To address the issue that directly using historical datasets as input in photovoltaic (PV) power prediction models struggles to meet the required prediction accuracy, a short-term PV power prediction model, namely CSSA-LSTM model was proposed based on time-segmented convolutional dynamic time warping (CDTW) for similar day selection and the chaotic sparrow search algorithm (CSSA) for long short-term memory (LSTM) optimization. After preprocessing the data, the model identified key meteorological factors affecting power generation through Pearson correlation analysis, thereby avoiding interference from irrelevant data. It then divided time periods based on the total solar radiation sequence of the day to be predicted, applied the CDTW algorithm to each time period for selecting similar time segments, and reconstructed these segments into a set of similar days. Meanwhile, CSSA was used to optimize the hyperparameters of the LSTM model, enabling adaptive search for the optimal hyperparameters. Simulation analysis results using measured data from 2021 of a PV power station in southern China show that compared with traditional similar day methods and LSTM models optimized by the sparrow search algorithm (SSA), the similar day selection method and CSSA-LSTM model proposed in this study exhibit higher prediction accuracy and robustness.