To address the challenge of difficult identification caused by the high coupling between leakage signals and load fluctuations in natural gas pipelines, an intelligent leakage identification method based on CNN (convolutional neural network) fused with SENet (squeeze-and-excitation network) was proposed. In this method, the spatiotemporal features of pressure and flow data was fused through a dual-input branch structure. The channel weights of leakage-sensitive features was dynamically enhanced by SENet to suppress load interference. And the data scale was compressed via GMM (Gaussian mixture model) clustering combined with orthogonal test design. Spatiotemporal thermal map analysis, ablation experiment analysis, and comparative test analysis were conducted based on the simulation data of three real pipeline networks in China. The results show that the proposed method exhibits excellent performance in tree-shaped, looped, and mixed topological structures, with an accuracy rate of 96.8%, a 3.4%~5.2% improvement compared with traditional leakage identification models. Furthermore, the sample size is reduced from tens of millions to thousands through data compression, which achieves a balance between accuracy and computational efficiency. The research results provide an intelligent solution for high-precision and low-latency monitoring of natural gas pipeline leakage.