A grid operator cuts unplanned outages by 27% with predictive maintenance

The challenge

The operator runs a network of substations, transformers and overhead lines serving more than eight million customers. Maintenance followed fixed schedules set decades earlier: every transformer inspected on the same cycle, regardless of age, load or condition. Crews spent much of their time on healthy equipment while failures on older assets caused most of the outages.

The regulator had set tighter reliability targets with financial penalties attached. At the same time, the operator's capital plan for grid expansion meant the maintenance budget could not grow. Previous analytics pilots had produced accurate models that nobody used.

Our approach

We began with the decision, not the data. Planners and crew leaders told us what they needed: a weekly list of the assets most likely to fail, with a reason they could check in the field. We built toward that.

The team combined twelve years of maintenance records, sensor readings from about 3,000 monitored transformers, load data and weather history to build failure-risk models for the six asset classes behind 80 percent of outages. Our engineers worked inside the operator's data team, so the models ran on its own platform from the start.

The harder work was changing how maintenance was planned:

• Weekly planning moved from fixed calendars to a risk-ranked list, reviewed by planners and crew leaders together

• Each prediction came with the three factors driving it, so crews could confirm or overrule it on site

• Overrules were logged and fed back into the models every month

• Inspection cycles for low-risk assets were lengthened, with the regulator's agreement

We piloted the approach in two regions for twelve weeks, compared results against two control regions, and rolled it out nationally once the difference was clear.

The impact

In the first full year, unplanned outages fell by 27 percent and the operator met its regulatory reliability target for the first time in four years. Maintenance spend fell by 18 percent as crews stopped servicing equipment that did not need it, saving about $23 million a year.

The models now score 11,000 assets daily. Crew overrules have dropped from one in four predictions to fewer than one in ten, a sign that the field teams trust the list. The operator's own data team now maintains and extends the models, and has added two further asset classes without outside help.