⚡ electrical engineering

Permutation-Equivariant Deep Equilibrium Networks for Real-Time Single-Bus DER Dispatch with O(1) Parameter Scaling

This paper proposes a permutation-equivariant Transformer Deep Equilibrium Network (DERDEQN) that achieves O(1) parameter scaling and real-time inference for dispatching thousands of heterogeneous distributed energy resources by learning competitive equilibria and clearing prices through an unsupervised, label-free training process based on market-clearing and KKT conditions.

Original authors: Jiang Zhan, Guofeng Liang, SiQi Chen, Tong Su, Fanrong Feng

Published 2026-07-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiang Zhan, Guofeng Liang, SiQi Chen, Tong Su, Fanrong Feng

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Conducting a Gigantic Orchestra

Imagine a power grid as a massive orchestra. Instead of musicians, the "players" are thousands of different devices: electric cars (EVs), home batteries, and flexible appliances like air conditioners. The conductor is the Real-Time Price, which changes every 15 minutes to tell everyone when to play (use energy) and when to rest (save energy).

The goal is to get all these devices to work together perfectly so that the total energy used matches the total energy available, without anyone running out of battery or missing a deadline.

The Problem:
When there are only a few players, the conductor can easily tell each one what to do. But when you have 5,000 players, trying to give individual instructions to every single one using traditional math is like trying to solve a puzzle with a million pieces while running a marathon. It takes too long (too slow for real-time) and requires a computer brain that gets bigger and heavier as you add more players.

The Solution: A "Smart" Conductor with a Magic Memory

The authors created a new type of AI called a Permutation-Equivariant Deep Equilibrium Network (DER-DEQN). Think of this as a super-smart conductor who has learned a special trick: they don't need to memorize the name of every single violinist. They just need to know the type of instrument and the section of the orchestra.

Here is how their solution works, broken down into three main superpowers:

1. The "O(1)" Brain (The Flat Memory Trick)

The Old Way (MLP): Imagine a traditional AI as a student who has to write a unique instruction manual for every single device. If you add 100 more electric cars, the student has to write 100 more pages. The manual gets huge, and the student gets overwhelmed.
The New Way (Transformer): The authors' AI is like a master chef with a single, perfect recipe. Whether they are cooking for 50 people or 5,000, the recipe book stays the exact same size. They just apply the same rules to the new ingredients.
The Result: The AI's "brain size" (number of parameters) stays tiny and constant, no matter how many devices are added. It can handle 5,000 devices with the same mental effort as 50.

2. The "Symmetry" Rule (The Identical Twins Analogy)

The Concept: If you have two identical electric cars with the same battery level and the same deadline, they should get the exact same instruction.
The Old Way: Traditional AIs often get confused. If you swap the names of the two cars, the AI might give them different instructions, which is unfair and inefficient.
The New Way: This AI is "Permutation Equivariant." It understands that if you swap two identical twins, the instructions should just swap with them. It treats identical devices as identical. The paper shows this works with near-perfect precision (like a machine that never makes a mistake in this specific rule).

3. Learning Without a Teacher (The "Self-Correcting" Loop)

The Challenge: Usually, to teach an AI, you need a "teacher" who shows it the perfect answer (labeled data). But finding the perfect answer for 5,000 devices is so hard that even the teacher doesn't know it yet.
The Trick: This AI learns by checking its own homework. It doesn't need a teacher. Instead, it has a set of rules (the "Equilibrium Conditions"):
1. Does the power balance? (Supply = Demand)
2. Are the devices happy? (Did they get the best deal for their battery?)
3. Are the constraints met? (Did the car charge in time?)
The AI makes a guess, checks these rules, sees where it went wrong, and adjusts itself. It keeps doing this until the "mistakes" (residuals) are tiny. It's like a musician practicing alone until they can play perfectly without a conductor.

The Results: Speed and Scale

The authors tested this on two powerful graphics cards (like high-end gaming computers):

Speed: When there are 5,000 devices, the old methods (solving complex math equations) take minutes or even hours. This new AI makes a decision in 8.8 milliseconds. That is faster than a human eye blink. It is fast enough to be used in real-time, second-by-second dispatch.
Scale: The AI works perfectly for 5,000 devices. The old methods would crash or take too long at that size.
Accuracy: The AI gets the "big picture" right (the market clears, and constraints are met). However, the authors admit one small flaw: while it gets the general price right, it doesn't get the perfect theoretical profit for every single device 100% of the time. It's "good enough" for real-world use, but not mathematically perfect in every tiny detail.

Summary

The paper introduces a new AI conductor that can manage thousands of energy devices instantly.

It's efficient: Its brain size doesn't grow as the crowd gets bigger.
It's fair: It treats identical devices identically.
It's fast: It makes decisions in less than 10 milliseconds.
It's self-taught: It learns by checking its own rules rather than needing a teacher to show it the answers.

This allows power grids to manage massive fleets of electric cars and batteries in real-time, something that was previously too slow and complex for computers to handle.

Technical Summary: Permutation-Equivariant Deep Equilibrium Networks for Real-Time Single-Bus DER Dispatch

1. Problem Formulation

The paper addresses the challenge of real-time dispatch for heterogeneous Distributed Energy Resources (DERs)—specifically Electric Vehicles (EVs), behind-the-meter battery storage (ST), and flexible demand-response loads (DR)—under a single-bus real-time pricing scheme.

The core difficulty lies in solving a high-dimensional constrained equilibrium every dispatch interval (e.g., every 15 minutes). In this setting:

Device Dynamics: Each device $i$ optimizes its intertemporal energy usage subject to physical constraints (State-of-Charge limits, charging windows, power ratings) and device-specific parameters (e.g., urgency for EVs, elasticity for DR).
Market Clearing: A single real-time price $\lambda(t)$ must clear the aggregate power balance, coupling all $N$ devices.
Scalability Bottlenecks:
1. Curse of Dimensionality: Centralized mixed-integer or convex programs become prohibitively slow as $N$ scales into the thousands.
2. Cross-Device Interaction: Devices are coupled only through the aggregate balance and price, a structure standard fully connected architectures fail to exploit efficiently.
3. Real-Time Latency: Re-solving optimization problems every interval fails to meet tight latency budgets (5–15 minutes) for large fleets.

Existing learning-based approaches are typically supervised, requiring expensive labeled data generated by offline solvers, and often suffer from poor scaling with $N$ .

2. Methodology: DER-Transformer-DEQN

The authors propose a Permutation-Equivariant Transformer Deep Equilibrium Network (DER-DEQN). This approach casts the dispatch problem as a competitive equilibrium and learns the policy and clearing price simultaneously without labeled data.

A. Competitive Equilibrium Formulation

The problem is modeled as a competitive equilibrium where:

Devices are price-takers maximizing intertemporal value given $\lambda(t)$ .
The price $\lambda(t)$ is the dual variable of the market-clearing constraint.
This structure is analogous to the heterogeneous-agent bond-market equilibrium solved by original DEQN methods, but adapted for power systems.

B. Architecture: O(1) Parameter Scaling

The network utilizes a Transformer architecture to achieve parameter efficiency independent of the number of devices $N$ :

Tokenization: The state is serialized into an aggregate context vector $g(t)$ and $N$ per-device tokens $h_i$ .
Shared Weights: Unlike a Multi-Layer Perceptron (MLP) where input/output layer widths scale with $N$ (resulting in $O(N)$ parameters), the Transformer shares weights across all device tokens. The parameter count remains $O(1)$ regardless of fleet size.
Composite Positional Encoding: To inject necessary physical structure while maintaining permutation equivariance, a composite positional encoding is added:
$PE_i = \text{TypeEmbed}(\tau_i) \oplus \text{WindowEmbed}(w_i) \oplus \text{FeederEmbed}(f_i) \oplus \text{SinRank}(r_i)$
This encodes device type, availability window, feeder class, and a sinusoidal rank (tie-breaker), breaking symmetry only where physical identity dictates.

C. Unsupervised Equilibrium Loss

The model is trained without labeled optimal dispatches by minimizing the residual of the equilibrium conditions. The loss function $L$ comprises five components:

Intertemporal Marginal-Value (FOC): Analogous to Euler equations, ensuring marginal value equals price for storage, EVs, and DR.
KKT Power-Box: Complementarity conditions for power limits.
KKT SoC-Bounds: Complementarity for State-of-Charge limits.
KKT EV-Deadline: Conditions for meeting target charge by departure.
Market Clearing: The residual of the aggregate power balance equation, which pins the real-time price.

The loss uses a relative-floor normalization to ensure scale invariance and stable training.

3. Key Contributions

Problem Formulation: Casting real-time single-bus DER dispatch as a competitive equilibrium structurally analogous to heterogeneous-agent macroeconomic models.
Architecture Design: Introducing a permutation-equivariant Transformer-DEQN with $O(1)$ parameter scaling in $N$ , contrasting with the $O(N)$ scaling of standard MLP baselines.
Symmetry Injection: Redesigning positional encoding to incorporate device-type, temporal-window, and feeder priors as the only symmetry-breaking terms.
Label-Free Training: Designing an unsupervised loss based on equilibrium residuals (Euler analogues, KKT complementarities, and market clearing) that requires no optimal dispatch labels.
Empirical Validation: Demonstrating the method's properties at two scales:
- Architecture Scale ( $N=5000$ ): Validating parameter efficiency and latency.
- Equilibrium Quality ( $N=50$ ): Validating convergence of residuals and welfare comparison.

4. Experimental Results

Experiments were conducted on two NVIDIA RTX 3090 GPUs.

A. Parameter Scaling and Efficiency

Parameters: The Transformer model maintains a flat 119,142 parameters across all fleet sizes ( $N=50$ to $5000$). In contrast, an MLP baseline scales from $1.87 \times 10^6$ to $8.60 \times 10^7$ parameters at $N=5000$ (a 720x gap).
Permutation Equivariance: The Transformer achieves machine-precision equivariance ( $\sim 2 \times 10^{-6}$ error) across permutations, whereas the MLP exhibits errors growing linearly with $N$ .

B. Real-Time Latency

The learned policy executes in 8.8 ms for $N=5000$ (median latency), well within the 5–15 minute dispatch budget.
This is two orders of magnitude faster than centralized solvers (CVXPY/CLARABEL) and significantly faster than Mixed-Integer Linear Programming (MILP) approaches, which become intractable for large $N$ .

C. Equilibrium Quality and Welfare

Residual Convergence: At $N=50$ , the market-clearing and KKT residuals converge to $\sim 10^{-4}$ , indicating stable label-free training.
Welfare Comparison: The learned dispatch achieves competitive welfare compared to the centralized convex optimum. However, the intertemporal marginal-value (FOC) residual settles to a non-zero floor ( $\sim 0.6$ ).
Limitation: Due to this FOC floor and the unanchored price scale (Proposition 1), the dispatch is approximately, but not exactly, welfare-optimal. The paper notes that the policy's price path does not perfectly track the shape of the dual price from the centralized solver, suggesting the gap is driven by the FOC residual rather than simple scale mismatch.

D. Training Constraints

While parameters scale $O(1)$ , the attention activation memory scales $O(N^2)$ . Full-batch training at $N=5000$ exceeds GPU memory (24 GB) due to the $[B, H, N, N]$ attention matrix. Consequently, equilibrium quality results are reported at $N=50$ , the largest scale where the unsupervised loss converges under the current training budget.

5. Significance and Claims

The paper claims that this work bridges the gap between computational economics and power system operations by enabling real-time, fleet-scale dispatch on commodity hardware.

Scalability: The $O(1)$ parameter scaling allows a single fixed-capacity network to serve fleets of any size, eliminating the need to retrain or resize models as devices join or leave the grid.
Real-Time Feasibility: Sub-10 ms inference times make the method viable for dispatch cycles where centralized optimization fails.
Theoretical Alignment: By satisfying equilibrium conditions directly, the method respects the economic structure of the problem (price-mediated coordination) without relying on supervised imitation of offline solvers.

Modest Limitations Acknowledged:
The authors explicitly state that exact welfare parity with centralized solvers remains an open challenge due to the non-zero FOC residual floor. Furthermore, the current $O(N^2)$ memory bottleneck for training large fleets requires future adoption of linear or flash-attention variants to fully realize the method's potential at the largest scales. The current study is limited to a single-bus model without network constraints.