Predictive Monitoring of SF6 Gas Leaks in GIS Substations Using Sensor Technology and Machine Learning Models

The project began with the installation of industrial sensors for measuring gas pressure, density, and temperature. The hourly data collected from these sensors enabled the development of predictive models for the early detection of incipient leaks. Two predictive models are currently being implemented at different stages of deployment.

The first model, already in production within a real‑time industrial data analytics and historian platform, is based on linear regression and calculates the density slope of each GIS compartment using a one‑month moving window. Additionally, the duration for which this slope remains negative is measured. When experimentally defined thresholds (derived from real leak events) are exceeded, a leak alert is triggered. This model has been validated by successfully detecting leaks within a timeframe of several days.

The second, more complex model employs a TFT (Temporal Fusion Transformer), a Deep

Learning architecture specifically designed for multivariate time‑series forecasting. Trained with one year of data, the model predicts future density values for each GIS compartment.

Alerts are generated when the deviation between predicted and actual values exceeds one standard deviation and persists for more than 12 hours, a statistically defined threshold established during testing across all sensors, where no instance showed > 12 hours of consecutive deviation without a subsequently confirmed leak, thereby outlasting diurnal temperature cycles and preventing false positives.

Both models aim to reduce the detection time of SF₆ leaks, which are difficult to identify using traditional densimeters due to significant temperature variations throughout the day (from –5°C to 25°C). Operational experience has shown that waiting for the activation of low‑density protection systems in GIS substations can result in the loss of approximately 10 kg of SF₆ per event.

The objective of this article is to describe the development of both models, present the implementation results, and propose recommendations and future steps.