High-resolution spatial interpolation of weather generator parameters using local weighted regressions

The time-domain stochastic models for daily weather data known as “weather generators” are very useful for producing time series of arbitrary length that statistically resemble real weather data. One limitation to their use is that synthetic weather series may be needed at locations for which no real data exist, on which to base parameter estimates. This paper describes an approach to defining weather generators at such locations through interpolation of their parameters using weighted local regressions. This flexible method can capture nonlinear parameter variations in space and allows objective and automatic selection of both the regression predictor and the size of each local neighborhood of influence. It is found to perform well for a moderately large region of the northeastern U.S.