TIME SERIES PREDICTION OF CYCLE RESERVOIR WITH STEP JUMPS NETWORK

The traditional chaotic time series prediction based on echo state network has the problems of uncertain network structure and redundant internal structure of the reserve pool, resulting in low network prediction accuracy. To solve these problems, an improved deterministic cyclic hopping network is proposed. The topology of unidirectional ring connection was constructed, and the connection weights were shared to avoid network instability caused by random connections in the reserve pool and ensure the improvement of prediction accuracy. We designed a bidirectional step hopping mode to reduce the redundancy of network internal connections, reduced the complexity of the reserve pool, and effectively improved the speed of network construction. The experimental results of short-term prediction on chaotic time series show that the proposed algorithm has better performance in one-step prediction of chaotic time series.