Analysis of dredging productivity for trailing suction hopper dredger based on a CNN-TCN-Transformer model optimized with a GWO strategy

Accurately predicting Trailing Suction Hopper Dredger (TSHD) productivity is pivotal for efficiency but challenging due to high-dimensional time series data characterized by non-linear interactions and complex spatial-temporal dependencies. Standard Transformers often struggle to capture local spatial features and multi-scale temporal patterns within this data. To bridge these specific gaps, this paper proposes a novel hybrid CNN-TCN-Transformer model optimized via the Grey Wolf Optimizer (GWO). The architecture leverages a Convolutional Neural Network (CNN) to extract local spatial features, a TCN to capture multi-scale temporal dependencies efficiently, and a Transformer encoder to model global interactions. The framework utilizes dual-smoothing preprocessing and identifies key features through CatBoost and Spearman correlation analysis. Validated on real TSHD project data, the proposed model achieves a superior coefficient of determination (R²) of 0.9855, outperforming several benchmark models. GWO optimization further enhances convergence speed and robustness, providing a highly accurate framework for intelligent decision-support in dredging operations.