Carbon emission prediction in China considering new quality productive forces using a deep and cross learning modeling framework

New quality productive forces (NQPF), digital economy development, and artificial intelligence (AI) are increasingly discussed as potential drivers of sustainable urban development. This study develops a Multi-head Attention Deep and Cross Network (MADCN) framework to predict urban carbon emissions and examine the relative associations of technological factors with those emissions. The framework integrates feature-interaction modeling, attention mechanisms, and SHapley Additive exPlanations (SHAP) for model interpretation. Using an annual city-level dataset covering 275 Chinese cities from 2009 to 2021, which is further expanded into a continuous monthly sequence for predictive training, MADCN outperforms conventional statistical, machine-learning, and deep-learning baselines, achieving a Mean Squared Error of 406151.1, a Mean Absolute Error of 612.3, and an R? of 0.991 under the study’s random holdout design. Additional forward-validation tests on the cleaned sample also retain R? values above 0.99, further supporting predictive stability. SHAP results indicate that population, city size, urbanization rate, and GDP remain the dominant predictors, whereas NQPF, the digital economy index, and AI technology level display differentiated association patterns. In particular, AI is more consistently associated with lower predicted emissions, whereas NQPF and the digital economy show weaker direct associations but clearer interaction patterns when combined with AI. These findings should be interpreted as predictive and exploratory evidence rather than causal estimates, and they motivate further investigation into how technology integration relates to urban decarbonization.