High-Dimension Bayesian Parameter Estimation for WECC Composite Load Model Using Realistic Event Measurements

Calibration of the composite load model with distributed generation (CMPLDWG) can be challenging due to its nonlinear and high-dimensional characteristics. To address this issue, this paper proposes a Bayesian load calibration approach for the CMPLDWG model considering realistic measurement errors. The critical parameters that have large sensitivities to the load model are selected through sparse polynomial chaos expansion (SPCE)-based global sensitivity analysis. Subsequently, a deep neural network (DNN)-kernelized vector-valued Gaussian process (VVGP) is developed to learn the relationship between these critical parameters and the load response. Consequently, the DNN-VVGP is utilized to substitute the CMPLDWG model simulator, enabling efficient gradient-based posterior distribution sampling of load parameters via Hamiltonian Monte Carlo (HMC). Numerical results on realistic event measurements verify that the proposed method is more robust to measurement error and achieves higher estimation accuracy than existing state-of-the-art methods.