Spatio-temporal PV forecasting sensitivity to modules' tilt and orientation

Using deployed PV generation as inputs for spatio-temporal forecasting approaches has the potential for fast and scalable very short-term PV forecasting in the urban environment but one has to consider the effect of their tilt and orientation on the forecasting accuracy. To address this issue, tilted irradiance data sets were simulated using state of the art solutions on a horizontal irradiance data set from a pyranometer network deployed in Oahu, Hawaii, and used as inputs to train a 10-s ahead linear ARX model. Results showed that the mismatch in tilt/orientation degrades the forecast skill, justified by the difference in the diffuse fraction of each surface and, thus, how each reacts to changes in cloud cover. From 4000 simulated sets, it was shown that using information from more sites led to better forecasts and made the model performance less sensitive to the PV modules' tilt and orientation. Forecast skill showed to be quite sensitive to the tilt and orientation ensemble when the inputs consisted of only rooftop or facade systems (between 18.1-29.6% and 8.2-19.4%, respectively). Forecasting a rooftop system with vertically tilted neighbors lead to considerably lower skill values (9.8-16.2%) and benefitted when all shared the same orientation. On the other hand, forecasting a vertically tilted system with rooftop neighbors had a lower impact (9.2-14.7%) and benefitted from diversely oriented neighbors.