Heterogeneous distribution energy resources, such as renewable energy sources and electric vehicles (EVs) impose prominent challenges to active distribution network planning. This paper proposes a bi-level stochastic optimization framework that enables the coordinated optimization of renewable generation, flexible loads, grid capacity, and energy storage through an EV integration mechanism with rejection sets. A mixed-integer non-linear programming model is constructed to minimize total planning costs under technical and operational constraints, where upper-level investment decisions are coupled with lower-level operations. For the long-short term uncertainties in EV demand and renewable energy output, a scenario-based decision-dependent distributionally robust joint chance-constrained approach is employed for effective characterization. Finally, the proposed bi-level model is transformed into a single-level optimization problem via linearization, significantly reducing computational complexity. Simulation results on the 25-node system, IEEE 141 and IEEE 300 system demonstrate that the proposed approach significantly reduces the total planning cost and alleviates voltage violations and line congestion, while maintaining computational scalability.
Power systems are required to operate consistently within a safe frequency range. In this paper, we propose a safe optimal frequency control framework based on reinforcement learning to generalize and improve the classical linear droop-based primary frequency control. We formulate an optimal frequency control problem with frequency limit constraints and reformulate it with chance constraints. To solve this problem, we develop a reinforcement learning framework based on the twin delayed deep deterministic (TD3) policy gradient algorithm. This framework learns actor-critic neural networks for control policy design via primal-dual updates on a Lagrangian-form actor loss function, where the dual variables adapt to constraint violations while stabilizing the policy optimization. Simulations of a 39-bus test system verify that in response to disturbances in uncontrolled (e.g., renewable) generation or load, our framework effectively learns nonlinear decentralized controllers, which can stabilize the system faster and reduce oscillation compared to representative existing methods, while restricting frequency within limits.
Natural and human-induced disasters can trigger widespread outages in distribution systems, which makes resilience improvement a pressing concern. However, existing methods are inadequate due to unaddressed disaster uncertainty and overlooked short-circuit current limitations that underpin system stability. This paper develops a short-circuit current constrained preventive-restorative resilient control for distribution systems under disaster uncertainty, which is modelled as two-stage stochastic programming problem with conditional value-at-risk. The model jointly optimizes the preventive deployment carried out before a disaster and the restorative scheduling performed afterwards for mobile energy storage systems and other resources. The first stage fixes the optimal pre-disaster siting, while the second stage responds to faulted-branch topology changes and coordinates the charging and discharging profiles. This model considers short-circuit current constraints across both prevention and restoration stages. Case studies on a modified IEEE 33-node system reveal significant impacts of these short-circuit current limits. Validation on the IEEE 123-node system demonstrates the scalability of the proposed framework.
To address the difficulty of source-load balance matching in power systems with high penetration of renewable energy, this paper proposes a multi-time scale automatic "scenario-strategy" matching method for power system dispatch. Firstly, a scenario matrix containing the time series of system states and temporal constraints of flexible loads is constructed, and typical scenarios are identified and prioritized using multidimensional criteria. Secondly, a multi-objective dispatch optimization model is designed to generate flexible load strategy sequences adapted to state evolution. Finally, a nonlinear mapping between scenarios and dispatch strategies is established based on the LSTM model and deep learning. Case studies show that, compared with traditional single-time scale strategies, the proposed method can effectively improve dispatch performance and provide important support for secure operation and dispatch of power systems under massive and complex operation scenarios caused by uncertainties of renewable energy.
The increasing penetration of distributed renewable energy and the mass integration of electric vehicles (EVs) in distribution networks (DNs), particularly the proliferation of high-power EV superchargers, introduces sudden and high-power spikes that jeopardize scheduling efficiency and voltage stability. Consequently, there is an urgent need for rapid and accurate optimal scheduling methods to ensure safe DN operation. To address this challenge, a physics-informed voltage-aware deep reinforcement learning (PIVA-DRL) scheduling framework for DN optimal scheduling is developed. Firstly, to provide precise and robust real-time voltage perception faced with incomplete DN power flow model, a physics-informed compensated residual network (Pi-CORNet) that combines physics model with residual learning is proposed. Next, to achieve faster policy convergence and improved adherence to operational constraints, the Pi-CORNet is used to lead DRL agents to make informed decisions. Case studies conducted on the refined IEEE 141-bus DN demonstrate the proposed method can capture accurately voltage profiles under dynamic conditions and achieve faster convergence speed, thus improve the stability of DN operation.
