Conference Agenda

Overview and details of the sessions of this conference. Please select a date or location to show only sessions at that day or location. Please select a single session for detailed view (with abstracts and downloads if available).

Please note that all times are shown in the time zone of the conference. The current conference time is: 19th May 2022, 13:32:29 CEST

 
 
Session Overview
Session
Session F2.7 (Online Track): Buildings paving the way for the energy transition / optimisation /control
Time:
Friday, 03/Sept/2021:
10:30 - 12:00

Session Chair: Raul Fernando Ajmat, University of Tucuman
Location: Virtual Meeting Room 3

External Resource: Click here to join the Zoom Meeting
Show help for 'Increase or decrease the abstract text size'
Presentations
10:30 - 10:48

Prediction of HVAC loads at different spatial resolutions and buildings using deep learning models

Antonio Liguori1, Shiying Yang1, Romana Markovic2, Thi Thu Ha Dam1, Andreas Wagner2, Jérôme Frisch1, Christoph van Treeck1

1RWTH Aachen, Germany; 2KIT, Germany

Aim and Approach

(max 200 words)

This paper explores the applicability of deep learning-driven heating ventilation and air conditioning (HVAC) energy consumption models for agnostic commercial buildings. For that purpose, the modeling is conducted using two data sets. The first data set consisted of the energy consumption profiles from a 12-floor office building located in Seattle, USA. The second data set is collected on a significantly different 3-floor office building located in Aachen, Germany.

The modeling approach relies on recurrent neural networks (RNNs), while the input consists of physical data streams such as indoor air temperatures and data obtained from the central HVAC. The research steps include the implementation of an existing RNN-based model for energy consumption and further model optimization using training and validation set. Eventually, the final model was evaluated using the data from two data sets and the evaluation performance was evaluated in case of the varied spatial and system granularities. For that purpose, the energy consumption boundaries were defined as a) HVAC consumption in single office b) multiple thermal zones (floor-wise consumption) as well as c) building-wise HVAC consumption.

Scientific Innovation and Relevance

(max 200 words)

Significant proportion of the worldwide energy consumption is caused by buildings. In particular, HVAC systems are responsible for 50 % of the total building energy consumption in United States (Pérez-Lombard et al. (2008)), while the similar trend is confirmed across different locations (Al Amoodi and Azar, (2018); Knight, (2012)). Due to its’ amplitude and autoregressive properties, the energy consumption required for HVAC represents a major optimization potential towards a better use of renewable energy and more energy efficient buildings.

One of the approaches to achieve the latter two goals is to predict the energy improve the prediction of the required energy loads. In past, the energy loads were commonly predicted by applying analytical simulation-based approaches. However, such physical simulation models are not well scalable to a large number of buildings and they require extensive model calibration. As an alternative, data-driven methods present a promising approach that is more applicable to agnostic buildings or energy prediction system boundaries (such as thermal zone, building or district level). In that regard, the presented deep RNN model is identified as a suitable modeling approach that leads to satisfying predictive accuracy.

Preliminary Results and Conclusions

(max 200 words)

The results showed, that the existing RNN-based model proposed by Chen et al. (2017) is well applicable to different buildings, given repeated model training using data from the target domain (i.e. data set from Germany). The accuracy for building-wise HVAC energy consumption ranged between 3 % and 8 % in terms of normalized root mean squared error, while the spatial granularity had a direct impact on the predictive performance. Furthermore, the data analysis showed that the energy consumption at more coarse spatial resolution is less stochastic and it shows higher autocorrelation. According to the conducted analytics, these properties could be accounted for simplier model formulation and improved accuracy in case of the building-wise predictive modeling. Further results as well as conclusions will be presented in the main paper.

Main References

(max 200 words)

Al Amoodi, A., & Azar, E. (2018). Impact of Human Actions on Building Energy Performance: A Case Study in the United Arab Emirates (UAE). Sustainability, 10(5), 1404.

Chen, Y., Shi, Y., & Zhang, B. (2017). Modeling and optimization of complex building energy systems with deep neural networks. In 2017 51st Asilomar Conference on Signals, Systems, and Computers (pp. 1368-1373). IEEE.

Knight, I. P. (2012). Assessing electrical energy use in HVAC systems. REHVA Journal (European Journal of Heating, Ventilating and Air Conditioning Technology), 49(1), 6-11.

Kusiak, A., Li, M., & Tang, F. (2010). Modeling and optimization of HVAC energy consumption. Applied Energy, 87(10), 3092-3102.

Pérez-Lombard, L., Ortiz, J., & Pout, C. (2008). A review on buildings energy consumption information. Energy and buildings, 40(3), 394-398.



10:48 - 11:06

Data-driven calibration of joint building and HVAC dynamic models using scalable Bayesian optimization

Ankush Chakrabarty1, Emilio Maddalena2, Hongtao Qiao1, Christopher Laughman1

1Mitsubishi Electric Research Labs, United States of America; 2École polytechnique fédérale de Lausanne, Switzerland

Aim and Approach

(max 200 words)

Physics-informed simulation models of heating, ventilation, and cooling (HVAC) systems play a critical role in predicting system dynamics and enabling analysis, control, and optimization of buildings and equipment. The predictive performance of these simulation models is strongly linked to calibration mechanisms: algorithms that systematically select parameter values that optimize a given calibration-cost map (e.g., L-2 error). Poorly selected parameter values typically result in large deviations between measured building data and simulated data, limiting the utility of the simulation model in subsequent design.

State-of-the-art calibration methods explore the parameter space by computing numerical gradients that are susceptible to measurement noise or employ population-based search mechanisms that require exorbitant data. To improve robustness and curtail data requirements, one can ‘learn’ or approximate the calibration-cost map and subsequently leverage the topology of the approximated function to find good search directions despite noisy measurements.

Concretely, we employ machine learning to construct a calibration-cost map to direct model calibration for systems with joint dynamics of buildings and HVAC equipment. The learner explores subregions of the parameter space with high uncertainty and queries the model only where collecting simulation data yields useful information. This leads to lower simulation data-requirements compared to widely used calibration mechanisms.

Scientific Innovation and Relevance

(max 200 words)

Calibration-cost maps are not always differentiable/convex due to coupled interactions in joint building/equipment dynamics. Furthermore, measured data is corrupted by environmental and process noise, limiting the effectiveness of gradient-based methods. Population-based, gradient-free searches are effective, but incur high computational expenditure as they require an exorbitant number of simulations. Dynamical estimators such as Kalman filters also underperform due to multi-rate dynamics and limited generalizability of state-space models used to design these estimators.

Our contributions are as follows.

- We study the problem of model calibration for joint building/equipment dynamics.

- We employ data-driven Gaussian processes (GPs) for learning a parameter-to-calibration-cost function that contains the true cost with high probability. The GP also generates confidence bounds around predicted function values that quantify the prediction uncertainty at various regions in the parameter space.

- We utilize confidence bounds to explore the parameter space without requiring exorbitant simulations by collecting simulation data for parameters in regions with large uncertainty bounds and high likelihood of containing global optima, both of which are estimated by designing an appropriate acquisition function (as per standard Bayesian optimization).

- After sufficient exploration, we exploit the topology of the learned function to obtain optimal parameters without requiring additional simulation data.

Preliminary Results and Conclusions

(max 200 words)

We employ Modelica to design a physics-informed model of a building system with closed-loop controllers embedded at the equipment level. We calibrate solar and infrared emissivity coefficients for the roof along with an infiltration/volumetric flow rate of the building. The calibration is performed using measured temperature data (space and ambient) obtained at 15-minute intervals over 7 days. A challenging aspect of using temperature data to calibrate building parameters is that the calibration-cost map is confounded by environmental and climactic factors. Despite this difficulty, our proposed Bayesian optimization method estimates these 3 parameters to >95% accuracy within 100 simulation runs. In contrast, gradient-based methods report <80% accuracy with 2 parameters and diverge when searching for all 3. Population-based methods with a modest population size (e.g., 20 agents) require >100 iterations (thus, >2000 simulation runs) to find good estimate, which is far more computationally demanding than our proposed methodology.

Encouraged by preliminary results using Bayesian optimization for selecting building parameters, our ongoing efforts are devoted to developing and testing scalable Bayesian optimization frameworks in joint models of buildings and equipment. Evolution of system dynamics at multiple timescales, along with coupled nonlinear interactions between these systems, makes this joint problem particularly challenging.

Main References

(max 200 words)

[1] D. Coakley, P. Raftery, M. Keane, A review of methods to match building energy simulation models to measured data, Renewable Sustainable Energy Rev., Vol. 37, pp. 123–141, 2017.

[2] A.-T. Nguyen, S. Reiter, P. Rigo, A review on simulation-based optimization methods applied to building performance analysis, Appl. Energy, Vol. 113, pp. 1043–1058, 2014.

[3] S. Asadi et al., Building energy model calibration using automated optimization-based algorithm, Energy and Buildings, Vol. 198, pp. 106–114, 2019.

[4] J. Snoek, H. Larochelle, and R. P. Adams. Practical Bayesian optimization of machine learning algorithms. Adv. in Neural Information Processing Systems, pp. 2951-2959. 2012.

[5] A. Chakrabarty, M. Benosman. Safe Learning-based Observers for Unknown Nonlinear Systems using Bayesian Optimization. arXiv preprint arXiv:2005.05888 (2020).

[6] C. R. Laughman, C. Mackey, S. A. Bortoff, and H. Qiao. Modeling and Control of Radiant, Convective, and Ventilation Systems for Multizone Residences. Proc. 16th IBPSA Conference, 2019.



11:06 - 11:24

A performance-driven simulation workflow for PV integration into the design process: application on an innovative building project in Switzerland

Sergi Aguacil Moreno

Building2050 group, Ecole Polytechnique Fédérale de Lausanne (EPFL), Fribourg, Switzerland

Aim and Approach

(max 200 words)

In the construction sector, the integration of active elements functioning both as building envelope material and on-site electricity generator is identified as a key measure to achieve the 2050 targets and carbon neutrality [1,2]. Despite an increasing confluence of the photovoltaic (PV) industry and the building glass manufacturers, which offers high design freedom in relation to size, colour and texture, architects and engineers continue to address the issue in a traditional way, by limiting themselves to the application of "standard" catalogue products that particularly constrain their designs [3,4]. This way of working, which makes it difficult to integrate PV systems into the building envelope and, therefore, into the overall building design, often leads to a method of sizing that largely makes abstraction of the building’s electricity needs. This article presents the application – in an ongoing design process – of an innovative active-surface selection method adapted to the design phases using 3D-modeling and hourly-step simulations. Through an integrated-design approach, the proposed workflow allows quickly obtaining visual and quantitative results on the building envelope, and supports a design decision-making process that proposes an alternative approach to usual practice.

Scientific Innovation and Relevance

(max 200 words)

The literature review shows a lack of reliable design-driven methods to support the sizing and implementation of building-integrated (BI)PV installations in projects [5,6]. Most existing methods are highly time-consuming and based on complex optimisation algorithms not tailored to the workflow of architects/designers [7].

We propose the application of a novel method that relies on BIM-integrated parametric tools to allow architects to get instant feedback on the impact of their decisions on building energy performance along the design process. The method consists in a streamlined and automated workflow supporting BIPV design (sizing and positioning) in coherence with the building’s context, architectural features, and electricity needs. It uses the Rhino.Inside Revit add-on [8], an open source Rhino WIP project that allows Rhino [9] and Grasshopper [10] to run inside other 64-bit Windows applications such as Revit, by which we are able to apply, to any Revit informed-3D model, our Grasshopper definition for the calculation and optimisation of PV systems upon building envelopes. This in-depth analysis from the early design stages aims to reduce the costs due to design changes [11]. The method’s added-value is here tested in a real project and compared to more common PV-sizing approaches.

Preliminary Results and Conclusions

(max 200 words)

Our method was applied in the ongoing design of the future Smart Living Lab building in Fribourg, Switzerland. Prior to this test, we observed that the design of the PV installation for the building tended to follow a common approach, i.e. adjust standard, building-applied (BA)PV panels on the least visible surfaces (roof), while aiming to produce the maximum amount of energy possible. This approach is partly explained by the desire to achieve the Minergie-A label [12], which requires the production of more energy than the building’s needs in absolute annual value. Moreover, the most economical solution was to be favoured, to the detriment of architectural integration. The method was then applied, focusing on reaching a self-sufficiency and self-consumption adapted to the needs of the building, regardless of the annual amount of energy produced. This resulted in a better-adapted installation that not only produces the needed energy, but also reduces the building’s environmental impact since the integrated panels substitute part of the materials that would have been otherwise used for the building envelope. Results show that although the building is compact and located in a dense urban context, high-level of performance such as the 2000-Watt society targets for 2050 are reachable.

Main References

(max 200 words)

[1] S. Aguacil, S. Lufkin, E. Rey, Active surfaces selection method for building-integrated photovoltaics (BIPV) in renovation projects based on self-consumption and self-sufficiency, Energy Build. 193 (2019).

[2] S. Aguacil, Architectural Design Strategies for Building Integrated Photovoltaics (BIPV) in Residential Building Renovation (Thesis N°9332), EPFL Lausanne, (2019).

[3] P. Bonomo et al., BIPV: building envelope solutions in a multi-criteria approach. A method for assessing life-cycle costs in the early design phase, Adv. Build. Energy Res. 11 (2017) 104–129.

[4] I. Zanetti et al., BIPV- Product overview for solar building skins- Status Report, (2017).

[5] C. Ballif et al., Integrated thinking for photovoltaics in buildings, Nat. Energy. 3 (2018) 438–442.

[6] P. Heinstein et al., Building Integrated Photovoltaics (BIPV): Review, Potentials, Barriers and Myths, Green. 3 (2013).

[7] T. Østergård et al., Building simulations supporting decision making in early design - A review, Renew. Sustain. Energy Rev. 61 (2016) 187–201.

[8] Robert McNeel&Associates, Rhino.Inside, (2020). https://www.rhino3d.com/inside.

[9] Robert McNeel&Associates, Rhinoceros software, (2020). https://www.rhino3d.com/.

[10] S. Davidson, Grasshopper, Algorithmic modelling for Rhino, (2020). http://www.grasshopper3d.com/.

[11] A. Hollberg, A parametric method for building design optimization based on Life Cycle Assessment, Bauhaus-Universität Weimar, (2016).

[12] Minergie, Minergie standard certification label, (2020)



11:24 - 11:42

An inquiry into the accuracy of the energy model calibration process

Ipek Yilmaz1, Burak Gunay1, Guy Newsham2, Adam Wills2

1Carleton University, Canada; 2National Research Council, Canada

Aim and Approach

(max 200 words)

This paper aims at quantifying the uncertainty inherent in the building energy model calibration process. The following research questions are investigated: (a) how accurately can the unknown parameters of a calibrated energy model be estimated through the model calibration process?; (b) how do the imperfections in the model calibration process affect the calibrated model’s ability to make operational decisions? To answer these questions, we needed a building with all thermophysical properties accurately characterized. To this end, a three-storey office building in Ottawa, Canada is simulated with different occupancy and envelope property scenarios to generate metered energy data for calibration. Subsequently, several input parameters (e.g., air infiltration, thermal conductances) are then assumed unknown. The model is coupled to a custom optimization script which searches for these unknown parameters by minimizing the deviation to the metered energy data subject to practical and physical constraints. The process is repeated with monthly and hourly metered energy data while introducing varying levels of systematic error to the meter data. The performance of the model calibration process is assessed based on the optimization algorithm’s ability to estimate the unknown parameters. Further, the calibrated energy models are used to make operational decisions related to air handling units.

Scientific Innovation and Relevance

(max 200 words)

Energy models for existing buildings are needed to make operation and retrofit decisions. However, pure data-driven modelling approaches may not be suitable to many existing buildings due to issues in obtaining metadata and limitations in the sensing and data collection infrastructure. Thus, physics-based energy models calibrated with metered energy use data have gained popularity among the building simulation community. In practice, a calibrated physics-based energy model is largely assumed to be an adequate representation of the actual building when it complies with a measurement and verification standard’s criteria for fitness (e.g., ASHRAE Guideline 14). To this end, the analysis presented in this paper advances our understanding about the uncertainty in the energy model calibration process. It demonstrates the imperfections in the model calibration process and highlights how these imperfections translate into the building performance simulation-based operational decision-making process.

Preliminary Results and Conclusions

(max 200 words)

Preliminary results of this investigation point out that calibration with monthly energy use data does not generate accurate estimates of the unknown parameters – even though the goal was to estimate only seven parameters. Due to the multicollinearity amongst many of the parameters, it was challenging to acquire correct estimates, particularly for the window and wall thermal conductances, and the air-infiltration rate. The unknown parameters were accurately estimated through calibration with hourly energy use data, albeit only when the systemic error in the metered energy data was less than ±5%. Imperfections in the model calibration process appeared to affect only some of the operational decisions. Thus, a calibrated model, despite not being able to adequately represent the thermophysical properties of a building, may support the operational decision-making process.

Main References

(max 200 words)

Gunay, H. B., Ouf, M., Newsham, G., & O'Brien, W. (2019). Sensitivity analysis and optimization of building operations. Energy and Buildings, 199, 164-175.

O’Neill, Zheng, and Bryan Eisenhower. "Leveraging the analysis of parametric uncertainty for building energy model calibration." Building simulation. Vol. 6. No. 4. Springer Berlin Heidelberg, 2013.

Sun, Kaiyu, et al. "A pattern-based automated approach to building energy model calibration." Applied Energy 165 (2016): 214-224.

Lim, Hyunwoo, and Zhiqiang John Zhai. "Influences of energy data on Bayesian calibration of building energy model." Applied Energy 231 (2018): 686-698.

Guyot, Dimitri, et al. "Building energy model calibration: a detailed case study using sub-hourly measured data." Energy and Buildings (2020): 110189.

Chong, Adrian, et al. "Bayesian calibration of building energy models with large datasets." Energy and Buildings 154 (2017): 343-355.

Mihai, Andreea, and Radu Zmeureanu. "Bottom-up evidence-based calibration of the HVAC air-side loop of a building energy model." Journal of Building Performance Simulation 10.1 (2017): 105-123.



11:42 - 12:00

Deep reinforcement learning-based optimal building energy management strategies with photovoltaic systems

Minjeong Sim, Geonkyo Hong, Dongjun Suh

Kyungpook National University, Korea, Republic of (South Korea)

Aim and Approach

(max 200 words)

<Research abstract>

Owing to the spread of solar photovoltaic (PV) systems, a significant amount of research has been conducted on the development of efficient energy management methods. Significantly, the energy operation strategies are essential for residential buildings due to the difference between peak demand and solar power generation time.

Therefore, we proposed a novel deep reinforcement learning-based model considering both, direct use of the generated energy to the buildings and selling to utilities to minimize the building's total energy operating cost in a residential building with PV-energy storage system (ESS) installed.

To verify the performance of the proposed model, case studies such as rule-based, selling-only case, and consumption-only case were conducted, showing that the proposed model minimized energy operating costs.

Scientific Innovation and Relevance

(max 200 words)

-Practical Implications

This study supports a deep reinforcement learning-based optimal PV-ESS management system for residential facilities considering economic efficiency based on several energy management strategies.

Furthermore, this study can provide proper guidelines for deriving the optimal energy management methods of residential buildings equipped with PV-ESS systems, according to various operation strategies, energy policies, and building types.

- Research framework

PV power generation is considered for the purpose of reducing the total operating cost and aims to minimize the operating cost of buildings by lowering the peak of energy consumption of buildings.

The energy management system is then modeled using the Deep Reinforcement Learning(DRL)-based method. The energy stored in the ESS can be determined by discharging used to lower the peak load of the building and selling to grid. Based on this behavior, the operating cost of the building can be optimized.

In summary, PV/ESS scheduling method is presented through time series-based building energy flow analysis. In addition, we derive an energy management algorithm based on deep reinforcement learning through the exchange of PV/ESS and external grid energy.

Preliminary Results and Conclusions

(max 200 words)

This study proposes an optimal operational management model of PV-ESS system based on DRL to minimize the building's total operating costs in residential buildings where a difference between peak demand and solar power generation time is observed. The proposed model minimizes the operating cost of the building through the charging and discharging scheduling of ESS. In the proposed DQN-based model, the ESS agent learns actions until it maximizes the total cumulative reward while preventing overcharging and undercharging.

A simulation-based case study was conducted to verify the performance of the proposed model. The experimental results show that the proposed model reduced the energy operation cost by $21,609 compared to the rule-based approach, confirming the monetary benefits. Thus, the proposed model can provide a guideline for optimal energy management methods of residential buildings equipped with PV-ESS systems.

Further research is needed to improve the PV and ESS optimally linked energy management methods, considering several energy policies in the development of energy management techniques of the general-purpose PV-ESS system.

Main References

(max 200 words)

IEA (2017) Market Report Series Renewables 2017

Gul, M., Kotak, Y. and Muneer, T. (2016) Review on recent trend of solar photovoltaic technology, Energy Exploration and Exploitation

Kim, B. and Suh, D. (2020) ‘A hybrid spatio-temporal prediction model for solar photovoltaic generation using numerical weather data and satellite images’, Remote Sensing, 12(22), pp. 1–21.

Teleke, S. et al. (2010) ‘Rule-based control of battery energy storage for dispatqianng intermittent renewable sources’, IEEE Transactions on Sustainable Energy. IEEE, 1(3), pp. 117–124.

Hong, T. et al. (2014) ‘An economic and environmental assessment for selecting the optimum new renewable energy system for educational facility’, Renewable and Sustainable Energy Reviews. Elsevier, 29, pp. 286–300.



 
Contact and Legal Notice · Contact Address:
Privacy Statement · Conference: Building Simulation 2021
Conference Software - ConfTool Pro 2.6.143+TC
© 2001–2022 by Dr. H. Weinreich, Hamburg, Germany