Application of Hybrid Neural–Statistical Models in Other Infrastructure Systems

Abstract

Ageing infrastructure presents a persistent and growing challenge across many parts of the world, with water distribution networks, power grids, and transportation systems increasingly susceptible to unplanned failure and costly service disruption. This paper examines the application of hybrid neural–statistical models to predictive maintenance in these three sectors, building on earlier work focused specifically on pipelines. The proposed framework integrates four neural network architectures—Artificial Neural Networks (ANN), Long Short-Term Memory networks (LSTM), Convolutional Neural Networks (CNN), and Graph Neural Networks (GNN)—with Generalised Additive Models (GAMs) to estimate the Remaining Useful Life (RUL) of infrastructure assets and to assess failure risk. Case study evaluations demonstrate that the hybrid model consistently outperforms conventional approaches, including linear regression, ARIMA, and Support Vector Machines (SVM), across all three sectors. Specifically, the hybrid framework achieved Root Mean Square Error (RMSE) values ranging from 0.598 to 0.654, Mean Absolute Error (MAE) values between 0.462 and 0.528, and R² values of 0.913–0.926, compared with RMSE of 1.034–1.312 and R² of 0.654–0.711 for the best-performing conventional baselines. Classification accuracy ranged from 87.5% to 89.4%, with corresponding F1-scores between 0.849 and 0.876. These results point to a framework that is both accurate and interpretable, offering a practically applicable tool for guiding maintenance planning decisions at operational and policy levels. The findings have direct implications for national infrastructure management strategies, supporting a transition from reactive to proactive maintenance regimes.