A deep recurrent neural network-based droop control strategy for frequency stabilization in low-inertia power systems with high renewable energy penetration

• Proposed DDRNN-based droop control for dynamic frequency stabilization in low-inertia, high-renewable power systems. • Combines temporal learning and droop principles to adapt parameters in real time for nonlinear, time-varying grid dynamics. • Validated on IEEE 118-bus system achieving 98.9% prediction accuracy and ±0.06 Hz deviation limit. • Outperforms conventional droop and basic neural networks with a 47% improvement in stabilization performance. • Ensures fast stabilization, computational efficiency, and adaptability for practical real-time grid applications. In modern low-inertia power systems with high renewable energy penetration (REP), reduced system inertia and fluctuations in renewable power make it increasingly challenging to maintain frequency stability, exposing the limitations of conventional droop control methods. To address this challenge, a deep droop recurrent neural network (DDRNN) is developed by embedding recurrent neural learning within the classical droop framework. The DDRNN leverages temporal state dependencies to dynamically adjust droop coefficients, thereby capturing both short- and long-term frequency interactions. This formulation provides an adaptive, data-driven extension of traditional primary control that remains compatible with established grid operation practices while enhancing resilience against rapid fluctuations. The proposed approach preserves interpretability by retaining the physical droop structure, while augmenting it with data-driven adaptability that allows accurate representation of nonlinear system responses. The framework is generalizable to a wide range of grid operating conditions and scalable from microgrids to bulk power systems. The proposed DDRNN was validated using the IEEE 118-bus system with REP, demonstrating that the DDRNN approach achieves a 98.9 % prediction accuracy in frequency stabilization and confines frequency deviations within ± 0.06 H z of the nominal frequency, meeting stringent grid code requirements. The proposed DDRNN scheme significantly outperforms traditional droop control and basic neural network models, with a 47 % improvement. It not only reduces post-disturbance stabilization time but also demonstrates computational efficiency, stable convergence, and seamless adaptation to changing grid conditions and disturbances. These characteristics make the DDRNN approach suitable for real-time implementation, thereby enhancing its practicality and applicability.