Page - 135 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

Figure1. ComparisonbetweenthepredictionoftheLSTMandtheactual temperatureofthefirst200instances of test1 (fold1)of the3-foldcrossvalidationexperiments. Figure2. ComparisonbetweenthepredictionoftheLSTMandtheactual temperatureofthefirst200instances of test2 (fold2)of the3-foldcrossvalidationexperiments. RMSE MAE PCC R2 %Error1 %Error2 10%Test 0.6592 0.4582 0.9964 0.9928 1.46% 0.99% Table3. Resultsobtained for theexperimentsofpredicting24hours fordifferentdays As it can be seen, the results obtained are better than those of the previous exper- iment. The RMSE obtains an average error of 0.66 degrees Celsius. The fit of the R2 model is similar to the previous experiment. It isworthmentioning that the type 1 and type2errors increasea little.The type1errors reachalmost1.5%while the type1errors are almost1%.Thesmall increaseof theseerrorshas its origin in that topredict the last hoursof theday Idonothaveclose referenceson thebehaviorof the temperature. Finally,Figure4 showsgraphicallyhowthe trend inpredictionandactual tempera- turesare similar.Thisfigure shows the24hoursofFebruary17,2019. M.Á.Guillén-Navarroetal. /AnLSTMDeepLearningScheme forPredictionofLowTemperatures 135