Prediction of chiller power consumption using time series analysis and artificial neural networks.

Number: pap. 539

Author(s) : FAN C., XIAO F., WANG S.

Summary

This paper compares the accuracy and computation times of two widely used modelling techniques, i.e. time series analysis and artificial neural network (ANN), for one-hour-ahead prediction of chiller power consumption. A linear Autoregressive Integrated Moving Average (ARIMA) model, a nonlinear Self-Excited Threshold Autoregressive (SETAR) model and two ANN models were built for the chillers installed in the tallest commercial building in Hong Kong. Clustering analysis is employed to preprocess the chiller power data prior to the building of time series models, resulting in four typical 2-day clusters from one-month of chiller power consumption data. Two ANN models are developed from the unclustered data.
The results show that the modified ANN model is the best performer, with a mean absolute percentage error (MAPE) below 5% for the out-of-sample predictions, but consumes the most computation time of over 200s compared with 10s for time series models. It was also found that clustering analysis clearly improves the performance of the time series model. The nonlinear SETAR model, however, does not perform better than the linear ARIMA model. Trade-off between prediction accuracy and computation time must be considered when selecting the appropriate prediction model of chiller power consumption in practice.

Available documents

Format PDF

Pages: 11 p.

Available

  • Public price

    20 €

  • Member price*

    15 €

* Best rate depending on membership category (see the detailed benefits of individual and corporate memberships).

Details

  • Original title: Prediction of chiller power consumption using time series analysis and artificial neural networks.
  • Record ID : 30009340
  • Languages: English
  • Source: Clima 2013. 11th REHVA World Congress and 8th International Conference on Indoor Air Quality, Ventilation and Energy Conservation in Buildings.
  • Publication date: 2013/06/16

Links


See other articles from the proceedings (424)
See the conference proceedings