首页>
外文期刊>Journal of Natural Gas Geoscience
>Transparent machine learning provides insightful estimates of natural gas density based on pressure, temperature and compositional variables
【24h】
Transparent machine learning provides insightful estimates of natural gas density based on pressure, temperature and compositional variables
Machine-learning algorithms are widely used to predict the physical properties of natural gas, but most techniques involve establishing correlations between the influencing variables. In the case of neural networks these correlations tend to be obscured or hidden. The recently-introduced transparent open box (TOB) learning network applies an optimized, data-matching methodology that does not involve correlations. A large dataset, consisting 4512 compiled public-domain records incorporating just three of several influencing variables: molecular weight; pseudo-reduced pressure (Ppr); and, pseudo-reduced temperature (Tpr), is evaluated to predict gas density. For a mixture of gases, averages of the critical temperatures and critical pressures of its component gases are used to derivePprandTpr. A gas cannot be readily liquefied by applying pressure when it is above its critical temperature. The critical pressure of a gas is the vapor pressure at its critical temperature. The three influencing variables are related to natural gas density by complex non-linear relationships. An artificial neural network (ANN) and a TOB network both achieve highly accurate predictions of natural gas density for this dataset, but the ANN achieves slightly higher statistical prediction accuracy across the full gas density range evaluated. However, ANN and TOB complement each other in terms of the information and insight they provide regarding the underlying system. TOB offers tangible benefits when predicting natural gas properties by revealing details of the matched data records involved and their exact contributions to each prediction. The mathematical basis of TOB's rigorous two-stage approach is straightforward, yet it yields predictions of high accuracy for the highly non-linear systems associated with natural gas behaviour. Its methodology also identifies regions of a dataset prone to underfitting and/or overfitting and avoids unjustified extrapolations.
展开▼