|
Abstract:
|
We consider the problem of predicting solar irradiance (power per unit area) as a function of location and time using weather station data in addition to data from two weather models. Challenges include very large data size, non-stationarity of the unknown response surface, and non-space-filling weather station locations. Modeling approaches such as local Gaussian process and multi-resolution functional ANOVA, which have potential to work well for this problem, are briefly discussed. We explore a neural network approach to modeling in more depth. Issues include data-driven choice of the number of basis functions, computational efficiency, estimation of the shape, location, and coefficients of basis functions, and inference.
|
