The energy storage system and power configuration of a port charging station are based on the power load of the station. The ultra-short-term load prediction of the port charging station is essential for its efficient scheduling and stable operation. To enhance the accuracy of ultra-short-term load prediction for port charging stations, a dual-model approach is proposed. This method involves modifying the results of a conventional daily load prediction model by incorporating similar-day feature data modeling. Firstly, the load data is decomposed into features using EEMD, and the fused features from the decomposition and DBN processing are inputted into the LSTM prediction model to obtain the daily load prediction results. Then, the PSO algorithm optimizes the k-density clustering parameters to cluster similar-day data. Finally, the cluster results of daily loads and the results from the similar-day model are combined to produce the final prediction. Through validation using real port data, the proposed method demonstrates an effective improvement in the accuracy of port charging station load prediction, with significant practical engineering significance.