Short-term multiple power type prediction based on deep learning

摘要: This paper proposes a method based on 4-layer deep neural network model by stacked denoising auto-encoders to analyze four types of power data: current (I), voltage (U), active (P) and reactive (Q). We collect 7 days household data. In the beginning, prediction accuracy rate can reach 82.45% when 1-h historical data are used predict for following 5 min. order optimize parameters this model, over 3-month period collected. The is 95.52% three-day next hour. Finally, supplemental experiments added verify that change has greater impact model. set as training set. Extract 2 weeks from 3 months data, 2-week divided into two test sets. effect 7:00 in morning 24:00 evening, 0:00 evening studied. rates 95.05% 99.02%, respectively. It shows higher with lower frequency consumption than consumption, we prove better 3-layer, 5-layer 7-layer models.