Petroleum Production Forecasting Based on Machine Learning

Abstract

Reservoir numeric simulation is the most commonly used method for oilfield petroleum production forecasting, but its accuracy is based on accurate geological models and high-quality history matching. In order to overcome the shortcomings of numeric simulation requires, like time consuming, high cost, and lot of data required, an machine learning method was adopted and trained for predicting oilfield production using static and dynamic developing parameters. Since the traditional BP neural networks cannot accurately capture the time correlation between data, a long short-term memory model was used to establish production prediction model that can consider the trends and context correlations of production data. Mean Decrease Impurity method was first conducted to analyze the relative importance of predictor variables. Relative unimportant features then can be excluded according to their relative importance. The dimension reduction of predictor variables was combined with production data to train and optimize LSTM network. Thereby predictive model for production prediction was established after the training. The actual oilfield data was used to verify the proposed approach and conducting application effect analysis. The results show that the predicted production computed by LSTM network is highly consistent with the actual production, which can accurately reflect the dynamic variation of production.