Wednesday, 15 January 2025
Hall C (New Orleans Ernest N. Morial Convention Center)
High ozone concentrations can occur during some winters in the Uinta Basin, Utah. These episodes occur when heavy snowfall and rising pressure cultivates growth of a cold pool. In turn, this traps emissions from oil and gas operations, and these precursors lead to unhealthy levels of ozone under sunlight.
The Bingham Research Center runs an email list (Ozone Alert) that communicates risk of high ozone to stakeholders in public health, oil/gas, government, etc. These emails are manually written, guided by numerical weather prediction (NWP) forecasts, despite these models’ difficulty in capturing important small-scale features such as shallow temperature inversions and spatial variation in snowfall accumulation. This unreliable guidance is worsened by data scarcity limiting skill of nowcasts. Nonetheless, the ability for human forecasts to tolerate error (i.e., smooth irrelevant noise) in forecast variables pertinent to the well understood Uinta Basin winter-ozone mechanism suggests a rule-based structure may be apt.
A mixed but promising performance of this concept ozone-prediction model was shown in a recent paper describing a prototype. Leading from this, we present the system’s core: an unfamiliar but low-complexity fuzzy-logic inference whose construction was informed by the observational archive. Our prediction system (CLYFAR) generates ozone forecasts with input from numerical weather prediction (NWP) models. CLYFAR uses rule-based fuzzy inference (e.g., “high snowfall” and “high pressure” leads to “elevated ozone”) where variables (e.g., “snow”) have partial membership of a category (e.g., “high”) between 0 and 1. We argue herein that this better accounts for uncertainty for risk-averse decision-makers. We choose the simplest system that generates useful forecasts — i.e., encouraging explainable AI (XAI) behaviour — to avoid deployment of complex black-box (e.g., deep-learning) systems whose behaviour may be surprising and therefore less trustworthy. We present preliminary evaluation of performance of CLYFAR version 1.0.0, and lay out a roadmap of development in coming projects.
The Bingham Research Center runs an email list (Ozone Alert) that communicates risk of high ozone to stakeholders in public health, oil/gas, government, etc. These emails are manually written, guided by numerical weather prediction (NWP) forecasts, despite these models’ difficulty in capturing important small-scale features such as shallow temperature inversions and spatial variation in snowfall accumulation. This unreliable guidance is worsened by data scarcity limiting skill of nowcasts. Nonetheless, the ability for human forecasts to tolerate error (i.e., smooth irrelevant noise) in forecast variables pertinent to the well understood Uinta Basin winter-ozone mechanism suggests a rule-based structure may be apt.
A mixed but promising performance of this concept ozone-prediction model was shown in a recent paper describing a prototype. Leading from this, we present the system’s core: an unfamiliar but low-complexity fuzzy-logic inference whose construction was informed by the observational archive. Our prediction system (CLYFAR) generates ozone forecasts with input from numerical weather prediction (NWP) models. CLYFAR uses rule-based fuzzy inference (e.g., “high snowfall” and “high pressure” leads to “elevated ozone”) where variables (e.g., “snow”) have partial membership of a category (e.g., “high”) between 0 and 1. We argue herein that this better accounts for uncertainty for risk-averse decision-makers. We choose the simplest system that generates useful forecasts — i.e., encouraging explainable AI (XAI) behaviour — to avoid deployment of complex black-box (e.g., deep-learning) systems whose behaviour may be surprising and therefore less trustworthy. We present preliminary evaluation of performance of CLYFAR version 1.0.0, and lay out a roadmap of development in coming projects.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner