Aiming at the shortcomings of large computation load and poor real-time performance for deep learning models with monitoring data, combining the advantages of board learning system (BLS) and bi-directional long short term memory (Bi-LSTM) model, a structural deformation prediction model was proposed based on bi-directional gate board learning system (Bi-G-BLS). The cyclic feedback structure and the forget-gate structure were added to the feature nodes of BLS to improve the dependence of the current node on the previous node, and the internal features of the time series were extracted from the forward and reverse respectively to make full use of the bidirectional data characteristics. As a result, the prediction accuracy was improved effectively while the prediction time was greatly reduced. The test results of the measured monitoring data for the subway foundation pit settlement showed that compared with the gated recurrent unit (GRU), BLS, Bi-LSTM, and G-BLS models, the root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) of the proposed model were reduced by 21.04%, 12.81% and 24.41%. The prediction time of the proposed model was reduced by 99.59% compared to Bi-LSTM model with similar prediction accuracy. The results showed that the prediction speed and accuracy of the proposed model were significantly improved over the comparison models.