Heterogeneous analysis of residential electricity demand with machine learning-based causality models

In this thesis study, the factors affecting the prediction of energy consumption in residential buildings were examined using causality-based artificial intelligence models. Within this scope, one of the causal AI methods—the Causal Forest method—was used. This method identifies which variables influence electricity consumption predictions and determines how these variables should be modified in order to increase or decrease consumption. Before applying the causal AI method, the Isolation Forest method was used to detect and remove variables in the dataset that could negatively affect the prediction results. The dataset was compiled from the websites of TUIK (Turkish Statistical Institute), ECMWF (European Centre for Medium-Range Weather Forecasts), and the Ministry of Treasury and Finance of the Republic of Türkiye. The variables included in the dataset were: year, city, level of development of the city, annual average budget revenues of cities, temperature, population under age 14, population aged 15–64, population over age 65, single-person households, households consisting of at least one elementary family and other persons, households consisting of a single elementary family, households with multiple persons but no elementary family, net migration by city, average household size by city, and population growth rate. As a result of the study, when the treatment variable was selected as households consisting of a single elementary family, successful results were obtained. The analysis revealed that an increase in the number of such households led to an increase in electricity consumption in residential areas of the cities.