Power quality approximation for household equipment load combinations using a stepwise growth in input parameters of AI models

Detached off-grids, subject to the generated renewable energy (RE), need to balance and compensate the unstable power supply dependent on local source potential. Power quality (PQ) is a set of EU standards that state acceptable deviations in the parameters of electrical power systems to guarantee their operability without dropout. Optimization of the estimated PQ parameters in a day-horizon is essential in the operational planning of autonomous smart grids, which accommodate the norms for the specific equipment and user demands to avoid malfunctions. PQ data for all system states are not available for dozens of connected / switched on household appliances, defined by their binary load series only, as the number of combinations grows exponentially. The load characteristics and eventual RE contingent supply can result in system instability and unacceptable PQ events. Models, evolved by Artificial Intelligence (AI) methods using self-optimization algorithms, can estimate unknown cases and states in autonomous systems contingent on self-supply of RE power related to chaotic and intermitted local weather sources. A new multilevel extension procedure designed to incrementally improve the applicability and adaptability to training data. The initial AI model starts with binary load series only, which are insufficient to represent complex data patterns. The input vector is progressively extended with correlated PQ parameters at the next estimation level to better represent the active demand of the power consumer. Historical data sets comprise training samples for all PQ parameters, but only the load sequences of the switch-on appliances are available in the next estimation states. The most valuable PQ parameters are selected and estimated in the previous algorithm stages to be used as supplementary series in the next more precise computing. More complex models, using the previous PQ-data approximates, are formed at the secondary processing levels to estimate the target PQ-output in better quality. The new added input parameters allow us to evolve a more convenient model form. The proposed multilevel refinement algorithm can be generally applied in modelling of unknown sequence states of dynamical systems, initially described by binary series or other insufficient limited-data variables, which are inadequate in a problem representation. Most AI computing techniques can adapt this strategy to improve their adaptive learning and model performance.