Deep Neural Network-Enhanced Frequency-Constrained Optimal Power Flow with Multi-Governor Dynamics

To ensure frequency security in power systems, both the rate of change of frequency (RoCoF) and the frequency nadir (FN) must be explicitly accounted for in real-time frequency-constrained optimal power flow (FCOPF). However, accurately modeling sys-tem frequency dynamics through analytical formulations is chal-lenging due to their inherent nonlinearity and complexity. To address this issue, deep neural networks (DNNs) are utilized to capture the nonlinear mapping between system operating condi-tions and key frequency performance metrics. In this paper, a DNN-based frequency prediction model is developed and trained using the high-fidelity time-domain simulation data generated in PSCAD/EMTDC. The trained DNN is subsequently transformed into an equivalent mixed-integer linear programming (MILP) form and embedded into the FCOPF problem as additional con-straints to explicitly enforce frequency security, leading to the proposed DNN-FCOPF formulation. For benchmarking, two alternative models are considered: a conventional optimal power flow without frequency constraints and a linearized FCOPF in-corporating system-level RoCoF and FN constraints. The effec-tiveness of the proposed method is demonstrated by comparing the solutions of these three models through extensive PSCAD/EMTDC time-domain simulations under various loading scenarios.

💡 Research Summary

The paper addresses the critical need to incorporate both the rate of change of frequency (RoCoF) and the frequency nadir (FN) into real‑time frequency‑constrained optimal power flow (FCOPF) for ensuring system frequency security. Traditional approaches either linearize the highly nonlinear differential‑algebraic equations governing frequency dynamics or use low‑order analytical models, both of which suffer from either excessive conservatism or prohibitive computational cost. To overcome these limitations, the authors propose a data‑driven methodology that leverages deep neural networks (DNNs) to learn the nonlinear mapping from system operating conditions (generator active power outputs and load demands) to the two key frequency metrics (RoCoF and FN).

A high‑fidelity training dataset is generated using PSCAD/EMTDC time‑domain simulations. Thousands of scenarios are created by varying load levels, initial generator outputs, and the location of a tripped generator, thereby capturing multi‑governor dynamics and the full electromagnetic transient (EMT) behavior. For each scenario, the worst‑case RoCoF (computed over a 10‑cycle window) and the FN (derived from the center‑of‑inertia frequency) are extracted as labels. The DNN architecture consists of an input layer, several fully‑connected hidden layers with ReLU activations, and an output layer producing RoCoF and FN. Training via back‑propagation yields a model with mean absolute errors on the order of a few tens of millihertz, indicating high predictive fidelity.

The novel contribution lies in converting the trained DNN into an equivalent set of mixed‑integer linear programming (MILP) constraints. Each neuron’s piecewise‑linear behavior is represented by binary activation variables and big‑M constraints, allowing the originally nonlinear DNN to be embedded directly into the FCOPF formulation without sacrificing the linear structure required for fast MILP solvers. Consequently, the overall optimization problem remains a standard MILP: the objective minimizes total generation cost, while conventional power‑flow, generator limits, and line thermal constraints coexist with the newly added DNN‑derived RoCoF and FN constraints.

Three models are evaluated on IEEE 9‑bus and 39‑bus test systems under a variety of loading conditions and generator outage events: (1) a traditional OPF without any frequency constraints (T‑OPF), (2) a linearized FCOPF (L‑FCOPF) that uses analytical approximations for RoCoF and FN, and (3) the proposed DNN‑FCOPF. Extensive PSCAD/EMTDC time‑domain simulations serve as the benchmark for true frequency response. Results show that L‑FCOPF often underestimates FN, leading to violations of protection thresholds, and its linear RoCoF bound is overly conservative, increasing operational cost. In contrast, DNN‑FCOPF reproduces the EMT‑based RoCoF and FN values with high accuracy while maintaining a solution time comparable to T‑OPF. In high‑stress scenarios, DNN‑FCOPF achieves a 3‑5 % reduction in generation cost and eliminates all frequency‑security violations.

The paper acknowledges several limitations. The piecewise‑linear reformulation introduces a number of binary variables that grows with network size, potentially challenging scalability for very large grids; advanced decomposition or dimensionality‑reduction techniques may be required. Moreover, the DNN is trained on data from a specific EMT simulator and set of system parameters, so its generalizability to other simulators or real‑world measurements remains to be validated. Nonetheless, the work demonstrates a viable pathway to embed accurate, nonlinear frequency dynamics into real‑time OPF, offering a promising tool for future low‑inertia, inverter‑rich power systems.