Comparative prediction of runoff in the Beiluo River, Shaanxi Province of northwestern China based on TCN-BiLSTM and LSTM models

Objective The aim of this study was to investigate the performance of coupled TCN-BiLSTM model and traditional LSTM model in runoff simulation and prediction, specifically the effect of different input variables on the accuracy of machine learning hydrological model and performance of the model under different foresight periods.

Method A new coupled model TCN-BiLSTM for runoff prediction was established based on bi-directional long short-term memory network (BiLSTM) and temporal convolutional network (TCN) with the Beiluo River Basin as the study area. Using correlation analysis, the input factors for predicting runoff were screened, and four different input schemes were identified to be applied to the coupled TCN-BiLSTM model and the conventional LSTM model, each of which predicted runoff volumes for 1, 2, and 3 d, respectively. Mean absolute error (MAE), root mean square error (RMSE) and goodness of fit (R²) were used to assess the predictive performance of the model.

Result (1) The overall prediction performance of the TCN-BiLSTM coupled model was better than that of the LSTM model, and the R² of TCN-BiLSTM can reach 0.91, which was higher than that of the LSTM, 0.89. Compared with LSTM, TCN-BiLSTM was more capable of capturing peaks and mutation points, and was better at predicting complex data with large fluctuations. (2) In the runoff prediction for the next 1−3 d, the predictive effectiveness of TCN-BiLSTM and LSTM models under the four scenarios decreased with the extension of the foresight period, and compared with the prediction of 1 d, the TCN-BiLSTM and LSTM R² for the prediction of 3 d decreased on average by 0.17 and 0.14, respectively, and the RMSE increased on average by 4.59 and 4.40, respectively, and the MAE increased on average by 1.26 and 1.31, respectively. (3) Among the four input scenarios, the best model predictions were obtained when daily precipitation data and daily runoff data were used as input variables. The inclusion of precipitation data improved the R² of the TCN-BiLSTM and LSTM models for 1, 2, and 3 d runoff predictions by 15%, 14%, 6%, and 18%, 14%, and 1%, respectively, compared with the single daily runoff data as an input variable.

Conclusion Both the TCN-BiLSTM coupled model and the LSTM model R² can reach more than 0.85, and the TCN-BiLSTM R² is improved by 2% compared with the LSTM. In comparison, the TCN-BiLSTM model performs better in fitting the flood process, and the prediction performance for flood season is better than that for non-flood season. The input variables have a greater impact on the model, and effective and high-quality meteorological data can improve the prediction performance of the model.