Utilization of Machine Learning for Flow Assurance in the Oil and Gas Sector: a Focus on Annular Flow Prediction
There is insufficient literature for annular flow and applications of Machine Learning in the petroleum industry; thus, this thesis is centred on annular flow prediction. Liquid holdup and flow behaviour during annular flow was accurately predicted using the Neural Network toolbox on MATLAB (Machine Learning). The experimental data contained measurements of liquid holdup at three different probes across an air – water system, over a period of 15 seconds. Measurements were noted at intervals of 0.001s; thus, a total number of 15,000 time steps. The superficial gas velocity within the system was changed 17 times (ranging from 6.17 ms-1 to 16.05 ms-1), while the liquid superficial velocity was constant in all cases (0.02 ms-1); data exists for 17 different velocities, for the 3 different probes. Effective neural networks that yielded >90% validation accuracy were noted to occur when the architecture was designed to have 10 hidden neurons and greater than 50 delays. An efficient architecture was further solidified by analysis of the “Autocorrelation error” chart; all non-zero lags should be within the confidence limit. This exceptional performance was ascertained via the calculation of average liquid holdup values. The average liquid holdup of experimental and simulated values was noted to be the same or greatly similar, proving accuracy of the implemented neural network design. This research has proved that the trend in annular flow time series data can be identified by a neural network; further research could be carried out to determine the relevant variables that are needed for accurate liquid holdup computations with time.