⚡ electrical engineering

Coherent Load Profile Synthesis with Conditional Diffusion for LV Distribution Network Scenario Generation

This paper proposes the use of Conditional Diffusion models to synthesize coherent daily active and reactive power load profiles for low-voltage distribution substations, addressing the need for realistic, spatially-correlated load scenarios to improve distribution network planning and congestion management.

Original authors: Alistair Brash, Junyi Lu, Bruce Stephen, Blair Brown, Robert Atkinson, Craig Michie, Fraser MacIntyre, Christos Tachtatzis

Published 2026-02-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alistair Brash, Junyi Lu, Bruce Stephen, Blair Brown, Robert Atkinson, Craig Michie, Fraser MacIntyre, Christos Tachtatzis

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 you are a city planner trying to design a new electrical grid for a massive, growing neighborhood. To do this right, you need to know exactly how much electricity everyone will use—when they’ll turn on their kettles, when their electric cars will charge, and when their heat pumps will kick in.

The problem? You can’t put a smart meter on every single house in the entire country; it’s too expensive and creates too much data to manage. So, you have to "guess" what the electricity demand looks like.

This paper describes a new, high-tech way to make those guesses using Artificial Intelligence.

The Problem: The "Average" Trap

Traditionally, engineers used "Typical Load Profiles." Think of this like planning a massive dinner party by assuming every guest will arrive at exactly 7:00 PM, eat exactly 500 calories, and leave at 9:00 PM.

In reality, people are messy. Some arrive early, some stay late, and some eat huge meals while others snack. If you plan for the "average" guest, you’ll end up with a kitchen that’s too small for the sudden rush at 7:00 PM, or a dining room that sits empty most of the night. In the power grid, if you only plan for the "average" load, you might face blackouts when everyone happens to use their appliances at once.

The Solution: The "Digital Mimic" (Diffusion Models)

The researchers used a type of AI called a Conditional Diffusion Model.

Think of this AI like a master sculptor working with a block of marble.

The Noise (The Marble): The AI starts with a block of pure "static" or "noise"—just random, chaotic data that looks like a TV with no signal.
The Sculpting (The Diffusion): The AI has been trained by looking at thousands of real-world examples of electricity use. It knows what a "real" day looks like. It slowly chips away at that random noise, refining it, shape by shape, until a beautiful, realistic "statue" of a daily electricity profile emerges.
The Instructions (The "Conditional" Part): This is the secret sauce. Instead of just making any random day, you can give the sculptor instructions: "Make a day that is freezing cold, it's a Tuesday in December, and this neighborhood has 200 houses." The AI then sculpts a profile that specifically fits those conditions.

The Three Levels of "Guessing"

The researchers tested three different versions of this "sculptor":

Level 1 (The Blind Sculptor): It just makes random, realistic-looking days. It’s okay, but not very helpful for specific planning.
Level 2 (The Weather-Aware Sculptor): You tell it the weather and the day of the week. It gets much better at matching the "vibe" of a real day.
Level 3 (The Super-Detailed Sculptor): You give it the weather plus some basic stats (like the highest and lowest power used that day). This sculptor is incredibly accurate—it can recreate the "extremes" (the massive spikes in usage) that are most dangerous for the grid.

Why does this matter? (The Stress Test)

To prove it worked, they didn't just look at the "statues" (the data); they put them to work. They took these AI-generated "fake" days and plugged them into a digital simulation of a real power network.

They wanted to see if the "fake" electricity would cause the digital wires to "overheat" or the "voltage" to drop.

The result? The most advanced AI (Level 3) was so good that the power grid behaved almost exactly as if it were running on real, measured data. It successfully predicted the "stress points" where the grid might struggle.

The Big Picture

As we move toward a world of electric cars and heat pumps, electricity use is becoming more unpredictable. This paper provides a way for utility companies to "hallucinate" incredibly realistic, high-stakes scenarios. By practicing on these AI-generated "what-if" worlds, they can build a stronger, more reliable grid for the real world.

Technical Summary: Coherent Load Profile Synthesis with Conditional Diffusion for LV Distribution Network Scenario Generation

1. Problem Statement

As power grids transition toward decarbonization, Low Voltage (LV) distribution networks face increasing complexity due to the integration of Low-Carbon Technologies (LCTs) like electric vehicles and heat pumps. A critical challenge for Distribution Network Operators (DNOs) and Distribution System Operators (DSOs) is the lack of observability at the LV level.

Traditional methods for modeling LV loads, such as Typical Load Profiles (TLP), rely on seasonal averages that fail to capture the high degree of diversity and non-stationary behavior inherent in individual substations. This lack of diversity leads to two major risks in network planning:

Underestimation of extremes: Failing to capture peak demand (thermal risks) or trough depth (voltage risks).
Lack of Coherence: Inability to model how independently behaving substations collectively impact higher-voltage (Medium Voltage) network stability.

2. Methodology

The authors propose a generative AI framework using Conditional Diffusion Models to synthesize daily active ( $P$ ) and reactive ( $Q$ ) power profiles for LV substations. Unlike previous studies focusing on smart meters, this approach targets the substation level to capture aggregated behavioral routines.

A. Model Architecture:
The study utilizes the SSSDS4 (Structured State Space Diffusion Model) architecture. This model is specifically chosen for its ability to capture long-term temporal dependencies (diurnal and seasonal trends) through structured state-space models. The authors modified the training loop to ensure conditional inputs are correctly integrated after the noise corruption process.

B. Proposed Models (Increasing Complexity):

LVGenU (Unconditional): Generates profiles from pure Gaussian noise without external guidance.
LVGenWC (Weather & Calendar): Conditions generation on weather data (temperature, humidity, etc.), calendar variables (day/month), and substation metadata (number of customers).
LVGenWCS (Weather, Calendar, & Statistics): The most advanced model, which adds daily summary statistics (minimum, mean, and maximum power) as conditional cues to guide the model toward realistic extremes.

C. Evaluation Framework:
The methodology employs a dual-layer evaluation:

Statistical Fidelity: Using Mean Squared Error (MSE), Maximum Mean Discrepancy (MMD), Wasserstein distance, and a novel Tail-Focused MSE (measuring error at the 95th percentile) to assess how well the model replicates the distribution of real data.
Power System Coherence: A "post-hoc" evaluation where synthesized loads are propagated through a UK Generic Distribution System (UKGDS) 77-bus network using Newton-Raphson power flow modeling to check if the generated loads produce realistic voltage magnitudes and phase angles.

3. Key Contributions

First Application of Diffusion to LV Substation Synthesis: Moving beyond smart-meter-level modeling to the substation level.
Joint Active and Reactive Power Modeling: Recognizing that reactive power is critical for voltage stability and is increasingly variable due to LCTs.
Multi-Tiered Conditioning Strategy: Demonstrating that while metadata and weather improve general shape, daily statistics are essential to capture the "tails" (extremes) of the distribution.
Power-System-Centric Validation: Proving that statistical similarity is insufficient; a model must also produce "coherent" results that satisfy physical network constraints (voltage/phase angle limits).

4. Results

Statistical Performance:
- LVGenU showed good temporal coherence but failed to capture the distribution's tails (high Tail-Focused MSE).
- LVGenWC improved the general distribution but still struggled with extreme values.
- LVGenWCS achieved near-perfect alignment with real data, significantly outperforming all other models in MMD and Tail-Focused MSE.
Power Flow Analysis:
- Benchmark models (GMM, WGAN, Tao Vanilla) failed to provide reliable voltage estimates, often resulting in $R^2$ values that were negative or very low.
- LVGenWCS was the only model capable of maintaining voltage magnitudes within the statutory $\pm 0.05$ p.u. limit, demonstrating high accuracy in both voltage magnitude and phase angle prediction.

5. Significance

This research provides a scalable solution for scenario generation in modernizing grids. By using conditional diffusion, DNOs can generate highly realistic, diverse, and physically plausible load scenarios even when full monitoring infrastructure is not present. This enables more accurate long-term network planning and short-term operational flexibility management, ultimately supporting the safe integration of renewable energy and electric vehicle infrastructure.