深层煤层气直丛井生产解吸规律、游离气与解吸气的转换时机尚不明确，产能差异的主控因素仍不确定，影响产能的提高。为进一步评价产能主控因素，基于36口直丛井的生产动态参数，结合神经网络预测井底流压，形成了以初期米采气指数为核心，综合多种机器学习算法的产能影响因素评价新方法。结果表明：①Beggs & Bill模型和Gray模型对深层煤层气井底流压预测适用性差。单相气体模型随着产水量下降，整体误差变小。采用神经网络方法预测效果较好，与实测相对误差小于10%。②采用Kendall’s tau-b（肯德尔相关系数）分析，离散型主控因素为微构造位置，主要位于抬升区正向构造区；其次为裂缝发育程度，以较发育和发育为主。③联合套索回归-随机森林-决策树逐步剔除非相关影响因素，确定影响产能的连续型主控因素从大到小排序为：灰分含量、平均施工排量、入地总砂量、见气时返排率、垂厚、声波时差、伽马、平均施工压力、百目砂占比、气测平均值，工程因素对气井产能影响不能忽视。该方法结合多种机器学习算法的优点，操作性强，提高煤层气动态预测精确度，有助于优化压裂设计参数，为提高煤层气压后的产能提供科学依据。

The desorption production patterns of deep coalbed methane（CBM） vertical cluster wells, as well as the transition timing between free gas and desorbed gas, remain unclear. The dominant factors causing productivity differences are still uncertain, which hinders productivity improvement. To further evaluate the primary factors controlling productivity, a new method for assessing productivity-influencing factors was developed, based on the production dynamic parameters of 36 vertical cluster wells and using neural networks to predict bottom-hole flowing pressure. This method centered on the initial meter gas production index and integrated multiple machine-learning algorithms. The results showed that: 1） The Beggs & Bill model and Gray model exhibited poor applicability for predicting the bottom-hole flowing pressure of deep CBM wells, while the single-phase gas model demonstrated reduced overall error as water production declined. Predictions using the neural network method were more accurate, with a relative error of less than 10% compared to measured values. 2） Using Kendall's tau-b correlation analysis, the discrete dominant factor was identified as the microstructural position, primarily located in uplifted positive structural zones, with the secondary factor being fracture development, categorized mainly as “well-developed” or “developed.” 3） By combining lasso regression-random forest- decision tree algorithm to iteratively eliminate irrelevant factors, the continuous dominant factors influencing productivity were ranked in descending order as: ash content, average construction discharge rate, total sand volume pumped, flowback rate at gas breakthrough, net pay thickness, acoustic travel time, gamma ray log value, average construction pressure, percentage of 100-mesh sand, and average gas measurement value. Engineering factors were found to have a significant impact on gas well productivity and cannot be overlooked. This method leverages the advantages of multiple machine-learning algorithms, demonstrating strong operability and improving the accuracy of CBM dynamic predictions. It aids in optimizing fracturing design parameters and provides a scientific basis for enhancing post-fracturing productivity in CBM wells.