energymodelmachine(emm)papers.cumincad.org/data/works/att/ecaade2017_269.pdf · 2017-09-13 ·...

EnergyModel Machine (EMM)

Instant Building Energy Prediction usingMachine Learning

Mohammad Rahmani Asl1, Subhajit Das2, Barry Tsai3, Ian Molloy4,Anthony Hauck51,3,4,5Autodesk Inc. 2Georgia Institute of Technology1,3,4,5{mohammad.asl|barry.tsai|ian.molloy|athony.hauck}@[email protected]

In the process of building design, energy performance is often simulated usingphysical principles of thermodynamics and energy behaviour using elaboratesimulation tools. However, energy simulation is computationally expensive andtime consuming process. These drawbacks limit opportunities for design spaceexploration and prevent interactive design which results in environmentallyinefficient buildings. In this paper we propose Energy Model Machine (EMM) asa general and flexible approximation model for instant energy performanceprediction using machine learning (ML) algorithms to facilitate design spaceexploration in building design process. EMM can easily be added to design toolsand provide instant feedback for real-time design iterations. To demonstrate itsapplicability, EMM is used to estimate energy performance of a medium sizeoffice building during the design space exploration in widely used parametricallydesign tool as a case study. The results of this study support the feasibility ofusing machine learning approaches to estimate energy performance for designexploration and optimization workflows to achieve high performance buildings.

Keywords:Machine Learning, Artificial Neural Networks, Boosted DecisionTree, Building Energy Performance, Parametric Modeling and Design, BuildingPerformance Optimization

INTRODUCTIONThe building sector as the largest consumer of theUnited States primary energy has been seeking totake necessary actions to reduce its energy use. Dueto the considerable impact of the buildings on theenvironment andwith the rise in environmental con-cerns, designers are increasingly expected to con-form appropriate minimum requirements regarding

energy efficiency (Rahmani Asl et al. 2015). Theyneed to identify design parameters with the signifi-cant impact on the building energy performance andoptimize them in the process of design to achievebuilding models with higher energy efficiency (Yu etal. 2010).

Multiple software applications have been devel-oped for simulating building energy performance,

MATERIAL STUDIES - ENERGY - Volume 2 - eCAADe 35 | 277

renewable energy, and sustainability in buildings(Zhao and Magoulès 2012). While these simulationtools produce high accuracy results, the simulationprocess is computationally expensive and time con-suming. On the other hand, these tools are basedon physical principles and they require high levels ofexpertise and detail building and environmental pa-rameters as input. As a result, building energy perfor-mance analysis is typically performed for final designvalidation and at the later phases of design process.

During the early design stages, designers oftenneed to quickly explore multiple design alternativesand optimize multiple performance factors at thesame timemake preliminary decisions. At this phase,designers do not require high fidelity simulations fordecisionmaking and they just need to comparemul-tiple designoptions andfind themost appropriate al-ternatives for their problem. This process is difficultwith whole building energy simulation tools due toslow feedback from conventional energy simulationengines (Tsanas and Xifara 2012). Therefore, buildingscience researchers proposed different approachesfor developing practical surrogate models to replaceactual simulation in the early stages of design. Signif-icant efforts have been done tomake conceptual de-sign tools environments interactive, so that design-ers can get instant feedback for continuous designiterations. One of the alternatives to high accuracyenergy simulation is the use of fast surrogate models(Guo et al. 2016). Among these models, data-drivensurrogate models become more and more practicaland important because getting access to large vol-ume of simulation data and computation power isgetting easier. Over the past years, machine learn-ing approaches and in specific deep learning meth-ods were very successful in learning from data.

In this paper we introduce Energy Model Ma-chine (EMM), aMachineLearning (ML)based tool thatuses Artificial Neural Networks (ANNs) and BoostedDecision Tree (BDT) methods trained on the existingsimulation results to predict the energy performanceof the buildings without the need to perform the ac-tual simulation. EMM is developed to provide instant

feedback to designers in the early phases of buildingdesign andguide them tobetter building energy andenvironmental performance. EMM uses a data set ofsimulatedmodels for optimising theweight parame-ters of ANNs algorithm and and BDT variesmaximumdepth of pruning and number of models for the en-semble set for each experiments. The trained mod-els is then used to predict the energy performanceof the newly submitted models by users and pro-vide instant feedback to help them design energy ef-ficient buildings. Basically EMM is an Artificial Intelli-gence that uses the results of the existing simulationsto predict the performance of the user models as aservice. It can be integrated with generative designand parametric performance analysis workflows toenable designers study a large number of design al-ternatives in a short period. In this paper, we providedetails about the EMMdevelopment process, andwecompare the EMMpredicted results of arbitrarymod-els and actual building energy simulation results. Thepaper provides a case study of use of EMM in a gener-ative design application to explore the design spaceof a medium size office building considering annualenergy use as the main performance factor.

RELATEDWORKBuilding energy simulation engines are widely usedfor energy performance prediction to help designersin the process of high performance building designsince practice has shown that these tools can oftengenerate resultswhichaccurately reflect actual build-ing energy use (Tsanas and Xifara 2012). These toolsuse physical rules and principles to calculate thermaldynamics and building energy use. Most of the ini-tial work on developing building energy simulationalgorithms was done a few decades ago. Neverthe-less, these tools became more accessible to design-ers over the past few years with the advancementof computational services. The U.S. Department ofEnergy (2012) has been publishing the “Building En-ergy Software Tools Directory” that provides detailinformation for over four hundred simulation toolsfor evaluating energy efficiency, renewable energy,

278 | eCAADe 35 - MATERIAL STUDIES - ENERGY - Volume 2

and sustainability in buildings. Crawley et al. (2008)provided a report comparing the features and capa-bilities of twenty major building energy simulationapplications. These resources can be used as a ref-erence for detail information about the widely usedbuilding energy simulation tools.

Using advanced building energy simulationtools may provide reliable solutions to estimate theenergy performance of building design alternatives;however this process can be very time-consumingand requires user-expertise in a particular program.Hence, in practice designers have to rely on surrogatemodels to study the energy performance of buildingdesign alternatives in generative design and opti-mization workflows to explore building design spaceand optimize building performance. In the literaturethere are various studies that used surrogate modelsto predict building energy performance. In multiplestudies, regression model is used to correlate oneor multiple building parameters to building energyconsumption (Bauer 2008; Ansari et al. 2005; Catalinaet al. 2008; Lam et al. 2010; Al Gharably et al. 2016).Other machine learning methods such as SupportVector Machine (SVM) (Dong et al. 2005; Lee et al.2009), Decision Tree (Yu et al. 2010), and ArtificialNeural Network (ANNs)(Kalogirou et al. 1997; Ben-Nakhi and Mahmoud 2004; Ekici and Aksoy 2009;Zhang et al. 2010) has been used to predict buildingenergy performance. In this study we chose ANNsthey are the most widely used artificial intelligencemodels in the application of building energy predic-tion they are good for solving complex problems.EMM also uses BDT as it is able to generate predic-tivemodelswith interpretable flowcharts that enableusers to quickly understand the model.

ENERGYMODELMACHINE (EMM)Building EnergyModel RepresentationIn this study, we use building design parameters torepresent building energy model. These parametersare identified based on their architectural and func-tional relevance and their potential impact on build-ing energy use. These parameters can be categorised

into three groups: 1) geometry parameters, 2) con-struction parameters and 3) load parameters. Table 1lists all of the parameters studied for EMM.

Table 1List of parametersstudied in EnergyModel Machine

In this study, we considered building wall, roof, andfloor area and window area by direction. Further-more, we studied the impact of cross-terms param-eters that consider wall, window, roof, floor area withtheir level height in the building as potential vari-ables. These parameters are designed as a set of newgeometry related parameters that help to track thevariations in geometry more accurately and they arenot studied in the previous studies focused on ML-based building energy performance prediction. Also,we gathered directional resolution for walls and win-


dows to increase the accuracy of the regressionmod-els. Parameters such as U-Value and Solar Heat GainCoefficient (SHGC) of building objects are includedin the model as construction parameters. Light-ing Power Density (LPD), Equipment Power Density(EPD), and Infiltration are some of the load parame-ters that are included in the model.

Building energy models are created using Au-todesk Insight energy analytical model creator fromarchitectural (conceptual and detailed) models andexported as Green Building XML (gbXML) files. Weprepared the input parameters for machine learningalgorithm by running a parser on gbXML files for themodels in the data set and prepare a data set of de-sired parameters. The ML algorithm is then run onthis data set. The following section describes the un-derlying working principles of some of the ML algo-rithms implemented on the data set.

Machine LearningWe used two machine learning methods, ArtificialNeural Networks (ANNs) and Boosted Decision Tree(BDT), for EMM. These two methods have been suc-cessfully used to predict building energy perfor-mance in other studies. ANNs are very good in mod-eling complex relationships between input parame-ters and outputs to predict accurate results. How-ever, ANNs usually operate as black box and are verydifficult to interpret especially for architects and de-signers in case they want to understand more aboutthe generatedmodel and theway it works. EMM alsouses BDT method which is much simpler to utilizeandits result can be interpreted easily compared toANNs.

Data. Our dataset comprises of over 180,000 datapoints, each representing a building energy simula-tion output. The Data has over 97 features includ-ing both categorical and quantitative data types. Forour experiment and prototype we stored the dataas a local csv file for data transformation, cleaningand building our machine learning model. However,for future use cases with more data points, we envi-sion to setup adatabase anduse the current software

architecture on it. We cleaned the data of any un-wanted features like, project ids, weather id, or sparsefields which did not addmuch to the information setany way. After removing such features from the data,we were left with 87 features.

DataModeling.Webuilt ourmachine learningmod-els with an iterative approach. On each iteration ofmodel building we verified their performance by 10folds’ cross validation on the whole data set. Withmultiple iterations, we compared Root Mean Squarevalues and R2 Scores across different instances ob-tained by tweaking model parameters. We pickedthe model instance with best performance based onthe abovemetrics. Nextwe split the dataset into 67%training set and 33% test sets. We trained our se-lected model instance on the training set. Our case,is a regression problem, where out of the 87 fea-tures, we have 86 independent variables (’X’). Fromour dataset, our dependent variable or unknownout-put ’Y’, is the Energy Simulation Output. We com-puted our model performance based on how accu-rate the prediction of the energy simulation outputis with respect to correct results as present in the testdata set. Weconductedexperimentson twomachinelearning models ANNs and BDT. We also tested theresults with and without feature selection or dimen-sionality reduction algorithms. Our intent was to ex-plore the trade-off between speed gain and accuracyfrom the various experiments.

Technology Stack. We used Python for all the datacleaning, data transformation and model buildingprocesses, using numpy, pandas, and scipy pack-ages (Walt et al. 2011). For machine learning mod-els, we used scikit learn (Pedregosa et al. 2011) andKeras deep learning toolkit using Theano as a back-end to handle the computational heavy lifting forANNs. For data visualisation to evaluate model per-formance and to compare model output, we usedPython’s Matplot (Hunter 2007) library package. Be-low we describe our machine learning models andexperiment setup in detail.


Artificial Neural Networks (ANNs)ANNs are the most widely used artificial intelligencemodels in the application of building energy pre-diction since it is good for solving nonlinear prob-lems with complex interdependencies. In the pasttwenty years, researchers have applied ANNs to ana-lyze various types of building energy consumption ina variety of conditions, such as heating/cooling load,electricity consumption, sub-level componentsoper-ation, and estimation of usage parameters. For moreinformation please refer to Krarti (2003) and Dounis(2010) research which provide review studies of arti-ficial intelligencemethods in the applicationof build-ing energy systems.

Interest in this direction resurfaced in, 1982,when JohnHopfield (Hopfield 1982) in his paper pre-sented methods to create useful systems, using bidi-rectional connections between neurons. The ideawas to learn from biological processes and mimicthem on a computer system to help model systemsbetter. With the current advent of Big Data sourcesand commendable advances in hardware and com-pute speed availability, use of these systems, has po-tential to show very good results. Neural networksare capable of generating their implicit rules by learn-ing from provided instances. Their ability to gener-alise the model has been proved superior to othercomparable machine learning systems spanning ap-plication in awidearrayof fields (Ayat andPour 2014).

ANNs - Experimental Setup.We created ANN mod-els using Keras library with Scikit Learn package inpython. We followed the same pipeline of cross val-idation, model building on train set and then verify-ing performance on test data as mentioned before.Keras library runs by building a ANN model first. Wecreated several such models and compared modeloutput and compute time. We describe the widthof such a model by the number of features it uses asneuron for each ANN layer. Similarly, we describe thelargeness of such amodel as number of hidden layersit contains to compute the output. As our problemis a regression problem, unlike a classical classifica-tion problem on ANN, our model has only one node

on the output layer (here output, is a number, ratherthan yes or no values for different classes ). Output isthe predicted energy simulation result based on theinput independent variables ‘X’ in the model.

ANNs - Experimental Result. Below, we briefly de-scribe the range of models we have experimentedto build our model on the training data. For all themodels below we used Rectified Linear Unit (RELU)as the activation function. However, we also havetested models with other activation functions likesoftmax, sigmoid and linear, but RELU’s performancewas much better than others for our data set.

• Base Model: This model was the simplest ofthe models studied. It had only one hiddenlayer. The number of neurons used were 13.This model was much faster than others be-cause of its simplicity and gave accepted re-sult with R2 Score 0.999777.

• Wide Model: This model was bit more com-plex than the base model. It had only onehidden layer but we increased the numberof neurons to 50. This model was slowerthan the previous with a marginal incrementin performance, having output with R2 Score0.999899.

• Wide and Large Model: This model was themost complex we tried. It tested with 3 hid-den layers having over 50 neurons on eachlayer. This model was way slower than allprevious models with a very slight incrementin performance, having output with R2 Score0.999977. Owing to its toomuch dependenceon input features on train data, this modelshowed signs of overfitting.

The charts in Figures 1 and 2 explain some ofthe model performance comparison of our ANNModel. After experimenting with different modelswith changeable parameters, for this dataset, we rec-ommended the baseline model. However, we fore-see a scenario, where end users can select the param-eters and the model type from Dynamo Node whenwe query our model engine.


Figure 1Prediction (x-axis)and simulation(y-axis) resultscomparison (ANNs)

Figure 2Prediction andsimulation resultscomparison (y-axis)and data points(x-axis) (ANNs)

Boosted Decision Trees (BDT)Further, we tested our dataset on Boosted DecisionTrees. This method is able to generate predictivemodels with interpretable tree flowcharts that en-able users to quickly understand the model and ex-tract useful informationwhich is an advantage of thismethod over other widely used techniques (Yu et al.2010). We used a list of weak learners and iterativelypredicted scores on our dataset and tested its accu-racy. On each iteration, we upweighted those datapoints with predicted valuewas different from actualresults. Our final predictor is theweighted average ofall the predictors from the weak learners. BDT worksby incrementally fixing those data points with pre-dicted score was off from the actual results (Roe etal. 2005).

BDT - Experimental Setup. We used Scikit Learn’sDecisionTreeRegressor and AdaBoostRegressor (Pe-dregosa et al. 2011) to model simple decision treesandboosteddecision trees respectively. We followedthe same pipeline of cross validation, model buildingon train set and then verifying performance on testdata as mentioned in the previous section.

BDT - Experimental Result.• Decision tree models: We experimented withsimple decision trees with varying maximumdepth of pruning for each of our experiments.From model comparison, we realized a maxi-mumdepthof 10-12gavegood results. As de-cision trees were simple yet powerful models,the runtime for our experimentwasway fasterthan ANN models. Our best result from deci-sion treemodel was output with a R2 Score of0.999715.

• Boosted Decision tree models: We exper-imented with boosted decision trees withvaryingmaximumdepth of pruning and vary-ing number of models for the ensemble setfor each of our experiments. From modelcomparison, we realized amaximumdepth of10-12, with 4 models in the ensemble, gavegood results. As boosted decision trees weremore complex than decision trees, the com-pute runtime for our experiment was slowerthan that of decision trees. However, BDT run-time was significantly better than ANN mod-els with almost same performance output.Our best result from BDT model was outputwith a R2 Score of 0.999985. Figures 3 and4 show the comparison of the predicted re-sults with energy simulation results for BDTmethod.


Figure 3Prediction (x-axis)and simulation(y-axis) resultscomparison (BDT)

Figure 4Prediction andsimulation resultscomparison (y-axis)and data points(x-axis) (ANNs)

Feature Selection or Dimensionality Reduc-tion for ANNs and BDTWe tested both ANN and BDT models with featureselection to test if we can decrease model computeruntime without losing performance, especially forANN. We tested feature selection algorithm PrincipalComponent Analysis (PCA) by setting number of de-sired features as a model parameter. Out of avail-able 87 features we testedwith 20, 30 and 50 neededfeatures. There was little improvement on the run-time, but our results were not as precise as the onewe got from themodel without feature selection. Wealso tested other feature selection algorithms like Se-lect K Best and Recursive Feature Elimination (RFE).In some cases, the results were significantly bad andunacceptable with very low R2 score. For this use

case and data set, we recommend modelling with-out using feature selection, but in future use cases,wewould like to test feature selection further and re-fine our model.

CASE STUDYIn order to show the usefulness of the EMM in theearly design process, we created a case study of anoffice building. In this case we use Autodesk Dy-namo Studio, a graphical programming interface asthe design tool, to parametrically design the modeland study its energy performance as a measurementfactor. In this study, Window to Wall Ratio (WWR) ofNorth, South, East, West directions as well as ShadingSize, WindowWidth, and Balcony Extension parame-ters are evaluated as parametric buildingparameters.All of these parameters are controlling the amount ofthe lights that enters in the building (balconies of theupper level act as horizontal shading for the lowerlevel). Annual energy use (kBtu) and average WWRare the decision making parameters that are stud-ied in this case. Average WWR is a simple averageof the WWRs per direction. To evaluate the energyperformance, we pushed the trained model (in thiscase BDTmodel) as a service that is accessible byRep-resentational State Transfer (REST) Application Pro-gramming Interface (API) as a node in Dynamo Stu-dio. Figure 5 shows the view of Dynamo graph andthe geometry in Dynamo interface.

In order to parametrically study the design, Dy-namo Studio enables users to publish their modelon the web and open the model in project Fractal.Project Fractal enables users to manage explorationof design space providing different generation op-tions and facilitates decision making by Parallel Co-ordinate Chart (PCC) and visualization of design op-tions in design grid. Figure 6 shows the same modelinproject Fractalwhichgenerated about 7000designoptions using cross product generation method.

This case study shows the usefulness of EMMas a energy prediction approximation model whichenables design space exploration in a timely man-ner. Evaluating this large number of options by simu-


Figure 5Dynamo graph andthe geometry inDynamo interface

Figure 6Building model inproject Fractal


lating models using conventional energy simulationtools would have taken many days and lots of re-sources. Using EMM, and other rapid performanceanalysis factors such as view and comfort factors, thedesigner can narrow down the design options andthen perform detail simulation to get more accurateand detail results.

FUTUREWORK AND CONCLUSIONThe objective of the research is to find potential waysto bypass the process of energy simulation for build-ing designers in the early stage of building design.This is mainly for the fact, that energy simulationtakes a significant amount of time, is resource inten-sive and needs a certain level of detail in the buildingmodel, which is usually low in early stages. EMMusesan extensive set of parameters to track all of the im-pactful variables on energy performance of buildingmodel and learns from existing data by using ML al-gorithms providing quick and real-time energy simu-lation feedback todesigners. Theseparameters covergeometry, construction, and load related buildingvariables. Considering these parameters from all ofthe three categoriesmentioned above in the trainingprocess of the machine learning methods and mak-ing the service available to be easily used in designapplications is the main contribution of this paper.

Current implementationof EMMusesANNs sinceit can predict accurate results for complex interde-pendent problems. EMM also includes BDT since itenables users to quickly understand the model andextract useful information which is very useful for ar-chitects. As it was demonstrated in the body of themanuscript, both models have acceptable accuracyin predicting the energy performance factor. EMMmakes the trained model available as a service andmake it easy to be accessed by any parametric de-sign tool. The user would only need to call the ser-vice through the provided API and add it to the de-sign process.

The case study of the office building model de-sign exploration using two of the common paramet-ric design tools demonstrates the usefulness of this

approach. Using EMMwewere able to explore about7000 building design options and their energy per-formance and make informed design decision in theconceptual design phase.

As part of the future work, we are also addingmultiple images of the building model from variousangles as training parameters to be able to track thegeometry of the buildingmodelmore accurately. Weare training the models on the top of these imagesto categorize building model geometry and increasethe accuracy of the results. One of the drawbacksof the current system implementation is that it isfeeding all of the features in the dataset into the MLModel for prediction. However, fromexpert’s domainknowledge, we know every parameter of the build-ingdesigndoesnot have the same impact inbuildingenergy use computation. Thus, our in-progress workincludes implementing feature selection and dimen-sionality reduction algorithms to the data set, beforewe pass it on to the selected ML algorithm.

ACKNOWLEDGEMENTWe would like to acknowledge and thank OmidOliyan Torghabehi for his work and support onpreparing the case study for this paper.

REFERENCESAnsari, F, Mokhtar, A, Abbas, K and Adam, N 2005, ’A sim-

ple approach for building cooling load estimation’,Am. J. Environ. Sci, 1(3), pp. 209-212

Asl, MR, Zarrinmehr, S, Bergin, M and Yan, W 2015,’BPOpt: A framework for BIM-based performanceoptimization’, Energy andBuildings, 108, pp. 401-412

Bauer, M and Scartezzini, JL 1998, ’A simplified corre-lation method accounting for heating and coolingloads in energy-efficient buildings’, EnergyandBuild-ings, 27(2), pp. 147-154

Catalina, T, Virgone, J and Blanco, E 2008, ’Develop-ment and validation of regressionmodels to predictmonthly heating demand for residential buildings’,Energy and buildings, 40(10), pp. 1825-1832

DOE, US 2012, Building energy software tools directory,Department of Energy

Dong, B, Cao, C and Lee, SE 2005, ’Applying support vec-tor machines to predict building energy consump-


tion in tropical region’, Energy and Buildings, 37(5),pp. 545-553

Dounis, AI 2010, ’Artificial intelligence for energy conser-vation in buildings’, Advances in Building Energy Re-search, 4(1), pp. 267-299

Guo, X, Li, W and Iorio, F 2016 ’Convolutional neural net-works for steady flow approximation’, Proceedingsof the 22nd ACM SIGKDD International Conference onKnowledge Discovery and DataMining, pp. 481-490

Hopfield, JJ 1982, ’Neural networks and physical sys-tems with emergent collective computational abil-ities’, Proceedings of the national academyof sciences,79(8), pp. 2554-2558

Hunter, JD 2007, ’Matplotlib: A 2D graphics environ-ment’, Computing In Science \& Engineering, 9(3), pp.90-95

Krarti, M 2003, ’An overview of artificial intelligence-based methods for building energy systems’,TRANSACTIONS-AMERICAN SOCIETYOFMECHANICALENGINEERS JOURNAL OF SOLAR ENERGY ENGINEER-ING, 125(3), pp. 331-342

Li, Q, Meng, Q, Cai, J, Yoshino, H and Mochida, A 2009,’Applying support vector machine to predict hourlycooling load in the building’, Applied Energy, 86(10),pp. 2249-2256

Pedregosa, F, Varoquaux, G, Gramfort, A, Michel, V,Thirion, B, Grisel, O, Blondel, M, Prettenhofer, P,Weiss, R andDubourg, V 2011, ’Scikit-learn: Machinelearning in Python’, Journal of Machine Learning Re-search, 12(Oct), pp. 2825-2830

Pour, ZAandAyat, S 2014, ’Comparisonbetweenartificialneural network learning algorithms for predictionof student average considering effective factors inLearning and educational progress’, Journal ofMath-ematics and Computer Science, 8, pp. 215-225

Roe, BP, Yang, HJ, Zhu, J, Liu, Y, Stancu, I and McGregor,G 2005, ’Boosted decision trees as an alternative toartificial neural networks for particle identification’,Nuclear Instruments andMethods in Physics ResearchSection A: Accelerators, Spectrometers, Detectors andAssociated Equipment, 543(2), pp. 577-584

Tsanas, A and Xifara, A 2012, ’Accurate quantitative esti-mation of energy performance of residential build-ings using statistical machine learning tools’, Energyand Buildings, 49, pp. 560-567

Walt, Svd, Colbert, SC and Varoquaux, G 2011, ’TheNumPy array: a structure for efficient numericalcomputation’, Computing in Science \& Engineering,13(2), pp. 22-30

Yu, Z, Haghighat, F, Fung, BC and Yoshino, H 2010, ’Adecision tree method for building energy demand

modeling’, Energy and Buildings, 42(10), pp. 1637-1646

Zhang, J and Haghighat, F 2010, ’Development of Arti-ficial Neural Network based heat convection algo-rithm for thermal simulation of large rectangularcross-sectional area Earth-to-Air Heat Exchangers’,Energy and Buildings, 42(4), pp. 435-440

Zhao, Hx and Magoulès, F 2012, ’A review on the pre-diction of building energy consumption’, Renewableand Sustainable Energy Reviews, 16(6), pp. 3586-3592


energymodelmachine(emm)papers.cumincad.org/data/works/att/ecaade2017_269.pdf · 2017-09-13 ·...

Documents