The fined forecasting of visibility is difficult due to its local scale and complicated nonlinear variations. ANN (Artificial Neural Network) performs well in simulating complicated variation processes and thus is feasible for solving this problem. The study employs RNN (Recurrent Neural Network) to build a short-time forecasting and nowcasting model of visibility at a single station using surface observations of Fuzhou Meteorological Observation Station and its forecasting skill is evaluated. The results from random test samples show that the variation tendency forecast of visibility by RNN is basically in conformity with the observed data in the 1-h, 3-h, and 6-h forecast; compared with the forecast based on the actual situation, the RMSE (root mean square error) decreases by 15.75%, 31.66%, and 41.26%, respectively; compared with the forecast based on the traditional BPNN (Back Propagation Neural Network), the MAE (mean absolute error) decreases by 12.90%, 24.45%, and 38.99%, respectively. The results indicate that RNN has advantages in the forecasting of visibility, providing a new method for the refined short-time forecasting and nowcasting of visibility.