The Autonomous Profiling Sensors Trajectory Prediction Method Based on Deep Learning

Autonomous Profiling Sensors (APS) are autonomous marine observation devices capable of conducting stable, long-term monitoring of the ocean environment at preset time intervals. However, influenced by complex oceanic environmental factors, their drifting trajectories exhibit significant nonlinearity and randomness, posing considerable challenges for accurate trajectory prediction. Precise prediction of APS drifting trajectories is vital for optimizing deployment strategies, enhancing spatial coverage of monitored marine areas, and reducing deployment costs. By determining the optimal deployment locations based on predicted results, the probability of APS reaching targeted observation regions can be maximized, thus providing critical support for marine scientific research and environmental monitoring. Traditional numerical simulations and statistical models struggle to achieve accurate trajectory predictions due to low computational efficiency, heavy parameter dependence, and stringent data quality requirements. To address these issues, this study constructs a high-precision trajectory prediction model specifically for APS in the South China Sea region, based on the Informer deep learning model integrated with spatiotemporal fusion features, multi-source ocean environmental data, and a transfer learning strategy. Firstly, this study preprocesses data from the Global Drifter Program (GDP). A random forest model is utilized for feature selection, identifying 14 critical features affecting drifter drifting, including latitude difference, azimuth angle, and meridional velocity difference. Subsequently, the selected features are used to construct a source-domain dataset, which serves as the standard for establishing a target-domain dataset, ensuring consistency in data features. Building on this, a unimodal trajectory prediction model integrating spatiotemporal characteristics is developed. By embedding spatial and temporal information into the positional encoding module of Informer, the model effectively captures spatiotemporal dependencies within the trajectory data, overcoming the issue of information loss typically associated with long-sequence predictions in traditional models. Compared with traditional RNN, LSTM, GRU, Transformer, and Informer models, the proposed unimodal trajectory prediction model reduces the overall prediction error by 36.9%, 31.2%, 48.7%, 53.4%, and 18.6%, respectively, demonstrating significantly superior performance. Secondly, the study further integrates ocean environmental data and constructs a multimodal trajectory prediction model based on temporal features extracted from the unimodal model, accurately capturing complex relationships between APS trajectories and environmental variables. Experimental results indicate that compared to the unimodal trajectory prediction model, the multimodal model reduces the mean absolute errors for 6-hour, 12-hour, 18-hour, and 24-hour forecast intervals to 4.0633 km、4.8155 km、8.4299 km and 12.9916 km, respectively, resulting in an overall error reduction ofapproximately 47.8%. This highlights the substantial enhancement in predictive capability after integrating ocean environmental data. Furthermore, ablation experiments systematically evaluate the impact of different oceanic environmental variables on the model's predictive performance, further validating the effectiveness of the multimodal trajectory prediction model. Finally, employing a transfer learning approach, this study fine-tunes the multimodal trajectory prediction model using target-domain data from SVP3GI drifters deployed in the South China Sea. To comprehensively evaluate model performance, comparisons are conducted among two traditional numerical prediction methods and the original multimodal model. The optimized model demonstrates a 75.0% reduction in prediction error compared to the original multimodal model and achieves a 57.0% error reduction compared to numerical methods. Real-world APS data inputted into the optimized model yield a mean absolute error of 2.3700 km and a root mean square error of 0.03° within 24 hours, effectively confirming the reliability and accuracy of the multimodal APS trajectory prediction model in practical engineering applications. These outcomes not only illustrate the strong generalization capability of the multimodal trajectory prediction model but also provide a robust theoretical foundation for applying APS trajectory prediction technology in real-world ocean observations.