2025 Volume 81 Issue 16 Article ID: 24-16189
Our study reports whether or not it is possible to make good predictions even in spatiotemporal computational domains where there are no training datasets using physics-informed neural networks (PINN), which are the deep neural networks (DNN) that extendedly incorporate physical laws. To predict flood events in drainage canal networks in low-lying agricultural land, the PINN that incorporates the Saint-Venant equations (i.e., PINN-SVE) was applied to a straight canal in the drainage networks. Owing to no in-situ observed data, simulated flood data generated by a rainfall-runoff model with pseudo heavy rainfalls were used as the correct values to predict flow velocity and water depth in canal cells without training datasets. The predicted PINN-SVE results using 100 patterns of flood events showed that the errors and peak differences from the correct values were generally better than those of the conventional DNN. Good prediction accuracy was obtained for waveforms with smoothly temporal changes in the flow velocity and water depth; however, the poor accuracy was likely caused by the waveforms with rapidly temporal changes.
