Hybrid residual reservoir computing using dynamic memristor for time series prediction

Reservoir Computing (RC), as an efficient neural network framework, has demonstrated exceptional performance in time-series data processing and complex dynamic system modeling. For conventional memristor-based RC architectures, once the reservoir reaches a certain size, its prediction performance is difficult to further improve. This study integrates the characteristics of parallel and cascaded reservoirs while introducing a residual connection approach, proposing a hybrid residual reservoir computing architecture based on dynamic memristors. The performance of the proposed hybrid residual RC was evaluated and compared with the single, parallel, and hybrid architectures through tests on the Hénon map time series and chaotic Lorenz system prediction tasks. Numerical results show that, under the same reservoir scale exceeding a certain size, the proposed hybrid residual RC structure significantly outperforms other architectures. The normalized root mean square error (NRMSE) achieved 0.012 in the Hénon map prediction task, and 0.021 NRMSE was also attained in the Lorenz system prediction task, demonstrating good prediction accuracy. This study provides new insights into the development of high-performance implementations for dynamic RC system modeling for time series prediction.