Highly Robust South Fork Zumbro River Flow Forecasting Based on Deep Temporal Modeling

Abstract

River flow forecasting plays a vital role in water resource management and ecological conservation. Accurate flow forecasting enables decision makers to allocate resources efficiently, implement early flood prevention measures, and protect ecosystems. However, environmental noise interferes with forecasting, reducing the precision and reliability of decision making. To address this, we propose a Highly Robust River Flow Forecasting Model (HRRFFM). The model comprises two components: data preprocessing and deep temporal modeling. Data preprocessing involves interpolation, environmental noise simulation, and Wiener filtering to improve data quality and model robustness. The deep temporal modeling integrates Bidirectional Long Short-Term Memory (BILSTM) networks and Transformer architecture to capture river flow dynamics. BILSTM captures bidirectional features, enhancing the model’s capacity to learn complex flow sequences. Transformer utilizes self-attention and multi-head attention mechanisms to model global dependencies and amplify subtle time-series variations, significantly improving feature extraction efficiency and forecasting accuracy. In this paper, Gaussian noise is employed to simulate environmental disturbances. The model’s performance is validated through ablation studies across varying noise levels and forecast horizons. At the noise intensity of $\sigma$ = 0.05, for the three-hour-ahead predictions, HRRFFM outperforms baseline models with average improvements of 11.56%, 13.59%, and 4.98% in RMSE, MAE, and $R^2$, respectively.