Deep Neural Network Based Prediction of Leak-Off Pressure in Offshore Norway

Leak-off pressure (LOP) is an important parameter to determine a weight of drilling mud and in-situ horizontal stresses. When the well pressure become higher than the LOP, it can cause a wellbore instability during drilling, such as a mud loss. Thus, accurate prediction of LOP is important for safe and economical drilling for the oil and gas industry. In this study, we present a novel prediction model for the leak-off pressure (LOP) offshore Norway. The model uses a deep neural network (DNN) applied on a public wellbore database provided by the Norwegian Petroleum Directorate (NPD). We used a Python-based web scrapping tool to collect data from more than 6400 wells (1800 exploration wells and 4600 development wells) from the NPD factpages. Then, we analyzed the collected data to investigate impacts of spatial and regional factors on the collected LOPs. The DNN model was structured to predict the leak off pressure offshore Norway using open source libraries Keras and Tensor Flow. The model tests have various hidden layers (i.e. 3, 5, and 10 layers). In order to avoid overfitting, we specified an early-stop algorithm. In our study, we took 80% of the data as the training set keeping the remaining 20% to test the model. In total, the database consists of around 3000 leak-off pressure data from about 1800 exploration wells, and grouped in geographical area (North Sea, Norwegian Sea, Barents Sea groups). The LOPs of the North Sea and the Norwegian Sea show a bi-linear trend with depth. The LOPs that are measured from deeper than 2-3 km below sea level show clear a deviation in trend, with a steeper increase compared to the shallower section. The steeper part of the bi-linear trend at greated sub-surface depths can be related to a coupling with tectonic stresses from base rocks. The data from the Barents Sea shows more scattered LOP compared to the other regions offshore Norway. The scattered data seem to relate to the complex geological history on the Barents sea. In general, the accuracy of the prediction increases with the number of hidden layers. However, when the number of the hidden layer exceed 5, there was no significant improvement in the accuracy of prediction. The validation test shows relatively good prediction of LOP with an MAE (Mean Absolute Error) of less than 0.07 even for areas experiencing complex geological history such as the deep subsurface of the Norwegian Sea and the shallow subsurface of the Barents sea. This study clearly demonstrates how a data-driven approach combined with machine learning algorithms can provide hidden patterns of not only LOP itself but also the additional information about the lithology, the stress history and the geographical frequency of exploration.