Assessing the impact of digitalization and artificial intelligence on urban energy consumption using an interpretable deep learning framework

Predicting urban energy consumption (EC) under digitalization and artificial intelligence (D&AI) remains challenging because feature interactions obscure input–output relationships. Using data from 275 Chinese cities, this study integrates the Decision Deep and Cross Feature Transformation Network (DDCFTN) with SHapley Additive exPlanations (SHAP). DDCFTN combines deep and cross networks, decision-tree-based feature processing, and feature transformation to capture higher-order interactions and nonlinearities. It outperforms evaluated baselines, achieving test-set RMSE, MAE, and R 2 values of 18,360.16, 12,915.71, and 0.97. Nadeau-Bengio corrected paired t-tests with Holm adjustment confirm significant improvements across all metrics, with robustness across seeds, target specifications, regional subsets, and city-grouped ten-fold cross-validation (mean R 2 = 0.89). Six D&AI variables rank among the ten most important predictors; DIF contributes 6.75% of total SHAP importance. DIF interactions are heterogeneous and nonlinear, sometimes reversing from negative to positive contributions. City-level out-of-sample sensitivity analyses broadly corroborate these patterns and indicate possible turning points, supporting city-specific energy policies.