Power quality estimations for unknown binary combinations of electrical appliances based on the step-by-step increasing model complexity

Abstract

Smart detached houses, contingent on renewable energy (RE), are subjected to an unstable power supply of the intermitted nature. The power quality (PQ) norms define allowable variances in the characteristics of electrical systems to ensure their functioning without malfunction. The estimation and optimization of PQ parameters on day bases are inevitable in the regulation of systems to comply with the specified standards and allow the fault-free operation of electrical equipment. Measurements of all PQ states are impossible for dozens of eventual grid-attached power consumers defined by their binary load patterns. Specific demands and uncertain RE can lead to system instability and unacceptable PQ. Self-optimizing models based on artificial intelligence (AI) can estimate the next PQ states in real off-grids where power is induced only by chaotic RE sources. A new proposed multistage prediction scheme allows incremental improvements in the accuracy of AI models beginning their development with binary coded data only. The number of selected PQ inputs gradually increased in the next estimate for the initial equipment in demand. Historical records include complete training PQ data for all parameters, but only "1/0" switch-on load sequences are available at prediction times. The most valuable PQ outputs are modeled in the previous stages to process their supplementary series in the next prediction. More capable models, applied to previously approximated PQ data, are able to better compute the PQ output in the secondary steps. Complementary PQ inputs are supplied with the new processing data, which were unknown in the previous stage. The growing number of input features enables a more complex representation of the target quantity in each iteration. Advanced input selection and data reevaluation can additionally improve model discriminability for unseen active load patterns. It can be applied in modeling unknown states of various dynamical systems, initially defined only by series of binary or inadequate input data, to improve the results.