Data-driven upscaling methods for regional photovoltaic power estimation and forecast using satellite and numerical weather prediction data

•Complete assessment of the regional PV forecast accuracy at different horizons.•Construction of two upscaling methods based on data-driven models.•Prediction intervals computed using a model to predict the forecast error.•Analysis of the smoothing effect of forecast error due to regional upscaling. The growing photovoltaic generation results in a stochastic variability of the electric demand that could compromise the stability of the grid, increase the amount of energy reserve and the energy imbalance cost. On regional scale, the estimation of the solar power generation from the real time environmental conditions and the solar power forecast is essential for Distribution System Operators, Transmission System Operator, energy traders, and Aggregators. In this context, a new upscaling method was developed and used for estimation and forecast of the photovoltaic distributed generation in a small area of Italy with high photovoltaic penetration. It was based on spatial clustering of the PV fleet and neural networks models that input satellite or numerical weather prediction data (centered on cluster centroids) to estimate or predict the regional solar generation. Two different approaches were investigated. The simplest and more accurate approach requires a low computational effort and very few input information should be provided by users. The power estimation model provided a RMSE of 3% of installed capacity. Intra-day forecast (from 1 to 4 h) obtained a RMSE of 5%–7% and a skill score with respect to the smart persistence from −8% to 33.6%. The one and two days ahead forecast achieved a RMSE of 7% and 7.5% and a skill score of 39.2% and 45.7%. The smoothing effect on cluster scale was also studied. It reduces the RMSE of power estimation of 33% and the RMSE of day-ahead forecast of 12% with respect to the mean single cluster value. Furthermore, a method to estimate the forecast error was also developed. It was based on an ensemble neural network model coupled with a probabilistic correction. It can provide a highly reliable computation of the prediction intervals.