The energy produced by photovoltaic farms has a variable nature depending on astronomical and meteorological factors. The former are the solar elevation and the solar azimuth, which are easily predictable without any uncertainty. The amount of liquid water met by the solar radiation within the troposphere is the main meteorological factor influencing the solar power production, as a fraction of short wave solar radiation is reflected by the water particles and can't reach the earth surface. The total cloud cover is a meteorological variable often used to indicate the presence of liquid water in the troposphere and has a limited predictability, which is also reflected on the global horizontal irradiance and, as a consequence, on solar photovoltaic power prediction. This lack of predictability makes the solar energy integration into the grid challenging. A cost-effective utilization of solar energy over a grid strongly depends on the accuracy and reliability of the power forecasts available to the Transmission System Operators (TSOs). Furthermore, several countries have in place legislation requiring solar power producers to pay penalties proportional to the errors of day-ahead energy forecasts, which makes the accuracy of such predictions a determining factor for producers to reduce their economic losses. Probabilistic predictions can provide accurate deterministic forecasts along with a quantification of their uncertainty, as well as a reliable estimate of the probability to overcome a certain production threshold. In this paper we propose the application of an analog ensemble (AnEn) method to generate probabilistic solar power forecasts (SPF). The AnEn is based on an historical set of deterministic numerical weather prediction (NWP) model forecasts and observations of the solar power. For each forecast lead time and location, the ensemble prediction of solar power is constituted by a set of past production data. These measurements are those concurrent to past deterministic NWP forecasts for the same lead time and location, chosen based on their similarity to the current forecast and, in the current application, are represented by the one-hour average produced solar power. The AnEn performance for SPF is compared to a quantile regression (QR) technique and a persistence ensemble (PeEn) over three solar farms in Italy spanning different climatic conditions. The QR is a state-of-the-science method for probabilistic predictions that, similarly to AnEn, is based on a historical data set. The PeEn is a persistence model for probabilistic predictions, where the most recent 20 power measurements available at the same lead-time are used to form an ensemble. The performance assessment has been carried out evaluating important attributes of a probabilistic system such as statistical consistency, reliability, resolution and skill. The AnEn performs as well as QR for common events, by providing predictions with similar reliability, resolution and sharpness, while it exhibits more skill for rare events and during hours with a...

a b s t r a c tThe efficient integration of wind in the energy market is limited by its natural variability and predictability. This limitation can be tackled by using the probabilistic predictions that provide accurate deterministic forecasts along with a quantification of their uncertainty. We propose as a novelty the application of an analog ensemble (AnEn) method to generate probabilistic wind power forecasts (WPF). The AnEn prediction of a given variable is constituted by a set of measurements of the past, concurrent to the past forecasts most similar to the current one. The AnEn performance for WPF is compared with three state-of-the-science methods for probabilistic predictions over a wind farm and a 505-day long period: a wind power prediction based on the ensemble wind forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) Ensemble Prediction System (ECMWF-EPS), the Limitedarea Ensemble Prediction System (LEPS) developed within the COnsortium for Small-scale MOdelling (COSMO-LEPS) and a quantile regression (QR) technique. The AnEn performs as well as ECMWF-EPS, COSMO-LEPS and QR for common events while it exhibits more skill for rare events. A comparison with the performances obtained with a deterministic forecasting method based on a Neural Network is also carried out showing the benefits of using AnEn.

No abstract

This work explores a Principal Component Analysis (PCA) in combination with two post-processing techniques for the prediction of wind power produced over Sicily, and of solar irradiance produced over the Oklahoma Mesonet. For wind power, the study is conducted over a 2-year long period, with hourly data of the aggregated wind power output of the island. The 0-72 hour wind predictions are generated with the limited-area Regional Atmospheric Model System (RAMS), with boundary conditions provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) deterministic forecast. For solar irradiance, we consider daily data of the aggregated solar radiation energy output (based on the Kaggle competition dataset) over an 8-year long period. Numerical Weather Prediction data for the contest come from the National Oceanic & Atmospheric Administration -Earth System Research Laboratory (NOAA/ESRL) Global Ensemble Forecast System (GEFS) Reforecast Version 2. The PCA is applied to reduce the datasets dimension. A Neural Network (NN) and an Analog Ensemble (AnEn) post-processing are then applied on the PCA output to obtain the final forecasts. The study shows that combining PCA with these post-processing techniques leads to better results when compared to the implementation without the PCA reduction.

Solar energy production is steadily growing in several countries. Depending on meteorological variables such as solar irradiance, cloud cover and temperature, solar power production predictability is limited by the chaotic nature of these parameters. This work presents an application of the European Centre for Medium-Range Weather Forecasts (ECMWF) Ensemble Prediction System (EPS) to produce short-term probabilistic solar power forecasts (SPF). The EPS is based on running a meteorological model multiple times, starting from slightly perturbed initial conditions.The distribution of these different runs allows estimating the prediction uncertainty. We use about two years of power data from three solar farms located in different parts of Italy, characterized by different climatic conditions. We retrieve ensemble members of meteorological variables to build a power probability density function (PDF) for 0-72 h forecast horizons. A neural network (NN) is applied to reduce model bias and to generate a PDF of power starting from the ensemble of meteorological variables. Then two statistical methods are applied to calibrate further the ensemble of power predictions, which results in the final probabilistic power forecast. The first technique has been already tested for wind power forecasting (WPF), and it is based on estimating the variance deficit (VD) or "lack of spread" of the ensemble distribution. The second is a well-established method to calibrate probabilistic predictions based on Ensemble Model Output Statistics (EMOS). Also, a persistence ensemble (PE) is used as a baseline technique. PE builds the ensemble members from the most recent power measurements available at the same lead time. It is shown that EPS coupled with VD and EMOS is a valid method to produce skillful probabilistic SPF in different climatic conditions. VD and EMOS show a similar level of statistical consistency and forecast skill for each test case. Overall, the performances of VD and EMOS appear significantly better than those of PE.