Combining distribution‐based neural networks to predict weather forecast probabilities. (23rd October 2021)

• Record Type: Journal Article
• Authors: Clare, Mariana C.A.; Jamil, Omar; Morcrette, Cyril J.
• Is Part Of: Quarterly journal of the Royal Meteorological Society. Volume 147:Number 741(2021)
• Page Start: 4337
• Page End: 4357
• Publication Date: 2021-10-23
• Subjects: data exploration -- deep learning -- ensemble dropout -- probabilistic weather forecasting -- probability density functions -- ResNet -- stacked neural network; Meteorology -- Periodicals; 551.5
• DOI: 10.1002/qj.4180
• Languages: English
• ISSNs: 0035-9009

Abstract

Abstract: The success of deep learning techniques over the last decades has opened up a new avenue of research for weather forecasting. Here, we take the novel approach of using a neural network to predict full probability density functions at each point in space and time rather than a single output value, thus producing a probabilistic weather forecast. This enables the calculation of both uncertainty and skill metrics for the neural network predictions, and overcomes the common difficulty of inferring uncertainty from these predictions. This approach is data‐driven and the neural network is trained on the WeatherBench dataset (processed ERA5 data) to forecast geopotential and temperature 3 and 5 days ahead. Data exploration leads to the identification of the most important input variables. In order to increase computational efficiency, several neural networks are trained on small subsets of these variables. The outputs are then combined through a stacked neural network, the first time such a technique has been applied to weather data. Our approach is found to be more accurate than some coarse numerical weather prediction models and as accurate as more complex alternative neural networks, with the added benefit of providing key probabilistic information necessary for making informed weather forecasts. Abstract : An example probabilistic forecast generated by the neural network. Using probability contours, the figure shows the cumulative distribution functions for three different thresholds for the temperature at the 850 hPa level (T850) from a 3‐day hindcast at 0000 UTC 17 October 2017 (during storm O p h e l i a ). These probability contours indicate the locations where the probability of T850 being lower than 263.15 K (blue), 273.15 K (green) and 283.15 K (red) is 10, 50 and 90%.