A new method is developed by using machine learning technique to forecast single well production in both conventional and unconventional reservoirs. Unlike investigating and analyzing existing wells production rate and decline curve, this method predicts new well production rate according to reservoir properties such as matrix permeability, porosity, formation pressure and temperature, as well as hydraulic fracture parameters including fracture half length, fracture width and fracture conductivity.

In this paper, an inversion scheme is coupled with decline curve model, Logistic Growth Model (LGM), to obtain a set of decline curve parameters by fitting with production data. Both the Principal Component Analysis (PCA) and sensitivity study are applied to analyze the variance and identify key factors that influence production rate from reservoir and hydraulic fracture parameters. The sensitivity analysis results and scree plot from PCA serve as references to select key factors. Lastly, Neural Network (NN) technology is applied to investigate the pattern and correlation of selected reservoir and hydraulic fracture parameters and decline curve parameters. Therefore, the NN model can be applied to forecast production rate for a new well according to given reservoir and hydraulic information.

There is a good agreement between the available production data and decline curve model predicated production data based on the inverted decline curve model parameters. The scree plot and bi-plot generated by PCA provide the weight percentage of each component and help to identify factors that should be considered. Field production data is used to verify the feasibility of this method. This field case study is conducted by fitting the predicted production data (decline curve) based on NN model with field production data. The Mean Squared Estimation (MSE) of NN model is 0.013 Mscf/D and the overall R value is 0.917. This indicates that NN model is reliable to study the dataset and provide proper production (decline curve) prediction. The results illustrate that the predicted production data (decline curve) has good accuracy.

This paper proposes a statistical way for production forecasting based on machine learning. Instead of forecasting future production of existing wells, it provides meaningful reference for the evaluation of a new well and decision making.

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