Prediction of well production event using machine learning algorithms

In this paper, a new approach was identified and tested to detect abnormal events in producing wells when a labeled dataset is unavailable or the number of instances are below 10% and are insufficient for conventional modelling methods. Autoencoders (AE), a type of unsupervised learning, are trained to learn normal behavior by trying to reconstruct the input data that is fed into the model. When run in prediction mode, low reconstruction errors are classified as Normal behavior whilst higher errors are classified as anomalous behavior. Different model structures were tested. An average accuracy of 94% with a precision and recall rate of 70% was achieved using a 6-Layered AE-NN model. The results of the models created show encouraging results and can help detect events and notify engineers when the well is deviates from expected behavior.