Visualization and prediction of energy consumption in smart homes

Abstract

Visualization and prediction of electrical energy can play an important role in managing the energy consumption at building level. Precise modeling of energy consumption is necessary in order to reduce consumption and thus reduce carbon emission. This paper focuses on the energy consumption of appliances normally used in a low energy consumption house. The dataset considered in this paper is collected from the freely available UCI machine learning repository. This dataset contains the records of 19735 instances of 29 attributes. Firstly, this paper uses a number of visualization tools such as box plot, correlation plot, commutative curves, and Pearson correlation map to find the impact of temperature, weather and humidity on energy consumption. It is found here that temperature and weather can contribute significantly to energy consumption. Secondly, the energy consumption in a smart house is predicted using a number of regression analysis such as using support vector regression (SVR), linear regression (LR), random forest (RF), multilayer perceptron regression (MLP) and elastic net. For this, both holdout and cross validation methods are performed. Results show that among these five models, RF exhibits the highest regression score or coefficient of determination and the lowest mean absolute percentage error. Thus, RF is a good choice for reliably predicting the household energy consumption.