Estimating area inertia of power systems with a high share of RES using deep learning

Inertia, the ability to maintain frequency stability, is crucial for power system secure operation. With the increasing integration of inverter-interfaced renewable energy sources (RESs), represented by wind and solar power, the inertia of the power system rapidly declines and exhibits spatial–temporal variation. Estimating area inertia becomes more complex, yet more vital for power systems with high share of RESs. Most model-based inertia estimation approaches rely on a linearized and simplified representation to system frequency dynamics, limiting their accuracy in presence of large amount of virtual inertia provided by RESs due to its time-varying feature. Here, we propose time-series-based residual neural network (TS-ResNet), a deep learning model integrating one-dimensional convolution operation and residual blocks to estimate area inertia with a mix of synchronous inertia and virtual inertia. TS-ResNet extracts frequency dynamic features from nodal frequencies and tie-line powers utilizing probing signals without affecting system stable operation. Additionally, to enable model’s robustness to complex scenarios, a loss function with elastic net regularization is introduced for the training process. Numerical results on a 3-region AC/DC hybrid system demonstrate its high accuracy and low computation efficiency. It also generalizes to unseen time delays of virtual synchronous generators (VSGs), DC power transmission variations, and different RES shares, and demonstrates strong robustness under various noise levels. Our findings suggest that TS-ResNet offers a fresh perspective on incorporating data-driven approaches to inertia estimation.