A bi-directional long short-term memory (Bi-LSTM)-based technique was proposed for subgrade settlement prediction in order to achieve early and accurate prediction of post-construction settlement of subgrade. The observational data was transformed into equally spaced observations using interpolation, namely the Akima method, for time series analysis. Six contributing factors were extracted from the ‘fill height-time-foundation settlement’ curves, and were used as input variables for training a Bi-LSTM model. Settlement predictions can be sequentially updated by combining with the rolling forecast technique. Results show that deep learning can efficiently use subgrade construction information, increase the sample size, and improve the reliability of early prediction. The Bi-LSTM model achieves bidirectional feature extraction on observational data and makes more accurate predictions in the same sample size conditions. The information during construction and 3-month post-construction period was used based on observations of six subsoil foundations of medium compressibility and one composite foundation. Predicted results with mean values of root mean square error (RMSE) and mean absolute percentage error (MAPE) were 1.19 mm and 1.04%, respectively.