Data Science @ Statnett

Gas Insulated Switchgear (GIS) is a type of high-voltage equipment where key electrical components are enclosed in sealed compartments filled with insulating gas. The insulating gas is typically sulfur hexafluoride (SF₆), a very potent greenhouse gas. Monitoring the gas compartments is therefore essential, not only to ensure correct operation of critical components, but also to address a significant environmental concern.

The compact design of GIS significantly reduces the land area required for substations, making it particularly suitable for locations where space is limited. The SF₆-gas plays a critical role in ensuring safe and reliable operation due to its excellent dielectric strength. At the same time, 1 kg of SF₆ has a global warming impact equivalent to roughly 23,500 kg of CO₂, making GIS-leakage detection essential to support emission reduction targets and regulatory compliance.  

Traditionally, GIS-chambers have been monitored through monthly manual inspections at site, sometimes requiring technicians to climb to reach the manometer (which measures pressure). Today, digital solutions change that. By using sensors and centralized, digital dashboards, operators can continuously monitor gas pressure in near real time and follow trends. This makes it possible to detect issues earlier, reduce the need for field visits, lower HSE risk, and move towards more proactive, condition-based maintenance of critical infrastructure.  

In this blog post, we will demonstrate how we exploit real time pressure data to determine leakages, which we define as a constant decline in the gas pressure.  The task might seem straightforward – we fit a straight line to the data and interpret a negative slope as a leak rate. In practice, this is too fragile. A single erroneous measurement can dominate the fit, and the pressure levels in GIS chambers may “randomly” be adjusted by technicians (see Figure 1). Furthermore, the manometers produce noise. The noise levels have to be taken into account when evaluating the authenticity of the detected leak. 

The purpose of the detection algorithm is therefore not only to fit a line. It is to create a robust and operationally meaningful subset of the data; remove values that are physically impossible, detect changes in the pressure level, ignore data before the most recent pressure level shift, remove remaining outliers, and only then estimate the pressure trend. 

We explain the steps of the leakage-detection-algorithm through a fictious example of raw data, shown in Figure 2.

The pressure time series in Figure 2 is contaminated by a couple of extreme data points.  The 100 bar point is not realistic since the operating pressures of GIS-chambers are usually between 4 -7 bar. Any value below 1 bar is not realistic either since the chambers are surrounded by air (air pressure is 1 bar).  As a first data processing step, we safely rule out all pressure values outside the range of 1-10 bar. This leaves us with Figure 3. 

Next, we pinpoint the time of the pressure level shift. This allows us to apply a linear fit to the data after the last shift only. The gas level shifts are characterized by a sudden increase in the average value of the pressure. To differentiate a sudden increase in the average from a sudden increase from a single extreme data point, we calculate the rolling median by dividing the time series into segments and calculate the median pressure in each segment. The median is less sensitive to outliers than the mean and gives a more stable description of the pressure evolution (see Figure 4). 

The difference between neighboring data points in the rolling median reveals the level shift as a prominent spike, as seen in Figure 5. This spike is localized using the Z-score outlier detection method. 

A data point’s Z-score reflects its rarity in its dataset (here, the rolling median of the pressure time series). The formula is 

where x_i  is the data point, μ is the dataset’s average value, and σ is the dataset’s standard deviation. As a result, the timing of the level shift in the gas pressure time series can be estimated by inspecting the time of the data points in Figure 5 that has a Z-score above a certain threshold. The data point responsible for the prominent peak in Figure 5 has a very high Z-score compared to all other data points.

After detecting the shift and discarding all data points before it, we use the Z-score method again to remove outliers in the remaining time series. In Figure 6, the final processed dataset is highlighted in orange with its linear fit and extrapolation as a green dotted line. The negative slope indicates a gas leakage, and with the extrapolation we can estimate the time until the pressure level reaches a critical level. In Figure 6, the horizontal red dotted line represents such a hypothetical critical level, and the vertical red dotted line represents a hypothetical future deadline. When the extrapolation crosses the critical level before the deadline, the algorithm will throw a pro-active alarm such as “There is a leak and it should be fixed to avoid critical pressure levels!”. 

Assuming the SF₆-gas behaves as an ideal gas, we can also use the pressure leak rate to estimate the mass leakage. In an ideal gas, where the temperature of the gas and the volume of the chamber stay constant, the relative fall in pressure over time is equal to the relative fall in mass over time  The mass of leaked SF₆-gas is useful for evaluating the environmental impact of GIS-leakages.

However, before throwing an alarm, we must make sure that the linear fit is good. The time series may be non-linear, and it may be noisy. We address this issue by calculating the Pearson correlation coefficient of the data. A coefficient of -1 means the data has a perfect negative linear relationship. 0 means no linear relationship. In Figure 7 there are some examples of data and their corresponding Pearson correlation coefficient, along with a linear fit. In the GIS-leakage detection algorithm, we enforce a Pearson correlation coefficient threshold that needs to be met for the algorithm to signal a leakage. The threshold increases with the length of the time series we analyze because the more data, the better signal-to-noise ratio, and we can expect a stronger linear correlation. With a high coefficient threshold, we can be more confident that small leakage rates are real and not just noise or other systematic effects like daily or seasonal changes. 

For the GIS-chambers, we run the leakage analysis algorithm on two different time-series lengths. One spanning the last 30 days and one spanning the last year. A year of data enables us to detect small leakages while the 30 day-analysis detects recent leakages. Figure 8 shows an example where the algorithm has successfully detected a level shift and fitted the line therafter. This is an example were the data-driven detection allows us to detect leakages in time before critical pressure levels are reached, and allows us to track the climate impact of SF₆-gas leakages.

At Statnett, we still have a way to go to digitalize all our GIS stations, with around 30% of the fleet currently covered by digital monitoring. As adoption increases, the value of continuous data and centralized insights becomes even more evident, with a potential that extends beyond GIS. In all of Statnett’s air-insulated systems, the circuit breakers are filled with SF₆, representing a significant opportunity to scale digital monitoring solutions more broadly. By detecting leaks early, operators can intervene before gas levels become critical, refill compartments when needed, and locate and repair leaking components. Monitoring data may also provide valuable insight into equipment condition, supporting more proactive maintenance. Expanding digital monitoring coverage across the fleet will therefore help reduce SF₆ emissions while improving asset reliability and maintenance planning.