Spatiotemporal Prediction Using Deep Models
Data that extend over both time and space, such as those observed in social infrastructure systems and natural phenomena, are continuously accumulated across many fields. However, real-world spatiotemporal data are often noisy and strongly affected by uncertainty and external factors. As a result, simple prediction methods tend to suffer from limited accuracy and reliability. In this research, we aim to capture the inherent spatiotemporal correlations in such data using deep learning and to predict future states with high accuracy. The outcomes of this research are expected to contribute to advanced maintenance planning, improved railway operation safety, and a better understanding of meteorological and environmental phenomena.
Deep Learning–Based Weather Forecasting
Ayumu Ueyama, Kazuhiko Kawamoto, Hiroshi Kera, VarteX: Enhancing Weather Forecast through Distributed Variable Representation, ICML 2024, Workshop on Machine Learning for Earth System Modeling, 2024 [arXiv].
This research aims to improve the accuracy of weather forecasting using deep learning. Weather prediction requires handling many meteorological variables, such as temperature, wind, and humidity, simultaneously, which poses challenges in computational cost and learning efficiency for existing methods such as ClimaX.
We propose a new approach called VarteX, which efficiently represents and integrates multiple weather variables. Experimental results show that VarteX achieves higher forecasting accuracy with a smaller model size, enabling accurate predictions even under limited computational resources.
Prediction of Degradation–Recovery Processes Using Binary Maintenance Intervention Records
Katsuya Kosukegawa and Kazuhiko Kawamoto, Long Short-Team Memory for Forecasting Degradation Recovery Process with Binary Maintenance Intervention Records, IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, Vol. E107.A, No.4, pp. 666-669, 2024 [paper].
The condition of civil infrastructure deteriorates over time but can recover through maintenance activities. In this research, maintenance actions are represented as simple binary time-series data indicating whether maintenance was performed or not. By combining these records with condition data, we study deep learning–based methods to predict future conditions and show how maintenance information can effectively improve prediction accuracy.
Spatiotemporal Prediction of Track Irregularity Using Exogenous Factors
Katsuya Kosukegawa, Yasukuni Mori, Hiroki Suyari, Kazuhiko Kawamoto, Spatiotemporal forecasting of vertical track alignment with exogenous factors, Scientific Reports 13, 2354, 2023 [paper][GitHub].
Ensuring railway safety requires continuous monitoring and prediction of track condition changes. This research focuses on capturing spatial relationships along railway lines and the influence of external factors such as maintenance history. Using deep learning models that integrate spatial information and exogenous factors, we demonstrate improved prediction of track irregularities. This work is conducted in collaboration with Central Japan Railway Company (JR Central).
2D Convolutional Neural Markov Models for Spatiotemporal Prediction
Calvin Janitra Halim and Kazuhiko Kawamoto, 2D Convolutional Neural Markov Models for Spatiotemporal Sequence Forecasting, Sensors 20, no. 15: 4195, 2020 [paper].
Spatiotemporal data in the real world often contain noise and uncertainty, making stable prediction challenging. In this research, we propose a probabilistic deep learning model that preserves spatial structure by combining convolutional neural networks with temporal modeling. The proposed approach achieves robust and accurate predictions, particularly for data with large variability and over longer prediction periods.
