.png){kind=icon}
.png){kind=icon}
Return
Interpretable WTConv1D-BiLSTM monthly-scale precipitation prediction model based on novel multilevel and multi-scale decomposition
.png){kind=icon}
PRE
AI
M
R
W
R
Y
DOI：10.1016/j.atmosres.2026.108948.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate monthly precipitation forecasting is essential for water-resources planning, flood-risk mitigation, and climate adaptation, as precipitation exhibits variability spanning long-term, seasonal, and sub-seasonal time scales. To disentangle these hierarchically coupled signals, this study proposes a scale-aware multilevel deep learning framework integrating Neural Seasonal-Trend Decomposition with Adaptive Multiband Filters and Crested Porcupine Optimizer tuned Variational Mode Decomposition. It is first employed to separate long-term climatic trends and dominant seasonal components to alleviate nonstationarity and reduce scale mixing. CPO-VMD is then applied to the residual high-frequency signals to extract band-limited intrinsic mode functions associated with sub-seasonal and mesoscale atmospheric processes, with adaptive parameter optimization ensuring frequency-localized signal separation consistent with precipitation dynamics. Each decomposed component is modeled independently using a WTConv1D-BiLSTM predictor to capture localized multiscale features and long-range temporal dependencies. Experiments on monthly precipitation data from 30 provinces in mainland China demonstrate Nash Sutcliffe Efficiency values above 0.95 in most regions and consistent improvements over conventional decomposition methods and baseline deep-learning models in MAE and RMSE. Gradient-weighted Class Activation Mapping further reveals scale-dependent feature contributions across climatic zones, providing transparent interpretability into multiscale temporal contributions. These results suggest that multilevel decomposition is physically plausible under scale consistency for precipitation modeling and enables robust, interpretable monthly precipitation prediction.
Keywords:
Monthly precipitation prediction
Multilevel multi-scale decomposition
WTConv1D-BiLSTM
Interpretable deep learning
Journal
IF:
4.4
Papers:
1.1K
Citations:
2.2W
Organization
Cited Papers