A Dynamic Ensemble and Replaying Model for Online Marine Sensor Data Prediction

Abstract

Deep learning excels in time-series data mining, yet offline-trained models often degrade when faced with dynamic marine observation data. To address this, we propose a brain-inspired online learning and replay framework for efficient marine time-series data prediction. The proposed framework tackles concept drift not only by updating the parameters of its internal modules but also by employing an attention mechanism to adaptively assign importance to these modules, and incorporating a neuroscience-inspired memory replay mechanism for reinforcing past knowledge. Unlike traditional deep learning models reliant on extensive historical data, our framework enables cold-start learning and inference, making it ideal for environmental monitoring stations with limited data where offline models struggle to generalize. We further introduce the first marine data prediction benchmark dataset MarineDrift-1.0, covering key marine environmental indicators with natural conecpt drift. Experiments on this dataset demonstrate the model’s superior performance over state-of-the-art methods. Notably, the framework is model-independent, allows seamless integration with various models, delivering strong results even with simple architectures.