Home Facts & Figures Knowledge base Gas Analysis of End of Field Life Techniques and predicting Liquid Loading using Artificial Neural Networks (2017)

Analysis of End of Field Life Techniques and predicting Liquid Loading using Artificial Neural Networks (2017)

This MSc thesis has been written by former intern Ellis Bouw.

Liquid loading of a gas well is the inability of the produced gas to remove the produced liquids from the wellbore. It is one of the major issues that decreases flow production substantially or even stops flow completely for wells that are in the mature or tail-end production phase. To sustain gas production, the problem can be overcome with End of Field Life (EoFL) techniques. Of these techniques, velocity strings and foam injection are the two most popular ones in the Netherlands. To date, the availability of data on the potential of EoFL techniques is limited and this study aims to quantify the potential volume gain from these two EoFL techniques. Moreover, it is difficult to determine when to install these techniques due to the prediction uncertainty of the liquid loading moment. A solution can be found in artificial intelligence. Using big data may predict future instability of production rates and may therefore be very useful in predicting the liquid loading moment in advance. In this thesis a first step is undertaken to use artificial intelligence, in particular artificial neural networks (ANN), to predict the onset of liquid loading applying actual field data.

In total, sixty-four liquid loading wells were examined in terms of production quantity. It can be concluded that the volume gain using a velocity string or foam injection has a very wide spectrum. About half of the wells produced 10 to 15 million Nm3 more due to one of these techniques, while other wells even produced or more. The wide range is due to the dependency on well and environmental conditions. Next to this it may be noted that operators had expected a larger volume gain from these EoFL techniques.

Of the sixty-four wells, fifteen wells were examined in more detail, particularly in terms of economic gain. The techniques show a high success rate for over 70% of the wells with a NPV of some 20 euros per well. The remaining wells showed a negative NPV, but only due to external factors, for example a leaking tubing. Therefore, EoFL techniques have shown to be valuable but more research should be undertaken to enhance knowledge on improving volume gains for wells in the mature ortail-end production phase in the Netherlands.

When predicting the onset of liquid loading, ANN forecasted future gas rates from historical monthly production data using the tubing size as a variable input parameter. The Nonlinear Autoregressive with External Input (NARX) is trained using the Levenberg-Marquardt algorithm. In order to improve training performance pre-processing was undertaken, the Lowess filter was applied and normalization was conducted.

The network showed a satisfactory prediction of the gas flow rate. Having the lowest mean squared error, the network was constructed with two layers, each containing 75 neurons. Coleman criterion was introduced to indicate the onset of liquid loading. When the predicted flow rate falls below the Coleman rate, the liquid loading alerter is triggered. The alerter forecasts the month in which liquid loading may occur up to a maximum period of twelve months. The prediction becomes more precise as the alerter approaches the the liquid loading moment and it may be concluded that the alerter is accurate up to two months in advance. Moreover, the alerter is able to predict nine months prior to liquid loading with a deviation of up to two months either side. Nevertheless the Coleman Criterion has the tendency to calculate lower critical rates than other methods. Therefore it may predict liquid loading to occur at a future month than the exact month of liquid loading.