Abstract: To address the challenges posed by strong noise interference in photovoltaic (PV) sequences and the issues of low accuracy and poor generalization of a single model in PV power prediction, we propose a short-term PV power prediction method. This method uses feature optimization and a hybrid improved Grey Wolf algorithm to optimize the BiLSTM network. Firstly, the mutual information algorithm is employed for variable selection in input data to eliminate redundant variables. Subsequently, the Complementary Ensemble Empirical Mode Decomposition and an improved wavelet threshold algorithm are applied to the selected data for feature reconstruction to reduce noise interference in the data and optimize input variable features. Finally, the standard Grey Wolf Optimizer algorithm underwent hybrid optimization by integrating improved Tent chaotic mapping, nonlinear decreasing factors, dynamic weight strategies, and the differential evolution algorithm. This process aimed to ascertain the optimal hyperparameter combination for the BiLSTM network. The attention mechanism was also employed to extract key temporal information from the data, developing a novel short-term photovoltaic power prediction model. Simulation experiments demonstrate that, compared to the least squares support vector machine, long short-term memory network, and bidirectional long short-term memory neural network, the proposed model achieves an average reduction of 12.45%, 7.95%, and 5.37% in root mean square error under various weather conditions such as sunny, cloudy, overcast, and rainy, showcasing excellent predictive performance, good generalization ability, and promising engineering application value.