Abstract
JOURNAL OF ASIAN ARCHITECTURE AND BUILDING ENGINEERING 2021, VOL. 20, NO. 3, 326–339 https://doi.org/10.1080/13467581.2020.1783271 ENVIRONMENTAL ENGINEERING A new design exploring framework based on sensitivity analysis and Gaussian process regression in the early design stage Yun Gao, Masayuki Mae and Keiichiro Taniguchi Department of Architecture, Graduated School of Engineering, University of Tokyo, Tokyo, Japan ABSTRACT ARTICLE HISTORY Received 12 February 2020 With building energy codes getting strict, quantitative analysis is necessary in the early design Accepted 29 May 2020 stage of high-energy-performance buildings. To fully explore the design space, a highly efficient method is necessary. In the research, a new design exploring framework was pro- KEYWORDS posed. Parameters are screened and separated into groups based on the sensitivity analysis, to Sensitivity analysis; reduce the dimension of inputs. Gaussian process regression, able to deal with the uncertainty interaction analysis; Gaussian in inputs, was used to make metamodels to reduce the calculation cost and fully explore the process; uncertainty; design design space. Dashboards were made to visualize the data interactively, to help designers dashboard make decisions and make communications smooth. This framework was demonstrated with a case study and showed its efficiency. 1. Introduction process, the possibilities of the design are often not fully explored (Attia 2018). Meanwhile the communica- Buildings are responsible for more than 40% of global tion cost, within the design team and between energy use and one-third of greenhouse gas emissions designers and clients is considerable. Automatic opti- globally (Hirst 2013). In Japan, since the oil crisis in the mization, with Genetic algorithm, Particle Swarming 1970s, the energy consumption by building sector Optimization, etc., is another way to efficiently improve considerably increased by about 250%, which the energy performance of the design (Wang, Rivard, accounts for 34.5% of the total energy use in the and Zmeureanu 2006). However, it is not very suitable country (Ministry of Economy, Trade and Industry, JP in the early stage, as there are a lot of factors and 2015). With. the progress of urbanization, the energy criteria that cannot quantify be such as aesthetical or use of building sector will continuously increase by historical things. The optimized results might be unde- about 30% in the coming 20 years (U.S. Energy sirable to designers. It requires thousands of iterations Information Administration 2017). of simulation to run a successful optimization, which is On the other hand, building sector has the largest quite calculation costly and takes too much time. On potential for a reduction in energy consumption and the other hand, in the early stage, it is better to explore greenhouse gas emission. It has been proved that the the design possibility as much as possible, rather than architecture design, especially in the early stage, has get accurate analysis or an optimized result. a significant impact on the energy performance The interactions between parameters are rarely ana- (Naboni et al. 2015), while the cost of changing design lyzed in the design process. When the designers study will get higher and higher in the later stage (WBDG a certain parameter, they usually assign constant 2019). However, the most important decisions are values to other parameters based on some assump- made in the early design stage by architects tions. The decisions made in earlier phases are often (Schlueter and Thesseling 2009), usually with some influenced by those made in the latter stages and rules of thumbs. With the energy codes getting more become no more optimal, due to the interactions and more strict in recent years and Zero Energy House between the parameters. As a result, the designers gradually becoming mandatory in Japan, it has been have to move back and repeat the studies in previous appointed out by some researchers that guild lines or phases. rule of thumbs are not enough to ensure the energy This study focuses on four questions, (1) How to performance (Attia 2018). Integrated Design Process is explore the design possibilities, (2) how to keep the required to achieve ultra-low-energy design (Ferrara, calculation cost at a low level at the same time, (3) how Sirombo, and Fabrizio 2018). Quantitative energy ana- to utilize the data to help the designers when making lysis should be carried out in the early stage. However, decisions and (4) how to avoid duplicated work. This limited by the low efficiency of the trial and error paper has six sections, including this introduction. CONTACT Yun Gao chuyun0316@hotmail.com The University of Tokyo, Tokyo, Japan Supplemental data for this article can be accessed here. © 2020 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group on behalf of the Architectural Institute of Japan, Architectural Institute of Korea and Architectural Society of China. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. JOURNAL OF ASIAN ARCHITECTURE AND BUILDING ENGINEERING 327 Related works, including published tools and extracted the information from the BIM model to carry researches, are reviewed in Section 2. Section 3 out energy, daylight and structure analysis. The design describes the framework of design exploring, as well would be optimized by Optimp, an optimization tool as the sensitivity analysis (SA), metamodeling and data integrated in Dynamo. The whole process was highly visualization methods used in this study. Section 4 integrated, nevertheless, the calculation was still very provides a case study as a demonstration of this frame- costly. Jalaei and Jrade (2015) integrated BIM and LEED work. Conclusion and future work are presented in certification system for the conceptual stage. This inte- Section 5. grated tool could generate LEED certification point using the information from BIM and help the designer selects the proper materials. However, LEED certifica - 2. Related works tion required a highly completed design. A lot of infor- mation necessary for LEED was lacked in the 2.1. Dashboard tools Dashboards, able to visualize the building parameters conceptual stage. As a result, it was still a question and the energy performance, are useable when that whether LEED provides good criteria in the early stage. designers make design decisions. Insight 360 (Autodesk 2019) is a building energy performance assessment tool published by Autodesk. 2.3. Sensitivity analysis This designer-oriented dashboard tool is intuitive and SA has been adopted by many researchers to reduce easy to use. Good data visualization helps the designer the number of parameters and calculation cost. understands the energy performances well. Heiselberg et al. (2009) introduced Morris method Parameters can be modified, and the feedback is real- into the design of sustainable buildings. They screened time. However, there are still some limitations. The the parameters by their importance and dropped less parameters and their range that can be studied are important parameters in the latter stage. They com- limited to the preset. Only the average value of energy mented that SAshould be performed in the early stage consumption of all cases is displayed. Other criteria, when important factors were not decided. Garcia such as thermal comfort, are not included or Sanchez et al. (2014) performed both first- extendable. and second -order analysis in building energy simula- Designer Explorer (Tomasetti 2019) is another tool tions. They commented that higher-order effects could under the concept of design dashboard. It consists of help better understanding the results of SA. Østergård, two parts, Grasshopper3D components and online Jensen, and Maagaard (2017) described an approach data visualization. Designers can do the analysis in to explore the multi-dimensional design space. They Grasshopper3D and output the results into a CSV file. used Morris method to screen parameters. The impor- The online data visualization tool visualizes the data tance of each parameter helped designers to make and makes dashboards once this CSV file is uploaded. decisions in the latter stage. Nguyen and Reiter This process takes quite a long time to execute the (2015), Rivalin et al. (2018) and Gagnon, Gosselin, and simulations if the designer wants to explore the possi- Decker (2018) compared several SAtechnics applied to bilities of design widely. building energy performance assessment. Linear regression was reported not good enough by most researchers. Morris method showed both good effi - 2.2. BIM-based tools ciency and enough reliability. Variance-based meth- The authors have made several interviews with profes- sional architects. They said that they prefer to stick to ods, though had the best reliability, were very BIM tools from the very beginning, even though they calculation costing. It was also pointed out that Latin hypercube outperformed the standard Monte Carlo in know that Grasshopper3D might be more convenient Morris method (Figure 1). in the early stage, so that they can push the work more smoothly to latter stage. Several Building Information Modeling (BIM)-based tools, aiming at the early design 2.4. Metamodeling stage, have been developed by researchers. Schlueter Metamodeling is a very effective way to reduce calcu- and Thesseling (2009) developed a plug-in for lation cost. Autodesk Revit called Design Performance Viewer, Hygh et al. (2012) presented a Monte Carlo frame- which could evaluate the energy performance of the work to develop a multivariate linear regression model building, using the information from BIM. Once the based on 27 parameters. The coefficient of each para- designers modified the model, they could get nearly meter in the regressed model could be used as the real-time feedbacks. Though modeling and analyzing sensitivity. Østergård, Jensen, and Maagaard (2017) were integrated, designers still had to work in the trial- also made metamodels with multivariate linear regres- and-error process. It was difficult to compare a lot of sion. They made a “what-if” dashboard, with the meta- alternatives simultaneously. Rahmani Asl et al. (2015) models running in background, to help the designers integrated BIM and multi-objective optimization. They making decisions. Gossard, Lartigue, and Thellier 328 Y. GAO ET AL. Figure 1. Comparison between sampling methods with 25 sample points. (2013) trained artificial neural network (ANN) with the simulation results of energy performance of the build- ing. They used this trained ANN as a part of objective function in GA optimizations. Asadi et al. (2014) also combined ANN and GA in their optimization process and applied it in retrofit projects. Rivalin et al. (2018), Wei et al. (2015) and Østergård, Jensen, and Maagaard (2018) compared several techs of metamodeling applied to building energy perfor- mance assessment. Gaussian process was reported to be the most robust and easy-to use, neural network and multivariate adaptive regression splines also had good performance. In some cases, polynomial chaos showed better accuracy than Gaussian process did. 3. Methodology 3.1. Targeted parameter and non-targeted parameter In order to make this research more understandable, we defined two concepts, targeted parameters and non-targeted parameters. Target parameters, which is actually an alternative way to say parameters of inter- est, means the parameters that can be decided by the design team, such as window to wall ratio (WWR), insulation, etc. Other parameters influencing the Figure 2. The proposed framework of design exploring. energy performance but can hardly be decided or controlled by the design team, such as climate, occu- pancy, etc., are called non-targeted parameters in this research. We think that both targeted and non- volume or the topology of the functional spaces. targeted parameters should be considered, as well as Based on the conceptual design, the design team, the interactions between them. including architects and experts from different disci- plines, will list the targeted and non-targeted para- 3.2. Framework meters and their ranges or distributions. The second In this research, we proposed a framework of design step is to analyze the sensitivity of each parameter. exploring based on SA (Figure 2). Samples will be made with Latin Hyper-cube. An IDF Architects will firstly work on the conceptual file will be made for each sample parametrically and schema of the design, decide the rough shape, the simulated with EnergyPlus. Morris method will be used JOURNAL OF ASIAN ARCHITECTURE AND BUILDING ENGINEERING 329 Figure 3. The movement of a sample in design space, Morris method. Figure 4. The movement of a sample, Expanded Morris method. to get the sensitivity of each parameter based on the each phase, those non-targeted parameters, which simulation results. Parameters will be screened based interact with the targeted parameters in this phase, on their sensitivities, or importance in another word. will also be studied together. The parameters will be The third step is to analyze the interactions between sampled with Monte Carlo method and simulated with parameters with Expanded Morris method, by quanti- EnergyPlus. A meta-model will be made with Gaussian fying the higher-order effects. An interaction matrix process. The targeted parameters and results from will be made. simulations will be used to train the model. Non- The design team can separate the targeted para- targeted parameters will be regarded as noise of the meters into groups based on their interactions, so that model. one problem is divided into several problems and the With a trained meta-model, the design team is able dimension of each problem is reduced. The design to get real-time feedback when any parameter is mod- team can study the problems phase by phase. In ified, and test thousands of cases in minutes. A design 330 Y. GAO ET AL. Figure 5. The movement of a sample, Expanded Morris method. dashboard, visualizing the data, can help the design after moving, divided by the distance is used to repre- team understand the energy performance, make deci- sent the local sensitivity of the jth parameter (Eq.4), sions and communicate in teamwork or with clients. which can be regarded as an approximation of the After decision made, the design team can move into absolute value of the partial derivative. The local sen- the next phase, do simulations, train a metamodel, sitivity of each parameter will be calculated, and yðXÞ expand the design dashboard and make decisions on will be calculated kþ 1 time for each row. another group of targeted parameters. This procedure repeats r times for all the rows in the input matrix, so the calculation of yðXÞ will be executed totally rðkþ 1Þ times. The average value of local sensi- 3.3. Sensitivity and interaction analysis tivity of jth parameter in all the rows will be used as its 3.3.1. Morris method. global sensitivity (Eq. 5). The SA can be roughly divided into variance-based 2 3 method and one-step-at-a-time method (OAT). As in 1 k x x ��� x ��� 1 1 the early stage of design, efficiency was more impor- 6 7 . . 6 . . . . . 7 . . . . . 6 7 tant than accuracy, we employed Morris method, . . . 6 7 1 j k 6 7 a global OAT for global SA, to analyze the sensitivity Mo ¼ x x # (1) ��� x ��� i i 6 i 7 6 7 . . of each parameter. 6 7 . . . . . . . . . . 4 5 . . . The first step is to generate the input matrix in the 1 j k x x ��� x ��� r r r range of (0,1]. In this research, Latin hypercube sam- pling (LHS) was used to generate the input matrix Mo � � (Eq.1), as LHS can better fill the exploring space than 1 k X ¼ Mo ¼ x ; . . . ; x ; . . . ; x # (2) i i i i i random sampling. The shape of Mo is r in row and k in column. Column count k is equal to the number of � � parameters of interest. Row count r represents the j j 1 j jþ1 1 k X ¼ x þ Δ ; . . . ; x þ Δ ; x þ Δ ; x ; . . . ; x 1 j 1 j i i i i i i number of times that OAT will be repeated. j 1 ¼ X þ Δ þ . . .þ Δ i 1 One row in the input matrix can be regarded as j¼X þΔ # a k-dimensional vector, X . (Eq. 2) An element in the i (3) vector, x , will move a small distance of equal size Δ in j j 1 its dimension (Eq. 3). The absolute value of the differ - jyðXÞ yðX Þj i i � � � � (4) LS ¼ # j 1 j i Δ ence in two results, y X and y X , before and i i JOURNAL OF ASIAN ARCHITECTURE AND BUILDING ENGINEERING 331 the training data. Radial basis function kernel (RBF) (Eq. GS ¼ LS # (5) 16) is used in this research. The prediction given by i¼1 a trained GPR is a normal distribution, rather than a single value, which makes GPR a very robust method 3.3.2. Extended Morris method. to regress noisy data. In this study, we also studied the interactions between � � �� � � �� y mðxÞ K K parameters. Interaction between two parameters N ; # (12) K K y mðx Þ � � �� means the influence of one parameter on the sensitiv- ity of the other one. For example, if the interaction pðy jx ; x; yÞ ¼ Nðy jμ ; V Þ# (13) � � � � between x and x is strong, it means that the value � 1 2 of x strongly influents the relationship between x and 1 2 T 1 μ ¼ mðx Þþ K K ðy mðxÞÞ# (14) y, also the relationship between x and y is strongly � influenced by x . Extended Morris method was T 1 (15) V ¼ K K K K # employed in this study to quantify the interactions � �� � between parameters. Like Morris method, an input � � 0 2 jjx x jj matrix Mo (Eq. 1) is generated firstly. For each pair of 2 Kðx; x Þ ¼ exp # (16) 2σ j 1 parameters, jth and j’th, the vector X moves a small distance of equal size Δ along axis j (Eq.7), then along 3.5. Interactive data visualization axis j’ (Eq. 8), then along both axis j and j’ (Eq. 9). The We think the data visualization should be interactive. local interaction between jth and j’th parameters can Once the designers modify some parameters, the cor- be calculated with Eq.10, which quantifies the influ - responding results should feedback in real-time. This ence from a parameter on the local sensitivity of the allows designers, and even clients, quickly test their other. For each row, yðXÞ will be calculated idea. In this research, Microsoft Power BI is used. Power dðd 1Þ=2þðd 1Þþ 2 times. So, the calculation BI is an interactive data visualization tool based on will be executed r½dðd 1Þ=2þðd 1Þþ 2� times in Power Query, a data retrieval technology. Power BI is total. The average value of local interaction between able to retrieve data from a huge amount of data jth and j’th parameters from all rows is used as the resources based on the filters set by users and redraw global interaction of them (Eq.11). � � the graphs immediately. Once the data visualization ðj;j Þ1 j 1 y ¼ y X # (6) i i has been published, it can be accessed from PC, iPad or smartphones, which is a big benefit during meetings. � � ðj;j Þ2 j 1 y ¼ y X þ Δ # (7) i i 4. Case study � � ðj;j Þ3 j 1 y ¼ y X þ Δ 0 # (8) j 4.1. Description of the case study i i The case study in this research is a midrise office � � building located in Tokyo. There is a small lake to the ðj;j Þ4 j 1 y ¼ y X þ Δ þ Δ # (9) j j i i south of the site. The main entrance will be located on the east, facing a road. The west side is adjacent to � � 0 0 0 0 ðj;j Þ4 ðj;j Þ3 ðj;j Þ2 ðj;j Þ1 � � y y y þy ðj;j Þ another building. The plan, consisting of two office i i i i (10) LI ¼ # i Δ rooms in south and north and one core, is quite widely used in energy performance study in Japan (Takizawa ðj;j Þ GI 0 ¼ LI # (11) 1985). The design team decided to make big openings ðj;j Þ i¼1 on the south façade to make full use of the nice view from south. However, based on the climate in Tokyo, big opening on south, even with overhangs, would 3.4. Gaussian process result in higher energy consumption. How to balance In this research, we use Gaussian process regression the view and energy performance would be a key (GPR) to build metamodels. Gaussian process, also point in this design. Fins would be attached to east known as Kriging method, is a kind of supervised façade, to shade the building from the rising sun, learning method based on the Bayesian inference. which would also be important visual elements. How A prior probability is defined based on the covariance to decide to size of the fins, balance the appearance matrix of the observations (Eq. 12). Predictions are and performance is another question. made by interpolation governed by this probability (Eq. 13). So, GPR can be simply described as, the closer two points are located in the input space, the similar 4.2. Parameters and criteria the predicted outputs will be. Usually, a kernel function In this study, we mainly focused on the envelop per- is used to generate a prior covariance matrix based on formance. Though orientation and aspect ratio had 332 Y. GAO ET AL. a huge impact on the energy performance, limited by top 10 heating/cooling load from the 8760 results the site, these two parameters were not included. The were picked up to get the heating/cooling peak size of the opening was represented by WWR. As this load. Operative temperature was used to study the study aimed at the early stage, we simply represented thermal comfort. The percentage of comfortable, the performance of the window with Solar heat gain hot and cold hours was calculated. We also picked coefficient (SHGC) and U-factor of the glass. The ratio up the top 10 and bottom 10 operative tempera- between overhang depth and window height was ture from 8760 results to study the extreme situa- used to represent the shading on the south façade. tions. Table 3 shows the criteria. The depth and width of the fins on the east façade EnergyPlus 8.8 was used to execute the simulations. were also studied. Thickness of insulation material was The template of office building in climate zone 3a from used to represent the insulation of opaque part of the ASHRAE 90.1 2010 was used. We extract one floor with envelop. The solar absorptance of the outside layer of the standard plan from the building to do the calcula- the wall was also studied, as it was related to the color tion. The constructions and schedules followed the of the façade, which designers might be interested in. template. The roof and floor were set to adiabatic. Non-targeted factors, heating and cooling setpoints, The surroundings were reproduced with shading internal heat gain and air change rate, were also taken surfaces. into consideration as they had a direct influence on the energy performance. In the GPR, however, these fac- 4.3. Parametric modeling tors were regarded as noise, as they could not be A BIM model was built with Revit. The geometric precisely predicted in early stage. Tables 1 and 2 information was extracted with Dynamo, show the parameters and their ranges. a parametric modeling plug-in. A template IDF file Both energy performance and thermal comfort was made with OpenStudio[], based on the ASHRAE were studied. Hourly heating and cooling load 90.1, including the information of materials, construc- were calculated and converted into annual energy tions, occupancy schedules, etc. With a Python mod- density to evaluate the energy performance. The ule called EPPY, the geometric parts, walls, windows Table 1. Targeted parameters and ranges. Parameter Unit Range Description WWR (E, S, W, N) [0.2, 0.8] Window to wall ratio SHGC (E, S, W, N) [0.1, 0.9] Solar heat gain coefficient of the glass. U-factor of windows (E, S, W, N) W/m2 [0.5, 6] Represent the insulation of the transparent part of the façade. Thickness of gypsum board (E, S, W, N) m (0, 0.2] Represent the insulation of the opaque part of the façade. The conductivity of gypsum board is 0.16 W/mK Solar absorptance of external surface (E, [0.1, 0.9] Related to the color of the opaque part of facade S, W, N) Overhang scale (S) (0, 2] Overhang depth/window height Fin width (E) m (0, 0.9] The scale of a fin on east-west axis Fin depth (E) M (0, 0.9] The scale of a fin on south-north axis Table 2. Non-targeted parameters and ranges. Parameter Unit Range Description Cooling set point C° [23, 30] The room will be cooled if [room temperature ≥ clsp], from 6:00 to 22:00. Heating set C° [16, 23] The room will be cooled if [room temperature ≤ htsp], from 6:00 to 22:00. point Internal heat W/m2 [3, 15] The heat generated by people, lighting and machines in unit area, from 6:00 to 22:00. gain Air change rate times/hour [0, 6] The room will be ventilated if [room temperature > (clsp+htsp)/2] and [outdoor temperature < clsp] Table 3. Criteria. Criterion Unit Description Annual cooling energy density kWh/m The cooling energy consumed by unit area. COP is 3.2. Annual heating energy density kWh/m The heating energy consumed by unit area. COP is 3.5. Cooling peak kW The average value of the top 10 hourly cooling rate from 8760 results. COP is 3.2. Heating peak kW The average value of the top 10 hourly heating rate from 8760 results. COP is 3.5. Annual mean operative temperature °C The average value of hourly operative temperature of all 8769 results. Hot operative temperature °C The average value of the top 10 hourly operative temperature from 8760 results. Cold operative temperature °C The average value of the bottom 10 hourly operative temperature from 8760 results. Comfort OT percentage % Cooling period: [23, 26]; non-AC period: [21.5, 25]; Heating period: [20, 24] Hot OT percentage % Cooling period: (0, 23); non-AC period: (0, 21.5); Heating period: (0, 20) Cold OT percentage % Cooling period: (26, 100); non-AC period: (25, 100); Heating period: (24, 100) JOURNAL OF ASIAN ARCHITECTURE AND BUILDING ENGINEERING 333 generated with LHS. Each case was expended into 28 cases as there are 27 parameters. An EPW file of TMY from Meteonorm was used. As shown in Figure 7, the sensitivities of non- targeted factors were very high, which was disturbing. On the other hand, designers cannot really control them. For these reasons, we did not take them into consideration when determining the thresholds of sen- sitivities. For each criterion, the threshold was set to the half of the average value of the sensitivities of targeted parameters. A parameter would be skipped in further phases if its sensitivity was lower than the threshold. Appendix 1 shows the sensitivity of each parameter on each criterion. It was found that in this case, as this was a cooling- main building, the influence of insulation and solar absorptance on the total annual electricity consump- tion was almost negligible, which meant that the designer could feel free to decide the color and the material of the external surface. The sensitivity of the Figure 6. IDF files generation. insulation of the south façade on the operative tem- perature, however, passed the threshold. So, we and shadings were made and non-geometric proper- ignored all the insulation and solar absorptance para- ties were adjusted, such as SHGC, AC setpoints, etc. meters besides the insulation on the south in latter Figure 6 shows the flow that IDF files could automa- phases. The fin depth just passed the threshold a little. tically generated. Nineteen parameters left. Twenty cases, though enough to screen parameters, were too rough to visualize the sensitivities in latter 4.4. Sensitivity analysis steps. We did additional 250 cases for the left 19 para- The sensitivity of each parameter on each criterion was meters. Each case was expanded into 20 cases. In this analyzed with Morris methods. Twenty cases were Figure 7. Sensitivities of all parameters on annual energy density. 334 Y. GAO ET AL. step, we also took climate change into consideration. Each case was analyzed under four climates, TMY, 2020, 2050 and 2080 from Meteonorm. Twenty thou- sand simulations were executed in total. A sensitivity database was made. 4.5. Interaction analysis The interactions of the 19 parameters with respect to each criterion were analyzed with Expanded Morris method. Twenty cases were generated using LHS, and each case was expanded into 191 cases. Appendix 2 to 6 show the interaction matrices. The data in the matrices have been divided by the average value of interactions between WWR and U-factor on Figure 8. Extracting mean, bottom and top values from con- four orientations. The threshold was set to 0.5. The fidence interval. parameters of each orientation have strong interac- tions with each other. The interactions between para- thousand results. Symmetrically about the mean meters belonging to different orientations are quite value, 60% Confidence interval was used (Figure 8). weak. The 19 parameters were separated into 4 groups We extracted the bottom, mean and top value of the (Table 4). When we study a parameter, it is better to confidence interval to represent the possible energy study all the interactive parameters simultaneously. performance and thermal comfort of a case. For example, when the designers try to decide the A database was made. WWR on the south, it is recommended to take SHGC We repeatedly made the GPR for other three groups and U-value of the window on the south also into and add the predicted data into the database. consideration. However, designers can study the south façade and the east façade independently. 4.7. Data visualization and design dashboard Power BI was used to visualize the data and make 4.6. Metamodels by Gaussian process regression design dashboards. We firstly studied Group 1, the parameters related to For demonstration, we made three dashboards. the south façade. The values of the parameters in this Designers can adjust the parameters by moving the group were decided randomly with LHS, other para- sliders of filters. The variation of criteria and sensitiv- meters kept constantly to the mid value of their range. ities will be reflected in real-time. The database can be 2000 cases were generated and simulated as the train- utilized continuously throughout the design process. ing set, other 500 cases as testing set. Each case was New pages in dashboards can be easily made. simulated with four climates, TMY, 2020, 2050 and In the energy density and thermal comfort dash- 2080. In this phase, non-targeted parameters were board (Figure 9), designers could move the sliders on regarded as noise and not included in the training. the left side to adjust the ranges of parameters. The GPy (Sheffield ML 2019) was used to make the GPR EPW years could also be selected. Power BI would then models and train them. filter all the cases that the parameters were within this Other 200 thousand cases were generated by LHS. ranges from the database and display their results on The energy performance and thermal comfort of these the right side. For each case, three indices of energy cases are predicted with the trained GPR. As the output density, the bottom, mean and top of the 60% con- of GP for each case was a normal distribution, it was fidence interval, are displayed, named “Min predic- very difficult to analyze and visualize all the 200 tion”, “Mean prediction” and “Max prediction” in this dashboard. The average values of these three indices Table 4. Parameter groups. of all the filtered cases are displayed on the upper right Group 1 Group 2 Group 3 Group 4 in the form of a speedometer. Furthermore, we also Fin depth Overhang scale WWR W WWR N displayed the distributions of the three indices of all WWR E WWR S U-factor W U-factor N U-factor E U-factor S SHGC W SHGC N the filtered cases, which can help designers better SHGC E SHGC S Cooling Cooling understand the possibilities and risks of their decisions. setpoint setpoint Cooling Thickness of Heating Heating Due to the interactions between parameters, the setpoint gypsum S setpoint setpoint sensitivity of each parameter would change with the Heating Cooling setpoint Internal heat Internal heat setpoint gain gain values of other parameters changed. Another dash- Internal heat Heating setpoint Air change rate Air change rate board was made to display the sensitivity of each gain Air change rate Internal heat gain parameter on each criterion (Figure 10). By seeing the Air change rate sensitivities, the designers can be informed that which JOURNAL OF ASIAN ARCHITECTURE AND BUILDING ENGINEERING 335 Figure 9. Energy density and thermal comfort dashboard. Figure 10. Sensitivity dashboard. parameter is important, and they can make more the buildings. For example, we can see shading- efforts on those. This could be an important hint for related parameters getting more important but the designers. insulation-related parameters getting less important. As a building would last for decades, climate So, the designer should pay more attention to the change should also be considered in design stage. shadings. A climate change dashboard (Figure 11) was made. We thought that showing the influence from climate 4.8. Additional predictions and simulations change on the sensitivities of parameters, rather than When designers trying to explore the design space that on the energy consumption or thermal comfort, with the dashboard, it was found that, with the range can better help designers understand the impact on of parameters getting small, the cases met the filters 336 Y. GAO ET AL. Figure 11. Climate change dashboard. got quite a few and the distribution of the results higher than we expected. It is reasonable because this became unsmooth. We did additional 10 thousand building is 15 degrees East of South, the solar radiation predictions (for each climate and totally 40 thousand) might get into the room in the summer afternoon. The with trained GPRs within the shrunk range and append design team decided to attach a fin to the north the new data into the database. The distribution got façade. We did other 1000 simulations additionally to smooth as shown in Figure 13. Designers could do study the opening on the north. WWR, SHGC, U-value more specific tests on each parameter. of the north window and the newly added fin depth on From the sensitivity page, we found a blind spot. north were adjusted with LHS, other parameters kept The sensitivity of SHGC of north opening was much fixed. Another GPR was trained and new dashboard Figure 12. Added North facade dashboard. JOURNAL OF ASIAN ARCHITECTURE AND BUILDING ENGINEERING 337 Figure 13. Distribution got smooth with added data. (Figure 12) was made based on the prediction. As the Using the proposed process, a database contain- whole process was highly automated, it took just sev- ing 200 thousand cases could be made in two eral hours to run additional simulations and make new days. Comparatively, it takes more than 1 month dashboard. to simulate so many cases even using a high-end workstation. The non-targeted parameters, such as AC setpoints 5. Results or internal heat gain, are actually noise, rather than In the last section, we demonstrated the proposed parameters, in the early design stage. In this study, process with a case study. The CPU of the computer GPRshowed its robustness when dealing with noisy we used in this demonstration was an 8-thread one, data. By outputting the boundaries of the confidence which was consumer-grade and widely used in archi- interval, designers can also notice the possibility and tecture design studio. Eight simulations could be exe- risk of each decision, rather than only watch on the cuted simultaneously. It took around 10 to 30 second mean value. the run one simulation. A blind point, the shadings on the north façade, All the IDF files were generated parametrically using was found by widely exploring the design possibi- Python codes. It took several minutes to generated lity and showing the sensitivities. As the whole more than thousands of IDF files. process is highly automated after the first run, There were 27 parameters at the beginning. To do even there was modification in the base IDF file, the SA with Morris method, the simulations were exe- new database and dashboards were made with cuted 560 times, which took about 20 minutes. The 1 day. number of parameters reduced from 27 to 19 based on Interactive data visualization, with real-time feed- their sensitivities. Then, the interactions between the back, allows designers to test their ideas more effi - 19 parameters were analyzed with Extended Morris ciently. It also benefits the communication, within the method. The simulations were executed 3820 times, design team between different disciplines, or between which took about 2 hours. Based on the interaction, the design team and clients. the parameters were separated into four groups. In this The whole process (Figure 14) can be concluded as, way, the 27-dimensional problem has been simplified into one 8-dimensional, one 9-dimensional and two (1) Conceptual design 7-dimensional problems, which helped the designers (2) Sensitivity analysis and parameter screening avoid the curse of dimensionality. (3) Interaction analysis and parameter grouping For each group, 2500 IDF files were generated and (4) Metamodeling and prediction simulated with the weather data of TMY, 2020, 2050 (5) Data visualization and decision-making and 2080. The simulations were executed 40,000 times, which took about 20 hours. The results were used to Designers can return to step 4 or step 1 when mak- train and test Gaussian process models. With the ing decisions if necessary. trained GP models, the energy performance and ther- The proposed process was summed up into mal comfort of other 200 thousand cases were pre- a template of Python codes, commented with details, dicted within about 2 minutes. and uploaded to the GitHub. 338 Y. GAO ET AL. targeted parameters can help to improve the predic- tion, but a very narrow range is against the original idea of this research. As Power BI is a database-based tool and unable to utilize the metamodel, designers have to add data into the database when looking closer to a certain range of the design space. With the volume of data getting huger, it becomes very heavy and memory consuming to run the dashboard. The further steps of this research would be, Make the whole process more automatic Make further use of interactions and study higher- order effects Reduce the variance of the non-targeted para- meters but keep their ranges reasonable Introduce the variance of inputs into the GPR, for example, use Sparse GPR (Snelson and Ghahramani 2006) Make cross-platform data visualization which can run metamodel in background Disclosure statement No potential conflict of interest was reported by the authors. Figure 14. The revised framework. Notes on contributors Yun Gao is a Ph.D. student in the Department of Architecture, Faculty of Engineering, the University of Tokyo. His research 6. Conclusion and future works interests are data-driven issues in environmental design. A new design exploring process based on SA and GPR Masayuki Mae is an Associate Professor in the Department of was proposed in this research and demonstrated with Architecture, Faculty of Engineering, the University of Tokyo. a case study. Sensitivity and interaction analysis can His research interests are sustainable design and building reduce the dimension of the problem. Informed of the environment, especially passive house issues. sensitivities, designers can concentrate on the important Keiichiro Taniguchi is a Project Assistant Professor in the parameters. Based on the interaction analysis, parameters Department of Architecture, Faculty of Engineering, the separated into several groups, which can be studied inde- University of Tokyo. His research interests are building envir- onmental engineering and architectural environment pendently phase by phase. What is more, it can also design. prevent earlier decisions from being influenced by latter ones, avoiding repeated works. Trained Gaussian process models can predict the energy performance of a huge References number of cases in a moment. This process can help Asadi, E., M. G. Silva, C. H. Da, Antunes, L. Dias, and designers widely exploring the design possibilities. L. Glicksman. 2014. “Multi-objective Optimization for However, we found some new problems in this Building Retrofit: A Model Using Genetic Algorithm and research. Artificial Neural Network and an Application.” Energy and In the case study, it was found that the parameters Buildings 81: 444–456. doi:10.1016/J.ENBUILD.2014.06.009. Attia, S. 2018. Net Zero Energy Buildings (NZEB) 1st Edition belonging to different orientations were independent, Concepts, Frameworks and Roadmap for Project Analysis which could also be inferred by common sense and and Implementation. Oxford, UK: Butterworth-Heinemann. was disappointing. Further studies about how to make Autodesk. “Insight – High Performance and Sustainable better use of the interaction matrices are needed. Building Design Analysis.” accessed 25 December 2019. Though GPR showed its robustness when dealing https://insight360.autodesk.com/oneenergy Ferrara, M., E. Sirombo, and E. Fabrizio. 2018. “Automated with noisy data, it was also found in this study that if Optimization for the Integrated Design Process: The the variance of criteria caused by the non-targeted Energy, Thermal and Visual Comfort Nexus.” Energy and parameters was close to or even more than that by Buildings 168: 413–427. doi:10.1016/J.ENBUILD.2018.03.039. the targeted ones, GPR very likely failed to make Gagnon, R., L. Gosselin, and S. Decker. 2018. “Sensitivity proper predictions. Shrinking the ranges of non- Analysis of Energy Performance and Thermal Comfort JOURNAL OF ASIAN ARCHITECTURE AND BUILDING ENGINEERING 339 Throughout Building Design Process.” Energy and Buildings Østergård, T., R. L. Jensen, and S. E. Maagaard. 2018. 164: 278–294. doi:10.1016/J.ENBUILD.2017.12.066. “A Comparison of Six Metamodeling Techniques Applied Garcia Sanchez, D., B. Lacarrière, M. Musy, and B. Bourges. to Building Performance Simulations.” Applied Energy 211: 2014. “Application of Sensitivity Analysis in Building 89–103. doi:10.1016/j.apenergy.2017.10.102. Energy Simulations: Combining First- and Second-order Rahmani Asl, M., S. Zarrinmehr, M. Bergin, and W. Yan. 2015. Elementary Effects Methods.” Energy and Buildings 68: “BPOpt: A Framework for BIM-based Performance 741–750. doi:10.1016/J.ENBUILD.2012.08.048. Optimization.” Energy and Buildings 108: 401–412. Gossard, D., B. Lartigue, and F. Thellier. 2013. “Multi-objective doi:10.1016/j.enbuild.2015.09.011. Optimization of a Building Envelope for Thermal Rivalin, L., P. Stabat, D. Marchio, M. Caciolo, and F. Hopquin. Performance Using Genetic Algorithms and Artificial 2018. “A Comparison of Methods for Uncertainty and Neural Network.” Energy and Buildings 67: 253–260. Sensitivity Analysis Applied to the Energy Performance of doi:10.1016/J.ENBUILD.2013.08.026. New Commercial Buildings.” Energy and Buildings 166: Heiselberg, P., H. Brohus, A. Hesselholt, H. Rasmussen, 489–504. doi:10.1016/j.enbuild.2018.02.021. E. Seinre, and S. Thomas. 2009. “Application of Sensitivity Schlueter, A., and F. Thesseling. 2009. “Building Information Analysis in Design of Sustainable Buildings.” Renewable Model Based Energy/exergy Performance Assessment in Energy 34 (9): 2030–2036. doi:10.1016/j.renene.2009.02.016. Early Design Stages.” Automation in Construction 18 (2): Hirst, N. 2013. “Buildings and Climate Change: Summary for Decision 153–163. doi:10.1016/J.AUTCON.2008.07.003. Makers.” Design and Management of Sustainable Built Sheffield ML. “GPy by SheffieldML.” accessed 25 December Environments 9781447147817: 23–30. doi:10.1007/978-1-4471- 2019. https://sheffieldml.github.io/GPy 4781-7_2. Snelson, E., and Z. Ghahramani. 2006. “Sparse Gaussian Hygh, J. S., J. F. DeCarolis, D. B. Hill, and S. Ranji Ranjithan. Processes Using Pseudo-inputs.” In Advances in Neural 2012. “Multivariate Regression as an Energy Assessment Information Processing Systems, edited by Y. Weiss, B. Tool in Early Building Design.” Building and Environment Schölkopf, J. C. Platt, 1257–1264. Vol. 18. Cambridge, MA, 57: 165–175. doi:10.1016/j.buildenv.2012.04.021. USA: Bradford Series. MIT Press. Jalaei, F., and A. Jrade. 2015. “Integrating Building Information Takizawa, H. (1985). “A Proposal of Standard Plan for Energy Modeling (BIM) and LEED System at the Conceptual Design Research (Office Building).” The 15th symposium of thermal Stage of Sustainable Buildings.” Sustainable Cities and Society analysis, the state and topics, AIJ. Tokyo, Japan. 18: 95–107. doi:10.1016/J.SCS.2015.06.007. Tomasetti, T., “Design Explorer 2.” accessed 25 December Ministry of Economy, Trade and Industry, JP. (2015). 2019. https://tt-acm.github.io/DesignExplorer “Definition of ZEH and Future Measures Proposed by the U.S. Energy Information Administration. (2017). “International ZEH Roadmap Examination Committee [PDF File].” https:// Energy Outlook 2017 [PDF File].” https://www.eia.gov/out www.enecho.meti.go.jp/category/saving_and_new/sav looks/ieo/pdf/0484(2017).pdf ing/zeh_report/pdf/report_160212_en.pdf Wang, W., H. Rivard, and R. Zmeureanu. 2006. “Floor Shape Naboni, E., A. Malcangi, Y. Zhang, and F. Barzon. 2015. “Defining Optimization for Green Building Design.” Advanced the Energy Saving Potential of Architectural Design.” Energy Engineering Informatics 20 (4): 363–378. doi:10.1016/J. Procedia 83: 140–146. doi:10.1016/J.EGYPRO.2015.12.204. AEI.2006.07.001. Nguyen, A. T., and S. Reiter. 2015. “A Performance WBDG. “Architectural Programming. WBDG| WBDG Whole Comparison of Sensitivity Analysis Methods for Building Building Design Guide.” accessed 25 December 2019. Energy Models.” Building Simulation 8 (6): 651–664. https://www.wbdg.org/design-disciplines/architectural- doi:10.1007/s12273-015-0245-4. programming Østergård, T., R. L. Jensen, and S. E. Maagaard. 2017. “Early Wei, L., W. Tian, E. A. Silva, R. Choudhary, Q. Meng, and Building Design: Informed Decision-making by Exploring S. Yang. 2015. “Comparative Study on Machine Multidimensional Design Space Using Sensitivity Analysis.” Learning for Urban Building Energy Analysis.” Procedia Energy and Buildings 142: 8–22. doi:10.1016/J. Engineering 121: 285–292. doi:10.1016/J.PROENG.2015.0 ENBUILD.2017.02.059. 8.1070.
Journal
Journal of Asian Architecture and Building Engineering
– Taylor & Francis
Published: May 4, 2021
Keywords: Sensitivity analysis; interaction analysis; Gaussian process; uncertainty; design dashboard
