A Novel Hybrid Modeling and Optimization Approach for Energy Management in Industrial Utility Systems

Kurzfassung

Industrial energy utility systems must reduce operating costs and CO2 emissions while integrating variable renewable energy sources. Their volatile power generation complicates reliable system operation and requires continuous re-optimization. The rolling horizon approach addresses this challenge by enabling adaptive decision-making through iterative optimization with regularly updated input data. However, reliable real-time operation requires accurate component models and advanced optimization methods. On the modeling side, purely physics-based models become highly complex for systems with many interacting components. Simulation-based surrogate models often fail to fully capture real operating conditions, whereas experimental data-driven models have limited extrapolative capability under scarce data or changing operating conditions. To overcome these limitations, this work develops a hybrid modeling methodology that combines simulation and experimental data within a single surrogate model. Adaptive weighting prioritizes the most reliable information according to data quantity and noise, while the model can be retrained online as new data become available. The validation of this model using benchmark functions and a pilot-scale high-temperature heat pump demonstrates high accuracy, robustness, scalability, and extrapolative generalization. On the optimization side, nonlinear, nonconvex, and constrained energy management problems must be solved accurately within short computation times. To address the limitations of existing solvers, this work introduces BO-IPOPT, a hybrid optimization method combining the global exploration capabilities of Bayesian optimization with the fast local convergence of the Interior Point OPTimizer. Methodological advances include an efficient candidate selection strategy for high-quality IPOPT starting points, a unified surrogate modeling approach for equality and inequality constraints using an augmented Lagrangian with slack variables, adaptive trust regions to improve convergence, and a reduced number of user-defined hyperparameters, lowering computational effort while improving global search performance. Applied to benchmark problems and conceptual case studies of varying complexity, BO-IPOPT achieves high accuracy, robustness, and computational efficiency, particularly for high-dimensional problems. Compared with state-of-the-art optimizers, it obtains up to 120 % lower objective function values within the same CPU time. Finally, BO-IPOPT is applied to the energy management of a conceptual renewable steam generation system and a real-world case study from the German food and cosmetics industry. In both applications, the optimizer consistently outperforms state-of-the-art methods in solution accuracy and robustness, achieving up to 35 % lower objective function values within the same CPU time while ensuring feasible solutions. The influence of key parameters, including optimizer CPU time per rolling horizon iteration, forecast uncertainty, and optimization horizon length, is further analyzed to assess their effects on system performance and decision quality.