Continuous inertia optimization for PV-storage virtual synchronous machines via lightweight RBF networks: eliminating parameter discontinuity at the electromechanical edge

Abstract

Conventional piecewise-adaptive control for virtual synchronous machines in photovoltaic (photovoltaic)-storage microgrids suffers from non-differentiable jumps in virtual inertia at threshold boundaries, which excite high-frequency mechanical-electrical coupled oscillations and degrade the dynamic stability of electromechanical energy conversion systems under stochastic disturbances. To address this structural flaw rooted in discontinuous parameter scheduling, a continuous inertia optimization strategy empowered by a lightweight three-layer Radial Basis Function neural network is proposed. The RBF-NN maps real-time frequency error dynamics to smooth virtual inertia trajectories aligned with the rotor motion equations of synchronous machinery, while a holonomic coupling mechanism constrains the virtual damping to the theoretically optimal ratio of 0.707, ensuring the Virtual synchronous machine operates within the optimal damping region of its second-order mechanical dynamic model. Comparative simulations demonstrate that the proposed scheme reduces active-power overshoot by up to 84.9% relative to fixed-parameter VSMs and lowers the root mean square of the inertia time derivative by 76.1%, effectively eliminating parameter-induced discontinuities in the electromechanical control loop. DSP-level profiling on a TMS320F28335 controller gives a mean RBF-NN inference latency of 85.8 μs, while the maximum measured latency is 115.7 μs. Experimental measurements give a mean output-current total harmonic distortion of 2.31% ± 0.39%, with a maximum value of 3.32% under the investigated conditions. These results demonstrate favorable average computational feasibility and improved harmonic performance, although strict IEEE 519 compliance is not claimed.