The ability to monitor dissolved gases in transformers in near real time has been around for years and is proving its worth. Utilities not continuously monitoring their critical transformers — where, in this case, continuously means several times a day — may be reluctant to do so now for two reasons: cost and confidence, with perhaps the bigger reason being confidence. They are not yet convinced continuous monitoring yields significantly more relevant data than yearly or biannual laboratory analysis. This belief is motivated by the misconception that transformer failures are the result of a gradual deterioration taking place over several years. Arizona Public Service (APS) had that perception until it began collecting data on transformer failures within its fleet.

APS described its initial experience in the feature article “Data Mining Pinpoints Incipient Failures” (Transmission & Distribution World, July 2009). Since that article was published, on-line monitoring of the utility's transformers has resulted in saving two more transformers from imminent catastrophic events. APS now has approximately 150 of its most critical transformers being checked for failure conditions every four hours.

Catastrophic failures are generally caused by the buildup of volatile gases due to internal arcing. Failures also can be caused by overheating of the transformer's cooling oil or the failure of cellulose insulation that wraps the transformer coils. As APS has found, these conditions can develop, subside and then occur again in a matter of hours. Since there is no way an individual can monitor 150 transformers continuously, APS has come to rely on the use of automated monitoring equipment and software systems.

The utility's commitment to developing a reliable means of monitoring its transformer fleet began in earnest after it had one of the largest transformer fires in the United States back in 2004. One large transformer failed and caught four others on fire. To provide advanced warning and perhaps prevent future catastrophic events, APS developed the transformer oil analysis and notification (TOAN) system.

Mining the Data

When the level of dissolved gases in transformer oil is monitored, most of the time there is no significant change between readings. The TOAN system automatically weeds out the mundane day-to-day information from APS's fleet of on-line transformer monitors using an exception-based process in combination with artificial neural networks and fuzzy logic that analyzes the data looking for specific signatures that could indicate a problem is developing. The analysis uses multiple artificial neural networks to classify the sample results in four categories: overheating oil, cellulose degradation, low energy discharge and high energy discharge.

Once classified, the fuzzy logic system determines the severity level based on the level and rate of change of key gases within each of the four fault types. If the combination of fault types and severity levels are unchanged from previous samples, then nothing is sent. This encompasses greater than 99.5% of all samples taken. When there is a change in behavior, an e-mail is sent to maintenance personnel advising them of the problem. Using this software, APS has been able to prevent two catastrophic transformer failures from occurring during the most recent 12-month period.

TOAN determines the most recent gas rates using a piecewise linear regression technique — fitting multiple straight lines to the data — and immediately detects and determines the slope of the line when gassing behavior changes. A rapidly upward changing slope signifies that conditions within the transformer are becoming more volatile. It has been APS's experience that slow changes in dissolved gas values can mean trouble, but catastrophic events typically happen after periods of fluctuations that look like zigzags on the charts. TOAN is on the lookout for these conditions.

Looking for Zigzags

For example, prior to the inclusion of the piecewise linear regression algorithm in TOAN, APS had a large transformer failure in 2007 that was observed by an on-line dissolved gas monitor (Fig. 1.). When the data from the transformer monitor was analyzed, the utility saw the gassing went from 0 ppm of acetylene to 9 ppm within the course of four successive four-hour samples. Then the gassing stopped. A week later, it started again just before the catastrophic event. If APS had been using the current TOAN algorithm, the zigzagging signature of intermittent arcing within the transformer would have enabled the utility to predict the fast-acting fault that caused the failure. Conventional manual sampling and analysis of the transformer oil once or twice per year, or even continuous sampling without APS's new data processing methods, would not have enabled the utility to predict the failure when it occurred.

Using the TOAN software, APS has had four transformer saves over the past three years. Three of those saves followed detection of the zigzagging situations, after which APS was able to successfully take the transformers off-line. A fourth transformer had an overheating problem the utility was able to detect and successfully fix through near-real-time monitoring, preventing a failure. The most recent saves, involving two 525/345-kV units, happened in October 2009 (Fig. 2) and May 2010. One of the transformers was about 40 years old and the other four months old, suggesting that utilities should not just monitor their oldest transformers. In both cases, acetylene and hydrogen were shown to spike quickly.