Production forecasting and hydrocarbon reserve estimation play a major role in production planning and field evaluation. Traditional methods of production forecasting use historical production data and do not account for completion and geolocation attributes that limit their prediction ability, especially for wells with a short production history. In this paper, we present a novel data-driven approach that accounts for the completion and geolocation parameters of a well along with its historical production data to forecast production.

In this work, we used supervised learning to develop an ensemble of machine learning (ML) based models to forecast production behavior of oil and gas wells. The developed models account for historical production data, geolocation parameters, and completion parameters as features. The dataset used to create the models comprises publicly available data from 80,000 unconventional wells in North America. The developed models are rigorously tested against 5% of the original data set. The models are systematically studied and compared against traditional forecasting techniques and results are presented here.

The created ensemble of models was tested by forecasting the production of 3,700 wells and the obtained results were compared against real production data. We show that the models clearly capture the natural decline trend of the produced hydrocarbon. In cases where the natural decline of the well has been temporarily modified, possibly due to operations, the production during other periods of the time series matches the prediction. This indicates that, unlike in traditional methods, such changes don't adversely impact the forecasting ability of our method.

We also conducted a systematic investigation and compared the forecast from the developed model against the forecast from a traditional method (Arps, 1945). During the comparison, it was observed that for short-production history wells (available production data from 2 to 12 months), the error rate in the predicted production behavior from traditional methods was higher when compared with the developed method. As the quantity of historical production data increases, the forecasting ability of traditional methods improves. By comparison, the decline from the developed method matches the real production data for both short- and long-production history wells, and clearly outperforms the traditional methods based on blind tests.

In this work, we present a novel ML based approach for forecasting production. This approach overcomes the challenge of the traditional time-series forecasting techniques that use only the past data for forecasting. It also incorporates static parameters (completion and geolocation parameters) in its architecture. The developed method leverages statistical averaging by employing an ensemble of random forest models, making the developed approach better than traditional ML based methods (ARIMA and LSTM) for forecasting time-series data.

You can access this article if you purchase or spend a download.