A machine-learning approach to thunderstorm forecasting through post-processing of simulation data

TL;DR

This work tackles the challenge of reliable thunderstorm forecasting by post-processing high-resolution ensemble forecasts with a neural-network named SALAMA. SALAMA ingests 21 physically related NWP predictors to output a pixelwise thunderstorm probability, calibrated to climatology and trained on simulated data tied to LINET lightning observations. It demonstrates superior skill to a reflectivity-based baseline across lead times up to $11\ \mathrm{h}$ and reveals how prediction skill scales with the spatiotemporal label resolution and the intrinsic NWP spread. The findings suggest operational applicability with real-time data and highlight the importance of scale and ensemble uncertainty in shaping forecast reliability and usefulness.

Abstract

Thunderstorms pose a major hazard to society and economy, which calls for reliable thunderstorm forecasts. In this work, we introduce a Signature-based Approach of identifying Lightning Activity using MAchine learning (SALAMA), a feedforward neural network model for identifying thunderstorm occurrence in numerical weather prediction (NWP) data. The model is trained on convection-resolving ensemble forecasts over Central Europe and lightning observations. Given only a set of pixel-wise input parameters that are extracted from NWP data and related to thunderstorm development, SALAMA infers the probability of thunderstorm occurrence in a reliably calibrated manner. For lead times up to eleven hours, we find a forecast skill superior to classification based only on NWP reflectivity. Varying the spatiotemporal criteria by which we associate lightning observations with NWP data, we show that the time scale for skillful thunderstorm predictions increases linearly with the spatial scale of the forecast.

Figures (11)

• Figure 1: Days (from 8 UTC to 8 UTC) during the summer of 2021 which were used for compiling the datasets for training (dark brown), testing (light blue with bold numerals) and validation (light green). The days have been distributed at random among the three sets.
• Figure 2: (Color online) The architecture of SALAMA: Input features are scaled to order 1. We use rectified linear units as activation functions in the hidden layers. A sigmoid function maps the output layer to the open interval $(0,1)$.
• Figure 3: Reliability diagram of SALAMA, evaluated for the test set with the label configuration $\Delta r = 15km, \Delta t = 30min$ (\ref{['sec:lightning']}). (a) Calibration curve after applying probability correction \ref{['eq:probability_test']} (black solid line), and before (grey light dotted line), and histogram of examples per bin. Perfect reliability is indicated by a dashed diagonal. Shaded band corresponds to the symmetric 90% confidence interval obtained by 200.0 bootstrap resamples. (b) Bin-wise resolution and reliability (eq:reseq:rel) and their relation to the Brier skill score (BSS, \ref{['sec:skill_scores']}) as a function of model probability.
• Figure 4: Training of the baseline model. (a) Reliability diagram panels as in \ref{['fig:reliability_diagram']}, but for the baseline model. (b) Learned relationship between the baseline input field and the corresponding probability of thunderstorm occurrence.
• Figure 5: (Color online) Probability of thunderstorm occurrence for June 23, 2021 from 19 UTC on, for SALAMA (upper row) and the baseline model (lower row). The model lead times for the three hours are 1h, 2h, and 0h, respectively. The color maps display the result for the first ensemble member of ICON-D2-EPS, while lightning labels ($\Delta r = 15km, \Delta t = 30min$, \ref{['sec:lightning']}) are shown as black contours. A jump in the color maps indicates the decision thresholds used for the evaluation of the skill scores in \ref{['fig:identification_skills']}.