<p>Electric energy consumption is a critical indicator in steel production using electric arc furnaces (EAFs), determining production efficiency and costs. However, predicting EAF energy consumption faces significant challenges due to complex nonlinear metallurgical processes. Traditional EAF energy consumption predictions rely on purely data-driven algorithms, which may produce predictions that violate physical laws due to overreliance on data and exhibit poor generalization for data beyond historical scope. To address these limitations, this study proposes a novel hybrid modeling framework that synergistically integrates mechanistic principles with advanced data-driven architectures, enhancing prediction accuracy for electric energy consumption in EAF steelmaking. First, a mechanistic prediction model based on energy balance was constructed as the hybrid model’s foundation, ensuring predictions comply with physical principles and resolving poor generalization for edge data. Since mechanistic models inevitably contain assumptions, a data-driven architecture was developed to correct model errors. Through mutual information analysis, key input features were identified. Twenty advanced data-driven methods were comprehensively evaluated, covering traditional machine learning, deep learning, ensemble learning, and tabular deep learning. The Fick’s Law Algorithm provided intelligent hyperparameter optimization, marking the first application of diffusion optimization principles in metallurgical process modeling. Shapley additive explanations validated the data-driven architecture’s reliability. The proposed hybrid model was validated on an independent dataset of 500 industrial productions. MM-FLA-CatBoost achieved RMSE of 4.7126&#xa0;kWh&#xa0;t⁻¹, MAE of 3.0107&#xa0;kWh&#xa0;t⁻¹, MAPE of 0.0088, and <i>R</i>² of 0.9014. This approach provides valuable electric energy consumption prediction for EAF steel production and intelligent energy management.</p> Graphical Abstract <p></p>