Neural Network Voltage Estimation in Power Grids

The patent applies preliminary action by pre-training artificial neural networks and Gaussian process models using historical load flow calculation data. The models learn the complex nonlinear relationships between input features (power values, network topology) and output status values (node voltages, branch currents) during an offline training phase. During online operation, the trained models directly predict status values without performing computationally intensive load flow calculations, thus resolving the contradiction between accuracy and computational effort.

The patent creates simplified computational copies of the load flow calculation system using machine learning models. Instead of repeatedly executing the full load flow calculation algorithm, the system uses trained neural networks and Gaussian processes that replicate the input-output behavior of load flow calculations. These models capture the essential physics and network characteristics while requiring minimal computational resources during deployment.

The patent introduces machine learning models as intermediary systems that bridge the gap between limited available measurements and the full set of required status values. The models use readily available input data (power injections, network topology, weather conditions for renewable generators) and compute the complete state of the network including node voltages and branch currents at all locations, eliminating the need for physical sensors at every node.

The patent replaces the mechanical/physical measurement system (sensors, meters, and measurement devices at each node) with an information-processing system based on artificial intelligence. Instead of physically measuring every status value with hardware, the system uses trained algorithms to infer all necessary measurements from a small set of available inputs, thus substituting complex measurement infrastructure with computationally efficient modeling.