Reactive power regulation of power system using an SVC with fixed, optimized, and CNN-based PI controllers

Abstract Fluctuating wind generation and rapid load variations in an isolated Wind-Diesel Hybrid Power System (WDHPS) cause oscillations in generator reactive power and bus voltage, thereby affecting the stability and reliability of the system. The present work proposes a hybrid optimization-learning control framework for WDHPS employing a Static Var Compensator (SVC) where a PI controller is used for generating the SVC control signal. The three different controllers are Fixed PI, Enhanced Snake Optimization (ESO)-PI, and Convolutional Neural Network (CNN)-PI, thereby determining the effectiveness of SVC reactive power regulation and voltage stability enhancement. First, the ESO algorithm is implemented for obtaining the optimal PI gains offline initially by minimizing the performance index Integral Time Absolute Error (ITAE). The gains from ESO, along with the deviations in terminal voltage, synchronous generator reactive power, induction generator reactive power, and SVC reactive power are then used in training CNN for learning the non-linear mapping between optimal controller gains and system dynamics. This work combines metaheuristic-based offline optimization (ESO) with data-driven gain prediction (CNN) and maintains an identical PI controller structure for fair performance comparison. The ESO-PI achieves the fastest settling time (≈0.010 s) of voltage deviation and minimum ITAE. A near-optimal performance is shown by the CNN-PI controller with a settling time (≈0.013 s). Under an additional unseen operating condition, CNN-PI demonstrates faster settling than freshly re-optimized ESO-PI, and confirms genuine generalization beyond the training dataset rather than mere reproduction of ESO-tuned gains. The robustness is further validated under a substantially larger (5%) disturbance magnitude, where ESO-PI and CNN-PI consistently outperform Fixed-PI. These results demonstrate that the CNN has learned a generalizable gain-mapping behaviour rather than merely reproducing ESO’s condition-specific optimum, underscoring the practical value of an adaptive/predictive tuning approach for varying operating conditions.