Traditional prediction model only focuses on time factors. Aiming at that deficiency, a prediction model which introduces weather factor and considers date attribute was proposed. Firstly, the degree of correlation between weather factors and passenger flow was determined by significance test. Then, to gradually screen the low-relevant weather factors, grey relation analysis (GRA) was adopted to calculate the non-linear correlation between various weather factors and passenger flow. After each screening, the bi-directional LSTM neural network was used to forecast and the GRA-BiLSTM prediction model was proposed. The results show that taking weather factors with GRA value less than 0.6 as input will reduce the prediction accuracy. Compared with the traditional LSTM, the prediction error of GRA-BiLSTM，which is obtained by gradually eliminating the weather factors with low correlation, is significantly reduced on both work days， and non_work days and the convergence speed and robustness are better than the traditional machine learning as well.