Intelligent Prediction of Thermal Regimes in Engineering and Technological Systems

Abstract

Recently, surrogate or metamodeling has been actively used in design, optimisation, control, and predictive analytics tasks in modern technological industries such as power engineering, mechanical engineering, and metallurgy. Accurate modelling of temperature fields in metal structures, working elements, and units of technological installations is critically essential for assessing their reliability, performance, remaining service life, and resistance to thermal and mechanical wear. In the study, a neural network based on the multilayer perceptron architecture was developed for training. The training sample was a synthetic dataset of 105 unique simulations, produced by numerical integration of the 1D heat conduction equation via the Tridiagonal Matrix Algorithm (TDMA). Verification of the model on an independent test sample demonstrated its high predictive capability: $\text{MAE}=1.1415 \mathrm{K}, \mathrm{R}^{2}=0.9988$. The average direct inference time was 2.3 ms, which is significantly lower than the computational cost of classical numerical methods, whose accuracy and performance are determined by the discretization parameters. The findings demonstrate that the proposed method serves as a viable substitute for conventional numerical techniques and is applicable to real-time optimisation and thermal regime prediction in engineering and technological systems.