Predicting changes in Southern California’s residential electricity consumption due to urban warming using machine learning models

Abstract:

Decades of urbanization has shifted over half of the global population to cities, a fraction that is expected to grow to 70% by 2050. Concurrently, cities are experiencing rising average temperatures and more frequent extreme heat waves from global climate change and the urban heat island effect. Together, these trends will alter the way energy is consumed, produced, and delivered and increase the need for cooling. Understanding energy-climate interactions can improve electric load forecasting, which is essential to maintain grid reliability and plan for future infrastructure investments, and can help craft policies that mitigate the economic and environmental consequences of rising cooling needs.

Currently, the spatiotemporal distributions of future energy consumption are not well understood and analytical techniques to anticipate these trends with available data are limited. Machine learning models are capable of learning complex relationships amongst factors that influence energy use behavior, but previous studies have been restricted by 1) limited access to high spatiotemporal resolution data, which reduces a model’s ability to accurately define the relationship between electricity consumption and input variables, and 2) a lack of statistically representative electricity data at a regional scale. In this study, we develop and optimize a general machine learning model that can be used to explore electricity-temperature relationship using smart-meter data from approximately 200,000 homes in Southern California, as well as site weather, building characteristics, and socioeconomic data. The impact of outliers, data resolution, and training with subsets (i.e. by climate zone, year of building construction, and season) on model performance is examined. The improved model is used to predict residential electricity consumption for the region under different climate scenarios. Results from the study provide insight into 1) the usefulness of machine learning for electricity demand forecasting, 2) the extent to which residential electricity consumption responds to ambient temperature, and 3) how urban warming is projected to impact future residential demand. The knowledge gained from this study can serve as a reference for energy management policies and climate adaption and mitigation plans.