A framework for classification of non-linear loads in smart grids usingArtificial Neural Networks and Multi-Agent SystemsFilipe de O. Saraivan, Wellington M.S. Bernardes, Eduardo N. AsadaDepartment of Electrical and Computer Engineering, So Carlos School of Engineering, University of So Paulo. Av. Trabalhador So-carlense, 400,13566-590 So Carlos SP, Brazila r t i c l e i n f oArticle history:Received 1 September 2014Received in revised form13 January 2015Accepted 16 February 2015Available online 10 July 2015Keywords:Smart gridsNon-linear loads classificationHybrid intelligent systemsMulti-Agent SystemsArtificial Neural Networksa b s t r a c tThis paper proposes a general framework that uses the Artificial Neural Networks (ANNs) as aclassification tool of nonlinear loads in a simulated smart grid environment by using Multi-AgentSystems (MAS). The increasing of communication and computation infrastructure on devices installed onmodern power distribution systems allows new automated and coordinated control actions. This ismainly due to the ability to manage and process information and deploy actions in real-time mode. Oneimportant measurement tool is the smart meter, which will be present with all customers. Besides themeasurement function, it has the communication feature and also some computational processingcapability. Considering this base structure, the objective is to present methods to classify/identifynonlinear loads based only on current or voltage profiles measured by smart meters in this distributedcomputing environment. In this work, the MAS will manage the data and the tasks related to theclassification and the ANN will perform the classification, both tools have been developed in JADE/JAVAand Matlab environment, respectively. Test case using 4000 input signals distributed in eight classescorresponding to nonlinear medical electromedical loads have been used and 98.7% of the samples havebeen identified correctly.& 2015 Elsevier B.V. All rights reserved.1. IntroductionThe availability of new monitoring devices, communication andautomation infrastructure on power distribution systems intro-duces the new generation of distribution systems known as smartgrids. This is an evolution that allows the application of intelligentmethods on new functions such as primary feeder reconfiguration,system self-healing, management of different power sources, real-time pricing monitoring, demand-side management, power qual-ity improvement, to name a few 1. The smart grid is not limitedto distribution systems, in fact its general idea is considered to allGeneration, Transmission and Distribution infrastructure. How-ever, its visibility is strong on distribution systems, where thedirect interaction with consumers exists.Due to many remote controllable devices scattered on all overthe system, the smart grids that make use of them can beconsidered as a distributed computing system 2 and its modellingand simulation must follow those characteristics. Improving thepower quality is an important objective for smart grids and animportant characteristic is the voltage waveform. There are variousevents that affect the power quality in the distribution systems andthe presence of nonlinear loads introduces harmonics, responsiblefor the degradation of the waveform 3.This paper proposes a method for classification of nonlinearloads in a smart grid. The technique uses the Artificial NeuralNetwork (ANN) that discriminates and classifies signals from theloads. The classification is managed with agents which areintelligent devices modelled in a Multi-Agent System (MAS). Twoproposals (clientserver type 4) of the MAS have been devel-oped: (1) the model where the classification is performed by theagent present in the smart meter agent and the other agentreceives the result at the substation; and (2) a model where theagent in the smart meter sends directly the measured signals to aclassification agent situated in the substation. Benefits and dis-advantages of both models are discussed.The paper is presented as follows: Section 2 discusses thenonlinear loads identification in the context of smart grids.Section 3 presents the loads used in this study. The frameworkutilised to simulate the identification of nonlinear loads is describedin Section 4. Section 5 shows the computational simulations and,finally, Section 6 presents the conclusions.Contents lists available at ScienceDirectjournal homepage: 2015 Elsevier B.V. All rights reserved.nCorresponding author. Tel.: 55 16 3373 8152; fax: 55 16 3373 9372.E-mail addresses: filipe.saraivausp.br (F.d.O. Saraiva),wellingtonmayconusp.br (W.M.S. Bernardes), easadausp.br (E.N. Asada).URL: http:/www.sel.eesc.usp.br/lasee (F.d.O. Saraiva).Neurocomputing 170 (2015) 3283382. Smart meters and identification of nonlinear loadsThe presence of nonlinear loads in the power system intro-duces harmonics that can degrade the quality of the energy. Thistype of load can be found at domestic, commercial and industria