In the utilities sector, data is everywhere. Every meter, every reading, and timestamp tells a story. But without the right tools in place, most of it gets lost, and the story goes untold. The challenge for most utility companies isn’t a lack of data; it’s the ability to make sense of it at scale. With meters generating consumption data continuously across an entire network, manually identifying irregularities is not only time-consuming, but it’s also virtually impossible. That’s where anomaly detection comes in.
The Gap in Current Operations
Most utility companies have made significant investments in data infrastructure and reporting. Consumption data is being collected, dashboards exist, and teams are actively monitoring their networks. But there’s a difference between having visibility into data and having a systematic way to detect when something is wrong.
Without a defined algorithm or detection logic in place, identifying anomalies still relies heavily on manual review or customer-reported issues. The data is there, but the mechanism to automatically surface irregularities at scale (usually) isn’t. That gap is where problems slip through.
What Anomaly Detection Actually Looks Like
In a recent engagement with one of our North America-based utility clients, we were tasked with building a solution to identify anomalous consumption patterns across a utility client's meter network. Using SQL logic within Snowflake, we developed detection logic targeting three specific categories, each with its own methodology: zero consumption and low/high usage.
1. Zero Consumption
Zero consumption refers to a premise (a physical property or service address) recording zero water usage over a defined period. For a utility, this on an active premise is a red flag. It can indicate a faulty or stuck meter that has stopped recording, a potential bypass where water is being used but not measured, or in some cases unauthorized usage. However, not all zero consumption is created equally. A premise with no water usage, but normal electricity consumption tells a very different story than one that has zero water usage but has been listed for sale or rent.
To distinguish between legitimate and non-legitimate zero consumption, we built a scoring model in SQL that assigns each premise a weighted score based on a set of configurable scenarios. For example, a premise showing normal electricity use alongside zero water consumption might contribute 100 points toward the overall model score, while a known vacancy indicator like a for-sale or for-rent listing would subtract 100 points. The cumulative score determines whether the zero consumption is worth investigating, giving field teams a prioritized, context-aware signal rather than a raw list of zero readings.
To ensure accurate matching between the utility data and rental listings, we integrated property listing data from the RentCast API to identify premises currently listed for sale or rent. Because address formats between the utility data and RentCast listings were inconsistent, we used regex to standardize and align addresses across both datasets. This enrichment step was critical to the accuracy of the scoring model. Without it, legitimate zero consumption cases would have been incorrectly flagged, sending field teams to premises that didn’t warrant investigation.
2 & 3. Low and High Usage
Abnormal water consumption patterns can be early indicators of serious issues that would otherwise go unnoticed. Unusually low usage may point to a partial meter blockage, malfunction, or bypass, while sustained high usage could signal an undetected leak or irrigation system malfunction. For low and high usage detection, we calculated a rolling median in SQL across the previous 120 days for each meter and compared current consumption against that baseline.
Rather than applying fixed thresholds, the solution was designed to be flexible. Users could adjust the number of consecutive days to evaluate as well as the percentage increase or decrease in consumption that would update the set of flagged meters. This made the detection logic adaptable to the client’s operational needs. The consecutive days parameter was particularly important for ruling out legitimate short-term drops in consumption, such as a customer going on vacation, a holiday period, or a temporary absence. By requiring anomalous readings to persist over a defined number of days, the solution avoided flagging one-off scenarios that would generate false positives and unnecessary field visits.
It’s worth noting that meter data is rarely clean out of the box. In this engagement, we encountered large negative consumption readings, physically impossible values that are a clear indicator of data integrity issues. Rather than applying explicit exclusion logic, the use of a rolling median as the baseline provided a natural buffer. Medians are inherently more robust to outliers than averages, meaning a handful of erroneous readings wouldn’t skew the detection results.
Using Visualization as a Discovery Tool
One of the most valuable parts of this process was building an exploratory Tableau dashboard early on. Not as a final deliverable, but as a tool to drive conversation.
The choice of chart types was deliberate. To visualize low and high consumption patterns, a barbell chart was selected to showcase the difference between the lowest or peak consumption within the evaluation period and the median baseline for each meter. The barbell format makes this spread immediately readable at a glance; you can quickly see which meters have the widest gap between their normal usage and their extreme values, making outliers visually obvious without having to dig into the underlying data.
From there, the dashboard was designed to let stakeholders drill deeper. Selecting a meter from the barbell chart would filter a trend chart showing that meter’s consumption over time, allowing users to move from spotting an anomaly to understanding its pattern in a single interaction. Was the spike a one-off or has usage been climbing steadily? Has the meter been flat for weeks, or did it drop suddenly? The trend chart answered those questions instantly.
The goal wasn’t a polished, production-ready dashboard. It was to spark discussion, get stakeholders engaged, validate that the analysis was heading in the right direction, and surface questions that wouldn’t have come up from looking at raw data alone.
How You Can Get Started with Utility Anomaly Detection
Many utilities already have the data needed to begin anomaly detection. The challenge is less about technology and more about applying the right business logic. A practical starting point is to focus on a single, high-value use case such as zero consumption, unusually high usage, or sustained drops in consumption. From there, organizations can establish a baseline for normal behavior using historical meter data and create simple rules to identify deviations.
The most effective programs don't stop at raw consumption data. Adding context from sources such as customer, property, weather, or operational data helps reduce false positives and prioritize the anomalies most likely to require action. Another equally important part is validating findings with business stakeholders. Simple visualizations can help teams understand patterns, refine thresholds, and build confidence in the results before detection logic is operationalized.
Most importantly, anomaly detection should be connected to a clear response process. The goal isn't to identify anomalies for the sake of reporting, but to generate actionable insights that improve field operations, reduce revenue loss, and increase visibility into network performance. Organizations don't need sophisticated AI models to get started. In many cases, a combination of quality data, well-defined business rules, and iterative refinement can deliver meaningful results and establish a foundation for more advanced analytics in the future.
TL;DR Pick a single use case, establish a baseline, get stakeholders involved, connect to a clear response process.
Conclusion: From Detection to Action
Anomaly detection is only valuable if it drives action. In this case, the solution was designed so that flagged meters could be directly translated into prioritized field activities. That shift, from reactive to proactive operations, is where the real value lies. It reduces revenue loss from undetected issues, improves response times, and builds a more accurate picture of what’s actually happening across the meter network.
What makes this work particularly compelling is its scalability. The views and detection logic built in Snowflake aren’t a one-time analysis; they're a reusable foundation. As the client’s data matures and more historical patterns emerge, the thresholds can be refined, new anomaly categories added, and the solution can evolve toward more sophisticated predictive analytics.
For utilities sitting on years of meter data, that foundation is enormously valuable. The data is already there. The question is whether the infrastructure exists to unlock it.
Anomaly detection is a practical, high-impact starting point, and as this engagement demonstrates, you don’t need a complex AI model to get started. Clean data, well-structured SQL logic, and a willingness to explore what the data is telling you can go a long way.
Keyrus & Utility and Energy Orgs
Through comprehensive data platforms and AI-driven insights, Keyrus helps utility organizations gain actionable insights, ultimately increasing predictive maintenance, accelerating ESG and emissions reporting, and improving billing insights.
Keyrus is an international consulting group that turns data and intelligence into tangible, sustainable performance. Data is the organization's genetic code, and our legitimacy is built on 30 years of data foundations. That foundation is why we’re built for the new category now emerging: businesses that run on intelligence, not software, where distributed intelligence becomes the infrastructure of the enterprise. Value is moving into the underlying architecture of organizations, and Keyrus architects the Operating System of this new Intelligent Organization.
If your organization is looking to move from reactive to proactive and start running on intelligence, contact our Utility & Energy team to learn how we can help.
Anomaly detection in utilities is the process of identifying unusual consumption patterns that may indicate issues such as meter failures, leaks, unauthorized usage, data quality problems, or operational inefficiencies. It helps utility companies proactively investigate and resolve problems before they impact customers or revenue.
Anomaly detection helps utilities move from reactive to proactive operations by automatically identifying abnormal meter behavior. This can reduce revenue loss, improve customer service, prioritize field investigations, and increase confidence in consumption data.
Utility anomaly detection solutions can identify several types of irregularities, including: - Zero consumption on active premises - Abnormally low water usage - Unusually high water consumption - Potential leaks - Meter malfunctions - Unauthorized usage or meter bypasses - Data quality issues
Utilities can use scoring models that combine multiple data sources and business rules. Factors such as electricity usage, property occupancy status, and rental or sale listings help determine whether zero water consumption is legitimate or requires investigation.
A rolling median is less sensitive to extreme values and data errors than an average. This makes it more effective when working with utility data that may contain outliers, missing values, or erroneous readings that could distort anomaly detection results.
External data sources, such as property sale and rental listings, provide additional context that helps distinguish legitimate consumption patterns from potential issues. This reduces false positives and helps field teams focus on the highest-priority investigations.
No. Many utility organizations can achieve significant value using existing data, SQL-based business logic, cloud data platforms like Snowflake, and visualization tools such as Tableau. Advanced AI can be added later, but it is not required to start detecting anomalies effectively.
By automatically flagging and prioritizing suspicious meters, anomaly detection enables field teams to focus on the highest-risk locations first. This improves operational efficiency, reduces unnecessary site visits, and accelerates issue resolution.
Visualization helps stakeholders quickly identify trends, outliers, and patterns that may not be obvious in raw data. Interactive dashboards also support collaboration, validation of business logic, and faster decision-making during solution development.
Utilities can begin by identifying key consumption anomalies, establishing business rules, leveraging historical meter data, and building detection logic using existing analytics platforms. Starting with a focused use case creates a scalable foundation that can evolve into more advanced predictive analytics over time.