Towards Climate Variable Prediction with Conditioned Spatio-Temporal Normalizing Flows

TL;DR

The paper develops a conditioned spatio-temporal normalizing flow (ST-Flow) for climate-variable forecasting by conditioning a memory-augmented, invertible model on a past-state representation from a GatedConvLSTM. This approach enables exact likelihood computation and uncertainty quantification for spatio-temporal climate data, and is evaluated on ERA5 datasets where ST-Flow demonstrates superior long-horizon stability compared to deterministic and stochastic baselines. A key finding is that using coarser spatial representations via conditional NF auto-encoding can stabilize long rollouts and improve efficiency, though dataset characteristics modulate performance. The work highlights the potential of invertible, memory-conditioned flows for fast, uncertainty-aware climate scenario exploration and underscores directions toward physics-informed calibration for improved distributional fidelity.

Abstract

This study investigates how conditional normalizing flows can be applied to remote sensing data products in climate science for spatio-temporal prediction. The method is chosen due to its desired properties such as exact likelihood computation, predictive uncertainty estimation and efficient inference and sampling which facilitates faster exploration of climate scenarios. Experimental findings reveal that the conditioned spatio-temporal flow surpasses both deterministic and stochastic baselines in prolonged rollout scenarios. It exhibits stable extrapolation beyond the training time horizon for extended rollout durations. These findings contribute valuable insights to the field of spatio-temporal modeling, with potential applications spanning diverse scientific disciplines.

Figures (10)

• Figure 1: ST-Flow architecture with K flow steps and L scales. For further explanation, see Appendix A.
• Figure 2: Visualization of rollout trajectories starting from the same initial conditions on the ERA5 temperature and geopotential dataset from the conditional normalizing flow model. The last row shows the squared absolute error. For rollouts from other methods, see appendix.
• Figure 3: Dataset summary with their resolutions, time spans, and training/validation/test samples.
• Figure 4: RMSE curves over 100 test samples on the ERA5 daily temperature and hourly geopotential dataset computed for different models. The vertical line indicates the length of the context window size during training, which we set to 2.
• Figure 5: RMSE curves over 100 test samples on the ERA5 daily temperature and hourly geopotential dataset for different input resolutions of the data (orig,4x,8x,16x). The vertical line indicates the length of the context window size during training, which we set to 2.