9:00 AM - 9:15 AM
[MGI35-01] Development of the 3-D Nowcast Method Using LSTM with Dual Polarization Phased Array Radar observations
Keywords:Convolutional LSTM, nowcast
In order to predict the movement of rain clouds in a short period, the 3-D advection nowcast model has been developed, which predicts a three-dimensional space at 30-sec intervals along with changes in rain clouds observed by a meteorological radar. However, this model is based on advection vectors and its accuracy decreases rapidly as time passes. This study aims to improve the prediction accuracy with lead times longer than 10 minutes by using a deep neural network based on convolutional Long Short-Term Memory (LSTM) network. The LSTM is a kind of Recurrent Neural Network (RNN).
In this study, the dataset observed by the Multi-Parameter Phased Array Weather Radar (MP-PAWR) installed at Saitama University in Saitama prefecture, is used to predict the evolution of convective rain cells during summertime. Precipitation nowcasting experiments using the convolutional LSTM are conducted for a convective rain event that occurred near Saitama Prefecture at around 15:00 JST on August 1, 2019. By using this radar, we can observe a 3-D structure of rain clouds (x:40 km, y:40 km, z:15 km) with a horizontal and vertical resolution of 100 m at every 30 seconds. The observations of radar reflectivity (Z) and prediction results by the 3-D Nowcast are input to the convolutional LSTM network that can learn sequences of data from past observations. For this, we use the LSTM to learn the development trend of rain clouds by using radar observations acquired during a period of 1 to 2 hours before predictions.
In order to verify the accuracy of the predicted results, we compare them with predictions from 3-D advection nowcast and 3-D CNN models. Correlation coefficients, mean average error, and probability of detection (POD) are used as evaluation indices. POD is a ratio of correctly forecast rain area to the observed rain area above a reflectivity threshold. We employ 37.5 dBZ and 10 dBZ as Z threshold values, which correspond to 10.1 mm h-1 and 0.2 mm h-1 in rainfall intensity, respectively.
From this evaluation, we obtained a preliminary result that indicates the POD of 0.16 by 37.5 dBZ, the POD of 0.93 by 10 dBZ, the correlation coefficient of 0.65, and the mean average error of 2.98 dB.
In this study, the dataset observed by the Multi-Parameter Phased Array Weather Radar (MP-PAWR) installed at Saitama University in Saitama prefecture, is used to predict the evolution of convective rain cells during summertime. Precipitation nowcasting experiments using the convolutional LSTM are conducted for a convective rain event that occurred near Saitama Prefecture at around 15:00 JST on August 1, 2019. By using this radar, we can observe a 3-D structure of rain clouds (x:40 km, y:40 km, z:15 km) with a horizontal and vertical resolution of 100 m at every 30 seconds. The observations of radar reflectivity (Z) and prediction results by the 3-D Nowcast are input to the convolutional LSTM network that can learn sequences of data from past observations. For this, we use the LSTM to learn the development trend of rain clouds by using radar observations acquired during a period of 1 to 2 hours before predictions.
In order to verify the accuracy of the predicted results, we compare them with predictions from 3-D advection nowcast and 3-D CNN models. Correlation coefficients, mean average error, and probability of detection (POD) are used as evaluation indices. POD is a ratio of correctly forecast rain area to the observed rain area above a reflectivity threshold. We employ 37.5 dBZ and 10 dBZ as Z threshold values, which correspond to 10.1 mm h-1 and 0.2 mm h-1 in rainfall intensity, respectively.
From this evaluation, we obtained a preliminary result that indicates the POD of 0.16 by 37.5 dBZ, the POD of 0.93 by 10 dBZ, the correlation coefficient of 0.65, and the mean average error of 2.98 dB.