Fracturing in horizontal wells influenced by high tectonic effects is challenging in terms of achieving rock breakdown and fracture propagation. Near-wellbore complexities also lead to insufficient injection rate, post-breakdown, to place proppant. A machine-learning (ML) model based on in-depth multidomain analysis can assist in such cases in the design and execution phase. Part I of the paper here covers the extensive ML modeling. The following Part II will cover the full implementation scheme applied on full well logs and complete data.
A total of 106 fracturing stages were analyzed across 12 wells with a structured database created with 52 fracturing-relevant parameters. The dataset for ML modeling was skimmed down to 24 inputs and 4 output parameters. These included different phases of the well, such as drilling and completion, processed openhole logs, perforation details, fracturing treatment parameters, and pressure diagnostics data. A placement quality index (PQI) was calculated with mass of proppant placed, rate achieved, pressures experienced, etc. with application of appropriate weights on each. The PQI used weighting techniques such as the analytic hierarchy process and entropy weight method. Multiple classification and regression algorithms were tested and used to learn from these inputs to predict stage placement and proppant placement success. An algorithm comparison was done to select the best performing algorithms for each of the different prediction tasks. A detailed data exploration, feature engineering, and data preprocessing was conducted to study the correlations, establish causality, scale the data and prepare it to train/test the models.
The proposed ML workflow in the study consists of a three-step process: (1) a classification model used to predict stage skipped, which is crucial as it influences the subsequent regression models. Results showed an excellent result in the predictions with an accuracy of 94%. (2) Multiple regression models were implemented to predict injectivity index, total proppant, proppant load, and the PQI. Predictions were evaluated using several evaluation metrics including R2 (varying from 0.86 to 0.93), root mean square error (RMSE), and mean absolute error (MAE). Results showed a good performance that varied across the different models. (3) A particle swarm optimizer algorithm was used downstream to optimize the perforation and treatment design to enhance the success ratio based on PQI prediction. The algorithm aimed to maximize the PQI by varying the parameters in the search space within reasonable and practicable ranges that was divided by completion type. Results showed an enhancement of 93% and 63% on low PQI section; 8% and 11% on mean values, for cased hole completion and for openhole completion, respectively.
This work is a first attempt to use ML in enhancing proppant placement. This approach can be used with the existing reservoir quality, completion quality, and geologic quality indices to append the understanding and design of treatments and perforations. The deployment plan will be conducted into existing commercial numerical models to assist the engineers during the design process.
