Oil production prediction based on a machine learning method

Numerical reservoir simulation is the most common method for oilfield production prediction, but its accuracy is based on exact geological model and quality history matching. In order to overcome the shortcomings of numerical simulation technique (e.g. time consuming, high cost and numerous data), a machine learning based production prediction method was established, which can quickly and accurately predict oil well production based on abundant static reservoir data and dynamic development parameters that can be easily obtained on site. The traditional BP neural network method cannot describe the correlation of the production change in the time dimension, while the production prediction model which is established based on long-short-term memory (LSTM) and takes into account the change trend and correlation of production performance data is a more effective method for predicting oil well production. Firstly, the influential degree of each factor on single-well production was analyzed by means of mean decrease impurity (MDI) method. Then dimension reduction was carried out based on the importance of characteristic parameters, the uncorrelated redundancy characteristics were removed, and the main parameters influencing oil well production were determined. Finally, combined with screened out characteristic parameters and daily oil production, the LSTM model was trained and optimized, and the final oil well production prediction model was established. In addition, this newly established model was verified based on actual oilfield data and its application effect was evaluated. It is shown that the production predicted on the basis of LSTM model is highly accordant with the actual value. It is indicated that this model can reflect the dynamic production change law accurately and provides a new method for predicting oil well production.