Privileged information-driven random network based non-iterative integration model for building energy consumption prediction

Abstract Accurate building energy consumption (BEC) prediction plays an increasingly significant role in energy control and conservation. However, owing to the high level of randomness of BEC data, acquiring accurate prediction results is difficult. To further enhance the predicted precision and robustness, we herein present a non-iterative decomposition–integration model to realise BEC prediction. Our proposed method merges the ensemble empirical mode decomposition (EEMD), learning using a privileged information (LUPI) paradigm-based random vector functional link network (RVFL+), and support vector regression (SVR) to generate satisfactory results. In this model, EEMD is first adopted for the decomposition of historical energy consumption data. Subsequently, the decomposed subsignals are input into the RVFL+ network, and the features of high correlation with BEC are selected as privileged information to constrain the weight of the output layer of RVFL+ network in order to obtain their corresponding output RVFL+ models of each subsignal. Finally, the resulting RVFL+ output models are aggregated and input into the SVR model to acquire the prediction results, thereby breaking through the limitation of the single-prediction model and improving prediction accuracy. To verify our proposed model, five actual BEC datasets were employed for experiments: Jinan in China, Fairbanks and California in the USA, Vancouver in Canada, and Sydney in Australia. Experiment results indicated that our proposed EEMD-RVFL+-SVR method had better accuracy and anti-noise performance. Specially, the mean absolute percentage error of the proposed method, compared with the EMD-SVR, EEMD-SVR, EEMD-PSO-GA-SVR, EEMD-RVFL, ARIMA-SVM, FLS-SVM, EMD-RVFL+-SVR, CEEMD-RVFL+-SVR, SVR, RVFL, RVFL+, BPNN, and Wavelet neural network, was reduced by 65.07%, 37.20%, 40.61%, 52.27%, 44.66%, 34.76%, 50.01%, 39.06%, 43.04%, 53.77%, 46.38%, 41.82% and 56.14% respectively.