Search

水位計による河川水位のオリジナル観測データには，各種の異常値が含まれる．観測水位の異常値は，防災対応の判断や洪水予測システムに致命的なエラーを引き起こす可能性があるが，リアルタイムでの異常検知は十分に行われていない．本研究では，10分毎に観測所から配信される河川水位データを対象として，リアルタイムに異常値を検出する技術を開発した．機械学習による水位推定手法の技術を用いて，周辺の水位・雨量状況から対象とする観測地点の現時刻の水位を推定し，実観測データとの乖離度合いから異常度を算出した．さらにルールベースによる異常検知と組み合わせ，検知性能の向上を図った．九州管内の実データを用いて提案手法の精度検証を行い，既存手法と比較して高い検知性能を確認した．, Real-time observation data of the river water level includes various anomalies. Such anomalies may cause fatal errors in judgements on disaster prevention activity and flood forecasting systems, but reat-time anomaly detection has not been sufficiently implemented. In this study, we developed the model to detect anomalies in real-time for river water level data sent from observation station every 10 minutes. By using machine learning, the water level at the current time of the objective observation station was estimated from the neighboring water level and rainfall. Then the anomaly score was calculated from the degree of deviation between the estimated water level and actual obserbation. Furthermore,the model was combined with the rule-based anomaly detection model. The proposed method was verified using actual obsevation data, and better performance was cnfirmed compared to the existing method.

The model parameters for flood prediction currently used by river management have been established based on the relationship between precipitation and runoff for only a few specific cases. The prediction accuracy can be low when the data of cases that differ from the specific cases used for parameter-setting are applied, even when the subject cases are from the same basin as the specific cases. In this report, to improve the flood prediction accuracy of river management, a technique to calculate the run off rate corresponding to the basin storage capacity, which represents the wetness/dryness of the basin, was introduced by quantifying the basin storage capacity. By using the newly developed technique, reproduction of hydrographs and simulations of flood prediction were done. The technique compared with the currently used technique in which the runoff rate is fixed.

近年，世界各地で津波や洪水といった水災害の発生が報告されており，それらを予測する技術の向上が期待されている。なかでも，河川下流部への人口・資産の集中が進む現代においては，洪水予測の精度を上げることは重要である。一方，データの電子化や計算機技術の向上が今後も進むことを考えると，過去のデータの誤記や欠損は非常に少なくなっていくと考えられる。そこで本研究では，過去のデータを蓄積した事例ベースを利用して洪水予測を行うことを試みる。具体的には，事例ベースモデル(TCBM)を利用して実時間洪水予測である。TCBM の導入にあたり，適用性の拡大のため分布型流出モデルHydro-BEAM と時系列予測モデルLLM を利用する。, The aim of this study is to improve TCBM (Topological Case-Based Modeling) and to apply it for flood forecast by introducing other two concepts. One is LLM (Local Linear Model) thinking to use time-series information which is generally thought to be important for flood forecast. The other is to increase the data volume of flood events using the distributed runoff model (Hydro-BEAM) to cover the weak point of limited flood-case data. This enhanced TCBM is called FFM (Flood Forecast Model) for real-time flood forecasting system with six hour ahead forecast in this study. As a result, the effect of improving TCBM is found and the versatility of TCBM thought to be expanded.