Conceptualization of approaches to the development of statistical and neural network models for the New Safe Confinement, the Shelter Object Complex

Abstract

This article presents conceptual approaches to the development of a digital twin of a complex engineered facility — the New Safe Confinement, the Shelter Object (NSC-SO), which plays a pivotal role in the nuclear and radiation safety framework within the Chornobyl Exclusion Zone. Special attention is paid to the input data sources, which originate from an extensive sensor infrastructure that captures temperature, neutron flux, gamma radiation, and other parameters monitored within the NSC–SO. The article emphasizes the problem of incomplete or noisy data, which are typical of operational systems, and proposes algorithmic solutions for their processing and reconstruction. The research investigates classical statistical approaches, including regression analysis, moving averages, STL decomposition, and anomaly detection techniques. These methods offer high interpretability and are effective for early-stage diagnostics but show limitations in adapting to non-linear patterns and high-dimensional data structures typical of real-world industrial systems. The use of neural network models, in particular RNN, LSTM, GRU, and CNN, is proposed, as they enable the identification of temporal and spatial patterns in large data sets. The combination of statistical and neural network approaches makes it possible to build a hybrid model that better adapts to changing operating conditions of the object. Architecturally, the digital twin is envisioned as a modular, microservice-based system composed of dedicated layers for data collection, preprocessing, analytics, and visualization. The use of containerization and a service mesh is proposed to ensure scalability and component isolation. The article outlines a framework for evaluating model performance based on historical data fidelity, resilience to disturbances, and error metrics.