Metering, Smart Grid, T&D

Improving Grid Reliability Through Machine Learning

Issue 1 and Volume 20.

human circuit

by Uma Sandilya, C3 Energy

During the next five years, the number of electric smart meters is expected to more than double worldwide. While representing only a portion of the sensored devices on the grid, the escalating number of smart meters indicates the rapid penetration and growth of the smart grid.

Fueled by much needed upgrades to legacy power networks and digitization in a more connected home, as well as new electrification within growing economies, this development is poised to unleash a flood of data that utilities never have seen before. The data volumes’ coming from smart meters, advanced sensoring in the form of phasor measurement units, and distribution automation systems will approach one terabyte per day at a typical utility and drive utilities to leverage data analytics to make sense of this data and turn it into actionable insights.

Virtually all utilities today, however, use siloed, enterprise software products for managing grid-side asset operations. These disconnected systems arose for legitimate operational purposes, including meter data management (MDM), asset management, work management, supervisory control and data acquisition (SCADA) for transmission and distribution, outage management (OMS) and generation control systems. But these independent systems were not designed to support data sharing and analysis and thereby preclude achieving a unified view of the state of all utility-operated assets.

As utilities grapple with the influx of data from these source systems and growing number of sensored devices, big data rapidly is becoming an asset. And utilities are realizing that interpreting these data has the potential to transform business processes from customer service operations to engineering and maintenance and beyond.

Advanced real-time analytics applied to an aggregated and federated set of data can improve grid reliability by revealing the operational health and use of energy components in unprecedented detail, allowing predictive maintenance and planning not previously possible. For a typical utility, this translates to considerable potential efficiencies and cost savings on the order of $300 per meter per year.

More specifically, enterprisewide data analytics focused on the maintenance and reliable operation of transmission and distribution assets can drive significant improvements in grid reliability and unlock some $200 per meter each year for a utility and its customers. Unlocking this potential requires three important steps:

1. Aggregate and federate data to understand the state of grid assets. To realize the full economic potential of their assets, utilities must begin by seeking solutions capable of aggregating, federating and analyzing data at a scale that far exceeds the complexity of their systems of record and which move beyond simple visualization techniques for managing network health and predictive maintenance.

At a 5 million-customer utility, for example, analyses of the distribution transformer network, enabled by 15-minute interval smart meter data at the premise-level, revealed that more than 80 percent of distribution transformers were used at less than 10 percent of their rated capacity. Incorporating such information into future asset planning has the potential to allow capital expenses to be shifted to other portions of the network, resulting in capital efficiency to the tune of $50 million per year.

In choosing a data aggregation and analytics platform, utilities should exercise caution about scalability and future growth potential. Some traditional point solutions focus on data from one or two systems and simply aggregate the data into a visualization layer. This makes it easy for a utility operator to view the locations of grid assets and possibly even their basic operational status, but it neither significantly affects overall grid reliability nor can it produce noticeable returns. Utilities require a platform that integrates all the data from multiple source systems, performs complex analytics powered by state-of-the-art machine learning algorithms, and produces insights that can be prioritized and visualized in an intuitive manner to the end user.

2. Expand scope of predictive maintenance using data. Utilities globally have adopted a run-to-failure approach to asset management for all but critical assets (e.g., nuclear power plant components, high lead time assets such as step-up/step-down transformers). As a result, the average utility spends most of its time and resources on scheduled and reactive maintenance and little on predictive maintenance.

There is tremendous value in shifting most, if not all, of the reactive maintenance and a carefully selected set of scheduled maintenance to the predictive maintenance category. This ensures utilities only maintain assets that need attention and effectively reallocate capacity to respond to real-time events, thereby potentially reducing the duration of outages related to asset failures.

3. Use machine learning to improve prediction accuracy. Machine learning is simply using mathematical algorithms that can learn from data. It is the ability for a computer to learn without being explicitly programmed to do so. Using the same technology premise aircraft engine manufacturers use to monitor engines and ground a plane before a malfunction occurs, smart grid analytics can learn from historical data and predict asset failures to improve grid reliability.

Machine learning algorithms are trained using actual utility data on historical failure patterns, incorporating expert rules that are being used by decision-makers at utilities along with as much historical data as the utility has available. An added benefit of this approach is that utilities can codify expert knowledge that eventually will be lost through the attrition of tenured talent.

By integrating billions of data points from the headend, MDMS, OMS, AMS, WMS and other source systems and then applying relevant machine learning techniques, the algorithms can pinpoint asset health issues and identify patterns to predict potential failures.

Take the example of distribution transformers. A predictive maintenance solution, incorporating advanced machine learning techniques, can calculate and keep current in near real time a health score for each transformer based on numerous factors such as historical loading, through-fault statistics, winding and oil temperature, trends in DGA results, on-site diagnostic test results and forecasts of relevant parameters such as load levels, ambient conditions, and the projected as-operated network state. The health score can be configured to consider not only the probability of failure, which also can be expressed as a mean time to failure to enable operational decisions, but also the impact of the failure, both to customers and the utility. Operators can use this health score to prioritize assets on the system for predictive maintenance.

Another example is related to the use of a comprehensive set of feature inputs that allow the machine learning algorithm to identify and use only the features with the highest predictive power. For instance, most utilities do not analyze event data from their SCADA systems related to faults that do not result in breaker operations. A good example of such an event is a transient fault along a feeder that is extinguished by the variable impedance in the grounded neutral of the transformer in the primary substation supplying the feeder. This feature, combined with other relevant features for the feeder such as vegetation index, ambient conditions along the feeder, kVA loading, and historical and planned maintenance, become inputs to a machine learning classifier. The output of the classifier is the probability of future sustained faults that might result in a breaker operation and loss of a portion of load along the feeder. Utilities can use this insight to dispatch resources preemptively to take care of the failure before it happens.

Similar techniques can be applied to assets throughout the distribution network, and the impact on grid reliability is substantial. Utilities benefit directly from fewer outages because of early identification and resolution of at-risk assets. They also see financial benefits in reduced capital expenditures by some 4.5 percent from a reliability-centered maintenance and replacement approach, taking the place of the traditional emergency replace-on-failure approach. In addition, utilities can decrease operational costs some 2 percent through the systematic identification of specific assets that require replacement ahead of failure.

Conclusion

As major smart meter deployments are completed in the next few years and the grid transforms with more sensored devices, utilities must ensure reliability remains a priority. Measurable increases in reliability and value can be achieved only with advanced analytics that enable predictive maintenance. Utilities need a solution to monitor system health and predict asset failure to improve reliability at scale.

An important catalyst for this change across utilities globally is the presence of regulatory incentives that promote the shift to predictive maintenance by providing economic incentives to reduce operating costs (on a $/kwH served/year basis) while improving system reliability (measured through indices such as System Average Interruption Duration Index (SAIDI) and System Average Interruption Frequency Index (SAIFI)). A handful of U.S. and European utilities are at the forefront of this trend with support from their local regulatory agencies. These utilities are seeking solutions that support improved diagnostics with multiple data sources and leverage machine learning techniques to facilitate predictive maintenance.

With the right regulatory incentives, the global utility industry soon will see an inflection point in the use of advanced analytics to transform the way they deploy, maintain and operate their assets, resulting in a more efficiently operated grid for their customers.

Uma Sandilya is director of product marketing at C3 Energy. Prior to C3 Energy, he was an engagement manager at McKinsey & Co., where he was a core member of the firm’s electric power and natural gas practice. Prior to joining McKinsey, he managed transmission operations and system restoration at Entergy Corp.

More POWERGRID International Issue Articles
POWERGRID International Articles Archives
View Power Generation Articles on PennEnergy.com