Seite - 133 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

verywellwhenashort-termdependencies,however, ifwehave to recognizea long-term dependency, theyarenotasefficient.Thismakes theLSTMsveryefficient inmodeluni- variate time series forecastingproblems.These typesofproblemsarebasedon learning fromtheseriesofpastobservations topredict thenextvalue in the sequence.Therefore, LSTMsaredesignedtoremember informationfor longperiodsof time.While theRNNs haveastructureofchainsofneuralnetworks, theLSTMfollowasimilar structure,how- ever, each part of the chain, instead of being a layer of neural network, are multiple layers,which interactbetween them. TheLSTMused for this study isdesignedwithTensowFlowandKeras. 3.3. Experiment configuration Toevaluateandvalidate the resultsof theLSTMmodelproposed in this study twotypes of experimentshavebeenproposed.On theonehand, a3-foldcrossvalidationhasbeen carriedout inorder toevaluateat ageneral level thebehaviourof theLSTM.Theresults of thisexperimentareshownanddiscussedinsection4.1.Ontheotherhandwecarryout a24-hourprediction.For thiswe train thenetworkwith90%of thedataand the remain- ing10%hasbeenusedasa test, choosingwholedaysof24hours randomly.The results of this experiment are shownanddiscussed in the section4.2.Forbothexperiments, the sameLSTMparameter settings areused.Table1 shows theexecutionparametersof the LSTM,after apreviousadjustmentcarriedout toobtain theoptimumparameters. Parameter Value Numberof inputneurons 100 Lot size 32 Numberofepochs 150 Learning factor 0.001 Optimizer Adam Activation function hyperbolic tangent Numberofdelaysequences 6 Table1. LSTMexecutionparameters for theexperimentsperformed 4. Resultsanddiscussion In this section the results obtained from the LSTM for the two proposed experiments are shown. Inorder to carryout the evaluationof thegoodness of the technique and the modelobtained the followingmeasurementsareused: • theRootMeanSquareError (RMSE) • theMeanAbsoluteError (MAE) • thePearsonCorrelationCoefficient (PCC) • Determinationcoefficient (R2) M.Á.Guillén-Navarroetal. /AnLSTMDeepLearningScheme forPredictionofLowTemperatures 133