Improving estuarine discharge forecasting with a KAN-augmented LSTM model: A case study of the Yangtze River Estuary

Xuliujing section, Yangtze River Estuary, China. Predicting discharge in tidal rivers is challenging due to complex flow dynamics influenced by topography, tides, runoff, and weather. Traditional forecasting methods struggle with fixed parameters, limiting their adaptability and accuracy over time. To address this, we propose an enhanced deep learning model, A KAN-augmented LSTM framework, which integrates a Kolmogorov-Arnold network (KAN) with a long short-term memory (LSTM) network. This model retains LSTM's ability to handle long-term dependencies while replacing the fully connected layer with a KAN layer. A learnable B-spline activation function in the KAN layer improves the model's capacity to capture nonlinear dynamics and long-term dependencies in time series data, enhancing forecasting accuracy. This paper applies the LSTM-KAN model to the Xuliujing section of the Yangtze River Estuary and compares its performance with traditional harmonic analysis (HA) and four neural network models: LSTM, XGBoost, DLinear, and Informer. The results demonstrate that the LSTM-KAN model significantly enhances discharge forecasting accuracy, outperforming all comparative methods across short-term (6 h), medium-term (12–24 h), and long-term (36–48 h) forecasts. Specifically, it achieved relative accuracy improvements of 12.1 %–35.2 % over HA and 7 %–52.8 % over the traditional LSTM model. These findings suggest that the complex interplay of tidal forcing, runoff, and weather in the Yangtze Estuary is better represented by the adaptive, function-learning paradigm of KAN than by models with fixed nonlinearities. The model's superior performance offers new insights for studying complex flow dynamics, indicating that deep learning techniques with learnable activation functions provide a more powerful and accurate tool for operational forecasting in highly dynamic tidal river environments. • This study developed an LSTM-KAN model for more accurate discharge forecasting in tidal rivers. • KAN leverages adaptive edge-weight activation to capture tidal-hydrodynamic nonlinearities with fewer parameters. • This study first embeds the architecture in hydrologic forecasting, using the Yangtze Xuliujing reach as the pilot. • The proposed model significantly outperforms traditional approaches in streamflow-forecasting accuracy.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.