Tsinghua's new power-grid method cuts steelmaking optimization from 24h to 30min

Ruike Lyu's PhD thesis from Tsinghua University, "Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems," tackles a critical challenge: industrial loads consume over 60% of China's electricity and offer massive flexibility for balancing renewable-heavy power grids, yet their market participation has been blocked by complex production constraints, incomplete data, and the computational cost of coordinating giant facility portfolios. The dissertation, released on arXiv in August 2026, delivers a full pipeline that makes industrial demand response computationally tractable.

The first key breakthrough is in modeling. By unifying Linearized State Task Networks and continuous Resource Task Networks, the method represents discrete and continuous industrial processes (like steelmaking's electric arc furnaces) in a form power-system optimizers can handle. In a representative steelmaking case, solution time collapsed from over 24 hours to under 30 minutes — a 48x speedup — while preserving modeling accuracy. For parameter estimation, a privacy-preserving identification approach fuses process knowledge with hourly smart-meter data, achieving errors of 5.2%-8.5% for cement and steel-powder production, more than halving the errors of conventional machine-learning baselines.

Beyond modeling, the work tackles flexibility aggregation and real-time coordination. A data-driven method transforms high-dimensional, nonconvex flexibility regions into compact linear representations: a steelmaking process with more than 10,000 binary variables is reduced to just 24-48 continuous variables with a modest 3.6%-10.3% error. On top of that, a co-optimization framework pairs dimension-reduced bidding with exact disaggregation, allocating power among tens of thousands of resources in milliseconds while keeping every device locally feasible. In a benchmark comparison, this reduces interaction costs by 40% relative to a simplified bidding strategy.

The combined contribution is a tractable, end-to-end pipeline — process modeling, parameter identification, flexibility aggregation, and grid interaction — that could unlock industrial demand response at national scale. For factory operators, this means the ability to bid into electricity markets without exposing proprietary process data or sacrificing operational feasibility. For grid operators, it represents a new, reliable tool to balance variable renewables using the vast flexibility embedded in heavy industry.