纇神經網路結合衛星影像預報颱風降雨量

Abstract

本研究以西元2000~2004年蒐集之23場颱風事件為例，架構淡水氣象站颱風時期下一時刻之定量降雨預報模式。首先藉由不同之輸入項組合以建構三種不同之輸入模型，輸入變數包含地面氣象站所測得的時雨量資料、颱風特性資料以及GMS-5衛星資料，透過複迴歸分析與BPNN二種模式預報淡水氣象站下一時刻的颱風降雨量，其結果以使用GMS-5三個紅外光波段相對於淡水氣象站的9格雲頂溫度及雨量資料，作為模式之輸入項目，所得結果最佳，模式改善率最大。研究再針對27個雲頂溫度以不同的處理方式縮減變數個數，例如平均值法及主成分分析法等，進一步提出二種不同之輸入模型組合，透過複迴歸分析、BPNN及RBFNN三種模式預報下一時刻颱風降雨量，其測試結果以利用主成份分析縮減變數之方案表現最佳，其中以RBFNN模式最好(相關係數0.51、RMSE值4.74mm)，相較於僅以雨量做為輸入項之方案有8.477%之改善，並且RBFNN所建立之颱風降雨預報模式更是明顯優於複迴歸模式12.45%。

Rainfall forecast is very important for improving the efficient management of water resources systems. Nevertheless, accurate rainfall forecasting is still a great challenge faced by hydrologists. In this study, a station-based rainfall forecast model is constructed to forecast one-hour-ahead rainfall values during typhoon events. The developed model is constructed based on artificial neural networks (ANN) techniques which are capable of handle complex and non-linear systems. The available data are constituted by hourly rainfall values from 23 different events observed at the DanShui observation station and GMS-5 remote sensed data collected during 2000 to 2004. Firstly, to investigate the influence of the input information, three different schemes (schemes I, II and III) are proposed based on hourly rainfall, characteristics of typhoon and GMS-5 remote sensed data , respectively, and then applied to two different models, backpropagation neural network (BPNN) and multiple regression method (MRM) . The results showed that the BPNN model with scheme III, which includes nine cloud-top-temperatures of three thermal infrared and hourly rainfall measured data, presented the best performance. Furthermore, we have processed the input data reduction by two methods, (a) the average method and (b) the principal component analysis, and investigated their effectiveness through by three models-BPNN, MRM and Radial Basis Function Neural Network (RBFNN). The results suggest that the RBFNN model with input data reduction by the principle component analysis presented the best performance with smallest root mean square error (RMSE=4.74mm) and highest correlation coefficient (CC=0.51) when compared to all investigated schemes and forecast models.