Hybrid short-term forecasting of the electric demand of supply fans using machine learning. (May 2020)

• Record Type: Journal Article
• Authors: Runge, Jason; Zmeureanu, Radu; Le Cam, Mathieu
• Is Part Of: Journal of building engineering. Volume 29(2020)
• Publication Date: 2020-05
• Subjects: Multi-step-ahead forecasting -- Artificial neural network -- Ensemble -- Supply air flow rate -- Hybrid grey-box -- Electric demand -- Sliding window; Building -- Periodicals; 690.05
• DOI: 10.1016/j.jobe.2019.101144
• Languages: English
• ISSNs: 2352-7102

Abstract

Abstract: This paper presents the development and application of multi-step-ahead short-term forecasting models targeting supply fans installed in an institutional building. The models applied in this work consist of an artificial neural network (ANN) applied in order to forecast the future supply air flow rate of the fans (black-box approach), and a physical model coupled with the ANN applied in order to forecast the future electric demand of the supply fans (hybrid grey-box approach). The forecasting models use measurement data obtained at 15-min intervals in order to forecast the target variables over the next 6 h. The architecture of the ANN was found through an automated search in the training data set. The paper compares the results of selected ANN models with those from other machine learning techniques (support vector regression and ensemble methods) along with a simple forecasting approach. The results of this study show a better forecasting performance when compared with the results from other publications: the CV(RMSE) is 1.8–3.4% for the air flow rate, and 4.8–7.3% for the electric demand for all new models. The results demonstrate that automating the hyperparameter search of the ANN architecture can help alleviate the difficulty of manual parameter setting and achieve a high performing model. Highlights: Forecasting the performance of HVAC equipment over the next 6 h is challenging. Proposed optimized artificial neural network (ANN) model uses 15-mn measurements. Results show better forecasting performance compared with other studies. Batch and iterative techniques for retraining of ANN models are compared. Batch updating with random weight initialization gives best results.