Wang et al. (2026) Interpretable WTConv1D-BiLSTM monthly-scale precipitation prediction model based on novel multilevel and multi-scale decomposition

Identification

• Journal: Atmospheric Research
• Year: 2026
• Date: 2026-03-21
• Authors: Menghao Wang, Rui Yan, Hao Wang, Ru Zhang, Yiyang Li
• DOI: 10.1016/j.atmosres.2026.108948

Research Groups

• College of Artificial Intelligence, North China University of Science and Technology, Tangshan, China
• College of Sciences, North China University of Science and Technology, Tangshan, China
• College of Water Science, Beijing Normal University, Beijing, China

Short Summary

This study proposes an interpretable deep-learning framework, WTConv1D-BiLSTM, for accurate monthly precipitation prediction by integrating novel multilevel and multi-scale decomposition techniques to address nonstationarity and scale mixing. The model demonstrates superior performance and interpretability in predicting monthly precipitation across 30 provinces in mainland China.

Objective

• To develop a scale-aware, interpretable multilevel deep-learning framework for accurate monthly precipitation forecasting by effectively disentangling hierarchically coupled signals (long-term trends, seasonal components, and high-frequency variability) to alleviate nonstationarity and reduce scale mixing.

Study Configuration

• Spatial Scale: 30 provinces in mainland China
• Temporal Scale: Monthly precipitation data

Methodology and Data

• Models used:
  • Neural Seasonal–Trend Decomposition with Adaptive Multiband Filters (NSTDAF)
  • Crested Porcupine Optimizer (CPO) tuned Variational Mode Decomposition (VMD)
  • WTConv1D–BiLSTM (Wavelet Transform Convolutional 1D - Bidirectional Long Short-Term Memory)
  • Gradient-weighted Class Activation Mapping (Grad-CAM) for interpretability
• Data sources: Monthly precipitation data from 30 provinces in mainland China.

Main Results

• The proposed WTConv1D-BiLSTM model, integrated with multilevel and multi-scale decomposition, achieved Nash Sutcliffe Efficiency (NSE) values above 0.95 in most regions for monthly precipitation prediction.
• The model consistently demonstrated improvements in Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) compared to conventional decomposition methods and baseline deep-learning models.
• Gradient-weighted Class Activation Mapping (Grad-CAM) revealed scale-dependent feature contributions across different climatic zones, providing transparent interpretability into the multiscale temporal contributions.
• The multilevel decomposition approach was found to be physically plausible under scale consistency for precipitation modeling, enabling robust and interpretable predictions.

Contributions

• Introduction of a novel scale-aware multilevel deep-learning framework that integrates NSTDAF and CPO-VMD for effective decomposition of precipitation time series into long-term trends, seasonal components, and high-frequency signals.
• Development of a WTConv1D-BiLSTM predictor for independent modeling of each decomposed component, capturing both localized multiscale features and long-range temporal dependencies.
• Significant improvement in monthly precipitation prediction accuracy (NSE > 0.95, reduced MAE and RMSE) compared to existing methods, addressing the challenges of nonstationarity and scale mixing.
• Provision of transparent interpretability through Grad-CAM, revealing scale-dependent feature contributions and enhancing the physical plausibility of the model.

Funding

• [No funding information was explicitly mentioned in the provided text.]

Citation

@article{Wang2026Interpretable,
  author = {Wang, Menghao and Yan, Rui and Wang, Hao and Zhang, Ru and Li, Yiyang},
  title = {Interpretable WTConv1D-BiLSTM monthly-scale precipitation prediction model based on novel multilevel and multi-scale decomposition},
  journal = {Atmospheric Research},
  year = {2026},
  doi = {10.1016/j.atmosres.2026.108948},
  url = {https://doi.org/10.1016/j.atmosres.2026.108948}
}