Empirical Analysis of Rainfall Prediction System using an DGRNN-SAM Approach

Precipitation forecasting, especially for shorter timescales, is a critical issue in meteorological service. Most current research aims to improve prediction methods by using radar or satellite imagery. On the other hand, there is a situation in which numerous sensors at different observation sites acquire the same set of weather features. There is room for improvement in existing work due to the fact that imperfect site observations might still yield useful information for weather prediction at neighboring locations. This study proposes a multitask neural network model to automatically extract characteristics from time series gathered at observation sites and to utilize the correlation between different sites for weather forecasting, thus resolving the aforementioned problem. As far as currently aware, this is the first attempt to use multi-task learning and deep learning approaches to predict the amount of short-term rainfall based on characteristics from many places. In particular, it recast the task as an end-to-end multi-site neural network model, which allows to mimic inter-site correlations and transfer learning from one site to another. The method that has been suggested begins with the preprocessing step of cleaning and normalizing the data. When training the model with DGRNN-SAM, the process of feature selection and evaluation is carried out utilizing differential evaluation. When contrasted with other two models, such as LSTM and GRU, the suggested method fares quite well in terms of its performance.