Conference Agenda

Session
Session F2.2: Ensuring high quality building simulations
Time:
Friday, 03/Sept/2021:
10:30 - 12:00

Session Chair: Joel Neymark, J. Neymark & Associates
Session Chair: Ádám Bognár, Eindhoven University of Technology
Location: Cityhall (Belfry) - Room 2


External Resource: Click here to join the livestream. Only registered participants have received the access code for the livestream.
Presentations
10:30 - 10:48

CNN-based quick energy prediction model using image analysis for shape information

Manav Mahan Singh1, Philipp Geyer1,2

1KU Leuven; 2TU Berlin

Aim and Approach

(max 200 words)

An early stage of design requires a few hundreds of simulations to support informed decision-making for energy-efficient building design (Clarke & Hensen, 2015; Lee et al., 2016). However, high computational time and re-modelling efforts required by dynamic energy simulation tools limit their use at the early design stage (Geyer & Schlüter, 2014). It requires quick energy prediction approach integrated into the design process. Previously, other researchers have proposed use of data-driven approaches such as deep learning (DL) model and their integration into building information modelling (BIM) tool (Amasyali et al., 2018). A designer experiments with building shape, envelope and technical specifications at an early design stage (Picco et al., 2014). Thus, we aim to develop a DL model which predicts the energy demand for various combinations of building shapes and other design parameters. Since there is a lot of information which is uncertain at the early design stage, we aim to develop BIM-integrated tool to make probabilistic energy prediction based on the range of undecided design parameters (Singh & Geyer, 2019). The tool should be able to evaluate several design options for energy performance within a few seconds to make use of energy prediction results at the early design stage.

Scientific Innovation and Relevance

(max 200 words)

The building shape evolves continuously at early stages and representing it with a fixed set of parameters is cumbersome. However, we can represent a building having the same floor plan on each floor, by a two-dimensional image. We are using a convolutional neural network (CNN) to capture the information related to the building shape. Besides the building shape information, there is a set of technical specifications and other parameters which are adequately represented by numeric values. Thus, we need a model architecture which utilises both the information of building shape and other design parameters. We propose a CNN-DL model architecture, in which CNN captures the information from image (floor plan) and DL captures the information about technical parameters. Both models will be merged to get CNN-DL model, which will be trained to make quick energy predictions. The developed model will be useful for predicting energy demand of different building configurations within a few seconds. This reduction in time will make the energy prediction information available quickly as expected to make an informed decision at the early design stage.

Preliminary Results and Conclusions

(max 200 words)

The approach is tested on a medium-size office building in Munich. We developed the EnergyPlus model and calibrated against the real energy consumption of a three-storey office building. Then, it is extended to predict the energy demand of various shapes and combinations of design parameters. We generated the training dataset of 5000 samples and test dataset of 1000 samples. Building samples in both training and test dataset have random rectilinear floor plans. We trained CNN-DL model for several combination of hyper-paramters and the best model based on the least validation loss is selected. The model shows an accuracy of 0.983 in terms of R2 and root mean square error (RMSE) of 7 MWh/a.

We integrated these energy prediction model into a BIM tool to make energy prediction under uncertainty. The developed Approach makes probabilistic energy prediction five design options with 500 samples within a minute with the accuracy. Thus, in this research, we developed an approach to make use of energy prediction results to steer the design decisions at the early design stage.

Main References

(max 200 words)

Amasyali, K., Gohary, N., & El-Gohary, N. M. (2018). A review of data-driven building energy consumption prediction studies. Renewable and Sustainable Energy Reviews, 81, 1192–1205. https://doi.org/10.1016/j.rser.2017.04.095

Clarke, J., & Hensen, J. (2015). Integrated building performance simulation: Progress, prospects and requirements. Building and Environment, 91, 294–306. https://doi.org/10.1016/j.buildenv.2015.04.002

Geyer, P., & Schlüter, A. (2014). Automated metamodel generation for Design Space Exploration and decision-making – A novel method supporting performance-oriented building design and retrofitting. Applied Energy, 119, 537–556. https://doi.org/10.1016/j.apenergy.2013.12.064

Lee, B., Pourmousavian, N., & Hensen, J. L. M. (2016). Full-factorial design space exploration approach for multi-criteria decision making of the design of industrial halls. Energy and Buildings, 117, 352–361. https://doi.org/10.1016/j.enbuild.2015.09.028

Picco, M., Lollini, R., & Marengo, M. (2014). Towards energy performance evaluation in early stage building design: A simplification methodology for commercial building models. Energy and Buildings, 76, 497–505. https://doi.org/10.1016/j.enbuild.2014.03.016

Singh, M. M., & Geyer, P. (2019). Statistical decision assistance for determining energy-efficient options in building design under uncertainty. In P. Geyer, K. Allacker, M. Schevenels, F. De Troyer, & Pieter Pauwels (Eds.), 26th International Workshop on Intelligent Computing in Engineering. http://ceur-ws.org/Vol-2394/paper08.pdf



10:48 - 11:06

Application of MyBEM, a BIM to BEM platform, to a building renovation concept with solar harvesting technologies

Mathias Bouquerel1, Kevin Ruben Deutz1, Benoît Charrier1, Thierry Duforestel1, Mickael Rousset1, Bart Erich2, Gerrit-Jan van Riessen3, Thomas Braun4

1EDF R&D, Moret-Loing-et-Orvanne, France; 2TNO, Eindhoven, Netherlands; 3Emergo, Netherlands; 4Pilkington Germany, Gelsenkirchen, Germany

Aim and Approach

(max 200 words)

MyBEM is an innovative and modular platform designed to generate and simulate a Building Energy Model (BEM) from a Building Information Model (BIM), through an automated workflow. The platform is partly the result of the MERUBBI ANR project [1]. Building energy modelling is based on the Modelica library BuildSysPro [2].

The platform has been used to assess the energy performance of a case study within the ENVISION H2020 project [3][4]. The Envision project aims to develop and demonstrate a building renovation concept, which integrates solar harvesting technologies on the whole building envelope, including opaque and glazed parts of the vertical façades.

This case study required the development of specific models for solar harvesting elements (opaque solar collectors and PV windows), and their integration in the building energy model. The simulation of both the energy demand and the energy production can be simulated, to assess how local production can meet the energy needs.

Scientific Innovation and Relevance

(max 200 words)

The design of MyBEM is modular, in order to allow as much flexibility as possible [5]. The platform functionalities are handled by several independent modules, which communicate through file exchange. Open standards are used as much as possible for the purpose of interoperability.

A pre-processing module is able to import geometrical and construction data from various sources, and to produce a gbXML file – an extraction of BIM files dedicated to building energy modeling. This module is also able to run annual solar calculations through efficient ray tracing algorithms, so that for each façade element, solar masking and solar reflections are accurately taken into account [6].

A second module generates a building energy model in Modelica from the previously exported gbXML file. The high versatility of the Modelica library BuildSysPro is used to adapt the model to the specific needs of the Envision project study case. Indeed, Modelica models for solar harvesting technologies have been developed and integrated in the model generation algorithm.

The Modelica model is then simulated under Dymola. Post-processing routines are used to extract energy and bioclimatic indicators from the simulation.

Preliminary Results and Conclusions

(max 200 words)

The MyBEM platform has been applied to a demonstration building of the Envision project. It is a 3-storey residential building with 24 dwellings, located in the Netherlands, and built in the 1950s. The renovation configuration integrates:

- Insulation of the façades, roof and lower floor

- Change of windows

- Closure of balconies

- Integration of solar collectors on the vertical façades

- Integration of PV windows on the staircase

From a SketchUp geometric file and additional technical data, a gbXML file has been automatically generated within the pre-processing module. From this gbXML, the Modelica model has also been automatically generated.

Then simulations with Dymola have been carried out to calculate the energy needs and the energy production through solar harvesting. The coupling of both the space heating system and the solar harvesting system (based on solar collectors) is used to assess the self-consumption potential, the remaining energy needs that require an additional energy source and the unused harvested energy that can be injected in the energy networks (electrical grid or district heating network).

Main References

(max 200 words)

[1] Mathieu Schumann et al. Interdisciplinarity around design tools for new buildings and districts: the ANR MERUBBI project. Proceedings of 33rd PLEA International Conference, pages 2148-2155, Edinburgh, UK, July 2017.

[2] Gilles Plessis, Aurélie Kaemmerlen, Amy Linday. BuildSysPro: a Modelica library for modelling buildings and energy systems. Proceedings of Modelica Conference 2014, Lund, Sweden, March 2014

[3] ENVISION Project description : https://cordis.europa.eu/project/id/767180/fr

[4] Bart Erich et al. Energy harvesting by invisible solar façade collector. 14th Conference on Advanced Building Skins, Bern, Switzerland, Oct. 2019.

[5] Mathias Bouquerel, Sébastien Bermes, Adrien Brun, Hassan Bouia, Régis Lecussan, Benoît Charrier. Building Energy Modeling at District Scale through BIM Based Automatic Model Generation – Towards Building Envelope Optimization. Proceedings of Building Simulation 2019 Conference, Roma, Italy, Sept. 2019.

[6] Clément Ribault, Mathias Bouquerel, Adrien Brun, Mathieu Schumann, Gilles Rusaouen, Etienne Wurtz. Assessing tools relevance for energy simulation at the urban scale: towards decision-support tools for urban design and densification. Energy Procedia 122:871-876, September 2017



11:06 - 11:24

Optimal energy system sizing with independent load and weather time series.

François Lédée1,2, Gaëlle Faure1,2, Curran Crawford1,3, Ralph Evins1,2

1Institute for Integrated Energy Systems, University of Victoria, British Columbia, Canada; 2Energy in Cities group, Department of Civil Engineering, University of Victoria, British Columbia, Canada; 3Sustainable Systems Design Lab, Department of Mechanical Engineering, University of Victoria, British Columbia

Aim and Approach

(max 200 words)

To size an energy system with stochastic programing, hourly energy demand and weather variables over multiple years are required. Unlike weather data, hourly building’s energy demand over multiple years is usually not available. Therefore, some authors [1,2] highlight the possible use of stochastic generators. These algorithms do not jointly generate load and weather data [1,2,3], resulting in a loss of their natural correlation, possibly affecting further applications.

This study aims at assessing if an energy system sizing ran with an energy hub, an optimization framework for multi-source energy systems [4], can be conducted with independently stochastically generated time series. In this aim, we investigate if the proposed sizing for a reference simulation, where the building electric load and the solar data used are issued from the same year, differs significantly from other simulations, where the years of employed data differ. If so, we would conclude that jointly generated time series are required for multi-source energy system sizing.

We focus on the sizing of energy systems comprising a PV installation, for residential buildings in different climate zones. For each building, one year of hourly electrical load is used along with 20 different years of hourly solar data of the same location.

Scientific Innovation and Relevance

(max 200 words)

Optimization-based energy models is a growing research area aiming at supporting strategic energy planing at various scales and for different time horizons. These models require the use of time series, commonly with an hourly granularity, for example for the energy demand or the resource's availability.

Most of the commonly used models do not consider uncertainty and rely whether on past data or forecasts, often inaccurate [5]. To deal with optimization under uncertainty, two main approaches are considered: the use of robust optimization or stochastic programming. While the first “aims to find a solution with the best worst performance” [6], the second “optimizes the expected value of the objective over all possible realizations” [5].

In a multi-source energy model, the time series represent the most important parameters to consider. In an energy system, the demand and the weather conditions are often related. The current work proposes to investigate the importance of this relation in the optimization of an energy system, to see if its optimal sizing through stochastic programming can be based on the use of independent weather and building's load stochastic generators.

Preliminary Results and Conclusions

(max 200 words)

Our first results regarding a residential building in Austin (Texas, USA) reveal a undersizing of the solar facilities of 10% in average with the use of random years of sun availability, compared to the reference simulation. The sizing of other energy facilities varies between -6% and +10% and expected costs are slightly overestimated (+1.5% in average).

To assess if these differences are significantly acceptable, two analysis approaches are adopted. The first one is based on statistical tests. For a p-value of 5%, the results for the first simulated case reveals no significant difference between the reference case and other simulations for the cost and all facility size, except for the installed capacity of photovoltaic modules.

A second analysis is based on geometric approaches. Both a principal component analysis and a clustering-based analysis fail at rejecting the hypothesis of a significant overall difference between the reference simulation and other optimization sizing processes.

We will consolidate these preliminary results by multiplying the cases of study in various climate zones and for buildings with different characteristics. This would help drawing appropriates conclusions regarding the influence of the use of independent time series on the sizing of multi-source energy systems.

Main References

(max 200 words)

[1] Sun et al. (2019), Using Synthetic Traces for Robust Energy System Sizing, e-Energy 19, pp. 251-262.

[2] Sharafi M, ElMekkawy TY. Stochastic optimization of hybrid renewable energy systems using sampling average method. Renewable and Sustainable Energy Reviews. 1 déc 2015;52:1668‑79.

[3] Patidar S, Jenkins DP, Peacock A, McCallum P. Time Series decomposition for simulating electricity demand profile. In: Proceedings of Building Simulation 2019: 16th Conference of IBPSA. Rome (Italy); 2019.

[4] Evins et al. (2014), New formulations of the ‘energy hub’ model to address operational constraints. Energy 73, pp. 387–398.

[5] Moret et al (2020), Decision support for strategic energy planning: A robust optimization framework. European Journal of Operational Research 280, pp. 539-554.

[6] Moazeni et al. (2019), A Risk-Averse Stochastic Dynamic Programming Approach to Energy Hub Optimal Dispatch. IEEE Transactions on power systems 34 (3), pp. 2169-2178.



11:24 - 11:42

Evaluating the Performance of different Window Opening Styles for single-sided buoyancy-driven natural Ventilation using CFD Simulations

Akshit Gupta, Annamaria Belleri, Francesco Babich

Institute for Renewable Energy, Eurac Research, 39100 Bolzano, Italy

Aim and Approach

(max 200 words)

The aim of this research is to investigate the effectiveness of different types of windows such side-hung, top-hung, and horizontal pivot for natural and mixed-mode ventilation. This paper will focus on the CFD analysis that was performed as a preliminary step to better define the windows’ prototypes before being tested in full-scale experimental facilities.

In this study, both wind- and buoyancy-driven ventilation was considered. For both cases, the reference room included in the CFD model was one of the chambers (4m x 8m x 3m, width x depth x height) of the lab that will be later used for the tests. However, different levels of simplifications and different ways to model the boundary conditions were tests.

Firstly, the use of an external air domain was investigated to evaluate whether this was required or not, what would be its optimal size and its most appropriate boundary conditions. Secondly, different levels of geometrical simplification of the opening were compared. Thirdly, transient and steady-state simulations were performed.

For this CFD analysis, the preliminary validation of the models was based on similar previously published studies. A further and more comprehensive validation will be made after the completion of the tests with the prototypes.

Scientific Innovation and Relevance

(max 200 words)

Natural and mixed-mode ventilation (i.e. combination of natural and mechanical ventilation) are effective means to provide comfortable indoor environments (such as thermal comfort and good indoor air quality) while minimizing the energy consumption [1,2,3,4].

However, the use of different types of windows and control strategies usually lead to different indoor thermal conditions [1,5]. CFD is a powerful modelling technique to compare the air distribution within a room for varying scenarios. The boundary conditions and modelling assumptions should be carefully evaluated to ensure that the differences among the simulated scenarios are due to variations in the actual systems, and not due to other aspects such as oversimplification of the geometry or driving forces [2,4]. For instance, wind pressure coefficients are often given on the plane of the wall, but the same wind is likely to have a different indoor effect depending on the type of opening and its position (e.g. outer edge, inner edge or in the middle of a wall) [1,2,4,6].

Hence, this work aims to make a step forward by showing the effect of using different modelling assumptions and opening geometrical simplifications, in terms of results’ accuracy and required computational power.

Preliminary Results and Conclusions

(max 200 words)

In all cases, initial results indicate that, the use of an external air domain provide better results when comparing the CFD predictions with published similar studies [1,3,4].

For buoyancy-driven flows, the size of the external domain may be limited, while its boundary-conditions play an essential role. The most realistic results were achieved with an external domain as big as the testing chamber. The boundary conditions of the external domain with the top surface as opening and all the others as No-slip were most reasonable.

For wind-driven flows, the complexity increases depending on the wind direction and speed value, and on the position of the simulated room within the buildings (e.g. ground floor, mid floor, on an edge, etc.) [3,7].

Once the initial conditions are correctly defined, transient simulations usually provide the most accurate results as they can better capture the unsteady and cyclic nature of these flows, which is especially true for buoyancy-driven flows.

Lastly, the oversimplification of the geometry of walls and windows (e.g. overlooking the wall thickness, or the way in which the opened and closed sections of a window are modelled) is likely to cause unreliable results, and therefore to question the validity of the CFD analysis.

Main References

(max 200 words)

1. Wang, J., Wang, S., Zhang, T., & Battaglia, F. (2017). Assessment of single-sided natural ventilation driven by buoyancy forces through variable window configurations. Energy and buildings, 139, 762-779.

2. Cook, M. J., Ji, Y., & Hunt, G. R. (2003). CFD modelling of natural ventilation: combined wind and buoyancy forces. International Journal of Ventilation, 1(3), 169-179.

3. Gan, G. (2010). Simulation of buoyancy-driven natural ventilation of buildings—Impact of computational domain. Energy and Buildings, 42(8), 1290-1300.

4. Allocca, C., Chen, Q., & Glicksman, L. R. (2003). Design analysis of single-sided natural ventilation. Energy and buildings, 35(8), 785-795.

5. von Grabe, J., Svoboda, P., & Bäumler, A. (2014). Window ventilation efficiency in the case of buoyancy ventilation. Energy and Buildings, 72, 203-211.

6. Favarolo, P. A., & Manz, H. (2005). Temperature-driven single-sided ventilation through a large rectangular opening. Building and Environment, 40(5), 689-699.

7. Bangalee, M. Z. I., Lin, S. Y., & Miau, J. J. (2012). Wind driven natural ventilation through multiple windows of a building: A computational approach. Energy and Buildings, 45, 317-325.



11:42 - 12:00

Beyond normal: Guidelines on how to identify suitable model input distributions for building performance analysis

Giorgos Petrou1, Anna Mavrogianni2, Phil Symonds2, Mike Davies2

1UCL Energy Institute, United Kingdom; 2UCL Institute for Environmental Design and Engineering, United Kingdom

Aim and Approach

(max 200 words)

This work presents a step-by-step guide on how to identify the probability distribution function that best describes a given dataset of building-related parameters. We demonstrate the process for a set of wall U-value measurements. Firstly, histograms and cumulative distribution plots are used to visualise the data to establish whether the data appears to be normally distributed. Second, data cleaning is performed through observation of histogram extremes and removal of outliers. This is preferred over automated procedures based on the data’s interquartile range or standard deviation when the data does not appear to be normally distributed. Third, a set of candidate distributions are selected using the data’s empirical distribution and the ‘Cullen and Frey’ graph of kurtosis and square of skewness. Next, the candidate distributions are then fitted to the data using Maximum Likelihood Estimation – this is easily achieved with the R package fitdistrplus [1]. Finally, drawing from Information Theory, the Akaike Information Criterion (AIC) and its derivates are used to identify the best fitting distribution [2]. Density plots, Q-Q plots, P-P plots and Cumulative Distribution Function plots provide a supplementary measure of goodness-of-fit and inform the modeller whether the best-fitting distribution is satisfactory.

Scientific Innovation and Relevance

(max 200 words)

The importance of uncertainty propagation and model calibration in the built environment is widely recognised and distributions are an integral part of this process [3]. The normal distribution is commonly assumed in building performance simulation due its convenience and familiarity. Similarly, the uniform distribution is often used to express lack of knowledge about the possible value or distributional form of a model input. However, with data availability on the rise, distributions used for uncertainty quantification could in some cases be based on empirical evidence. If the modeller identifies the distribution that best describes the observed data relating to a model input, they can capture its expected value and shape more accurately. Alternative, the use of inappropriate distributions could contribute to the ‘performance gap’. Examples of using non-normal and non-uniform distributions exist within the field of building modelling [4, 5], however, no clear guidance on how to identify the most suitable distributions for a given dataset could be found. Therefore, by providing a detailed, step-by-step guidance of this process using code snippets in R, this paper aims to enable building performance modellers to make the best use of available data, potentially improving their modelling workflow and the accuracy of their predictions.

Preliminary Results and Conclusions

(max 200 words)

We demonstrate the step-by-step process using open source wall U-value data from English homes [6]. The data cleaning process did not reveal any outliers. Based on the empirical distribution and the Cullen and Frey graph, the candidate distributions were chosen to be the normal, Weibull, gamma and lognormal. The AIC was lowest for the gamma distribution with a value of -16.1. The difference in AIC between the best fitting distribution and the normal was 8, suggesting fairly weak support for the normal distribution to be a plausible alternative to the gamma [2]. By estimating the Akaike Weights, the probability that the gamma best describes the data amongst the candidate distributions was 0.75 while the normal had a probability of 0.01. As AIC is a relative goodness-of-fit measure, Q-Q plots, P-P plots and Cumulative Distribution Function plots were used to confirm that the gamma provides both the best fit amongst the candidate distributions and a sufficiently good fit for the given data and its intended use. To conclude, this paper shows that an alternative distribution better represents wall U-values in English homes than the more popular normal distribution. The steps taken could be followed by others to improve their modelling assumptions.

Main References

(max 200 words)

[1] Delignette-Muller ML, Dutang C. fitdistrplus: An R Package for Fitting Distributions. J Stat Softw 2015; 64: 1–34.

[2] Burnham KP, Anderson DR. Multimodel Inference: Understanding AIC and BIC in Model Selection. Sociol Methods Res 2004; 33: 261–304.

[3] Tian W, Heo Y, de Wilde P, et al. A review of uncertainty analysis in building energy assessment. Renew Sustain Energy Rev 2018; 93: 285–301.

[4] Kristensen MH, Choudhary R, Pedersen RH, et al. Bayesian Calibration Of Residential Building Clusters Using A Single Geometric Building Representation. 2017; 10.

[5] Booth AT, Choudhary R, Spiegelhalter DJ. Handling uncertainty in housing stock models. Build Environ 2012; 48: 35–47.

[6] Hulme J, Doran S. In-situ measurements of wall U-values in English housing. 290–102, Building Research Establishment (BRE) on behalf of the Department of Energy and Climate Change (DECC), 2014.