Weather and climate forecasting with neural networks: using general circulation models (GCMs) with different complexity as a study ground

Abstract

Abstract. Recently, there has been growing interest in the possibility of using neural networks for both weather forecasting and the generation of climate datasets. We use a bottom–up approach for assessing whether it should, in principle, be possible to do this. We use the relatively simple general circulation models (GCMs) PUMA and PLASIM as a simplified reality on which we train deep neural networks, which we then use for predicting the model weather at lead times of a few days. We specifically assess how the complexity of the climate model affects the neural network's forecast skill and how dependent the skill is on the length of the provided training period. Additionally, we show that using the neural networks to reproduce the climate of general circulation models including a seasonal cycle remains challenging – in contrast to earlier promising results on a model without seasonal cycle.

Can Machines Learn to Predict Weather? Using Deep Learning to Predict Gridded 500‐hPa Geopotential Height From Historical Weather Data

“…in favor of the fully convolutional architecture. Furthermore the success of this type of architecture applied to gridded atmospheric fields (Larraondo et al, 2019;Scher, 2018;Scher & Messori, 2019) further validates its use. We performed limited validation of the CNNs using varying numbers of convolutional layers and convolutional filter stencil sizes (and dilation) before obtaining this architecture.…”

“…However, the "weather" generated by their GCM was idealized in comparison to the real world, because processes like chaotic upscale error growth and factors like seasonality were not included. An extension of this work, which applied CNNs to GCM output including seasons and at higher horizontal resolution, showed a more complicated story (Scher & Messori, 2019). The CNNs performed slightly worse on model simulations including seasons, while their performance was more severely degraded on higher-resolution input, albeit more due to the complexity of the resolved weather than the computational cost of increasing the number of grid points.…”

“…Adding more inputs generally requires more training data, and it remains to be seen how much improvement can be obtained when adding additional fields while training with reliable reanalysis data sets that only go back about 60 years. One indication that this data record could be enough for substantial progress is provided by Scher and Messori (2019) who suggested that there is little benefit to using more than 100 years of GCM data to train CNNs to forecast the GCM analog atmospheres. Important improvements might also be obtained by refining the DLWP architecture.…”

“…Here, we approach ML weather prediction head-on by developing models that use CNNs to learn to predict 500-hPa geopotential heights and 700-to 300-hPa thickness from 24 years of atmospheric reanalysis on a 2.5 • grid over the Northern Hemisphere. We improve upon the prior work of Dueben and Bauer (2018) by using more advanced NN architectures, more years of reanalysis data, and higher resolution (albeit over the Northern Hemisphere only), while in contrast to Scher (2018) and Scher and Messori (2019), we predict observed weather instead of idealized GCM forecast states. Not surprisingly, our ML models do not compare in forecast accuracy to current operational NWP models, which have been refined by decades of research, operate at much higher resolution, and use far more data to describe the initial condition for each forecast.…”

Optimization of Deep Learning Precipitation Models Using Categorical Binary Metrics

“…It is increasingly common in meteorology to use machine learning approaches for identifying patterns in the atmosphere using large amounts of historical data (Dueben & Bauer, ; Scher & Messori, ; Ukkonen & Mäkelä, ; Weyn et al, ). This approach, of extracting the underlying physical relationships in the atmosphere from data, opens an opportunity to explore new algorithms that optimize the output based on different verification metrics.…”