Cluster-wise deep learning framework for scenario-adaptive energy consumption and productivity prediction of cutter suction dredger

In the construction process under complex geological conditions, accurately real-time predict energy consumption and productivity of cutter suction dredger (CSD) is a crucial foundation for promoting the realization of low-carbon emission reduction goals in the dredging field. This paper proposes a data-driven framework that combines unsupervised clustering with multivariate time series prediction to achieve scenario-adaptive predictions of CSD construction performance. First, a feature selection strategy based on non-linear mutual information and linear correlation is used to identify key parameters from 256-dimensional construction data strongly related to energy consumption and productivity. Next, K-means clustering is applied to the two-dimensional data of cutter cutting torque and winch swing torque. This analysis defines different dredging difficulty scales, enabling classification and mapping various dredging scenarios. Finally, a multi-output prediction model named Dynamic ProbSparse TimesNet (DPSTimesNet) is developed, embedding a sliding window Fast Fourier Transform dynamic weight adjustment mechanism and integrating ProbSparse self-attention into TimesNet architecture. The framework performance is validated using construction data from Tian Jing Hao CSD in the Pinglu Canal project. Results show that the clustering closely correlates with cutter depth and geological data, with four dredging scenarios distinguished objectively. The DPSTimesNet model accurately predicts energy consumption and productivity across different scenarios, achieving R2 values above 0.9824 and 0.9704, respectively, outperforming baseline models. Compared to traditional global modeling methods, this framework dynamically adjusts the prediction strategies based on geological conditions, demonstrating excellent scenario adaptability. Operators can enhance energy efficiency and reduce construction costs by adjusting key control parameters based on real-time feedback from energy consumption and productivity predictions.