In this paper… we gonna chane the way of thinking of the Energy managers all over the world. First of all we gonna prove that we are the best then :-)

You can get started by double clicking this text block and begin editing. You can also click the Text button below to add new block elements. Or you can drag and drop an image right onto this text. Happy writing!

Introduction and review

The integration of renewable energy into power distribution grids and their increasing diffusion in Europe, from photovoltaic (PV), Concentrated Solar Power (CSP) and Concentrating Photovoltaic (CPV) plants in particular, highlights the problems linked to their spread in both the technological and economic fields. The variable and non-programmable nature of the solar resource, the still high cost associated with energy storage, and the difficult balance between the flexibility of the network and its reliability impose a back-up for the renewable power systems, by auxiliary generators and / or storage systems that can intervene during periods of high variability. This inevitably increases the overall cost of production from solar sources.

Regardless of the current regulations of the electricity market in different regions, which favor, in general, the definition of the production profile with one day of advance (Kraas 2013), (Luoma 2014), (Nonnenmacher 2016), an accurate forecast is critical to the proper dimensioning and optimal management of any storage facility for the distribution grid (Hanna 2014), (Delfanti 2015).

Several approaches have been developed for the forecasting of power production from renewable sources in the literature, a recent and fairly comprehensive review is contained in (Inman 2013) and (Kleissl 2013).

Different approaches are used for solar power forecast depending on the time horizon for which the forecast is required. The short-term forecasts (\(T_{F}<1h\)) are based mainly on sky imaging techniques or on time series models, while the satellite data can give an improvement in the prediction when the forecast horizon increases (\(1h<T_{F}<6h\)) (Perez 2010) (Zagouras 2015). Accurate forecasts for longer time horizons (\(T_{F}>6h\)) strongly depend on the processing of Numerical Weather Prediction (NWP) models results.

In this paper we focus on a forecast technique for the production of a photovoltaic plant the day before, with the ultimate target of the optimal management of an energy storage system, and based solely on information that may be obtained from the numerical prediction systems of the weather conditions, and from the past data on weather condition and power production of the system.

The proposed method bases the power forecast mainly on the NWP systems’s output. The Global Horizonthal Radiation (GHI) is of particular relevance among the available data, being it strongly related to the instantaneous power obtained by a photovoltaic system. The available literature on the prediction of the GHI with a day in advance is quite wide and is not only related to the specific issue of energy production. In contrast, the literature available on the application of the provision of GHI and other output variables obtained from a meteorological model to forecast of energy production using photovoltaic systems does not enjoy the same abundance, presumably due to the poor availability of quality data from PV plants production (Larson 2016).

Recent works on the forecast of the power output of PV systems using NWP output include (Bacher 2009) in which 21 PV station in Denmark were analyzed using the the High Resolution Limited Area Model of the Danish Meteorological Institute.

A single PV plant in in Spain was studied using again a high resolution model in (Masa-Bote 2014).

In (Pelland 2011) forecasts from the Canadian Meteorological Centre, validated against ground measurements from the United States’s SURFRAD network, were used to evaluate the performance of three small PV systems, the data from the weather forecast were treated through spatial averaging and bias removal through Kalman filtering.

The input of the Japan Meteorological Agency forecasts was used in (Junior 2014) and (Junior 2014a) to evaluate data from PV plants of four and two regions in Japan using Support Vector Machine regression techniques (SVR) to obtain the power forecast. The method is applied to a single photovoltaic power plant in (Fonseca 2011).

A method based on global and meso-scale weather prediction models and artificial neural network was adopted and compared with other methods in (Fernandez-Jimenez 2012)where it is applied to a PV plant located in Spain.

The power output of five tracking plants in Spain using a non-parametric model built upon the Quantile Regression Forests as machine learning tool has been recently analyzed in (Almeida 2015) using as inputs several forecasts of meteorological variables from a Numerical Weather Forecast model, and actual AC power measurements of PV plants.

Recently the Ensemble Prediction System (EPS) of ECMWF and artificial neural networks were use in (Sperati 2016) to produce a probabilistic forecast of the power production of three solar farms located in Italy.

In this paper we present a new forecasting procedure for the power produced by a medium sized PV plant, located in the German island of Borkum. The work is in the framework of the European Union’s Horizon 2020 project NETfficient, aimed at improving energy and economic efficiency for today’s smart communities through integrated multi storage technologies, the forecast is used for the optimal management of the storage facilities.

The procedure is based on a regression techniques known as Multivariate Adaptive Regression Splines and to time interpolation through a clear sky model. The regression is based on the output of Numerical Weather Predictions models only, without the need of measurements of the the power production of the plant in the recent history. The power production history is solely used for the training phase of the model.

The paper is structured as follows, at first the NWP models are described, then a brief overview of the forecasting methods, of the solar irradiation on a surface and of the clear sky models is presented. The regression and interpolation technique is described, and the results of its application and an evaluation of the performance of the method is presented. Finally in the conclusions we will also discuss the future improvements that may be implemented.

Numerical Weather Prediction

The main purpose of this work was to verify the maximum degree of skill achievable in the prediction of PV energy production using as predictors the meteorological fields produced by a numerical weather forecast model and to quantify the benefit that can be obtained in comparison with the simplest predictive methods based on persistence and on the knowledge of the climate.

Our first objective was that of separating the uncertainty introduced by the predictors itself from that related to the length of the forecast time. To achieve this result we used, two different data sets for the calibration and verification of the procedure. In the first one (named from now on GFS1) we used forecast data with the the shortest possible forecast time, instead for the second (GFS2) we used data with forecast time greater than 24 hr.

Moreover the “ideal” weather forecast data set for our scopes, had also to be:

  • available for 2014, the year for which we disposed of the data production of the photovoltaic plant under study

  • be updated constantly, for the future, so as to support a real-time operating procedure —item possibly publicly available and updated multi-day

These features are fully satisfied by the the global model forecasts GFS (Global Forecasting System) operated by the US National Meteorological Service ( As for many of the activities funded by the US, the GFS model data are made publicly available and form the basis for many commercial and research activities, including private. The GFS model is a spectral model operated 4 times a day, starting from the time of analysis 00, 06, 12 and 18 UMT, and provides global forecasts, with an average of about 13km horizontal resolution and 64 vertical levels. Forecasts are available at the maximum spatial resolution, for forecasting time up to +120hr, every 3 hours (actually, since May 2016 forecast times up to +120hr are available hourly). Data of the operational model for the last year, although at reduced resolution compared to the actual one, are available for download from servers operated by NOAA (National Oceanic and Atmospheric Administration).

More precisely for data set GFS1, used to characterize the forecast error in energy production due to the predictors, we downloaded data for year 2014 at 0.5° resolution for +03hr and +06hr forecast time. This allowed to reconstruct the variability of meteorological fields during the day with a temporal resolution of 3 hours. In detail, the fields for each of 365 days of the year, we used are:

  • for 03 and 06 UTM, those of the forecast + 03hr and +06hr of the analysis of 00 UTM

  • for 09 and 12 UTM, those of the forecast + 03hr and +06hr of the analysis of 06 UTM

  • for 15 and 18 UTM, those of the forecast + 03hr and +06hr of the analysis of 12 UTM

  • for 21 and 00 UTM, those of the forecast + 03hr and +06hr of the analysis of 18 UTM

We decided to use forecast fields only, in view of a future operational implementation of the procedure and because, in general, not all fields are available for the analysis time.

To estimate the skill of the procedure when using predictors at longer forecast time, we build up the GFS2 data set, covering the second part of the year from 12 July 2014 till the end of the year, with fields for times of forecast +27hr and +30hr. The GFS2 data set has been organized as the GFS1, in which however the +27hr (+30hr) forecast has been used for each day in place of the +3hr (+6hr) and the analysis time is that of the day before (24hr ahead).