Application of artificial neural networks for the spatial estimation of wind speed in a coastal region with complex topography

Abstract

The impact of topography on the diurnal patterns of surface flow is of great importance, significantly modifying the wind speed spatial distribution and the vertical structure of the lower atmosphere. In this work, two feed forward neural network architectures are examined for their ability to estimate the hourly wind speed in a coastal environment which is characterized by complex topography. Additionally, the spatial average, the nearest and natural neighbor along with the inverse distance and square distance weighted average interpolation methods are employed and the results are compared for the area of study. These schemes utilize wind speed measurements from six meteorological sites and they are evaluated for their predictive ability in each location using the “leave-one-out cross-validation” technique. The predictive accuracy of the neural network which incorporates wind direction in the form of wind vector as an input is found to be statistically superior compared with the five traditional interpolation schemes and with the network which utilizes only the wind speed intensity. An insight on the underlying input–output function approximation of the neural networks is obtained by examining their ability to incorporate the mean wind variability characteristics of the study area.