Page - 124 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3283 3.2.3. EstimatingtheYear-AheadConsumption Theyear-aheadloadaimstoutilize the trendof theannualelectrical loadbyshowingthepower consumptionof thesameweekof thepreviousyear.However, theelectrical loadof theexact same weekof thepreviousyear isnotalwaysusedbecause thedaysof theweekaredifferentandpopular Koreanholidaysarecelebratedaccordingto the lunarcalendar. Everyweekof theyearhasaunique weeknumberbasedonISO-8601 [32].Asmentionedbefore, theaverageofpowerconsumptionsofall holidaysorworkdaysof theweekarecalculatedtowhichthepredictiontimebelongsanddepending on theyear, oneyear comprises52or53weeks. In thecaseof an issue suchas theprediction time belongs to the53rdweek, there isnosameweeknumber in thepreviousyear. Tosolve thisproblem, thepowerconsumptionof the52ndweekfromthepreviousyear isutilizedsince the twoweekshave similarexternal factors. Especially,electrical loadsshowvery lowconsumptiononaspecialholiday like theLunarNewYearholidaysandKoreanThanksgivingdays [35]. Toshowthisusagepattern, theaveragepowerconsumptionof thepreviousyear’s specialholidayrepresents theyear-ahead’s special holiday’s load. Theweek number can differ depending on the year, so representing the year-ahead’sspecialholidaypowerconsumptioncannotbedonedirectlyusingtheweeknumberof theholiday. This issuecanbehandledeasilybyexchangingthepowerconsumptionof theweekand theweekof theholiday in thepreviousyear. Figure3showsanexampleofestimating theyear-ahead consumption. If thecurrent timeisMondayof the33rdweek2016,weuse the33rdweek’selectrical loadof the lastyear. Toestimate theyear-aheadconsumptionofSundayof the33rdweek,weuse the averageof theelectrical loadsof theholidaysof the33rdweekof the lastyear. Figure3.Exampleofestimatingtheyear-aheadconsumption. 3.2.4. LoadForecastingBasedonLSTMNetworks Arecurrentneuralnetwork(RNN) isaclassofANNwhereconnectionsbetweenunits forma directedgraphalongasequence.Unlikea feedforwardneuralnetwork(FFNN),RNNscanuse their internal stateormemorytoprocess inputsequences [36]. RNNscanhandle timeseriesdata inmany applications, suchasunsegmented, connectedhandwriting recognitionor speech recognition [37]. However,RNNshaveproblemsinthat thegradientcanbeextremelysmallor large; theseproblems are called the vanishing gradient and exploding gradient problems. If the gradient is extremely small,RNNscannot learndatawith long-termdependencies. Ontheotherhand, if thegradient is extremely large, itmoves theRNNparameters farawayanddisrupts the learningprocess. Tohandle the vanishing gradient problem, previous studies [38,39] have proposed sophisticatedmodels of RNNarchitectures. One successfulmodel is long short-termmemory (LSTM),which solves the RNNproblemthroughacell stateandaunit calledacellwithmultiplegates. LSTMNetworksuse amethod that influences the behinddata by reflecting the learned informationwith theprevious data as the learning progresseswith time. Therefore, it is suitable for time series data, such as 124