A system of direct radiation forecasting based on numerical weather predictions, satellite image and machine learning. 31st Annual International Symposium on Forecasting Lourdes Ramírez Santigosa Martín Gastón Romeo Solar Thermal Energy Department CENER 1 A system of direct radiation forecasting based on numerical weather predictions, satellite image and machine learning. CSPP Overview Meteorological models Mathematical models 2. Forecasting system Forecasting scheme Cloud forecasting Type of data and DNI prediction 3. Results 2 2
Plant energy simulation model Design and forecasting tool Solar thermal power plants have a series of advantages for their integration in the electricity it system and power market The solar resource can be forecasted better than other renewable energies. The energy can be stored and converted into electricity on demand or to adjust to forecasts. STPPs can easily be hybridized with conventional energy systems. All this enables better grid manageability even without storage. Among the disadvantages: Complex models are needed to forecast electricity production at any time based on the system input variables and parameters DNI it is needed!! (and not forecasted by actual NWP models) 3 Need for a complex simulation model : Large number of parameters Optical, thermal, and hydraulic characteristics of every component, Performance curves, parasitic consumption, etc Meteorological conditions (radiation, temperature, wind ) Complex physics models Probabilistic optical model (Monte-Carlo method) or by convolution, Thermohydraulich model with differential algebraic systems, High variability of input parameters (meteorological data). Variety of operating strategies Managing focusing/defocusing of the solar field Managing use of the auxiliary boiler Managing use of storage Managing g power block startup/shutdown Production optimization based on economic criteria Cost of electricity generation (Levelized Electricity Cost), Profitability, economic return. 4
Need for Direct Normal Irradiation forecasting: CIRCUMSOLAR RADIATION DIRECT NORMAL RADIATION DIFUSSE RADIATION SKY RADIATION HORIZONT BAND Global = DNI * cos θ + Diffuse 5 Radiation Plant energy simulation model Production forecast estimate + Thermal storage system status + Hybridization strategy Forecast for market operator Real STPP production can be approximated to estimated production by using Operating strategies There are three possible cases according to the goodness of the radiation forecast: Case A: Radiation is as forecast. (objective: high) Case B: The forecast deviates but electricity production can be corrected by varying the operating strategy. Case C: The forecast deviates significantly or there is no margin for correction. (objective: low) 6
Tools for forecasting meteorological phenomena Meteorological Models, also known as numerical weather e prediction models (NWP) are physics models that forecast the atmospheric conditions in a region from certain initial conditions. Statistical Models are mathematical models that use previous knowledge and the place information to model the behavior and generate prognoses. They require historical measurement series or online measurements. Satellite images provide an overall view of the atmosphere in real time. Measurements of variables provide knowledge of local behavior and the evolution of site characteristics. 7 Meteorological models Global Forecast System (GFS): Global forecasting model used as input data by many mesoscale models. It has a spatial resolution of about 1º It is executed four times a day: 00:00, 06:00, 12:00 and 18:00 It is American and access is free of charge. ECMWF. European Centre e for Medium-Range Weather e Forecasts. s This is the Central European model, used for both global and mesoscale model ing It has a spatial resolution of 0.25ºx0.25º It generates forecasts every three hours Access is not free of charge 8
Meteorological models WRF, Weather Research and Forecasting: US model developed by the NCAR,,(National Center for Atmospheric Research), NCEP (National Centers for Environmental Prediction) and FSL (Forecast Systems Laboratory) agencies. The model is executed by a multitude of users, mainly locally by universities. It is an open access code SKIRON: Model developed in different stages by the University of Athens, the NCEP and the National Meteorological Service of Greece. It is operationally executed at CENER since the end of 2005. It has been adapted to forecast in different parts of the world at different spatial and temporal resolutions It forecasts in different domains daily with a 0.1º resolution. 9 Mathematical models Weather forecasting offers acceptable precision on a general scale, but loses effectiveness when the zone of interest is very specific (that is the case for a specific CSPP). Global radiation forecasting is parameterized in the models, that is, it is a derived variable. Global radiation is forecast, but no DNI. Post processing is necessary for DNI estimation: Globa2DNI models are not solved and depends on the location. Historical measurement series collected at the site are essential. Artificial intelligence techniques and statistical learning become importance. Clear sky models are included for DNI forecasts in unclouded days. 10
A system of direct radiation forecasting based on numerical weather predictions, satellite image and machine learning. CSPP Overview Meteorological models Mathematical models 2. Forecasting system Forecasting scheme Cloud forecasting Type of data and DNI prediction 3. Results 11 11 2. Forecasting solar radiation. Forecasting diagram Global Forecast System Meteorological l Model Meteorological Forecast Mathematical post process Historical measurement series Daily forecast Time series models Online measurements Satellite images Intraday forecasts and nowcasting 12
CENER SYSTEM: Cloud cover prediction: Skiron model generates three levels of cloudy information We used a historic of satellite images to combine these three predictions to obtain a simple cloud cover forecast. 13 Combined cloud cover Satellite image 14
2. Forecasting solar radiation. Kind of hour Module: A C-SVM is implemented to predict the type of hour: Clear or cloudy hour Meteorological predictions of sea level pressure, grown temperature, relative humidity and cloudy cover percentage play the role of set of features Historical ground data are classified by comparison between real data and theoretical maxima beam radiation in two categories. 1. Clear sky days 1. Cloudy days 15 2. Forecasting solar radiation. DNI form Clear Sky model When clear type is predicted a clear sky model is used to forecast the direct radiation The worst error of this module will be predict clear data when it was cloudy register The case of clear data correctly identified generates the lowest error levels 16
2. Forecasting solar radiation. DNI from Global to direct model When cloudy type is predicted a Global to Beam model is used to forecast the direct radiation Our Global to Beam model is based on two steps: 1. Select the historical subset of data nearest to the available meteorological prediction 2. Training a nu-svm as regression model between pressure, temperature and Global predicted by Skiron and the Beam radiation of the site Beam Global l Sk 17 A system of direct radiation forecasting based on numerical weather predictions, satellite image and machine learning. CSPP Overview Meteorological models Mathematical models 2. Forecasting system Forecasting scheme Cloud forecasting Type of data and DNI prediction 3. Results 18 18
3. Results CENER BSRN station ti Real data from the BSRN station managed by CENER. June 2010-April 2011 It is sited at Sarriguren (North of Spain). Real time knowledge of measurement records are crucial for these models. Satellite images provide highfrequency global information which can be used for the forecast system. 19 3. Results Prediction error. Evolution across measure level Measure Error Level (RME%) >0 0.5 >100 0.3 >200 0.24 >300 0.20 >400 0.17 CSPP Objective >500 0.15 >600 0.14 20
3. Results Monthly prediction error. Month Error (RME%) 201006 0.11 201007 0.15 201008 014 0.14 201009 0.18 201010 0.22 201011 0.29 201012 0.22 201101 0.20 201102 0.19 201103 0.20 201104 0.16 21 3. Results RELATED TO THE CLASIFICATION: The kind of data was correctly forecast in a 65% RELATED TO THE ERROR LEVEL: The error level in the clear sky prediction is around 12% The error level in cloudy forecast is near to 35% NEXT IMPROVEMENTS RELATED TO THE MONTLY ERROR: In summer months prediction error is lower to 15% Winter months present worst errors, between 20% and 29% 22
THANK YOU VERY MUCH info@cener.comcom www.cener.com T: + 34 948 252 800 23