Building demand management, machine-learning model, electric water heater, power prediction
The recent increase of smart meters in the residential sector has led to large available datasets. The electricity consumption of individual households/devices can be accessed in close to real time, and allows both the demand and supply side to extract valuable information for efficient energy management. Predicting electricity consumption should help utilities improve planning generation and demand side management, however this is not a trivial task as consumption at the individual household level varies with occupant behavior. In residential buildings, many loads have some power flexibility. One of them is water heater, which accounts for up to 20% of home daily electricity use. Conventional methods for water heater power prediction, which heavily rely on physical principles, have limited applicability as their performance is subject to many physical assumptions. Recently, black-box models have gained huge interest due to their flexibility in model development and the rich availability of data in modern buildings. Black-box modelling methods can be further categorized into two types, i.e., statistical methods and supervised machine learning (ML). While statistical methods are relatively easy to implement, they can only capture linear relationships among building variables. Since building operations are typically complicated and nonlinear, the resulting accuracy can be poor. Many ML-based, data-driven approaches have the ability to characterize and forecast total energy consumption of commercial data. However, a paucity of research applying data-driven methods have been tested on the hour ahead energy consumption forecasts for typical detached residential houses in the US. With the advances in smart metering, sub meter usage forecasts at the household-level is also gaining popularity for smart building control and demand response programs. This led us to develop a hybrid model to address the problem of residential hour and day ahead load forecasting through the integration of data-driven techniques. The developed forecasting models are built using three common ML algorithms, support vector machines(SVM), Gaussian Naire Bayes, and Random Forest. Performance comparison among these ML methods was carried out. The results suggest that all models were able to correctly predict a greater proportion of the actual power consumption with prediction accuracy yields between 94% ~ 96%. The SVM model performs the best, while the RF works the worst.