Existing studies conduct general cost analyses for prefabricated components, yet structural heterogeneity results in distinct cost drivers. Most studies concentrate on the technical performance of prefabricated staircases, with insufficient investigation into dedicated cost-estimation methods. This study establishes a hybrid prediction framework integrating GAN-based data augmentation and SHAP-empowered Multilayer Perceptron (SHAP-MLP) modeling, using prefabricated straight staircases as empirical objects for multidimensional analysis. Total cost is classified into production, transportation, and on-site installation phases, followed by systematic screening of 33 influencing factors for predictive modeling. The Analytic Hierarchy Process (AHP), with a 1–9 scale, is adopted to quantify indicator weights and prioritize features. Triple verification (multi-expert consistency test, group opinion coordination test, and sensitivity analysis) removes five weakly correlated parameters to form a preliminary indicator system. Based on 240 original engineering data samples, the GAN generates 60 high-fidelity synthetic samples. Distribution consistency between synthetic and original data is validated via the Kolmogorov–Smirnov (KS) test, p-value verification, and kernel density estimation (KDE). SHAP interpretability analysis identifies four core determinants: prefabrication rate, total staircase area, standardization level, and number of floors. Eight low-impact parameters are excluded to optimize model input, leaving 20 validated indicators. The GAN-SHAP-MLP model maintains superior performance in testing, with a test-set RMSE of 49.538, representing improvements of 41.3%, 22.5%, and 25.7% over LSTM (89.33), CNN (67.59), and standard MLP (70.56), respectively. The difference between its test-set and overall R2 is only 0.69%, significantly lower than 2.06% for LSTM and 5.47% for MLP. Empirical validation with real engineering cases from four different regions further confirms the model’s high prediction accuracy, with a minimum error of only 1.49%. The integration of data augmentation and interpretable deep learning provides a high-precision, interpretable cost prediction tool for prefabricated straight staircases, promoting methodological progress in construction economics.
Zhang et al. (Thu,) studied this question.