Accurate photovoltaic power prediction results are of great significance for the dispatching optimization of the power grid. Therefore, the existing prediction methods only focus on the time domain or frequency domain characteristics, resulting in the failure to fully extract the time and frequency domain information in the photovoltaic data. A photovoltaic power prediction model TDFDAFF integrated with time-frequency features was proposed, aiming to capture the complex characteristics of photovoltaic sequence data more comprehensively. First, the feature correlation analysis was conducted on the original photovoltaic sequence data was conducted, and the meteorological factors strongly related to the photovoltaic power were selected as the input features, and the filtered feature data were divided by sliding window technology to build the input feature matrix. Subsequently, it was input to the timing feature extraction module and the frequency feature extraction module respectively to extract the time-frequency features in the photovoltaic sequence data. Finally, the attention feature fusion module was used to fuse the time-frequency features extracted from the two parts, and assign different weights to the dynamics of each feature, so as to emphasize the attention of important features, and map the fused features through the full connection layer to obtain the final prediction output. Through validation on photovoltaic datasets in two regions, the TDFDAFF model has better prediction accuracy compared with other prediction models.