Modeling Reservoir Release Using Pseudo-Prospective Learning and Physical Simulations to Predict Water Temperature

Xiaowei Jia, Shengyu Chen, Yiqun Xie, Haoyu Yang, Alison P. Appling, Samantha Oliver, Zhe Jiang

TL;DR

This paper introduces a novel data-driven approach, the State-Aware Graph (SAG) model, to predict water temperature in river networks with reservoirs. It tackles the challenge of missing reservoir release data by employing a "pseudo-prospective" learning method and further enhances predictions by incorporating physics-based simulations for reservoirs with available data.

What Problem Does This Paper Address?

Accurately predicting water temperature in river networks is crucial for aquatic ecosystems and human water management. A significant challenge arises from the influence of reservoirs, which can drastically alter downstream water temperatures, especially when releasing cold water. However, obtaining detailed reservoir release data (volume and depth) is often difficult due to privacy concerns or lack of monitoring. Existing data-driven models struggle to capture these reservoir impacts, leading to inaccurate predictions, particularly during periods of significant release. This paper aims to bridge this gap by developing a method that can effectively model reservoir influences even when release data is unavailable, and improve predictions by leveraging available data and physical insights.

Methods

The core of the proposed solution is the State-Aware Graph (SAG) model. This model represents the river network and reservoirs as a heterogeneous graph, where nodes represent river segments and reservoirs, and edges capture their interactions. The SAG model maintains distinct "state variables" for both river segments and reservoirs, allowing it to learn how their temperatures evolve and influence each other over time.

To address the missing reservoir release data, the authors introduce a Pseudo-Prospective (PP) learning approach. This method mimics how water managers might make release decisions by anticipating future downstream water temperatures. It uses a separate forecasting model to predict future temperatures and then uses this anticipation to infer reservoir release information.

For reservoirs where release data is available, the paper proposes a Simulation-based Embedding (SE) approach. This involves using a physics-based General Lake Model to simulate reservoir temperature profiles at different depths. This simulated data is then combined with actual release flow information to create a more accurate "release embedding" that captures the physical impact of the release.

Crucially, the paper proposes a parallel learning structure where the knowledge gained from the SE approach (for reservoirs with data) is transferred to improve the PP approach (for reservoirs without data). This allows the model to learn from "transparent" reservoirs and generalize that knowledge to others.

Datasets

The study utilizes data from the Delaware River Basin. This includes:

  • Daily climate drivers (precipitation, air temperature, solar radiation).
  • Geometric features of river segments.
  • Static meta-features of reservoirs (height, width, etc.).
  • Observed water temperature data for numerous river segments.
  • Reservoir release data (volume and depth) for a subset of reservoirs (specifically, Cannonsville and Pepacton Reservoirs).

The paper focuses on two spatially disjointed subsets of the Delaware River Basin, S₁ and S₂, to test generalization capabilities.

Key Insights and Ideas

  • Reservoir release data is a critical, yet often missing, piece of the puzzle for accurate river temperature prediction. The paper demonstrates that ignoring this can lead to significant over-prediction, especially during summer.
  • Mimicking human decision-making (pseudo-prospective learning) can be a powerful way to infer missing data. By anticipating future conditions, the model can effectively "guess" what a human operator might do.
  • Combining physics-based models with data-driven methods offers a synergistic advantage. Physics provides crucial domain knowledge that can compensate for data scarcity and improve the interpretability and accuracy of data-driven models.
  • Knowledge transfer between reservoirs is feasible and beneficial. Even if some reservoirs are "black boxes," learning from "transparent" ones can significantly improve predictions for the black boxes.

Key Results

  • The proposed SAG model with the PP approach (SAG-pp) achieved over 10% accuracy improvement compared to existing data-driven models (ANN, RNN, RGrN) when reservoir release data was unavailable for any reservoirs.
  • When incorporating release data and physical simulations for a subset of reservoirs (using the SE approach), the performance was further improved, with methods like SAG-sim achieving the lowest RMSE.
  • The model demonstrated good generalization capabilities to spatially disjointed regions, outperforming baseline models.
  • The paper shows that the SE approach (SAG-flow and SAG-sim) can better predict temperature even under sudden increases in reservoir release, outperforming the PP-only approach.

Limitations and Future Work

  • Limited scope of PP approach: The current PP method only considers anticipated temperature for the next day. Water managers might consider multiple future days, which could be incorporated for better accuracy.
  • Varied release criteria: The model assumes uniform release criteria across reservoirs. In reality, different reservoirs may have distinct management strategies, which could be modeled for further improvement.
  • Data availability: While the paper addresses missing data, the SE approach still relies on the availability of release data for some reservoirs to enable knowledge transfer.

Future work could explore incorporating multi-day forecasting for the PP approach, modeling reservoir-specific release criteria, and investigating more sophisticated methods for transferring knowledge between reservoirs with different characteristics.

Summary updated

BibTeX
@inproceedings{jia2022modeling,
  author = {Xiaowei Jia and Shengyu Chen and Yiqun Xie and Haoyu Yang and Alison P. Appling and Samantha Oliver and Zhe Jiang},
  title = {{Modeling Reservoir Release Using Pseudo-Prospective Learning and Physical Simulations to Predict Water Temperature}},
  booktitle = {SDM},
  year = {2022},
  doi = {10.1137/1.9781611977172.11},
}

Loading related content...

Figure