MUMBAI, India, Aug. 31 -- Intellectual Property India has published a patent application (202641100484 A) filed by Dr Subash Chandra Bose S; Dr Kohila N; and Dr. P. V. Praveen Sundar on August 20, 2026, for Adaptive Graph Neural Network For Spatiotemporal Local Rainfall Prediction.
Inventors include Dr Subash Chandra Bose S; Dr Kohila N; and Dr. P. V. Praveen Sundar.
The application for the patent was published on August 28, 2026, under issue no. 35/2026.
Abstract: Adaptive graph neural network for spatiotemporal local rainfall prediction ABSTRACT The present invention relates to an adaptive deep learning system for accurate local rainfall prediction using dynamically evolving atmospheric spatial graphs. The system represents weather stations, satellite grid points, and IoT sensors as graph nodes, with dynamically reconfigurable edges determined from wind velocity vectors, pressure gradients, and learned meteorological interactions. A complementary graph neural network predicts inter-node impact weights to control real-time edge rewiring. Multi-modal inputs including radar reflectivity, soil moisture, sea surface temperature, and orographic elevation are fused for atmospheric representation. A dual-attention transformer encoder captures hourly and sub-hourly temporal dependencies, while a spatial attention module characterizes mesoscale convective system dynamics. A multi-task prediction head simultaneously predicts rainfall magnitude, onset time, and spatial extent, enabling improved precipitation forecasting over complex terrain where conventional numerical weather prediction models inadequately represent nonlinear topographic effects.
Disclaimer: Curated by HT Syndication.
