Based on forecast data obtained from the China Meteorological Administration, Taiwan Meteorological Center, Japan Meteorological Agency, and Joint Typhoon Warning Center of the USA, 24-hour, 48-hour, and 72-hour ensemble forecasts of typhoon tracks of 31 typhoons over the Northwestern Pacific in 2013 were conducted using the weighted bias-removed ensemble mean method. Error analysis indicates that the results of ensemble forecasts using the fixed training method with a constant weighting factor are quite sensitive to the selection of training samples, showing significant fluctuation of error for different training samples, while the error of ensemble forecasts using the moving training method with a variable weighting factor shows a declining trend with the increase of the number of training samples and tends to be stable when the number reaches 40. Comparisons show that twin typhoons in the training period have a large influence on the forecast accuracy, and the accuracy is significantly increased by removal of the influence of twin typhoons. Based on comprehensive comparisons of forecast results of 2013s typhoon tracks, the moving training method with 40 training samples and without consideration of twin typhoons in the training period is recommended.