Evaluating Reservoir Performance Using Transformer Based Proxy Models with Real Field Data

Feng Zhang


In this work, two Transformer based model were built to predict the production of each well in a real field case integrating production data, geological data, and dynamic data. Comparing to Regression model, Encoder-Decoder model has higher accuracy in both training and testing dataset. Long-term prediction depends on the training accuracy even Transformer model can capture more information in the longer sequence data. More features (i.e., well change and constraints changes) may be added to improve the training accuracy.