Robust Interpretable Hourly Runoff Forecasting Based on High-Performance Neural Networks
- Publication Type:
- Thesis
- Issue Date:
- 2025
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Scientific development and management of river require collaborative efforts among multidisciplinary stakeholders and the integration of cross-disciplinary knowledge, wherein runoff forecasting plays a pivotal role. With the widespread adoption of data-driven deep learning models, particularly neural networks, in the hydrological domain, high-precision runoff forecasting, which was previously difficult to achieve using traditional physically based models, has now become feasible. However, current neural network-based runoff forecasting models still encounter various challenges in real-world applications. To address limitations in efficiency, accuracy, temporal robustness, and interpretability, this thesis progressively proposes a series of diversified and high-performance runoff forecasting frameworks.
First, to address the overemphasis on accuracy in existing runoff forecasting models, we develop a lightweight and robust framework based on Temporal Convolutional Network (TCN), which employs dilated and causal convolutions to expand the receptive field and prevent information leakage. An attention module is integrated to enhance accuracy with low computational cost, and an improved Snapshot ensemble strategy is used during training to boost robustness under extreme perturbations. Moreover, to overcome the limitations of mainstream neural networks, such as limited receptive fields and long-term dependencies modeling, we further propose two high-performance forecasting frameworks. These incorporate an enhanced ResNet with dual pathways and dense shortcuts to optimize information flow and benefit from deeper network structures. Conventional attention mechanisms that focus on one single dimension are further extended to both temporal and spatial dimensions based on channel-dependence (CD) and channel-independence (CI) strategies. Bidirectional architecture and temporal shortcuts are also integrated to capture richer context and mitigate vanishing gradients in long sequences. Additionally, in response to the performance degradation observed in mainstream models during extended forecasting horizons, we propose a multi-lead-time forecasting framework grounded in the state space model (SSM), characterized by its dual attributes of convolution and recursion while still maintaining linear complexity. The effectiveness of the proposed framework is validated through quantitative evaluation, revealing strong temporal robustness across multiple forecasting horizons within the upcoming 24-hour. Notably, a model-specific local post-hoc explanation technique based on interpretable machine learning (IML) is also used to enhance the interpretability of the model's forecasting process.
In summary, this thesis proposes a set of runoff forecasting models that achieve state-of-the-art overall performance. Through progressive architectural enhancements across models, the structural limitations of existing approaches are effectively overcome, enabling highly accurate multi-lead-time forecasting of fine-grained hourly runoff sequences. These contributions provide robust and reliable solutions for stakeholders in hydrology.
Please use this identifier to cite or link to this item: