Risk Model Improvement Initiative

This study will curate threat and consequence types, field names, attribute lists, definitions, and data sources based on provided risk results or models into an interactive document. The purpose of the data dictionary is to support normalization of participant naming conventions and agree on the types of threats and consequences per industry standards and regulations. Each participant will have a map to this dictionary, so they are able to tie their naming conventions to project results. The overall objective is to organize a common risk language for efficient integration of current and future technologies, data systems, and other processes.

This study will apply machine learning and statistical methods to each participant’s model and\or risk results to reveal statistical contributions and variation of inputs to each threat and consequence type. Methods will also be applied to reveal the inputs to results considering interactions and non-linear relationships not detectable with simple EDA (exploratory data analysis). The objective of this analysis is to satisfy the overall requirements of 192.917(c) and 195.452, but more importantly provide guidance and transparency on preventive and mitigative (P&M) opportunities based on underlying data.

This study will normalize, anonymize and aggregate results from participant model analysis in deliverable #2 into an overall analysis. The objective of this analysis is to identify useful insights participants may want to consider for improving their risk models based on results across all participant models. The deliverables will be the same as the participant model analysis yet presented in a format to contrast and compare model structures and insights across different pipeline types, and categories of threats and consequences.

The pipeline industry is being encouraged to work towards a more quantitative approach to risk. Consultants are proposing a data driven machine learned approach to performing quantitative risk analysis (QRA) and will demonstrate how this can be performed leveraging existing models, data structures and PHMSA incident data. The deliverable will be a documented process for performing QRA including examples for common threats and consequences.

This is an optional deliverable available to each participant. The objective is to work with a representative sample of participant observational data to validate selected threat components of their risk models. Consultants will work with the participant to identify appropriate threat observation samples typically found through inspections, assessments and operational experience. The process is to machine learn predictive models based on these observations for purposes of statistical validation of current model results to observations. The overall objective is to validate the model with actual observations and provide insights for potential model improvement. The cost of this option will vary depending on the availability, structure and quality of observational data, the required sample size and the number of threats to be validated.