Modeling and Forecasting the Daily Maximum Temperature Using Abductive Machine Learning

Abstract

The abductory induction mechanism (AIM?) is a modern machine-learning modeling tool that draws from the fields of neural networks, abductive networks, and multiple regression analysis. This paper introduces AIM as a useful weather modeling and forecasting utility and reports on its use with daily maximum temperatures in Dhahran, Saudi Arabia. Compared with other statistical methods and neural network techniques, this approach has the advantages of faster and highly automated model synthesis as well as improved prediction and forecasting accuracies. AIM models developed using daily data for 18 weather parameters over 1 yr that were used to predict the maximum temperature on a given day from other parameters on the same day. Evaluated on data for another full year, these models give 97% yield in the ±3°C error category. Various models for 3-day forecasting have been developed and evaluated. First-day forecasts give 77% yield in the same error category, and they compare favorably with official forecasts for the region, particularly for the warm seasons, as well as with forecasts based on persistence and climatology. Model relationships and performance statistics are compared with those previously obtained for the minimum temperature. The effect of increasing the AIM model complexity is investigated for both modeling and forecasting.