A hybrid convolutional neural network for multi-station water level prediction: Enhancing navigation safety through spatial-temporal modelling

Abstract

Accurate prediction of multi-station water levels is crucial for mitigating flood risks and optimizing navigation management in complex riverine environments. Existing approaches often fail to capture the dynamic spatiotemporal interdependencies between monitoring stations, limiting their predictive performance and utility in operational decision-making. To address this challenge, we propose MSWLCN (Multi-Station Water Level Convolutional Network) model, a novel deep spatio-temporal convolution framework tailored for simultaneous and accurate prediction of daily water levels over consecutive days at multiple stations. This architecture integrates multi-layer Convolutional Long Short-Term Memory (ConvLSTM) and three-dimensional convolutional (Conv3D) networks with strong adaptations. The model explicitly extracts temporal dependencies and spatial correlations across stations through the spatio-temporal modelling architecture, enabling simultaneous prediction of multi-station water levels with complex changing characteristics. And we validate the framework using a comprehensive dataset spanning 1826 consecutive days from 19 hydrological stations along the Yangtze River, a globally significant navigational corridor. Experimental results demonstrate that the proposed MSWLCN outperforms conventional modelling methods in terms of prediction accuracy and computational efficiency. This research advances environmental modelling practices by offering a scalable solution for multi-station water level forecasting, with direct applications in water resource management and navigation safety assurance.