Multiple nonlinear regression modeling of appliances energy use in a low-energy house

The energy consumption of appliances plays an important role in the aggregated electricity demand of the residential sector. This paper aims to develop a high-performance forecasting model to predict the appliances energy consumption of a low-energy house located in Stambruges, Belgium. The study utilizes multivariate analysis and machine learning techniques to construct linear and nonlinear multiple regression models. The data involves measurements of temperature, humidity, and weather from a nearby airport station and lighting fixtures. Temperature and humidity are recorded every ten minutes using sensors from a wireless network in different rooms of the house. This data includes 27 attributes and 19,735 records. Data prepossessing including log-function, square-root, and box-cox has been conducted to control nonlinearity. Principle Component Analysis (PCA) is used with the regression model to reduce dimensionality and eliminate collinearity. The result shows that the Gradient Boosting Regression (GBR) improves the model to an adjusted R-squared of 41.97%, and the nonlinear third order polynomial regression model raises the percentage to 55.12%, which duplicates the accuracy to third fold compared to published work. The residuals chart shows some patterns that may lead to potentially further improvement based on this regression model.