Machine Learning: The New Intelligence in Utility Operations

Artificial intelligence and machine learning are the hottest things going in computer science, showing up in everything from cars to kitchen appliances.

Content Dam Up En Articles Print Volume 22 Issue 1 Product Focus Substation Management Maintenance Machine Learning The New Intelligence In Utility Operations Leftcolumn Article Thumbnailimage File

Artificial intelligence and machine learning are the hottest things going in computer science, showing up in everything from cars to kitchen appliances. The industrial world is benefiting from this revolution as well, and utilities are no exception. Machine learning analytics are transforming the way utilities operate and optimize their networks. We will discuss a few of the ways machine learning can change the way you monitor and maintain your assets:

1. Using anomaly detection to find hidden problems in equipment before they cause failures,

2. Using unsupervised machine learning to predict when assets will fail and when they won’t, and

3. Using reinforcement learning to optimize operations and workforce scheduling.

1. Anomaly Detection

Machine learning anomaly detection is used extensively across many industries for fraud detection, process control, manufacturing, cyber-security, fault detection in critical systems, and military surveillance. At utilities, anomaly detection is used in advanced analytics applications addressing varied use cases such as asset maintenance and revenue loss detection. A critical advantage of using machine learning for anomaly detection is the ability to learn what “normal” is for each individual asset, rather than relying on engineering models of an “ideal” unit.

1711uppf Spacetime P01

Anomaly detection is a useful tool for industrial operations, made more accurate and powerful using machine learning.

A typical machine learning anomaly detection approach defines a region in sensor data from the asset being studied that represents normal behavior, develops a model that describes this normal behavior, then concludes that any observation not belonging to this model is an anomaly. Machine learning applications can ingest enormous amounts of data and find the tiniest needles in a data haystack. In our experience with clients, machine learning anomaly detection will identify problems well before human operators or engineers notice anything out of the ordinary, resulting in less downtime and lower expenses when combined with optimized maintenance programs.

Common issues with data, however, can make anomaly detection challenging. With some ingenuity, there are paths to overcome the challenges:

• Defining a normal region in the data that encompasses every possible normal behavior is difficult. For example, one SpaceTime customer seeking to predict anomalies with a sophisticated renewable energy asset had less than a year of operating data available to train the model. To overcome this challenge, the company focused on short time periods following the asset type’s maintenance intervals to help define “normal,” which provided a higher level of assurance that the data used for training did not contain operating anomalies.

• Normal behavior may evolve over time. To overcome this issue, we applied a regularized model that reduces the variability of the estimated model and avoids overfitting to the training data. Regularization deliberately adds bias to the estimated model parameters, which is essential in making the normal fitted model generalize well to time periods that were never observed before.

• Finally, noisy data can lead to a high rate of false positives. To overcome this, SpaceTime used a technique using the Irwin-Hall distribution that scores each observation for anomalies, but combines a sequence of observations to trigger an alert. In other words, if the data is noisy, the alert will not be triggered, but if a sequence of anomalous observations occurs for an extended period of time, an alert will be triggered.

1711uppf Spacetime P02

Machine Learning for Optimized Asset Lifecycle

Using machine learning, failure times can be predicted for individual assets and with greater confidence. Machine learning also optimizes repair schedules, increasing ROI.

Anomaly detection, in its most general form, is not an easy problem to solve. With an understanding of the project’s subject, the business it supports, the assets’ characteristics and operations, and the relationship of the data among them, proven statistical techniques can be applied to overcome many of the common issues with data.

2. Asset Failure Prediction

Utilities, especially those with an aging asset fleet, are moving from scheduled maintenance to condition-based maintenance (CBM). Often CBM compares a statistical model for an asset class to a specific asset’s rated condition to determine which asset needs attention. As previously discussed, using machine learning anomaly detection is an even better way to determine an asset’s condition and whether it needs maintenance, because it can define normal and anomalous conditions for each asset individually. For some assets, though, when the right time-series data is available, there is a better way.

While most CBM systems attempt to predict imminent failure, they aren’t designed to reveal the probability that an asset will survive an unhealthy condition and continue operating. Unsupervised machine learning models can “see over the hill” and predict failure probability at different times in the future. The models learn and understand a specific asset’s possible paths to failure, the probability distribution of it taking any of those paths, and the failure times associated with those probabilities. Another machine learning technique, reinforcement learning, can then balance risk vs. cost and provide an optimal window for replacing or repairing an asset. This maximizes the operating life of the asset, optimizing capital expenditure. That same approach also provides the basis for optimizing the repair work, automatically determining the optimal schedules and routes for repair crews, improving operating expense efficiency as well.

1711uppf Spacetime P03

Global Optimization of Operations

Optimizing workforce field operations must account for essential random inputs such as weather. Stochastic optimization is a machine learning technique that deals with uncertainty.

3. Workforce Optimization

Scheduling crews to maintain and repair utility asset networks requires managing multiple variables, some of which, such as the weather, traffic and real-time levels of power generation and consumption, are constantly shifting. Making optimal real-time decisions with this much uncertainty requires manipulating more variables, constraints and tolerances than the human mind can handle.

Reinforcement learning can be used to optimize maintenance crew schedules to minimize operations and maintenance costs, asset downtime and safety incidents. Doing so requires being able to optimize under uncertainty, which can be accomplished using a technique called stochastic optimization. Stochastic optimization explicitly accounts for the uncertainty inherent in the system by using continuous variables (a variable that can take on infinitely many, uncountable values) in its calculations. By applying stochastic optimization, we can calculate a marginal on all of the attributes of the problem—such as the price of gasoline, the number of people per crew, work hours, travel time, spare parts inventory, market prices for the system output, and so on. We can then ask our model, for example, if marginal profits will increase or decrease if we increase total crew hours per week to reduce the time it takes to complete repairs. Similarly, we can ask whether having a crew stay at a job site to perform a repair scheduled for the following week, delaying their next scheduled job, increases or reduces profits.

Another aspect of machine learning relevant to optimization is online or continuous learning. Some machine learning techniques (e.g. deep learning) use batch learning, where they are trained on a large set of data all at once to learn the best predictors of future data. Online learning, by contrast, operates on a continuous, sequential stream of data and updates its best predictor for future data at each step. This method is particularly useful in the optimization problem, since crews operate in an environment that is stochastic (e.g. changing weather and traffic conditions) and the application is receiving a steady stream of time-series data (e.g. state of the equipment, location of crews, status of ongoing repairs and crew availability, weather reports and traffic reports).

Combining stochastic optimization and online learning techniques allows us to build a reinforcement learning model to continuously optimize maintenance and repair crews to minimize crew hours and reduce downtime, maximizing profits.


Mature, proven machine learning analytics are now available that can measurably improve how utilities monitor and maintain their assets. Many utilities are just beginning to digitize their operations, while others have rapidly adopted these new technologies. Wherever your utility is, now is the time to plan for and implement machine learning applications that will transform every area of utility operations. UP

More in Safety