To achieve good time-frequency concentration in short time Fourier transform (STFT), window length should be selected according to the characteristic of the processed signal. The selection mechanism of existing rules is analyzed and their merit and defect are discussed. It is found that normalization third order Renyi entropy rule and Stankovic rule often reach the maximum of the window length range while the window lengths determined by other rules are also not optimal. The reason for this phenomenon is analyzed. A new window length selection rule based on logarithmic window energy is put forward. The relationship between logarithmic window energy and window length is non-linear, which is quite different from the linear increase characteristic of common window energy with respect to window length. The logarithmic window energy increases with different speeds during the short and long window and the increasing speeds are independent of the type of window, thus good compromise between the short and long window length can be achieved. The processing examples of simulation and practical signals are provided, which demonstrate that STFT can obtain good time frequency concentration using the window length determined by the proposed rule,and the effect is better than that of the existing rules.