Dynamic Pricing Control for Constrained Distribution Networks With Storage

Mathias Bürger, Claudio De Persis, Frank Allgöwer

TL;DR

This paper introduces a dynamic feedback controller for distribution networks that manages time-varying supply and demand in real-time. By leveraging the "dual variables" of network optimization—often interpreted as prices—the authors create a system that simultaneously minimizes transportation costs and balances inventory levels across all storage nodes.

What Problem Does This Paper Address?

In traditional network optimization, we usually solve for the best way to move goods from point A to point B given a static set of requirements. However, real-world systems like supply chains, power grids, and data networks are rarely static. Supply and demand fluctuate constantly.

The challenge is twofold:

  1. Optimal Routing: How do we adjust flows instantaneously to minimize costs as conditions change?
  2. Inventory Balancing: How do we ensure that no single warehouse is overflowing while another is empty?

Existing solutions often rely on "one-step" predictive models or struggle with hard capacity constraints on transportation lines. This paper seeks a continuous-time control law that is robust, handles time-varying inputs, and respects physical capacity limits.

Methods

The authors model the network as a graph where nodes are inventories and edges are transportation lines. The core of their approach involves:

  • Legendre-type Cost Functions: Instead of simple quadratic costs, they use a specific class of convex functions. These are clever because they can represent hard capacity constraints; as the flow approaches a physical limit, the "cost" shoots toward infinity, naturally pushing the controller away from breaking the limit.
  • Dual Variable Characterization: The authors show that the optimal flow can be described using time-varying dual variables ($\zeta$). In economic terms, these variables act as "shadow prices" for the goods at each node.
  • The Generalized Dirichlet Problem: To make the controller work in real-time, the authors derive an explicit differential equation for these dual variables. This involves solving a version of the "Dirichlet problem" on a graph to find how prices should evolve to keep the network optimal.
  • Dynamic Feedback Control: They propose a controller that looks like a continuous-time Newton algorithm. It combines a "feed-forward" part (which anticipates changes in supply/demand) with a "feedback" part (which corrects inventory imbalances).

Datasets

This paper is theoretical and does not use large-scale empirical datasets. Instead, the authors validate their approach using a numerical simulation of a three-node inventory system with three transportation lines. The supply and demand are modeled as harmonic signals (sine waves) generated by an external dynamical system ("exo-system").

Key Insights and Ideas

  • Prices as Controllers: One of the most elegant "aha moments" is the realization that the dual variables of the optimization problem can serve as the state of the controller. By regulating these "prices," the system naturally finds the most efficient flow.
  • Consensus and Balancing: The authors link the problem of balancing inventory levels to the well-known "consensus" or "output agreement" problem in multi-agent systems. Balancing isn't just a nice-to-have; it's mathematically integrated into the optimization.
  • Soft Constraints, Hard Limits: By using Legendre-type functions, hard capacity constraints are handled "softly" within the math but "strictly" in practice. This avoids the need for complex switching logic or discontinuous control laws.

Key Results

  • Asymptotic Optimality: The authors provide a rigorous mathematical proof that their controller will cause the network flows to converge to the mathematically optimal values over time, even if the system doesn't start in an optimal state.
  • Robustness to Shocks: In simulations, when the phase of the supply/demand signal was changed instantaneously (a "shock" to the system), the controller successfully readjusted the flows and re-balanced the inventories without violating capacity bounds.
  • Efficiency: The simulation demonstrated that the controller intelligently prioritizes low-cost routes but shifts to higher-cost routes only when the cheaper lines approach their capacity limits.

Limitations and Future Work

  • Information Requirements: Currently, the controller assumes that the external supply and demand (and their derivatives) are known or can be measured. The authors note that future work should focus on designs that don't require direct measurement of these external factors.
  • Centralization: The current derivation is centralized. For massive networks (like the global internet or national power grids), a distributed implementation—where each node only talks to its neighbors—would be a significant next step.
  • Balanced Supply/Demand: The paper assumes the total supply equals the total demand across the whole network. Handling net surpluses or deficits in a dynamic way remains an open challenge.

Summary updated

Abstract

This paper studies the problem of optimal flow control in dynamic inventory systems. A dynamic optimal distribution problem, including time-varying supply and demand, capacity constraints on the transportation lines, and convex flow cost functions of Legendre-type, is formalized and solved. The time-varying optimal flow is characterized in terms of the time-varying dual variables of a corresponding network optimization problem. A dynamic feedback controller is proposed that regulates the flows asymptotically to the optimal flows and achieves in addition a balancing of all storage levels.
BibTeX
@article{burger2015dynamic,
  author = {Mathias Bürger and Claudio De Persis and Frank Allgöwer},
  title = {{Dynamic Pricing Control for Constrained Distribution Networks With Storage}},
  journal = {IEEE Trans. Control. Netw. Syst.},
  year = {2015},
  url = {https://openreview.net/forum?id=Q1JpDp0lfZK},
}

Loading related content...

Figure