Density peak clustering has the advantages of fast convergence, strong robustness, and there is no need to manually set the optimal clustering number, etc. It can be used for pattern recognition and classification of user electricity behavior in industrial load forecasting, and then, makes prediction, which has high practical application value. However, this algorithm has a poor clustering effect under the condition of small samples, which is easy to "miss" the clustering center in the sample. In view of this situation, inter-class optimization and intra-class optimization are carried out for the density peak clustering to enhance the data separability and conduct in-depth mining and classification of electricity consumption behaviors of users. Then, gray relational degree is used to determine the category to be predicted, and GRNN neural network is used for load prediction. Through Matlab simulation, it can be concluded that the method proposed in this paper can effectively improve the clustering effect of power load data of industrial users and improve the accuracy of load prediction.