Heat load prediction in flow boiling using boiling-induced vibrations aided with machine learning

Abstract

This study delves into the analysis of boiling-induced vibrations observed during a flow boiling experiment and explores their potential in predicting heat load through machine learning models. The frequency spectral analysis revealed that the dominant frequency ranges between 6.5 – 12.5 kHz, delineated into three distinct bands. Principal Component Analysis (PCA) underscored the significance of the 8 – 9 kHz peak, encapsulating around 65% of dataset variance. Diverse machine learning algorithms including decision tree regression, random forest regression, support vector regression, and multi-layer perceptron were rigorously evaluated. The Multi-Layer Perceptron (MLP) architecture with specific neuron configurations and a learning rate of 0.2 emerged as the superior model based on its minimal Mean Squared Error (MSE) and high R2 score. Notably, all models exhibited inference times within the microsecond range. This amalgamation of vibration spectral analysis, machine learning model assessments, and inference time evaluations underlines the promising prospect of utilizing boiling-induced vibrations for real-time heat load prediction, showcasing superior performance compared to conventional methods.