicts the likelihood of observing an arrival (including the intermediate arrivals) or a departure (including the intermediate departures) event in the next 30 minutes by inputting the time of day as the predictor (see Figure 6d).
The control algorithm was implemented in the EMS application of the BPS tool EnergyPlus to decide whether or not the setback should be applied. During the setback period, the heating setpoint was 18[degrees]C (64.4[degrees]F) and the cooling setpoint was 27[degrees]C (80.6[degrees]F). Otherwise, the temperature setpoints specified by the simulated occupants were maintained. It is worth to recall that the daytime setpoints are determined by the simulated occupants (defined by the stochastic thermostat use models shown in Figure 6).
RESULTS AND DISCUSSION
The first arrival and the last departure of a day in the monitored office were shown in Figure 7. Results indicate that arrivals tend to occur between 07:30 and 10:00, whereas departure events were scattered between 09:00 and 19:00. Figure 7 also presents the evolution of the recursively learned arrival and departure times. Because it was hard to distinguish an intermediate vacancy period (e.g., a lunch break) from an early departure, conservatively, only departures after 16:00 were used to update the mean and standard deviation of the departure time. The parameter estimates defining recurring occupancy patterns converged to stable and meaningful (i.e., consistent with the individual observations) values in less than 15 days.
[FIGURE 6 OMITTED]
[FIGURE 7 OMITTED]
The ability of the building learning algorithm to learn its parameters in recursion and to make predictions for the indoor air temperature were assessed subject to the loading conditions presented in Figure 8.
When the building learning algorithm is trained with the inputs shown in Figure 8, the parameter estimates converged to physically meaningful and stable values within two weeks (see Figure 9). For example, if the outdoor temperature is 40[degrees]C (72[degrees]F) colder than indoors, the model suggests that the envelope losses alone will cause a temperature drop of about 0.07[degrees]C (0.13[degrees]F) in 30 minutes ([u.sub.1] x [x.sub.2][DELTA]t). If the light intensity sensor reading is 100 fc (1076 lux), the model predicts that the solar gains and electric lighting will cause a temperature increase about 0.60[degrees]C (1.08[degrees]F) in 30 minutes ([u.sub.2] x [x.sub.3][DELTA]t). If the occupants are present, the model indicates that the occupant-driven activities (e.g., losses due to door position or gains from plug-in appliance loads) will likely cause a minor temperature decrease about 0.010[degrees]C (0.018[degrees]F) in 30 minutes ([x.sub.4][DELTA]t). When the radiant panels turn on, the model suggests that this will cause a temperature increase of 0.60[degrees]C (1.08[degrees]F) in 30 minutes ([x.sub.5][DELTA]t). If the discharge air temperature is 10[degrees]C (18[degrees]F) colder than the indoor air temperature and the discharge air pressure is at 75 Pa (0.3 in.w.g.) (i.e., the damper is open), the model suggests that this would reduce the indoor temperature by about 1.75[degrees]C (3.15[degrees]F) in 30 minutes (([T.sub.da] - [x.sub.1])[square root of ([p.sub.da]) [x.sub.6][DELTA]t).
To assess the predictive accuracy of the model, the parameter values on March 14,15, and 16, as annotated in Figure 9, were used to make offline predictions for the indoor temperature over two-day time horizons. The model predictions were in agreement with the timing of the local extrema and the inflection points (see Figure 10). The mean absolute errors were less than 0.7[degrees]C (1.2[degrees]F) (see Figure 10). Note that the number of model inputs and complexity were determined through an independent sensitivity analysis, and the current model was found as the simplest feasible model that can make accurate offline predictions over at least a 24 h time horizon.
Because we were trying to isolate the model-induced errors from the uncertainties inherent in predicting the model inputs over the prediction time horizon (e.g., what will the light intensity sensor readings be in the next two days?), the loading conditions were assumed to be known over the prediction time horizon. In recognition of the uncertainties in retrieving weather and occupant-driven model inputs, the prediction time horizon in zone level control applications should be far less than two days (Gunay et al. 2014b). In this control algorithm, the adapting agent was only permitted to compute the nighttime setback-to-setpoint temperature transition time [DELTA][t.sub.warm-up] after 05:00 and before the estimated time of arrival ([[micro].sub.arr] - 1.5 (1 - [p.sub.abs])[[sigma].sub.arr]). Typically, this results in prediction time horizons shorter than three hours. As the prediction time recedes from 05:00 to the estimated time of arrival ([[micro].sub.arr] 1.5(1 - [p.sub.abs])[[sigma].sub.arr]), the prediction time horizon becomes shorter and shorter. The adapting agent, by assuming inputs remain the same over this short period of time, keeps making predictions to decide whether or not to end the temperature setback.
The terminal heating and cooling systems serving the BPS model presented in the section titled "The BPS Model of the Monitored Office" were employed to contrast comfort and annual heating/cooling load implications of various static setback strategies with the adaptive setback periods selected by the control algorithm.
To this end, 100 different weekday temperature setback schedules were randomly generated from a uniform distribution. They were static, meaning that they start and end at the same time on each weekday for each of the 100 annual simulations. The beginning time of these temperature setback periods ranges from 16:00 to 22:00 and the ending time ranges from 05:00 to 10:00. These static setback periods represent a wide range of temperature setback schedule options that can be selected by different operators.
The comfort of the simulated occupants was analyzed from their interactions with the thermostats. If the simulated occupants experience thermal discomfort too frequently, they would need to adjust their thermostat settings frequently as well. A similar causal relationship between a simulated occupant's comfort and its adaptive behavior patterns were used elsewhere (Gunay et al. 2014c). Figure 11a presents the number of thermostat interactions observed at each of the 100 static weekday setback schedules. When the setback was employed during weekends only, the setback would cover about 30% of the year. If the setback is employed on weekends and on weekdays before 06:00 and after 22:00, the setback periods would cover about 50% of the year. If the setback is employed on weekends and on weekdays before 08:00 and 20:00, the setback periods would be enough to cover about 60% of the year. Results indicate that simulated occupants tend to interact with their thermostat (i.e., they are uncomfortable and they will remain uncomfortable until the indoor temperature reaches their preferred setpoint), when the setback schedules cover more than 55% of the year. In other words, even the most diligent operator after analyzing the office occupancy patterns and the nighttime setback-to-setpoint transition periods would choose a setback schedule that covers less than 55% of the year. Clearly, this value is peculiar to the studied office space. In reality, it depends on the characteristics of the occupancy, building, and HVAC system, which are rarely available to the operator.
In contrast, the control algorithm, by dynamically adapting the time temperature setbacks start and end on weekdays, maintained the temperature setback for more than 70% of the year without affecting the occupant comfort. This can be interpreted that the algorithm was able to recursively identify some of the redundancies in the operating hours and include them to the setback schedules. Instead of choosing a static setback schedule that covers 55% of the year, this resulted in 20% reduction from the annual cooling loads and 10% reduction from the annual heating loads (see Figure 11b). When the simulations were repeated by employing the BPS model with 200 mm thick concrete slab, the cooling load reductions decreased to 15% and the heating load reductions decreased to 8% (see Figure 12). This can be interpreted that as the thermal capacitance increases, the effectiveness of the nighttime setback strategies slightly diminishes.
[FIGURE 8 OMITTED]
[FIGURE 9 OMITTED]
The results shown in Figure 11b and 12b also underlines the sensitivity of heating and cooling loads to the temperature setback choices. The fact that the occupancy in this office covered only about 15% of the year underlines the potential for improvement with other learning algorithms and/or more aggressive setback strategies (e.g., assuming the estimated time of arrival ([[micro].sub.arr] - (1 - [p.sub.abs]) [[sigma].sub.arr]) instead of ([[micro].sub.arr] - 1.5(1 -- [P.sub.abs])[[sigma].sub.arr]) or applying a 4[degrees]C setback instead of 3[degrees]C [7.2[degrees]F instead of 5.4[degrees]F]). Future work should investigate these options.
[FIGURE 10 OMITTED]
Employing a traditional setback schedule in an office building tends to create abrupt changes in space heating and cooling loads during the setpoint transition times, because all subspaces experience this transition simultaneously (Masoso and Grobler 2010). A potential benefit of employing recursively learned individual setback schedules in each thermal zone can be diversifying the heating and cooling load profiles. This may help reducing the peak loads. Future work is planned to study this aspect through a field implementation with 16 identical offices (eight with the adaptive setback schedules and eight with traditional static setback schedules).
Unlike the monitored office in this paper, the terminal heating and cooling units often serve more than one room. This would reduce the uniqueness of the occupancy profiles and the nighttime setback-to-setpoint transition periods in each thermal zone and reduce the effectiveness of the control algorithm. Future work should study the performance of the control algorithm where the terminal systems serve multiple office spaces. However, because a thermal zone would always be a subset of a large office building (which consists of multiple thermal zones), the occupancy, building, and HVAC system characteristics of these thermal zones would be relatively individual and the control algorithm should still be able to reduce the space heating and cooling loads.
[FIGURE 11 OMITTED]
[FIGURE 12 OMITTED]
In this study, occupants' presence was inferred through their movement within a 30-minute time frame that was detected by the PIR motion sensors. The 30-minute time frame was selected based on that even a 15-minute time frame can be inadequate for capturing at least one occupant movement (O'Brien and Gunay 2014). However, this assumption might have caused false interpretations, when the occupants were immobile for 30 minutes and when the occupants were present less than 30 minutes. Recently, Nagy et al. (2015) showed that the optimal motion sensor delay time is unique in each office space and depends on a number of factors such as sensor position, office layout, and types of occupant activities.
In this study, recursively identified occupancy, building, and HVAC characteristics were used autonomously to decide when temperatures can be set back, leaving the operators outside the decision-making process. In a recent case study Fedoruk et al. (2015) reported that if the archived operational data can be made accessible, interpretable, and actionable, the operators can identify the operational faults and redundancies. Therefore, alternatively, the learned occupancy, building, and HVAC information could be made available to the operators, so that they can undertake informed decisions while selecting the operating hours.
A control algorithm that learns recurring occupancy patterns in tandem with the parameters of a simple model that describes the heat transfer process in a thermal zone from a small number of low-cost sensors was implemented in a southwest facing shared office space in Ottawa, Canada. Results from this implementation indicate that the parameters describing the occupancy, building, and terminal HVAC system characteristics converge to stable and physically meaningful values in less than two weeks. The indoor temperature predictions made upon these parameters resulted in less than 0.7[degrees]C (1.2[degrees]F) mean absolute error over a two-day prediction horizon.
The algorithm was also implemented in the EMS application of the BPS tool EnergyPlus to dynamically adapt the temperature setback schedules of a BPS model of the monitored office. The comfort and heating/cooling load implications of 100 static setback strategies were compared with the adaptive setback periods selected by the control algorithm. Results indicate that, with this control algorithm, the frequency of occupant overrides to the thermostat setpoints was the same as it was with a weekday setback schedule that covers 55% of the year with 15 to 20% lower cooling and 8 to 10% lower heating loads.
A potential benefit of employing recursively learned individual setback schedules in each thermal zone can be diversifying the heating and cooling load profiles. This may help reduce peak loads. Future work is planned to study this aspect through a larger scale field implementation.
This research is supported by a research funding provided by the Natural Sciences and Engineering Research Council (NSERC) of Canada, Delta Controls, ASHRAE, and Regulvar.
ASHRAE. 2013. Standard 90.1-2004, Energy standard for buildings except low-rise residential buildings. Atlanta: ASHRAE.
Bordass, B., R. Cohen, M. Standeven, and A. Leaman 2001a. Assessing building performance in use 3: Energy performance of the Probe buildings. Building Research & Information 29(2): 114-28.
Bordass, B., A. Leaman, and P. Ruyssevelt. 2001b. Assessing building performance in use 5: Conclusions and implications. Building Research & Information 29(2): 144-57.
Davis III, J.A. and D.W. Nutter. 2010. Occupancy diversity factors for common university building types. Energy and Buildings 42(9): 1543-51.
Dong, B., B. Andrews, K.P. Lam, M. Hoynck, R. Zhang, Y.S. Chiou, and D. Benitez. 2010. An information technology enabled sustainability test-bed (ITEST) for occupancy detection through an environmental sensing network. Energy and Buildings 42(7): 1038-46.
Fedoruk, L.E., R.J. Cole, J.B. Robinson, and A. Cayuela. 2015. Learning from failure: Understanding the anticipated-achieved building energy performance gap. Building Research & Information 43(6).
Golden, L. 2001. Flexible work schedules: Which workers get them? American Behavioral Scientist 44(7): 1157-78.
Grewal, M.S. 2008. Kalman filtering: Theory and practice using MATLAB. Hoboken, NJ: John Wiley and Sons.
Gunay, H.B., J. Bursill, B. Huchuk, W. O'Brien, and I. Beausoleil-Morrison. 2014a. Shortest-prediction-horizon model-based predictive control for individual offices. Building and Environment 82(0): 408-19.
Gunay, H. B., W. O'Brien, I. Beausoleil-Morrison, R. Goldstein, S. Breslav, and A. Khan. 2014b. Coupling stochastic occupant models to building performance simulation using the discrete event system specification formalism. Journal of Building Performance Simulation 7(6): 457-78.
Gunay, H. B., W. O'Brien, I. Beausoleil-Morrison and B. Huchuk 2014c. On adaptive occupant-learning window blind and lighting controls. Building Research & Information 42(6): 739-56.
Gunay, H. B., W. O'Brien, and I. Beausoleil-Morrison. 2015. Development of an occupancy learning algorithm for terminal heating and cooling units. Building and Environment 93: 71-85.
IES. 2010. The lighting handbook. 10th ed. New York: Illuminating Engineering Society.
Mahdavi, A., A. Mohammadi, E. Kabir, and L. Lambeva. 2008. Occupants' operation of lighting and shading systems in office buildings. Journal of Building Performance Simulation 1(1): 57-65.
Masoso, O.T., and L.J. Grobler. 2010. The dark side of occupants' behaviour on building energy use. Energy and Buildings 42(2): 173-7.
McMenamin, T.M. 2007. A time to work: Recent trends in shift work and flexible schedules. Monthly Labor Review 130: 3. http://heinonline.org/HOL/Page ?handle=hein.journals/month130&g_sent=1&id=1 269.
Motegi, N., M.A. Piette, S. Kinney, and J. Dewey. 2003. Case studies of energy information systems and related technology: operational practices, costs, and benefits. ICEBO--International Conference for Enhanced Building Operations. http://hdl.handle.net /1969.1/5195.
Nagy, Z., F.Y. Yong, M. Frei, and A. Schlueter. 2015. Occupant centered lighting control for comfort and energy efficient building operation. Energy and Buildings 94(0): 100-8.
O'Brien, W., and H.B. Gunay. 2014. The contextual factors contributing to occupants' adaptive comfort behaviors in offices--A review and proposed modeling framework. Building and Environment 77(0): 77-87.
Page, J., D. Robinson, N. Morel, and J.L. Scartezzini 2008. A generalised stochastic model for the simulation of occupant presence. Energy and Buildings 40(2): 83-98.
Parys, W., D. Saelens and H. Hens. 2011. Coupling of dynamic building simulation with stochastic modelling of occupant behaviour in offices--a review-based integrated methodology. Journal of Building Performance Simulation 4(4): 339-58.
Salsbury, T. 2005. A survey of control technologies in the building automation industry. Presented at the 16th IFAC World Congress. www.nt.ntnu.no/users/skoge/ prost/proceed ings/ifac2005/Fullpapers/02117.pdf.
Salsbury, T. 2009. The smart building. Springer handbook of automation. pp. 1079-93. S. Y. Nof, ed. Berlin, Heidelberg: Springer.
Sun, K., D. Yan, T. Hong, and S. Guo. 2014. Stochastic modeling of overtime occupancy and its application in building energy simulation and calibration. Building and Environment 79(0): 1-12.
Zeytinoglu, I.U., G.B. Cooke, and S. Mann. 2009. Flexibility: Whose choice is it anyway? Department des Relations Industrielles, Universite Laval: 64(4): 555-74.
Meysam Razmara, Research Assistant, Mechanical Engineering--Engineering Mechanics Department, Michigan Technological University, Houghton, MI: 1) How much is the computation time for parameter estimation?
2) A good control algorithm for building systems means not only low energy consumption but also low resulting discomfort level for occupants. Authors should study/mention the discomfort level in their paper.
3) UKF and EKF methods have been previously used for state-parameter estimation for building applications in the literature (e.g., Maasoumey et al. 2014). The authors should clearly clarify the contribution/difference of their work with respect to these studies.
Maasoumey, M., M. Razmara, M. Shahbakhti, and A.S. Vincentelli. 2014. Handling model uncertainty in model predictive control for energy efficient buildings. Energy and Buildings 77: 377-92.
H. Burak Gunay: 1) The parameter estimation is handled inside a commercial building controller (Delta DSC-1146) with a 32 bit processor and 2 MB flash memory in real time. Meaning that--despite the analytical and computational challenges in solving state and parameter estimation problems-- the computations were finished within a database scan period (2-5 Hz).
2) This paper presents the first stages of an adaptive occupancy and building learning temperature setback algorithm inside a laboratory environment (a shared office space with a stand-alone controls network). Therefore, we were able to infer the occupants' comfort through their simulated frequency of interactions with the thermostat. In the field implementation stage of the research project, we look into other metrics such as the ratio of arrivals below heating or above cooling setpoints, the occupied time spent below heating or above cooling setpoints, and the institutional complaint logs. This information will be presented in a future paper.
3) The authors are aware of the Maasoumey et al. paper (2014) and we acknowledge its contribution. In this endeavor, unlike the prior work, our emphasis is on developing practical indoor climate control strategies by incorporating occupants more actively and by acknowledging the computational, analytical, and sensing and metering-related limitations of the existing building controls networks.
H. Burak Gunay
Student Member ASHRAE
William O'Brien, PhD
Associate Member ASHRAE
Ian Beausoleil-Morrison, PhD, PE
Associate Member ASHRAE
H. Burak Gunay is a doctoral student and William O'Brien is an assistant professor in the Department of Civil and Environmental Engineering and Ian Beausoleil-Morrison is an associate professor and Jayson Bursill is a master's of science student in the Department of Mechanical and Aerospace Engineering, Carleton University, Ottawa, Canada.
|Printer friendly Cite/link Email Feedback|
|Author:||Gunay, H. Burak; O'Brien, William; Beausoleil-Morrison, Ian; Bursill, Jayson|
|Date:||Jan 1, 2016|
|Previous Article:||A general approach for predicting the thermal performance of metal building fiberglass insulation assemblies.|
|Next Article:||Investigating the effects of turbulence and premixed air/methane fuel combustion on the performance of a miniature gas turbine--computer numerical...|