Seite - 132 - in Short-Term Load Forecasting by Artificial Intelligent Technologies

Energies2018,11, 3283 (c) Cluster C Figure6.Feature importance inrandomforest. Table7shows theelectrical load forecastaccuracy for thepatternclassificationof similar time seriesfor2016. Inthetable, thepredictedresultswithabetteraccuracyaremarkedinbold. Forinstance, in thecaseofClusterA,while randomforest showsabetterpredictionaccuracy forpatterns1 to4, MLPshowsabetter accuracy forpatterns 5 to 8. Using this table,we can choose amore accurate predictionmodel for thepatternandcluster type. Table7.MAPEresultsof loadforecasting in2016. 2016 ClusterA ClusterB ClusterC Pattern MLP RF MLP RF MLP RF 1 3.339 3.092 3.705 2.901 2.736 2.475 2 2.199 1.965 4.395 3.602 2.987 2.731 3 2.840 2.712 3.343 2.990 2.853 2.277 4 4.165 3.472 3.794 3.978 3.517 2.568 5 7.624 9.259 8.606 15.728 4.229 10.303 6 4.617 5.272 5.404 6.172 5.159 4.894 7 3.816 4.548 9.199 8.860 3.686 4.718 8 6.108 6.402 5.844 6.768 2.152 2.595 Table8showspredictionresultsofourmodel for2017. ComparingTables7and8,wecansee thatMLPand random forest (RF) have amatched relative performance inmost cases. There are twoexceptions inClusterAandoneexception inClusterBandtheyareunderlinedandmarkedin bold. In thecaseofClusterC,MLPandRFgave thesamerelativeperformance. This isgoodevidence thatourhybridmodelcanbegeneralized. Table8.MAPEresultsof loadforecasting in2017. 2017 ClusterA ClusterB ClusterC Pattern MLP RF MLP RF MLP RF 1 2.914 2.709 4.009 3.428 2.838 2.524 2 1.945 2.587 3.313 3.442 2.622 2.474 3 2.682 2.629 3.464 3.258 3.350 2.583 4 5.025 4.211 4.005 5.116 2.694 2.391 5 7.103 11.585 9.640 20.718 3.300 15.713 6 4.503 6.007 5.956 7.272 6.984 6.296 7 3.451 3.517 13.958 12.386 3.835 4.443 8 6.834 6.622 7.131 8.106 2.562 3.722 132