Trained On LEAD
An Isolation Forest Learned Normal Consumption From LEAD Hourly Meter Records Using Log Energy, Time Of Day, And Each Building's Own Past-Only 168-Hour Deviation. Training Used Normal Rows Only, So The Model Knows Ordinary Behavior, Not Labeled Faults.
Performance Was Measured On 40 Held-Out Buildings Kept Strictly Disjoint From Training, Reaching 0.79 ROC-AUC. The Resulting Model Was Then Frozen And Never Retrained.
test.csv As Validation
The Kaggle Test Set Carries No Anomaly Labels, So Classical Accuracy Could Not Apply. Instead It Served As An Unlabeled Generalization Check On 206 Completely Unseen Buildings.
The Frozen Model, With Its Fixed Threshold, Flagged About One Percent Of Test Hours — Matching Its Calibration And Confirming The Detector Transfers Beyond The Buildings It Trained On.
Evaluated On IITD
Delhi Campus Minute-Level Mains Were Resampled To Hourly Energy, Low-Coverage Hours Masked, And Long Gaps Left Unfilled. Each Building Was Scored Against Its Own History Only.
Because Campuses Differ, Thresholds Were Recalibrated Per Building Instead Of Reusing The LEAD Cutoff. With No Ground-Truth Labels, Flags Are Anomaly-Health Indicators, Never Confirmed Faults.