Text–model generation–machine learning framework and performance of three-sided protection steel beam temperature model

The three-sided protection steel beam is commonly used in offshore and onshore oil and gas processing facilities. The top surface of the upper flange is exposed, causing a rapid temperature decrease in the vertical direction, influenced by section depth, insulation thickness, and conductivities. Deriving an equation to encompass these factors is analytically challenging. A machine learning model with fast prediction is a viable approach, despite the labour-intensive modelling and database training processes. To address this, a machine learning-based prediction programme connects parameter inputs, modelling, simulation, and data extraction using Python, ABAQUS, and MATLAB. A model generation agent embedded with a 2D contact detection algorithm is produced to establish the beam heat transfer models. For performing the large-scale simulation, an interruptible MATLAB terminal is adopted to control the sequence. Results are processed through a data pipeline and trained with Gradient Boosting and Deep Neural Network (DNN). In addition, a data-batch generator is designed to reduce memory usage and enable training on a single GPU. The database comprises 414 models of IPE sections, Universal Beam sections, and 63 welded sections, with depths ranging from 127 to 1500mm. 400°C and 538°C steel limit temperatures are considered for the insulation configuration. Results show a root mean squared error of 1.34°C between the test and prediction results, with gradient boosting prevailing for regression problems under similar training times. The test calculation time reflects high computational efficiency, with 83.3% of predictions completed within 60 seconds.