Optimal Horizontal Well Placement with Deep-Learning-Based Production Forecast in Unconventional Assets

Abstract Horizontal wells placement and production forecast of unconventional assets play critical roles in the success or failure of any given operation. Traditional reservoir simulation workflows are ineffective for unconventional assets and often lead to erroneous results in addition to being both cost- and time-prohibitive. This paper presents a streamlined data-driven workflow of optimal horizontal target placement in unconventionals coupled with deep-learning (DL) techniques to accurately forecast production rates. The presented framework relies on automated geologic and engineering workflows to map remaining oil, advanced algorithms to perform an optimized global search with 3D pay tracking, and statistical and DL-based techniques to assess neighborhood performance and geologic risks. The workflow handles multiple types of constraints, including configuration constraints like length, azimuth, and deviation range, as well as path constraints like zone, baffle, and fault-surface crossing. For production forecasting, we developed an Encoder-Decoder Long Short-Term Memory Networks (LSTM) architecture. The model combines three types of inputs, and the output is multi-step multiphase forecasts. The input data for each well consists of time-variant information (i.e., historical production), static well features (e.g., geology and spacing parameters), and known-in-advance control variables. In addition, we use quantile regression to estimate the confidence interval around the forecasts. The post-prediction process then aggregates the results by combining economic analysis, risk assessment, and operational restrictions. We successfully deployed this technology to a giant unconventional play in North America with more than 4000 wells. We identified an inventory of 700 potential horizontal targets with optimized completion design; 90 of them were in the low-risk category with estimated additional reserves of 55.6 MMSTB. After establishing a database of tens of thousands of historical hydraulic fractures using advanced data mining techniques, we defined key impacting features using advanced feature engineering techniques (combining key fracture features as well as deconvoluting geological effects using unsupervised learning). We then developed a predictive DL model using selected features and quantified the impact of each feature on the production of each well. Moreover, the model generates a probabilistic production forecast that allows operators to model future activities. This technology provides a robust, streamlined, fast, and accurate approach to identifying optimal horizontal well targets as well as examining historical hydraulic fracturing performance, using state-of-the-art machine learning workflows augmented by domain expertise. It provides a domain-infused feature engineering process, absorbed by an explainable DL architecture. It uncovers non-linear dependences on well features and provides fast prediction and uncertainty quantification.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.