Road grade is important for fuel cell hybrid vehicle (FCHV) in energy management strategies. However, the accurate road grade is difficult to obtain in real time. This study proposed a road grade estimation method based on an intelligent algorithm (long short-term memory recurrent neural network, LSTM). Based on the vehicle dynamic model, appropriate input variables were selected for the network input. For comparison, multilayer perceptron (MLP) algorithm was applied and the normalized root mean square error (NRMSE) of the estimation results in different literatures were listed. The results show that this method can estimate the road slope accurately without installing additional sensors with the RMSE (root mean square error) value and NRMSE value of the estimation error of 0.65 degree and 4.6%, respectively.