Day-ahead prediction of hourly electric demand in non-stationary operated commercial buildings: A clustering-based hybrid approach

As the crucial foundation of distributed energy management systems, a reliable prediction of dynamic energy consumption is urgently needed in the building sector. Advances in metering technology and machine learning methods provide both opportunities and challenges. In particular, a clustering-based hybrid model is established in this paper to predict the day-ahead hourly electric demand in hotel buildings, which were initially classified as the non-stationary operated buildings (NOB) due to the irregular electric temporal features. Specifically, with multi-dimensional attributes of 100 days considered by fuzzy c-means (FCM) clustering, similar days and hours were extracted stepwise into several homogeneous groups. Afterwards, the online modified predictor was established by the combination of support vector regression (SVR) and wavelet decomposition, taking the corresponding training samples extracted by FCM as the input. Clustered results of similar days testified the clear identification of stationary subsequences (Group IV) and non-stationary subsequences (Group I). Based on the validation, it can be concluded that the introduction of days-clustering can always improve the accuracy of day-ahead prediction, while the recognition of similar hours is necessary only for non-stationary subsequences. Finally, this paper offers an online short-term prediction approach for model-based energy management, based on the limited and fluctuating historical records.