Abstract
<jats:p>The subject matter of the article is the methods and tools for scenario-based (what-if) recalculation of the estimated reconstruction cost of a damaged infrastructure object using a digital twin. The goal of the work is to develop a digital twin of a damaged real estate object that provides forecasting of the estimated recovery cost, scenario-based recalculation when object characteristics change, and interpretation of results. The following tasks were solved in the article: analysis of existing approaches to digital twin modeling and reconstruction cost forecasting; development of an architecture that integrates a destruction recording system, a machine learning microservice, and a scenario recalculation module; implementation of a prediction interpretation mechanism; creation of a digital twin prototype with support for what-if analysis. The following methods are used: architectural design for building a hybrid web system; machine learning for automated cost calculation based on the XGBoost ensemble model; prediction interpretation using SHAP (SHapley Additive exPlanations) analysis to assess the contribution of individual object characteristics to the final cost estimate. The following results were obtained: a digital twin prototype was developed that combines a module for maintaining a digital object profile, a scenario calculation module for modeling alternative reconstruction options, an ML microservice for cost prediction, and an interpretation module based on TreeSHAP (algorithmic implementation of SHAP for tree-based models, in particular XGBoost). The system provides an end-to-end process from damage recording to justified comparison of recovery scenarios. Conclusions: the proposed approach is aimed at solving the problem of fragmentation of existing solutions, where damage registration, forecasting, and scenario analysis operate in isolation. The practical value lies in the ability to quickly compare reconstruction alternatives and increase transparency of budget planning. The limitations of the study are related to the small size of the training sample, validation of results in the range of UAH 20–90 million, and the descriptive nature of certain categorical features.</jats:p>