Determination of highway maintenance and repair costs with soft computing and classical methods

This study combines classical regression methods and soft computing techniques in a hybrid structure to accurately predict the maintenance and repair costs of road infrastructures. Increasing traffic volume, climate variability, and economic factors necessitate advanced prediction models for infrastructure management. A multidimensional dataset containing 21 independent variables related to physical infrastructure, economic indicators, meteorological data, and traffic density for the years 2004–2022 was used. In the first stage, without variable reduction, classical methods such as Multiple Linear Regression (MLR), LASSO, Ridge, and Stepwise AIC were compared with soft computing methods such as Artificial Neural Networks (ANN), Support Vector Machines (SVM), Random Forests (RF), and Gradient Boosting (GBOOST). ANN yielded the best results with MSE (410,294), R² (0.964), and MAPE (3.756). In the second stage, the most influential variables were identified using LASSO, Ridge, RF, Genetic Algorithm (GA), Sensitivity Analysis, and Stepwise AIC methods, and prediction models including LASSO, Ridge, RF, MLR, and ANN were developed and tested with these variable groups. The Ridge-ANN hybrid model demonstrated the highest prediction performance with MSE (127,967), R² (0.989), and MAPE (1.237). The superior performance was achieved by combining Ridge's ability to reduce multicollinearity with ANN's power to model complex relationships. As a result, the proposed Ridge-ANN model provides an effective and innovative solution for sustainable infrastructure management with high accuracy and generalisability.