Kangping Li

Recognition model of air conditioning electricity consumption behaviors for demand response based on k-means clustering

As a major part of household electricity load in summer, air conditioning (AC) consumption demonstrates enormous demand response (DR) potentials and numerous AC based DR programs are carried out annually. Understanding electricity consumption behaviors is one of the key elements for utilities to design more cost-effective DR programs. So studying on the AC electricity usage habits is of great significance to make full use of AC DR resources. In order to analyze the residential AC electricity consumption behaviors, a recognition model is proposed in this paper. First, the duty cycle curves (DCCs) are generated using non-intrusive load disaggregation (NILD) results drawn from the AC operating data. Second, those customers rarely using AC are identified and separated from the data set, meanwhile the remaining customers are classified into irregular and regular customers based on a density index. Third, k-means clustering is utilized to group regular customers into different clusters based on their typical duty cycle curves (TDCCs). Based on the clustering results, AC usage behaviors of customers in each cluster are analyzed. Finally, the effects of varies NILD accuracies on clustering results are taken into consideration, and the analysis results demonstrate that the recognition model is still feasible under the existing NILD algorithm precisions.

A classified irradiance forecast approach for solar PV prediction based on wavelet decomposition

A classified irradiance forecast approach for solar PV prediction is proposed based on wavelet decomposition. The Daubechies wavelet is chose to decompose the irradiance series measured in the PV plant into approximate component and detailed component. The trend and variability of irradiance series are estimated respectively based on the two components. Then all the available irradiance data are labeled according to the features extracted from the approximate and detailed components. In the end, multiple forecast models are built and trained to adapt to the irradiance series of different labels. The simulation results show the effectiveness of the proposed approach.