The short term forecast period is critical to dispatch planning for utilities. During this 0-2 hour forecast period, tactical decisions must be made as to the allocation of the energy portfolio. The variable nature of wind power complicates this process. Improvements in the accuracy of short term forecasts reduce the uncertainty of these operational decisions. A short term forecast system is distinct from a ramp forecast system. While ramps have an major impact on a utility's operations, they are relatively rare events. A ramp forecast system would warn the user of the timing and magnitude of such phenomena. On the other hand, actions which depend on an accurate short term forecast occur every hour of every day for utilities. NCAR has developed a wind energy forecast system for Xcel Energy which forecasts out to 7 days. Within the first 3 hours, the forecast interval is 15 minutes due to the operational need for a high temporal resolution forecast in the short term. Two modules contribute to the power forecast during this period. First is an automated adaptive consensus forecast system. Hub height wind speed forecasts from this system are passed to an advanced wind-to-power conversion module. This paper will describe the evolution of the NCAR/Xcel system that has led to significantly reduced errors in the power forecast within the first two hours. These reductions in error have been partially due to the learning process of the consensus forecast system. However, several algorithmic changes within the wind-to-power module have led to distinct incremental improvements. The power forecast error for the Xcel system now beats persistence at well within a one hour lead time.