Chasing the Clouds: Irradiance Variability and Forecasting for Photovoltaics

Abstract

The stochastic part of the variability of irradiance is captured through the variability of the clear-sky index. If variability is correlated for close locations, the aggregated PV-output will be relatively high compared to aggregate PV-output of locations that are far apart (and more likely to be de-correlated). Aggregate fluctuations of PV power output will eventually be added up at a local LV substation in the case of a residential neighborhood, or MV network when large PV-plants are concerned. Successful integration of high-levels of PV capacity depends on the correct evaluation of the expected correlate peaking highs and lows. Short-term forecasting deals with the question “what if the aggregation does not reduce variability enough?'': knowing what peak or low will come in the electricity network can help to prepare for this by e.g. planning the smart charging of electric vehicles or short-term spinning reserves, respectively in a smart grid. In general, daily correlation of PV-system pairs is retained longer at increasing temporal resolution from 0.5 km for 5 s data to 3--6 km for 10 min. averaged data, depending strongly on the pattern of clouds (variability). Measured over sub-day periods and specified to azimuthal directions, lagged temporal correlation downwind is retained over much longer distances than in the direction perpendicular to the cloud motion. A trade-off exists between building and maintaining a high-resolution pyranometer sensor-network at this (province) scale and the quality of the derived GHI data of using existing PV-systems as we have done. The principle of using a network of sensors for short-term forecasting is an extension to existing methods, and the application on real-world PV-systems was worthwhile to investigate. The P2P method showed to be useful especially on days with variable irradiance with clear, quasi-statically moving clouds for which the shortest distances retain aggregate variability the most strongly and thus can benefit from short-term detailed forecasting if supply and demand are to be coupled in a future smart grid. The relative Root Mean Square Error of the P2P forecasting method in the studied validation period is in the order of 50%, but performs 6-9% better than naive clear-sky index persistence on the shortest forecast horizon of 5 to 8 min. at 1 min. data resolution, depending on the method of cloud motion detection. The presented technique to infer the direction of motion of the clouds within the sensor-network itself, that would otherwise rely on cloud camera, satellite- or meteorological data, is an essential addition to using the PV-sensor network. The PV-sensor network is, in principle a great tool to have high-resolution measurement data over a large geographic area at relatively low effort of connecting and measuring these existing PV-systems instead of developing an irradiance sensor network from scratch. Using the Inverse PV model (with RMSE of 15.1%), the measured power-output was translated into GHI and hence clear-sky index such that measurement results become independent of PV-system meta data (tilt, orientation, rated power).