A hierarchical decomposition and component-specific framework for forecasting non-stationary mine water inflow: Pingdingshan mining region, China

Study Region: A coal mining region in Pingdingshan City, Henan Province, China. Study Focus: To address the challenges posed by the non-stationary and complex nature of mine water inflow, this study proposes a novel hybrid deep learning framework, STL-VMD-TCN-PatchTST, for high-precision, short-term forecasting. The framework's methodology integrates a hierarchical signal decomposition strategy with component-specific predictors. First, Seasonal-Trend decomposition using Loess (STL) isolates the dominant long-term trend. The volatile remainder is then further deconstructed using Variational Mode Decomposition (VMD). A lightweight Temporal Convolutional Network (TCN) is used to efficiently forecast the smooth trend, while a Whale Optimization Algorithm (WOA) optimized and strong-performing Patch Time Series Transformer (PatchTST) is deployed to model the complex, high-frequency components. New Hydrological Insights for the Region: The proposed framework provides new insights by demonstrating superior forecasting accuracy (R² = 0.957, RMSE = 4.669) compared to established benchmarks. The key hydrological insight for this region is that a component-specific prediction strategy, where predictors are tailored to the statistical properties of decomposed signals, significantly outperforms monolithic or “one-size-fits-all” hybrid approaches. Critically, the model excels at capturing sharp peaks and volatile dynamics where other models fail. This offers a more robust and reliable scientific basis for developing early warning systems, directly enhancing the safety and management of underground mining operations in the region.