(403e) Production Forecast and Surveillance Using Data-Driven Methods

Production analysis is used to identify well issues and predict well performance and life. Historically, several empirical decline models have dominated the applications. These models are widely accepted; however, they do not account for any physical or operational parameters. A data-driven model that combines physical and operational information and production forecast and surveillance is built and validated based on Gulf of Mexico (GOM) field case studies.

This paper describes a workflow to analyze production from multiple wells in a field. The new method for production decline analysis is based on neural networks (NNs). NNs comprise a layered structure of interconnected “neurons” that have the predictors (inputs) on the bottom layer, the forecasts (outputs) on the top layer, and “hidden neurons” on intermediate layers. NNs are trained from historical data and can uncover hidden dependencies. Prediction models using NN can identify trends and seasonal patterns caused by perturbing effects.

Monthly production data from hundreds of wells recorded over several years were partitioned into training. The goal was to predict future production and identify potential issues in advance. A lag-dependence estimate showed that a given month's value correlates with the previous 6 months data but little beyond that. Therefore, the NN was devised to take inputs based on this finding. Traditionally used models do not account for flowing pressure data, changing production conditions, or reservoir pressure with time. Whereas, this model can be tailored to individual fields and can take the physical parameters’ effect into account through training patterns (e.g., downhole flowing pressure). Each training pattern comprised seven values from the training set. The first six values were the input nodes and the last value was the output node. The network was then tested on the validation set.

The models developed in this work were validated with numerous wells from the GOM. Production data from these wells were analyzed and used to demonstrate the application of the proposed model. The advantage of using NNs for prediction is that they learn from training sets only. These sets can be categorized by fields, and once trained properly, they can detect hidden and nonlinear dependencies, even when there is significant noise in the training set.