🔢 mathematics

Trading in residential energy systems with storage: a kinetic mean-field approach

This paper proposes a kinetic mean-field control framework for coordinating large ensembles of residential storage devices, modeling battery dynamics through position-velocity-acceleration variables to capture ramp-rate constraints and endogenous population interactions, and solves the resulting complex five-dimensional stochastic optimization problem using deep learning-based numerical methods.

Original authors: Margherita Fabini, Andrea Pascucci, Alessio Rondelli

Published 2026-03-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Margherita Fabini, Andrea Pascucci, Alessio Rondelli

Original paper licensed under CC BY 4.0 (http://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: The Energy Conductor

Imagine a massive orchestra of thousands of homes. Each home has a solar panel (a musician) and a battery (a storage box). Some days, the sun is bright, and the home has extra energy. Other days, it's cloudy, and the home needs power from the grid.

Usually, every home acts alone. They charge their batteries when the sun shines and discharge when it's dark, reacting only to their own needs. This is chaotic. Sometimes, everyone tries to charge at the exact same time, causing a traffic jam on the power grid.

The Problem: The electricity market is like a volatile stock market. Prices change wildly every hour. If you could buy electricity when it's cheap (night) and sell it when it's expensive (day), you could make money. But individual homes are too small and too "dumb" to do this efficiently without breaking their batteries.

The Solution: Enter the Aggregator. Think of the Aggregator as a Conductor for this orchestra. Instead of letting every house play its own tune, the Conductor directs them all to play in harmony. The Conductor controls the batteries of all these homes to:

Smooth out the grid's bumps (so the lights don't flicker).
Buy low and sell high (arbitrage) to make money for the homeowners.

The Core Innovation: The "Kinetic" Approach

Most math models for batteries treat them like light switches: you can flip them from "off" to "full power" instantly. But real batteries are more like cars.

The Analogy: You can't instantly go from 0 to 100 mph. You have to accelerate. You also can't brake instantly; you have to slow down. If you try to change speed too fast, you damage the engine (the battery).
The Paper's Trick: The authors treat the battery's State of Charge (how full it is) as Position. The Power (charging/discharging speed) as Velocity. And the Control (the decision to speed up or slow down) as Acceleration.

This is called a Kinetic approach. It forces the math to respect the physical reality that you can't jerk a battery's power up or down instantly. It creates smooth, safe trajectories, just like a good driver.

The "Mean-Field" Crowd Control

Now, imagine the Conductor isn't just talking to one car; they are talking to 10,000 cars on a highway.

The Challenge: If the Conductor tells everyone to charge at 6:00 PM, the grid might crash because everyone is doing it at once.
The Solution: The Mean-Field approach. The Conductor doesn't look at every single car individually. Instead, they look at the average behavior of the crowd.
- "The average battery is 50% full. Let's tell the group to charge a little more."
- "The average battery is full. Let's tell the group to stop."

This creates a feedback loop. The decisions of the group influence the individual, and the individual's actions influence the group's average. The math in this paper solves the puzzle of how to control this massive crowd so that they don't all panic at the same time, but instead move together like a school of fish.

The "Physics" of the Math

The paper uses a very specific type of math called McKean-Vlasov Langevin dynamics. That sounds scary, but here is the simple version:

Langevin Dynamics: This is a way of modeling things that move randomly (like gas particles or stock prices) but are also being pushed by a force (the Conductor).
Hypoellipticity: This is a fancy math word that basically means "even though we can't control every single part of the system directly, the randomness and the connections between parts eventually smooth everything out." It ensures the solution is stable and doesn't blow up.

Think of it like steering a giant, foggy ship. You can't see the whole ocean, and the wind is blowing randomly, but because the ship is so big and connected, your small steering adjustments eventually guide the whole vessel safely to port.

The "Brain" of the Operation: Deep Learning

Solving this math problem is incredibly hard. It involves a 5-dimensional puzzle (Time, Price, Load, Battery Speed, Battery Level) that is constantly changing based on the crowd's behavior. Traditional computers can't solve this fast enough.

The authors used Deep Learning (Artificial Intelligence) to act as the Conductor's brain.

They trained a Neural Network (a digital brain) to look at the current situation (price, weather, battery levels) and predict the best move.
Instead of calculating the answer for every single house, the AI learns the "pattern" of the perfect strategy.
The Result: The AI learned to charge batteries when prices were low and discharge when prices were high, all while keeping the batteries safe and the grid stable.

The Results: Why It Matters

When the authors tested this model with real data from Italy (using actual electricity prices and solar data):

Money Saved: The optimized strategy made significantly more money (or saved more cost) than just letting batteries run on autopilot.
Battery Health: Because the model respected the "acceleration" limits (the kinetic part), the batteries didn't get beaten up by sudden jolts of power. They lasted longer.
Grid Stability: The "Mean-Field" part prevented the crowd from acting in a panic. The energy flow was smooth, reducing stress on the power lines.

Summary

This paper is about teaching a Conductor how to direct a massive orchestra of home batteries.

They use Physics-based math (Kinetic) to ensure the batteries aren't jerked around.
They use Crowd Psychology math (Mean-Field) to ensure the group doesn't panic.
They use AI to solve the impossible math equations in real-time.

The result is a smarter, cheaper, and safer way to manage the future of our energy grid, turning millions of home batteries into a single, powerful, money-making machine that keeps the lights on.

1. Problem Statement

The paper addresses the stochastic optimal control problem of an energy aggregator managing a large population of grid-connected residential prosumers (households with photovoltaic panels and battery storage). The aggregator operates under a contractual arrangement where it directly controls the storage assets in exchange for remuneration.

The objective is twofold:

Grid Stability: Reduce local supply-demand imbalances and smooth fluctuations.
Economic Value: Exploit intraday electricity price fluctuations through energy arbitrage (buying low, selling high), particularly when solar generation is low (e.g., winter months).

Key Modeling Challenges:

Physical Constraints: Batteries cannot change power output instantaneously; they are subject to ramp-rate limitations.
Scale and Interaction: The system involves a large population of heterogeneous devices where individual actions affect aggregate statistics (mean-field effects).
Market Dynamics: The aggregator must navigate stochastic electricity prices and net household loads.

2. Methodology: Kinetic Mean-Field Approach

The authors propose a novel kinetic mean-field formulation that bridges physical battery constraints with large-scale coordination.

A. State Space and Kinetic Analogy

Unlike standard models that treat power as the control variable, this model adopts a kinetic (Langevin-type) framework:

Position ( $X_t$ ): State of Charge (SOC) of the battery.
Velocity ( $V_t$ ): Charging/Discharging power.
Acceleration/Control ( $a_t$ ): The rate of change of power (ramp rate).

The dynamics are governed by a controlled McKean-Vlasov system:
$\begin{cases} dV_t = b(t, Z_t, [Z_t], a_t)dt + \sigma(t, Z_t, [Z_t])dW_t \\ dX_t = V_t dt \end{cases}$
where $Z_t = (V_t, X_t)$ and $[Z_t]$ denotes the time-marginal law (distribution) of the state. This formulation naturally encodes ramp-rate constraints, ensuring smooth power trajectories rather than discontinuous jumps.

B. Objective Functional

The aggregator minimizes a cost functional $J(a)$ comprising:

Grid Exchange Cost: $S_t(H_t + V_t)$ , where $S_t$ is the price and $H_t$ is the net load.
Penalties: Quadratic terms for battery degradation ( $|V_t|^2$ ) and control effort ( $a_t^2$ ).
Mean-Field Terms: Penalties for deviating from the population average (e.g., $h(X_t - \mathbb{E}[X_t])$ ), promoting collective stabilization.
Terminal Cost: Value of residual energy and end-of-horizon coordination.

C. Mathematical Framework

The problem is formulated as a Markovian stochastic control problem involving:

Controlled McKean-Vlasov FBSDEs: A fully coupled Forward-Backward Stochastic Differential Equation system where the backward component depends on the law of the forward state.
Hamilton-Jacobi-Bellman (HJB) Equation: A 5-dimensional (time + 4 state variables) ultraparabolic (kinetic) PDE. The operator is hypoelliptic due to the degenerate diffusion (noise acts only on power $V$ , not directly on SOC $X$ ).
Functional Spaces: The analysis relies on anisotropic and intrinsic Hölder spaces ( $C^\alpha_{L,T}$ ) tailored to the hypoelliptic structure of the Langevin operator, ensuring classical solutions despite the degeneracy.

3. Key Contributions

Kinetic Mean-Field Formulation:
- First application of a kinetic (acceleration-controlled) framework to residential energy storage aggregation.
- Explicitly models physical ramp-rate constraints, yielding smooth power profiles that are more realistic than standard instantaneous power control models.
Theoretical Well-Posedness:
- Proves the existence and uniqueness of a strong solution to the McKean-Vlasov optimal control problem.
- Establishes the connection between the stochastic control problem, the coupled FBSDE, and the kinetic HJB equation.
- Utilizes hypoelliptic regularity theory (specifically for ultraparabolic operators) to prove that the value function is a classical solution ( $C^{3+\alpha}$ ), allowing the use of Itô's formula and classical verification theorems.
Fixed-Point Argument for Mean-Field:
- Solves the mean-field interaction via a contraction mapping on the flow of marginal laws.
- Uses the Kantorovich-Rubinstein duality ( $W_1$ distance) rather than $W_2$ , which is crucial for obtaining linear estimates and ensuring the contraction property in the presence of the specific structure of the coefficients.
Deep Learning Numerical Solver:
- Develops a Deep BSDE solver (FBSNN) to handle the high dimensionality (5D) and the fully coupled nature of the system.
- Architecture: Uses a neural network to approximate the value function $u(t, Z_t)$ .
- Key Innovation: The martingale term $\zeta_t$ (gradient of the value function) is recovered via automatic differentiation rather than a separate network, ensuring structural consistency and stability.
- Mean-Field Handling: Approximates the law $[Z_t]$ using an empirical measure from a particle system (Monte Carlo simulation) during training.

4. Results

Theoretical Results

Theorem 2.8: Under specific Lipschitz and regularity assumptions, the control problem admits a unique optimal control characterized by a decoupling field (the solution to the HJB equation).
Regularity: The value function possesses sufficient regularity ( $C^{3+\alpha}$ in intrinsic norms) to justify the feedback control law $a^*_t = \hat{a}(t, Z_t, \nabla_v u)$ .

Numerical Experiments

Setup: Calibrated using Italian day-ahead electricity prices (PUN) and real household load/PV data (SolarEdge).
Comparison: The optimal kinetic strategy is compared against a "passive" benchmark (battery only compensates net load without strategic arbitrage).
Findings:
- Cost Reduction: The optimal strategy significantly reduces the accumulated cost compared to the passive benchmark.
- Stability: The policy yields tighter dispersion in costs across trajectories (lower variance).
- Battery Preservation: Despite lower costs, the optimal strategy systematically minimizes battery usage (reducing degradation) by only charging/discharging when price spreads justify it.
- Grid Stabilization: The mean-field penalties successfully reduce the amplitude of grid fluctuations (narrower confidence bands in net load).
- Arbitrage: The control successfully tracks price fluctuations, charging during low-price periods and discharging during peaks.

5. Significance and Impact

Bridging Physics and Economics: The paper provides a rigorous mathematical bridge between the physical constraints of battery actuation (ramp rates) and the economic incentives of large-scale energy management.
Regulatory Relevance: The model aligns with modern regulatory frameworks (e.g., FERC Order No. 841) that encourage the aggregation of behind-the-meter resources into wholesale markets.
Computational Feasibility: It demonstrates that complex, high-dimensional, hypoelliptic mean-field control problems can be solved numerically using deep learning, opening the door for real-time implementation in smart grid management.
Mathematical Novelty: It extends the theory of McKean-Vlasov control to kinetic (ultraparabolic) settings, addressing a gap in the literature regarding degenerate diffusions with mean-field interactions.

In summary, this work offers a sophisticated, mathematically rigorous, and computationally validated framework for optimizing residential energy storage at scale, balancing grid stability, economic profit, and hardware longevity.