Dynamic modelling of low frequency oscillations in swirl burner/furnace systems using artificial neural networks. Xue, Y.; Syred, N.; Rodriguez-Martinez, V.; O'Doherty, T. Journal of the Energy Institute, Volume 79, Number 4, December 2006 , pp. 232-240

The present paper is concerned with modelling techniques for combustion excited oscillations in swirl burner/furnace systems (SBF) for control strategy development. Artificial neural networks are used to model the combustion dynamics in an SBF which can be made to oscillate under various conditions. A methodology for building system models is proposed where dynamic flowrate measurements of gas and air are absent, the inputs being pressure time records, measurements of mean gas and air flowrates and geometry of the furnace. The modelling has two objectives: one is to model the system fluctuating pressure frequencies and rms amplitude, which can then be used to build a core stimulation model and the other to model the time domain dynamic response to pressure oscillation. Experiments were undertaken on a 100 kW swirl burner/furnace systems (SBF) test rig. This system could be made to oscillate regularly by parametric changes such as mean equivalence ratio, mean gas and air flowrates and small geometry changes. The pressure fluctuations generated by low frequency oscillations are considered when f<70 Hz. Fast fourier transform analysis of the data was carried out to derive the dominant first and second harmonic frequencies for various combustion conditions. Multilayer perceptron neural networks are applied and a solution to enhance the training converging speed applied. This is a rule of instant learning performance based dynamic learning rate. The effectiveness of the non-linear models trained is verified by prediction and comparison with experimental data. It is shown that the model can give predictions of oscillation frequency and amplitude beyond that of the experimental data upon which it was based.