Multi-Reservoir Computing with Ordered Aggregation for Time Series Analysis

Reservoir Computing, known for its simple yet efficient architecture, excels in time series prediction tasks. In this study, we propose an "Ordered Aggregation Multi-Reservoir Computing Model (OAMRC)" for nonlinear time series prediction and classification. Built on a multi-layered architecture, the OAMRC model systematically aggregates reservoir states across time steps, capturing both short-term and long-term temporal features. By adjusting the number of layers and aggregation step sizes, the model flexibly captures multi-scale temporal patterns. Compared to existing multi-reservoir models, OAMRC enhances flexibility through precise aggregation weight adjustments. The model was applied to ECG signal classification in a five-class task, achieving a maximum accuracy of 92.7% with proper tuning of reservoir size and layers. Results highlight its capability to discern complex physiological patterns. Additionally, experiments on time series prediction tasks, such as the Mackey-Glass and Lorenz systems, demonstrate the OAMRC model's superior predictive performance over other state-of-the-art RC models, including Deep Reservoir Computing. In real-world applications requiring extended memory, such as traffic flow prediction, the OAMRC model achieves a normalized root mean square error of 0.04, showcasing its ability to capture prolonged trends and intricate correlations. A comparative analysis reveals its advantage over alternative models in addressing diverse temporal challenges, making it a robust tool for complex time series tasks. The implementation code of the proposed OAMRC model is publicly available at GitHub (https://github.com/yangxuesong556/OAMRC).