Error-Driven Chained Multiple-Subnetwork Echo State Network for Time-Series Prediction

Hybrid echo state networks (ESNs), a type of modified ESN, have been developed to improve the prediction accuracy of ESNs. However, they have been criticized for their computational complexity, which makes it difficult to use them directly in industrial applications. In this article, an error-driven chained multiple-subnetwork ESN (CESN) is proposed to build a simple structured hybrid network and improve its prediction accuracy. For this reason, a chain topology is generated to gradually reduce the residual error, while each subnetwork is trained separately. The weight matrix for each subnetwork does not need to be optimized, which reduces the computational cost. Meanwhile, the optimal number of subnetworks is determined on the basis of a given application. The efficiency of the proposed CESN is tested on a Santa Fe Laser and a public building dataset. Compared with ESN, 70% of the test data have been optimized by CESN for the public building dataset.