Rainfall Amount Forecast Using GNSS-PWV Based on Machine Learning Fusion Strategy and the Constraint of Rainfall Event
Mingkun Su, Cong Chen, Zhao Li, Weiping Jiang, Yang Gao, Junna Shang, Xingyu Zhou
Abstract
Accurate rainfall forecasting plays a crucial role in weather monitoring. Currently, the application of Global Navigation Satellite System-derived precipitable water vapor (GNSS-PWV) has mainly focused on forecasting rainfall event occurrence, while neglecting the forecasting of rainfall amount. In this study, a new method based on machine learning fusion strategy and the constraint of rainfall event is proposed. The machine learning fusion strategy is used to improve the accuracy of rainfall amount forecasting by considering the difference in rainfall types and machine learning algorithms, while the rainfall event constraint strategy is used to reduce the rainfall amount forecasting error during periods without rainfall event. In the new method, the Long Short-Term Memory (LSTM) algorithm is adopted to forecast rainfall event by considering the temporal correlation of rainfall event. Then, the Support Vector Regression (SVR), Group Method of Data Handling (GMDH), and Harmony search (HS) algorithms are combined to forecast rainfall amount with the constraint of rainfall event forecast results. The SVR performs better at forecasting little and medium rainfall, while GMDH is better at forecasting heavy rainfall. HS is mainly used to optimize the parameters of the forecasting model. The data sets collected from 2019 to 2023 at Hong Kong GNSS stations are used to evaluate the performance of the proposed method. The experimental results show that after adopting the proposed rainfall event forecasting method, the total True Positive Rate (TPR) is approximately 90.45% and the prediction errors for without rainfall event is only about 4.09%. Moreover, after adopting the proposed rainfall amount forecasting method, the average Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) are approximately 0.42 mm and 0.13 mm, respectively. The average improvement rate of MAE is approximately 24.39% for the internal experiment and 25.27% for the external experiment compared to the single SVR model. In conclusion, the proposed method can effectively improve the accuracy of rainfall amount forecast, which can further provide support for meteorological monitoring.