Application of Machine Learning Models to the Development of SF6-free High Voltage Gas Insulated Switchgear

Towards this goal, this paper presents a systematic methodology utilizing data analysis and machine learning (ML) modelling techniques to capture the thermal clearing behavior of a puffer-type HVCB with a C4F7N + CO2 based gas mixture for current interruption, using High

Power Laboratory (HPL) SLF90 test data and 1D simulation results for SLF90 and SLF75 test duties. The design parameters considered are nozzle material type and geometry, pin/tulip diameter, dead volume length and overpressure (OP) valve cracking pressure. The test circuit parameters are di/dt, input current and arcing time. The test result parameters include CV pressure at peak and current zero (CZ), arc voltage peak, arc resistance (R200 and R500) and slope of arc resistance dR/dt. The iterative methodology follows a well-defined sequence of steps:

1. Data mining stage: Deploy correlation, classification and regression data analysis / small dataset ML techniques to identify and extract underlying patterns linking design parameters, test circuit parameters, and test result parameters to the HVCB thermal clearing capability using test data 2. Interrupter modelling stage: Using the measured puffer travel, input current and pressure build-up in the compression volume (CV) from the test data, calibrate and validate the 1D interrupter model parameters. Derive additional constraints on thermal clearing capability based on the simulation result parameters.

3. CB modelling stage: Generate synthetic data for SLF90 and SLF75 IEC test sequence through additional simulations using a 1D CB model (drive model + validated interrupter model). Check compatibility between the synthetic data and the test data by comparing the relative importance of each parameter in determining the CV pressure build-up.

4. Surrogate model stage: Create a large-scale synthetic dataset through exploration of the parametric space within the physics-based limits defined by the compatible 1D CB model. Use advanced deep learning models to capture the sensitivity of each variable parameter within the permissible parametric space.

The present study builds on previous work [1] to incorporate steps (1)-(3) of the methodology outlined above. The test reproductions and predictions are performed using Amesim, a commercial 1D simulation software, while the data analysis and machine learning methods are implemented using Python.

ML analysis demonstrates that design parameters can be used to predict HVCB thermal clearing performance, thereby revealing patterns that cannot be obtained from correlation methods alone. Correlation analysis indicates that arc resistance is closely linked to thermal clearing.

Simple regression models effectively predict CV pressure build-up using design parameters, arcing time, and applied current. Successful classification of thermal clearing performance through ML methods requires the incorporation of test result parameters such as arc resistance.

The results provide a guideline for HVCB initial design, and the compatible CB model can be used to train and validate a neural network-based surrogate model.