How can we predict the accuracy of weather forecast with machine learning?

Abstract

A better weather forecast (WF) is of public interest. Over the past decades, the precision of WFs has been significantly improved by developments of more advanced forecasting techniques with state-of-the-art numerical models and machine learning. However, validation results of such precision hardly go public, and thus individual WF user tends to evaluate the accuracy of every WF through his ‘rules of thumb’. Predictive values of a forecast inevitably vary with changes in temporal-spatial dynamics of the atmosphere which is simulated by a set of numerical models that have both strength and weakness in their performances. Largely using web-scraping, we examine the causality of simulated atmospheric dynamics (inputs) and predictive values (outputs) through deep learning, in order to evaluate the accuracy of WFs in a daily basis. It is anticipated that our evaluation results will allow users to choose better forecasting systems.