In order to solve the problem that random selection of clustering number k can easily lead to instability and low computational efficiency of k-means clustering algorithm in the zoning calculation of wind load on the surface of long-span roof structures, an improved Canopy-k-means clustering algorithm was proposed. Firstly, the Canopy algorithm was introduced, and the selection of its initial threshold and clustering center was improved to reduce the blindness of the initial value selection so as to improve the reliability of the results of wind load zoning. Secondly, the improved Canopy algorithm was used to preprocess the wind load dataset to determine the cluster number k quickly and accurately. Thirdly, the improved Canopy algorithm was combined with k-means to achieve the accurate identification of the optimal classification number k, so that the improved Canopy-k-means clustering algorithm can get the zoning results quickly and accurately when the wind load on the surface of long-span roof structure was divided. Finally, taking a long-span cylindrical roof dry coal shed structure as an example, based on the test results of the surface wind load data obtained from the wind tunnel test, the improved Canopy-k-means clustering algorithm was used to calculate the surface wind load. The results show that by using the improved Canopy-k-means clustering algorithm, the wind loads on the surface of long-span roofs at 0 °, 50 °and 90 °wind angles are divided into three different zones, and the corresponding SD values are 2.36,3.51 and 2.52,respectively, which are significantly lower than those obtained by the traditional k-means clustering algorithm, and the intra-class compactness and inter-class dispersion are obviously improved. Therefore, the improved Canopy-k-means clustering algorithm can obtain the optimal zoning results quickly and accurately, and has good engineering application value for wind load zoning on the surface of long-span roofs.