With the development of the census and monitoring of national geographical conditions in China, the availability of information has sharply increased. Progress in data mining methods and social application tools has provided a way for solving the problems of low resource allocation and high uncertainty in decision-making regarding planning. To relieve non-capital functions and serve the healthy development of the Beijing Metropolitan Area, we propose a new model of self-adaptive cellular automaton based on ensemble learning (EL-CA). The method is based on the data collected by monitoring geographical conditions and is guided by complex geocomputing that simulates city-scale evolution in Beijing. A comparison of predicted and real data for Beijing in 2015 demonstrated that the predictions made by the EL-CA model proposed significantly outperformed those by traditional cellular automaton (CA) models based on empirical statistics. Data on the geographical conditions in Beijing in 2007 and 2015 were employed in model simulation and training to predict the scale of the city in 2023. The urban agglomeration points in Beijing tended to be dense, the overall construction land tended to be saturated, and the growth rate of land use areas slowed. Results from the model also established that the construction land in Beijing is close to saturation from a quantitative perspective, and the potential urban expansion hotspots in the future are mainly concentrated in the Tongzhou District, the Daxing District, the Fangshan District, the south side of the fourth and fifth ring roads, and the southwest side of Pinggu District. These results can provide decision-makers in urban planning with supporting data and support Beijing to relieve Beijing of functions nonessential to its role as China’s capital.