⚡ electrical engineering

Demand Response Under Stochastic, Price-Dependent User Behavior

This paper proposes and validates a stochastic, feedback-based pricing strategy for residential demand response that addresses the inherent uncertainty and estimation errors in customer behavior by modeling responses as price-dependent random variables within a decision-dependent distribution framework.

Original authors: Guido Cavraro, Andrey Bernstein, Emiliano Dall'Anese

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guido Cavraro, Andrey Bernstein, Emiliano Dall'Anese

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

Imagine the electrical grid as a massive, bustling highway system. The cars are the electricity being used by homes, and the traffic cops are the utility companies trying to keep traffic flowing smoothly.

Sometimes, too many cars try to enter the highway at once (a "peak load"), causing a traffic jam that could crash the whole system. To prevent this, the traffic cops usually try to send a message to the drivers: "Hey, if you wait 15 minutes to leave, we'll give you a discount on your toll!" This is called Demand Response.

However, there's a big problem: No one knows exactly how drivers will react.

Some drivers are super sensitive to price and will wait immediately.
Some don't care at all.
Some might be distracted, tired, or just in a bad mood.
Some might have a sudden emergency and have to leave now, regardless of the price.

If the traffic cop guesses the wrong discount, they might send too few cars (wasting money) or not enough (causing a crash). Traditional methods try to guess the "perfect" discount based on a rigid math model, assuming everyone is a robot that reacts perfectly. But humans aren't robots.

The Paper's Big Idea: "The Feedback Loop"

This paper proposes a smarter way to manage the traffic. Instead of guessing the perfect price once and sticking to it, the utility company should use a stochastic feedback loop.

Here is how the paper breaks it down using simple analogies:

1. The "Guessing Game" (The Old Way)

Imagine a chef trying to bake a cake for 1,000 people. The chef assumes everyone eats exactly 2 slices. So, the chef bakes exactly 2,000 slices.

The Risk: If the guests are actually hungry and eat 3 slices, the chef runs out of cake (grid crash). If they are full and eat 1 slice, the chef wastes food (wasted money).
The Paper's Critique: The old way assumes we know exactly how "hungry" (price-sensitive) every customer is. But we don't. We only have a vague idea.

2. The "Stochastic" Model (The New Way)

The authors say, "Let's stop pretending we know the future." Instead, they model customer behavior as random variables.

The Analogy: Instead of assuming everyone eats 2 slices, the chef assumes: "Most people will eat 2, but some might eat 1, and some might eat 3, and the exact number depends on how much the cake costs."
The Math: They use a fancy type of math called "stochastic optimization with decision-dependent distributions." In plain English: They acknowledge that the uncertainty of the customers' reaction changes depending on the price they set.

3. The "Feedback Loop" (The Solution)

Since we can't predict the future perfectly, the paper suggests a real-time feedback system.

The Analogy: Imagine the chef puts out a small batch of cake first.
1. Step 1: Set a price (toll).
2. Step 2: Watch what happens. Did the traffic slow down? Did the power usage drop?
3. Step 3: If the traffic is still too heavy, adjust the price slightly higher. If it's too light, adjust it lower.
4. Step 4: Repeat this cycle rapidly.

This is the Stochastic Inexact Performative Optimum (Stochastic InPO) algorithm mentioned in the paper. It's like a thermostat that doesn't just guess the temperature; it constantly checks the room and tweaks the heat until it hits the perfect spot, even if the window is open or the sun is shining.

Why is this better?

The paper compares their new method against three other "old school" strategies:

The "Perfect Oracle" (PO): This assumes the chef knows exactly how many slices everyone will eat. It works great in theory but is impossible in real life because we can't read minds.
The "Static Guess" (InPO): This uses a rough guess of how people react. It's better than nothing, but if the guess is wrong, the system fails.
The "Ignorant Driver" (PS): This strategy sets a price and ignores whether the traffic actually slowed down. It often fails to solve the problem.

The Winner: The Stochastic InPO (the feedback loop).

Robustness: Even if the chef's guess about how "hungry" people are is wrong, the feedback loop corrects the mistake in real-time.
Stability: It guarantees that the system won't crash, even with uncertain customers.
Near-Optimal: It gets very close to the perfect price without needing to know the future.

The Takeaway

The paper argues that in a world full of unpredictable human behavior, rigid plans fail, but flexible feedback succeeds.

By treating customer behavior as a "cloud of possibilities" rather than a single fixed number, and by constantly listening to the grid's reaction (the feedback), utility companies can keep the lights on and the bills low, even when they don't know exactly what their customers are thinking. It turns a chaotic guessing game into a smooth, self-correcting dance.

1. Problem Statement

The paper addresses the challenge of implementing Price-Based Residential Demand Response (DR) in distribution networks. While DR is a key tool for system operators (SOs) to manage peak loads and grid stability, existing strategies often fail due to two primary limitations:

Epistemic Uncertainty: Accurately modeling customer price elasticity (responsiveness to price changes) is difficult. Behavioral responses are small, confounded by external factors (weather, occupancy), and data is often limited.
Limitations of Deterministic/Open-Loop Models: Traditional approaches assume a deterministic relationship between price and consumption or add fixed noise. These "open-loop" strategies rely on perfect knowledge of user sensitivity parameters ( $\beta$ ). When these estimates are inaccurate, open-loop strategies fail to meet operational constraints (e.g., failing to reduce load sufficiently) or cause system instability.

The core problem is to design a pricing strategy that can steer aggregate power demand to a target reference ( $p_{ref}$ ) despite:

Stochastic User Behavior: User response is modeled as a random variable whose distribution depends on the price signal (decision-dependent distribution).
Unknown Parameters: The exact price sensitivity parameters ( $\beta$ ) are unknown or difficult to estimate accurately.
Real-time Constraints: The SO lacks real-time, node-level load data and relies on aggregate measurements.

2. Methodology

A. System and User Behavior Modeling

Grid Model: The authors model a distribution feeder with $N$ nodes. The total power $p_0$ is a linear function of flexible loads ( $w$ ) and non-dispatchable loads ( $\hat{d}$ ).
Stochastic User Response: Instead of a fixed function, user load deviation $w_n$ is modeled as a random variable $w_n \sim \mathcal{D}_n(x_n)$ , where the distribution depends on the price adjustment $x_n$ .
Linearized Mean Response: The paper assumes the mean response $\mu_n(x_n)$ is linearly decreasing with price for small adjustments:
$\mu_n(x_n) = \delta_n - \beta_n x_n$
where $\beta_n$ is the unknown price sensitivity.

B. Optimization Formulation

The objective is to find a price adjustment vector $x$ that minimizes a cost function (generation cost minus user payments plus a penalty for price deviation) while satisfying a chance constraint on the total power:
$\min_{x} \mathbb{E}_{w \sim \mathcal{D}(x)} [c_0(x, w)] + \frac{\kappa}{2}\|x\|^2$
$\text{s.t. } \mathbb{E}_{w \sim \mathcal{D}(x)} [\xi(w)] \leq 0$
where $\xi(w)$ represents the violation of the power target.

C. Proposed Solution: Stochastic Inexact Performative Optimum (Stochastic InPO)

Recognizing that solving the optimization problem directly is infeasible without perfect knowledge of $\beta$ and full load observability, the authors propose a feedback-based online algorithm.

Algorithm Structure: It is a primal-dual stochastic gradient descent algorithm.
- Primal Update (Price): Updates the price vector $x(t)$ using an estimated sensitivity matrix $\hat{B}$ and a gradient of a regularized Lagrangian.
- Dual Update (Lagrange Multiplier): Updates the multiplier $\lambda(t)$ based on the real-time measurement of the aggregate power error $(p_0(t) - p_{ref})$ , rather than a model-based expectation.
Key Innovation: The algorithm decouples the need for perfect model knowledge from constraint satisfaction. Even if the estimated sensitivity $\hat{\beta}$ is inaccurate, the feedback loop (using actual $p_0(t)$ measurements) corrects the trajectory to satisfy the power constraint.

D. Theoretical Analysis

The paper establishes rigorous convergence guarantees:

Static Case: If non-flexible loads are constant, the algorithm converges to a neighborhood of the optimal solution. The error bound depends on the step size, the mismatch between true and estimated sensitivities ( $\|\hat{\beta} - \beta\|$ ), and the stochastic noise in measurements.
Dynamic Case: If non-flexible loads vary over time, the algorithm tracks the time-varying optimal trajectory. The tracking error is bounded by the maximum variability of the loads and the estimation errors.

3. Key Contributions

Decision-Dependent Stochastic Framework: The paper moves beyond fixed-noise models by adopting a framework where the distribution of user response is explicitly dependent on the decision variable (price), utilizing tools from stochastic optimization with decision-dependent distributions.
Feedback-Based Pricing Strategy: It proposes a novel "Stochastic InPO" algorithm that combines performative optimization with real-time feedback. This allows the system to compensate for estimation errors in user elasticity without requiring pervasive smart metering at every node.
Theoretical Guarantees: The authors prove that the proposed algorithm is stable and achieves near-optimality, providing explicit bounds on the convergence error relative to model inaccuracies and stochastic noise.
Robustness to Uncertainty: The work demonstrates that feedback mechanisms can ensure constraint satisfaction (meeting load reduction targets) even when the underlying behavioral models are imperfect.

4. Results

The authors validated their approach using numerical simulations on an IEEE 37-bus feeder with two scenarios:

Constant Load Scenario:
- Compared against Open-Loop strategies (Performative Optimum with perfect info, Inexact PO with estimates, and Performative Stable).
- Finding: Open-loop strategies (InPO, PS) failed to meet the power constraint when sensitivity estimates were inaccurate. The proposed Stochastic InPO successfully regulated power close to the target $p_{ref}$ regardless of the estimation error, though operational cost efficiency slightly decreased with larger errors.
Time-Varying Load Scenario:
- Used real-world minute-level data (Pecan Street project) to simulate dynamic non-flexible loads.
- Finding: The Stochastic InPO maintained robust constraint satisfaction ( $p_0 \approx p_{ref}$ ) while other methods either violated the constraint (PS) or over-satisfied it leading to high costs (InPO). The PO (perfect info) was optimal but impractical; Stochastic InPO offered the best trade-off between robustness and cost.

5. Significance

This paper is significant for the transition toward grid-interactive energy systems.

Practicality: It addresses the "data gap" in DR programs. Utilities often lack granular data or accurate behavioral models; this method allows them to implement effective DR using only aggregate measurements and rough estimates.
Reliability: By integrating feedback, the method ensures that critical grid constraints (like peak load reduction) are met even under uncertainty, preventing the "adverse effects" (like load synchronization) that can arise from poorly tuned open-loop pricing.
Future Direction: It establishes a pathway for using stochastic optimization and feedback control to manage the inherent variability of human behavior in the energy sector, moving away from rigid, model-dependent control toward adaptive, data-driven strategies.