Adaptive period decomposition network with muti-scale temporal variation learning for water level prediction

Hanjiang River Basin, China This study introduces DLGNet, an advanced water level prediction model that employs an adaptive multi-resolution decomposition to decouple multiple periodic components. The model incorporates a 2D overlap patch attention mechanism and separable dilation causal convolution to effectively capture local and global temporal variations. Additionally, DLGNet utilizes multiscale 2D separable convolutions to capture temporal variations among period components and employs a novel aggregation strategy for feature interaction analysis, ultimately enhancing prediction accuracy through direct decoding. Compared to ten other advanced prediction models, DLGNet achieved an average reduction in mean square error (MSE) by 94. 82 % , 93. 35 % , and 83. 13 % for short-term, mid-term, and long-term predictions, respectively. The adaptive period decomposition capability and multiscale temporal variation representation learning of DLGNet ensure its robustness and accuracy across various prediction tasks. Moreover, in all prediction tasks, the output of DLGNet more closely approximates the ground truth, exhibiting rich local detail changes and aligning more accurately with the actual fluctuations of water levels over time. • We propose a novel model, DLGNet, for multi-scale water level prediction. • DLGNet learns temporal variation both in one- and two-dimensional spaces. • DLGNet achieved an average MSE reduction of 90.43 % across 11 prediction tasks compared to 10 other models.