ULeaD: A Hybrid Isolation Forest-LSTM Autoencoder Model for Unsupervised Gas Pipeline Leak Detection: The Kinyerezi-Ubungo Case Study
Abstract
Reliable leak detection on natural-gas transmission pipelines is limited by scarce confirmed leak records and by operating transients that resemble leaks. This study developed ULeaD, an unsupervised model that fuses Isolation Forest (point-wise) and LSTM Autoencoder (temporal) anomaly scores with validation-selected weights of 0.3 and 0.7 for the 12.3 km Kinyerezi–Ubungo gas transmission pipeline in Tanzania. Twelve months of one-minute SCADA records (2022) were split chronologically into training, validation and test partitions (70/15/15). Ten input parameters were selected through expert screening, redundancy analysis and validation-only ablation, and the alarm threshold and a three-of-five persistence rule were fixed before testing. Evaluation used three forms of evidence. On twenty-five unperturbed test windows of normal operation, ULeaD’s timestamp-level false-alarm rate was 1.73%, lower than that of either component. In 200 controlled synthetic-leak sensitivity experiments on the same windows (flow-only losses of 1–10% of inlet flow), ULeaD achieved mean precision of 0.89, recall of 0.94 and F1-score of 0.88 for 2–10% leaks, a pooled AUC-PR of 0.906 and a mean detection delay of 10 minutes, with statistically significant gains over both components in F1-score, AUC-PR and delay (p < 0.01). On the single confirmed field leak (28 December 2022), withheld from development, ULeaD alarmed 8 minutes after the reported onset, compared with 15 and 19 minutes for the components. ULeaD is a candidate decision-support tool for pipelines with little leak history, but its field performance requires validation against further confirmed events.