PREDICTION OF WIND SPEEDS BASED ON DIGITAL ELEVATION MODELS USING BOOSTED REGRESSION TREES |
Paper ID : 1057-SMPR-FULL |
Authors: |
Christophe Etienne1, Jiaojiao Tian2, Thomas Krauss2 1Secquaero Advisors Ltd. 2German Aerospace Center |
Abstract: |
Man-made objects are exposed to a broad range of natural hazards, like extreme rain, floods or winter storms. During the last decades we encounter an increase in the awareness of such geo-risks in the public. Stakeholders like governmental institutions but also private companies like insurances develop zoning systems which map potential dangers. The motivation is to minimize the risk for future investments like infrastructures and settlements to be affected, but also to raise the awareness of human beings who live in endangered regions. The degree of exposure is related to the surrounding orography of the object, the explicit dependencies can just be resolved to a certain degree. Digital Elevation Models and Land Cover maps are descriptions of the real world, and serve as a starting point for simulations in the natural hazard domain by various means. Estimating wind speeds and the occurrence of turbulences over complex terrain is a challenging task, thus different methodical approaches exist to make such predictions. The domain of meteorological models deals with atmospheric boundary layers and the description of interactions in the atmosphere to model the behaviour of airflows, detailed application examples are given in (Stepek & Wijnant, 2011), (Hofherr & Kunz, 2010), (Rooy & Kok, 2004) and (Verkaik, 2006). Computational Fluid Dynamics (CFD) simulations are also suitable for simulating wind speeds, but resolving the Navier-Stokes equation is very resource demanding from a computational point of view. Garcia and Boulanger (Garcia & Boulanger, 2006) use a CFD program for making discrete statements for an area of 108*108 km² covering Mt. St. Helens (USA). Besides of these physical models, techniques from the non-parametric regression domain are suitable for making spatial predictions which describe true world phenomena. As the strict functional dependencies between the independent Digital Elevation Model (DEM)-based parameters and the measured wind speeds are of no necessity for such models, such approaches are straight forward. Lehmann et al. (2002) developed a software package for making spatial predictions within the methodology of non-parametric regression by using Generalized Additive Models. In theirwork the conclusion is drawn that non-parametric regression techniques are suitable for uncovering spatial relationships. Also other techniques like MARS (Multivariate adaptive regression splines) were already used for making spatial predictions, like in (Leathwick, et al., 2005). In this paper we present a case study to predict wind speeds based on a DEM using regression Trees. Regression Trees are another accepted simple and stable method in the non-parametric regression domain. The basic algorithm can be enhanced by various ways; key words are bagging, pruning, and forests of regression trees (random forests). The basic regression tree algorithm is enhanced by an iterative procedure called Boosting, which aims to minimize the initial residuals. A cross validation is done to determine the quality of the estimates. The study site is Switzerland, a mountainous region in the heart of Europe. This test site has been recorded already in Etienne, et al. (2010) , and been tested with the same intention using the methodology results of (Lehmann, et al., 2002). The data for training and validation of the predictor are provided by the Swiss weather service (MeteoSwiss), who manages and maintains a network of weather stations since the early 1960s, recording a broad set of parameters like daily max. wind speed, mean wind direction, temperature etc. We use data ranging from 1980 to 2005 of 200 stations, recording the daily mean. wind speed in m/s. Figure 1 shows the DEM of Switzerland and a subset of the selected meteorological stations. A broad range of parameters like slope, aspect, terrain positioning index etc. is derived from the DEM and represent the independent parameters of the non-parametric regression problem. In conjunction with the wind speed measurements, which are supposed to be dependent from the DEM, a training data set is generated to train the BRT predictor. The trained predictor is then used to produce a wind speed map for Switzerland. |
Keywords: |
Wind Speeds, DEM, Spatial Predictions, Boosted Regression Trees |
Status : Paper Accepted (Oral Presentation) |