Building a top-down method based on machine learning for evaluating energy intensity at a fine scale.

Energy intensity is an important representative of energy efficiency. Currently, most countries lack fine-scale energy intensity data, taking China as an example, it only published provincial energy intensity data. However, the published large-scale energy intensity cannot support the formulation of local policies. What's more, the research work about the evaluation of fine-scale energy intensity is rare. To solve this problem, a "top-down" method based on machine learning is proposed to evaluate the fine-scale energy intensity. Appropriate features were extracted from multi-source satellite data, then the performances of multiple machine learning models were compared. It is found that deep neural network reaches the highest level among these models. Therefore, it was selected to estimate city-scale energy intensity from the year of 2001–2017. It turns out that the energy efficiency of southeast cities is higher than that of northwest cities in China, and most cities are developing towards the direction of improving energy efficiency. Among all cities, the central ones are the fastest to improve energy efficiency. However, the energy efficiency of a few cities is found to reduce during this period. The proposed method can also be used in other countries to help governments save energy and reduce emissions. • Energy intensity at city scale was estimated in China using satellite data and machine learning. • The temporal and spatial characteristics of Chinese cities' energy intensity were analyzed. • Nighttime light is most important feature in estimating energy intensity. [ABSTRACT FROM AUTHOR]