APPLICATION OF PHYSICS-INFORMED NEURAL NETWORKS (PINN) COUPLED WITH ONE-DIMENSIONAL UNSTEADY FLOW SIMULATION TO OPEN CHANNEL FLOW
Применение физически информированных нейронных сетей (PINN) в сочетании с одномерным моделированием неустановившегося течения к течению в открытом русле
2025-01-01
SCID: 54.1/epeb2ncv
Discuss with AI
Saint-Venant equationsflood event predictionone-dimensional unsteady flow simulationopen channel flowphysics-informed neural networks (PINN)
Figures from the paper
Abstract (AI)
本研究は,物理法則を組込んだDeep neural networks(DNN)であるPhysics-informed neural networks(PINN)を開水路の流れ解析に適用し,学習用のデータがない時空間領域でも,良好な予測が可能かどうかを検討する.具体的には,低平農地の直線の排水路における洪水イベントを予測対象とし,Saint-Venant equationsを組込んだPINN(PINN-SVE)を適用した.現地の観測値を収集することが困難なため,模擬豪雨と物理モデルから生成された疑似洪水データを正解値としてPINN-SVEの境界条件等に設定し,学習させた.その上で,学習用の正解値がない水路区間での流速と水深を予測した.100個の疑似洪水データを用いたPINN-SVEの予測結果は,従来のDNNと比べて,正解値との誤差やピーク差は概ね改善された.予測対象とする現象が滑らかな時間変化を伴う場合には,予測精度が高かったものの,急激な時間変化の場合には,予測精度が低くなった.
Key Findings
1
A physics-informed neural network coupled with the Saint-Venant equations (PINN-SVE) was applied to predict flood flows in a straight drainage channel in lowland agricultural areas.
2
PINN-SVE predicted flow velocity and water depth in channel sections without training target data, demonstrating spatial generalization.
3
Prediction accuracy was high for phenomena with smooth temporal variation but decreased for rapidly changing flows.
4
Synthetic flood data generated from simulated rainfall and a physical model enabled PINN-SVE training despite the lack of field observations.
5
Using 100 synthetic flood datasets, PINN-SVE generally reduced errors and peak discrepancies relative to a conventional deep neural network.
Research Object
Flood flows in a straight drainage channel on low-lying agricultural land
Research Subject
Prediction accuracy of flow velocity and water depth in ungauged channel reaches, including the effects of smooth versus abrupt temporal changes
Publication Details
Publication Date
2025-01-01
Journal
Publisher
ISSN
Cited by
0
Access Type
Author Information
Download PDF
Subscribe to digest