複数の気象要因を考慮した融雪期における河川の水位予測
分布画像を用いた深層学習の適用
融雪に伴う出水が発生する豪雪地帯において行われる河川工事では，河川工事エリア内の重機や建設資機材を安全，確実に避難，養生させるため，12時間以上先の出水の有無や水位変動を把握する必要がある．本研究では，融雪時期を対象に融雪に関する気象分布画像を用いて深層学習により1～24時間先の水位を1時間ごとに予測する手法を構築し，実測水位と予測水位の相関性を確認した．本手法は，入力値への重み付けを行うことで予測精度を高めることを特徴としている．8河川で検証したところ，全ての河川で重み付けをしない場合に比べて予測精度が向上し，出水を精度良く検知できる可能性があることを確認した。今後本手法を，出水警報システム「T-iAlert®River」に適用して，融雪時期の河川工事に積極的に活用していく予定である．
キーワード：融雪期，河川水位予測，安全管理，深層学習，気象分布
＊１ 技術センター 社会基盤技術研究部 水理研究室
Prediction of river water levels during snow-melting season using multiple types of weather distribution images
Application of deep learning with distribution images
At river work sites, it is necessary to predict water levels more than 12 hours in advance in order to evacuate heavy machinery and construction equipment and materials in a river work area and protect them in a safe and sure manner. In this study, a method to predict water levels from 1 to 24 hours in advance every hour has been developed by deep learning using weather distribution images related to thawing in the snowmelt season, and the correlation between the measured and the predicted water levels was confirmed. This method is characterized by weighting input values to enhance prediction accuracy. The method was applied to eight rivers, and it has been confirmed that the prediction of water levels with weighting shows higher prediction accuracy than prediction without weighting in all rivers and that the method may be able to accurately detect rises in water level. The authors plan to apply the method to the T-iAlert River flood warning system and actively use it in river works during the snowmelt season.
Keywords: snowmelt season, water level prediction, safety management, deep learning, weather distribution
*1 Hydraulic Research Section, Infrastructure Technology Research Department, Taisei Advanced Center of Technology