Le Ngoc Hieu creator Sutat Weesakul Chairperson Shrestha, Sangam Examination Committee Sarawut Ninsawat Examination committee Somchai Chonwattana Examination committee (External Expert) Kanoksri Sarinnapakorn Examination committee (External Expert) AITCV Silver Anniversary Scholarships Scholarship Donor text series technical report th Pathum Thani, Thailand Asian Institute of Technology 2019 continuing eng 155 leaves : ill. (some col.) + 1 online resource Because of the complexity of the atmospheric processes that can generate precipitation with variety of different factors over a wide range of scales both in space and time, rainfall can be considered to be one of the most complicated and unprecedented factors in the hydrology cycle to understand. Therefore, meteorologists have long been developing number of mathematical models in attempt to adapt with atmospheric dynamics which is extremely complicated. For the case of Thailand, heavy rainfall often occurs in Central and Northeast parts which can possibly lead to flood. Hence, this study aims to develop machine learning as a tool to serve as an early warning system that can support Weather Research and Forecasting - Regional Ocean Modelling System (WRF - ROMs) to increase the accuracy of prediction and detect extreme events. The study first selected the best qualified stations that met three specific criteria among more than 300 telemetering stations from Hydro Informatics Institute. After that, three different points that were nearest to station's coordinates were extracted from Weather Research and Forecasting - Regional Ocean Modelling System. Next, Pearson correlation test was performed to decide input features for the model. As a result, twelve stations were selected among more than three hundred stations and Pearson correlation tests indicated high correlation between variables in telemetering stations and low or no correlation between telemetering stations and WRF - ROMs model's outputs. Therefore, input features were included both station and WRF - ROMs data. Machine learning model for rainfall forecast was built using the concept of Decision tree with Adaptive boosting (Decision tree with Adaboost). Precipitation was forecasted based on two different concepts which are rain and no - rain conditions and multiples levels of rain. Each concept was based on different threshold in order to replace numeric precipitation with binary features. The results indicated better performance of rain and no - rain conditions over multiple levels of rain with a few stations has excellent accuracies. Prediction of multiple levels of rain were lower in accuracy, however, the model showed its capability capturing extreme events. Since machine learning needs input to generate output, future inputs were calculated based on their lagged features. Feature inputs were station temperature, humidity and pressure. Lagged temperature, humidity and pressure were the inputs to calculate future temperature, humidity and pressure. Furthermore, autocorrelation test was used to determine to best lagged hour that had the highest correlation to the current feature. To forecast future inputs, machine learning with concepts of Decision tree with Adaboost and Polynomial regression were used. As a result, lagged two hours of each feature has better correlation to the current feature than lagged three hours and onwards. Predicting temperature, humidity and pressure in two hours ahead had higher confidence than predicting them in three hours ahead. A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Water Engineering and Management Thesis (M.Eng.) - Asian Institute of Technology, 2019 Machine learning Development Rain and rainfall Forecasting Monsoons Thailand Asian Institute of Technology. http://203.159.5.9/ait-thesis/detail.php?q=B06380 http://203.159.5.9/ait-thesis/detail.php?q=B06380 200907 20260818090534.0