This paper puts forward the role of learning techniques in addressing the problem of an efficient and optimal centralized voltage control in distribution networks equipped with dispersed generation systems (DGSs). The proposed methodology employs a radial basis function network (RBFN) to identify the multidimensional nonlinear mapping between a vector of observable variables describing the network operating point and the optimal set points of the voltage regulating devices. The RBFN is trained by numerical data generated by solving the voltage regulation problem for a set of network operating points by a rigorous multiobjective solution methodology. The RBFN performance is continuously monitored by a supervisor process that notifies when there is the need of a more accurate solution of the voltage regulation problem if nonoptimal network operating conditions (expost monitoring) or excessive distances between the actual network state and the neuron's centres (ex ante monitoring) are detected. A more rigorous problem solution, if required, can be obtained by solving the voltage regulation problem by a conventional multiobjective optimization technique. This new solution, in conjunction with the corresponding input vector, is then adopted as a new train data sample to adapt the RBFN. This online training process allows RBFN to (i) adaptively learn the more representative domain space regions of the input/output mapping without needing a prior knowledge of a complete and representative training set, and (ii) manage effectively any time varying phenomena affecting this mapping. The results obtained by simulating the regulation policy in the case of a medium-voltage network are very promising. (c) 2007 Elsevier B.V. All rights reserved.

This paper puts forward the role of learning techniques in addressing the problem of an efficient and optimal centralized voltage control in distribution networks equipped with dispersed generation systems (DGSs). The proposed methodology employs a radial basis function network (RBFN) to identify the multidimensional nonlinear mapping between a vector of observable variables describing the network operating point and the optimal set points of the voltage regulating devices. The RBFN is trained by numerical data generated by solving the voltage regulation problem for a set of network operating points by a rigorous multiobjective solution methodology. The RBFN performance is continuously monitored by a supervisor process that notifies when there is the need of a more accurate solution of the voltage regulation problem if nonoptimal network operating conditions (ex post monitoring) or excessive distances between the actual network state and the neuron’s centres (ex ante monitoring) are detected. A more rigorous problem solution, if required, can be obtained by solving the voltage regulation problem by a conventional multiobjective optimization technique. This new solution, in conjunction with the corresponding input vector, is then adopted as a new train data sample to adapt the RBFN. This online training process allows RBFN to (i) adaptively learn the more representative domain space regions of the input/output mapping without needing a prior knowledge of a complete and representative training set, and (ii) manage effectively any time varying phenomena affecting this mapping. The results obtained by simulating the regulation policy in the case of a medium-voltage network are very promising.

### Voltage regulation in MV networks with dispersed generations by a neural-based multiobjective methodology

#####
*Vaccaro A;Villacci D*

##### 2008-01-01

#### Abstract

This paper puts forward the role of learning techniques in addressing the problem of an efficient and optimal centralized voltage control in distribution networks equipped with dispersed generation systems (DGSs). The proposed methodology employs a radial basis function network (RBFN) to identify the multidimensional nonlinear mapping between a vector of observable variables describing the network operating point and the optimal set points of the voltage regulating devices. The RBFN is trained by numerical data generated by solving the voltage regulation problem for a set of network operating points by a rigorous multiobjective solution methodology. The RBFN performance is continuously monitored by a supervisor process that notifies when there is the need of a more accurate solution of the voltage regulation problem if nonoptimal network operating conditions (expost monitoring) or excessive distances between the actual network state and the neuron's centres (ex ante monitoring) are detected. A more rigorous problem solution, if required, can be obtained by solving the voltage regulation problem by a conventional multiobjective optimization technique. This new solution, in conjunction with the corresponding input vector, is then adopted as a new train data sample to adapt the RBFN. This online training process allows RBFN to (i) adaptively learn the more representative domain space regions of the input/output mapping without needing a prior knowledge of a complete and representative training set, and (ii) manage effectively any time varying phenomena affecting this mapping. The results obtained by simulating the regulation policy in the case of a medium-voltage network are very promising. (c) 2007 Elsevier B.V. All rights reserved.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.