Considering the significant impact of weather factors such as solar radiation and temperature on photovoltaic power generation, this paper proposes a method for ultra-short-term photovoltaic power prediction based on an Improved Dung Beetle Optimization algorithm (IDBO) optimizing the structure of Bi-directional Long Short-Term Memory (BiLSTM) networks. Firstly, the historical dataset is partitioned into sample sets related to three types of similar day conditions, respectively sunny days, cloudy days, and rainy days through fuzzy C-means clustering (FCM) based on various meteorological feature variables. Then, an improved dung beetle algorithm is employed to optimize the hyperparameters of the BiLSTM model, and the training sets of the aforementioned three categories of similar day are selected to establish models to train the network and validation. Finally, the optimal network models trained under different similar day conditions are used for photovoltaic power prediction. The results show that regardless of weather conditions, the proposed model can effectively predict photovoltaic power and exhibits strong adaptability. Compared with other models, the model's high prediction accuracy is verified.