GridForge: An Open Physics-Informed Digital Twin and Benchmark Suite for UK 11 kV Underground Distribution Cables

GridForge is an open‑source, physics‑informed digital twin and benchmark suite that provides UK 11 kV underground cable data, a calibrated thermal model, and a PINN surrogate to standardise fault‑diagnosis research.

GridForge: An Open Physics-Informed Digital Twin and Benchmark Suite for UK 11 kV Underground Distribution Cables

Introduction

The UK’s underground 11 kV distribution network generates massive streams of telemetry, yet practitioners lack a shared dataset, reference model, and evaluation framework for data‑driven decision support. GridForge fills this void by delivering an open‑source digital twin that couples a rigorously verified thermal model with a synthetic benchmark dataset and a physics‑informed neural network (PINN) surrogate. The suite is designed to accelerate research on fault diagnosis, condition monitoring, and maintenance planning while ensuring reproducibility across Distribution Network Operators (DNOs).

A physics‑based thermal core

GridForge’s foundation is a steady‑state and transient thermal model derived from IEC 60287 (conductor heating) and IEC 60853 (thermal ageing). The model computes cable temperature profiles from electrical loading, soil properties, and burial conditions, and it has been validated against analytical solutions and industry‑standard calculators. By exposing all numerical parameters—conductor resistivity, soil thermal resistivity, heat‑transfer coefficients—the model serves both as a reference oracle and as a transparent teaching tool for engineers.

Synthetic benchmark dataset

To provide a common playground, the authors generated a calibrated synthetic dataset covering 64 distinct cables and 320 cable‑years of operation. The dataset spans four canonical UK thermal archetypes (e.g., urban, rural, high‑load, low‑load) and four condition modes (healthy, degraded insulation, moisture ingress, mechanical damage). Each record contains hourly loading, ambient temperature, soil moisture, and the resulting cable temperature computed by the IEC‑based model. Because the data are fully synthetic yet grounded in public standards, they can be freely redistributed under an MIT licence, enabling anyone to benchmark algorithms on identical inputs and ground‑truth outputs.

Physics‑informed neural network surrogate

Training a surrogate that respects the underlying physics reduces data‑efficiency requirements and improves interpretability. GridForge includes a PINN that incorporates the algebraic balance from IEC 60287 directly into its loss function. When evaluated on an independent held‑out set, the PINN achieves an RMSE of 0.148 °C, closely matching the analytical oracle while training in roughly two minutes on a modest GPU. This rapid convergence makes the surrogate practical for real‑time virtual‑sensor applications, where computational budgets are tight.

Five‑task benchmark suite

The authors propose five benchmark tasks that reflect common operational challenges:

  1. Steady‑state temperature prediction – estimate cable temperature from loading and environmental variables.
  2. Transient response forecasting – predict temperature evolution after a load step.
  3. Fault detection – classify abnormal temperature excursions indicative of insulation failure.
  4. Virtual‑sensor reconstruction (the only task where baselines diverge) – infer temperature at uninstrumented locations using surrounding sensor data.
  5. Maintenance prioritisation – rank cables for inspection based on predicted degradation risk.

For each task, three baseline models are provided: a pure physics solver, a gradient‑boosted tree trained on the synthetic data, and the PINN surrogate. The current benchmark split shows the virtual‑sensor task (T4) as the most discriminative, highlighting an opportunity for future work to enrich the other tasks with task‑specific logic.

Practical implications for the industry

GridForge offers DNOs a reproducible workflow: ingest raw telemetry, run the IEC‑based model to generate ground truth, train or fine‑tune a PINN on site‑specific data, and evaluate performance against the benchmark suite. Because the code and data are openly licensed, utilities can audit the methodology, adapt the thermal parameters to local soil conditions, and share results without legal barriers. Moreover, the suite encourages collaborative improvement—researchers can contribute new baselines, augment the dataset with real‑world measurements, or extend the benchmark to higher voltage levels.

Conclusion

By uniting a standards‑compliant thermal model, a richly annotated synthetic dataset, and a physics‑informed surrogate, GridForge establishes a common reference point for UK underground cable analytics. Its open‑source nature and lightweight training requirements lower the entry barrier for both academic and industry teams, fostering a more data‑driven, transparent approach to cable health management. As the benchmark suite evolves, it promises to become a cornerstone for evaluating next‑generation fault‑diagnosis algorithms across the power‑distribution sector.

Source

Read original source

Why we picked this

Open-source physics-informed digital twin and benchmark suite, relevant AI infrastructure.

← Back to all articles