Access the full text.
Sign up today, get DeepDyve free for 14 days.
A. Ruano, E. Crispim, E. Conceicao, M. Lúcio (2006)
Prediction of building's temperature using neural networks modelsEnergy and Buildings, 38
(2019)
Savings from your ecobee
Gaby Baasch, A. Wicikowski, Gaëlle Faure, R. Evins (2019)
Comparing Gray Box Methods to Derive Building Properties from Smart Thermostat DataProceedings of the 6th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation
Tianyu Zhang, Omid Ardakanian (2019)
A domain adaptation technique for fine-grained occupancy estimation in commercial buildingsProceedings of the International Conference on Internet of Things Design and Implementation
C. Chatfield (1990)
The Analysis of Time Series: An Introduction
W. Hu, Yong Luo, Zongqing Lu, Yonggang Wen (2019)
Heterogeneous Transfer Learning for Thermal Comfort ModelingProceedings of the 6th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation
P. Lauret, E. Fock, Rija Randrianarivony, Jean-François Manicom-Ramsamy (2008)
Bayesian neural network approach to short time load forecastingEnergy Conversion and Management, 49
C. Teodosiu, Raluca Hohotă, G. Rusaouen, M. Woloszyn (2003)
Numerical prediction of indoor air humidity and its effect on indoor environmentBuilding and Environment, 38
Datong Zhou, Qie Hu, C. Tomlin (2017)
Quantitative comparison of data-driven and physics-based models for commercial building HVAC systems2017 American Control Conference (ACC)
Nilavra Pathak, James Foulds, Nirmalya Roy, Nilanjan Banerjee, R. Robucci (2019)
A Bayesian Data Analytics Approach to Buildings' Thermal Parameter EstimationProceedings of the Tenth ACM International Conference on Future Energy Systems
F. Haldi, D. Robinson (2011)
The impact of occupants' behaviour on building energy demandJournal of Building Performance Simulation, 4
(2019)
Allied Market Research
M. Kane, Natalie Price, M. Scotch, P. Rabinowitz (2014)
Comparison of ARIMA and Random Forest time series models for prediction of avian influenza H5N1 outbreaksBMC Bioinformatics, 15
J. Braun, N. Chaturvedi (2002)
An Inverse Gray-Box Model for Transient Building Load PredictionHVAC&R Research, 8
A. Deconinck, S. Roels (2016)
Comparison of characterisation methods determining the thermal resistance of building components from onsite measurementsEnergy and Buildings, 130
Diederik Kingma, Jimmy Ba (2014)
Adam: A Method for Stochastic OptimizationCoRR, abs/1412.6980
M. Razmarab, M. Shahbakhtib, A. Vincentelli (2014)
Building Efficiency and Sustainability in the Tropics ( SinBerBEST ) Title Handling model uncertainty in model predictive control for energy efficient buildings Permalink
M. Hossain, Tianyu Zhang, Omid Ardakanian (2019)
Evaluating the Feasibility of Reusing Pre-trained Thermal Models in the Residential SectorProceedings of the 1st ACM International Workshop on Urban Building Energy Sensing, Controls, Big Data Analysis, and Visualization
Hao Huang, Lei Chen, E. Hu (2015)
A neural network-based multi-zone modelling approach for predictive control system design in commercial buildingsEnergy and Buildings, 97
P. Ferreira, A. Ruano (2002)
CHOICE OF RBF MODEL STRUCTURE FOR PREDICTING GREENHOUSE INSIDE AIR TEMPERATUREIFAC Proceedings Volumes, 35
D. Crawley, L. Lawrie, C. Pedersen, F. Winkelmann (2000)
EnergyPlus: Energy simulation programAshrae Journal, 42
Pablo Romeu, Francisco Zamora-Martínez, P. Botella-Rocamora, Juan Pardo (2013)
Time-Series Forecasting of Indoor Temperature Using Pre-trained Deep Neural Networks
Oluwaseyi Ogunsola, Li Song, G. Wang (2014)
Development and validation of a time-series model for real-time thermal load estimationEnergy and Buildings, 76
G. Mustafaraj, Glenn Lowry, Jyh-chen Chen (2011)
Prediction of room temperature and relative humidity by autoregressive linear and nonlinear neural nThe Lancet
C. Blundell, Julien Cornebise, K. Kavukcuoglu, Daan Wierstra (2015)
Weight Uncertainty in Neural NetworksArXiv, abs/1505.05424
C. Blundell, Julien Cornebise, K. Kavukcuoglu, Daan Wierstra (2015)
Weight Uncertainty in Neural Network
Amin Mirakhorli, B. Dong (2016)
Occupancy behavior based model predictive control for building indoor climate—A critical reviewEnergy and Buildings, 129
M. Fayazbakhsh, F. Bagheri, M. Bahrami (2015)
A Resistance–Capacitance Model for Real-Time Calculation of Cooling Load in HVAC-R SystemsJournal of Thermal Science and Engineering Applications, 7
G. Mustafaraj, Jie Chen, G. Lowry (2010)
Thermal behaviour prediction utilizing artificial neural networks for an open officeApplied Mathematical Modelling, 34
T. Lu, M. Viljanen (2009)
Prediction of indoor temperature and relative humidity using neural network models: model comparisonNeural Computing and Applications, 18
J. Moon, S. Jung (2016)
Algorithm for optimal application of the setback moment in the heating season using an artificial neural network modelEnergy and Buildings, 127
S. Salakij, Na Yu, S. Paolucci, P. Antsaklis (2016)
Model-Based Predictive Control for building energy management. I: Energy modeling and optimal controlEnergy and Buildings, 133
C. Buratti, E. Lascaro, D. Palladino, M. Vergoni (2014)
Building Behavior Simulation by Means of Artificial Neural Network in Summer ConditionsSustainability, 6
P. Blight (1991)
The Analysis of Time Series: An IntroductionJournal of The Royal Statistical Society Series C-applied Statistics, 40
Wilhelm Kleiminger, F. Mattern, S. Santini (2014)
Predicting household occupancy for smart heating control: A comparative performance analysis of state-of-the-art approachesEnergy and Buildings, 85
N. Zhu, Shengwei Wang, Zhenjun Ma, Yongjun Sun (2011)
Energy performance and optimal control of air-conditioned buildings with envelopes enhanced by phase change materialsEnergy Conversion and Management, 52
(2015)
Energy Savings from the Nest Learning Thermo-stat
J. Moon, S. Jung, Jong-Jin Kim (2009)
APPLICATION OF ANN ( ARTIFICIAL-NEURAL-NETWORK ) IN RESIDENTIAL THERMAL CONTROL
A. Pollard, A. Stoecklein (1998)
Occupant and Building Related Determinants on the Temperature Patterns in New Zealand Residential Buildings
V. Harish, Arun Kumar (2016)
Reduced order modeling and parameter identification of a building energy system model through an optimization routineApplied Energy, 162
Sooyoung Kim, Ji-Hyun Lee, J. Moon (2014)
Performance evaluation of artificial neural network-based variable control logic for double skin enveloped buildings during the heating seasonBuilding and Environment, 82
J. Moon, Sung-hoon Yoon, Sooyoung Kim (2013)
Development of an artificial neural network model based thermal control logic for double skin envelopes in winterBuilding and Environment, 61
D. Enescu (2017)
A review of thermal comfort models and indicators for indoor environmentsRenewable & Sustainable Energy Reviews, 79
S. Patil, H. Tantau, V. Salokhe (2008)
Modelling of tropical greenhouse temperature by auto regressive and neural network modelsBiosystems Engineering, 99
(1998)
The sustainable city; Volume 2; Electrotechnical: simulation and control; energy management: telecommunications. Institution of Professional Engineers New Zealand
Léopold Mba, P. Meukam, A. Kemajou (2016)
Application of artificial neural network for predicting hourly indoor air temperature and relative humidity in modern building in humid regionEnergy and Buildings, 121
J. Seem, S. Klein, W. Beckman, J. Mitchell (1989)
Transfer Functions for Efficient Calculation of Multidimensional Transient Heat TransferJournal of Heat Transfer-transactions of The Asme, 111
M. Gouda, S. Danaher, C. Underwood (2002)
Building thermal model reduction using nonlinear constrained optimizationBuilding and Environment, 37
Identifying Grey-box Thermal Models with Bayesian Neural Networks Md Monir Hossain Tianyu Zhang Omid Ardakanian University of Alberta University of Alberta University of Alberta Edmonton, Alberta Edmonton, Alberta Edmonton, Canada mdmonir@ualberta.ca tzhang6@ualberta.ca ardakanian@ualberta.ca ABSTRACT responsible for around 40% of the global energy use. For example, ecobee claims that eco+ customers can save up to 23% on their Smart thermostats are one of the most prevalent home automa- heating and cooling costs [8] which is on par with cost savings tion products. They learn occupant preferences and schedules, and reported by Nest, i.e., 10% on heating and 15% on cooling [30]. utilize an accurate thermal model to reduce the energy use of heat- To maintain a healthy and comfortable living and working envi- ing and cooling equipment while maintaining the temperature for ronment, smart thermostats rely on a black-box or grey-box heat maximum comfort. Despite the importance of having an accurate transfer model. This model relates changes in the room temperature thermal model for the operation of smart thermostats, fast and to a set of exogenous and control variables, e.g., ambient temper- reliable identification of this model is still an open problem. In this ature, and the amount of heat injected into or extracted from the paper, we explore various techniques for establishing a suitable ther- space by the heating or cooling system [20, 37]. The lumped param- mal model using time series data generated by smart thermostats. eter models — aka Resistance-Capacitance (RC) network models We show that Bayesian neural networks can be used to estimate for the analogy between heat flux and power flow [ 14] — are one of parameters of a grey-box thermal model if sufficient training data is the most popular grey-box thermal models due to their simplicity. available, and this model outperforms several black-box models in They divide the envelope and interior of a building into a number terms of the temperature prediction accuracy. Leveraging real data of temperature-uniform lumps and describe temperature dynamics from 8,884 homes equipped with smart thermostats, we discuss how inside the building using linear differential equations derived from the prior knowledge about the model parameters can be utilized the heat transfer theory. to quickly build an accurate thermal model for another home with While the accuracy of a thermal model may increase with its similar floor area and age in the same climate zone. Moreover, we order, training a high-order model is computationally expensive. investigate how to adapt the model originally built for the same Reduced-order thermal models can strike a balance between ac- home in another season using a small amount of data collected curacy and model complexity [12]. However, identifying even a in this season. Our results confirm that maintaining only a small low-order thermal model (e.g., 2R2C or 3R3C) requires several days number of pre-trained thermal models will suffice to quickly build of data. Should this data be available, the parameters of the RC accurate thermal models for many other homes, and that 1 day model can be estimated by solving a constrained optimization prob- smart thermostat data could significantly improve the accuracy of lem [12, 14], using a Kalman filter [ 23] or through supervised learn- transferred models in another season. ing [32]. While sufficient training data is being gathered by a smart thermostat to identify the building’s thermal model, the thermostat KEYWORDS may take suboptimal control decisions which violate the thermal System Identification, Transfer Learning, Bayesian Neural Network comfort requirements and increase the energy use. Thus, reducing ACM Reference Format: the amount of training data and the time necessary for building an Md Monir Hossain, Tianyu Zhang, and Omid Ardakanian. 2020. Identifying accurate thermal model is an important research problem which Grey-box Thermal Models with Bayesian Neural Networks. In Proceedings we study in this paper. of arXiv preprint. ACM, New York, NY, USA, 10 pages. Leveraging data from smart thermostats installed in 8,884 homes in Canada, we explore various techniques for establishing a well- 1 INTRODUCTION suited grey-box thermal model for each home. We show that a The demand for smart and programmable thermostats has been on first-order model, i.e., 1R1C, does not yield an acceptable level of the rise in the past decade thanks to improving living standards in accuracy in predicting indoor temperature for many homes in this developing countries, and increasing retail electricity prices around dataset. This motivates us to build a 2R2C model for a small number the world. The global smart thermostat market was pegged at $1.36 of homes (chosen by clustering) using several days of training data billion in 2018 and is anticipated to reach $8.78 billion by 2026 [1]. collected during one season. We utilize a Bayesian Neural Network Smart thermostat devices, such as ecobee, Nest, and Resideo, take (BNN) to estimate parameters of the RC model, and investigate into account their measurement of environmental variables and how to transfer the prior knowledge about the model parameters building occupancy along with weather forecasts to optimally con- to another season and across homes that belong to the same cluster. trol heating and cooling equipment while maintaining the room The clusters are formed using metadata (floor area and year built), temperature within desired limits. This optimal control could re- which is readily available for the homes in our dataset. This allows sult in significant energy and cost savings in buildings, which are us to select the most appropriate model from a library of pre-trained arXiv preprint, 2020 models for a given home knowing only its age and floor area, and arXiv:2009.05889v1 [eess.SY] 13 Sep 2020 arXiv preprint, 2020 Hossain et al. eliminates the need to collect several days of training data from Zhou et al. [40] compare a low dimensional RC model with a physics- each home. Thus, from the first day, the smart thermostat can use a based model, and conclude that the RC model can substitute the fairly accurate thermal model to optimize comfort and energy use. physics-based model with a negligible loss of accuracy. The main intuition for this work is derived from the observation The RC model can be arbitrarily complex. Several attempts have that despite all differences in the layout and design of homes, there is a been made to date to represent the building interior and its enve- limited number of combinations of RC parameters. This is especially lope using RC network models of different orders. For example, true, if we restrict our focus to one country with a narrow range of 4R1C [13, 23], 3R2C [31, 41], and 2R1C [12] networks have been climates and specific building codes. Thus, the knowledge acquired used to model the building envelope, while 1C [23], 1R1C [13, 41], in one home will be quite useful in another similar home. and 2R2C [31] networks have been used to represent the building Our contribution is threefold: spaces. The majority of studies that build an RC model assume • We propose a methodology based on Bayesian neural net- the knowledge of the building insulation, thermal mass, floor area, works for identifying the RC model of a home equipped with layout, and construction material [4, 7, 10, 38]. Leveraging this a smart thermostat. We compare the predictive power of dif- knowledge can indeed greatly reduce the complexity of model ferent RC models and show that a 2R2C model yields lower training. However, this contextual and physical information is typi- accuracy than other low-order models for most homes in cally hard to obtain without an intrusive energy audit, especially our dataset. for residential buildings that we study in this work. To overcome • We show that a grey-box model is more accurate than several this barrier, one approach is to estimate the parameters of the RC black-box models, including time series and neural network model from time series data generated by smart thermostats. Specif- models, in predicting the room temperature. ically, the model can be built by solving a constrained optimization • We assign homes in our dataset to a small number of clusters problem [12, 14], employing an unscented Kalman filter [ 23], and based on their floor area and age, and show that this clus- utilizing the genetic algorithm [31]. All these methods require a tering allows for transferring a pre-trained representative significant amount of data to build an accurate model. On the con- model of that cluster to this home with and without adap- trary, this paper focuses on building a suitable RC model for a given tation. We also discuss how a model trained for one season residential building utilizing its metadata and a small amount of can be transferred to another season. time series data. The dataset we use in this study contains time series data and Our work is similar to [2], which focuses on estimating the metadata obtained from a large number of ecobee smart thermostats. amount of heat loss through the building envelope using a 1R1C We describe this dataset in Section 3. model. The authors evaluate the balance point, decay curve, and energy balance models to cluster buildings and estimate the heat loss. Compared to that work, we form clusters based on metadata, determine the level of complexity of the RC model for each building, 2 RELATED WORK and use transfer learning to reduce the amount of data needed for The optimal control of the building Heating, Ventilation, and Air system identification. Pathak et al. [ 32] use Bayesian learning to Conditioning (HVAC) system is of great importance as it accounts estimates the RC parameters for one season and then utilize the for a large fraction of the building energy use, and is responsi- learned parameters in another season. Our work is similar to that ble for maintaining the temperature inside the building within a work but we explore the possibility of adapting and reusing an RC comfortable range. To optimally control this system, most related model in buildings that have similar characteristics to the building work adopts receding horizon control which relies on a model that for which this model was initially built. explains how the room temperature changes as a result of imple- Transfer learning allows for taking advantage of the knowledge menting a certain control policy. This has given rise to a large obtained in one model in another model. For example, Zhang et number of studies aiming to solve a system identification problem al. [39] employ transfer learning to transfer an occupancy estima- to infer this thermal model. tion model across different buildings. Although transfer learning A variety of data-driven techniques have been used in the lit- can provide initial estimates for the model parameters, there are erature to establish the thermal model. This includes Artificial other challenges such as over-fitting and unobserved patterns in Neural Network (ANN) [34], ANN with Levenberg-Marquardt the training data of the source domain. Bayesian neural networks (LM) [18, 26, 27], ANN with Backpropagation (BP) [5, 25], Mul- include uncertainty in the weights and biases to resolve this is- tilayer Perceptron (MLP) with LM [16, 24], Radial Basis Function sue [3]. The uncertainty in neural networks avoids over-fitting and (RBF) [11, 36], Autoregressive model with exogenous variables makes it possible to identify unseen patterns and give reasonable (ARX) [22, 29], and Autoregressive Moving Average model with predictions [21]. A recent study utilized BNN together with trans- exogenous inputs (ARMAX) [28, 33]. These black-box models es- fer learning to build thermal models [15]. Compared to that work, sentially map a number of features to the room temperature, and we discuss how to choose the order/complexity of the RC model, their accuracy highly depends on the selected features [9]. compare this RC model with several black-box models proposed Another type of thermal models is the RC model which is com- in the literature, and incorporate metadata in the model training monly used for heat transfer analysis in buildings. This grey-box process to reduce the need for temperature time series data. We model turns building spaces and multi-layered walls into a number also investigate how to transfer a model trained in one season to of latent thermal resistances and capacitances. Despite its simplicity, another season within the same home. it achieves a high accuracy in predicting the indoor temperature. Identifying Grey-box Thermal Models with Bayesian Neural Networks arXiv preprint, 2020 Table 1: A subset of time series data included in the dataset. Feature Data Type Unit Feature Data Type Unit HVAC Text N/A T Decimal F mode out T Decimal F Motion Binary N/A in Figure 1: A first-order RC model representing the building T Decimal F H Decimal (pct.) N/A setcool in envelope with a single thermal resistor (1R) and the building T Decimal F setheat interior with a thermal capacitor (1C). 3 DATASET To build and evaluate grey-box and black-box thermal models, we use the smart thermostat dataset released by ecobee – one of the key players in the smart thermostat market – as part of a program called ‘donate your data’. This program allows customers to voluntarily share their anonymized smart thermostat data with researchers to foster research and development. The dataset contains time series data generated by smart thermostats along with metadata about Figure 2: An RC model representing the building envelope homes where the thermostats are installed. The metadata contains with an nR(n− 1)C network and the interior of the building information about the homes and their occupants, e.g., location, using a thermal capacitor (1C). year built, floor area, typical occupancy level, number of cooling and heating stages, etc. There is a total of 104,693 homes in this dataset located in 51 is the relative humidity inside the building and motion is a binary countries with different climates and building codes. In this paper, variable which is 1 when motion is detected by the sensor and is 0 we just consider homes that are located in Canada and are therefore otherwise. To impute missing data (except for the motion data) we in the same climate zone (Zone 7). This yields 12,960 homes which use a simple linear interpolation technique. For motion data, we is a reasonable size sample. The data is available between 2015 set all missing values to zero. to 2019. The number of participating homes has increased over time as more homes equipped with ecobee thermostats opted in to 4 GREY-BOX THERMAL MODELS participate in this program. Hence, the lengths of time series data In this section we explain a grey-box modelling approach which are different for different homes. Moreover, the number of homes builds a thermal model for a building given time series data gathered varies based on the season. by the thermostat installed in that home. We initially restrict our study to winter due to the additional The RC model uses the analogy between thermal and electrical complexity that mixing data from different seasons presents [ 32]. conduction to model heat flux, Q , in a building. It models the heat For the winter season, we consider homes in our dataset that have transfer across the building envelope and inside the building using at least 3 months worth of data between November 2018 to February multiple constant thermal resistors and capacitors connecting spa- 2019. This gives us 8,884 homes to run the experiment on. We then tially temperature-uniform lumps. Specifically, the building exterior attempt to transfer models from the winter season to the summer and interior can be modelled by a number of constant resistances season considering homes that have reported at least 3 months and capacitances. The thermal resistor reduces the current heat worth of data between May and August 2019. This gives us 8,834 flux between its two terminals. The amount of heat that can pass homes to run the experiment on. through the resistor can be computed from Q = ∆ T/R, where ∆ T We identify and separate the homes located in Canadian using the metadata file that contains information regarding the country, is the temperature difference between the two terminals and R is province, and city a specific home is located in. We also use a subset the thermal resistance. The thermal capacitor stores heat according ∂T of time series data which exist in this dataset. In particular, for each to C = Q , where C represents the constant capacitance. ∂t home we use indoor and outdoor temperatures, humidity, motion To build an RC model, the first step is to decide on its order indicator, operation mode of heating and cooling equipment, heat- (determining the model complexity). In general, the higher the ing setpoint, and cooling setpoint. All these features are recorded at complexity is, the more accurate the model becomes as it does not 5-minute intervals. We note that this dataset does not contain solar lump together elements that do not have the same temperature. radiation, energy use of plug loads, or more accurate or fine-grained However, increasing the order of the model is a mixed blessing. occupancy information. Thus, we cannot incorporate them in our The higher the order of the model is the more difficult it becomes model. to estimate the RC parameters in terms of computational cost and Table 1 shows the features used in this study. Here, HVAC the amount of training data needed. More information would also Mode represents the state of the HVAC system, which can be one of be required about different spaces within the building (e.g., the the following textual values: ‘off ’, ‘heat’, ‘cool’, and ‘auto’. T is blueprint) to train a high-order RC model. Prior work suggests that in the average indoor temperature. T and T are indoor a 2nd-order RC model gives a negligible loss of accuracy compared setcool setheat cooling and heating setpoints, respectively. T is the outdoor to a 20th-order RC model [12]. Hence, we focus on building an out temperature that is collected from the nearest weather station. H accurate low-order RC model for residential buildings. in arXiv preprint, 2020 Hossain et al. nRnC Models: We model the building envelope using an nR(n−1)C In both models, we assume that the temperature sensor (which is network since new built homes typically have multi-layer walls embedded in the thermostat) is located in a place that can measure and insulated glazing. When n = 1 the 1R model lumps the building the indoor temperature with negligible error. Moreover, we use envelope into one thermal resistor, denote by R . The nR(n− 1)C constant values Q and Q to represent the heat injected heat cool model for n > 1 divides the building envelope into 2n − 1 com- or extracted by the HVAC system, assuming that it injects and ponents, including n resistors, denoted by R ,··· , R , and n − 1 extracts heat at constant rates in the heating and cooling stages. 1 n capacitors, denoted by C ,··· ,C . This model can represent the This assumption does not necessarily hold in practice; nevertheless, 1 n−1 n-layer structure of the wall [14]. The 2n − 1 parameters depend our experiments show that it does not introduce a significant error. on the thickness, materials used, and insulation of the wall. We also assume that Q = f(motion, humidity) where internal We use a single capacitor to represent the building interior as f is any function of motion and relative humidity, the quantities one thermal zone which absorbs and retains heat. The constant that are measured directly by smart thermostats. Unfortunately, capacitance is denoted by C in the 1R1C model and C in the nRnC the binary motion data is not typically a good proxy for home 1 n model. This thermal capacitance determines how much inertia the occupancy as a home might be occupied during a period with no building provides against temperature fluctuations, and depends recorded motion events. Our experimental results confirm that on the building’s floor area and ceiling height. Hence, the larger a incorporating the measured motion and relative humidity in the building is, the higher its thermal mass (or capacitance) would be. grey-box and black-box models described in the next section has Since the heating (or cooling) system can inject heat into (or an imperceptible impact on the predictive power of these models . extract heat from) the space, we connect it to the node that repre- Thus, we ignore the internal heat gain due to occupant presence sents the building space. We denote by Q the heat flux caused and activities, i.e., Q , in the rest of the paper. This simplifies HV AC internal by HVAC and denote by Q the internal heat gain due to the indoor heat gain or loss to Q . We describe how to select internal HV AC occupant presence and activities, and other latent variables (e.g., the order of the RC model and how to estimate model parameters appliance usage). Figure 1 and Figure 2 illustrate the 1R1C and using a Bayesian neural network in Section 5. nRnC models, respectively. The temperature dynamics in the 1R1C model can be described 5 METHODOLOGY by the following equation: In this section we discuss how to estimate parameters of an RC 1 Q + Q HV AC internal model and how to train various black-box thermal models for a ∂T = (T − T )∂t + ∂t (1) in out in C R C given home assuming that a sufficient amount of smart thermostat 1 1 1 time series data is available. where T represents the temperature inside the building, T in out represents the outside temperature, and Q = k Q − HV AC heat heat k Q where k and k are two binary control vari- 5.1 Parameter Estimation for RC Models cool cool heat cool ables determined by the operation mode of the HVAC system, i.e., To estimate the model parameters we first transform the linear dif- HVAC : mode ferential equations to linear difference equations. This allows us to use different techniques for model parameter estimation consider- 1, if T < T & HVAC ∈ {auto,heat} in setheat mode k = ing several consecutive measurements from the smart thermostat. heat 0, otherwise 5.1.1 1R1C model. We transform the linear differential equation (1) 1, if T > T & HVAC ∈ {auto,cool} in setcool mode k = to a linear difference equation by replacing the temperature differ- cool 0, otherwise ence between two consecutive time slots with ∆ T and the tem- in perature difference between the outdoor and indoor environment This simplified model incorporates the heat introduced or extracted with ∆ T : by the HVAC system and the internal heat gain due to occupancy, dif f but neglects solar radiation, and different wall insulation layers. 1 Q Q Similarly, the nRnC model can be expressed using a system of heat cool ∆ T = ∆ T + k · − k · in dif f heat cool linear differential equations describing heat flow inside the building C R C C 1 1 1 1 and across its envelope: Given m successive values of ∆ T and ∆ T , and leveraging in dif f 1 1 Q Q ∂T = (T − T )∂t + (T − T )∂t heat cool 1 out 1 2 1 the fact that R C , , cannot be negative, we solve the 1 1 C R C R C C 1 1 1 1 1 2 following non-negative least squares problem to determine the 1 1 ∂T = (T − T )∂t + (T − T )∂t ∀i∈{2, ...,n−1} i i−1 i i+1 i model parameters which are collected in vector x : C R C R i i i i+1 1 Q + Q HV AC internal ∂T = (T − T )∂t + ∂t (2) in n−1 in min||Ax − y|| C R C x n n n s.t. x ≽ 0 where T ’s are the envelope temperatures as depicted in Figure 2. This model incorporates the heat introduced or extracted by the HVAC system and internal heat gain due to occupancy, but neglects the heat flux from solar radiation as solar radiation and the window 1 The difference between the average RMSE with and without these features is 0.0066 area are not captured in our dataset. for 100 homes that are randomly selected from the dataset. Identifying Grey-box Thermal Models with Bayesian Neural Networks arXiv preprint, 2020 where [1] ∆ T [1] [1] [1] in ∆ T k −k dif f heat cool C R 1 1 [2] [2] [2] [2] ∆ T ∆ T k −k Q in heat dif f heat cool A = x = y = C ··· . cool [m] [m] [m] . ∆ T k −k dif f heat cool [m] ∆ T in Figure 3: Structure of the BNN-RC model used to predict the We use the active-set method to solve this convex optimization indoor temperature. problem. 5.1.2 nRnC model. To estimate the parameters of an nRnC model, Next we can calculate the inverse of (F · I− Φ) matrix, which we first find the time domain solution of the system of n + 1 linear is equal to the adjugate of (F · I − Φ) divided by determinant of differential equations expressed in (2). The state-space representa- (F · I− Φ). Assuming that this determinant is nonzero we have: tion of this system is: n−1 n−2 ∂x M F + M F +··· + M 0 1 n−1 −1 = Ax + Bu, (3) (F · I− Φ) = n n−1 n−2 ∂t F + e F + e F +··· + e 1 2 n y = Cx + Du, (4) where where M = I e = −Tr(ΦM )/i 0 i i−1 M = ΦM + e I 1 1 1 e = −Tr(ΦM )/n i i−1 i n n−1 − + 0 ··· 0 row 1 C R C R C R 1 1 1 2 1 2 and Tr() denotes the trace of a matrix. Hence, the system output ←−−−−−−−− v −−−−−−−−→ row 2 2 A = response can be written in a compact form as: ←−−−−−−−− v −−−−−−−−→ row n− 1 n−1 n n Õ Õ 1 1 y = S u − e y (6) 0 ··· 0 − row n t i i t−iδ t−iδ C R C R n n n n i=0 i=1 " # 1 1 1 1 where v = 0,··· , 0, ,− + , , 0,··· , 0 C R C R C R C R | {z } i i i i i i+1 i i+1 | {z } S = CM Γ 0 0 2 i−2 zeros n−i−1 zeros S = C(M (Γ − Γ ) + M Γ ) i ∈ {1,··· ,n− 1} i i−1 1 2 i 2 0 0 ⊤ C R S = CM (Γ − Γ ) ⊤ n n−1 1 2 1 1 0 . 0 0 0 . We get a linear function that maps the measured room tempera- . . . . ture in the past n time slots, and the outside temperature and heat . . . B = C = D = . . . . flux from the HVAC system in the current and past n time slots to 0 0 0 the room temperature in the current time slot: Q Q heat cool 0 − C C y = f(u ,u ,··· ,u ,y ,··· ,y ) n n t t t−1 t−n t−1 t−n Here the state vector x = T T ··· T T collects Here f is a linear function that can be approximated using a 1 2 n−1 in Bayesian neural network. The Bayesian neural network mimics the envelope and room temperatures, the output y denotes the indoor structure of the nRnC model; thus, we refer to this model as BNN- temperature (T ), the input vector u = T k k in out heat cool nRnC. Specifically, we use the terms in Equation (6) to decide about collects the outdoor temperature and the HVAC system control the neurons. Figure 3 illustrates the structure of our BNN-nRnC inputs, and A, B, C and D are constant coefficient matrices. As model. We use Gaussian scale mixture distribution with σ = 1, shown in [38] the response of this system can be written as: 1 σ = 0.1, and π = 0.2 for the weights and biases in our BNN. The −1 x = (F · I− Φ) (F · Γ + Γ − Γ )u , (5) t 2 1 2 t model learns 4n + 3 compound RC parameters for an nRnC model. where F denotes the forward shift operator (i.e., Fx = x where t+δ 5.1.3 Transfer learning. A large amount of time series data is re- Aδ δ denotes the time step), Φ = e is the matrix exponential, and quired to train a BNN-nRnC model which can accurately predict Γ the temperature inside the building. However, sufficient training −1 −1 Γ = A (Φ− I)B, Γ = A − B . 1 2 data is not always available for all homes and collecting this much data can be time consuming. More importantly, the HVAC system Substituting x in (4) yields the system output response: will be controlled in a suboptimal fashion before a good RC model −1 y = C(F · I− Φ) (F · Γ + Γ − Γ )u + Du , is trained and used for model predictive control. To address this t 2 1 2 t t problem, we select several representative homes as source homes −1 = C(F · I− Φ) (F · Γ + Γ − Γ )u . 2 1 2 t and use the posterior weights trained in those BNN-nRnC models The second equation above holds because D = 0 ··· 0 . as the prior for the new homes. Furthermore, we retrain the model arXiv preprint, 2020 Hossain et al. with the small amount of data that might be available in the target To estimate q, the parameter of the moving average part of the home. We discuss the strength of transfer learning in Section 6.3. ARMIAX model, we examine the autocorrelation function (ACF) Note that we do not transfer 1R1C models to other homes. This which measures how a time series is correlated with itself at differ- is because the 1R1C model describes the thermal resistance of the ent lags. The ACF plot of the temperature time series suggests that building envelope using a single resistor regardless of the material q should be set to 2 for the ARIMAX model. used, the number of layers and thickness of each layer. Hence, Although using the same parameters for all homes may not the value of R does not really represent any of the above factors. result in optimal ARIMAX models, but it significantly reduces Transferring this knowledge will not be helpful for training a model the complexity of building the time series model for each home. for a different home. We selected one home randomly and performed grid search with 125 parameter combinations to find the best ARIMAX model. For this specific home, the best ARIMAX model found by grid search 5.2 Developing Black-Box Models is ARIMAX(0, 1, 2) which has RMSE of 0.483 degree Fahrenheit, The temperature evolution inside a home can also be modelled using while using the same parameters for all ARIMAX models (i.e., a purely data-driven approach. In this section we introduce 4 models ARIMAX(1, 1, 2)) yields RMSE of 0.487 degree Fahrenheit for this adopted in related work. These models can be trained using time home. We also tested the parameters on multiple randomly selected series generated by the smart thermostat, namely Autoregressive homes and obtained similar RMSE values. This means that the Integrated Moving Average with exogenous variables (ARIMAX), ARIMAX models obtained using the same parameters achieve suf- Long Short Term Memory (LSTM), Deep Neural Network (DNN), ficiently high accuracy in predicting the temperature inside the and Random Forest (RF). These models are capable of predicting homes. Hence, in this work we use the same parameters(1, 1, 2) for the indoor temperature using the same features as the BNN-nRnC all ARIMAX models. model. We use these models as baselines to evaluate the grey-box RC model in the source domain and after transferring to the target 5.2.2 Long Short Term Memory (LSTM) model. LSTM model is domain. another time series model that is widely used to learn long-term 5.2.1 ARIMAX model. The temperature time series data exhibits dependencies in data. In particular, LSTM maintains a state to cap- temporal correlation and responds to changes in several variables ture the history of the input and output sequence, enabling it to including the ambient temperature, and the operation mode of the learn complex temporal dependencies. The LSTM network used for HVAC system, i.e., whether it is in the heating or cooling stage. indoor temperature estimation is a stacked LSTM model. It has two Thus, to model the temporal correlation and capture the effects of hidden layers which contain 20 cells each. This stacked model gives external variables on the room temperature, we fit an Autoregres- the best average RMSE result for 5 randomly sampled homes in our sive Integrated Moving Average model with exogenous variables dataset over 20 different stacked and single-layer LSTM models. (ARIMAX) [6] to this data. The ARIMAX model is basically a variant Our LSTM model contains one output node in the output layer of the widely used Autoregressive Moving Average (ARMA) model, representing the indoor temperature. The input layer has 4 cells which is more appropriate for modelling non-stationary time se- which are the previous indoor temperature, the current outdoor ries and can account for external variables which could affect the temperature, the current HVAC cooling operation state, and the time series. This model has three parts, namely the autoregres- current HVAC heating operation state. The loss is computed using sive part (AR), the integrated part (I), and the moving average part the mean squared error between logits and labels. For each batch (MA). These parts are characterized by three parameters, denoted of the data, we minimize the loss using a first-order gradient-based by p,d,q, respectively. optimization method [19]. This LSTM model is fed data from n The parameter d of the integrated part of the model denotes consecutive time steps as a sequence, where n depends on the the number of times we apply the differencing operator on the order of the respective BNN-nRnC model (for fair comparison). Our time series to ensure that the resulting time series is stationary. experiments show that the number of time steps has a negligible Hence, we need to analyze the stationarity of the time series for impact on the estimation result in terms of RMSE. estimating d . We use the Augmented Dickey-Fuller (ADF) test – a formal statistical test for stationarity – to check stationarity of the 5.2.3 Deep Neural Network (DNN) model. Deep neural networks differenced time series. This test enables us to identify if the time have been successfully used in the past for forecasting tempera- series has a unit root (i.e., a stochastic trend in the time series). ture [35]. Our DNN model consists of 400 hidden cells which are Running the ADF test on temperature data of one sample home divided evenly into two hidden layers. It adopts Rectified Linear reveals that the p-value is greater than 0.05. Hence, we need to Unit (ReLU) as the activation function for all hidden layers to have difference the time series at least once. We find that after one dif- low computational cost. The linear activation function is used in ferencing the time series becomes stationary (p-value is less than the output layer, which contains one output node for estimating 0.05). Therefore, we set d to 1. the indoor temperature. This model takes the same number of in- To estimate p, the parameter of the autoregressive part of the put features as the benchmarking BNN-nRnC model, i.e., 4n + 3 ARMIAX model, we draw the partial autocorrelation function features to learn the temporal correlations of the data over time. (PACF) which measures the correlation between the original time This architecture is determined by performing grid search over 5 series and the lagged time series after removing the effects of inter- different depths and 10 different numbers of units in each hidden vening past observations. The PACF plot of the temperature time layer. The DNN model utilizes the same loss function and is trained series suggests that p should be set to 1. using the same optimization method as the LSTM model. Identifying Grey-box Thermal Models with Bayesian Neural Networks arXiv preprint, 2020 5.2.4 Random Forest. Random forest can determine the feature im- portance by using different subsets of available features to construct decision trees. Each decision tree learns multiple binary rules based on the input data. Random forest takes the average of estimations from all decision trees and returns it as the final result. Random forest is computationally inexpensive, easy to imple- ment and also powerful. It gives promising results for time-series prediction [17]. We use the bisection method to find the most appro- priate number of trees, which is set to 100 based on results obtained from 5 randomly selected homes. We use the Gini impurity measure to decide on the optimal split. Figure 4: RMSE distribution for different BNN-RC models (left panel) and RMSE density for outliers only (right panel). 6 EVALUATION An outlier is defined as a home with an RMSE that is greater We build several grey-box and black-box models for the homes in than 1.5×IQR added to the third quartile or is less than 1.5× our dataset and compare them in terms of temperature prediction IQR subtracted from the first quartile. accuracy in the test data set using the root-mean-square error (RMSE). We answer the following questions in this section: (a) Which RC model has the best overall performance considering all square foot. This suggests that the first-order RC model typically homes in our dataset? (b) How much training data is needed to performs well in small homes with a low occupancy level that have build an accurate model? (c) Can we utilize a pre-trained RC model single-stage heating and cooling equipment. Since the 1R1C model with little or no adaptation (retraining) to estimate the temperature cannot be trained for 86% of homes using 75 days of training data inside a given home? (d) Can we transfer RC models across seasons? and even when it is trained it cannot be transferred to similar homes, (e) Does a Bayesian neural network offer any advantage over a we turn our attention to other RC models. standard neural network in the estimation of RC model parameters? 6.1.2 Selecting the order of the nRnC model. We investigate if (f ) How does the BNN-RC model perform compared to black-box higher-order models can give us a better result for most homes. We models? build four different RC models , namely 2R2C, 3R3C, 4R4C, and 5R5C, and estimate their parameters using BNNs. We use 75 days 6.1 Selecting the order of RC models of data for training the model and test it on the remaining 15 days. Determining the most appropriate order of RC models is essen- As shown in Figure 4 the 2R2C model outperforms higher-order tial. Higher-order models can represent a more complex building models in terms of RMSE obtained in the temperature prediction structure, yet identifying such models requires more parameters task. Considering the 8,884 homes in our dataset, 2R2C yields a to be estimated by BNN and increases the amount of training data lower average and a lower median RMSE. Moreover, there are fewer needed to train the BNN. In contrast, lower-order models can be homes with an RMSE that is outside the interquartile range (IQR) learned easily, but they may not accurately represent the tempera- multiplied by 1.5. The RMSE density of these ‘outliers’ is depicted in ture dynamics inside the building. To decide on the order of the RC the right panel of Figure 4. Although 4R4C and 5R5C give slightly model, we compare the ability of 5 different RC models to estimate lower RMSE values than 2R2C for a small number of homes, their in- the indoor temperature for all homes in our dataset. terquartile range is much wider than that of 2R2C and their average RMSE is larger too. 6.1.1 Establishing 1R1C models. As discussed in Section 5, iden- Next we examine the RC model that attained the lowest RMSE tifying a 1R1C model is different from other nRnC models as the in each home. We find that the 2R2C model gives the lowest RMSE parameters can be estimated directly by solving a non-negative least squares problem. We use 75 days of data to solve the optimiza- in 3,940 homes (44.35% of homes in our dataset), while 3R3C, 4R4C, tion problem and then test it on the remaining 15 days in the same and 5R5C models give the lowest RMSE in 2393 homes (26.94%), season. 1525 homes (17.17%) and 1026 homes (11.55%), respectively. This Our result indicates that in 7,698 homes (out of the 8,884 homes indicates that the best-fit RC model is different for each home, but if in our dataset), the obtained parameters are invalid, i.e., the R or C we are to develop RC models of the same order for all homes in our dataset we should pick the 2R2C model which is more parsimonious value is zero. This implies that a meaningful 1R1C model cannot be and gives a higher prediction accuracy on average than the other built for the majority of homes. We attribute this to the fact that models. Note that the main reason for following this one-size-fit-all the building envelope cannot be accurately modelled by a single approach is that it allows for transferring a model built for one home thermal resistor. Interestingly, the 1R1C model achieves a very low to any other home in the dataset as we do not need to change the RMSE in the remaining 1,186 homes. The average RMSE is 0.22 BNN architecture. Transfer learning is useful when sufficient smart degree Fahrenheit if we consider these homes only. In 120 homes, thermostat data is not available for the target home to determine the obtained 1R1C model could estimate the indoor temperature the proper order of the RC model and build it from scratch. In the with 100% accuracy. According to the metadata, over 71% of such homes have the maximum typical occupancy of 2 people or less, We do not consider nRnC models when n > 5 because it becomes more difficult to around 90% of them have one heating stage and one cooling stage train a BNN for estimating parameters of these models and that they are not suitable only, and around 64% of them have a floor area of less than 2,000 for transfer learning. arXiv preprint, 2020 Hossain et al. Figure 5: Accuracy of BNN-2R2C trained using different amounts of data. following we discuss how much data is needed to train an accurate Figure 6: Diminishing return for SSE. BNN-2R2C model and how this model can be transferred to other homes or to different seasons. where K is the total number of cluster, n is the number of members 6.2 Finding the amount of data needed to train of the k th cluster, x is the i th member of the k th cluster, and µ i,k k BNN-2R2C models is the mean of the k th cluster. We now explore how much data is needed to train a BNN-2R2C We use the elbow method to determine the number of clusters. that has an acceptable RMSE value. To this end, we build the BNN- We plot the diminishing return for the SSE values which can be 2R2C models using one day, one week, and 75 days of training written as: data and test them on the remaining days of the same season. As SSE − SSE κ+1 κ d = × 100 can be seen from Figure 5, the models built with 75 days of data SSE performs noticeably better than the other two cases. In particular, As it can be seen from Figure 6, the comparative decrease becomes the models built with 75 days of training data have an average RSME flat after 8 clusters. Therefore, we set the number of clusters to 8 of 0.50 compared to average RSME of 5.05 and 48.16 obtained for and assign every home to its corresponding cluster. Since k-means models built with 7 days and 1 day of training data, respectively. We calculates the cluster centre based on the arithmetic mean, the attribute this to the fact that the BNN needs to learn 11 parameters centre does not necessarily represent a real home. We identify the and 1 bias for the 2R2C model, hence it needs more than 1 day of closest real homes to the k-means cluster centres and treat them training data. Unfortunately, 75 days (or even 7 days) of training as representative homes of their clusters for transfer learning. We data is not readily available for some homes, especially the ones transfer the BNN-2R2C model of the representative home of each that have recently installed a smart thermostat. This motivates us cluster which is trained using sufficient training data (75 days) to to take advantage of transfer learning to reduce the amount of other members of that cluster which we refer to as target homes. training data needed for building an accurate 2R2C model. It can be seen from Figure 7 that the performance of the trans- ferred BNN-2R2C model is comparable with that of the BNN-2R2C 6.3 Transfer learning across homes model that could have been built from scratch if 75 days of smart As 75 days of training data may not be readily available in a home, thermostat data was available from the target home. The average we try to identify some generic BNN-2R2C models that can be trans- RMSE is 0.54 degree Fahrenheit when we transfer directly, i.e., we ferred to similar homes. We hypothesize that homes with similar do not retrain the transferred source model using time series data floor areas which are constructed around the same time should from the target home. We get a slightly better average RMSE of 0.51 have similar model parameters given that they are all located in the if we use 1 day data from target homes to retrain the transferred same country. This is because the heat capacitance of the indoor model. That said, even without adaptation, we get almost the same space shows a strong correlation with the size of the home, and average RMSE as we got if we had 75 days of training data. This insulated glazing and multi-layer wall structure are more common suggests that we can simply transfer the thermal model of the rep- among newly built homes. Based on these observations, we cluster resentative home of that cluster to a home that recently installed a homes according to their floor area and age. The advantage of us- smart thermostat and use this model for optimal control from the ing metadata for clustering is that they are available, and once the first day that the smart thermostat is installed. clusters are formed, they can be used to immediately assign a new home to one of these clusters based on some distance measure. 6.4 Transfer learning across seasons We use k-means clustering and utilize the features mentioned Our experiments show that the RC model trained in one season above. We run the clustering algorithm on 8,884 homes starting does not typically achieve the same level of prediction accuracy with 2 clusters and increasing it to a maximum of 30 clusters. We when used in another season. This is in line with what has been find that the sum of squared error (SSE) gradually decreases as the reported in [32] and can be attributed to latent variables that we did number of clusters increases. The SSE can be defined as: not capture in our model or changes in the effective RC parameters K k Õ Õ from one season to another. Nevertheless, it is imperative to update SSE = (x − µ ) i,k k or retrain the thermal model over time. To this end, we transfer the i=1 k=1 Identifying Grey-box Thermal Models with Bayesian Neural Networks arXiv preprint, 2020 Figure 7: Comparison of RMSE when the model is trained for each home using 75 days of training data and when the pre-trained model of the representative home of the respec- tive cluster is transferred to each home with and without Figure 9: Comparison of RMSE distributions of BNN-2R2C retraining. and baseline black-box models. uses point estimates unlike BNN) is also capable of estimating the parameters in Equation (6). To justify the need for a BNN to learn the model parameters, we estimate parameters of a 2R2C model using a neural network; we refer to this model as NN-2R2C. Our experiments suggest that this model performs poorly compared to the BNN-2R2C model. In particular, assuming that both models are trained 75 days of smart thermostat data, the average RMSE is 25.87 for NN-2R2C and 0.50 for BNN-2R2C. The uncertainty intro- duced by the BNN-2R2C model addresses the overfitting problem. Figure 8: Comparison of RMSE distributions when transfer- Moreover, the Bayesian approach offers other advantages when it ring models to the summer season. comes to building an RC model, e.g., it enables us to incorporate the prior knowledge regarding the model parameters when estimating model parameters in the target domain. individual and cluster representative models trained for winter to We also benchmark the BNN-2R2C model with other black-box summer to evaluate its performance. models introduced in Section 5, including ARIMAX, LSTM, RF, The homes in our dataset have different lengths of time series and DNN. All these models are trained using 75 days of data and data and not all of them include data from both seasons. Thus, in tested using 15 days of data from the same season. Figure 9 shows this experiment we only consider homes that had data for both the performance of all models when used to predict the indoor summer and winter. We consider two scenarios. The rst fi scenario is temperature. As it can be seen, the BNN-2R2C outperforms all black- where the winter model of the representative home of the respective box models and has an average RMSE of 0.50. The ARIMAX model cluster is transferred and used (with and without retraining) as the is the second best model with an average RMSE of 1.08 and a narrow summer model in the target home. The second scenario is where spread of RMSE values. The BNN-2R2C is superior to the ARIMAX the winter model developed for the target home using 75 days of model (in terms of RMSE) in 94.4% of homes. LSTM, RF, and DNN data is transferred to summer with and without retraining. In both give an average RMSE of 2.59, 2.56, and 2.59, respectively. This scenarios, time series data from 1 day in summer is utilized when the shows the efficacy of the proposed method for building accurate model is retrained. Data from the remaining days in the summer thermal models. season is used to test the models. Figure 8 shows that in both scenarios, we obtain mean RMSE of 2.25 and 0.51 without retraining, 7 CONCLUSION which is markedly high in the case that models of representative This paper studies the problem of identifying grey-box thermal mod- homes are transferred. The mean RMSE and its variance decrease els (RC-network models) with Bayesian neural networks leveraging considerably using 1 day of retraining data from summer. This time series data generated by smart thermostats and metadata about shows that retaining is necessary when transferring across seasons, the homes. These models have superior performance in estimating especially if the transferred model does not belong to the same home. the indoor temperature; thus, they are suitable for model-based Another observation is that if adaptation is performed using 1 day control of heating and cooling equipment. We argued that since of data from the target home, transferring the pre-trained winter building accurate grey-box models requires at least several days model of the representative home is as effective as transferring the of training data, a library of pre-trained thermal models from rep- accurate winter model that is trained for the same home. resentative homes can be built and one model from this library can be chosen and transferred to the target home to achieve high 6.5 Comparing RC models and black-box accuracy. The representative homes are selected via clustering of models the metadata that is available in our dataset. Bayesian neural network is not the only model that can estimate the Using real data collected by ecobee smart thermostats installed parameters for the RC model. A standard neural network (which in over 8,000 homes in Canada, we investigated which order of arXiv preprint, 2020 Hossain et al. the developed RC model can better describe the heat flux in the on Urban Building Energy Sensing, Controls, Big Data Analysis, and Visualization. 23–32. many homes in our dataset. We found that on average the 2R2C [16] Hao Huang et al. 2015. A neural network-based multi-zone modelling approach model can perform better than other nRnC models. We showed for predictive control system design in commercial buildings. Energy and buildings 97 (2015), 86–97. that it is crucial to use BNN to estimate the parameters of the RC [17] Michael Kane et al. 2014. Comparison of ARIMA and Random Forest time series model, and compared the performance of the BNN-2R2C model models for prediction of avian influenza H5N1 outbreaks. BMC bioinformatics with various black-box thermal models proposed in the literature. 15, 1 (2014), 276. [18] Sooyoung Kim et al. 2014. Performance evaluation of artificial neural network- Furthermore, we explored the idea of transferring the BNN-nRnC based variable control logic for double skin enveloped buildings during the model across seasons in the same home and across homes that have heating season. Building and environment 82 (2014), 328–338. similar characteristics. Transfer learning can greatly reduce the [19] Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014). need for training data and would ensure achieving higher accuracy [20] Wilhelm Kleiminger et al. 2014. Predicting household occupancy for smart heat- targets in estimating the indoor temperature. ing control: A comparative performance analysis of state-of-the-art approaches. Energy and Buildings 85 (2014), 493–505. One limitation of this work is that we cannot uniquely identify [21] Philippe Lauret et al. 2008. Bayesian neural network approach to short time load the R and C parameters of an nRnC model given the compound RC forecasting. Energy conversion and management 49, 5 (2008), 1156–1166. parameters estimated by the Bayesian neural network. We have [22] Tao Lu and Martti Viljanen. 2009. Prediction of indoor temperature and relative humidity using neural network models: model comparison. Neural Computing to at least know the true value of Q or Q to solve for R heat cool and Applications 18, 4 (2009), 345. and C , but this information is not included in the ecobee dataset. [23] Mehdi Maasoumy et al. 2014. Handling model uncertainty in model predictive Should we know the amount of heat flux from the HVAC system, control for energy efficient buildings. Energy and Buildings 77 (2014), 377–392. [24] Leopold Mba et al. 2016. Application of artificial neural network for predicting these parameters can be uniquely identified and possibly utilized to hourly indoor air temperature and relative humidity in modern building in humid conduct virtual energy audits, detect faults, and offer energy saving region. Energy and Buildings 121 (2016), 32–42. [25] Jin Woo Moon et al. 2009. Application of ANN (artificial-neural-network) in recommendations. residential thermal control. In Proc. 11th International IBPSA Conference, Glasgow, In future work we plan to incorporate the grey-box models built Scotland. 27–30. for the homes in our dataset to implement various control algo- [26] Jin Woo Moon et al. 2013. Development of an artificial neural network model based thermal control logic for double skin envelopes in winter. Building and rithms. Using a co-simulation platform, we will simulate the result- Environment 61 (2013), 149–159. ing control policy (i.e., adjust the temperature setpoints over time) [27] Jin Woo Moon and Sung Kwon Jung. 2016. Algorithm for optimal application to calculate the HVAC energy use and study thermal comfort. This of the setback moment in the heating season using an artificial neural network model. Energy and Buildings 127 (2016), 859–869. enables us to compare the BNN-nRnC model with other competing [28] G Mustafaraj et al. 2010. Thermal behaviour prediction utilizing artificial neural thermal models in terms of potential energy savings and impact on networks for an open office. Applied Mathematical Modelling 34, 11 (2010), 3216–3230. occupant comfort. [29] Giorgio Mustafaraj et al. 2011. Prediction of room temperature and relative humidity by autoregressive linear and nonlinear neural network models for an open office. Energy and Buildings 43, 6 (2011), 1452–1460. REFERENCES [30] Nest Labs. 2015. Energy Savings from the Nest Learning Thermo- [1] Allied Market Research. 2019. Smart Thermostat Market. https://www. stat. https://storage.googleapis.com/nest- public- downloads/press/documents/ alliedmarketresearch.com/smart- thermostat- market. Online; accessed on 7 energy- savings- white- paper.pdf . Online; accessed on 1 January 2020. February 2020. [31] Oluwaseyi T Ogunsola et al. 2014. Development and validation of a time-series [2] Gaby Baasch et al. 2019. Comparing Gray Box Methods to Derive Building model for real-time thermal load estimation. Energy and buildings 76 (2014), Properties from Smart Thermostat Data. In Proc. 6th ACM International Conference 440–449. on Systems for Energy-Efficient Buildings, Cities, and Transportation . 223–232. [32] Nilavra Pathak et al. 2019. A Bayesian Data Analytics Approach to Buildings’ [3] Charles Blundell et al. 2015. Weight Uncertainty in Neural Network. In Proc. 32nd Thermal Parameter Estimation. In Proc. 10th ACM International Conference on International Conference on Machine Learning, Vol. 37. PMLR, 1613–1622. Future Energy Systems. ACM, 89–99. [4] James E Braun and Nitin Chaturvedi. 2002. An inverse gray-box model for [33] SL Patil et al. 2008. Modelling of tropical greenhouse temperature by auto transient building load prediction. HVAC&R Research 8, 1 (2002), 73–99. regressive and neural network models. Biosystems engineering 99, 3 (2008), 423– [5] Cinzia Buratti et al. 2014. Building behavior simulation by means of Artificial 431. Neural Network in summer conditions. Sustainability 6, 8 (2014), 5339–5353. [34] Andrew Pollard and Albrecht Stoecklein. 1998. Occupant and building related [6] Chris Chatfield. 2003. The analysis of time series: an introduction. Chapman and determinants on the temperature patterns in New Zealand residential buildings. Hall/CRC. In IPENZ Conference 98: The sustainable city; Volume 2; Electrotechnical: simulation [7] An-Heleen Deconinck and Staf Roels. 2016. Comparison of characterisation and control; energy management: telecommunications. Institution of Professional methods determining the thermal resistance of building components from onsite Engineers New Zealand, 62. measurements. Energy and Buildings 130 (2016), 309–320. [35] Pablo Romeu et al. 2013. Time-series forecasting of indoor temperature using [8] ecobee. 2019. Savings from your ecobee. https://www.ecobee.com/savings/. pre-trained deep neural networks. In International conference on artificial neural Online; accessed on 1 January 2020. networks. Springer, 451–458. [9] Diana Enescu. 2017. A review of thermal comfort models and indicators for indoor [36] Antonio Ruano et al. 2006. Prediction of building’s temperature using neural environments. Renewable and Sustainable Energy Reviews 79 (2017), 1353–1379. networks models. Energy and Buildings 38, 6 (2006), 682–694. [10] M.A. Fayazbakhsh et al. 2015. A Resistance–Capacitance Model for Real-Time [37] Saran Salakij et al. 2016. Model-Based Predictive Control for building energy Calculation of Cooling Load in HVAC-R Systems. Journal of Thermal Science and management. I: Energy modeling and optimal control. Energy and Buildings 133 Engineering Applications 7, 4 (2015), 041008. (2016), 345–358. [11] PM Ferreira and AE Ruano. 2002. Choice of RBF model structure for predicting [38] JE Seem et al. 1989. Transfer functions for efficient calculation of multidimen- greenhouse inside air temperature. In IFAC World Congress. 91–96. sional transient heat transfer. Journal of heat transfer 111, 1 (1989), 5–12. [12] M.M. Gouda et al. 2002. Building thermal model reduction using nonlinear [39] Tianyu Zhang and Omid Ardakanian. 2019. A domain adaptation technique for constrained optimization. Building and Environment 37, 12 (2002), 1255–1265. fine-grained occupancy estimation in commercial buildings. In Proc. International [13] Frédéric Haldi and Darren Robinson. 2011. The impact of occupants’ behaviour Conference on Internet of Things Design and Implementation. 148–159. on building energy demand. Journal of Building Performance Simulation 4, 4 [40] Datong P Zhou et al. 2017. Quantitative comparison of data-driven and physics- (2011), 323–338. based models for commercial building HVAC systems. In 2017 American Control [14] V.S.K.V. Harish and Arun Kumar. 2016. Reduced order modeling and parameter Conference (ACC). IEEE, 2900–2906. identification of a building energy system model through an optimization routine. [41] Na Zhu et al. 2011. Energy performance and optimal control of air-conditioned Applied Energy 162 (2016), 1010–1023. buildings with envelopes enhanced by phase change materials. Energy conversion [15] Md Monir Hossain et al. 2019. Evaluating the Feasibility of Reusing Pre-trained and Management 52, 10 (2011), 3197–3205. Thermal Models in the Residential Sector. In Proc. 1st ACM International Workshop
Computing Research Repository – arXiv (Cornell University)
Published: Sep 13, 2020
You can share this free article with as many people as you like with the url below! We hope you enjoy this feature!
Read and print from thousands of top scholarly journals.
Already have an account? Log in
Bookmark this article. You can see your Bookmarks on your DeepDyve Library.
To save an article, log in first, or sign up for a DeepDyve account if you don’t already have one.
Copy and paste the desired citation format or use the link below to download a file formatted for EndNote
Access the full text.
Sign up today, get DeepDyve free for 14 days.
All DeepDyve websites use cookies to improve your online experience. They were placed on your computer when you launched this website. You can change your cookie settings through your browser.