Risk management

Operation Planning Tool Improves Quality, Efficiency of Offshore Risk Management

The complete paper discusses Equinor’s operation planning tool, developed to present planners with the technical conditions of a platform, identify potentially dangerous combinations of concurrent activities, and propose learnings from 8 years of incident recordings.


The complete paper discusses Equinor’s operation planning tool (OPT), developed to present planners with the technical conditions of a platform, identify potentially dangerous combinations of concurrent activities, and propose learnings from 8 years of incident recordings. The OPT provides a single interface detailing a plant’s technical conditions, all planned work orders, and relevant lessons from previous incidents. By reducing reliance on personal experience, the tool has improved risk identification and handling, achieved faster knowledge transfer to new employees, and focused cross-platform knowledge sharing.


The operator’s requirement that all incidents or near-misses be reported has resulted in a database of over 108,000 incidents. Alongside lessons learned from previous incidents, planners must consider the suitability of concurrent jobs and any technical status that may affect the safety of executing a planned job at a specific time.

Before 2018, planners had to manually consult eight data sources, each with a separate user interface (UI), to obtain an overview of the plant’s technical status, concurrent jobs planned, and relevant lessons learned from previous incidents. This work flow had two clear bottlenecks:

  • Manual investigation of the separate data sources was time-consuming, often involving keyword searches or queries on structured data.
  • No easy way existed to obtain an overview of full asset status, with discipline leaders focusing individually on their jobs.

A high volume of planned jobs and the time-consuming nature of investigating multiple data sources dictated heavy reliance on personal experience of the production platform to accelerate searching of relevant data segments. However, this posed challenges to rapid onboarding of new employees and capture of cross-platform learnings. The OPT was developed to provide a single interface displaying data from all eight data sources.
Using natural language processing (NLP) techniques, the tool leverages unstructured data to supplement structured data, creating additional information that previously did not exist in a machine-readable format. Because of the technical vocabulary used to describe offshore operations, language models had to be generated to capture descriptions of equipment used and tasks being performed. This information was used to help determine relevance of an incident to a planned activity.

The complete paper describes in detail the methodologies used to generate specialized language models that can extract desired entities of interest. The paper also describes how, by incorporating such entities into a knowledge graph alongside entities derived from structured data, relevant matches can be made between planned work and historic incidents that can be reviewed.

The OPT has reduced time spent formulating a work plan and subjectively has improved risk identification and handling. Overall, the tool reduces reliance on personal experience and therefore provides a platform for faster knowledge transfer to new employees as well as focused cross-platform knowledge sharing.


The different data sources consulted during the planning of offshore activities and incorporated into the OPT are presented in tabular format in the complete paper.


The OPT has three main elements—platform status, operational plan with integrated risk information, and lessons learned from previous incidents. The data are overlain onto an outline of the physical locations of the platform to provide all decision-makers with a common view of the risks. The physical view is familiar to all participants at the same time because it is specific enough to show jobs and risks for each discipline. Most of the data also contain reference to a physical location and can be linked to the image.

The operational plan element focuses on the details of the jobs planned for execution during a time window. The plan view provides the possibility of filtering groups of jobs by discipline, location, or material needs, while linking jobs to the related risks. Cross-disciplinary jobs are highlighted with a special focus on well-known risk combinations to be avoided. Linking both frame-condition and job-related risks to a single job, together with flexible grouping and sorting functionality, enables dynamic, focused discussions on the largest risks.

The learning element focuses on connecting past incidents with current planned work to reduce the likelihood of a similar incident in the future. A complication of matching incidents to work orders is that a number of features that would be natural to link, such as equipment used or the type of work being executed, are not captured in the structured data, but instead stored only within the text fields describing the incidents or planned work.

To optimize information use and supplement the structured data, features can be extracted from the unstructured data fields using regular expressions and entity extraction. The learning element computes a relevance between the 108,000 historic incidents and a planned work order to recommend the most-relevant incidents to learn from before executing the order.

The complete paper discusses two subsets of NLP: regular expressions and entity extraction. Regular expressions entail identification of functional locations, which detail the system on which the work is to be performed, and provide a key indicator of whether a previous incident is highly relevant to a planned activity. Entity extraction identifies key elements in text and classifies them into predefined categories. These processes, and the process of connecting incidents and work orders in a knowledge graph, are discussed in ­detail in the complete paper. A schematic of the OPT learning knowledge graph also is presented.

Fig. 1 provides two examples of how an incident and work order can be connected. Fig. 1a provides an example of how the incident and work order are connected through two paths—one through the functional location identified through the use of regular expressions, and one through the equipment term identified by the use of the trained entity extraction models. The second example illustrates a more-complex connection between an incident and work order that share the same system and equipment type, but not the same equipment instance.

Fig. 1—Graph connections between incidents and work orders. In (a), the incident and work order both share a functional location and extracted entity term. In (b), the incident and work order share the same equipment type—firehouse cabinet and firewater and foam—but not the same functional location.


For each platform, the landing page of the OPT displays the current technical conditions of that platform. A sketch of the platform provides users with the ability to select an area and focus on its current status in terms of technical conditions, dispensations and well status, and other frame conditions affecting that area. The linked information displays a highlight of the risks. With a single click, the user can jump into the source system to see more details. The information shown in the heat map of the platform easily can be switched on and off to support a focused discussion on a single topic without losing the context of the high-level risks. Sample sketches are presented in the complete paper.

The Operations Plan page provides a dynamic overview of the operation plan with integrated risk information, with options to filter by key elements, making it easier to identify and discuss high-risk activities and either make changes to the plan or find compensating measures.

The Learnings page details all planned activities that have been linked to past incidents and provides users with clues as to whether further time investment is warranted. If so, a single click will connect to the full report of that specific ­incident with suggested actions.


At the time of writing, the OPT had been adopted by a number of assets to plan offshore operations for all Equinor-operated platforms on the Norwegian Continental Shelf. Significant interest has also been expressed in adopting the solution to help plan refinery activities and by the company’s international assets to enhance global learning.

Cross-platform learnings constituted a key delivery. Because of changing ownership of the platforms, equipment tag numbers are constructed across the platforms in several different ways. A catalogue of regular expressions was created to extract equipment tags within text fields.

The current version of the OPT focuses on preparation for, and during, the operational planning meeting. The tool’s functionality is being expanded to include creation of the operational plan. New functionality will include optimal scheduling of jobs on the basis of job prioritization and workers available, as well as tracking of ordered materials required for jobs. These features will involve inclusion of additional data sources, such as personnel-on-board information, boat and helicopter arrivals, and material availability. Additionally, new logic will be incorporated to highlight planned jobs for which personnel or materials are not available and to identify combinations of jobs to make efficient use of specialist personnel or construction equipment.

This article, written by JPT Technology Editor Judy Feder, contains highlights of paper SPE 195750, “Improving the Quality and Efficiency of Operational Planning and Risk Management With Machine Learning and Natural Language Processing,” by Claire Emma Birnie, SPE, Jennifer Sampson, and Eivind Sjaastad, Equinor, et al., prepared for the 2019 SPE Offshore Europe Conference and Exhibition, Aberdeen, 3–6 September. The paper has not been peer reviewed.