HydroDCM: Hydrological Domain-Conditioned Modulation for Cross-Reservoir Inflow Prediction

TL;DR

HydroDCM addresses the domain shift challenge in cross-reservoir inflow forecasting by treating reservoir spatial metadata as pseudo-domain cues and learning domain-invariant temporal representations via adversarial training and contrastive learning. It then reintroduces domain-specific nuance through a lightweight FiLM-based modulation conditioned on reservoir attributes during inference, enabling accurate predictions for unseen reservoirs with limited data. Evaluated on 30 Upper Colorado Basin reservoirs, HydroDCM outperforms state-of-the-art domain generalization baselines across multiple forecast horizons while remaining computationally efficient. This approach offers a practical, scalable solution for reliable multi-reservoir hydrological forecasting under distributional heterogeneity.

Abstract

Deep learning models have shown promise in reservoir inflow prediction, yet their performance often deteriorates when applied to different reservoirs due to distributional differences, referred to as the domain shift problem. Domain generalization (DG) solutions aim to address this issue by extracting domain-invariant representations that mitigate errors in unseen domains. However, in hydrological settings, each reservoir exhibits unique inflow patterns, while some metadata beyond observations like spatial information exerts indirect but significant influence. This mismatch limits the applicability of conventional DG techniques to many-domain hydrological systems. To overcome these challenges, we propose HydroDCM, a scalable DG framework for cross-reservoir inflow forecasting. Spatial metadata of reservoirs is used to construct pseudo-domain labels that guide adversarial learning of invariant temporal features. During inference, HydroDCM adapts these features through light-weight conditioning layers informed by the target reservoir's metadata, reconciling DG's invariance with location-specific adaptation. Experiment results on 30 real-world reservoirs in the Upper Colorado River Basin demonstrate that our method substantially outperforms state-of-the-art DG baselines under many-domain conditions and remains computationally efficient.

Figures (3)

• Figure 1: Overview of HydroDCM architecture for reservoir inflow prediction. Given the observations $\mathbf{X}_i$, (1) the feature encoder $f_\phi(\cdot)$ extracts temporal representations $\mathbf{h}_i$, which are refined into invariant features $\mathbf{z}_i$ through adversarial learning with $d_\theta(\cdot)$ using spatial metadata $\mathbf{s}_i$ as pseudo-domain labels; (2) modulation parameters $\{\gamma(\mathbf{s}_i), \delta(\mathbf{s}_i)\}$ in the FiLM adapter $m_\beta(\cdot)$ adjust $\mathbf{z}_i$ into $\tilde{\mathbf{z}}_i = \gamma(\mathbf{s}_i)\odot \mathbf{z}_i + \delta(\mathbf{s}_i)$, which is then passed to the predictive head $p_\omega(\cdot)$ for estimating future inflow $\hat{y}_i$.
• Figure 2: Geographical Position and River Network among 30 reservoirs in the Upper Colorado River Basin.
• Figure 3: Comparison of model performance on daily NSE scores across three reservoirs (MCR, JVR, and MCP) from Day 3 to Day 7. Each subplot shows the NSE (%) for DANN, MLDG, CondAdv, IRM, and HydroDCM on a specific reservoir–day combination. Error bars indicate standard deviations over five independent runs. The horizontal black, blue, and red dash lines represent the Base, Few-shot, and Oracle standards for observing enhanced robustness, respectively.