A Sequential Modelling Approach for Indoor Temperature Prediction and Heating Control in Smart Buildings
AA Sequential Modelling Approach for IndoorTemperature Prediction and Heating Control inSmart Buildings
Yongchao Huang
University of Oxford [email protected]
Hugh Miles
Atamate
Ltd. [email protected]
Pengfei Zhang
Facebook, London [email protected]
Abstract
The rising availability of large volume data, along with increasing computingpower, has enabled a wide application of statistical Machine Learning (ML) al-gorithms in the domains of Cyber-Physical Systems (CPS), Internet of Things(IoT) and Smart Building Networks (SBN). This paper proposes a learning-basedframework for sequentially applying the data-driven statistical methods to predictindoor temperature and yields an algorithm for controlling building heating sys-tem accordingly. This framework consists of a two-stage modelling effort: in thefirst stage, an univariate time series model (AR) was employed to predict ambientconditions; together with other control variables, they served as the input featuresfor a second stage modelling where an multivariate ML model (XGBoost) wasdeployed. The models were trained with real world data from building sensornetwork measurements, and used to predict future temperature trajectories. Ex-perimental results demonstrate the effectiveness of the modelling approach andcontrol algorithm, and reveal the promising potential of the mixed data-driven ap-proach in smart building applications. By making wise use of IoT sensory dataand ML algorithms, this work contributes to efficient energy management andsustainability in smart buildings.
Internet of Things (IoT) technologies introduce a new generation of smart building solutions [1, 2].It embeds network platforms and advanced data analytics with the Building Management System(BMS) which integrates lighting, heating, ventilation, and air conditioning (HVAC), safety, and se-curity, turning sensory data into to actionable decisions. IoT has a wide spectrum of applicationssuch as smart facilities maintenance and efficient energy management [3], an example is using tem-perature, CO and humidity sensors to provide the feedback for HVAC systems, or using motiondetectors for lighting control. An IoT ecosystem normally consists of a comprehensive sensor net-work for data collection, communication, and networking, and data consuming solutions. MachineLearning (ML) plays an increasingly significant role in IoT decision-making (AI-IoT). The goal ofbuilding intelligence is to build an automated, intelligent, and decentralised decision-making systemthat monitors building service status and interprets them as machine-readable commands. In thissetting, sensor network sits in the centre; along with data consuming platforms, it can yield signifi-cant savings for operations and maintenance, without degrading occupants’ comfort level. Take for Workshop on machine learning for engineering modeling, simulation and design @ NeurIPS 2020 a r X i v : . [ ee ss . SP ] N ov xample the heating and cooling systems, which can account for up to 60% of a building’s overalluse [4], by efficiently learning use patterns and managing on-off controls, it can reduce energy usageby 30% as reported. This serves as the main motivation for heating control.Traditional building management solutions use complex building dynamics to achieve control ob-jectives; more recently emerging techniques such as big data analytics [5], ML [6, 7, 8] and deepreinforcement learning techniques have been deployed [9]. A typical scenario of application is tomeet a target temperature at a pre-set time. In this pursuit, we aim to decide the best time to switchon the radiator, and to switch off thereafter without degrading comfort. These are the two optimi-sation tasks: the first helps avoid wasting utility in case the valve is turned on too early; the secondgives obvious energy-saving bonus. This is particularly the case when one finishes work and asksBMS to heat up a room before actually arriving home. This paper focuses on the first task, while thesolutions provided can be safely extended to the second. The challenge is to accurately predict thewarm-up time, and we introduced a sequential, data-driven modelling approach to tackle it. Applying ML [6, 7] and big data analytics [5] to IoT (particularly smart buildings) has been anincreasingly hot area in recent years. Some previous work on modelling and controlling buildingenergy systems were summarised, for example, in [10, 11, 12, 13]. Mathematical (structured) mod-elling (also termed the ’white box’ approach) has long been used in simulating natural phenomenonsuch as heat diffusion, it’s well constructed based on physical laws of thermal dynamics, and nor-mally comes in the form of partial differential equations with some stochastic characteristics ( e.g. the state-space models describing a dynamical system). For example, [14] describes a mathemati-cal model for heating modelling in residential buildings, their results demonstrate the effectivenessof the model-based heating control strategy. Non-linear systems can also be designed by purposefor special buildings such as [15]. Mathematical models have intuitive interpretability and maturesolvers. However, they may not be as flexible as statistical models which are learning-based, data-driven, and independent of the physical system. In statistical modelling (also termed ’black box’approach), each signal is treated as an input feature, it’s thus easier to take control over the vari-ables fed in. The models are based on statistical principles which analyse the relation between inputfeatures and the output variable. Since statistical models are physics-independent, they can be uni-versally used to model various phenomena by learning from data. The disadvantage, however, isalso obvious: they depend on the partial info contained in the data used to train it. Following thedata allows fast adjustments but also introduces bias. Results from a data-driven approach can yieldgood accuracy, but care should be taken on generalisation. [16] emphasizes the importance of spe-cific tuning for each ML model used in the prediction of buildings heating and cooling loads. [13]gives a review on the use of several ML methods in forecasting building energy performance, andconcluded that all models provide reasonable accuracy given large samples and fine tuned hyper-parameters. A compromise could be combining both the ’white box’ and ’black box’ approachesto form a semi-structured modelling approach (the ’grey box’ or hybrid approach), in which knownlaws are represented by deterministic terms, and unknown effects or uncertainties can be embeddedvia learning-based models ( e.g. a neural network architecture). This bridges the gap between physi-cal and statistical modelling: it enables prior info to be embedded, the building dynamics captured,and gives a reliable description of the uncertainties [17], and results in a simpler models as comparedto pure physics based models [18].While the power of the data driven approach has been proved in many smart building settings ( e.g.
HVAC control [19, 20]), it also has some drawbacks with estimation [21] and forecasting. CommonML models,
Random Forest for example, rely on instantaneous, cross-sectional ambient info tomake inference about a response. In most situations, however, collection of the input features maynot be in a timing manner. Even if the data can be collected and processed in near real-time, thereis inevitably a delay in forecasting, i.e. the algorithm can only make inference about the currentstate where surrounding information is provided. Therefore, these methods are more suitable forinferencing a currently unobservable response variable rather than making swift predictions. Thegame changes when future information becomes available. Future information itself can be predictedin a structured way if it follows some dynamics, and in sensor network analysis particularly, mostsignals are time series sequences and there are mature techniques well suited for forecasting. Forexample, there are analytical forms such as an autoregressive process, and architectures such as2n
LSTM network for analysing sequential data. This modelling approach serves as a pathway toobtaining future information about the quantity of interest. Once the future information is estimatedto certain accuracy, they can be fed into common ML models to make predictions for the responsevariable. This motivates the two-stage forecasting framework.
We proposed a two-stage modelling procedure for household indoor temperature prediction andheating control. The framework was shown in Fig.1. It’s a mixed data-driven method which com-bines input control and learning. The sensory data were classified into two categories: control andnon-control variables. In the first stage, each non-control variable time series was treated as a uni-variate process and time series models ( e.g.
AR) were employed to make predictions for a futureperiod; in the second stage, the control variables (with future values set according to some controlrules such as on/off), together with the non-control variable predictions, were fed into a multivariateML model (pre-trained using historically measured data) to yield final predictions for the outputquantity (indoor temperature). In later contexts, the first practice was labelled ‘ambient conditionsmodelling’, while the second was termed ‘indoor temperature modelling’.Figure 1: The two-stage sequential modelling procedure
A BMS is a comprehensive network enabling data flow between different functional componentsand making decisions based on pre-defined rules and algorithms. The front ends are sensors, whichmonitor and detect changes in environmental and device variables such as temperature, air quality,valves and occupancy to optimise a building’s operation services like HVAC over time. Data wascollected by a sensor unit which accommodates an array of sensors such as temperature, lighting,fan, radiator meters, Passive Infrared motion (PIR) sensor, rain detector, and relative humidity (RH)reader. It flows into data containers and processors where data storage, aggregation, transformation,inference and prediction happen. The challenge is to deploy the system in a real-time environmentat scale. For example, data may be collected overtime in an augmented manner so that time-variantadjustments can be made. Retraining a predictive model could be either done online (streamingdata) or offline (batch data). For large learning tasks, transfer or incremental learning could beleveraged via a de-centralised data consuming architecture. These put requirements on both thephysical and software sides. For example, sensors may need to talk to each other wirelessly, dataneeds to be transported either locally ( e.g. offline training) or on cloud, and processors may needto communicate for collaboration purpose. Terminal devices such as water tanks, controllers ( e.g .radiators) and air conditioners need to be plugged in BMS to enable an state-action feedback loopfor services control. In addition, extensibility is desired so that new sensors, gateways (5G highwayfor example) and devices, can be flexibly added, upgraded and integrated into the existing BMS.Sensors play a key role in engineering scenes such as Cyber-Physical Systems (CPS), IoT and SmartBuilding Networks (SBN). Thanks to the increasing capacity of sensor network, large volume ofdata becomes available for analysis. Making wise use of sensory data could potentially improveenergy use and achieve better building management. In this research, multiple functional sensorswere installed in a house environment (Fig.2) to monitor quantities such as temperature, lighting,fan, heating, occupancy, rain and humidity statuses. In total there were 54 sensors used. The details3f the set of sensors deployed were shown in Table.1 and Table.2. These sensors together formed anetwork. Table 1: Sensor categories
Sensor type Sensor code (functionality abbreviations)
Sensory data type Example instance(s)
Temperature TMP numeric
Lighting AMB/LTC Boolean(0/1)
Fan FAN Boolean(0/1)
Radiator HTC/HTV Boolean(0/1)
Occupancy PIR/EPIR/OPC Boolean(0/1)
Rain RAIN Boolean(0/1)
Relative humidity HYGR numeric ∗ Sensor naming convention: the first number implies type of space, i.e. 0 refers to outside space and 1 is house. Thesecond number gives the zone area in the building (see the floor layout plan). The strings are sensor code, followed bysensor number.
Table 2: Physical mapping of sensor layout
Zone area of sensor location Sensors deployed
Figure 2: Building floor layout (left: Floor 0, right: Floor 1)
The data was sliced from a large dataset spanning over the whole year of 2019 in the house; onlyDecember data (‘2019-12-01 00:00:00’ to ‘2019-12-30 14:39:00’) was chosen because it demon-strates the use of heating. The raw data consists of 44640 datapoints and 54 sensor signals. Priorto formal modelling, data pre-processing was performed to clean data and perform exploratory dataanalysis.In this task, we aim to predict the indoor temperature (the response variable) in the master bedroomsitting on the first floor, i.e.
Random Forest (RF) model was trained and the resulted permutation importance wascalculated to rank the 53 raw features, the top 15 ranked features are shown in Fig.4. As expected,the downstairs room temperature 1-8-TMP1 is important in predicting the response variable, i.e. thetemperature in bedroom 1-15 (1-15-TMP1). Other close friends are 1-14-TMP1 and 1-13-HTV1:the next-door temperature and the status of the valve controlling underfloor heating of Room 1-15,respectively. A roughly monotonic relation between 1-14-TMP1 and 1-15-TMP1 is observed inFig.4. Besides, 1-15-TMP1 also exhibits some patterns with certain weekdays and hours in a day.Based on the importance analysis, the top 15 features, as listed in Fig.4, plus the remaining weekdayfeatures, which totals 19 variables, were selected. As stated, these 19 features are classified intotwo categories: control and non-control variables. In a first trial, the underfloor heating valve status(1-13-HTV1) was employed as the single control variable. A second trial also included the next-door temperature 1-14-TMP1 and downstairs temperature 1-8-TMP1 as extra control variables forachieving more aggressive control. The rest 18 (16 in the second case) variables served as non-control features.Figure 4: Left: top 15 RF-based permutation importance, right: correlation between 1-14-TMP1and 1-15-TMP1 5 .2 Ambient Conditions Modelling
The first step in the two-stage modelling practice is to model the non-control variables, i.e. the ambi-ent conditions. All the 18 features, except for weekdays and hours, need to be learned from historyand predicted for a future period. Each variable trajectory was treated as a single univariate timeseries and modelled using autoregressive regression (AR). AR is a mature technique for modellingsequential data, the general form of an AR(p) model can be described as X t = µ + α X t − + α X t − + ... + α p X t − p + (cid:15) t (1)where X is the time series, µ is the constant offset, and (cid:15) is the error term. α s are the AR parametersto be estimated using historical realisations. For stationary signals, α lies in (-1,1).We iterated through all predictable non-control features using original scale values. By trial anderror, a uniform, absolute threshold of 0.1 was employed for partial autocorrelation function ( pacf )to identify all statistically significant lags for each AR model (Fig.5c). All AR models were trainedwith the training dataset which spans from ’2019-12-01 00:00:00’ to ’2019-12-30 14:39:00’, whilethe test set, which spans from ‘2019-12-30 14:40:00’ to ‘2019-12-31 23:59:00’, was reserved forcomparison purpose. The time span for future prediction was from ‘2019-12-30 14:40:00’ to ‘2019-12-30 18:00:00’ (totalling 201 datapoints). The future prediction scope partially overlapped withthe test set horizon. They were however used for different purposes: the future prediction was usedto test control variables and the control strategy, while the test set was mainly designed for eval-uating model performance. In the AR model, a rolling forecasting strategy was used, in which asingle prediction was made at each time step, and the prediction was appended to historical data toincrementally augment the training size. An example of the predicted trajectories versus the real testsignal was shown in Fig.5a and Fig.5b, also accompanied is the control variable 1-13-HTV1. Theunderfloor heating valve history roughly coincides with the indoor temperature trajectory, showinga delayed and less evident influence on 1-14-TMP1. The estimated AR model gives reasonably wellpredictions agreeing with the average trend due to its memory characteristic, although deviations ex-ist, we will see later in the boosting tree case that it doesn’t affect the indoor temperature predictionsto a significant level, as trees are grown in a way such that only interval values matter due to thenature of tree splitting. The control variable 1-13-HTV1 was binary and takes values 0/1 (off/on).Its value was set to 1 for all future timestamps, representing a constant on status, which is in linewith our goal of warming up the room. (a) Historical and controlled/predicted values for and (b) predictions (zoom-in of (a) on date 30/12/2019) (c) pacfs vs lags (selected: 1, 2, 3, 7) Figure 5: Ambient conditions modelling6 .3 Indoor Temperature Modelling and Control
In the second stage, all historical and forecasted future feature values were fed into a multivariateML model to produce predictions for the response variable 1-15-TMP1 on the same future scale(‘2019-12-30 14:40:00’ to ‘2019-12-30 18:00:00’). Before building the ML model, a standardizationprocedure was applied to scale the numeric variables in Table.1. This is useful when distance-basedmethod ( e.g. nearest neighbors) or dimension reduction techniques are to be used. For simplicity, weonly use one model, i.e.
Extreme Gradient Boosting Machine (XGBoost), to demonstrate the secondstage modelling process. We first described the case where single control variable (1-13-HTV1) wasused; two extra control variables, i.e. max.
50 trees, each tree with a max. depth of 4, learning rate of 0.3, row and columnsubsampling rates of 0.6 and 0.8, respectively. The trees essentially meshes the feature space andproduces intervals for each variable. As an example, part of the first tree boosted is shown inFig.6. The model sequentially and frequently employs 1-8-TMP1, 1-14-TMP1, 1-13-HTV1, etc . tosegment data instances, which is consistent with the feature importance ordering in Fig.4.Figure 6: Depth-two diagram of the first boosted tree generated by XGBoost algorithmOn a future time scale, four control experiments were designed, as shown in Table.3. Three featureswere employed as control variables: 1-13-HTV1, 1-14-TMP1, and 1-8-TMP1. They are closelyrelated to 1-15-TMP1, as evidenced in Fig.3 and Fig.4, thus may give hints on inferring 1-15-TMP1.Other non-control variable values were directly obtained from the first stage AR predictions. In allthe four cases, control(s) were applied starting at the time 16:00 and onwards. The experimentsresults were presented in Fig.7. The curves were smoothed using a 20-min moving average.Table 3: Control experiments design case with control no control ◦ C, 1-8-TMP1 = 22 ◦ C 1-13-HTV1 = 0, 1-14-TMP1 = 26.5 ◦ C, 1-8-TMP1 = 22 ◦ C2 1-13-HTV1 = 0, 1-14-TMP1 = 26/27 ◦ C, 1-8-TMP1 = 22 ◦ C 1-13-HTV1 = 0, 1-14-TMP1 = 26.5 ◦ C, 1-8-TMP1 = 22 ◦ C3 1-13-HTV1 = 0, 1-14-TMP1 = 26.5 ◦ C, 1-8-TMP1 = 23 ◦ C 1-13-HTV1 = 0, 1-14-TMP1 = 26.5 ◦ C, 1-8-TMP1 = 22 ◦ C4 1-13-HTV1 = 0/1, 1-14-TMP1 = 27 ◦ C, 1-8-TMP1 = 23 ◦ C 1-13-HTV1 = 0, 1-14-TMP1 = 26.5 ◦ C, 1-8-TMP1 = 22 ◦ C Fig.7(a) shows the effect of control of 1-13-HTV1. The valve remained silent first ( e.g. setting thevalue of ‘1-13-HTV1’ to be 0), till a certain chosen timestamp ( e.g. ‘2019-12-30 16:00:00’) it wastriggered. The differences between the controlled and uncontrolled scenarios are evident: generally,there is an temperature up lift of ∼ ◦ C after valve control was applied, which evidences that thecontrol is effective. This is justified by the fact that 1-13-HTV1 controls the underfloor heating ofthe large Room 1-15 (although the valve sits in the wardrobe). Bringing in 1-14-TMP1 (temperaturein the ensuite bathroom) as a control variable can also be justified: the bathroom has smaller spaceand it accommodates some fast heating devices ( e.g. electric radiators, heated towel rail, infraredheating panel and fan) which facilitate control. Fig.7(b) presents the results obtained by raising ordecreasing the temperature in Room 1-14: cooling down the room seems to have greater effect on1-15-TMP1 than heating it up. Fig.7(c) assumes we are able to swiftly control the temperature inReception1 (1-8-TMP1, downstairs of 1-15-TMP1), and we can see the XGBoost model accuratelycaptured the collaborative relation between these two variables: by increasing 1-8-TMP1 from 22 ◦ Cto 23 ◦ C, a jump of approximately 0.4 ◦ C was induced in 1-15-TMP1. This also shows 1-8-TMP1 has7reater impact than 1-13-HTV1 and 1-14-TMP1 in influencing 1-15-TMP1, as has been previouslyobserved from Fig.4. The biggest impact may result from a hybrid control strategy combining all ofthe three, by which we expect to achieve faster, even though more aggressive heating control. Theresults yielded in Fig.7(d) are in line with our expectations: a jump of 0.6 ◦ C was witnessed.The temperature changes (0.1 ∼ ◦ C) in Fig.7, induced by the changes in the values of input controlvariables, are small but evident, implying that the XGBoost model is very sensitive in detecting vari-ations of input controller values, and can act correspondingly to generate differentiated responses.The change is not huge as the Room 1-15 is spacious, and the influence of 1-14-TMP1 and 1-8-TMP1 are indirect. These patterns of course are obvious from a human’s perspective; however, itbecomes appreciable when they are identified by a learning algorithm without human intuition. (a) case 1: control of 1-13-HTV1 (b) case 2: control of 1-14-TMP1(c) case 3: control of 1-8-TMP1 (d) case 4: hybrid control of 1-13-HTV1, 1-14-TMP1 and 1-8-TMP1
Figure 7: Indoor temperature predictions by XGBoost, with single or hybrid controls
With the capacity of predicting future temperature trajectories, we can calculate the warm-up timeneeded to hit a target temperature, when the heating valve is switched on at an arbitrary timestamp.The inverse problem being, if we want the temperature to be at or above a desired value at certaintimestamp, when should the heating valve be switched on? A brute-force strategy is to first predictambient conditions over the entire period from the current time to that timestamp, and work out theclosest starting point that leads to an exact match or close surpass of the target temperature at thatdesired timestamp. If the time gap between now and the desired point is small (during which theambient conditions may not necessarily change much), another simpler strategy is to work out thetime needed for the indoor temperature to arise to the target temperature if we switch on the valve atany future preferred time, and do an equal-length shift to the desired timestamp by assuming that theambient conditions remain stationary or static over this time period. For simplicity, the later staticapproach was demonstrated using the hybrid control experiment (case 4 in Fig.7).We simulated a scenario to achieve heating control using the predicted indoor temperature trajectory.Imagine the following heating event was pre-set on a calendar ( e.g. pre-saved as .ical or .yml files): Event: HeatingTarget Zone: Room 1-15Target Timestamp: 2019-12-30 18:00:00Target Temperature: 22.40 ◦ CControls: 1-13-HTV1, 1-14-TMP1, 1-8-TMP1
Suppose the current time is ‘2019-12-30 16:00:00’, so we want the temperature in Zone 1-15 to beat least, ideally around, 22.40 ◦ C at 18:00 tonight. The task being, when should the heating valvebe switched on to optimize utility use. The task fails if the temperature is below that at the desiredtime (the valve responds late); a waste of energy occurs if the target temperature is achieved earlier8 lgorithm 1
Learning-based heating control
Require: historical data, pre-set future event
Step 1: Initialization
Scan future events on calendar and obtain the future event timestamp t (‘2019-12-30 18:00:00’) and target temperature T(22.40 ◦ C).
Step 2: Ambient conditions forecasting
1. use a sliding window to retrieve historical data (both ambient conditions and indoor temperature).2. train a univariate model ( e.g. time series or
LSTM ) to forecast future values for the selected features, spanning the time gapbetween now and t . Step 3: Indoor temperature prediction
1. from now on, choose a before-event point t ( e.g. now) as simulation starting point.2. set the control variables (single/hybrid control) to their desired values ( e.g. set valve status = on).3. use historical ambient conditions and indoor temperature to train a multivariate, supervised ML model ( e.g. XGBoost or neuralnetwork).4. plug in the forecasted future ambient conditions to predict the indoor temperature trajectory from t to t . Step 4: Warm-up time estimation and heating control
1. Smooth the predicted value ( e.g. using moving average), record the first timestamp t when the predicted indoor temperaturehits the target temperature T and remains above for a consecutive period.2. calculate ∆ t = t − t as the warm-up time.If ∆ t ≥ t − now , trigger the heating at current time. If ∆ t < t − now , either set the control start time as t − ∆ t (staticapproximate control, assuming time-invariant characteristics of future ambient conditions), or go back to Step 3 and repeat theprediction-estimation process (dynamical, iterative search-based control).3. a buffer time can be further deducted from the control start time to allow extra flexibilities (sufficient warm-up time to furtherguarantee the target temperature is achieved).4. finally, set the control variables to be the desired values at the derived control start time. (the valve responds too early). The ideal case is for the temperature to arrive around the targettemperature at the specified time. The complete control procedure was described in Algorithm.1.Algorithm.1 was implemented using the hybrid control strategy blending variables 1-13-HTV1, 1-14-TMP1 and 1-8-TMP1. We tentatively switch on the heating valve (1-13-HTV1=1), set 1-14-TMP1= 27 ◦ C and 1-8-TMP1 =23 ◦ C at ‘2019-12-30 16:00:00’, and ask the trained XGBoost modelto predict how they influence the indoor temperature trajectory 1-15-TMP1. It is observed fromFig.7(d) that, given the prior temperatures 1-14-TMP1=26.5 ◦ C and 1-8-TMP1=22 ◦ C, the targetroom temperature 1-15-TMP1 was predicted to be about 21.9 ◦ C initially. After heating controlswere applied, the temperature started to rise and hit the target temperature 22.40 ◦ C at around 16:41and stay above it for a consecutive period of over 20 mins (this user-defined consecutively over-timedefines a rule to find whether the target temperature has been achieved or not). This gives a warm-uptime of 19 mins, i.e. given the future predicted ambient conditions, if we adopt the hybrid controlstrategy, it takes 19 mins to arrive at the target temperature 22.40 ◦ C. As the estimated warm-uptime (19 mins) is less than the two-hour gap between the current time (‘2019-12-30 16:00:00’) andthe desired timestamp (‘2019-12-30 18:00:00’), if ambient conditions stay approximately stationary,then we can assume the warm-up time is time-invariant, and a static approximate control strategyadvises to switch on the heating valve at time ‘2019-12-30 17:41:00’; if we want to give moreconfidence in achieving the target temperature at the desired time, an extra 5 mins buffer time canbe added, and a safe switch-on time would be ‘2019-12-30 17:36:00’. Alternatively, if variations offuture ambient conditions are to be considered, this prediction-estimation process can be repeatedlyexercised, starting at any arbitrary time closer to the desired timestamp to yield a potentially moreaccurate estimation of the warm-up time.Fig.8 recaps the complete graph of implementing the control algorithm. The AR and XGBoost mod-els were first trained using data from ‘2019-12-01 00:00:00’ to ‘2019-12-30 14:39:00’. As statedbefore, even though the AR model does not give perfect accuracy, still in the second stage mod-elling, the performance of the XGBoost model was very impressive: given real measured ambientconditions, the indoor temperature predictions (green and magenta dots) agree very well with groundtruths (blue dots) on both the training (in-sample) and test (out-of-sample) datasets. To implementthe hybrid control strategy, we manually masked the real ambient conditions data from ‘2019-12-3014:40:00’ to ‘2019-12-30 18:00:00’ (the ‘future prediction’ scope), and asked the previously trainedAR model to forecast the ambient conditions over this period, then XGBoost model was employedto further predict the indoor temperature 1-15-TMP1, with the hybrid control strategy applied. Theresults (red dots) provide sound advice on the most appropriate time to switch on the heating valves.Note that, the future predictions (red dots) are not directly comparable to the ground truth (bluedots), as the prior ambient conditions are different.9igure 8: Learning-based heating control implementation
The two-stage modelling practice is an example of empirical illustration. Once the two-stage frame-work is built, it can generalises to other building environments with flexible sensor inputs. In suchcases, a model calibration process, along with feature selection, can be performed based on local datalearning, which ensures models trained in one building can quickly adapt to new environments viapathways such as transfer learning. While AR is a mature technique in time series modelling, otherpotential methods could be explored. Among them, good candidates are those capable of sequen-tial data processing, these include ARIMA, HMM and LSTM. We have researched these methodsand they provide very good performance over AR; however, they also consume more computing re-sources and pose more requirements on real-time computing. The performance of the second stagelearning, i.e. the XGBoost, was impressive. This is built on the success of feature forecasting inthe first stage. Both stages have different levels of error tolerance, due to the nature of tree-basedmodels (weak learners are ensembled sequentially into a strong learner), good performance wasachieved. It would be good practice to compare this two-stage modelling approach with other mod-elling techniques such as vector AR (VAR) which directly predicts the future using multiple timeseries inputs.The control strategy developed answers the question of when to turn on the control variables; how-ever, it can be used to decide when to turn off as well, by repeating the same algorithm in thepost-control stage ( i.e. after the target temperature has been achieved), intervally turning of someof the control variables and checking the resulting future temperature trajectories. So it’s a solutionfor both tasks. An intelligent BMS can automatically manage different operation variables such asradiators, lights, water tanks, air conditioners, and even appliance. Theoretically, adding in morepositively correlated (with the response variable) control variables could benefit the control strategy,and accelerate the heating process. This was demonstrated in the case where the heating valve 1-13-HTV1, the next door temperature 1-14-TMP1, and downstairs temperature 1-8-TMP1, join togetherto form a hybrid control strategy. However, more investigations are needed to study the join effectof mixed control variables.The significance of this work lies in following aspects: first, it builds up an automated, AI-powered,data-driven forecasting system for heating control, this framework contributes to efficient energyutilisation in line with user request, and improve sustainability in smart buildings. It learns fromenvironments (ambient conditions) and the learned patterns makes this system ‘well-behaved’ (sav-ing energy while retaining comfort level) in a real-time environment. It’s more flexible comparedto traditional rule-based control systems, also more efficient than conventional thermal dynamics-10ased mathematical models as it evolves dynamically, reducing the labor of reconstructing structuredmodels. Secondly, this work combines time series modelling and ML to form a two-stage sequen-tial modelling approach. It allows some flexibilities for control (via the set of control variables) aswell as dynamical learning from environments (ambient conditions). Time series model capturesthe temporal dependence in each feature, while ML model reveals the cross-sectional correlationsbetween features and response. This mixed temporal-spatial approach enables real time control ofthe quantity concerned.
This research introduces a sequentially pipelined, mixed data-driven approach to modelling buildingdata, from which indoor temperature patterns can be learned and optimal heating control decisionscan be made. Sensory data was collected through a wireless sensor network, information were fedinto an intelligent building management system to be analysed and to advise control actions. Thetwo-stage modelling framework sequentially connects time series and machine learning methods,and combines learning and input control. A heating control algorithm was designed to maximizeenergy utilization, and the effectiveness of the single and hybrid control strategies was demon-strated. This work contributes towards scalable, automated, and energy-efficient intelligent buildingmanagement system design and effective service control to meet human comfort expectation.
References [1] Intel. Smart buildings with internet of things technologies, 2020.[2] Eugenia Alexandropoulou Egyptiadou Chrysi K. Metallidou, Kostas E. Psannis. Energy effi-ciency in smart buildings: Iot approaches.
IEEE Access , 8, 2020.[3] K. A. Kelly, P. Sassi, and J. Miles. In-use energy performance study of automated smart homes.
SDAR* Journal of Sustainable Design & Applied Research , 7(1), 2019.[4] Morgan Stanley. Green buildings power savings and returns, Jun 2017.[5] Wenzhuo Li, Choongwan Koo, Taehoon Hong, Jeongyoon Oh, Seung Hyun Cha, and Sheng-wei Wang. A novel operation approach for the energy efficiency improvement of the hvacsystem in office spaces through real-time big data analytics.
Renewable and Sustainable En-ergy Reviews , 127, Jul 2020.[6] Luis M.Candanedo and Véronique Feldheim. Accurate occupancy detection of an office roomfrom light, temperature, humidity and co2 measurements using statistical learning models.
Energy and Buildings , 112:28–39, Jan 2016.[7] Manar Amayria, Abhay Arorab, Stephane Ploix, Sanghamitra Bandhyopadyay, Quoc-DungNgo, and Venkata Ramana Badarlab. Estimating occupancy in heterogeneous sensor environ-ment.
Energy and Buildings , 129:46–58, Oct 2016.[8] Ján Drgoˇna, Damien Picard, Michal Kvasnica, and Lieve Helsen. Approximate model predic-tive building control via machine learning.
Applied Energy , 218:199–216, May 2018.[9] Ruoxi Jia, Ming Jin, Kaiyu Sun, Tianzhen Hong, and Costas Spanos. Advanced buildingcontrol via deep reinforcement learning.
Energy Procedia , 158:6158–6163, Feb 2019.[10] V.S.K.V. Harish and Arun Kumar. A review on modeling and simulation of building energysystems.
Renewable and Sustainable Energy Reviews , 56:1272–1292, April 2016.[11] Xiwang Li and Jin Wen. Review of building energy modeling for control and operation.
Re-newable and Sustainable Energy Reviews , 37:517–537, September 2014.[12] Hrvoje Krsti´c and Mihaela Teni. Review of methods for buildings energy performance mod-elling.
IOP Conf. Series: Materials Science and Engineering , 245, 2017.[13] Ivan Glesk Saleh Seyedzadeh, Farzad Pour Rahimian and Marc Roper. Machine learningfor estimation of building energy consumption and performance: a review.
Visualization inEngineering , 6, 2018. 1114] Degurunnehalage Wathsala Upamali Perera and Nils-Olav Skeie. Comparison of space heatingenergy consumption of residential buildings based on traditional and model-based techniques.
Buildings , 7, 2017.[15] Yuguo Li Zhi Zhuang and Bin Chen. A mathematical model for a house integrated with anelevated chinese kang heating system.
Proceedings: Building Simulation , 2007.[16] Parag Rastog Saleh Seyedzadeh, Farzad Pour Rahimian and Ivan Glesk. Tuning machinelearning models for prediction of building energy loads.
Sustainable Cities and Society , 47,2019.[17] Rune Juhl Henrik Madsen, Peder Bacher and Jan Kloppenborg Møller. Grey-box modeling:an approach to combined physical and statistical model building. 2015.[18] Kumar Saurav and Vikas Chandan. Gray-box approach for thermal modelling of buildings forapplications in district heating and cooling networks.
Proceedings of the Eighth InternationalConference on Future Energy Systems , pages 347–352, May 2017.[19] Fritsde Nijs Donald Azuatalam, Wee-Lih Lee and Ariel Liebman. Reinforcement learning forwhole-building hvac control and demand response.
Energy and AI , 2, November 2020.[20] Yanzhi Wang Tianshu Wei and Qi Zhu. Deep reinforcement learning for building hvac control. , June 2017.[21] Sten Bay Jorgensen Niels Rode Kristensen, Henrik Madsen. Parameter estimation in stochasticgrey-box models.