Combining TITAN and LSTM Model in a New Tool of Nowcasting

This study presents a new weather radar nowcasting methodology that combines TITAN (Thunderstorm Identification, Tracking, Analysis and Nowcasting) scheme and an artificial intelligence procedure known as Long Short-Term Memory (LSTM) model. TITAN is a worldwide nowcasting algorithm that uses radar reflectivity factor and vertical liquid water content fields to advect rainy systems in time and to extract the life cycle of rainy systems. On the other hand, LSTM recurrent neural network is a type of network capable of learning long-term dependencies in data using memory cells, which is useful for modeling and forecasting time series.
For the proposed methodology, TITAN has been used to extract the temporal evolution of different properties of the rainstorms (area, rain volume, mean and maximum precipitation, major and minor precipitation radius, velocity and storm orientation), and LSTM recurrent neural network is employed to parametrize the temporal evolution of the rainstorms. This model consists of two LSTM layers followed by a layer that interprets the results obtained by the LSTM layers. The network was trained with 200 nodes and the loss was calculated by mean squared error (MSE). Moreover, we used the Adaptive Moment Estimation (ADAM) algorithm as an optimizer of the model with a learning rate of 0.001.
To train the LSTM model we have used the life cycle of 388 rainstorms that were extracted with TITAN that was configured to use radar measurements from the Dual Polarization S-band Doppler weather radar from FCTH/DAEE - São Paulo, Brazil, during the period of 2016 to 2019. Next, we have investigated the best parameter configuration (area, rain volume, mean and maximum precipitation, major and minor precipitation radius, velocity and storm orientation) that presents the lowest nowcasting error. Basically, we grouped 9 combinations of parameters and analyzed the performance as a function of nowcasting time (5, 10, 15, 20, 25 and 30 min) and input time intervals (starting at 10 minutes and going up to 50 minutes with 5 minute intervals). Preliminary results revealed: a) as the input time interval increases the forecasting error decreases; b) the longer the forecasting time the higher the root mean square error (RMSE). As a result, we identified two combinations that presented errors lower than the average error at each forecasting time: the first combination uses rain area, rain volume and major and minor precipitation radius and it presented a maximum RMSE of36km2 and minimum of 5km2; the second combines the first combination model with the velocity and propagation direction of the storm and it presented a maximum and minimum RMSE of 38.3km2 and 4.7km2, respectively. The two proposed LSTM models present acceptable performance in terms of forecast accuracy in most of the storms analyzed. For the conference, we will also present an analysis of the performance of the LSTM model with the results forecasted by TITAN.