💧 Water Consumption Anomaly Detector 💧

The system analyzes history notes for each account to identify meters that previously had leaks or dead batteries. It examines cases where new consumption values were significantly higher or lower than normal for the account, indicating either a water leak (unusually high consumption) or a dead meter (unusually low or zero consumption). This historical data provides the foundation for supervised learning.

Previous confirmed leaks and meter failures are used to train a supervised learning model using scikit-learn. The model learns patterns from historical data including consumption patterns, meter age, property characteristics, and seasonal variations to predict potential leaks and dead meters.

The system employs Isolation Forest algorithms from scikit-learn's ensemble module to perform unsupervised anomaly detection. This method isolates data points that are statistically different from the majority, identifying unusual consumption patterns that may indicate leaks or meter malfunctions without requiring labeled training data.

Results from both supervised and unsupervised learning algorithms are combined and filtered using preset parameters. Additional filtering is applied based on account details such as meter age, property ownership changes, seasonal patterns, and neighborhood consumption averages to reduce false positives and improve detection accuracy.

The system creates a comprehensive Excel output containing flagged accounts identified by both collective algorithms (supervised + unsupervised) as having potential leaks or dead meters. The report includes account details, consumption patterns, confidence scores, and recommended actions, enabling city staff to efficiently identify and contact affected property owners.