A neural network approach for determining spatial and geometry dependent Green's functions for thermal stress approximation in power plant header components

Abstract

The trend in power generation to operate plant with a greater frequency of on/partial/off load conditions creates several concerns for the long term structural integrity of many high temperature components. The Green's function method has been used for many years to estimate the thermal stresses in components such as steam headers by attempting to solve the un-coupled thermal stress problem for a unit temperature step. Once a Green's function for a unit temperature step has been determined, realistic or actual component temperature profiles can be discretised and the time dependent stress profile reconstructed using Duhamel's theorem. Stress fluctuations can therefore be estimated and damage due to fatigue mechanisms can be quantified. A potential difficulty with this method is that Green's function approximations are determined for a single analysis point in a structure. This is because Green's functions are approximated by fitting a trial function to the results of finite element (FE) simulations. While a user can make some judgement on which point in a structure will give the "worst case" (or life limiting) conditions, it is foreseeable that points of interest will be dependent on the specific analysis conditions, such as the stub penetration geometry and the loading condition considered. The neural network approach described in this paper provides a means where transient thermal stress models of complex components (here taken to be steam headers) can be generated relatively quickly and used pro-actively to assess and modify plant operation. A range of header geometries have been considered to make the network applicable over an industry relevant envelope. Coefficients of determination (R 2 ) are typically above 0.92 when reconstructed (from neural network results) unit temperature step stress profiles are compared against "true" FEA results. Mean errors in the stress profiles are, for the majority of cases, less than $ This document is a collaborative effort.

A review of thermal, microstructural and constitutive modelling of 9Cr steel for power plant applications: Towards a through-process model for structural integrity of welded connections

“…Efficient alternatives to FE analysis, e.g. neural networks and Green's functions [154] or welding process parameter optimisation frameworks [155], could potentially also be adopted to increase the efficiency of computation. Therefore, structural integrity prediction using throughprocess modelling concepts is possible without the need for very long, computationally expensive simulations.…”

A process-structure-property model for welding of 9Cr power plant components: The influence of welding process temperatures on in-service cyclic plasticity response

“…Attempts to model entire 9Cr power plant components have included modelling of welding-induced residual stress [28,29] in pipe sections, as a guide to select appropriate PWHT conditions, but in-service operation was not considered. Rouse et al [30] used neural networks to predict thermal stresses in power plant components but all weld regions used the same material properties. Li et al [24] looked at a welded power plant tee-joint, treating the weld as an idealised three-material region, comprising of HAZ, WM and PM.…”