Multimodel AI Prediction Network (MAPNet) Applied to Temperature Forecasts

Abstract Accurate forecasts of maximum and minimum temperatures play a fundamental role in the planning of daily activities, strategic decision-making, and mitigation of impacts from extreme weather events. In this context, the development of more advanced prediction systems becomes crucial to enhance the reliability and accuracy of temperature forecasts. Therefore, this study aims to develop a multimodel (MM) system using a convolutional neural network (CNN) approach with the U-Net architecture to predict daily maximum and minimum temperatures, hereinafter referred to as the MM AI Prediction Network (MAPNet). MAPNet utilizes temperature forecasts (24, 48, and 72 h) from atmospheric models as input data and observed temperature products as labels, both integrated and made available operationally by the National Institute for Space Research (INPE). The period used for the study ranged from January 2022 to December 2023. The final configuration of MAPNet was obtained through extensive hyperparameter exploration using an automation framework. In addition to comparing with the models used as inputs, MAPNet results were also evaluated against a statistical MM, also operationally integrated by INPE. The results indicate that combining forecasts from various atmospheric models using a CNN with U-Net architecture significantly reduces inherent errors in the atmospheric models and, consequently, substantially improves the quality of forecasts. Furthermore, the results of this study have demonstrated that MAPNet not only presents improvements compared to the statistical MM but also requires a significantly lower amount of processing resources during its execution, suggesting that it may be considered a potentially effective substitute for the statistical MM.