Optimal Pump Control for Water Distribution Networks via Data-Based Distributional Robustness

Yi Guo, Shen Wang, Ahmad F. Taha, Tyler H. SummersIEEE Trans. Control. Syst. Technol 2020 · OpenReview

TL;DR

This paper introduces a novel data-driven approach for optimizing water distribution network operations by explicitly accounting for uncertain water demand forecasts. It uses a distributional robustness framework with Wasserstein distance to balance operational costs against the risk of violating critical constraints, leading to more reliable and efficient water management.

What Problem Does This Paper Address?

Municipal water distribution networks (WDNs) are complex systems facing increasing challenges due to factors like energy costs, reliability demands, and the inherent unpredictability of water demand. Traditional optimal control methods often rely on deterministic demand forecasts, which are unrealistic and can lead to inefficient operations or constraint violations when actual demand deviates. While stochastic methods exist, they often make strong assumptions about demand distributions or can be overly conservative. This paper tackles the problem of developing a robust and efficient control strategy for WDNs that can adapt to fluctuating water demands with limited information about forecast errors, without assuming a specific probability distribution for these errors.

Methods

The core of this work is a data-driven Model Predictive Control (MPC) framework that incorporates distributional robustness using the Wasserstein metric. Instead of assuming a known probability distribution for demand forecast errors, the authors define an "ambiguity set" of possible distributions centered around an empirical distribution derived from historical data. This allows the controller to optimize for the worst-case scenario within this ambiguity set, providing a tunable level of conservativeness.

To handle the nonlinear hydraulic dynamics of WDNs, the paper employs a Laplacian-based linearization approach. This simplifies the complex flow-head relationships into a more manageable linear system, enabling the use of efficient successive linearization algorithms for solving the optimization problem. The MPC controller then iteratively solves this linearized, distributionally robust optimization problem over a finite time horizon, updating its decisions as new data becomes available.

Datasets

The authors use a portion of the Barcelona drinking water network as a case study. The nominal water demand patterns are generated using EPANET, and demand forecasting errors are simulated based on a persistence forecast. The electricity price is also provided as a time-varying input.

Key Insights and Ideas

  • Data-driven Robustness: The use of Wasserstein distance to define ambiguity sets allows for robust control strategies that are directly informed by historical data, without requiring strong distributional assumptions. This is crucial for real-world systems where demand patterns can be complex and non-Gaussian.
  • Tunable Conservativeness: The Wasserstein radius ($\epsilon$) and risk aversion parameter ($\rho$) provide explicit knobs to tune the trade-off between operational efficiency (lower costs) and robustness (lower risk of constraint violations). A larger $\epsilon$ leads to more conservative decisions, improving out-of-sample performance.
  • Laplacian Linearization for Efficiency: Simplifying the nonlinear hydraulic model using a Laplacian-based approach makes the computationally intensive MPC problem tractable, enabling real-time or near-real-time control.
  • Balancing Cost and Risk: The framework explicitly balances pump operational costs against the risk of violating tank head constraints, offering a practical approach for operators to manage these competing objectives.

Key Results

The paper demonstrates the effectiveness of their proposed framework through case studies on the Barcelona WDN. Key findings include:

  • Improved Robustness: Increasing the Wasserstein radius ($\epsilon$) leads to more conservative pump schedules and significantly lower risk of tank constraint violations, while ensuring stronger out-of-sample performance.
  • Cost-Risk Trade-off: Higher risk aversion ($\rho$) emphasizes constraint violation mitigation, leading to higher operational costs but reduced risk.
  • Comparison to RBC: The proposed data-driven robust MPC outperforms a traditional Rule-Based Control (RBC) mechanism in managing tank heads under uncertain demand scenarios, highlighting the benefits of a model-based, data-driven approach.
  • Accurate Linearization: The Laplacian-based linearization approach provides a good approximation of the nonlinear hydraulic dynamics, with linearized errors being very small (e.g., less than $2.5 \times 10^{-7}$).

Limitations and Future Work

The authors acknowledge several limitations and suggest avenues for future research:

  • Scalability: The computational complexity increases with network size and the number of demand samples. Future work aims to improve computational affordability for large-scale WDNs.
  • Communication: The current framework requires global communication for optimization. Future work will explore distributed control with local communication.
  • Network Topology: The Laplacian approximation is most effective for tree-like network topologies. More general approximations are needed for mesh networks.
  • Theoretical Feasibility: The feasibility of the proposed data-based MPC is numerically validated. Theoretical analysis of feasibility and convergence of the successive linearization algorithm is a future direction.
  • Other Future Work: Developing data-based distributionally robust control for water contamination, incorporating actuator operational status as controllable variables, and exploring the impact of changing the CVaR risk level ($\beta$).

Summary updated

Abstract

In this article, we propose a data-based methodology to solve a multiperiod stochastic optimal water flow (OWF) problem for water distribution networks (WDNs). The framework explicitly considers the pump schedule and water network head level with limited information of demand forecast errors for an extended period simulation. The objective is to determine the optimal feedback decisions of network-connected components, such as nominal pump schedules and tank head levels and reserve policies, which specify device reactions to forecast errors for accommodation of fluctuating water demand. Instead of assuming that the uncertainties across the water network are generated by a prescribed certain distribution, we consider ambiguity sets of distributions centered at an empirical distribution, which is based directly on a finite training dataset. We use a distance-based ambiguity set with the Wasserstein metric to quantify the distance between the real unknown data-generating distribution and the empirical distribution. This allows our multiperiod OWF framework to trade off system performance and inherent sampling errors in the training dataset. Case studies on a three-tank WDN systematically illustrate the tradeoff between pump operational cost, risks of constraint violation, and out-of-sample performance.
BibTeX
@article{guo2020optimal,
  author = {Yi Guo and Shen Wang and Ahmad F. Taha and Tyler H. Summers},
  title = {{Optimal Pump Control for Water Distribution Networks via Data-Based Distributional Robustness}},
  journal = {IEEE Trans. Control. Syst. Technol.},
  year = {2020},
  url = {https://openreview.net/forum?id=behOXbcVEb},
}
Figure