Cuili Yang

A self-organizing cascade neural network with random weights for nonlinear system modeling

A self-organizing cascade neural network is proposed for nonlinear system modeling.A method is derived to select the input and hidden units for the network.A simple way is derived to train the weights of the network.The proof of the convergence has been given.Predict the key parameters in a wastewater treatment plant using the network. In this paper, a self-organizing cascade neural network (SCNN) with random weights is proposed for nonlinear system modeling. This SCNN is constructed via simultaneous structure and parameter learning processes. In structure learning, the units, which lead to the maximal error reduction of the network, are selected from the candidates and added to the existing network one by one. A stopping criterion based on the training and validation errors is introduced to select the optimal network size to match with a given application. In parameter learning, the weights connected with the output units are incrementally updated without gradients or generalized inverses, while the other weights are randomly assigned and no need to be tuned. Then, the convergence of SCNN is analyzed. Finally, the proposed SCNN is tested on two benchmark nonlinear systems and an actual municipal sewage treatment system. The experiment results show that the proposed SCNN has better performance on nonlinear system modeling than other similar methods.

Design of polynomial echo state networks for time series prediction

Abstract Echo state networks (ESNs) have been widely used in the field of time series prediction. In conventional ESNs, the spectral radius of reservoir is always scaled to lower than 1 to satisfy the necessary condition for echo state property (ESP), while the sufficient condition is unregarded. Meanwhile, the output weights are always trained without considering the high order statistics of input signals. To solve above problems, the original ESN is extended to polynomial ESNs (PESNs) by employing the polynomial functions of input variables into output weights. Firstly, the reservoir of the PESN is built by the singular value decomposition (SVD) method. Secondly, the prime PESN (P-PESN) is implemented and its polynomial output weights are augmented by the high order statistics of inputs. Thirdly, the simplified PESN (S-PESN) is constructed by decomposing the polynomial output weights of the P-PESN into randomly generated polynomial nodes and tuned output weights. Furthermore, the regression matrix properties of P-PESN and S-PESN are theoretically analyzed, respectively. Finally, the simulation results show that the proposed PESNs obtain better performance than other methods in terms of prediction accuracy and learning speed.

Adaptive lasso echo state network based on modified Bayesian information criterion for nonlinear system modeling

Echo state network (ESN), a novel recurrent neural network, has a randomly and sparsely connected reservoir. Since the reservoir size is very large, the collinearity problem may exist in the ESN. To address this problem and get a sparse architecture, an adaptive lasso echo state network (ALESN) is proposed, in which the adaptive lasso algorithm is used to calculate the output weights. The ALESN combines the advantages of quadratic regularization and adaptively weighted lasso shrinkage; furthermore, it has the oracle properties and can deal with the collinearity problem. Meanwhile, to obtain the optimal model, the selection of tuning regularization parameter based on modified Bayesian information criterion is proposed. Simulation results show that the proposed ALESN has better performance and relatively uniform output weights than some other existing methods.

Dynamical regularized echo state network for time series prediction

AbstractEcho state networks (ESNs) have been widely used in the field of time series prediction. However, it is difficult to automatically determine the structure of ESN for a given task. To solve this problem, the dynamical regularized ESN (DRESN) is proposed. Different from other growing ESNs whose existing architectures are fixed when new reservoir nodes are added, the current component of DRESN may be replaced by the newly generated network with more compact structure and better prediction performance.n Moreover, the values of output weights in DRESN are updated by the error minimization-based method, and the norms of output weights are controlled by the regularization technique to prevent the ill-posed problem. Furthermore, the convergence analysis of the DRESN is given theoretically and experimentally. Simulation results demonstrate that the proposed approach can have few reservoir nodes and better prediction accuracy than other existing ESN models.

A Hybrid Intelligent Optimal Control System Design for Wastewater Treatment Process

Due to the characteristics of large lag and high nonlinearity, the optimizing operation of wastewater treatment process (WWTP) is difficult to be designed. To solve this problem, a control optimization system based on hybrid intelligent technology is proposed in this paper. This system includes a feed-forward compensator, axa0feedback supervision module, a pre-setting module and a soft-sensor module. To obtain the minimum energy consumption (EC) under effluent standards, the set-points of the dissolved oxygen concentration and nitrate nitrogen concentration are adjusted through feed-forward compensation, feedback correction and online estimation. Finally, the proposed approach is applied on the WWTP simulationxa0model. Compared to the proportional-integral- derivative (PID) and data-driven adaptive optimal controller (DDAOC) methods, simulation results of the method proposed in this paper show better performance. DOI: http://dx.doi.org/10.5755/j01.itc.46.3.16061