To simplify the complexity of the distribution network model when typhoon hits and enhance the effectiveness of evaluating the reliability of distribution networks, the method of training long short-term memory (LSTM) with the results of mechanistic model and assessing reliability by LSTM is proposed. Firstly, the wind speed-based distribution grid element pole failure rate calculation model is proposed, through which the failure rates of distribution grid elements under different wind speeds can be calculated, based on which, a training dataset is obtained by constructing a model-driven short-term reliability assessment method that relies on sequential Monte Carlo simulation. Then, the LSTM model is constructed, and the LSTM model for mapping between system state and reliability index is obtained through offline training to achieve rapid system reliability assessment. Finally, the arithmetic example employs the IEEE-RBTS-BUS6-F4 feeder and the calculation is performed on Matlab, and the method's correctness and validity are validated by the calculation results.