Crude oil is crucial for economic growth, therefore accurate production estimation is vital, especially for unpredictable US shale reservoirs. Shale production depends on kerogen quality, maturity, fracture stimulation, and reservoir properties. Nano-Darcy permeability poses challenges, impacting oil properties over time and complicating forecasting. Decline Curve Analysis (DCA) is a common but, alone it may not be accurate, resulting in under- or over-estimation of Ultimate Recovery. Numerical simulations struggle due to shale complexity, slow convergence from nano-Darcy permeabilities, and small grid size. This paper combines DCA and machine learning to forecast crude oil production, analyzing US shale data to capture intricate relationships between reservoir parameters and production decline.
The dataset has over 1000 wells from a US shale gas play, with monthly production (up to first 12 months), GOR, WC, and key drivers (depth, porosity, velocity, modulus, proppant, rate, stages, ISIP). Workflow starts with DCA Arps method. Rigorous DCA fitting obtains 3 parameters: initial rate, b-factor, decline rate. ML regression models (MLR, RFR, KNN, XGBoost and SVM) have been built to predict DCA parameters as a function of key reservoir parameters. Error metrics (RMSE, MAPE, R2) help in selecting the best model. A hybrid DCA-ML approach helps develop consistent forecasting methodology and generate reliable forecasts based on reservoir properties.
Exploratory data analysis (EDA) handles outliers and missing values for a representative dataset. Production data was cleaned through a streamlined workflow for reliability. Cleaned data, fitted with DCA, b-factors, decline rates, etc., showed strong correlations with reservoir properties. Smaller ISIP values meant faster decline rates and smaller EURs, while higher GORs correlated with higher b-factors. Machine learning effectively characterized well production by correlating it with reservoir properties. The XGBoosting model with 3-segmented ARPS DCA proved best for forecasting with R2 exceeding 70% for high accuracy. Hypertuning and Monte-Carlo cross-validation avoided overfitting.
This study can be used to enhance reserves estimation and production forecasting using ML to identify correlations between reservoir properties and DCA parameters. The objective is to reduce uncertainty by establishing meaningful relationships with production drivers. The workflow handles noisy production data, identifies critical drivers, analyzes trends, and generates accurate DCA guidelines for forecasting shale wells.
