Forecasting electric demand of supply fan using data mining techniques

This paper presents the application of the process of KDD (knowledge discovery in databases) for the forecasting of the electrical power demand of a supply fan of an AHU (air handling unit). The case study uses trend data from the BAS (Building Automation System), which is recorded every 15 min in a...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Energy (Oxford) 2016-04, Vol.101, p.541-557
Hauptverfasser: Le Cam, M., Daoud, A., Zmeureanu, R.
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:This paper presents the application of the process of KDD (knowledge discovery in databases) for the forecasting of the electrical power demand of a supply fan of an AHU (air handling unit). The case study uses trend data from the BAS (Building Automation System), which is recorded every 15 min in an office building. Data mining techniques are used as a preprocessing step in the development of the forecasting model. A clustering analysis detects atypical operations and then partitions the whole dataset into three subsets of typical daily profiles of the supply fan modulation. A hybrid model, combining a closed-loop nonlinear ANN (autoregressive neural network) model and a physical model, forecasts the electric power demand over a horizon of up to 6 h. The optimum architecture of ANN, found by using a Simple Genetic Algorithm, is composed of 13 input neurons, 1 hidden neuron and 23-day training set size, for the cluster corresponding to working days except Mondays. The results show good agreement between the forecasts and measurements of fan modulation, and electric demand, respectively. The fan modulation was forecasted over the testing period with RMSE (Root Mean Squared Error) of 5.5% and CV(RMSE) of 17.6%. The fan electric demand was forecasted with a RMSE of 1.4 kW, CV(RMSE) of 30% over a 6-h time horizon. The sensitivity analysis indicated that the reduction of training data set size from 23 days to 4 or 8 days does not have a negative impact of the value of RMSE. •Clustering analysis for typical daily profiles and atypical events detection.•Optimization of ANN architecture for the forecasting of fan modulation.•Forecasting of fan modulation and electric demand over the time horizon of 6 h.•Sensitivity analysis of the forecasting error when using a reduced number of training days.
ISSN:0360-5442
DOI:10.1016/j.energy.2016.02.061