Deep Learning for Day Forecasts from Sparse Observations

TL;DR

This work introduces MetNet-3, an observation-based neural weather forecast model that extends lead times to 24 hours and predicts precipitation, temperature, dew point, and wind at high spatiotemporal resolution. It leverages a novel densification technique to convert sparse ground observations into dense forecasts, enabling accurate predictions across CONUS with minimal latency. MetNet-3 outperforms state-of-the-art probabilistic NWPs (ENS/HREF) for both instantaneous and hourly precipitation and for surface variables, validated on hold-out stations and case studies. The model’s architecture combines topographical embeddings, a U-Net backbone, and a MaxViT transformer, trained with a lead-time conditioning scheme and extensive multi-task losses, demonstrating a scalable, real-time observation-based alternative to traditional NWP.

Abstract

Deep neural networks offer an alternative paradigm for modeling weather conditions. The ability of neural models to make a prediction in less than a second once the data is available and to do so with very high temporal and spatial resolution, and the ability to learn directly from atmospheric observations, are just some of these models' unique advantages. Neural models trained using atmospheric observations, the highest fidelity and lowest latency data, have to date achieved good performance only up to twelve hours of lead time when compared with state-of-the-art probabilistic Numerical Weather Prediction models and only for the sole variable of precipitation. In this paper, we present MetNet-3 that extends significantly both the lead time range and the variables that an observation based neural model can predict well. MetNet-3 learns from both dense and sparse data sensors and makes predictions up to 24 hours ahead for precipitation, wind, temperature and dew point. MetNet-3 introduces a key densification technique that implicitly captures data assimilation and produces spatially dense forecasts in spite of the network training on extremely sparse targets. MetNet-3 has a high temporal and spatial resolution of, respectively, up to 2 minutes and 1 km as well as a low operational latency. We find that MetNet-3 is able to outperform the best single- and multi-member NWPs such as HRRR and ENS over the CONUS region for up to 24 hours ahead setting a new performance milestone for observation based neural models. MetNet-3 is operational and its forecasts are served in Google Search in conjunction with other models.

Figures (11)

• Figure 1: Abstract depiction of densification aspects. (a) During training a fraction of the weather stations are masked out from the input, while kept in the target. (b) To evaluate generalization to untrained locations, a set of weather stations represented by squares is never trained on and only used for evaluation. (c) To evaluate forecasts for the sparse locations for which data is available, these stations are fed as input during the evaluation as well. (d) The final forecasts uses the full set of training weather stations as input, and produces fully dense forecasts aided by spatial parameter sharing.
• Figure 2: Case study for Sat Apr 23 2022 12:00 UTC featuring the Rocky Mountains of Colorado showing the mean of the ENS and MetNet-3 6 hour wind speed forecasts (top, left and center) along with the OMO stations ground truth (top, right) and the error of ENS and MetNet-3 on the individual weather stations (bottom). Circles and squares denote, respectively, training and test stations with MAEs calculated on both training and test stations. This example shows MetNet-3's ability to densify the targets, the higher spatial resolution of MetNet-3 as well as forecast precision on the weather stations.
• Figure 3: An example of a precipitation rate distributions from MetNet-3 forecasts for a single location for different lead times. A black colored bar indicates MRMS precipitation ground truth rate.
• Figure 4: Comparison of basic characteristics of physics-based baselines used in this work and MetNet-3. MetNet-3 forecasts precipitation at 1 km / 2 min resolution and ground variables at 4 km / 5 min resolution. MetNet-3 can be run more frequently than NWP models because running the model is almost instant (about 1s for a single lead time) and requires fewer computational resources than NWP models.
• Figure 5: Performance comparison between the probabilistic MetNet-3 and NWP baselines for instantaneous precipitation rate on CRPS (lower is better) and the categorical CSI (higher is better); CRPS includes all precipitation rates, whereas the CSI plots are for light (1 mm/h), moderate (4 mm/h) and heavy (8 mm/h) precipitation. Deterministic baselines (HRRR and HRES) are ommited for clarity in the CRPS plot due to performing significantly worse than the probabilistic models. Note, the thresholds for turning the probabilistic forecasts of MetNet-3 and ENS into deterministic forecasts for use in the CSI calculation, have been optimized on a validation set. HREF is omitted in all plots due to instantaneous precipitation rate being unavailable. See Supplement E for the CSI plots for other rates, the CRPS plot with the deterministic baselines included and an explanation for the CRPS lines being non-monotonic.