No abstract

In reaction to the impacts of global warming, individuals are becoming increasingly conscious of their homes unchecked power usage, particularly the use of electrical energy for cooking, heating, refrigeration, dish-washing and drying. There is an increased concern about idle losses, expended by devices when not in use, which not only add to utility bills but also add to the waste of energy. Monitoring and controlling end-use electricity demand in residential buildings can have a significant impact on reducing peak demand and optimizing energy consumption that can be achieved in smart households with residential load control systems. This study benchmarked eight Machine learning based algorithms: Linear, Ridge and LASSO regression; Support Vector Machine; Multilayer Perceptron; Nearest Neighbor regression; Extra-Trees and XG-Boost on a pre-collected "appliance energy" data-set. The specified algorithms were benchmarked on error metric of: training and testing set Rsquared statistic; MAE; RMSE and also training time. Datapreprocessing and visualization was done to yield insight into data used. Firstly, un-tuned version of the eight algorithms were benchmarked, then model tuning via Grid-Search was carried for five algorithms and finally the effects of inclusion of certain features and varying parameters was tabulated and graphed. The least scores, on the specified error metrics, were obtained by the regression algorithms. The best scores were obtained by Extra-Trees and XG-Boost, which belong to ensemble algorithms of which Extra-Trees obtained best variance explanation (R-squared) scores of 98.94% on training set and 60.21% on testing set along with least scores on above specified error metrics.

No abstract