A MACHINE-LEARNING APPROACH TO MODELING AND FORECASTING THE MINIMUM TEMPERATURE AT DHAHRAN, SAUDI-ARABIA

We investigate the use of modern machine-learning techniques for weath er prediction. The AIM double dagger (Abductory Induction Mechanism) t ool for the Macintosh computer has been used for modelling and 3-day f orecasting of the minimum temperature in the Dhahran region. Compared to other statistical methods and neural network techniques, this appro ach has the advantages that model synthesis is much faster and highly automated, requiring little or no user intervention. The resulting mod els are simpler, requiring data for fewer weather parameters, and pred iction and forecasting accuracies are also superior. AIM models were d eveloped using daily data for 18 weather parameters over one year to p redict the minimum temperature from other parameters on the same day. Evaluated by using data for another full year, the models give over 99 % yield in the +/-3-degrees-C error category compared to approximately 67% for both a statistical model and a model based on back propagatio n artificial neural networks. AIM forecasting models give a correspond ing yield of 93% for the first forecasting day. Model relationships de scribing the synthesized AIM networks are compared with those derived through a statistical model previously developed for the region. The e ffect of increasing the model complexity is investigated for both mode lling and forecasting.