Oil-in-water concentration prediction for offshore oil & gas produced water treatment using deep learning approach with hyperparameter optimization

Mahsa Kashani, Stefan Jespersen, Zhenyu Yang,
Oil-in-water concentration prediction for offshore oil & gas produced water treatment using deep learning approach with hyperparameter optimization,
Energy Reports,
Volume 14,
2025,
Pages 4275-4290,
ISSN 2352-4847,
https://doi.org/10.1016/j.egyr.2025.11.045.
(https://www.sciencedirect.com/science/article/pii/S2352484725006651)
Abstract: Effective control of the produced water treatment (PWT) process is critical for offshore oil and gas operations, both to comply with environmental discharge regulations limiting oil-in-water (OiW) concentrations and to enhance process efficiency. This study proposes a novel deep learning-based prediction model for OiW concentration in a de-oiling hydrocyclone system, aiming to address the complex nonlinearities and temporal dynamics inherent in offshore PWT processes. The model integrates convolutional neural networks (CNN) to extract spatial features and long short-term memory (LSTM) networks to capture temporal dependencies, while employing an auto-regressive (AR) feedback structure to dynamically incorporate prior predictions, thereby improving robustness against process uncertainties and fluctuations. To reach the optimal design of the proposed deep learning model, the network optimization problem is presented to optimize both the network’s structure and hyperparameters. The proposed problem is presented as a multi-objective optimization problem (MOOP), considering minimizing prediction error and network training time. Moreover, the non-dominated sorting genetic algorithm-II (NSGA-II) is employed to reach the Pareto optimal set. To assess the performance of the proposed approach, a comprehensive comparison is conducted between the results obtained from the proposed MOOP and its single-objective form. Moreover, the comparison incorporates both traditional and advanced models, including ARIMAX, Hammerstein, LSTM, CNN–LSTM, AR–LSTM, and the proposed AR–CNN–LSTM. Results show that the proposed model effectively captures the system’s spatial hierarchies and temporal dynamics. Using experimental data from a scaled offshore de-oiling pilot plant, the proposed model achieves superior predictive accuracy and more accurately represents the system’s nonlinear and time-dependent behavior compared to the benchmark models.
Keywords: System identification; LSTM; CNN–LSTM; De-oiling hydrocyclones; Process control; Hyperparameter optimization; NSGA-II