Stochastic generation of multi-station daily temperatures using a neural network

Stochastic algorithms for generating daily weather elements have been developed over the past two decades to provide synthetic time series in support of a variety of agricultural and hydrological issues. Although often multivariate with respect to the number of weather elements, the models are usually designed to provide time sequences of weather at a single location.

In this study, a neural network is developed to generate a stochastic sequence of daily average temperatures at over 150 stations distributed across the lower 48 states. Close replication of the auto- and cross-covariance structure of the observed temperature fluctuations is the desired objective. The neural network operates on the space-time principal components of the temperature variability. The skill of the methodology is assessed through cross validation.