Study on Real-Time Forecasting of Hydrological Variables Using Technique of Non-Linear Time Series Analysis

为了有效地防洪，准确的实时洪水预报是非常重要的。为了支持洪水预报，准确的实时降雨预报也很重要。本研究利用非线性时间序列分析技术，即局部线性逼近(LL)法、最近邻(NN)法和线性输出映射自组织特征映射(SOLO)算法，开发了新的洪水暴雨预报系统，并对系统的预报精度进行了检验。首先，对日本冈山县北部黑木水坝流域1979-2002年间观测到的39个暴雨记录，应用LL方法进行了实时洪水预报。结果表明，该方法具有较高的洪水预报精度，与常规的采用卡尔曼滤波技术的水池模型实时预报系统的预报精度相当。其次，将SOLO算法应用到黑木水坝基地的实时洪水预报中，发现采用SOLO算法的洪水预报系统在使用由历史流量和降雨数据的所有主成分的分数组成的特征向量的情况下，达到了较高的精度。该方法的精度几乎与LL法相当。然而，当特征向量直接由流量和降雨数据组成时，预测精度变差，特别是在验证期间。最后，利用日本自动气象数据采集系统(AMeDAS)冈山周边地面雨量计观测到的雨量资料，将神经网络方法应用于冈山的实时降水预报。尽管降雨发生的预报准确率高达80%以上，但降雨深度预报的精度往往被低估，不能满足实际应用的需要。

For effective flood control, accurate real-time forecasting of flood discharge is very important. For supporting the flood forecasting, accurate real-time forecasting of rainfall is important as well. In this study, we developed new flood and heavy rainfall forecasting system using technique of non-linear time-series analysis, that is, local linear approximation (LL) method, Nearest Neighbor (NN) method and Self-Organizing feature map with Linear Output mapping (SOLO) algorithm, and examined forecasting accuracy of the developed system. First, we applied LL method to real-time flood forecasting for 39 storms records observed at Kuroki Dam Basin located in the northern part of Okayama Prefecture, Japan, during 1979 -2002. It is, as a result, found that forecasting accuracy of flood discharge by the LL method is so high that it is comparative to that by the conventional real-time forecasting system using the Tank Model with Kalman filtering technique. Second, we also applied the SOLO algorithm to real-time flood forecasting at Kuroki Dam Basing and found that the flood forecasting system by the SOLO algorithm achieved a high degree of accuracy in the case of using feature vectors composed of scores of all principal components of past discharge and rainfall data. The accuracy is almost equivalent to that by the LL method. However, when feature vectors are directly composed of discharge and rainfall data, forecasting accuracy became worse particularly during verification duration. Finally we applied the NN method to real-time rainfall forecasting at Okayama using the rainfall data observed at ground rain gauges around Okayama of Automated Meteorological Data Acquisition System (AMeDAS) of Japan. Although forecasting accuracy of rainfall occurrence was so high as over 80%, accuracy of forecasted rainfall depth was insufficient for practical use because it tends to be underestimated.