Simulation of daily temperatures for climate change scenarios over Portugal: a neural network model approach

Methods to assess the impact of global warming on the temperature regime of a single site are explored with reference to Coimbra in Portugal. The basis of the analysis is information taken from a climate change simulation performed with a state-of-the-art general circulation model (the Hadley Centre model). First, it is shown that the model is unable to reproduce accurately the statistics of daily maximum and minimum temperature at the site. Second, using a re-analysis data set, downscaling models are developed to predict site temperature from large-scale free atmosphere variables derived from the sea level pressure and 500 hPa geopotential height fields. In particular, the relative performances of linear models and non-linear artificial neural networks are compared using a set of rigorous validation techniques. It is shown that even a simple configuration of a 2-layer non-linear neural network significantly improves on the performance of a linear model. Finally, the non-linear neural network model is initialised with general circulation model output to construct scenarios of daily temperature at the present day and for a future decade (2090-99). These scenarios are analysed with special attention to the comparison of the frequencies of heat waves (days with maximum temperature greater than 35C) and cold spells (days with minimum temperature below 5C).