Power quality multi-step predictions with the gradually increasing selected input parameters using machine-learning and regression

Abstract

Autonomous off-grid systems dependent upon Renewable Energy (RE) sources are characterized by stochastic supplies of the fluctuating low short-circuit power. Power Quality (PQ) standards define load characteristics in electric power systems and their ability to function properly without failures. Monitoring, prediction and optimization of PQ parameters are necessary to maintain their alterations steady within the prescribed range, which allow fault-tolerant operation of various electrical devices. It is not possible to measure complete PQ data for all possible combinations of dozens of grid-connected appliances, whose load specifics and collisions primarily determine the course of PQ parameters and their eventual disturbances. Self-adapting PQ prediction models based on Artificial Intelligence (AI) are required as induced power is influenced particularly by changeable weather conditions in real off-grid operation mode of systems using RE. A novel multi-step PQ prediction algorithm is proposed, which develops AI models with the gradually increasing number of selected input PQ-parameters. In each next step a more complex model is formed, using an additional co-related PQ-input to calculate its target PQ-output with a better accuracy. PQ-models with the progressively growing PQ-inputs, using their data predicted in the previous step, can better approximate and estimate the target quantity. The presented results show this training and feature selection procedure can step by step improve accuracy of PQ-models for unknown combinations of off-grid connected household appliances.