The role of China is increasingly pivotal in climate change mitigation, and the formulation of energy conservation and emission reduction policies requires city-level information. The effectiveness of national policy implementation is contingent upon the support and involvement of local governments. Accurate data on final energy consumption is vital to formulate and implement city-level energy transitions and energy conservation and emission reduction policies. However, there is a dearth of data sources pertaining to China’s city-level final energy consumption. To address these gaps, we developed computational modeling techniques along with top-down and downscaling methods to estimate China’s city-level final energy consumption. In this way, we compiled a final energy consumption inventory for 331 Chinese cities from 2005 to 2021, covering seven economic sectors, 30 fossil fuels, and four clean power sources. Moreover, we discussed the validity of the estimation results from multiple perspectives to enhance estimation accuracy. This dataset can be utilized for analysis in various cutting-edge research fields such as energy transition dynamics, transition risk management strategies, and policy formulation processes.

Natural gas is often described as the cleanest fossil fuel. The consumption of natural gas is increasing rapidly. Accurate prediction of natural gas spot prices would significantly benefit energy management, economic development, and environmental conservation. In this study, the least squares regression boosting (LSBoost) algorithm was used for forecasting natural gas spot prices. LSBoost can fit regression ensembles well by minimizing the mean squared error. Henry Hub natural gas spot prices were investigated, and a wide range of time series from January 2001 to December 2017 was selected. The LSBoost method is adopted to analyze data series at daily, weekly and monthly. An empirical study verified that the proposed prediction model has a high degree of fitting. Compared with some existing approaches such as linear regression, linear support vector machine (SVM), quadratic SVM, and cubic SVM, the proposed LSBoost-based model showed better performance such as a higher R-square and lower mean absolute error, mean square error, and root-mean-square error.