The Impact Analysis of Different Activation Functions on CNN-based Reduced OPF

The Optimal Power Flow (OPF) obtains economic generation dispatch considering transmission line constraints. The solution of OPF for a large power system is computationally complex due to many variables and constraints. OPF with only critical constraints called Reduced OPF (ROPF) is used to reduce computational complexity without compromising the accuracy of the solution. Suitable deep-learning algorithms can identify critical constraints. The performance of such deep learning algorithms relies on activation functions. However, there is less focus on deep learning-based ROPF and impact analysis of activation function on deep learning-based ROPF. Therefore, this paper proposes a CNN-based ROPF using different activation functions. The CNN algorithm is used to predict the critical constraints and hyperparameters of CNN are optimized with the Stochastic Gradient Descent (SGD) optimization algorithm. The OPF is performed using the Gurobi optimizer. The performance of different activation functions like Exponential Linear Unit (ELU), Leaky Rectified Linear Unit (Leaky ReLU), and Parametric Rectified Linear Unit (PReLU) are analyzed to identify the best one. The performance analysis is carried out on the IEEE 73 bus system. The results show that CNN with PReLU performs better than other activation functions. Also, significant time is saved with deep learning-based ROPF.