Solar and wind quantity 24 h-series prediction using PDE-modular models gradually developed according to spatial pattern similarity

Abstract

The design and implementation of efficient photovoltaic (PV) plants and wind farms require a precise analysis and definition of specifics in the region of interest. Reliable Artificial Intelligence (AI) models can recognize long-term spatial and temporal variability, including anomalies in solar and wind patterns, which are necessary to estimate the generation capacity and configuration parameters of PV panels and wind turbines. The proposed 24 h planning of renewable energy (RE) production involves an initial reassessment of the optimal day data records based on the spatial pattern similarity in the latest hours and their follow-up statistical AI learning. Conventional measurements comprise a larger territory to allow the development of robust models representing unsettled meteorological situations and their significant changes from a comprehensive aspect, which becomes essential in middle-term time horizons. Differential learning is a new unconventionally designed neurocomputing strategy that combines differentiated modules composed of selected binomial network nodes as the output sum. This approach, based on solutions of partial differential equations (PDEs) defined in selected nodes, enables us to comprise high uncertainty in nonlinear chaotic patterns, contingent upon RE local potential, without an undesirable reduction in data dimensionality. The form of back-produced modular compounds in PDE models is directly related to the complexity of large-scale data patterns used in training to avoid problem simplification. The preidentified day-sample series are reassessed secondary to the training applicability, one by one, to better characterize pattern progress. Applicable phase or frequency parameters (e.g., azimuth, temperature, radiation, etc.) are related to the amplitudes at each time to determine and solve particular node PDEs in a complex form of the periodic sine/cosine components. The proposed improvements contribute to better performance of the AI modular concept of PDE models, a cable to represent the dynamics of complex systems. The results are compared with the recent deep learning strategy. Both methods show a high approximation ability in radiation ramping events, often in PV power supply; moreover, differential learning provides more stable wind gust predictions without undesirable alterations in day errors, namely in over-break frontal fluctuations. Their day average percentage approximation of similarity correlation on real data is 87.8 and 88.1% in global radiation day-cycles and 46.7 and 36.3% in wind speed 24 h. series. A parametric C++ executable program with complete spatial metadata records for one month is available for free to enable another comparative evaluation of the conducted experiments.