Dynamic classification and attention mechanism-based bidirectional long short-term memory network for daily runoff prediction in Aksu River basin, Northwest China.

Abstract

Inland river runoff variability is pivotal for maintaining regional ecological stability. Daily flow forecasting in arid regions is crucial in understanding water body ecological processes and promoting healthy river ecology. Precise daily runoff forecasting serves as a cornerstone for ecological evaluation, management, and decision-making. With the advancement of artificial intelligence technology, data-driven models have exhibited promising capabilities in runoff prediction. Nevertheless, the arbitrary selection of boundaries between different flow patterns without considering temporal changes across seasons limits the accuracy of runoff simulation. This paper proposed an integrated modeling approach encompassing a dynamic classification method, an attention mechanism, and a bidirectional long short-term memory network (CA-BiLSTM) to enhance flow prediction performance while accommodating diverse flow patterns. The classification boundary was determined by the dynamic change interval value of relevant hydrological variables, facilitating a more comprehensive exploration of the relationships and information within hydrological data. The performance of the CA-BiLSTM model was compared against a traditional machine learning model lacking data classification, utilizing data from the West Bridge station of the Aksu River Basin (ARB). The results indicate that the CA-BiLSTM model outperforms traditional LSTM and BiLSTM models across all seasons. The CA-BiLSTM model demonstrates superior performance in arid zones. Compared to the single LSTM model, CA-BiLSTM exhibits reductions of 42.99%, 36.89%, and 49.73% in MAE, RMSE, and MAPE, respectively, while enhancing R² and KGE by 10.47% and 11.76%. The proposed hybrid model effectively reduces runoff prediction uncertainty, offering valuable insights for water resource management in arid zones.