Publication: Learning based model for local weather nowcasting
Loading...
Date
2024-07-05
Authors
Huen, Jing Lum
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Accurate and localized weather nowcasting can provide important and timely information to the agriculture sector and extreme weather conditions for decision making and necessary mitigation to reduce losses. This thesis presents a learning-based model for local weather forecasting which focuses on rainfall prediction. The research addresses the limitations of current nowcasting services which were generally lack localized specificity, which would eventually lead to potentially misleading information and poor decision-making. This thesis explores the application of various machine learning models, including both regression and classification approaches for rainfall nowcasting. The modelling was subjected to multivariate single source numerical weather data from single point weather station at Bukit Mertajam and Cameron Higlhands, Malyaisa. For standardization, a general weather nowcasting framework was purposed. It involves comprehensive data preprocessing, data analysis, feature engineering, and the implementation of machine learning models to predict rainfall events. The models include K-Nearest Neighbors (KNN), Random Forest, Support Vector Machine (SVM), Decision Tree, Extreme Gradient Boosting (XGBoost), and the ARIMA model serving as a benchmark due to its popularity in time series forecasting. The experimental results demonstrate that single-step predictions outperform multi-step predictions. Random Forest regressor achieved the best performance with Probability of Detection (POD) of 0.575, False Alarm Ratio (FAR) of 0.312 and True Skill Statistic (TSS) of 0.326. The study identified key challenges in rainfall nowcasting using traditional machine learning models, including limited spatial and temporal resolution of weather data, lack of contextual data, issues with data imbalance and rare events and the limitations of the model itself.