Frost forecast - a practice of machine learning from data
Liya Ding, Yosuke Tamura, Kosuke Noborio, Kazuki Shibuya · International Journal of Reasoning-based Intelligent Systems · 2021
Among the efforts in frost forecast using machine learning techniques, a well-adopted method is to first apply time series forecast for the lowest temperature at future time points, such as the next a few days, and then apply predictive model to predict the event of frost at these time points using corresponding temperature forecasted. According to the domain understanding, there exists some 'cause-effect' between environment factors, including temperature and others, and the occurrence of frost in a few hours' period. A new modelling concept has been proposed by Ding et al. to capture such cause-effect. Preliminary experiments showed encouraging results with a sample of minute-level sensor data collected in Ikuta campus of Meiji University. In this article, as a continuation of the work, we shall further discuss methods of modelling, including causal models and associative models, and propose a framework of hybrid system in supporting frost forecast of short-term (e.g., a few hours) as well as that of relatively longer periods (e.g., a few days). More experiments are provided, and the issues of performance evaluation are discussed.