Abstract
Industrial demand response is essential for grid stability amid increasing variable renewable energy penetration. However, explicit demand response adoption is hindered by post-activation demand spikes, known as load rebounds. This paper proposes a methodology for transmission system operators to manage load rebounds in high-voltage networks with limited supply-side flexibility. First, a rolling-horizon model combining mixed-integer linear and constraint programming schedules an energy-intensive steel plant under implicit and explicit demand response programmes and generates feasible flexibility offers for balancing markets. Second, the industrial model is embedded in a modified high-voltage CIGRE stress-test network to simulate load rebounds after flexibility activations. Finally, in a safe deep reinforcement learning approach, a recurrent Proximal Policy Optimisation agent dispatches distributed demand-side flexibility and the reactive power of static var compensators to manage the network and mitigate load rebounds. A safety projection verifies each action of the agent with a power flow and, when an operational limit is violated, corrects it with a quadratic program built on exact power flow sensitivities; when no safe action exists, the action of minimal measured violation is applied. Compared with local voltage control alone, the agent reduces the hours with voltage violations from 12.2\% to 1.2\% and those with line loadings above the operational limit of 80\% of the thermal rating from 45.9\% to 3.6\%, with limited demand-side and no supply-side flexibility. The results support coordination schemes between transmission and distribution system operators in which distributed flexibility absorbs the rebounds of large industrial activations, enabling broader industrial participation in balancing markets through dynamic, voluntary flexibility.