Using of neural network to precipitation nowcasting based on weather radar data

Up to now neural networks have been used in radar meteorology to estimate measurements of precipitation. Next step is to take advantage of them in nowcasting. Preliminary version of model based on simple neural network’s structure will be presented. Precipitation nowcasting on a basis of weather radar data gives satisfying results in timescale of few hours ahead, especially in the cases of air advection. Convective precipitation is forecasted applying knowledge of physics of atmospherical processes, using radar data as input, whereas nowcasting with neural network is based on categorisation of various historical situations. It can give a possibility for improvements of short-term forecasting efficiency. A model’s tests were conducted on data from weather radar in Katowice derived every 15 min. and simplified to resolution 4x4 km. The full radar image size is 400x400 km. Different meteorological parameters are also included to the model. In the preliminary version the prepared model is multilayer with back-propagation. It includes three layers (one hidden). Process of training of this model is slow but efficient. The number of neurones in the hidden layer is chosen on the basis of wanted precision. Results of testing point out typical disadvantages of neural network like limited possibilities of training. Nevertheless it can be hope for more effective nowcasting.