Nearest-neighbor methodology for prediction of intra-hour global horizontal and direct normal irradiances

This work proposes a novel forecast methodology for intra-hour solar irradiance based on optimized pattern recognition from local telemetry and sky imaging. The model, based on the k-nearest-neighbors (kNN) algorithm, predicts the global (GHI) and direct (DNI) components of irradiance for horizons ranging from 5 min up to 30 min, and the corresponding uncertainty prediction intervals. An optimization algorithm determines the best set of patterns and other free parameters in the model, such as the number of nearest neighbors. Results show that the model achieves significant forecast improvements (between 10% and 25%) over a reference persistence forecast. The results show that large ramps in the irradiance time series are not very well capture by the point forecasts, mostly because those events are underrepresented in the historical dataset. The inclusion of sky images in the pattern recognition results in a small improvement (below 5%) relative to the kNN without images, but it helps in the definition of the uncertainty intervals (specially in the case of DNI). The prediction intervals determined with this method show good performance, with high probability coverage (≈90% for GHI and ≈85% for DNI) and narrow average normalized width (≈8% for GHI and ≈17% for DNI).