Source-linked AI summary
Analog forecasting of extreme-causing weather patterns using deep learning
Ashesh Chattopadhyay, Ebrahim Nabizadeh, Pedram Hassanzadeh
TL;DR
Extreme weather remains difficult to forecast with computationally demanding NWP models, motivating a data-driven alternative. The paper combines analog forecasting, CapsNets, and impact-based auto-labeling to predict North American temperature-extreme occurrence and region from earlier circulation patterns. CapsNets achieve useful multi-day skill from Z500 alone, improve substantially with Z500 and T2m, and outperform simpler methods.
Problem
Forecasting atmospheric circulation patterns associated with heat waves and cold spells remains difficult despite advances in computationally demanding NWP models.
Method
The framework uses analog forecasting with CapsNets trained on impact-based labels that connect earlier Z500 patterns to later North American T2m extremes.
Results
CapsNet Z500-only predictions reach 68.8%–45.1% winter accuracy and 61.6%–40.6% summer accuracy across 1–5-day leads, while Z500+T2m raises accuracies to 82.0%–76.7% in winter and 79.3%–75.8% in summer.
Takeaways & Limitations
The framework provides a proof of concept for predicting extreme-causing patterns and could support early warnings and NWP efforts.
Takeaways & Limitations
Extending the framework to planetary scales and prediction leads beyond five days requires architectures that account for Earth’s spherical geometry and remains future work.
Abstract
from arXiv · showhide
Numerical weather prediction (NWP) models require ever-growing computing time/resources, but still, have difficulties with predicting weather extremes. Here we introduce a data-driven framework that is based on analog forecasting (prediction using past similar patterns) and employs a novel deep learning pattern-recognition technique (capsule neural networks, CapsNets) and impact-based auto-labeling strategy. CapsNets are trained on mid-tropospheric large-scale circulation patterns (Z500) labeled $0-4$ depending on the existence and geographical region of surface temperature extremes over North America several days ahead. The trained networks predict the occurrence/region of cold or heat waves, only using Z500, with accuracies (recalls) of $69\%-45\%$ $(77\%-48\%)$ or $62\%-41\%$ $(73\%-47\%)$ $1-5$ days ahead. CapsNets outperform simpler techniques such as convolutional neural networks and logistic regression. Using both temperature and Z500, accuracies (recalls) with CapsNets increase to $\sim 80\%$ $(88\%)$, showing the promises of multi-modal data-driven frameworks for accurate/fast extreme weather predictions, which can augment NWP efforts in providing early warnings.
1 Introduction
Extreme weather prediction remains difficult because atmospheric circulation patterns that often drive heat waves and cold spells are challenging for computationally demanding NWP models to forecast. The paper introduces deep-learning-enhanced analog forecasting as an alternative pattern-recognition framework.
- Predicting heat waves and cold spells is scientifically and societally important, yet forecasting their driving atmospheric circulation patterns remains challenging.
- NWP models require substantial computational resources and still struggle to accurately predict the formation of extreme-causing circulation patterns.
- Analog forecasting predicts weather by finding similar patterns in the past.
- The proposed framework uses deep learning to classify circulation patterns rather than relying on one near-perfect analog identified by correlation or Euclidean distance.
- The paper presents the framework's data, labeling, training, testing, results, and discussion across Sections 2–5.
2 Data
The study uses daily LENS climate-model data to define North American temperature extremes and their onsets, cluster events geographically, and construct datasets designed to reduce imbalance and temporal-correlation contamination.
- The dataset consists of daily 1920–2005 simulations from 40 CESM1 ensemble members under historical radiative forcing.
- The study uses daily surface air temperature at 2 m and 500-mb geopotential height during boreal summer and winter.
- Heat waves and cold spells require summer or winter T2m anomalies crossing specified percentile and magnitude thresholds for at least five consecutive days.
- Extreme-event onsets are the first days of qualifying heat-wave or cold-spell sequences over North America between 30°N and 60°N.
- Onset samples are selected to minimize correlations within and across training and testing sets, with a separate ensemble-member split used to test contamination by temporal correlation.
- T2m onset patterns are reduced to 22 principal components explaining over 90% of variance, then clustered by K-means into four seasonal geographical regions.
- The prediction target is a five-class outcome: no extreme event or an event in one of four geographical clusters.
3 Methodology
The methodology combines ConvNets or CapsNets with impact-based auto-labeling: future T2m impacts label earlier Z500 patterns, which are then trained and tested to predict extreme-event occurrence and region.
- Deep Learning Techniques: ConvNet and CapsNet: CapsNets and ConvNets learn feature-extraction filters through backpropagation, with CapsNets expected to better handle spatio-temporal climate patterns through equivariance.
- Impact-based auto-labeling of daily Z500 patterns: Because expert labels poorly capture complex nonlinear extreme-causing patterns over several days, the method uses impact-based auto-labeling.
- Impact-based auto-labeling of daily Z500 patterns: Earlier Z500 patterns receive labels 0–4 according to whether later T2m produces no extreme onset or an event in one of four clusters.
- Impact-based auto-labeling of daily Z500 patterns: The strategy labels one feature map, T2m, while training and testing on another, Z500, to predict the original impact feature.
- Each season and lead time uses balanced datasets with M = 1000 samples for one- and two-day leads and M = 900 for longer leads.
- Separate models are trained for seasons and lead times using Z500 alone or stacked Z500 and anomalous T2m channels.
- Four train/test divisions support validation and hyperparameter selection before final testing of CapsNets or ConvNets.
- Accuracy measures correct predictions across all five clusters, while recall measures correctly identified extreme-event samples in clusters 1–4.
4 Results
CapsNet predicts extreme-causing weather patterns from Z500, with stronger performance when anomalous T2m is added. Performance varies by season, training-set size, and model architecture, while longer lead times and broader spatial scales remain open challenges.
- Z500-only prediction: 68.8%–45.1% winter and 61.6%–40.6% summer accuracies were achieved for 1–5-day Z500-only predictions, versus 20% random chance.Corresponding recalls were 77.2%–48.1% in winter and 72.8%–46.6% in summer.
- Training-set sensitivity: Reducing training samples nearly fourfold lowered winter accuracy by at most 4.7%, but reduced summer 5-day accuracy by 9.5%.The weak overall dependence on sample size suggests that additional training data alone may not yield substantially higher accuracies.
- Seasonal differences: Summer prediction skills were lower than winter skills because summer circulation patterns were more similar and heat-wave dynamics involve more processes than the available mid-tropospheric predictor.The cited processes include horizontal advection, adiabatic and clear-sky radiative warming, and land–atmosphere feedbacks.
- Future improvements: Including variables from other tropospheric and stratospheric altitudes, such as information related to polar-vortex variability, is expected to improve prediction skills.This expectation is stated even for winter cases where meridional advection dominates.
- Multivariate prediction: 82.0%–76.7% winter and 79.3%–75.8% summer accuracies were achieved when Z500 and anomalous T2m were combined, with recalls near 87%–88%.The multivariate approach particularly improved longer lead times and reduced false alarms.
- Future improvements: Extending the framework beyond 5-day lead times requires planetary-scale inputs and architectures that account for Earth’s spherical geometry.The paper leaves planetary-scale, longer-lead predictions for future work.
- Model comparison: CapsNet accuracies exceeded ConvNet accuracies by 2.8% in winter and 0.7% in summer for Z500-only predictions, and by 7.7% and 7.1% with Z500+T2m.ConvNet performance degraded more as the training set shrank, especially in the multivariate setting.
- Model comparison: CapsNet accuracies exceeded logistic regression by 11.4% in winter and 19.6% in summer for Z500 predictions.Logistic regression was consistently less accurate than ConvNet and CapsNet.
5 Discussion
The framework demonstrates skillful prediction of extreme-causing weather patterns from limited driver information and combines impact-based labeling with CapsNets for pattern recognition. Its potential uses include early warnings, precursor identification, and extension to other high-impact phenomena, while regional class design and interpretability remain open research areas.
- Results: The framework predicts high-impact surface temperature extremes using only Z500 patterns, without direct information about the impact itself.This provides a proof-of-concept for predicting extreme-causing weather patterns from a key driver.
- Method: Impact-based auto-labeling and CapsNets together decode relationships between T2m impacts, Z500 patterns, and their temporal evolution.The approach addresses challenges including nonlinear sensitivity and the rarity of perfect analogs.
- Future directions: The multi-modal framework using Z500 and T2m may support higher accuracies and longer lead times when equipped with spherical CapsNets for planetary-scale inputs.Additional atmospheric, land, ocean, and stratospheric variables are suggested as potential inputs for improving prediction.
- Applications: The framework could provide early warnings, guide public and NWP efforts, and help identify precursors through interpretable machine learning.Training on observational data is appealing, but short records and temporal non-stationarity pose challenges.
- Applications: High-resolution NWP simulations could provide training data for a data-driven framework that yields extremely fast and inexpensive predictions.
- Limitations: The study uses five classes formed from four K-means geographical regions over North America, while alternative clusterings and regional binary models remain to be evaluated.Determining which training data and framework designs best handle practical limitations requires further research.
- Future directions: Future work should examine CapsNet feature maps and filters because precursors of extreme-causing patterns such as blocking events remain poorly understood.Interpretation could improve dynamical understanding of extremes and potentially improve weather and climate models.
- Applications: The framework can be generalized to other high-impact phenomena, including extreme precipitation and severe air pollution, although domain expertise remains important for auto-label design.
A Convolutional Neural Network (ConvNet)
The ConvNet architecture extracts spatial features through stacked convolutions and pooling before classifying input patterns. Its training uses regularization and adaptive optimization, while full-size atmospheric patterns can cause overfitting because of chaotic small-scale variability.
- Architecture: ConvNets use convolutional filters to extract features, followed by ReLU activations and pooling operations.
- Input patterns: The atmospheric Z500 inputs use 97 × 66 or 157 × 66 longitude-latitude grids for small and large domains.
- Limitations: Training on full-size patterns is challenging because chaotic, high-frequency small-scale variability leads to overfitting and inaccurate predictions.
- Architecture: The model has four convolutional layers with 8, 16, 32, and 64 filters, each using 5 × 5 kernels.
- Classification: The final feature representation feeds a 200-neuron fully connected network, with softmax assigning the input to the highest-probability cluster.Cross-entropy, L2 regularization, dropout, and ADAM are used during training.
B Capsule Neural Network (CapsNet)
CapsNets are motivated by the need to preserve relative feature positions that pooling-based ConvNets may lose. The architecture encodes features as high-dimensional capsule tensors and routes them by agreement for pattern recognition.
- Motivation: ConvNets can struggle when correct classification depends on the relative spatial positions of features.
- Motivation: CapsNets replace pooling with capsules whose vector outputs encode information such as position, distances, poses, and orientation.CapsNets seek equivariance rather than the invariance emphasized by pooling-based ConvNets.
- Weather patterns: Blocking-event signatures depend on pressure-system configurations, whereas opposite configurations often progress eastward without causing extremes.
- Architecture: The CapsNet uses two convolutional layers with 32 and 64 filters, followed by ReLU layers and a primary capsule layer.
- Architecture: The primary capsule layer contains 8 capsules with 8 convolution layers each and sends information onward through routing-by-agreement.
C Other skill metrics: Persistence and climatology
Climatology and persistence are standard forecasting baselines, but the study does not compare them because its prediction objectives do not match their applicable formulations. In particular, Z500-only prediction targets T2m impacts from circulation inputs.
- Baseline methods: Climatology forecasts a variable using its long-term average for the same calendar date, while persistence assumes the variable remains unchanged.
- Baseline methods: The study does not compare its framework with climatology or persistence because the objective cannot be formulated in terms of those baselines.
- Applicability: For Z500-only inputs, climatology and persistence cannot predict T2m clusters from Z500 patterns.
- Applicability: The study's baseline rationale depends on the distinct input-output formulation and the absence of T2m extremes within five days of an extreme onset.