Implementing k-Nearest Neighborhood as a Forecast Method for Intra Hour Resolution with No Exogenous Outputs
Abstract
Forecasting solar irradiance is a big challenge for solar power plant and grid operators, and managing power inputs to the grid according to reality thus becomes a matter of not only scientific and technological interest, but of economic and strategic importance as well. Forecasting techniques are an important asset, with several methods with different singularities. In this paper, we discuss the k-Nearest Neighborhood method (kNN) in a real case scenario for a forecast horizon of one minute, using irradiance data from Florianópolis-SC in Brazil (27oS, 48oW) for the year 2018. The kNN presents itself as a simple and robust method, having accuracy results of: R² = 95% and NRMSE = 22.13%. Moreover, real case forecasting was simulated and the punctual error for each point was calculated and analyzed, thus providing indications about which factors were related to the accuracy loss, concluding, afterwards that the biggest error gather was irradiance ramps, caused by cloud covers and cloud edges.