Abstract:This paper introduces a novel data-driven constraint screening approach aimed at accelerating the solution of convexified AC optimal power flow (C-OPF) by eliminating non-binding constraints. Our constraint screening process leverages a novel mixture-of-experts architecture called MoGE (Mixture of Gradient Experts), which is trained to predict optimal dual variables based on problem parameters. The results demonstrate that subject to certain mild conditions on the C-OPF model, our proposed method guarantees an identical solution to the full problem but significantly reduces computational time. The solution's accuracy is shown through C-OPF properties. Even with incorrectly screened constraints, a recovery process is possible. This results in a problem with fewer constraints than the original. Extensive simulations conducted on the quadratic-convex model show that our method outperforms other constraint screening techniques.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to accelerate the solution process in the convexified alternating - current optimal power flow (C - OPF) problem by eliminating non - binding constraints. Specifically, the paper proposes a data - driven method that uses machine - learning techniques to predict the optimal dual variables, thereby identifying and removing the constraints that are inactive at the optimal solution, reducing the problem size and speeding up the solution. This method not only guarantees the same solution as the complete problem but also significantly reduces the computation time. Even in the case of incorrectly screening constraints, a recovery process is provided to ensure the accuracy of the final solution.
### Main contributions of the paper:
1. **Propose a new constraint screening method**: Use the Mixture of Gradient Experts (MoGE) framework to learn the mapping from problem parameters to optimal dual variables. The predicted dual variables are used to identify non - binding constraints, thereby reducing the problem size and accelerating the solution process.
2. **Establish model - independent conditions**: Verify the strong duality and the uniqueness of the dual solution of the general C - OPF model, which is the key to classifying constraints using dual variables. These conditions are verified in the commonly used quadratic convex relaxation ACOPF model.
3. **Propose an efficient data generation method**: Utilize the convexity of C - OPF to develop a new data - augmentation technique, accelerate the generation of synthetic data, and improve the training effect of MoGE.
4. **Extensive simulation verification**: Conduct simulations on multiple test cases in the PGLIB OPF library to verify the effectiveness of the proposed method, especially its performance in handling large - scale C - OPF problems is better than other constraint - screening techniques.
### Method overview:
- **Parametric representation of C - OPF**: Represent C - OPF as a parameterized convex optimization problem, where the parameters include load demand and line - flow limits, etc.
- **MoGE framework**: Design a new neural - network architecture, MoGE, which combines the Input Convex Neural Network (ICNN) and the Monotone Gradient Network (MGN), and softly selects a more accurate output through the Gate network.
- **Data generation and training**: Generate training data through uniform sampling and convex combination to ensure the feasibility of the problem, and train MoGE using the mean - squared - error loss function.
- **Accelerate C - OPF solution**: Use the trained MoGE to predict dual variables, identify non - binding constraints, reduce the problem size, and accelerate the solution process. If misjudged constraints are detected, re - solve through the recovery process.
### Conclusion:
The method proposed in the paper performs well in accelerating the C - OPF solution, especially when dealing with large - scale problems. Through efficient constraint screening, the computation time is significantly reduced while the accuracy of the solution is guaranteed. This method provides new ideas and technical means for power - system optimization.