Nowadays, echo state networks (ESNs) with a variety of model structures are available for industrial process time series predictions. However, most existing ESNs employ a uniform timescale for data knowledge learning, which obviously ignores the influences of multivariate status at different timescales to the prediction target, usually leading to unsatisfactory model approximation performances. In response to this problem, this paper proposes a multi-parallel cycle reservoir with jumps (MP-CRJ) which is embedded with the feature knowledge of different timescales contained in the multivariate data. The MP-CRJ uses a more concise and superior circular jump reserve pool and a more memorable leaky integral neuron-filled parallel structure able to reduce the spatial complexity resulted from parallel ESNs and relatively improve the dynamic diversity of reserve pools. In addition, grey relational analysis algorithms are used to select relevant variables contributing to the prediction in filtering unnecessary data information. Applying to practical plant data for methanol productions, it shows that the MP-CRJ can help increase the prediction accuracy while maintain prediction speeds, as well as enjoy better adaptions to dynamics of complex industrial processes.

No abstract

No abstract