Improved Blending of PV Power Forecasts in Case of Measurements with Limited Reliability

For forecasting the power output of a photovoltaic (PV) power plant, solar irradiance forecasts are an essential input. Forecasts generated from different sources and models such as satellite data, numerical weather models, or irradiance measurements, perform differently depending on the forecast horizon. An optimized forecast for each horizon can be derived by combining several different source forecasts via a machine learning (ML) model to create a resulting ‘blended’ forecast. The training of an ML blending model requires power generation data of the considered PV power plant as a target variable. Typically, power measurements from the plant are used for this purpose. However, poor quality or limited availability of power measurements will lead to low quality blended forecasts as a result. In this work we consider intra-day forecasts, obtained from blending numerical weather predictions with satellite-derived forecasts, for a large PV plant with a capacity of 1 GW. The plant is frequently subject to curtailment by grid operators, which limits the reliability of its power measurements. This poses a significant challenge for the optimization of the blending model, resulting in a notable underestimation of the power forecasted by the model in high-power situations. We present an approach to improve forecast blending by using satellite derived power as a replacement for measurements in model training. We evaluate the effect of this replacement for two different ML blending models: a Huber linear regressor and a neural network. The performance of the different models is characterized by employing commonly used metrics such as RMSE and BIAS, as well as a distribution-oriented evaluation framework.