Inertia Estimation for Power Generation Units Based on Dual Neural Networks and a Polynomial Ridge Regression Fusion Model

In modern power systems, accurate estimation of the inertia of individual generation units is essential for maintaining system stability and identifying weak-inertia areas that require additional inertia support. Conventional neural network ensemble methods based on equal-weight averaging neglect the performance differences among different neural network models. As a result, they cannot fully exploit the complementary strengths of heterogeneous neural network architectures. To address this limitation, this paper proposes an inertia estimation method for power generation units based on dual neural networks and a polynomial ridge regression fusion model. A dual-network framework consisting of a long short-term memory–multilayer perceptron (LSTM–MLP) model and a temporal convolutional network–long short-term memory (TCN–LSTM) model is first constructed. Polynomial ridge regression is then introduced to establish a nonlinear mapping between the inertia estimates produced by the two neural networks and the corresponding true inertia values. This mapping further exploits the complementary capabilities of the two models in time-series feature extraction. The proposed method is validated through comparative, ablation, and noise robustness experiments on the IEEE 14-bus system, and further evaluated on an IEEE 39-bus system containing PV-VSG units. The results demonstrate improved estimation accuracy over the individual models and the equal-weight ensemble, good robustness under Gaussian measurement noise, and relatively high estimation accuracy under different system and generation configurations.