The realization of many technologies in the industrial field is based on accurate load data, while the existing measurement system in factories often results in a large number of bad data due to communication and storage failures. Therefore, an industrial load data identification and correction method with improved k-prototypes clustering algorithm was proposed to reduce the impact of bad load data on the clustering results by introducing characteristics of non-load data, so as to make the identification and repair results more accurate. Through random initialization and parallel calculation, the improved k-prototypes algorithm overcomes the defect that standard algorithm tends to fall into the local optimal solution. And the problem of subjectively determining the number of clusters is solved by adaptive processing. Based on the clustering results, the feasible region of load data is obtained according to the principle of normal distribution, and the bad data is identified. The identified bad data is corrected by centroid vector replacing. Experiments show that the method outperforms the fuzzy C-means clustering method which only considers the load data.