An open source library for automatically analyzing dynamic systems

Presented at NTNU BRU21 2025.05.26 by Steinar Elgsæter

Example of advanced analytics applied to real-world industrial data at scale.
The library deals with dynamic modeling (including transients PID feedback loops), and solves many practical issues that one faces in applying methods to real-world data that are often in need of cleaning or pre-processing and may be less-than-perfect.
(advanced analytics is one of 17 emerging technologies under the "Industry 4.0" umbrellla term.)

The library is in use by Equinor for condition-based monitoring of feeeback loops at scale, and allows for "what-if simulations" and can be used to automatically detect changes or the need for mainentance in a large fleet of feedback loops.

This reduces reaction time and increases the productivity of domain experts, thereby improving margins by minimizing the losses and inefficiencies.

A novel open-source library was developed to do time-series analysis at large scaled autonomously or semi-autonomously, to generate deeper insights and reccomendations from large industrial data sets (based on C#,.NET)

The library is openly avilable on github, https://github.com/equinor/timeseriesanalysis

1. An open-source library for automatically analyzing dynamic systems Steinar Elgsæter Tech Lead Advanced Analytics Equinor

2. Motivation : «Advanced anlytics » 2 | An open-source library for automatically analyzing dynamic systems

3. How to maintain situational awareness systematically and effectivly in O&G ? Challenges : a lot of equipment to cover, constraints on manpower / expertise , pressure to improve the way we work systematically look at the right place at the right time systematically understand and react to changes systematically describe how systems interact Opportunity : Access to vast amounts of time-series data and computational power , two of 17 emerging technologies often referred to as «Industry 4.0» Answer : «Advanced analytics » and «digital twins »

4. Advanced analytics is using sophisticated techniques and tools, typically, beyond those of traditional business intelligence (BI), to autonomously examine data to discover deeper actionable insights , m ake predictions , or generate recommendations . typically at very large scale (“fleets”) 4 | github.com/equinor/timeseriesanalysis

5. 5 | An open-source library for automatically analyzing dynamic systems Digital twin (in this presentation, overloaded term): A model that has realistic predictions

6. What is « advanced analytics »? - google maps analogy Level Travel analogy Information provided / Value 1 Mapbook «as- built » 2 Traffic app « current » + «as built » 3 Google maps « actionable insights » equivalent Process simulator «Digital twin » «Advanced analytics » 6 | github.com/equinor/timeseriesanalysis

7. Open- source library 7 | An open-source library for automatically analyzing dynamic systems

8. Inspiration : Building on the success of Model- Based Predicitive Control (MPC) MPC Arguably , the most succesful use of «digital twin » models in Equinor , in terms of Number of models in use ( houndreds / thousands ) The predictive power of the models Correct level of detail /parameters for the datasets at hand in O&G Grey- box : human- understandable + automatically identifiable 8 | An open-source library for automatically analyzing dynamic systems

9. Our advanced analytics library Full scale full fidelity dynamic training simulator «AI» digital twins Dynamic Yes Yes No Training Automatic* Manual Semi-automatic Simulation speed >10.000x realtime 1-10x ? (Fast) Model complexity (parameter count ) Low-low High High Capex / Opex Low-low High High Re-training complexity Low-low High High « Simpler is faster, and cheaper » 9 | github.com/equinor/timeseriesanalysis 10x reduction in costs compared to alternatives

10. 10 | github.com/equinor/timeseriesanalysis

11. github 11 | github.com/ equinor / timeseriesanalysis code documentation

12. An extensive set of unit test A good way to understand the code Plotting of time-series results can be enabled An essential feature to develop the code (test driven development ) An essential feature for re- using the code and for others to extend the code 12 | github.com/equinor/timeseriesanalysis

13. Case study : PID- controllers 13 | An open-source library for automatically analyzing dynamic systems

14. Case: PID- control 14 | An open-source library for automatically analyzing dynamic systems PID Process Disturbance ( unmeasured ) Setpoint ( measured ) Manipulated ( measured ) Process variable ( measured )

15. a) «Discover deeper insights » 15 | An open-source library for automatically analyzing dynamic systems Based on disturbances estimated Systematically describing how PID- controllers interact Each node represents one PID-loop Automatically generated analysis and graph Gives a deeper insight compared to considering each PID-loop in isolation Loop! PID Process

16. b) Make predictions 16 | An open-source library for automatically analyzing dynamic systems PID Process Interactive simulator

17. c) Generate recommendations Recomending : looking at the right place at the right time reacting to changes automatically detecting and analyzing : incoherent data changes in the process changes in the pid-controller oscillations trend in disturbance 17 | An open-source library for automatically analyzing dynamic systems Example of incoherent data PID Process Model Measured

18. At « fleet » scale 18 | An open-source library for automatically analyzing dynamic systems 900 «digital twins » created so far with just a small team of 2-3 people re- simulating and re- identifying models every night

19. Conclusions 19 | An open-source library for automatically analyzing dynamic systems

20. Unique selling points tailored for “as is” O&G real-world data sets that are often sparse in information “control” approach to time-series modeling inspired by Model-Predictive Control(MPC) invite open collaboration (vendors, academia, other energy companies) modelling library open-source https://github.com/equinor/timeseriesanalysis 20 | models numerically robust millisecond fast dynamic representative of the past «advanced analytics» automate analysis understanding interactions «what if» on-the-fly models behind your dashboards understandable enables

21. Steinar Elgsæter Tech Lead Advanced Analytics An open-source library for automatically analyzing dynamic systems