A combined approach for predicting liquid loading onset via clustering and convolutional neural networks

Abstract Liquid loading is a common problem in gas wells, leading to decreased production and even cessation of production. Accurately predicting liquid loading onset is important for stable production. Existing models predict liquid loading onset based on different mechanisms, but no universally validated model exists for all gas well conditions. This paper presents a novel data‐driven approach for liquid loading onset prediction in various types of gas wells. The approach combines clustering‐based optimized model and convolutional neural networks (CNNs) to accurately classify the liquid loading state. A database of more than 1000 records was created by collecting data on gas wells from both the literature and fields, which was used to train and test the new model for predicting liquid loading onset. The comparison shows that the clustering‐based optimal model is superior to the traditional equation‐based empirical or semi‐empirical models for liquid loading onset predictions, especially when accounting for model applicability. The CNN model achieves a prediction accuracy of 92.5%–96.5% (average above 90%). When the two methods are combined, the overall prediction accuracy reaches 95%, exceeding the performance of either the optimized mechanistic model or the standalone CNN model. This combined strategy enhances robustness and demonstrates strong potential for broader field applications.