PREDICTION OF THE DEW POINT PRESSURE OF A GAS RESERVOIR USING ARTIFICIAL NEURAL NETWORK

ABSTRACT

In retrograde reservoirs, as reservoir pressure drops below the dew point during production, the pressure drop causes the formation of condensate liquids which accumulates at the bottom of the well, this condensate liquid saturation around the wellbore, results in what is known as condensate banking or blocking thereby resulting in the reduction of the relative permeability of the gas to flow, which in turn leads to well productivity decline. The dew point pressure which is the point at which condensate liquids begins to come out of gas is an important parameter that is required for fluid characterization, reservoir management and gas recovery. The experimental process of estimating dew point pressure is rigorous, expensive and time consuming. Hence this thesis proposes an alternative way of determining the dew point pressure using artificial neural network. For the development of this network, a total set of 1960 experimental data point of constant volume depletion test for different gas condensate fluids were used. The data sets include experimental values for reservoir temperature, hydrocarbon composition of 𝐶1 through 𝐶7 +, non-hydrocarbon composition (𝑁2,𝐶𝑂2,𝐻2𝑆), molecular weight of the heptane plus fraction (MW𝐶7 +), and dewpoint pressure. The data sets were divided into three sets. For training, validation, testing the neural network developed. The optimal neural network that had minimum errors was performed using the ADAM algorithm which has a network topology of one input layer with 1 neuron, 3 hidden layers with 32 neurons each and one output layer with one neuron. The accuracy of this developed model was tested with 392 data samples that was not used in training the model and also the result was compared to the results from already existing correlations and neural network model developed. The model from this work performed better with a lower RMSE (root mean square error) of 0.05218 and a lower MAE (mean absolute error) of 0.2079 than alternative correlations and existing model