.png){kind=icon}
.png){kind=icon}
Return
Modelling road maintenance expenditure dynamics under uncertainty: a hybrid VAE–random forest framework
.png){kind=icon}
OA
AI
T
E
J
E
DOI：10.1080/15623599.2026.2691236.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Road Maintenance Expenditure (RME) forecasting is essential for strategic budgeting, proactive asset management, and long-term infrastructure resilience. However, existing approaches often treat expenditure drivers in isolation and inadequately account for socio-economic conditions, pavement condition metrics, economic disruptions, natural shocks, and their delayed or compounded impacts. This study addresses this gap by developing a domain-integrated forecasting framework to model nonlinear, lagged, and shock-driven RME dynamics using a New Zealand case study. The framework combines Variational Autoencoder (VAE)-based latent feature extraction and nonlinear dimensionality reduction with a fine-tuned Random Forest regressor. Bayesian optimisation is used to identify optimal latent representations, while lagged variables and interaction terms explicitly represent delayed and compounded effects of external shocks. Comparative benchmarking across Local, Total, and New Zealand Transport Agency expenditure categories demonstrates strong in-sample modelling performance, with R2 values of 0.9818, 0.9842, and 0.9800, respectively, outperforming conventional linear, ensemble, and neural network-based models. A supplementary forward-rolling validation under chronological forecasting conditions showed reduced predictive performance, highlighting the challenge of temporal generalisation amid residual uncertainty, decentralised decision-making, and evolving external conditions. The study contributes a robust methodological framework for uncertainty-aware maintenance expenditure forecasting and demonstrates the importance of distinguishing structural relationship learning from realistic temporal prediction.
Keywords:
Road maintenance
socio-economic factors
shock
machine learning
forecasting
infrastructure planning
resilience
Journal
I
IF:
3.4
Papers:
255
Citations:
0