Abstract
As the lifeline of energy supply for various offshore projects, accurately evaluating and predicting the operation status of submarine cables are the foundation for the reliable operation of energy systems. Based on fully mining the dynamic and static characteristics of submarine cable operation and maintenance data, this paper proposes a submarine cable operation status prediction method based on a convolutional neural network—bidirectional gated recurrent unit (CNN-BiGRU) integrating attention mechanism. Firstly, the evaluation index system of the submarine cable operation status is established by considering three key influencing factors including online monitoring, routine inspection, and static test. Then, the operation condition evaluation model for submarine cable is constructed based on the cooperative game theory and the multi-level variable weight evaluation. Finally, the CNN-BiGRU combined neural network model integrating the attention mechanism is established, and the historical operation data and condition quantification results (health value) are used as input characteristic parameters to predict the evolution trend of the operation status of the submarine cable. The case study shows that the proposed method can effectively predict the operation status of submarine cables, and the root mean square error of the prediction is as low as 1.36%, which demonstrates the superior performance compared with the back propagation (BP) neural network, CNN, long short-term memory (LSTM), CNN-LSTM, and other algorithms.
Introduction
Submarine cables serve as the critical infrastructure for offshore energy transmission, linking wind farms and oil platforms to mainland grids. Operating in harsh, inaccessible underwater environments, these cables are subject to complex stressors. Traditional monitoring often relies on single-parameter thresholds, such as temperature, which fail to account for the multifaceted nature of cable degradation. Furthermore, standard time-series models often struggle to integrate static characteristics (e.g., cable age) with dynamic data (e.g., real-time current or mechanical strain). This research introduces a hybrid deep learning model, CNN-BiGRU-AM, designed to predict the health status of submarine cables with high precision, enabling proactive maintenance strategies.
Methodology
The proposed framework consists of a comprehensive evaluation system and a sophisticated neural network architecture.
Evaluation Index System
The researchers established a multi-level evaluation system comprising 21 indicators categorized into three primary layers: Online Monitoring, Routine Inspection, and Static Tests. To determine the importance of these factors, the study employed cooperative game theory, blending subjective expert insights (Improved Analytic Hierarchy Process) with objective data-driven weights (Entropy Weight Method).
The CNN-BiGRU-AM Model
The predictive engine utilizes three core components:
- Convolutional Neural Network (CNN): Extracts local, static features and identifies correlations between different indicators.
- Bidirectional Gated Recurrent Unit (BiGRU): Processes time-series data in both forward and backward directions. This bidirectional approach captures long-term dependencies more effectively than standard unidirectional models.
- Attention Mechanism (AM): Assigns weights to features, focusing the model on the most critical signals within the data stream and reducing the impact of noise.
Data and Training
The model was validated using 510 groups of real-world data from a 35 KV three-core submarine cable in the Bohai Sea, China, collected over 255 days. Data was normalized and cleaned before being split into training (88%) and testing (12%) sets.
Results and Discussion
The CNN-BiGRU-AM model outputs a "Health Value" (M) on a scale of 0-100, categorizing cable status into four tiers: Normal (≥90), Notice (90-75), Abnormal (75-60), and Serious (<60).
The hybrid approach significantly outperformed traditional algorithms (BP, CNN, LSTM, GRU, and CNN-LSTM) across several metrics:
- Accuracy: The model achieved a Root Mean Square Error (RMSE) of 1.36% and a Mean Absolute Percentage Error (MAPE) of 1.25%.
- Reliability: The maximum prediction error was capped at 3.35%, a substantial improvement over the standard BP neural network's ~14% error rate.
- Efficiency: Despite the added complexity of the Attention Mechanism, the GRU-based architecture remained more computationally efficient than LSTM-based alternatives due to its simpler gating structure.
Conclusion
This study demonstrates that integrating spatial feature extraction (CNN) with bidirectional temporal modeling (BiGRU) and attention-based weighting provides a robust solution for submarine cable health prediction. While the current model relies on some expert-derived weights and requires careful hyperparameter tuning to avoid overfitting, its superior accuracy marks a significant step forward in offshore energy reliability. Future research will focus on generalizing the model across different cable types and voltage levels.
Summary updated