An Ocean Time Series Prediction Framework Based on Improved PSO with Multiscale Learning

Accurate prediction of marine time series data is of great practical significance for the protection of marine ecosystems and the sustainable utilization of marine resources. Marine time series data, such as those containing marine parameters such as chlorophyll, turbidity, CDOM, etc., are characterized by high complexity and significant time periodicity. In this paper, oriented to the real ocean time series data prediction scenarios, we design an ocean time series prediction framework based on improved PSO and multi-scale learning for the problems of different inter- and intra-series correlations between ocean time series data on different time scales and the need to set the hyper-parameters individually on different datasets. The framework firstly extracts the significant periodicity in ocean time series data by frequency domain analysis technique and decomposes it to multiple time scales. Secondly, an adaptive graph convolutional network is incorporated to allow correlations between sequences to be independently explored and learned at different time scales, while a self-attention mechanism is incorporated to accurately capture correlations and dependencies within the time series. Finally, an adaptive and dynamic particle swarm optimization algorithm (ANDPSO) is proposed, which reduces the risk of the algorithm falling into a local optimal solution and helps the framework to select the key hyperparameters through the dynamic improvement of inertia weights and learning factors as well as the introduction of population diversity judgment mechanism. Experimental results show that the prediction performance of the framework outperforms the baseline model on a real ocean time series dataset, and ablation experiments are conducted to confirm the indispensability of each component, aiming to provide support for real ocean time series prediction scenarios.