GAN-KANInformer-Based Prediction of Underwater Temperature-Pressure Sensor Data
Abstract
Accurately predicting short-term changes in interconnected temperature and pressure variables is essential for effective monitoring of subsea oil and gas transportation systems. However, underwater sensor data collected from the field are often limited and show temporal dependencies, nonlinear relationships between variables, and varying operating conditions. To overcome these challenges, this study proposes a two-stage forecasting framework called GAN-KANInformer. First, after splitting the data chronologically, a Wasserstein generative adversarial network with gradient penalty is trained solely on the normalized real training data to generate additional multivariate sequences. The synthetic data are assessed for similarity in distribution, consistency in first differences, preservation of autocorrelation, and maintenance of inter-variable correlation before being combined with the real training samples. Next, the KANInformer model is trained on this augmented dataset, where ProbSparse attention captures temporal interactions and KAN-based nonlinear mappings model the coupled relationships among the thermodynamic variables. Experiments used 1,500 observations recorded every 15 minutes from a temperature and pressure monitoring system in a deepwater oil and gas transport environment at a depth of 1,500 meters. Across ten independent trials, the GAN-KANInformer achieved normalized root mean square errors (nRMSE) of <inline-formula> <tex-math notation="LaTeX">$0.080~\pm ~0.004$ </tex-math></inline-formula>, normalized mean absolute errors (nMAE) of <inline-formula> <tex-math notation="LaTeX">$0.062~\pm ~0.004$ </tex-math></inline-formula>, and mean absolute percentage errors (MAPE) of 4.820% <inline-formula> <tex-math notation="LaTeX">$\pm ~0.213$ </tex-math></inline-formula> for temperature prediction. For pressure prediction, the results were <inline-formula> <tex-math notation="LaTeX">$0.041~\pm ~0.002$ </tex-math></inline-formula> (nRMSE), <inline-formula> <tex-math notation="LaTeX">$0.033~\pm ~0.002$ </tex-math></inline-formula> (nMAE), and 4.190% <inline-formula> <tex-math notation="LaTeX">$\pm ~0.181$ </tex-math></inline-formula> (MAPE), demonstrating the proposed framework’s effectiveness on this complex dataset under challenging operating conditions.