Gas Well Liquid Loading Analysis with Classic and Hybrid Deep Learning Model Approaches

Abstract This paper presents a comparison of classic and hybrid deep learning models implemented for time series forecasting for predicting liquid loading time of gas wells in mature fields with reservoir pressure depletion and limited types of measurements. This type of forecasting can assist to determine the well remaining time, estimate the quality of applied workover operation (well blowing), optimize workover schedules, reduce NPT with corresponding minimizing of production losses. The initial approach relies on comparing wellhead and line pressures. Once it equalized, the well is recognized as loaded and included in workover plan. To add automated forecasting capabilities, classic and hybrid deep learning models were implemented. In both cases, the forecasting was based on historical high-frequency wellhead and line pressure measurements. To get additional estimation capabilities of the wellbore accumulated liquid volume, the dataset for hybrid model was enhanced with a data generated using a dynamic multiphase flow simulator, which calculated well behavior under various boundary conditions and scenarios. Both types of models were trained and tested on their respective datasets, and their forecasting capabilities were verified. The use of classic deep learning models made it possible to identify the well cycle status (stable, loading, or loaded), understand correlations between the parameters and their trends, forecast the values for several hours ahead, and estimate liquid loading time. Meanwhile, the hybrid model enabled the identification of liquid accumulation dynamics in the wellbore which would be useful for evaluating workover quality and its further scheduling. As additional output the unmeasured parameters such as water content, gas flowrate, casing head pressure, etc. were also estimated based only on wellhead and line pressure trends as input data. The study demonstrates that deep learning analysis of basic parameters, such as wellhead and line pressure, can provide valuable insights for gas production management and optimization. Additionally, the implementation of hybrid models integrates the advantages of dynamic multiphase flow simulators into real-time data analysis without requiring significant modeling and computational resources. These techniques offer a novel solution for real-time prediction of liquid loading time in mature fields with limited production data.