HydroFusionNet: CNN-LSTM-self-attention network for water level prediction in Xijiang River with Ascend CANN adaptation

Abstract

[Objective] Accurate water level prediction is essential for flood control, navigation safety, and water resource management. Traditional hydrological models often fail to capture the complex spatiotemporal dependencies and nonlinear characteristics in water level fluctuations. [Methods] To address these limitations, the Hydro Fusion Net deep learning framework was proposed, integrating CNN-LSTM module and self-attention mechanism in a parallel architecture to improve prediction accuracy. The CNN-LSTM module extracted local spatiotemporal dependencies, while the self-attention mechanism captured global contextual relationships. The outputs of both modules were fused for final prediction. This framework was optimized for deployment on the Ascend CANN platform, leveraging high-performance computing to achieve efficient training and prediction.[Results] Experiments on the Xijiang River dataset showed that Hydro Fusion Net outperformed conventional models, achieving an RMSE of 0. 394 and an R² of 0. 894. Furthermore, with optimization on Ascend CANN, the training speed increased by 2. 3 times, and energy consumption reduced by 17%, indicating significant computational advantages. [Conclusion] The proposed framework effectively captures complex hydrological dynamics, providing real-time water level prediction support for intelligent hydrological management and disaster prevention.

Keywords

water level prediction / deep learning / CNN-LSTM / self-attention mechanism / Ascend CANN / hydrological / flood control and disaster reduction / predictioin a ccuracy

Funding

National Natural Science Foundation(62466051)

Guangxi Young Teachers’ Basic Ability Improvement Program(2024KY0692)

Autonomous Region-level College Students’ Innovation and Entrepreneurship Training Program(S202411354227)