應用長短期記憶類神經網路於雙偏極化雷達推估降雨之研究

A Study of Dual-Polarization Radar Rainfall Estimation Using Long Short-Term Memory Neural Networks

林俊良(Jun-Liang Lin)

指導教授 ： 張麗秋

淡江大學/工學院/水資源及環境工程學系碩士班/碩士(2022年)

https://doi.org/10.6846/TKU.2022.00723

摘要

近年來受到氣候變遷與全球暖化之影響，世界各地方發生極端水文事件的頻率增加，對人類生命與財產安全造成極大的威脅。臺灣降雨在時間與空間分布呈現不均現象越趨嚴重，短延時強降雨的豪雨事件造成局部地區發生積淹水，甚至嚴重淹水災情頻發，相對縮短救災整備與應變時間。近年為提升天氣預報與降雨推估能力，臺灣設立多部C波雙偏極化防災降雨雷達，本研究目的是結合雙偏極化雷達參數與人工智慧技術，以精進短延時降雨推估。。
本研究以臺北市作為研究區，利用長短期記憶類神經網路(LSTM)與倒傳遞類神經網路(BPNN)學習雷達參數特徵，建置未來1小時內之10分鐘累積雨量，雷達參數包含比差分相位(KDP)、合成回波(Z)與都卜勒徑向風場(VR)。最後分析54站雨量站雨量推估結果並探討雨量站地文與水文條件差異對於雷達參數映射雨量之關係，將山區與非山區之單站雨量站雨量推估結果比較，應用LSTM模式與BPNN模式於樹林防災降雨雷達之掃描參數進行雨量估計的效果，由R2與RMSE顯示，相較於使用R(KDP)關係式與R(Z)關係式表現較佳，R2改善率達6.90%與5.40%，RMSE改善率達19.59%與18.78%，研究區域內54站雨量推估，T時刻之R2均接近0.85，可以說明利用合成回波(Z)與比差分相位(KDP)，能有效掌握降雨趨勢，有助於提升短延時強降雨發生時之降雨估計的精確度，並且在沒有設置雨量站的區域亦可以透過該區域上方2×2網格之Z與KDP即可估計該區域當下雨量，達到區域內所有網格估計降雨效果。

關鍵字

類神經網路 ； 長短期記憶類神經網路 ； 倒傳遞類神經網路 ； 雙偏極化雷達 ； 降雨推估

並列摘要

Extreme hydrological events have become more common in recent years as a result of climate change and global warming. The uneven distribution of rainfall over time and space in Taiwan is becoming increasingly serious. the frequent occurrence of short-duration high intense storm has also led to poor drainage of water sewers, resulting in localized flooding within a short period, or even widespread flooding, which shortens the preparation and response time for disaster relief units. With the advancement of weather radar, Taiwan has set up several C-band Dual-Polarization, which have improved the observation and analysis capability of weather systems in the lower atmosphere and become indispensable for the development of many numerical models in estimating rainfall and predicting changes in weather systems; On the other hand, artificial neural networks have become one of the most popular research methods in various research fields with the development of artificial intelligence technology. In this study, we used a Long Short-Term Memory (LSTM) Neural Network and a Back Propagation Neural Network (BPNN) to learn the radar parameters of short-duration high intense storm and forecast 10-minute rainfall in urban areas from the present to the next 10 to 50 minutes(T+1~T+6).
In this study, the parameters of the rainfall radars for disaster prevention in the forest areas and the rainfall of 54 stations from May 2021 to November 2021 in the Taipei City administrative districts were collected. These radar parameters include Specific Differential Phase (KDP), reflectivity(Z), and Doppler Radial Wind Field(VR). The rainfall was estimated by using the specific differential phase and the reflectivity of the adjacent 2 x 2 grids from the radar stations, and the output variable is the rainfall of the rainfall stations then; The Doppler Radial Wind Field was multiplied by the predicted time interval to simulate the future movement of particles in the aqueous phase, enabling us to estimate the position of these particles after displacement. If it had reached the 2 x 2 grid adjacent to the rainfall stations, we would take the value of reflectivity and specific differential phase before displacement as the input modes for rainfall prediction, and the output variable is the 10-min rainfall of the rainfall station at the time interval. Finally, we analyzed the estimated rainfall results of the 54 rainfall stations and explored the connections between the differences in geomorphological and hydrological conditions of rainfall stations and the rainfall their radar parameters reflected, and we compared the estimated rainfall results of single-rainfall-station between mountainous and non-mountainous areas.

並列關鍵字

Artificial neural networks ； Long Short-Term Memory Neural Networks(LSTM) ； Back Propagation Neural Network(BPNN) ； Dual-Polarization Radar ； Rainfall Estimation

參考文獻

7參考文獻

1. Altche, F., and de La Fortelle, A. (2018). An LSTM network for highway trajectory prediction. IEEE Conference on Intelligent Transportation Systems, Proceedings, ITSC, 2018-March, 353–359.

2. Aswin, S., Geetha, P., and Vinayakumar, R. (2018). Deep Learning Models for the Prediction of Rainfall. Proceedings of the 2018 IEEE International Conference on Communication and Signal Processing, ICCSP 2018, 657–661.

3. Bouktif, S., Fiaz, A., Ouni, A., and Serhani, M. A. (2018). Optimal Deep Learning LSTM Model for Electric Load Forecasting using Feature Selection and Genetic Algorithm: Comparison with Machine Learning Approaches †. Energies 2018, Vol. 11, Page 1636, 11(7), 1636.

4. Chang, F. J., Chiang, Y. M., Tsai, M. J., Shieh, M. C., Hsu, K. L., and Sorooshian, S. (2014). Watershed rainfall forecasting using neuro-fuzzy networks with the assimilation of multi-sensor information. Journal of Hydrology, 508, 374–384.