This paper describes a novel computational approach of monitoring network optimization and its application to a problem of wind power plants sites evaluation. In this complex spatial decision making process it is essential to build a model for the estimation of wind speeds and their constancy over time. Geostatistical models are often used to assess these properties from meteorological data to assemble wind power capacity atlases. However they often rely on incomplete information and insufficient number of measuring stations, especially in mountainous regions of complex topographies. They also require various corrections to account for the influence of topography on wind speeds.
In this study we apply a methodology based on support vector regression (SVR) to assess the areas of interest for wind power plants construction in Switzerland from the interpolation of monthly average wind speed data. The model incorporates a rich set of topographical indices as explanatory input variables. With this procedure, we produce a spatial map of average wind speeds on the Swiss territory, which helps assessing the suitability for the construction of new plants.
To enhance the accuracy of this map by taking of additional measurements, one has to consider the monitoring network optimization as an exploration of the high-dimensional space of combined geographical and topographical variables. We define an active learning criterion to achieve this goal and target the exploration at uncertain areas close to the decision threshold for power plants construction. Network design can be willingly biased towards the areas at risk of the phenomenon to be modeled, in our case average the wind speeds of about 4.5 [m/s] at a height of 50 [m] over the ground, which is the minimum monthly average wind speed required for the expediency of a power plant facility construction.
Using this criterion, we study the topographical features of the Swiss territory in terms of interest for a new monitoring station. We pay particular attention to sites capable to improve the accuracy of wind speed models in the areas of maximal uncertainty around the decision threshold of 4.5 [m/s]. This study enabled to extract topographical conditions related to model uncertainty and to extract possible locations of interest for new monitoring stations.