Physically Jacobian-Informed Encoder-Decoder ANNs for Nonlinear Power Flow Regression

Power flow (PF) is the basis of steady-state analysis and control of power systems. The conventional model-driven PF formulated by a set of implicated nonlinear equations is solved iteratively by using Newton-Raphson method. But, the speed and convergence of PF computation are influenced by proper initial values and the efficiency of iterative process. The data-driven PF regression method can overcome the above issues by learning an explicit mapping function from PF data set. However, it simply achieves the nonlinear map from PF input to output, overlooking the physical rules in the PF calculation, which may result in poor accuracy and generalization. This paper presents a physically jacobian-informed encoder-decoder neural networks (NNs) for PF nonlinear regression. Based on the forward and inverse PF model, a multi-task learning method with encoder-decoder NNs is constructed for data-driven PF regression. To introduce PF physical characteristics, the corresponding jacobian information is embedded into the regression model to improve the accuracy and generalization. The performance of high accuracy and generalization of the proposed method is validated by IEEE test systems.