Predict the failure. Call the peak. Price the storm.

A cooperative’s year comes down to three questions. Which transformer fails next? When does the peak land? What does a storm cost the members who live through it?

Each one has a price. A distribution transformer now takes about two years to procure, at prices up as much as 95% since 2019. For some cooperatives, a year of transmission charges is set by demand in four fifteen-minute intervals, the highest of each summer month.

Most cooperatives already collect the data to answer all three, in meters, line sensors and outage logs. It sits in systems never built to talk to each other.

Grid Intelligence brings meter, line sensor and outage data together on Snowflake and turns it into decisions a team can act on.

• Every monitored transformer gets a daily score for failing within 90 days, read from loading, oil temperature, arcing and heat stress.
• Each unit on the watch list carries a plain reason. Dispatching an inspection writes the work order back to Snowflake.
• At noon the platform forecasts tonight’s peak and calls demand response only if tonight looks like one of the month’s top three.
• A storm replay compares restoration as it happened against crews staged early on the 20 riskiest feeders. A dispatch board recommends the nearest qualified crew.
• An outage cost view uses Berkeley Lab estimates in 2013 dollars. Eight hours without power costs a household about $17 and a large business about $84,000.
• Anyone can ask in plain English. Snowflake Cortex writes one query, runs it read-only inside the account and answers in two sentences.

No cooperative’s internal data went into Grid Intelligence. We generated a test fleet of 1,200 transformers to match public utility records. Every screen labels it synthetic.

We graded the failure model the way it would be used, training on one stretch of time and testing on the months after. On the test fleet it scored an AUC of 0.81. Across 31 failures in the test window, the 50-unit watch list caught 58% and the riskiest tenth of units held 6.5 times their share.

The load forecast had to beat two simple baselines to stay. Over seven held-out months on the test fleet it missed by 8.8%, against 11.8% for repeating yesterday’s load and 16.3% for last week’s. In the late-afternoon hours that set the bill it held 8.5% against 10.8%. Replayed out of sample across one test-fleet summer, the noon rule called 21 events and caught all four monthly peaks.

• Every panel prints the Snowflake query ID behind it, so anyone can check exactly what ran.
• Run natively in Snowflake ML FORECAST, the same forecast missed by 6.0% against 10.0% for persistence on a two-week test-fleet holdout.

• No cooperative runs Grid Intelligence today. Every model result comes from the test fleet and must hold up again on real data.
• The failure model predicts wear-out failures. It does not pretend to know which units a storm will take down.
• Both forecast tests used actual weather in place of a weather forecast, which a production feed would supply.
• Regulators are reviewing how these summer peaks are charged. The lasting value is the forecast and the dispatch, not any one tariff.

• A transformer failure model with time-split validation and a watch list sized to inspections
• Load forecasting and a noon peak call, tested out of sample against simple baselines
• Snowflake ML FORECAST and ANOMALY_DETECTION running natively in the account
• A storm replay, a crew dispatch board and a member outage cost calculator
• Plain-English questions through Snowflake Cortex behind read-only guardrails
• A control room with the query ID on every panel, deployable as one container

Technologies

• Snowflake
• Snowflake ML FORECAST
• Snowflake ML ANOMALY_DETECTION
• Snowflake Cortex
• scikit-learn
• FastAPI
• React

• AI & Machine Learning
• Data & Analytics
• Generative AI