Reactor analytics drives nuclear industry towards machine learning

Nuclear digitization projects are tackling data quality and training challenges to implement predictive maintenance strategies and pave the way for machine learning capabilities, industry experts said.

Digitization projects are at the heart of U.S. nuclear industry efforts to reduce costs in the light of challenging wholesale market competition.

Cost reduction initiatives such as the Nuclear Energy Institute's (NEI's) Delivering the Nuclear Promise (DNP) plan have shown how digitization can help to boost asset value.

Launched in December 2015, the DNP plan called on operators to reduce generating costs by 30% between 2016 and 2018 through efficiency improvements. By December 2016, the DNP plan had identified $625 million in industry-wide savings through the issuance of more than 40 Efficiency Bulletins, NEI said.

Predictive maintenance is a key tool for identifying plant issues before they occur, increasing plant safety and providing engineers with more data and insights at increased speed to improve performance, Brent Shumaker, Senior Engineering Manager and 2nd Vice Chair of the American Nuclear Society, told Nuclear Energy Insider.

Recent advances in algorithm development, computing power and storage capability, as well as wireless technology, all provide the means for nuclear plant operators to remain competitive with other energy generating sources, Shumaker said.

“These technologies will play an important role in the existing and next generation of [nuclear plants] to maintain their safety and reliability for years to come, as long as plants take the initiative and foresight to take advantage of them,” Shumaker said.

     Profit margin advantage of AI adopters

                              (Click image to enlarge)

Source: McKinsey Global Institute report: “Artificial intelligence: The next digital frontier?” (June 2017). Copyright (c) 2017 McKinsey & Company.

While the cost benefits of digitization are clear, the shift from condition- and time-based asset management to the risk-based approach offered by predictive analytics has been a huge step for nuclear operators which have typically been risk-averse companies, Jim Newman, Senior Director Product Management at software group Bentley Systems, said.

The nuclear industry could take the next steps towards adopting machine learning technology in the next year, for specific processes that take advantage of the increased processing power, Newman said.

However, plant employees must be made aware they will not be handing over decisions to machines, particularly in the early phases, he said.

In the first implementations of machine learning, machines will sort through large data sets "and present us information so that we make decisions in minutes and not hours, not days and not weeks,” Newman said.

New solutions

Bentley’s AssetWise software solutions suite focuses on information management, asset reliability and operational analytics. Newman said the company is focused on growing and expanding machine learning into a more robust capability. 

Bentley announced a partnership with Siemens on October 9 to develop digital platforms that will accelerate digitization of power utilities. Jointly-developed digital solutions will integrate Siemens’ decision support tools with Bentley’s applications.

GE also recently unveiled additions to its suite of edge-to-cloud technologies and industrial applications to help its customers build complete asset strategies on its Predix platform. GE said these solutions can help companies move from intelligent asset management, to automation, to insights-led machine learning across a distributed system, enabling asset management and streaming analytics.

Earlier this year Areva announced a partnership with IBM Watson Internet of Things to develop machine learning tools, combining data from equipment and databases to improve maintenance schedule efficiency and optimize the use of limited resources.

Data quality

Global technology giants spent $20 to 30 billion on Artificial Intelligence (AI) in 2016 with 90% of that funding research and development (R&D) and deployment, according to McKinsey Global Institute.

Machine learning is a key driver to wider development and acceptance of AI, as it is seen as an enabler of other technologies such as robotics and speech recognition, McKinsey said in its June 2017 report: “Artificial intelligence: The next digital frontier?”

AI can deliver significant competitive advantages but only for firms that are fully committed, technologically and strategically, McKinsey warned.

“Take any ingredient away—a strong digital starting point, serious adoption of AI, or a proactive strategic posture—and profit margins are much less impressive,” it said.

Access to clean and consistent data sources remains a key barrier to adopting Big Data analytics in nuclear plants, Newman said.

The nuclear industry has a wealth of historical documentation required to retain licenses, but much of it is paper-based or on microfilm, which is not easily retrieved, searched or digitized, he said. While some operators have used computerized maintenance systems for some time, this does not guarantee the data will be of good enough quality to perform predictive analytics.

Without valid and reliable baseline data, predictive analytics can generate irrelevant or unusable insights, Newman warned.

For example, a machine analyzing incomplete or inconsistent data may not recognize a production-possibility frontier (ppf) reliability curve, which is used to expose the point at which a problem with an asset is predicted to occur, he said.

Valuable skills

Historical data also only reveals what happened rather than what should have happened and do not necessarily identify or analyze event drivers, Newman said. 

Engineering knowledge is critical to plotting the expected course of an event, understanding why there was a problem, as well as gauging the extent and impact of that problem, he noted.

The development of advanced diagnostic and predictive technologies such as advanced pattern recognition models employed by online monitoring (OLM) will require significant training and expertise, Shumaker said.

Engineers will need to understand the underlying algorithms used to provide predictive data analytics, as well as apply their inherent knowledge of plant equipment and operating systems, he said.

Experienced plant engineers might initially have limited knowledge of how OLM techniques function, while engineer’s familiar with OLM might not be skilled in recognizing and diagnosing particular plant faults, Shumaker added.

Focused spending

The nuclear industry should take calculated steps towards machine learning technology, starting with the processes that can take advantage of the increased processing power and maximize value, Newman said.

It is complex and potentially costly to build an asset model that allows the analytics solutions to perform useable, relevant and valuable predictions, he noted.

Utilities should decide on a case-by-case basis whether to build an asset model, he said. The operator can then prioritize areas for digitization, he said.

In one example, Bentley’s AssetWise ALIM digital solution replaced 50 mainframe applications at Ameren’s 1.3 MW PWR Callaway Energy Center in Missouri. The software platform enabled access to documents containing data from around 240,000 components, regardless of the information system or format.

Plant staff were able to view relevant data, drawings, maintenance records and design-basis documentation from a single platform for each component. AssetWise ALIM ensured timely and effective maintenance during a planned outage, significantly reducing revenue losses estimated at $1 million per day.

Forecasting power

Machine learning could help optimize outage scheduling according to external variables such as weather forecasts and the technologies are already being used in other sectors to optimize equipment purchasing and spares replacement, Newman noted.

Machine learning will also equip operators with more sophisticated tools to help ensure grid reliability. Increasing penetration of renewable energy capacity has increased the importance of grid resilience in the integration of centralized and rising distributed energy generation.

While digitization projects represent a significant challenge for nuclear operators, they must take advantage of the mass of plant data available to them, Schumacher said.

As well as reducing costs, predictive maintenance has been shown to improve plant safety and can therefore help to ensure existing and next-generation nuclear power plants are seen as a safe and viable source of energy, he said.

By Karen Thomas