Power quality statistical predictions based on differential, deep and probabilistic learning using off-grid and meteo data in 24-hour horizon

Abstract

Off-grid systems using renewable energy (RE) are dependent on the stochastic power supply, which results in a high level of uncertainty, noise, and variability in the operational conditions. Power quality (PQ) norms specify allowable variations in relevant parameters of grid systems, necessary to maintain certain limits to guarantee their fault-tolerant states. 24-hour PQ prediction is vital in planning power consumption and utilization in smart off-grid houses based on RE. PQ data for all possible combinations of grid-attached household appliances and variable outside conditions cannot be measured completely or described exactly by physical equations. PQ predictions on daily bases using artificial intelligence (AI) models are requisite because atmospheric fluctuations and anomalies in local weather primarily influence the induced power and potential operation mode in real off-grids. Load specifics and possible collisions of the switched-on power consumers together with alterations in RE production can lead to additional disturbances in PQ and the consequent instability of the autonomous system. An automatic selection algorithm can give preliminary suggestion or correct several eventual energy consumption variants based on load time-shifting to allow efficient utilization of the predicted photovoltaic power (PVP). PQ models can after examine feasible daily load scenarios, offered by the system or re-defined by the users, scheduling selected equipment in the optimal switch intervals to ensure the system stability primarily. Users will have several practical options to choose from or combine the best ones to meet their demand in the optimal PQ and load planning strategy. This 2-step self-determination and verification procedure can help in effective operation of smart-grids in consolidation of their future state, balancing its power demands with local RE potential. The AI statistical models were evolved with the pre-assessed lengths of daily data periods. After that, they are applied to the last unseen data series, used in testing, to predict one-step sequences of the target PQ parameters in the trained all-day horizon. Parametric C++ application software with applied PQ and weather data is available for free to allow reproducibility of the results.