摘要
Abstract
To address the limitations of traditional tidal level prediction models in deep extraction of hydrological features and modeling of long-term temporal dependencies, a forecasting method integrating deep learning techniques was developed to improve the accuracy of tidal level prediction under complex hydrological conditions in the lower Yangtze River. A coupled CNN-BiLSTM-transformer model is proposed based on the collaborative mechanism of Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and transformer. CNN is used to extract multi-scale spatial features from water level sequences, BiLSTM captures bidirectional temporal correlations, and the transformer attention mechanism is introduced to model long-range dependencies. An empirical study based on hourly observational data from 2012 to 2024 at the Tianshenggang Tidal Station demonstrates that the model's mean squared error (MSE) decreases by 88.38％, 89.85％, 91.04％, and 93.90％, respectively, for forecast horizons of 24 h, 48 h, 60 h, and 80 h, compared with the BiLSTM benchmark model. As the forecast horizon extends, the model's prediction accuracy continues to improve, validating its effectiveness in long-term forecasting. These improvements are mainly attributed to the collaborative modeling of spatial and temporal dependencies by the spatiotemporal attention mechanism and the effective capture of long-range dependencies by the transformer. This model effectively addresses the challenge of spatiotemporal feature coupling in hydrological forecasting and provides high-accuracy decisionmaking support for flood control regulation in the Yangtze River Estuary.关键词
潮位预测/卷积神经网络(CNN)/循环神经网络/双向长短期记忆神经网络(BiLSTM)/自注意力模型(transformer)/长江下游/潮位Key words
tidal level prediction/convolutional neural network(CNN)/recurrent neural network/bidirectional long short-term memory neural network (BiLSTM)/self-attention model (transformer)/lower Yangtze River/tidal level分类
建筑与水利
