Abaphandi abavela eGrisi baye baphuhlisa i-PV yokubikezela i-PV ye-prosumer schemes usebenzisa i-federated learning, indlela yokufunda umatshini othumela uhlaziyo lwemodeli yendawo kwi-server ephakathi ukuze ilungiswe. Ukulinganisa kwabo kubonisa iziphumo ezothusayo xa kuthelekiswa noqikelelo olusembindini.
Umfanekiso: Iifoto zeBlue Coat, Flickr, CC BY-SA 2.0
Izazinzulu ezivela kwiNational Technical University yaseGrisi yaseAthene zicebise ngenoveli yePV yokuqikelela indlela ekhusela ubumfihlo beprosumer. Izikimu zeprosumer ezisebenzayo zixhomekeke kwiimodeli ezichanekileyo zokuqikelela imveliso yelanga, efuna idatha ebanzi, ukwenza ubumfihlo kunye norhwebo oluluncedo luyimfuneko. Indlela yabaphandi yokulungelelanisa olu rhwebo lusekwe kwi-federated learning (FL).
"Inkqubo yeFL iqala ngemodeli yehlabathi ekwabelwana ngayo nazo zonke izixhobo. Isixhobo ngasinye siqeqesha imodeli ekuhlaleni kwaye sithumela uhlaziyo kwiseva esembindini, apho zidityanisiwe ukuphucula imodeli, ”abafundi batsho. “Le modeli ihlaziywayo emva koko ibuyiselwe kwizixhobo zoqeqesho olongezelelweyo. Umjikelo weFL uphindaphindwa amatyeli amaninzi de imodeli yehlabathi ifikelele kokuchaneka okunqwenelekayo. "
Imodeli yeqela iqhuba ekuhlaleni kumatshini ngamnye kwaye ibandakanya inkumbulo yexesha elifutshane elide (LSTM) uyilo, iyunithi yokuyeka, kunye neeleya ezimbini ezixineneyo ezidityaniswe ngokupheleleyo. I-LSTM ibamba idatha elandelelanayo, ngelixa iyunithi yokuyeka inciphisa ukugqithiswa, kunye neengqimba ezixineneyo zinceda ekwenzeni uqikelelo lokugqibela.
Imodeli ikwasebenzisa ii-hyperparameters ukulungelelanisa imifuziselo ye-LSTM yendawo kunye neqela labaxhasi abafanayo kumncedisi ophakathi. Ezi hyperparameters, zibekwe phambi kokuba uqeqesho luqale, zilawula inkqubo yoqeqesho lomatshini wokufunda.
Ezinye iimodeli
"I-dataset phantsi koviwo ifunyenwe kwigridi yombane yaseTerni, eItali, equka idatha evela kwi-30 encinci yeprosumers yombane esebenzisa iinkqubo ze-photovoltaic zokuvelisa amandla," kuchaza iqela. "Ukulandela ukulungelelanisa, sahlula i-dataset kwiiseti ezimbini: isethi yoqeqesho loqeqesho lomzekelo kunye neseti yokuvavanya ukuvavanya ukusebenza kwemodeli kwiinkcukacha ezingabonakaliyo. Eli candelo lithobela ulwahlulo lwe-80-20, kunye nedatha esuka ngoJanuwari 2015 ukuya kuDisemba 2017 ekhethelwe uqeqesho kunye nedatha esuka ngoJanuwari 2018 ukuya kuDisemba 2019 eyabelwe uvavanyo."
Abaphandi ke bathelekisa imodeli ye-FL-LSTM kwidathasethi efanayo ngokuchasene neendlela ezininzi zokufunda. Eyokuqala ibiyimfundo yendawo, esebenza kwindawo yabucala ngokupheleleyo, kwindawo ehlala abantu. Okwesibini ibikukufunda okuphakathi, okubonelela ngokuchaneka okuphezulu kodwa kuncame ubumfihlo. Imodeli yesithathu i-FL iphuculwe ngobumfihlo obuhlukeneyo (DP) ukunciphisa ithuba lokuchonga igalelo lomntu ngamnye, usebenzisa iziphindaphindi zengxolo ezibekwe kwi-0.2, 0.25, 0.3, okanye i-0.4.
"Ukuvavanya ukusebenza kweemodeli, ii-metrics ezimbini eziphambili zisetyenziswa: i-matric absolute absolute error (MAE) kunye ne-root mean error square (RMSE)," kuchaza iqela. “Ukhetho lwe-MAE luvumela uphononongo olubanzi lweempazamo zemida yeemodeli zethu, ngakumbi ngenxa yokuqina kwayo ngokuchasene nabangaphandle - uphawu olubonakalayo lwedatha yethu. Ngapha koko, i-RMSE igxininisa ubuntununtunu kwiimpazamo ezinkulu, ezibaluleke kakhulu ekuvavanyeni ukuchaneka koqikelelo lwesizukulwana, njengoko iqaqambisa ifuthe lokutenxa okungaphezulu kune-MAE. ”
Iziphumo zibonise ukuba imodeli ephakathi yenza kakuhle, kunye ne-MAE ye-0.00960 kunye ne-RMSE ye-0.01687. Imodeli yeFL yayine-MAE ye-0.01993 kunye ne-RMSE ye-0.02872. Imodeli ye-FL-DP ene-multiplier yomsindo we-0.2 irekhode i-MAE ye-0.01857 kunye ne-RMSE ye-0.02669. Imodeli yendawo ine-MAE ye-0.02436 kunye ne-RMSE ye-0.04679, ngelixa imodeli ye-FL-DP kunye ne-multiplier yomsindo we-0.25 ibonisa i-MAE ye-0.02651 kunye ne-RMSE ye-0.03375. Iziphumo zokuphindaphinda ingxolo ka-0.3 no-0.4 azizange zibonelelwe.
“Ekukhangeleni umgangatho wengxolo oza kubonelela ngokusebenza okufanayo kokuphunyezwa kwe-non-DP FL siye sadibana nesiphithiphithi esimangalisayo. Umlinganiselo ochanekileyo wokusebenza kwengxolo wokusebenza ubonwe kwi-multiplier ye-0.2, engalindelekanga iziphumo ezingcono kune-FL, "iqela laphawula. "Uvavanyo lwethu lokuphindaphinda ingxolo ephezulu kuno-0.2 lubonise ukuthotywa okulindelweyo ngokuchaneka kwangaphambili kunye ne-0.4 yophindaphinda eyenza imodeli ingakwazi ukudibana."
Iqela lathi “esona sinyanzelo sibandakanya ubungakanani obulinganiselweyo bedathasethi malunga nenani labathengi abathatha inxaxheba. Olu phononongo lusebenza njengesiseko; Ukongeza iiprosumers ezininzi ekuhambeni kwexesha ngokuqinisekileyo kuya kwandisa ukusebenza kwe-FL kunye ne-FL-DP. Ngaloo nto engqondweni, iziphumo zethu zibonisa ukuba kwiiseti zedatha ezincinci kunye nabaxumi abambalwa abathatha inxaxheba, ukufunda okuphakathi kugqwesa i-FL ngokuchaneka, nangona zombini iindlela ziphakamisa idatha edibeneyo ekhoyo. Ngaphandle koku, iFL ibonelela ngeenzuzo malunga neendleko zabucala kunye nonxibelelwano. "
Babonise iziphumo zabo "Ukuxhobisa iindlela zokufunda ezidityanisiweyo zokugcina uqikelelo lwe-PV yabucala," esanda kupapashwa Iingxelo zamandla.
Lo mxholo ukhuselwe yi-copyright kwaye awunakuphinda usetyenziswe. Ukuba ufuna ukusebenzisana nathi kwaye ungathanda ukuphinda usebenzise omnye umxholo wethu, nceda uqhagamshelane: editors@pv-magazine.com.
Umthombo ovela pv imagazini
Ukuziphendulela: Ulwazi oluchazwe ngasentla lunikezelwa yi-pv-magazine.com ngaphandle kwe-Cooig.com. I-Cooig.com ayenzi lumelo kunye neziqinisekiso malunga nomgangatho kunye nokuthembeka komthengisi kunye neemveliso. I-Cooig.com ikhupha ngokucacileyo naliphi na ityala lokwaphulwa kwelungelo lokushicilela umxholo.
