Uncertainty-aware demand management of water distribution networks in deregulated energy markets

Pantelis Sopasakis, Ajay K. Sampathirao, Alberto Bemporad, Panagiotis PatrinosEnviron. Model. Softw 2018 · DOI · OpenReview

Introduction

Operating a municipal water distribution network (DWN) is a complex, energy-intensive task that requires balancing reliable service with operational efficiency. Utilities must strategically pump water into reservoirs during periods of low electricity prices to meet consumer demand during peak hours. However, modern deregulated energy markets introduce significant volatility in electricity pricing, while consumer water demand remains inherently stochastic.

Traditional control strategies often rely on nominal predictions or conservative worst-case scenarios. The former risks service interruptions during demand spikes, while the latter incurs excessive operational costs. To address these challenges, researchers have proposed Stochastic Model Predictive Control (SMPC). While theoretically robust, SMPC is computationally demanding, often involving millions of decision variables, which has historically limited its application in real-time hourly decision-making. This paper introduces RapidNet, a GPU-accelerated software framework designed to make large-scale SSMPC feasible for modern water utilities.

System Modeling and Modular Framework

A water distribution network can be modeled as a collection of interconnected modular components. RapidNet leverages this structure to represent the flow dynamics within the system. The network typically consists of storage tanks, mixing nodes, controlled pumps, valves, and demand sectors.

Modular components of a water distribution network, including storage tanks, mixing nodes, and control elements like pumps and valves.

As shown in the figure above, each component is defined by specific flow and mass balance equations. For instance, controlled pumps and valves ($q^i = u^i$) serve as the primary actuators for managing the network's state, while demand sectors ($q^i = d^i$) represent the stochastic sinks in the system. By decoupling the network into these fundamental units, the framework allows for flexible modeling of diverse metropolitan infrastructures.

Stochastic Model Predictive Control via Scenario Trees

The core of RapidNet’s approach to uncertainty is Scenario-based Stochastic Model Predictive Control (SSMPC). Rather than relying on a single deterministic forecast, the system utilizes a branching scenario tree to capture potential future trajectories of water demand and electricity prices.

A scenario tree representing the branching evolution of uncertain demand and price paths over time.

This scenario tree structure (illustrated above) captures the "multistage covariance" of the system—the principle that errors in predictions at one time step are likely correlated with errors in subsequent steps. By optimizing over this tree, the controller can make decisions that are robust to a wide range of probable futures. The researchers found that a "sweet spot" of approximately 631 scenarios provides an optimal balance between reliability (the Safety Index) and operational cost (the Economic Index).

GPU-Accelerated Optimization

The primary technical innovation of RapidNet is its use of prox-friendly optimization implemented in CUDA-C++. Traditional "interior point" solvers used in commercial software are difficult to parallelize. In contrast, RapidNet employs an accelerated proximal gradient method. This algorithm decomposes the massive stochastic optimization problem into millions of simple, independent mathematical operations.

Because these operations are independent, they can be offloaded to a Graphics Processing Unit (GPU). By utilizing thousands of GPU cores to perform calculations in lock-step, RapidNet effectively equalizes the time required to solve a complex stochastic problem with that of a simple deterministic one. This allows the framework to solve problems with over 2.3 million primal variables in under 83 seconds.

Case Study: The Barcelona Water Network

The framework was validated using a high-fidelity simulation of the Barcelona Water Network, a massive system comprising 63 tanks, 75 pumps, 39 valves, and 88 demand sectors. Real-world hourly electricity price data from the Energy Exchange Austria (EXAA) and historical demand data were used to drive the simulation.

Key findings from the case study include:

  • Cost Efficiency: Accounting for price uncertainty led to a 4% reduction in total operating costs, saving approximately €199 per hour. Ignoring price volatility was found to make operations roughly 5% more expensive overall.
  • Computational Speed: RapidNet achieved a 10x speedup compared to Gurobi, a leading commercial CPU-based solver, while maintaining the same level of accuracy.
  • Architectural Flexibility: The framework decouples the forecasting engine from the optimization core, allowing utilities to integrate advanced machine learning models (such as Support Vector Machines) for demand prediction without altering the underlying control logic.

Conclusion and Future Work

RapidNet demonstrates that the inherent structure of water network equations is exceptionally well-suited for GPU architectures. By rethinking optimization algorithms to fit consumer-grade hardware, utilities can achieve significant energy savings and improved reliability without the need for supercomputing resources.

Future development of the framework will focus on expanding beyond "flow-based" models to include "pressure-based" networks, which involve non-convex constraints. The authors suggest integrating the PANOC algorithm and operator splitting techniques to handle these non-linearities. Additionally, scaling the library to support multi-GPU architectures will enable the management of even larger infrastructure systems, such as regional power grids.

Summary updated

BibTeX
@article{sopasakis2018uncertainty,
  author = {Pantelis Sopasakis and Ajay K. Sampathirao and Alberto Bemporad and Panagiotis Patrinos},
  title = {{Uncertainty-aware demand management of water distribution networks in deregulated energy markets}},
  journal = {Environ. Model. Softw.},
  year = {2018},
  doi = {10.1016/j.envsoft.2017.11.035},
}
Figure