Wind Speed NWP Local Revisions Using a Polynomial Decomposition of the General Partial Differential Equation

Precise daily weather forecasts are necessary for the utilization of renewable energy sources and their penetration into grid systems. Standard meteorological statistical post-processing methods relate local observations with numerical predictions to eliminate systematic forecast errors. Neural networks, trained with the last historical series, can model the current weather frame to refine a target forecast for specific local conditions and reduce random prediction errors. Their daily correction models can process numerical prediction model outcomes of the same data types (instead of the unknown data) to recalculate 24-hour wind speed forecast series. Global numerical weather models succeed generally in forecasting upper air patterns but are too crude to account for local variations in surface weather. Long-term complex forecast systems, which simulate the dynamics of the complete atmosphere in several layers, cannot exactly detail local conditions near the ground, determined by the terrain relief, structure, landscape character, pattern and other factors. Extended polynomial networks can decompose and solve general linear partial differential equations, being able to model properly unknown dynamic systems. In all the network nodes are produced series of relative polynomial derivative terms, which convergent sum combinations can directly define and substitute for the general differential equation to model an uncertain system target function. The proposed local forecast correction procedure using adaptive derivative regression model can improve numerical daily wind speed forecasts in the majority of days.