An Artificial Neural Network for Short Time Air Temperature Prediction

Abstract

Air temperature is an extremely important factor in agriculture, from planting to post-harvest processes, and having the ability to predict air temperature can be a valuable tool for avoiding damage, maximizing production quality, and optimizing resources. In this work, we propose a simple air temperature prediction model based on a small neural network with a relatively small volume of training data. This work uses data from the Climatology and Biogeography Laboratory of the University of São Paulo (USP), from the Experimental Meteorological Station in São Paulo City, Brazil. The dataset corresponds to air temperature data collected during the years 2018 and 2020. For machine learning, two types of artificial neural networks were adopted: one of the long short-term memory recurrent network and one feed-forward network. Three past air temperatures were used to predict the next hour’s air temperature, and chain predictions were used to predict up to 24 hours. The feed-forward neural network presented the best results, with most errors below 2°C. The results show that it is possible to use a simple neural network, using only air temperature as the meteorological variable, to predict air temperature for the next hours. The simplicity of the model makes its application more feasible for various problems in agriculture.