Natural Gas Transmission Forecasting Using CNN-LSTM Deep Learning Model

Oil and gas distribution through undersea pipelines requires continuous monitoring, especially when transmitting these resources through undersea pipelines, as operational failures can lead to significant economic losses and environmental damage. Accurate forecasting of natural gas pipeline conditions is essential for proactive maintenance and risk mitigation. This study presents a novel dual-input CNN-LSTM architecture enhanced with an attention mechanism to predict sink pressure in undersea natural gas pipelines one hour in advance. As one of the initial applications of this framework in such industrial settings, the model utilizes 12-hour sequences of historical sensor data and event-driven attention flags to enhance predictive accuracy. The CNN layer extracts spatial features from multi-sensor inputs, while the LSTM captures temporal dependencies across time steps. An attention module dynamically focuses on critical shutdown events, allowing the model to prioritize relevant segments of input data. To ensure optimal performance, hyperparameters were tuned using the Hyperband optimization algorithm, which balances computational efficiency and robustness. Additionally, the model incorporates separate input branches for sensor data and attention flags, enhancing both interpretability and predictive performance. The results showed about a 5.2% decrease in Mean Absolute Error (MAE), emphasizing how important it is to include spatial information for better predicting conditions in complicated industrial systems. These results highlight the value of hybrid deep learning architectures in complex industrial systems, particularly where both spatial and temporal patterns are crucial for accurate condition forecasting.