On the value of data fusion and model integration for generating real-time risk insights for nuclear power reactors

https://doi.org/10.1016/j.pnucene.2020.103497Get rights and content

Highlights

  • Integration of data science, model-based reasoning, and probabilistic risk assessment.

  • Modeling online risk using heterogeneous data and models.

  • Improved risk-informed decision-making for nuclear power plants.

  • Operational risk assessment and event response.

Abstract

The integration of data science, analytics, and model-based reasoning provides a mechanism for enhanced understanding of systems and improved decision-making, but its potential has not been thoroughly explored for improving the safety and operational efficiency of nuclear power reactors. Nuclear power owners, operators, regulators, and researchers have made significant investments in probabilistic risk assessments, numerical models, computational simulations, and development of databases that capture industry-wide component performance and operating experience. The nuclear industry is relatively unique in the size, variety, scope, technical sophistication, and quality of available data and models that capture system performance under normal operations and a wide-range of adverse event conditions. However, to date, these resources have been used in a largely static and siloed manner. Data science, analytics, and model-based reasoning provide a mechanism for fusing diverse data sources and models to develop new insights on a variety of topics. Of particular interest to the nuclear industry is the ability to leverage these resources to enhance the safety and operational efficiency of nuclear power reactors.

In this paper, we present a challenge to the nuclear energy community to better leverage the existing investments in data and models to enhance decision-making. In particular, we propose that integration of recent advances in data science, analytics, and model-based reasoning provides a valuable opportunity for the nuclear industry to build upon their existing investments by accessing the power of modern data integration and risk assessment tools. We begin by describing common data and model resources available in nuclear power operations and safety analysis and offer commentary on the potential power of using Bayesian networks as a structured framework for data fusion and model integration. Then we present two example problem structures for modeling risk-informed operational decisions using heterogeneous data and models to provide simple illustrations of the means by which information streams can be leveraged in real-time to provide online assessment of risk and to increase diagnostic capabilities. Illustrative model formulations are presented for decisions under adverse events and normal operational contexts. We conclude by identifying research activities that will enable the transformation of decision-making by applying new computational and modeling tools to existing data and models.

Introduction

Nuclear power facilities represent an important component of the energy production portfolio in the United States and internationally, serving as a significant carbon-free baseload energy source. As such, ensuring the continued safe, secure, and reliable operation of nuclear power plants (NPPs) is a national and international priority. However, meeting this objective is associated with a number of engineering and economic challenges.1

NPPs are comprised of complex systems involving dynamic, interconnected, and spatially dispersed structures and components as well as human performance elements. Moreover, NPPs, like many industrial facilities, are exposed to a number of natural and human-made hazards that may evolve over time and are associated with significant uncertainty. In addition to engineering challenges, the financial viability of nuclear power in the United States is being challenged by market conditions. Reduced revenues coupled with the relatively high operating costs has led to the shutdown of a number of NPPs in the United States (US EIA, 2018). The United States Department of Energy has identified several research and development areas of relevance to ensuring that nuclear power remains viable, including the need for information technologies that provide an enhanced understanding of plant operating conditions and support state-of-the-art NPP safety analysis that yield new insights regarding plant safety and operational margins (US DOE, 2015).

For decades, probabilistic risk assessment (PRA) has played a critical role in the safety and regulation of the nuclear power industry and has been used to develop safety insights for internal plant hazards as well as a limited number of external hazards. The Fukushima Dai-Ichi accident revealed and re-emphasized the important insights that can be gleaned through use of PRAs (i.e., sources of common cause vulnerabilities such as the placement of all emergency diesel generators at elevations potentially exposed to tsunami flooding) as well as highlighting several ways in which the current practices and tools used for PRA are not able to fully capture the complexities of event progressions including the need to treat complex event progressions and dynamic system behaviors (National Research Council, 2014), (Siu et al., 2013).

This paper proposes that data science, analytics, and model-based reasoning provide a mechanism to develop new insights on a variety of topics related to safety and operational efficiency. Machine learning and data analytics have the potential to improve the assessment of plant safety and support increased operational efficiency and productivity of NPPs (OSU, 2017). However, opportunities exist to go beyond data analytics by coupling data analytic approaches with the PRA tools, numerical models, computational simulations, and industry-wide databases that the nuclear industry has been investing in for decades. Despite this potential, there are only scarce efforts to rethink the field of nuclear power PRA by coupling it with new data analytic methods and technologies. To date PRA has largely been used as a static technology, but as noted by Goble and Bier (2013), dynamic monitoring and risk assessment has ‘‘game changing’’ potential, which has not been exploited.

In this paper, we present a challenge to the nuclear energy community to better leverage the existing investments in data and models to enhance the scope of PRA and risk-informed decision-making. We begin by describing common data and model resources available in nuclear power operations and offer commentary on the potential power of using Bayesian networks as a structured framework for data fusion and model integration. Then we present two example problem structures for online risk modeling using heterogeneous data and models to provide simple, conceptual illustrations of how information streams can be leveraged in real-time to provide online risk assessment and increased diagnostic capabilities. Model formulations are presented to illustrate the capabilities of real-time risk assessments for plant scenarios involving adverse events and normal operational contexts. The paper concludes with identification of research activities that will enable the transformation of safety and operational decision-making by applying new computational and modeling tools to existing data and models. This paper focuses primarily on data analytics from the perspective of online risk assessment (for event response and online assessment) and offers limited commentary on the potential extended applications to risk-informed decision-making.

Section snippets

NPP data resources

The nuclear industry is relatively unique in the size, variety, scope, technical sophistication, organization, and quality of available data and models that capture system performance under normal operations and a wide-range of adverse event conditions; see Fig. 2-1. However, to date, these resources have been used in a largely static and siloed manner and have not been fully leveraged using modern data processing tools.

It is noted that a key nuance of this paper involves the use of the word

Structured formulations for leveraging online data streams and observations

By coupling information from the aforementioned data streams with probabilistic and other physical or mechanistic models, powerful insights may be garnered. Through this coupling, it becomes possible to perform complex reasoning tasks such as diagnosing potential causes of changes in structure, system, or component (SSC) status via multi-directional probabilistic inference (e.g., see (Groth et al., 2015), (Darling et al., 2018)). Multi-directional inference encompasses three classes of

Moving forward

While machine learning and data science provide an opportunity to improve upon existing practices, there are challenges that must be overcome before realizing the benefits of data fusion and model integration for improving NPP operations. The first step toward resolving these challenges is to develop a strategy that is practical and aligns with current regulatory and operational paradigms. For example, similar to the risk-informed (rather than risk-based) approach utilized by the nuclear

Conclusion

This paper challenges the nuclear industry to better leverage its deep investments in nuclear power data, PRAs and other plant models to enable better decision-making from safety and operational efficiency perspectives. It is proposed that (near) real-time/online incorporation of diverse data streams with existing plant models can support improved risk-informed decision-making associated with multiple classes of decisions and provide more accurate representations of on-line plant risks. In

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements and Funding

The authors acknowledge and appreciate research support received from the University of Maryland and through the United States Nuclear Regulatory Commission Faculty Development grant program. The authors acknowledge the valuable feedback provided by the anonymous reviewers of this paper; reviewer comments improved the content and structure of the paper.

References (33)

  • K. Zwirglmaier et al.

    A discretization procedure for rare events in Bayesian networks

    Reliab. Eng. Syst. Saf

    (Sep. 2016)
  • M. Bensi et al.

    A Bayesian Network Methodology for Infrastructure Seismic Risk Assessment and Decision Support

    (2011)
  • M. Bensi et al.

    Framework for post-earthquake risk assessment and decision making for infrastructure systems

    ASCE-ASME J. Risk Uncertain. Eng. Syst. Part Civ. Eng.

    (Mar. 2015)
  • R. Boring et al.

    How to run A control room simulator research study

  • US NRC

    Risk Management Technical Specifications

    (Mar. 28, 2019)
  • Events assessment

  • Cited by (6)

    View full text