Page 18 - Gi_August2019

Digital guidance of uncertainty reduction in hydrogen value chains

FIGURE 2 INCREASING CONFIDENCE IN ESTIMATED FAILURE PROBABILITY BASED ON INCREASING FIGURE 3 H2 DETONATION
NUMBER OF SIMULATIONS

simulations. DNV GL is currently surrogate models and the prediction combine them directly.
running large scale experiments on uncertainty, it is possible to design The recent development in
the safety of hydrogen transport and adaptive learning algorithms that not probabilistic programming is a powerful
storage at its Spadeadam facility that only reduce the overall uncertainty of tool to include beliefs in a probabilistic
will reduce uncertainties related to the predictions, but which reduce the model. Through Bayesian inference, we
the hydrogen value chain. However, uncertainty where it matters the most can update our beliefs after considering
we quickly encountered the limitation (e.g., close to a safety critical limit)⁶. observed evidence, e.g., after field
that we cannot run experiments or Thus, we can design experiments with observations and by running simulations
simulations for all possible variations of the objective to reduce risk in a more or experiments (see Figure 2)⁸.
pipeline parameters and load scenarios. efficient way. DNV GL has, through continued
To ameliorate this obstacle, we need Having established such surrogate research over several years, successfully
predictions of relevant situations and models (combining results from both applied Bayesian thinking and
scenarios – even though they have not experiments and simulations), it is developed their Multi Analytic Risk
explicitly been tested or simulated. possible to combine these models Visualisation (MARV) to assess corrosion
Recent advancement in computational with real-time operational data and risks for gas pipeline operators, as
power and ease of implementing probabilistic models to assess, for well as successfully applied adaptive
advanced machine learning methods example, the real-time risk related to exploration to reduce uncertainty in
have enabled us to train surrogate models brittle fracture and fatigue accumulation failure probabilities related to, for
(fast running approximations) based on in a probabilistic digital twin⁷. instance, crack growth in pipelines⁹.
observed data to enable predictions of From a Bayesian worldview, we
not-yet observed scenarios. Good policy is based can rarely be certain about a result,
These approximations come with a on understanding uncertainty but we can be very confident. By
cost in the form of added uncertainty. When the epistemic uncertainty (related including beliefs of the outcome of
Some of the most popular machine to knowledge) is at a level where it is not different policy decisions, we can run
learning methods do not treat practical to reduce it further, or when it is ‘what if?’ type assessments and assess
uncertainty consistently. This might be apparent the physical system or scenario the outcome of a policy through the
acceptable when a lot of training data can be changed to move it towards an distribution of potential effects.
is available and the consequence of an acceptable level of uncertainty, we have
erroneous prediction is low. However, the foundation for fact-based rational
for capital intensive infrastructure decisions. However, we still need to References
with high consequences for failures, assess subjective uncertainties reflecting 1. Tannert et al. (2007) ‘The ethics of uncertainty’,
this is not acceptable. contextual policy, regulatory, technical, www.onlinelibrary.wiley.com/doi/full/10.1038/
sj.embor.7401072
To handle uncertainty consistently, and market aspects. In the case of 2. van Asselt and Rotmans (2002) ‘Uncertainty in
specifically where consequences are transporting pure hydrogen through integrated assessment modelling: from positivism to
high and we do not have enough data to transmission pipelines, it is assumed pluralis’, Climatic Change, 54: 75–105.
be reassured by big-data methods, we that hydrogen has a higher cost than 3. Hafver et al. (2016) ‘Enabling confidence’, DNV GL
position paper.
suggest utilising a class of probabilistic today’s transportation and consumption 4. Eldevik et al. (2017) ‘Risk, Uncertainty, and What
machine learning methods. of natural gas. Thus, to make hydrogen a if? – A practical view on uncertainty and risk in the
More specifically, we suggest using viable energy carrier, policies related to knowledge and physical domain’, ESREL.
Gaussian processes to establish these carbon taxes, subsidies and hydrogen 5. EIGA: EIGA Doc. 121/14 ‘Hydrogen Pipeline
Systems’ www.eiga.eu/publications/eiga-
fast running approximative models and pricing must be assessed. These documents/doc-12114-hydrogen-pipeline-systems
determine an uncertainty measure of the uncertainties need to be combined 6. Eldevik et al. (2018) ‘Safety+AI’. www.ai-and-
safety.dnvgl.com/#sec-critical
prediction. This has the added benefit that with the uncertainties of the technical 7. Hafver et al. (2019) ‘Probabilistic digital twins’.
we can adaptively prioritise experimental feasibility discussed above. As these www.ai-and-safety.dnvgl.com/probabilistic-twin
and simulation efforts based on the uncertainties are different in nature 8. PROB PROG: Cameron Davidson-Pilon
knowledge we have gained so far. (technical uncertainties are often et al., ‘Bayesian Methods for Hackers’. www.
camdavidsonpilon.github.io/Probabilistic-
The design of experiments (DOE) related to variability and knowledge, Programming-and-Bayesian-Methods-for-Hackers
is a well-known technique used to while subjective uncertainties are 9. Keprate et al. (2019) ‘Structural reliability assessment
prioritise experimental and simulation more often related to the beliefs of of pipeline girth welds using Gaussian process regression’,
Proceedings of AIM-PIMG2019, Houston.
efforts. However, utilising these the assessors) it might be difficult to

18

HydrogenValueChains.indd 3 18/07/2019 15:40