A Graph Neural Network with Spatio-Temporal Attention for Multi-Sources Time Series Data: An Application to Frost...

Area of Science:

• Climate Science
• Artificial Intelligence
• Agricultural Technology

Background:

• Frost events pose significant economic risks to various industries, necessitating accurate forecasting.
• Traditional forecasting methods struggle with the complex spatio-temporal dynamics of frost incidence.
• Advanced prediction models are required to mitigate economic losses associated with frost.

Purpose of the Study:

• To develop a novel graph neural network (GNN) model for enhanced frost forecasting.
• To predict minimum temperatures and frost occurrence with improved accuracy and lead times.
• To establish a new state-of-the-art in frost prediction using spatio-temporal data analysis.

Main Methods:

• Development of GRAST-Frost, a GNN with a spatio-temporal architecture.
• Utilization of an IoT platform for real-time weather data acquisition from an experimental site.
• Integration of data from 10 proximate weather stations for comprehensive spatial coverage.
• Simultaneous processing of multiple time series data, considering spatial and temporal dependencies.

Main Results:

• GRAST-Frost outperforms classical time series forecasting methods (linear, nonlinear ML, simple DL, non-graph DL).
• The model demonstrates superior performance in predicting minimum temperatures and frost incidence.
• Predictions were successfully made for 6, 12, 24, and 48-hour lead times.
• Significant improvement over existing state-of-the-art frost forecasting techniques was achieved.

Conclusions:

• The proposed GRAST-Frost model offers a significant advancement in frost forecasting capabilities.
• The GNN approach effectively captures complex spatio-temporal weather patterns for accurate predictions.
• This technology has the potential to substantially reduce economic impacts of frost events across industries.