水力压裂参数优化是提高非常规油气开发的关键。针对目前油田压裂设计趋同化、压裂设计难以充分反映储层物性差异等问题，以玛湖砾岩油藏为研究对象，通过利用极端梯度提升（extreme gradient boosting, XGBoost）算法对目标储层压裂改造效果进行了深入评价，并结合热力学图版和特征参数相关性分析，明确了不同类型油井压裂效果的主控因素。随后，通过模拟退火参数优化算法，得出了能够最大化产量的参数组合。最后基于长短期记忆（long short-term memory, LSTM）、卷积神经网络（convolutional neural networks, CNN）和KAN网络（Kolmogorov-Arnold networks）算法的组合神经网络模型，对单井产量进行了高精度预测，预测准确率达到94%，为现场提供了有力的技术支撑，同时也为复杂油气藏的开发提供了理论支撑与实践参考。

The optimization of hydraulic fracturing parameters is identified as the key to enhancing unconventional oil and gas development. To address current issues, such as the convergence of oilfield fracturing design and the challenge of fully reflecting reservoir physical property differences in fracturing design, the Mahu conglomerate reservoir was taken as the research object. The extreme gradient boosting(XGboost) algorithm was applied to conduct an in-depth evaluation of the fracturing and modification effect of the target reservoir. By combining this with thermodynamic plates and correlation analysis of characteristic parameters, the main controlling factors of the fracturing effect for different oil wells were identified. Subsequently, a simulated annealing parameter optimization algorithm was employed to derive the parameter combinations that can maximize production. Finally, a combined neural network model based on long short-term memory(LSTM), convolutional neural networks(CNN), and Kolmogorov-Arnold networks(KAN) algorithms was utilized to predict the output of a single well with high accuracy, achieving a prediction accuracy of 94%. This provides strong technical support for the field, as well as theoretical support and practical reference for the development of complex oil and gas reservoirs.