The Sub-Metered HVAC Implemented for Demand Response Dataset

As the power grid evolves away from fossil fuels, new approaches to balancing supply and demand will become necessary. One possible approach is demand response (DR). Traditionally, DR approaches manipulate the power consumption of flexible loads to reduce operational costs and/or improve grid reliability [1]. However, DR can also be used to match generation with demand, which is especially useful for grids with high penetration of variable renewables [2].

Commercial building heating, ventilating, and air conditioning (HVAC) systems have been identified as a flexible load that could provide both short timescale fast-acting DR [3–12]. Commercial buildings have high thermal inertia, meaning that the HVAC system power consumption can be modulated for short periods of time before the building temperature would noticeably change. Additionally, commercial buildings account for 37% of all electrical load in the United States, with 49% of their electrical consumption by the HVAC system or supporting systems [13], making them a large potential source of flexibility [12]. HVAC systems have been shown to have potential in both slower timescale grid balancing [11] and faster timescales, such as for frequency regulation which operates on a timescale of seconds [9,10]. Additionally, researchers have further broken down DR potential to specific components of the HVAC system, such as the fans [6] or chillers [14], where individual device control may offer more precise tracking of a power reference signal.

Traditionally, DR is implemented via open-loop control signals, for example, time-varying electricity prices that encourage load shedding/shifting or direct load control signals from the utility that manipulate load power consumption (e.g., by switching loads on/off, or modulating their power consumption) in response to grid needs. However, DR for grid balancing should ideally use closed-loop control to achieve accurate power tracking, which requires measurements used as feedback to the controllers. Unfortunately building automation systems (BAS), used to control and monitor building variables, do not usually collect power measurements, and whole-building electric meters typically measure power at intervals of 15 min, which is too slow for faster timescale grid balancing. Moreover, whole-building electric load measurements only give a coarse picture of building HVAC system dynamics and responses.

In this paper, we introduce the Submetered HVAC Implemented For Demand Response (SHIFDR) dataset, a massive dataset for measuring the response of individual HVAC system components. The dataset includes HVAC system component power measurements during normal system operation and during a large variety of DR events, along with associated whole-building electric load data and control/monitoring BAS data, such as room temperature, fan airflow, and temperature set-points. Therefore, this dataset provides the link between the HVAC system and the building electrical power system that is needed to close the loop around DR events. At the time of writing, the dataset contains submetered fan power data (at 1 min resolution), associated whole-building electric load (at 5 min resolution), and BAS data (at 5 min resolution) from 14 commercial buildings in southeast Michigan, during periods in which open-loop DR events occurred. We have collected this data over 5 years of experimental testing and monitoring, and are now making it publicly available for other researchers to analyze. We demonstrate the value of submetering HVAC components through a case study that demonstrates that BAS data alone are likely insufficient for closing the power tracking control loop.

Much DR research is based on simulations, rather than experiments. In some cases, researchers have used building data acquired during normal operations to develop building models for DR simulations [15] or conducted building experiments to gather data to develop building models for use in DR simulations or future DR experiments [4,12]. However, many of these studies focus on a single building and usually the resulting datasets are not made publicly available, in contrast to our study, which includes a large number of buildings and releases our data publicly. Other studies analyzed data from buildings participating in DR pilots or programs, e.g., [16], but these data are generally not publicly available. Open building datasets include the Building Data Genome Project [17,18], which includes power measurements from 1,636 buildings; however, it does not include submetering data, only provides hourly data, and does not include data from DR events. The Volttron dataset [19] also does not include data from DR events. Ref. [20] published hourly whole-building electric load, chilled water, and summary statistics on temperature zone data from DR experiments in Northern California, but did not submeter HVAC components. Only a few projects have conducted experimental studies of DR for grid balancing, e.g., [9,10,21–24] conducted frequency regulation experiments on commercial buildings or laboratory facilities meant to mimic commercial buildings. These studies generally include submetering. However, in contrast to our dataset, these studies focus on only 1–2 buildings, and again their datasets are not publicly available.

Past research has demonstrated the mismatch between building simulation models and real building behavior [25,26], pointing to the value of better models, which could be enabled through collection of more building data. Past research has also identified the need for submetering to enable fast DR via closed-loop control, which reduces the need to rely exclusively on inaccurate building models unique to each building [23,24]. Additionally, past work has argued that the limitations of existing building control and monitoring infrastructure to perform accurate DR are not well understood [27]. To the best of our knowledge, no past work has explored the mismatch between (1) power estimates obtained from simulation models leveraging only BAS data and (2) power measurements, one of which is needed to close the loop for fast DR. We note that many DR simulation models assume these power measurements are available, e.g., [28,29], even though they are typically not available.

In summary, the contributions of our paper are as follows. First, we present a new dataset that includes submetered fan power, BAS, and whole-building electric load data from 14 commercial buildings during normal operation and open-loop DR testing. The scale of this dataset in terms of measurement resolution, the number of data points, and years exceeds what is available in the literature. Second, we present a case study demonstrating the importance of submetered fan power data for closed-loop control and modeling. We show that fan power cannot be accurately estimated on fast timescales using common physics-based models nor a regression model leveraging BAS data alone, meaning closing the loop on fan power will likely require submetering of the fans. Third, we discuss the challenges associated with usage of the BAS for closed-loop control in fast DR applications, specifically that it is often slow and has limited communication network capabilities, which hinder the acquisition, accuracy, and usage of real-time measurements. These results highlight a gap in the literature between theoretical building modeling and control and the actual capabilities of real building infrastructure.

The rest of the paper is organized as follows: Sec. 2 describes how commercial building HVAC systems can provide fast DR and the need for submetering. Section 3 describes the dataset, including measurements and experiments conducted. Section 4 describes a case study comparing fan power data to estimated data. Section 5 discusses the limitations of the dataset. Section 6 concludes the paper and discusses future research directions.

In this section, we first give a high-level description of commercial building HVAC systems and describe the capabilities and limitations of the BAS. Then, we explain how commercial building HVAC systems could provide fast, closed-loop DR, the challenges associated with closing the loop with BAS data alone, and the need for submetering.

The basic operation of a typical commercial building cooling system uses a chiller to cool the supply air, supply/return fans to move air through the ducts, and variable air volume (VAV) boxes to control the air supply to individual rooms or zones [4]. Figure 1 shows an example of HVAC operation for a single duct VAV HVAC system. Fan speed is controlled to maintain positive duct pressure, where the pressure set-point can be adjusted to improve fan power efficiency. Fans in the air handler unit (AHU) blow air over cooling coils that are supplied with chilled water by the chiller. The chiller cools the supply air in the AHU to a nominal temperature. The VAV boxes attach to the pressurized ducts and contain a damper valve that controls the volume of supply air that is allowed into the room or zone. The VAV box is controlled by a room temperature controller to adjust the airflow into the room to achieve the room temperature set-point. A return duct brings the air out of the room and back to the AHU, where it is exhausted out of the building or mixed back in with the supply air to reduce the amount of cooling required. Many of the buildings included in our dataset are on chilled water loops fed by an external chiller plant that distributes chilled water to several nearby buildings. The power consumption of the fans and chiller are a function of the airflow needed to supply the VAV boxes, the intake air temperature, the pressure in the duct, the HVAC controller design, and the set-points.

Commercial buildings are usually controlled by a centralized BAS that continually measures and controls components in the HVAC system. The BAS can capture hundreds of points throughout the building including pressure, temperature, humidity, and airflow. Some of these data are used by digital controllers that actuate HVAC system components, such as fan speeds and valve positions. BAS controllers are developed for each building to achieve temperature and ventilation goals [31,32]. BAS data give insight into building operations; however, HVAC system/component power consumption is not typically measured because it is not needed for typical building control functions. Additionally, the BAS can have communication network limitations. For example, if too many data points are captured and communicated too often, data channels can become congested and data can arrive at the central computer delayed, or can be lost completely, which can negatively affect normal building operations. Both of these limitations—lack of power measurements and BAS communication network limitations—make it difficult to leverage BAS data alone for fast DR applications that require closed-loop control of HVAC power consumption.

Fast, closed-loop DR, which enables commercial building HVAC systems to provide grid balancing, can be achieved through different types of controls. These strategies generally target fan power consumption, because chillers respond more slowly than fans. If control actions are fast, and without a significant dc component (i.e., energy neutral with respect to the baseline consumption), then fan power manipulation should have minimal impact on chiller power consumption. For example, [9] developed controllers to control fan power by directly manipulating (1) fan speed or (2) air flowrate set-points. It is also possible to control fan power by employing global thermostat adjustment (GTA) though this is, generally, a slower and less-precise control approach that may have some impact on chiller power consumption [30]. However, the benefit of using a GTA approach is that it uses the existing BAS control inputs and loops to update HVAC temperature set-points.

Figure 2 shows different options for closed-loop control of building fan power for grid balancing through manipulation of BAS set-points (e.g., temperature set-points via GTA). A DR controller sends set-point adjustments to the BAS, which sends control signals to the building. Monitoring data is sent back to the BAS, and whole-building electric load data is captured by the meter. The BAS data together with the whole-building electric load data may be used by the DR controller to close the loop. However, the DR controller also needs a measurement or estimate of fan power consumption, i.e., the controlled output. Figure 2(a) shows the case without submetering; BAS data is used to estimate fan power via a power estimator. Figure 2(b) shows the case with submetering; submetered fan power data is captured and sent directly to the DR controller.

Estimating fan power is nontrivial. Fan power is a nonlinear function of BAS variables [33]. Physics-based models have been used to estimate fan power changes based on airflow measurements [4,6,25,34], and have been used within model predictive control approaches [35,36]. However, these models may not accurately capture fan power consumption behavior. In Sec. 4, we find that while typical models have reasonable accuracy during steady-state operation, during transients at timescales relevant for fast DR, these models fail to accurately estimate fan power consumption. Our past work has also demonstrated the mismatch between fan power consumption simulation models and experimental results, e.g., [25]. This points to the value of fan power submetering versus approaches that use BAS data alone for fan power estimation.

Our work relates to the broader issue of the difficulty of accurately modeling buildings. This has implications for building controls, which leverage building models. Some approaches combine machine learning and model predictive control to close the loop around building power consumption for DR [37,38], which can help compensate for inaccurate building models. Other approaches leverage model-free machine learning [39,40]. However, machine learning requires significant data, and there is often insufficient training data from periods when buildings underwent DR. Instead, training and control verification is often done through simulation, which, again, requires very accurate building models. More building measurements, for example, fan power consumption data, can help us build better models and/or more effectively leverage machine learning approaches, which is another benefit of HVAC submetering, beyond just providing a measured (rather than estimated) control output, as described above.

The SHIFDR dataset is available at Ref. [41]. It includes data from 2017 to 2021 from 14 commercial buildings in southeast Michigan. DR experiments were conducted on these buildings in the summers by modifying room temperature set-points through GTA. Table 1 contains information about the buildings. Table 2 summarizes the data collected from each building. BAS data for VAV box, supply/return fan, and chiller monitoring; whole-building electric load data; and single-phase submetered fan current data (from which fan power data can be computed, as described in Sec. 3.1) are included. Due to changing experimental goals, the range and number of BAS data points captured and fans submetered for each building differ and may not be consistent from year to year. To protect the security of the buildings, they have been anonymized with fictitious names.

The remainder of this section is organized as follows. In Sec. 3.1, we describe the process used to collect submetered fan power data. In Sec. 3.2, we discuss the methodology for selecting the BAS data, and in Sec. 3.3, we describe the whole-building electric load data. We give a brief overview of our experiment design and testing that took place each summer in Sec. 3.4. Finally, in Sec. 3.5, we describe a preprocessed data subset from the summer of 2021 that is in an easy-to-use format, which is also included in the dataset.

Fan power consumption can be computed from fan current measurements. For each fan that was selected for submetering, an Onset CTV-C 100 amp or Onset CTV-D 200 amp split core current sensor was placed on a single phase of the fan wiring within the electrical panel to which the fan was most directly connected. Data from the sensors were measured by an Onset HOBO 4-channel analog logger, with each sensor connected to a separate logger. The AC current was logged in 1-min intervals. Each minute, the logger took 4 separate current readings from the sensor and logged the “exact,” maximum, minimum, mean, and standard deviation of the readings. The “exact” current measurement is the last reading from the sensor each minute, reported at the time the data was logged. Periodically, data was downloaded from the loggers and their storage reset, resulting in some missing data points.

where $ij(t)$ is the average single-phase current magnitude measurement of fan j at time t and $pf¯$ is the average power factor of that type (supply or return) of fan. We have used these assumptions and average values when computing fan power in our past work [26,30]. Based on the collected data, we estimated a 1% error on voltage and a 4% error in power factor [30], which we will use in our case study to determine the error of our power “measurements.”

For all the buildings in the dataset, BAS data points are generally logged at 30 min intervals. However, we increased this sampling frequency for a subset of points, referred to as “microtrends.” Based on previous experiences with BAS network congestion due to aggressive microtrending, the BAS administrators imposed communication limits in terms of sampling frequency and number of microtrended data points. Microtrends were generally tracked at 5 min intervals with reports sent weekly. In some cases, we tracked data at 1 min intervals.

In general, BAS measurements target three areas: AHU, VAV box, and chiller. AHU measurements include air temperature and humidity around the AHU intakes, return ducts, supply ducts, and reheat ducts. We measured the fan airflow and pressure for each AHU we monitored. VAV box measurements targeted the damper and valve positions for VAV boxes associated with several rooms as well as the room temperature and set-point. Generally, only one VAV box was measured for each AHU. For the chiller, we measured flow rates and water (supply and return) temperatures as well as chilled water valve positions. We did not submeter chiller power consumption, as the chilled water is supplied by an offsite chiller plant. The chiller plant provides chilled water to many different buildings, and determining the chiller load for individual buildings is nontrivial [30]. Chilled water flow into and out of each building are included in the BAS data.

We tried to select points that we thought would be important when interpreting the power consumption of the HVAC fans or chillers, and the cooling service provided to the building. As discussed in Sec. 5, the points that we selected do not always depict the building dynamics holistically. For example, for each AHU, we selected one room for which to measure the VAV box temperature, airflow, and control values. However, after analysis, we found that data from one room are generally not sufficient for developing a holistic representation of AHU behavior.

Whole-building electric load data are measured every 15 min by the utility company for billing. Data points represent the average power consumption of the building during the 15 min interval prior to the measurement. This includes the fan power consumption but not the chiller power consumption for all buildings other than Vostok. Vostok has its own chiller, while all other buildings are on chilled water loops fed by external chiller plants.

We conducted DR experiments each summer from 2017 to 2021. The events and type of events varied each year, reflecting changes in our research goals, incorporating learnings from past years. The full test schedule for each building is included in the dataset. Not every building from which we were collecting data was experimented on each year. During periods when we were conducting experiments, we would also ensure there were some days with no experiments, i.e., “baseline days,” useful for reference modeling and analysis. These days are not specifically mentioned in the dataset, but are days in between days with experiments. No experiments were conducted on weekends, as buildings were generally in unoccupied mode and demand response events are uncommon on weekends.

All experiments conducted were open-loop temperature set-point adjustments through GTA. Specifically, we programed predetermined temperature set-points into the BAS. These events fall into three categories that represent common strategies for DR:

1. Unipolar events: the temperature set-point is changed by a fixed offset during the entire event. This might represent a load shedding event, in which the cooling temperature set-point is raised slightly to reduce the HVAC power consumption. We also tested lowering the temperature set-point to achieve a load increase, which can be useful for balancing excess generation.

2. UP–DOWN event: the temperature set-point is decreased for a period to increase power consumption, and then increased for a period to decrease power consumption. This represents load shifting through pre-cooling.

3. DOWN–UP event: the temperature set-point is increased for a period to decrease power consumption, and then decreased for a period to increase power consumption. This also represents load shifting, where energy is “paid back” after an initial load reduction.

Typically, events were conducted for an hour, with 30 min for each period of an UP–DOWN or DOWN–UP event. We conducted events with varying temperature set-point changes. We define symmetric events as UP–DOWN or DOWN–UP events in which the magnitude of temperature change was the same in both halves of the event. We also conducted repeated UP–DOWN and DOWN–UP events, in which the temperature increased/decreased over several consecutive cycles, which results in a square wave signal sent to the thermostat. Some tests also used a temperature set-point ramp, in which the set-point would linearly change over a short time instead of making a single step change.

In 2017 and 2018, our research was focused on conducting experiments to enable comparison of our experimental results to past experimental results [42] and simulation results [25] exploring the efficiency of building response to load shifting events. Therefore, we focused on symmetric events, and conducted some unipolar events. We initially measured few BAS data points.

In 2019, our research was focused on better characterizing building thermal response to set-point changes. Therefore, we focused on unipolar events. We also conducted repeated UP–DOWN and DOWN–UP events to explore the effects of repeated actuation, which would be needed to provide grid balancing services.

In 2020, we studied the response to asymmetric UP–DOWN and DOWN–UP events, in which the UP and DOWN temperature set-point changes were different magnitudes. Previous experimental results had found that symmetric temperature set-point changes led to approximately symmetric fan power consumption changes [4]; however, our experimental results showed that this is not often the case [30]. Asymmetric tests were designed to explore how asymmetric temperature set-point changes could achieve symmetric fan power consumption changes and/or pure energy shifting in which the total energy consumption during the event is identical to the baseline energy consumption. Experiments in 2020 were limited due to the onset of the COVID-19 pandemic.

In 2021, we conducted more asymmetric UP–DOWN and DOWN–UP events and unipolar events. Experiments in 2021 were more extensive than in 2020, but the buildings were still only partially occupied because of the COVID-19 pandemic.

All experiments were constrained based on agreements with facilities staff. Specifically, we could change the temperature set-points within the buildings by no more than $2°$F for less than 1 hr.

Though the dataset includes all of our raw data to enable maximal transparency and usage, we also include data from six buildings that have been lightly processed for easy use. The inclusion of these files is meant to act as a quick setup for anyone who wishes to access the data without writing their own code to interpret BAS data or calculate power data from the current measurements. The preprocessed data covers the experiments conducted in the summer of 2021. Preprocessing involved the following:

1. The mean fan current measurements were used to calculate the power consumption of each fan, leveraging eq. (1).

2. The BAS data were linearly interpolated to 1-min intervals (from 5-min intervals).

3. The BAS and fan power data were combined into a single file.

4. The total fan power (i.e., the sum of all measured fan power) is provided.

As described in Sec. 3.4, the test schedule for each building is included with the dataset. In 2021, all of these buildings followed the same testing schedule.

In a majority of commercial buildings that could effectively implement fast DR events, submetered fan power is not available but BAS data is. Since the SHIFDR dataset includes both BAS data and submetered fan power consumption data, we can evaluate the performance of fan power estimation methods that rely solely on BAS data, by comparing fan power estimates to measured fan power data.

In this section, we present a case study demonstrating the value of HVAC submetering. Specifically, we implement widely used physics-based models relating fan power to airflow as well as a linear regression model utilizing all available BAS measurements, and we compare the power estimates to the power measurements. We use the preprocessed 2021 summer data described in Sec. 3.5. Specifically, we use data from baseline days when no DR events occurred to identify the parameters of the physics-based and linear regression models. Then, we use the models and the BAS data from DR events to estimate the fan power during DR events, and compare the estimated and measured fan power.

The rest of this section is organized as follows. In Sec. 4.1 we describe the modeling and parameter identification methods used to estimate fan power from BAS data. In Sec. 4.2, we present the results, showing that fan power is not accurately estimated by these models using available BAS data. Then, in Sec. 4.3, we compare these results to building simulation results generated using the Modelica Buildings Library [43], and show that our results are consistent with the simulation results. Finally, in Sec. 4.4 we describe other possible uses of the dataset.

The most common models used for estimating fan power from BAS data rely on the physical relationship between fan power and airflow volume. There has been debate in the literature whether this relationship is best approximated as linear, quadratic, or cubic. The physics tells us it should be cubic [44]; however, BAS operation and control affect the empirical relationship, which often has a better fit to a linear or quadratic function [4]. We will examine each option. Additionally, we develop and test a linear regression model (LRM) that is linear in every available BAS measurement for each building. While not tied to a physical understanding of the building, this data-driven model allows us to explore whether or not BAS data beyond airflow can help us estimate fan power consumption.

where $β1,j, β2,j, γ1,j$⁠, and $γ2,j$ are constant parameters specific to fan j. The total fan power consumption is the sum of the individual fan power consumption, whether measured or estimated. We note that we also tested the following model variants: (1) removing the second parameter in each model, which requires the curves to pass through the origin; (2) including all intermediary polynomial terms and associated parameters in the quadratic and cubic models; and (3) computing the total building fan power consumption as the sum of the linear, quadratic, or cubic terms for each fan. We do not include these models/results here for brevity, and because the associated results do not provide significant additional insights.

where $Mbas,k(t)$ is a column vector containing all available BAS data from building k at time t and χ_k is a row vector containing constant parameters specific to building k.

To determine each model's parameters, we use ordinary least squares on a set of training data, specifically, 22 baseline days (i.e., weekdays without DR events) from the summer of 2021. We also explored the impact of using a subset of DR event days for training and a different subset for testing; however, our results were very similar to those obtained using baseline days for training, and so are not included here.

The fan power estimation models were used to estimate the total fan power associated with all UP–DOWN and DOWN–UP events that occurred during the summer of 2021. Our results show that the models can estimate the fan power well in steady-state but are unable to reliably track fan power during DR events.

We first summarize the results of model fitting. Table 3 lists the means of the identified parameters and R² values across all fans in each building for the linear, quadratic, and cubic models. We do not summarize the LRM parameters as there are many, they are different across each building, and, since each BAS data point has different units, comparing parameter values does not provide much insight.

Figure 3 shows the mean estimated total fan power calculated by each model compared to the mean measured total fan power for each building. Specifically, we first normalize the start time of each event to t = 0. Then, we compute the fan-power estimates for each individual event, and average them across events of the same type (UP–DOWN or DOWN–UP) to obtain the mean estimated total fan power associated with each model. Finally, we average the fan-power measurements across events of the same type to obtain the mean measured total fan power.

We recognize that our fan power “measurements” themselves are imperfect since they are calculated from current measurements using Eq. (1) leveraging assumptions on power factor and voltage. Therefore, we assume that the power factor has a $±4%$ error and the voltage a $±1%$ error, and we approximate the standard error on the mean as $±5%/n$⁠, where n is the number of events of the same type used to compute the mean. Then, instead of plotting the calculated fan power, we plot the expected range of mean measured fan power (yellow area) based on these error assumptions.

Most mean estimates generally fall near the expected (measurement) range during steady-state operation before and after DR events. However, during events, the mean estimated fan power can significantly deviate from the expected range. In general, we found that the estimated power ramps more quickly than the measured power. This is best seen in the Caspian building data. The mismatch may be due not only to the accuracy of the estimation model, but also the accuracy and resolution of the BAS data. Overall, the LRM performed the best of all the models.

where $pk(t)$ is the measured total fan power of building k at time t, computed as the sum of the individual fan power measurements from building k, which are calculated from Eq. (1). We use Eq. (6) to calculate the time series error for each event in each building. Figure 4 shows the median error across all events at each time for each building. As before, we see that the error is relatively small during steady-state operation before and after DR events. During events, we see that the error increases across all models. Specifically, we see large spikes in error at $t=0, 0.5, 1$ hr, corresponding to times when the temperature set-points in the buildings are changed by DR events. This indicates that these estimation models and/or BAS data are not able to capture fan power transients, which can be captured by submetering. We have chosen to use the median value of the error as it provides resilience to outlier events. The randomness present in individual events can cause extreme error values, and we have found that using the median error provides more representative results.

We also calculated the normalized root-mean-square error (RMSE) between the estimated and measured total fan power for each event (using data from t = –1 to t = 3 h, i.e., the same range as shown in Fig. 4) and each building. We normalized by the average measured fan power in the building over this same time frame. Table 4 shows the mean RMSE and standard deviation of the RMSE across all UP–DOWN and DOWN–UP events in each building, for each estimation model and each building. We highlight the model with the lowest mean RMSE for each building by emboldening these values. It is important to note that this does not necessarily represent the “best” model for that building as small average RMSE is only one indicator of model accuracy. These results show that the LRM often, but not always, outperforms the other models.

Figures 3 and 4 and Table 4 demonstrate that these simple estimation methods are not well-suited to estimating fan power from airflow volume and other BAS data during transients inherent to DR events. Three possible explanations for this result are as follows. First, all of the physics-based models are overly simplified and neglect salient physics. As the building set-points change, the dampers will change position to compensate, which could potentially cause turbulence within the ducts. The cubic model is derived assuming constant laminar airflow. Further, these models only use BAS data from the current time-step, and do not account for airflow dynamics such as those caused by recent damper actuation. During steady-state operation, this is likely not an issue as the BAS variables remain relatively constant. This could be why the steady-state estimation is relatively accurate, while the transient estimation is not.

Second, the BAS data used within the models themselves are problematic. BAS measurements are taken at 5 min intervals and linearly interpolated to 1 min intervals. The error could be associated with interpolation error. However, this is unlikely to be a significant issue as typically the error spikes in Fig. 4 last for around 10 min and, in most cases, we see an error spike at the 5 min mark, which is a point that is not interpolated. Additionally, as we will discuss in Sec. 5, we have not captured all of the BAS data from the building. It is possible that BAS data could be used to estimate fan power accurately with a LRM, but that we are not measuring the right points to do so and/or we should be considering higher-order dependencies on BAS data points. BAS communication delays and/or faulty HVAC equipment/sensors could also result in old/inaccurate data being used within the models.

Third, we cannot rule out the possibility that our power “measurements” are worse than we have assumed, and the mismatch is due, at least in part, to bad “measurements” rather than bad estimates. Voltage and power factor do fluctuate, especially when fans switch on and draw in-rush current.

In the context of closed-loop building control, estimation error during transients could be detrimental to accurately tracking grid balancing signals, making HVAC DR a poor resource at fast timescales. Therefore, we argue that one either needs to use more accurate fan power consumption models than the models considered here, or to submeter fan power. While the former may be possible, one needs to consider the impacts on controller computational tractability, the availability and accuracy of BAS data required by those models, and model accuracy during DR actuation.

To verify the results presented in Sec. 4.2, we performed the same analysis on simulated data generated using the Modelica Buildings Library [43]. Specifically, we simulated UP–DOWN and DOWN–UP events on the ASHRAE2006 five-zone example building with reheat, from version 7 of the library. We conducted the events at 1 pm on 10 consecutive days starting on July 31st and we use TMY3 weather data. We also simulated the same days with no DR events to obtain baseline data for training the models. The linear, quadratic, and cubic airflow models were fitted to the baseline data using ordinary least squares, the same way as was done with the real building data. We do not implement an LRM since it is not clear which simulation data it should include.

The simulation results are shown in Fig. 5. Again, we plot the median error between the estimated total fan power and the measured (i.e., simulated) total fan power. We see that the error spikes during the DR event, similar to Fig. 4. The main difference between the results of Fig. 4 and the simulation results is that the latter have much larger estimation errors. A possible explanation for this is the simulated building is much smaller than the real buildings. The simulated building has a single supply fan that provides airflow to only five dampers. Therefore, damper adjustments need to be larger to achieve the same temperature changes, and the unmodeled physics become more important. Despite this, the simulation results confirm the phenomena we observed within the real buildings, i.e., simple models leveraging BAS data are insufficient for estimating fan power consumption, especially during transients inherent to DR events.

These results also lead us to consider whether a high-fidelity building model like the ones included in the Modelica Buildings Library should be used to estimate fan power within building controllers, and if such a model could produce high quality fan power estimates sufficient to close the loop without submetering. However, these types of models were built for simulation, not control. They are computationally complex and require calibration with significant amounts of building data. Moreover, while calibrated simulations can perform well on average, they may struggle to predict individual events [45], presenting challenges for control.

In this paper, we examine a single case study of using this dataset to verify different methods for power estimation. However, the dataset is not limited to just this application. Parts of this dataset have been used in Ref. [30] to explore the effect of GTA on building operations. Other parts have been used in Ref. [26] to evaluate the efficiency of HVAC systems providing load shifting for grid services. Additionally, we foresee that this data could be useful for future research in HVAC efficiency studies, building modeling, building-grid interaction modeling and simulation, or as training data for data-driven approaches to building modeling and control.

While our dataset offers a large amount of building data, it still does not completely capture the full building dynamics. As mentioned previously, due to concerns about BAS communication network congestion, we had to be strategic about which BAS points we chose to microtrend. However, in some cases, we selected points that were not important, or we missed important points needed to understand building response and the impacts on building-wide temperatures.

To highlight this issue, we can look at the single VAV boxes that we chose to monitor for each AHU. We selected boxes that we expected would represent typical behavior for that AHU. Figure 6 shows the mean room temperature and set-point temperature for UP–DOWN events. We observe three cases: receiving a set-point change and matching the set-point (e.g., Michigan RM4), receiving a set-point change and not matching the set-point (e.g., Caspian RM3), and not receiving a set-point change at all (e.g., Huron RM4). For the latter two cases, we know that the DR event must have led to temperature changes somewhere in the portion of the building supplied by that AHU since the fan power changed, but we failed to capture those temperature changes since the room we monitored did not change temperature properly. Additionally, we did not capture occupancy or ventilation override signals, which may have explained the VAV box response.

Our BAS data may also be inaccurate and/or delayed. BAS communication network limitations can lead to situations in which building measurements lag behind their physical values. Building controllers work around this by operating slower than the potential bottleneck (of the order of around 10 min). Other experimental studies have noted the network limitations within BAS control [20,23,24]. Additionally, the presence of faulty equipment/sensors is a known issue; however, the scale and ramifications of these issues is unclear [46]. While data validation is possible, we have not employed a data validation method to detect these issues in our BAS data.

In summary, the most significant limitations of our dataset are a direct consequence of the limitations in the BAS. The limited number of data points and communication bottlenecks caused by the BAS relying on aging hardware and software gives us an incomplete picture of the building operation. This issue is not easily addressable in existing buildings, though software updates, network overhauls, addition of sensors, and system retrofits could help. When designing new buildings and associated controls, one should design for not only traditional building operation but also for opportunities around DR and other forms of grid interaction, and consider more versatile controls, sensing, and communication systems. Future experimentation via, for example, living learning labs could help identify innovations and designs enabling more effective building-grid interactions [47].

In this paper, we introduced SHIFDR, a massive dataset containing submetered fan power and BAS data for 14 commercial buildings in southeast Michigan. The scope of this dataset in terms of the number of buildings, the complementary data streams (fan power submetering, BAS, and whole-building electric load data), and the availability of data from both baseline days and subhourly DR events is beyond what is currently available publicly. Through a case study utilizing the submetered data, we show that common models used to estimate fan power consumption from BAS data do not accurately track the submetered fan power during event transients. Our results indicate that proposed DR controllers that use BAS data and models to close the loop on fan power estimates may struggle to provide accurate fast DR. Instead fan submetering is likely necessary. Our results highlight the gap in understanding between the idealistic view of buildings in which HVAC DR is discussed in the literature and the actual capabilities of real building infrastructure. We hope that future analysis of the SHIFDR dataset can help bridge this gap in knowledge. Lastly, while the SHIFDR dataset is already large, we would like to make it larger by adding more buildings from other climate zones.

We are immensely grateful to the facilities and operations staff of the University of Michigan for supporting the research team. We would like to especially thank Andrew Berki, Kevin Morgan, Timothy Kennedy, Connor Flynn, David Anderson, Matthew Peterson, Andrew Cieslinski, Scott DeMink, Joel Kennedy, JW Krantz, and Scott Wells for providing their expertise and guidance. We would also like to thank the students and postdocs who have been involved in this project including Sina Afshari, Paul Giessner, Ruikai Xu, Han Li, Jordan Dongmo Nzangue, and Miguel Siller.

• US Department of Energy Building Technologies Office (Award No. DE-AC02-76SF00515; Funder ID: 10.13039/100006161).

• US National Science Foundation Division of Graduate Education (Award No. DGE 1841052; Funder ID: 10.13039/100000082).

• Guangdong Science and Technology Department (Award No. 2022A1515110833; Funder ID:10.13039/501100007162).

The data and information that support the findings of this article are freely available at footnote².

E =

percent error between measured and estimated total fan power

i =

current measurement

j =

fan index

k =

building index

M_bas =

column vector of available BAS data

n =

number of DR events

p =

fan power computed from current measurement

p_est =

fan power estimated with models

$pf¯$ =

average power factor

t =

time index

V =

fan airflow volume

$v¯$ =

average voltage

α, β, γ =

model parameters for linear, quadratic, and cubic fan power models

χ =

row vector of linear regression model parameters