2026 Volume 7 Issue 3 Pages 386-396
This paper introduces a Temporal Convolutional Network (TCN) for inferring the coefficient matrices in a real-time flood inundation prediction method that combines dimensionality reduction with deep learning, and examines the speedup in prediction through comparison with LSTM. Singular Value Decomposition (SVD), Non-negative Matrix Factorization (NMF), and an Autoencoder (AE) were adopted as the dimensionality reduction techniques. The method was applied to 3,000 flood simulation cases assuming overtopping at two locations along the right bank of the Arakawa River. The results showed that inundation areas could be predicted with high accuracy even when the approximately 220,000-dimensional inundation data were compressed to 10 dimensions. Among the reduction techniques, AE was superior in terms of reconstruction accuracy and NMF in terms of reconstruction time. Furthermore, TCN maintained prediction accuracy comparable to that of LSTM while reducing the inference time to approximately one-fifth, demonstrating its effectiveness in improving real-time performance.