沁水盆地南部煤层气区块储层非均质性强，气井产能预测难度大，且压裂施工缺乏针对性设计，导致压裂后井间生产效果差异显著。为此，基于沁水盆地南部187口煤层气直井的地质、测井、压裂和生产数据，构建了基于多任务学习策略的随机森林算法的气井产能预测模型，并通过粒子群优化算法优化压裂参数。研究使用深度卷积自动编码-解码器处理测井曲线等非结构化数据，采用随机森林算法结合多任务学习策略，有效缓解了样本数据有限和泛化性能低的问题，使得模型在小样本数据下仍能保持较高的预测精度。分析结果表明：深度、施工液量和小粒径支撑剂用量是影响产能的主要因素；地质条件是决定气井长期产能的关键因素；压裂参数则主要影响气井的峰值产能。多任务学习的随机森林算法在小样本数据上表现出高预测精度，测试集中峰值30 d和5 a累产气量的决定系数（R²）分别为0.883和0.887。对6口新井的5 a累产气量预测R²达0.901，显示出模型在实际应用中的高准确性和稳定性。通过粒子群优化算法对压裂参数进行优化后的方案，能够显著提高气井的产能分类等级或提升气井的产能水平。优化后的预测单井产能比原实际方案提高了约153%至188%，显示出优化方案在实际应用中的显著效果。通过结合多任务学习和粒子群优化算法，成功解决了小样本数据下的产能预测及压裂参数优化问题。构建的产能预测模型和压裂参数优化算法为沁水盆地南部煤层气高效开发提供了理论支持和实践参考。

The coalbed methane (CBM) blocks in the southern Qinshui Basin exhibit strong reservoir heterogeneity, resulting in challenges for accurate productivity prediction of gas wells. Furthermore, the absence of tailored fracturing designs has caused substantial variations in post-fracturing production performance among adjacent wells. To address these issues, a predictive model for well production capacity was developed based on geological, well logging, fracturing, and production data from 187 vertical CBM wells in the southern Qinshui Basin. The model employs a random forest algorithm integrated with a multi-task learning strategy and utilizes a particle swarm optimization (PSO) algorithm to optimize fracturing parameters. A deep convolutional autoencoder-decoder was applied to unstructured data (e.g., well logs), and the integration of random forest with multi-task learning strategies effectively addressed limited sample sizes and poor generalization, ensuring high prediction accuracy under small-data conditions. The results indicate that well depth, fracturing fluid volume, and small-sized proppant dosage are the dominant factors affecting productivity. Geological conditions determine long-term productivity, whereas fracturing parameters predominantly affect peak production performance. The multi-task random forest algorithm achieved high accuracy on small datasets, with R² values of 0.883 for 30-day peak cumulative production and 0.887 for 5-year cumulative production in the test set. Furthermore, the R² for 5-year cumulative production predictions of six new wells reached 0.901, confirming the model’s robustness and reliability in field applications. The PSO-optimized fracturing parameters significantly improved the productivity classification and overall productivity levels of the gas wells. The optimized parameters increased single-well productivity by 153-188% compared to original designs, demonstrating substantial practical efficacy. The combined multi-task learning and PSO framework successfully resolves productivity prediction and fracturing optimization challenges under small-data constraints. The proposed model and fracturing parameter optimization algorithm provide theoretical support and practical references for efficient CBM development in the southern Qinshui Basin.