Winter Precipitation Type Diagnosis and Uncertainty Quantification with a Physically Consistent Machine Learning Method

TL;DR

The paper tackles the challenge of forecasting winter precipitation type with uncertainty by developing an evidential neural network that outputs four p-type probabilities plus epistemic uncertainty. It leverages quality-controlled crowd-sourced mPING observations and RAP soundings, and incorporates a physical XAI framework to ensure consistency and interpretability. Across bulk statistics, a central Plains case study, and interactive analyses, the method demonstrates calibrated probabilities and meaningful uncertainty representations, outperforming or matching traditional NWP-based and area-based approaches in key regimes. The approach supports impact-based decision making by providing a physically grounded, uncertainty-aware forecast tool and an interactive platform for exploring model behavior and limitations.

Abstract

Correctly forecasting the timing and location of changes in winter precipitation type could help decision makers mitigate the worst impacts of winter storms. Multiple precipitation type algorithms have been developed from both physical and statistical perspectives, but all of them struggle in certain scenarios, and most of them do not account for uncertainty with a single model. We developed an evidential neural network that can predict both the probability of each winter precipitation type as well as the epistemic uncertainty. We trained our model on quality controlled and curated observations from the crowd-sourced mPING dataset in conjunction with vertical profiles from the NOAA Rapid Refresh model analyses. Our static and interactive evaluation revealed that the data curation procedure resulted in meteorologically consistent forecasts and appropriately represents uncertainty in difficult regimes where predictability may be limited by the atmospheric representations of current NWP models. We compare our model to both the Rapid Refresh NWP model in addition to other thermodynamic area-based methods from June of 2020 through June of 2022 and from a High Resolution Rapid Refresh central plains case study from December 24-26, 2023.

Figures (12)

• Figure 1: Distributions of mPING observations before (blue) and after (red) quality control was performed.
• Figure 2: Conceptual diagram outlining the procedure from data origin to model output.
• Figure 3: Bulk Performance metrics for all models by precipitation type. For probabilistic models, the highest probability is considered the predicted class.
• Figure 4: Confusion matrices for all models. The top row of matrices is normalized by the mPING observations (rows), and the bottom row is normalized by the predictions (columns).
• Figure 5: Distributions of observations and model predictions. The top row has the thermodynamic energy calculated using the dry-bulb temperature inline with the original Bourgouin method, and has the derived threshold super imposed. The second row is calculated in wet-bulb space and had the modified Bourgouin method fit superimposed. The bottom row has the density estimates to more clearly show the data overlaps.