The installed photovoltaic (PV) power capacity is expected to grow by 300 to 500% in Switzerland until 2035 (Energy Act 2018) and to expand similarly strongly in other European countries (IEA 2019). The intraday forecasts of generated PV power that are in operational use today are mostly based on weather forecast models and tend to produce large forecast errors (Fig. 5), in particular in changeable and cloudy weather conditions. More accurate intraday forecasts of the generated solar power and a better understanding and quantification of the forecast uncertainties are required in order to enable a more cost-efficient solar plant operation and reduce the cost incurred by grid operators and utilities for stand-by and storage capacities as well as suboptimal bidding in the intraday and balancing markets.Most of today’s state-of-the-art probabilistic forecast models for solar irradiance and PV power are purely based on weather forecast products or historical in-situ point measurements rather than on near-real-time observations of the plant surrounding area (Hong et al. 2016, David et al. 2018), despite the fact that forecasting methods using sensor network and satellite data are believed to have great potential for higher accuracy intraday forecasts with well characterized uncertainties (Yang et al. 2018, Antonanzas et al. 2016). The existing satellite and sensor network based models only provide point forecasts, i.e. a single estimate of the forecast variable, and hence lack an explanation and characterization of the forecast’s uncertainty (Wang et al. 2019, David et al. 2018). The goal of this project is to provide a framework for more accurate and reliable intraday forecasts over the Swiss plateau where most current and additional future Swiss PV capacity will be located. Specifically, the project aims to develop more accurate, probabilistic forecasting methods of the global horizontal irradiance (GHI, W/m2) and the resulting generated PV power with quantified forecast uncertainties for forecast lead times of minutes to several hours. This time frame is most relevant for intraday operational risk management and decision support, and weather-model based PV forecasts tend to be outperformed by statistical and machine learning based forecasts including cloud advection models at this time horizon (Antonanzas et al. 2016). We will also investigate correlations between forecast errors across PV plants and scaling effects across the plants’ power generation.To achieve these goals, we will combine satellite-derived estimates of GHI and data from one of the densest radiation sensor networks worldwide and use probabilistic cloud advection schemes (Pulkkinen et al. 2019) to forecast GHI probability density distributions. We will use statistical and machine learning approaches including deep neural networks to derive probabilistic PV production forecast models and employ them in a case study with four major Swiss multi-MW photovoltaic power plants (Fig. 4).Our expected results include the provisioning and publication of the forecasting framework and its application and characterization in a case study with the four largest Swiss PV plants, where we will also analyze correlation and scaling effects between the plants. This project will be the first to perform probabilistic surface solar irradiance forecasts and probabilistic PV power forecasts based on satellite imagery, to the best of our knowledge. We expect to provide to the field a more accurate and reliable PV forecasting methodology and a deeper understanding of aggregated PV output modelling for a suite of plants. Extending beyond academic research, this project will facilitate an improved risk management and reduced operational cost through more accurate and reliable forecasts. The forecasting framework and results will be presented and provided to PV plant and grid operators in the course of the project.
Last updated:27.01.2023
