Objective: Central aortic blood pressure (CAP) is a critical factor associated with cardiovascular events. Current methods for converting peripheral blood pressure to CAP using specific algorithms suffer from significant inter-individual variability in estimation accuracy and poor stability. This study aims to investigate the feasibility of using a deep learning (DL) model to predict CAP by fusing non-invasively measurable peripheral hemodynamic information. Methods: To address the challenge of acquiring large-scale and high-quality clinical data required for DL model development, we used a computational model of the human cardiovascular system to generate virtual human samples, including 2000 samples for model development and 100 in-distribution samples for validation. Additionally, 100 samples with distinct cardiovascular properties were generated for generalization testing, defined as out-of-distribution dataset. The DL model employed a CNN framework with a multi-head attention mechanism, taking blood flow velocity waveforms from four peripheral arteries (with artificially introduced random measurement errors) and brachial systolic/diastolic pressures as inputs, and CAP as the output. Results: The DL model showed good convergence in training (loss function<5×10-3). Testing on the in-distribution validation dataset showed a CAP waveform prediction error of less than 1.5mmHg (R2>0.98), with the mean prediction errors for systolic and diastolic pressure below 1.3mmHg. However, testing on the out-of-distribution dataset revealed a significant increase in CAP prediction error, with the mean systolic pressure error exceeding 5.7mmHg. Conclusion: The DL model can rapidly predict CAP based on peripheral hemodynamic information, but its prediction accuracy is constrained by the coverage capability of the training samples regarding individual differences. Future research should focus on collecting large-scale clinical data covering a wide range of individual variations to further enhance the clinical value of DL models.