Scale-Aware and Causality-Driven Modeling for River Water Quality Prediction

Authors:

Abstract
River water quality time series exhibit pronounced multi-scale non-stationary characteristics. Under diffuse pollution conditions dominated by agricultural runoff and overland flow, processes at different temporal scales are governed by distinct driving factors and regulatory mechanisms. However, most existing models rely on a single temporal scale, leading to cross-scale information mixing and consequently limiting predictive accuracy and mechanistic interpretability.To address this limitation, we propose a multi-scale water quality prediction framework oriented toward scale-dependent driving mechanisms, integrating time-series decomposition, causal identification, and predictive modeling at each scale. Time-Varying Filter–based Empirical Mode Decomposition (TVF-EMD) was applied to decompose total nitrogen (TN), ammonia nitrogen (AN), total phosphorus (TP), and permanganate index (CODMn) into high-frequency, intermediate-scale, and long-term components. Granger causality analysis was conducted at each scale to identify dominant driving factors and lag structures, based on which scale-specific models were developed and reconstructed.The proposed framework significantly outperformed direct modeling based on the original series for all indicators, achieving test-set R² values of 0.850 (TN), 0.914 (AN), 0.967 (TP), and 0.732 (CODMn), while substantially reducing mean absolute error (MAE) and root mean square error (RMSE). This study advances water quality prediction from scale-independent statistical fitting to scale-dependent causal modeling and enables proactive downstream risk prediction based on upstream monitoring, providing practical implications for early warning and management in diffuse-pollution-dominated watersheds.

This publication has 21 references indexed in Scilit: