Article Details

Abstract

Aiming at the problems of neural network structure redundancy, insufficient identification accuracy and model selection dependence on experience in complex nonlinear system modeling, this paper proposes an intelligent modeling method based on lower-bound smooth group L1/2 regularization and alternating identification. Firstly, an alternating identification framework composed of Forgetting Factor Recursive Least Squares and Extreme Learning Machine is constructed, and the nonlinear system is decomposed into the linear main part and the nonlinear dynamic remainder. Secondly, L_GL1/2 regularization is introduced to sparsely prune the extreme learning machine. Furthermore, an adaptive model combination strategy is proposed to dynamically select the optimal single model. Experiments on the power load data set show that the proposed method has a coefficient of determination R2 of 0.994 2 and a Mean Absolute Error (MAE) of 96.51, which is 32 % ～ 88 % lower than comparison methods. Ablation experiments verify that each module contributes effectively. In the simulation of a stochastic nonlinear system, the method converges to the order of 10-13 after three iterations. This method provides an efficient modeling approach for a class of nonlinear discrete systems.

Keywords

nonlinear system identification; alternating identification; L1/2 regularization; Extreme Learning Machine; model combination

References

1. [1] Zhang Yajun, Chai Tianyou, Yang Jie Alternating identification algorithm and application for a class of nonlinear discrete-time dynamic systems [J]. Journal of Automation, 2017, 43 (1): 101-113
2. [2] Huang G B，Zhu Q Y，Siew C K. Extreme Learning Machine: Theory and Applications [J].Neurocomputing，2006，70（1）：489-501.
3. [3] Ding Feng New Theory of System Identification [M]. Beijing: Science Press, 2013
4. [4] Wang Linzhu Nonlinear System Neural Network Modeling Based on Extreme Learning Machine [D]. Shenyang: Shenyang University of Technology, 2023
5. [5] Tang A D，Niu L F，Miao J Y，et al. Training Compact DNNs with L1/2 Regularization [J/OL].Pattern Recognition，2023，136：109206.https://www.sciencedirect.com/science/article/abs/pii/S0031320322006859.
6. [6] Hochreiter S，Schmidhuber J. Long Short-Term Memory [J].Neural Computation，1997，9（8）：1735-1780.
7. [7] Wang D H，Li M.Stochastic Configuration Networks: Fundamentals and Algorithms [J].IEEE Transactions on Cybernetics，2017，47（10）：3466-3479.
8. [8] Quasdane M，Ramchoun H，Masrour T. Sparse Smooth Group L0∘L1/2 Regularization Method for Convolutional Neural Networks [J/OL].Knowledge-based Systems，2024，285：111327.https://www.sciencedirect.com/science/article/abs/pii/S0950705123010754?via%3Dihub.
9. [9] Lyu B S，Wang Q，Xu Y L，et al. Weighted Multi-innovation Extended Stochastic Gradient Identification for Multivariable Hammerstein Nonlinear Systems based on Multi-signal Processing [J/OL].Measurement，2025，252：117256.https://www.sciencedirect.com/science/article/abs/pii/S0263224125006153.
10. [10] Chen W，Liu Q Y，Li H C，et al. The Prediction Performance Analysis of the Lasso Model with Convex Non-Convex Sparse Regularization [J/OL].Algorithms，2025，18（4）：195.https://doi.org/10.3390/a18040195.
11. [11] Shanmugam L，Lim C P，Rihan F A，et al. Dynamical Analysis of Nonlinear Physical Systems: Mathematical Modelling, Artificial Intelligence, and Applications [J].The European Physical Journal Special Topics，2025，234：1331-1339.
12. [12] Xu Y，Kurths J，Li Y G，et al. Data-based Nonlinear and Stochastic Dynamics [J].The European Physical Journal Special Topics，2025，234（3）：409-413.

Funding

Project of Shanxi Police Vocational College (YB-202503)

About the authors

Wang Linzhu (1996-), female, Han ethnicity, from Jinzhong, Shanxi Province, teaching assistant, master's student, research direction: nonlinear system modeling and intelligent control.