Date of Award

8-17-2024

Document Type

Thesis

Abstract

Researchers have long been engaged in the challenging task of production forecasting and uncertainty quantification, often by combining Decline Curve Analysis (DCA) models with Probabilistic algorithms. Also, deep-learning approaches have been explored for production forecasting. In this work, we present the application of three probabilistic algorithms combined with various deterministic DCA models. We have also applied two machine learning (deep learning) algorithms for production forecasting and uncertainty quantification. Finally, we have compared the results of ML algorithms with the probabilistic algorithms based on the obtained MAPE values to conclude whether probabilistic or ML algorithms are better for production forecasting. Our analysis commences with the utilization of three probabilistic algorithms, namely Approximate Bayesian Computation (ABC), Conventional Bootstrap (CBM), and Modified Bootstrap methods (MBM). Each algorithm was integrated with three deterministic DCA models: Arps, Duong, and Logarithmic Growth Analysis (LGA). We then harnessed the power of machine learning (ML) algorithms called long-short-term memory (LSTM) and Transformer neural networks. The hyperparameters for the LSTM algorithm were chosen using Bayesian Optimization. We conducted a comprehensive study on 400 gas wells from the Marcellus Unconventional shale basin, evaluating LSTM, Transformer, and each probabilistic-DCA combination. Our hindcasting, which employed 12, 24, 36, 48, and 60 months of historical production data (hind-casts) to forecast up to 96 months, yielded forecasts that are essentially 10th, 50th, and 90th percentiles, providing the P10, P50, and P90 estimates, respectively. These estimates, which we refer to as uncertainty bands, vividly demonstrate uncertainty quantification as we progress from 12-60 months of hind-casts. In simple words, as we increase the use of historical data from 12 to 60 months, the P10 and P90 bands tighten, depicting that uncertainty decreases in the forecasts. Furthermore, we present the Mean Absolute Percentage Error (MAPE) for each probabilistic algorithm-DCA combination and LSTM algorithm for each hind-cast length for comprehensive comparison and conclusive insights. Our findings are helpful and should inspire confidence in the potential of ML algorithms for production forecasting and uncertainty quantification. We demonstrate the superior performance of both ML algorithms (LSTM and Transformers), particularly for 12 to 36 months of hind-cast for MAPE values and uncertainty bands. The compression of uncertainty bands with increasing hind-cast lengths indicates a decrease in uncertainty as production history increases; a promising result. Furthermore, the MAPE value decreases as we extend the hind-cast period from 12 to 60 months, suggesting improved accuracy with longer hind-casts. The uniqueness of this work is in the comparison of the ML algorithms (Transformers and LSTM) with the probabilistic DCA approach as discussed above. Also, the proportionality scaling function (PSF) that is introduced in this work allows the analyst to be able to capture the uncertainty associated with machine learning forecasts in a relatively simple manner.

Handle

http://hdl.handle.net/11122/15545

Download

Share

COinS