Accurate wind energy forecasts can be an essential component for economic viability of a wind project. Timely and accurate short-term (hours) forecasts can increase the electric grid efficiency, and minimize ancillary or other firming requirements, ultimately resulting in reduced costs. In this study the integration of observational data, at distances up to 200 km from a wind farm, and fine scale numerical weather predictions are shown to significantly increase short-term forecast accuracy. Methods for choosing the best predictor variables will be discussed. These methods are applied to linear regression, neural networks, and conditional neural networks. An example from the Pacific Northwestern United States is described. It is shown that significant forecast improvements are possible when using a combination of off-site observations, mesoscale numerical weather prediction forecasts, and conditional neural networks.