Development of a Prediction Method for Weather Types using a Deep Neural Network

Accurate weather forecasting is required since it is closely related to our daily lives and economic activities. The Japan Meteorological Agency (JMA) has been operating various kinds of “guidance” forecasts to correct biases in numerical weather prediction (NWP) outputs or to predict the elements that NWP models do not directly predict. The weather type guidance is one of the operating products, which predicts 3-hourly weather categories: fair, cloudy, rainy, sleety, and snowy. It determines the weather categories through a flowchart that uses sunshine duration, temperature, humidity, and precipitation.

We have been developing a deep-neural-network-based (DNN-based) weather guidance system that predicts weather categories at each grid point using NWP outputs. The explanatory variables are temperature at four vertical levels, humidity at seven vertical levels, and precipitation. The objective variable is weather types extracted from the estimated weather distribution products of the JMA, which is generated from geostationary satellite imagery, Radar/Rain gauge-Analyzed precipitation, estimated surface temperatures, and NWP outputs.

We evaluated the DNN-based weather type guidance using the weather types based on surface observations independent of the estimated weather distribution products or the DNN targets. We confirmed that the DNN-based weather type guidance outperformed the operational weather type guidance. In particular, the accuracy of fair/cloudy discrimination was improved in the cases dominated by upper-level stratiform clouds.

In this presentation, we introduce the prediction method and verification results of the DNN-based weather type guidance.