Analysis of deep learning neural network combined with experiments to develop predictive models for a propane...

Area of Science:

• Chemical Engineering
• Computational Fluid Dynamics
• Risk Assessment

Background:

• Industrial fires, particularly jet fires, pose significant risks in process industries, often leading to cascading failures.
• Artificial neural networks (ANNs) offer a computationally efficient approach for simulating complex processes, including risk analysis.

Purpose of the Study:

• To develop and compare artificial neural network models for predicting jet fire dimensions.
• To utilize experimental data for training ANNs to understand the relationship between jet fire parameters and flame geometry.

Main Methods:

• Experimental data on jet fire shape ratios (defined by the 800 K isotherm) were used for ANN development.
• Multi-layer perceptron (MLP) with Bayesian Regularization and Radial Based Functions (RBF) networks were employed.
• Mass flow rates and nozzle diameters served as input, while jet flame lengths and widths were the output.

Main Results:

• The optimized MLP network had twenty neurons in its hidden layers, and the RBF network had fifteen neurons in its single hidden layer.
• The MLP network achieved a Mean Squared Error (MSE) validation performance of 0.00286 at 100 epochs.
• The RBF network achieved an MSE validation performance of 0.00426 at 20 epochs.

Conclusions:

• ANNs, specifically MLP with Bayesian Regularization, can effectively simulate jet fire characteristics using readily available input parameters.
• The developed models provide a rapid method for estimating jet fire dimensions, contributing to improved risk assessment in process industries.