14:30 〜 14:45
[MIS04-04] Rainfall prediction in Metro Manila, Philippines using lightning data integrated machine learning method
キーワード：Rainfall prediction, Lightning, Machine learning, P-POTEKA
Metro Manila, Philippines is one of the cities that frequently experiences heavy rainfall disasters, leading to
significant economic losses and injuries. Therefore, it is essential to improve the accuracy of rainfall
forecasting and establish an early warning system crucial for disaster mitigation. Currently, numerical weather
prediction (NWP) models play a key role in weather forecasting, and some studies have suggested that
incorporating lightning information can improve prediction accuracy. However, the physical process linking
lightning and precipitation remain insufficiently understood for direct integration into NWP models.
Machine learning offers an alternative approach by learning patterns from a large number of past heavy rainfall
events. Our data analysis, based on electric field plate sensors of P-POTKEA, a system capable of obtaining
both lightning-induced electric field change and weather data, revealed that lightning frequency peaks precede
rainfall peaks by a few minutes. Additionally, lightning occurring near the plate sensor network has a higher
probability of being detected. In this presentation, I will introduce a method for reconstructing the position
and charge amount of lightning discharges using P-POTEKA plate sensor data. Furthermore, I will utilize
the SA-ConvLSTM model, which is one of the machine learning models, to capture spatiotemporal
dependencies in meteorological data. By integrating weather parameters highly correlated with precipitation,
including lightning position and neutralized charge, into the machine learning framework, I aim to improve
rainfall prediction accuracy. I will also present preliminary results from this approach.
significant economic losses and injuries. Therefore, it is essential to improve the accuracy of rainfall
forecasting and establish an early warning system crucial for disaster mitigation. Currently, numerical weather
prediction (NWP) models play a key role in weather forecasting, and some studies have suggested that
incorporating lightning information can improve prediction accuracy. However, the physical process linking
lightning and precipitation remain insufficiently understood for direct integration into NWP models.
Machine learning offers an alternative approach by learning patterns from a large number of past heavy rainfall
events. Our data analysis, based on electric field plate sensors of P-POTKEA, a system capable of obtaining
both lightning-induced electric field change and weather data, revealed that lightning frequency peaks precede
rainfall peaks by a few minutes. Additionally, lightning occurring near the plate sensor network has a higher
probability of being detected. In this presentation, I will introduce a method for reconstructing the position
and charge amount of lightning discharges using P-POTEKA plate sensor data. Furthermore, I will utilize
the SA-ConvLSTM model, which is one of the machine learning models, to capture spatiotemporal
dependencies in meteorological data. By integrating weather parameters highly correlated with precipitation,
including lightning position and neutralized charge, into the machine learning framework, I aim to improve
rainfall prediction accuracy. I will also present preliminary results from this approach.