Abstract
In the early stages of a project, the owner is faced with the challenge of deciding whether to invest in the project or not. The road tunnel project involves a large amount of investment, and in order to help the owner to make a selection decision, it is crucial to achieve accurate and efficient road tunnel cost prediction in the early stages. In this paper, the main factors affecting the cost of road tunnels are identified through literature research, and the least squares support vector machine is optimized using a particle swarm optimization algorithm. The results show that the least squares support vector machine model based on the particle swarm optimization algorithm performs well in road tunnel cost prediction, with high fit and low error, and meets the accuracy and practicality requirements of the pre-project.
Introduction
Project owners in civil engineering often face an "early-stage dilemma": the need to make massive investment decisions when available data is at its scarcest. Highway tunnels are particularly challenging due to their high capital requirements, technical complexity, and long construction timelines. Traditional cost estimation methods, which often rely on manual expertise or basic linear models, struggle to capture the non-linear relationships between complex geological conditions and final construction costs.
While modern machine learning techniques like Back Propagation (BP) neural networks can handle such complexity, they typically require massive datasets that are rarely available for niche infrastructure projects. This research addresses the need for a model that provides fast and accurate predictions using a small, high-quality sample set.
Methodology: The PSO-LSSVM Approach
The proposed model utilizes a hybrid approach combining Least Squares Support Vector Machines (LSSVM) with Particle Swarm Optimization (PSO).
Least Squares Support Vector Machine (LSSVM)
LSSVM is a variation of the standard SVM that simplifies the optimization process by solving a set of linear equations rather than complex quadratic programming problems. This modification makes the algorithm faster and more stable, particularly when dealing with the "small sample" problem inherent in construction data.
Particle Swarm Optimization (PSO)
The performance of an LSSVM model is highly sensitive to two hyperparameters: the penalty factor ($C$) and the kernel parameter ($\sigma$). To avoid the inefficiencies of manual "trial and error," the researchers employed PSO—a metaheuristic algorithm inspired by the foraging behavior of bird flocks—to automatically search for optimal parameter values.
The model incorporates three primary input categories identified through literature research:
- Tunnel Length: Ranging from short to extra-long.
- Rock Grade: The percentage distribution of rock quality (Grades III, IV, and V).
- Average Cross-sectional Area: A critical driver for construction methods and material volume.
Experimental Setup and Dataset
The study utilized a dataset of 18 highway tunnel projects. 16 samples were used for training, while 2 samples ($P_{17}$ and $P_{18}$) were reserved for testing. The target output variable was the cost per linear meter (measured in units of 10,000 RMB/m). Although the dataset is small compared to typical machine learning benchmarks, it accurately reflects the high-value, low-frequency data characteristic of heavy infrastructure research.
Key Results and Discussion
The PSO-LSSVM model was benchmarked against a Genetic Algorithm-based LSSVM (GA-LSSVM) model. The results demonstrated significant improvements in both accuracy and efficiency:
- High Precision: The PSO-LSSVM achieved a Coefficient of Determination ($R^2$) of 0.998, indicating a near-perfect fit for the training data.
- Low Error Rates: On the test set, the Mean Absolute Percentage Error (MAPE) was only 2.50%, significantly lower than the 10.08% recorded for the GA-based model. For test sample $P_{17}$, the relative error was remarkably low at 0.25%.
- Computational Efficiency: The PSO-LSSVM completed its optimization in 3.33 seconds, roughly six times faster than the GA-LSSVM (20.11 seconds).
These findings suggest that for specialized engineering fields, "smarter" algorithms like LSSVM are often superior to "deeper" algorithms like neural networks, which tend to overfit when data is sparse. Furthermore, including the percentage of different rock grades rather than a single average grade allowed the model to capture cost volatility more effectively.
Conclusion and Future Work
The PSO-LSSVM model provides a robust tool for early-stage highway tunnel cost prediction, meeting the accuracy and practicality requirements of project owners. However, future work should focus on expanding the dataset to validate the model across different geographic regions and incorporating "soft" variables such as market fluctuations in material prices and labor costs. Additionally, developing a user-friendly software interface would facilitate broader industry adoption beyond the research environment.
Summary updated