Urban agglomerations with a high economic activity and population density are key areas for carbon emissions and pioneers in achieving carbon peaking and the Sustainable Development Goals (SDGs). This study combines machine learning with an extended STIRPAT (Stochastic Impacts by Regression on Population, Affluence, and Technology) model to uncover the mechanisms driving carbon peaking disparities within these regions. It forecasts carbon emissions under different scenarios and develops indices to assess peaking pressure, reduction potential, and driving forces. The findings show significant carbon emission disparities among cities in the Yangtze River Delta, with a fluctuating downward trend over time. Technological advancement, population size, affluence, and urbanization positively impact emissions, while the effects of industrial structure and foreign investment are weakening. Industrially optimized cities lead in peaking, while others—such as late-peaking and economically radiating cities—achieve peaking only under the ER scenario. Cities facing population loss and demonstration cities fail to peak by 2030 in any scenario. The study recommends differentiated carbon peaking pathways for cities, emphasizing tailored targets, pathway models, and improved supervision. This research offers theoretical and practical insights for global urban agglomerations aiming to achieve early carbon peaking.

Accurately predicting future carbon emissions is of great significance for the government to scientifically promote carbon emission reduction policies. Among the current technologies for forecasting carbon emissions, the most prominent ones are econometric models and deep learning, but few works have systematically compared and analyzed the forecasting performance of the methods. Therefore, the paper make a comparison for deep learning model, machine learning model, and the econometric model to demonstrate whether deep learning is an efficient method for carbon emission prediction research. In model mechanism, neural network for deep learning refers to an information processing model established by simulating biological neural system, and the model can be further extended through bionic characteristics. So the paper further optimizes the model from the perspective of bionics, and proposes an innovative deep learning model based on the memory behavior mechanism of group creatures. Comparison results show that the prediction accuracy of the heuristic neural network is higher than that of the econometric model. Through in-depth analysis, the Heuristic neural network is more suitable for predicting future carbon emissions, while the econometric model is more suitable for clarifying the impact of influencing factors on carbon emissions.