TCN-attention: a temporal convolutional network with self-attention mechanism for streamflow prediction in Chinese River Basins
TL;DR
TCN-Attention is proposed, integrating Temporal Convolutional Networks with multi-head self-attention for weekly streamflow prediction, employing dilated causal convolutions to capture multi-scale temporal patterns, while the self-attention module dynamically identifies informative time steps.
Abstract
Accurate streamflow prediction is essential for water resource management and hydropower scheduling. Traditional RNNs (LSTM, GRU) suffer from sequential processing limitations in capturing long-range temporal dependencies. We propose TCN-Attention, integrating Temporal Convolutional Networks with multi-head self-attention for weekly streamflow prediction. The TCN employs dilated causal convolutions to capture multi-scale temporal patterns, while the self-attention module dynamically identifies informative time steps. Evaluated on the CCAM HydroMLYR dataset (102 Yangtze River catchments, 11 meteorological features, 12 static attributes), TCN-Attention achieves NSE of 0.2016 and RMSE of 0.9881, outperforming LSTM (NSE=0.1869), Transformer (NSE=0.1738), and GRU (NSE=-0.1044) baselines.