French utility EDF has built an automated watchdog that flags its own electricity production forecasts when something looks off.
Researchers describe TAMIS, a system that analyzes EDF's daily production time series and looks for atypical intra-day patterns compared to historical data. The system is built for human-in-the-loop review: instead of silently flagging problems, it surfaces its top-ranked anomalous days through an automated daily newsletter that staff can check. In tests on real industrial data, the researchers report TAMIS delivered the best accuracy-to-efficiency trade-off among the methods they compared it against. The team also released anonymized versions of the datasets used in the study so other researchers can reproduce the results.
Utilities like EDF generate new production forecasts daily to keep electricity supply and demand in balance, and an undetected anomaly, whether a data quality issue or an operational irregularity, can quietly corrupt decisions downstream. The research is a useful counterpoint to most AI headlines: it is not about a bigger model doing something flashy, but about making anomaly detection cheap and interpretable enough that a human will actually trust its output and act on it.
That distinction, between a system built to impress a benchmark and one built to survive a Tuesday at a power utility, is exactly the gap most production AI never clears.