Decentralized System Identification Method for Large-Scale Networks

Traditional identification methods, including the prediction error method and the subspace method, encounter difficulties when identifying large-scale networks due to computational complexity, interactions between nodes and local observation. To tackle these problems, we propose a decentralized and computationally efficient system identification method for large-scale networks. The method enables subsystems to estimate and continuously update the states to obtain the parameter matrices of interactions between nodes and identify the system when only local information is available. The proposed method is divided into two parts. The first part initially identifies the system based on a subspace method. We give the upper bound of the error of initial identification result. The second part is an online identification algorithm based on the least-squares method and regularized approximation. In the online identification algorithm, subsystems estimate the system parameter matrices by the least-squares method and use regularized perturbation of the state estimation to get more precise identification result. It is proved that our method obtains consistent and convergent identification results Simulations demonstrate the effectiveness of our method.