Modelling Residential End-Use Electricity Consumption Using Statistical and Artificial Intelligence Approaches and Determining the Effective Saving Measures

Abstract

Both appliance ownership and characteristics determine residential electricity consumption, the exact electricity consumption of each appliance depends on the usage patterns of the occupants. Various approaches have been used to determine the appliance specific electricity consumption at household level. One of the statistical approaches used since 1980s is the conditional demand analysis (CDA) which takes into account appliance ownership, appliance characteristics (such as volume, size, power, etc.), weather, and market data to disintegrate the billing data into appliance specific form. Since 1990s, neural network and fuzzy logic concepts which are artificial intelligence-based modelling approaches have been used to determine the appliance specific electricity consumption using various types of data available on appliance and occupant characteristics. An artificial intelligence-based approach which combines the prediction performance capabilities of neural networks and fuzzy logic is adaptive network based fuzzy inference system (ANFIS) models have been used in various subjects such as for the water quality, hourly electricity load demand, and air pollutant emission estimation. According to the investigations and researches, it has been observed that ANFIS approach have not been used to model end-use electricity consumption of the residential sector, yet. The aim of this study is to compare the prediction performances of CDA based and ANFIS based approaches to determine the electricity consumption based on the appliances at household level. An extended survey data which covers detailed information about 92 different types of appliances including all domestic and minor appliance properties, occupant characteristics, and billing information of 260 homes is used for developing the CDA and ANFIS models. Afterwards, prediction performances and capabilities of the CDA and ANFIS models to disintegrate the total electricity consumption into appliance specific forms are compared. According to this study, it has been found that after necessary multicollinearity check and regression analysis are performed, CDA model is developed with the MAPE of 12.1%. Variables which are received from the developed CDA model are used as inputs of ANFIS model as a first step of the model generation period. ANFIS model development period has been finalized with the seventh trial by calculating the MAPE of the testing data of ANFIS model as 17%. Besides the MAPE values, based on the error distribution of each approach, it has been observed that CDA model has significantly high prediction performance than ANFIS approach for the appliance-based electricity consumption estimation. As a last step of the study, different scenarios have been applied to estimate the appliance impacts on household electricity consumption. Scenarios are applied by considering the usage patterns of lighting, dishwasher and washing machine and percent decrease values in overall electricity consumption are observed as 0.7%, 4.3%,1.4% respectively.