Case study: CREATE Tower, Singapore
In warm climates, knowledge of wind behaviour around buildings supports building-design efforts to improve natural ventilation, reduce energy consumption and enhance pedestrian comfort. While computational fluid dynamics (CFD) simulations have become an important tool for wind predictions, large differences are often observed between simulation results and field measurements. This project led by Prof Dr Ian Smith of EPFL aims to improve simulation predictions of wind around buildings using field measurements.
The main findings are that wind predictions can be improved using a data interpretation approach that is based on multiple-model falsification and a specially tailored sensor placement strategy.
Computational fluid dynamic (CFD) simulations provide wind predictions and allow efficient parametric studies to be performed. However, predictions are often inaccurate due to uncertainties associated with:
i) modelling complex phenomena such as convection and turbulence
ii) simplification of urban geometries
iii) numerical challenges at wall boundaries
iv) definition of parameter values that are seldom known exactly, including inlet wind conditions.
Model-based data interpretation techniques reduce uncertainties in simulation predictions through identifying likely sets of parameter values using measurements. In this research, a population of CFD models is generated by varying values of input parameters that are not precisely known. In total, 15 parameters related to modelling, geometric simplifications, numerical methods and inlet wind conditions are studied. Sensitivity analysis and parameter selection techniques are employed to avoid computational overload.
In this study conducted at the CREATE Tower in Singapore, three parameters are selected following a sensitivity analysis: wind speed and direction at the atmospheric boundary, as well as building surface roughness. Values of these parameters are varied uniformly within plausible ranges defined by engineering judgment and available literature. 768 sets of parameter values are created and used to run individual CFD simulations. The simulation output is a discrete population of wind predictions at 186 potential measurement locations. Figure 01 presents the wind predictions at 30 meters height predicted with one CFD simulation.
Fig. 01 Wind speed predictions at 30 meter height around the CREATE Tower obtained from a CFD simulation.
Novel hierarchical sensor placement
Good sensor placement is an important task in model-based data interpretation and can ensure that predictions are improved in the most efficient manner. So far in wind studies, sensors have been placed mostly using educated guess and experience. No systematic methodology for sensor placement is available. A challenge is that flow properties vary considerably with space and time and as a result, the position of sensors and their number significantly affect the value of sensor information (Pavageau and Schatzmann, 1999). In addition, the number of feasible measurement locations can be limited (van Hooff and Blocken, 2012).
While no placement methodology is available for wind studies, sensor placement methodologies have been proposed for structural diagnosis. Earlier studies have demonstrated that (information) entropy can be successfully used as a criterion to select sensor locations prior to field measurement (Papadimitriou, 2004; Robert-Nicoud et al., 2005). In these studies, sensors were placed sequentially at positions that provided high values of entropy in model predictions. Sequential sensor placement is preferred over global search strategies, such as genetic algorithms, because of lower computational cost (Papadimitriou, 2004). However, most strategies propose redundant sensors that provide similar information content.
This research studies a novel sensor placement methodology using a hierarchical strategy and joint-entropy (Fig. 02). The methodology is applied to a discrete population of wind predictions obtained from multiple CFD simulations. Optimal sensor configurations are identified prior to field measurement assuming limited knowledge of wind behaviour around buildings. Configurations are identified by considering common information amongst sensors as well as measurement and modelling uncertainty.
Fig. 02 Flowchart of the measurement-system-design methodology
The methodology is applied to a full-scale case study at the CREATE Tower (Fig. 03). Optimal sensor configurations are evaluated for their ability to support model-based data interpretation and reduce the number of models and the prediction range, as well as increase the accuracy of predictions. Results show that a hierarchical sensor-placement strategy using joint entropy leads to sensor configurations that improve predictions of wind speed at unmeasured locations near buildings.
Fig. 03 Plan (left) and front view (right) of potential sensor locations displayed in the simulation environment; the red marks represent the potential locations and the yellow marks the selected configuration.
Model-based data interpretation
Model-based data interpretation allows candidate models to be inferred from measurements. Model calibration, whereby an “optimal” model is found by minimising the difference between simulation predictions and measurement data, is not appropriate for model-based data interpretation. While calibration approaches might give good predictions at sensor locations, they might not provide good predictions at unmeasured locations (Beven, 2008).
An alternative approach is model falsification, such as error-domain model falsification (Goulet et al., 2013), in which incorrect sets of parameter values are falsified. Error-domain model falsification was developed for applications of bridge diagnosis and leak detection. Error-domain model falsification involves falsification of CFD models for which the difference between measurement data and simulation predictions, for any measurement location, is larger than an estimate of uncertainty bounds. In this work, error-domain model falsification is adapted for time-dependent situations, such as wind behaviour around buildings.
The information content on wind behaviour obtained from measurement data depends on levels of measurement and modeling uncertainties at sensor locations. Strategies are proposed to evaluate sources of modeling uncertainties including
1) uncertainties associated with the use of the steady Reynolds-averaged Navier-Stokes (RANS) equations in predictions of mean wind variables (Vernay et al., 2014),
2) uncertainties associated with turbulent fluctuations and
3) uncertainties associated with thermal processes.
The approach is applied for predicting wind speed and direction around the CREATE Tower. Simulation predictions and field measurements at eight sensor locations are employed (Fig. 04). Results show that after model falsification, ranges of wind-speed predictions at unmeasured locations are decreased by more than 75 %.
Fig. 04 Wind-cup anemometer used to measure wind speed and wind direction around the CREATE Tower.