In this study, the performance of wind forecasts over Svalbard, located between the Arctic Ocean and the Norwegian Sea, was evaluated using the Polar Weather Research and Forecasting (PWRF) model and three-dimensional variational data assimilation (DA) system. The forecasts of the analysis-forecast cycling experiment using the PWRF 3DVAR were compared with those of the cold start experiment using reanalysis as the initial condition. Three strong wind cases that occurred during January and February 2011-2012 were selected, where polar lows were generated on the east coast of Greenland and generated a wind speed above 20 m s −1 in Svalbard. The wind speed forecasts for both cycling and cold start experiments were similar to the highest 10minute average wind speed for the last 1 hour (HAW). The average root mean square error (RMSE) of the forecasts in the cold start experiment from HAW was 3.78 m s −1 for three cases and was greater than that in the cycling experiment. The forecast performance in the cycling experiment was comparable to, or even better than, that in the cold start experiment, which implies that the cycling system with DA is more useful than the cold start system in forecasting polar weather to support research activities.