Abstract: Flow field reconstruction based on the convolutional neural network (CNN) is a new method in the field of wind resource assessment. It has the advantages of low cost, short evaluation and high precision for the wind farm construction and wind resource utilization, which provides values in both academic research and engineering practice. By selecting Bolund Island as a case study, the flow field reconstruction method based on a multi-layer fusion convolution neural network has been developed, which adopts sample labels of the input wind profile and the output flow field as used in the CFD. Aiming at shortening the computational time, this study examines the influence of the sample number on the accuracy of the flow field reconstruction method, and analyzes its extrapolation ability in the meantime. The results suggest that the flow field reconstruction method based on CNN can restore the flow field information well. The maximum reconstruction error is less than 2% when the reconstruction target is within the training sample range, and it keeps within 5% when the number of samples is reduced from 100 to 25. The reconstruction method has some extrapolation ability, but is limited within some range. When the reconstruction target is beyond the training sample range, the maximum reconstruction error rises from less than 5% to more than 20%.