Solar power forecasting

Solar power forecasting is the gathering and analysis of data in order to predict the optimal conditions for solar power generation.

This information includes the Sun´s path, the atmosphere's condition, the scattering processes and the characteristics of the solar energy plant. The power output depends on the incoming radiation and on the solar panel characteristics. Solar power forecast information is used for efficient management of the electric grid and for solar energy trading.

Common solar forecasting method include stochastic learning methods, local and remote sensing methods, and hybrid methods (Chu et al. 2016).

It is useful to classify these techniques depending on the forecasting horizon

• now-casting (forecasting 3–4 hours ahead),
• short-term forecasting (up to seven days ahead) and
• long-term forecasting (months, years...)

Nowcasting

Solar power nowcasting then refers to the prediction of solar power output (or energy generation) over time horizons of tens to hundreds of minutes ahead of time with up to 90% predictability. It has historically been important for electrical grid operators in order to guarantee the matching of supply and demand on energy markets. Solar power nowcasting services are usually related to temporal resolutions of 5 to 15 minutes, with updates as frequent as every 5 minutes.

The regular updates and relatively high resolutions required for nowcast require automatic weather data acquisition and processing techniques. These are accomplished by three primary means:

Statistical techniques

These techniques are usually based on time series processing of measurement data, including meteorological observations and power output measurements from a solar power facility. What then follows is the creation of a training dataset to tune the parameters of a model (I. Espino eta al, 2011), before evaluation of model performance against a separate testing dataset. This class of techniques includes the use of any kind of statistical approach, such as autoregressive moving averages (ARMA, ARIMA, etc.), as well as machine learning techniques such as neural networks, support vector machines (etc.). These approaches are usually benchmarked to a persistence approach in order to evaluate their improvements. This persistence approach just assumes that any variable at time step t is the value it took in a previous time.

Satellite based methods

These methods leverage the several geostationary Earth observing weather satellites (such as Meteosat Second Generation (MSG) fleet) to detect, characterise, track and predict the future locations of cloud cover. These satellites make it possible to generate solar power forecasts over broad regions through the application of image processing and forecasting algorithms. Key forecasting algorithms include cloud motion vectors (CMVs).

An example of sky-imager used for detecting, tracking and predicting cloud cover conditions in the vicinity of a solar energy facility of interest. Most often, these devices are used to make estimates of solar irradiance from the images using local calibration by a pyranometer. The solar irradiance short-term forecasts are then fed into PV power modelling routines to generate a solar power forecast.
Credit: UC San Diego

Ground based techniques.

These techniques are used to derive irradiance forecasts with much higher spatial and temporal resolution compared with the satellite-based forecasts. Local cloud information is acquired by one or several ground-based sky imagers at a high frequency (1 minute or less). The combination of these images and local weather measurement information are processed to simulate cloud motion vectors and optical depth to obtain forecasts up to 30 minutes ahead.

Short-term solar power forecasting

Short-term forecasting provides predictions up to seven days ahead. This kind of forecast is also valuable for grid operators in order to make decisions of grid operation, as well as, for electric market operators. Meteorological variables and phenomena are looked from a more general perspective, not as local as nowcasting services do.

Numerical weather prediction

Most of the short term forecast approaches use numerical weather prediction models (NWP) that provide an initial estimation of weather variables. These included the Global Forecast System (GFS) or data provided by the European Center for Medium Range Weather Forecasting (ECMWF). These two models are considered the state of the art of global forecast models, which provide meteorological forecasts all over the world.

In order to increase spatial and temporal resolution of these models, other models have been developed which are generally called mesoscale models. Among others, HIRLAM, WRF or MM5 are the most representative of these models since they are widely used by different communities. To run these models a wide expertise is needed in order to obtain accurate results, due to the wide variety of parameters that can be configured in the models. In addition, sophisticated techniques such as data assimilation might be used in order to produce more realistic simulations.

Some communities argue for the use of post-processing techniques, once the models’ output is obtained, in order to obtain a probabilistic point of view of the accuracy of the output. This is usually done with ensemble techniques that mix different outputs of different models perturbed in strategic meteorological values and finally provide a better estimate of those variables and a degree of uncertainty, like in the model proposed by Bacher et al. (2009)

Long-term solar power forecasting

Long-term forecasting usually refers to forecasting of the annual or monthly available resource. This is useful for energy producers and to negotiate contracts with financial entities or utilities that distribute the generated energy.

In general, these long-term forecasting is usually done at a lower scale than any of the two previous approaches. Hence, most of these models are run with mesoscale models fed with reanalysis data as input and whose output is postprocessed with statistical approaches based on measured data.

Energetic models

Any output from a model must then be converted to the electric energy that a particular solar PV plant will produce. This step is usually done with statistical approaches that try to correlate the amount of available resource with the metered power output. The main advantage of these methods is that the meteorological prediction error, which is the main component of the global error, might be reduced taking into account the uncertainty of the prediction.

As it was mentioned before and detailed in Heinemann et al., these statistical approaches comprises from ARMA models, neural networks, support vector machines, etc. On the other hand, there also exist theoretical models that describe how a power plant converts the meteorological resource into electric energy, as described in Alonso et al. The main advantage of this type of models is that when they are fitted, they are really accurate, although they are too sensitive to the meteorological prediction error, which is usually amplified by these models. Hybrid models, finally, are a combination of these two models and they seem to be a promising approach that can outperform each of them individually.

See also

Energy portal

• Energy forecasting

References

• Y. Chu, M. Li and C.F.M. Coimbra (2016) “Sun-Tracking Imaging System for Intra-Hour DNI Forecasts” Renewable Energy (96), Part A, pp. 792–799.
• Luis Martín, Luis F. Zarzalejo, Jesús Polo, Ana Navarro, Ruth Marchante, Marco Cony, Prediction of global solar irradiance based on time series analysis: Application to solar thermal power plants energy production planning, Solar Energy, Volume 84, Issue 10, October 2010, Pages 1772-1781, ISSN 0038-092X, doi:10.1016/j.solener.2010.07.002.
• Heinemann, D., Lorenz E., Girodo M. Forecasting of solar radiation. Oldenburg University, Institute of Physics, Energy Meteorology Group.
• Alonso, M, Chenlo F. Estimación de la energía generada por un sistema fotovoltaico conectado a red. CIEMAT. Laboratorio de sistemas fotovoltaicos.
• Alvarez, L., Castaño, C.A., Martín, J. A computer vision approach for solar radiation nowcasting using MSG images. EMS Annual Meeting Abstracts. Vol. 7, EMS2010-495, 2010. 10th EMS/8th ECAC.
• Espino, I., Hernández, M.. Nowcasting of wind speed using support vector regression. Experiments with Time Series from Gran Canaria. Renewable Energy and Power Quality Journal, ISSN 2172-038X, N9, 12 May 2011.
• Bacher, P., Madsen, H., Nielsen H.A. Online short-term solar power forecasting. Solar Energy. Vol 83, Issue 10, October 2009: 1772-1783.
• Diagne, H.M., David, M., Lauret, P., Boland, J. Solar irradiation forecasting: state-of-the-art and proposition for future developments for small-scale insular grids. In Proceedings of the World Renewable Energy Forum 2012 (WREF 2012), Denver, USA, May 2012.

External links

• How to predict solar energy production, by Rafał Rybnik, case study on predicting solar electricity production from weather data.
• Solar and Wind Forecasting projects, by National Renewable Energy Laboratory (NREL).
• https://www.meteoswift.fr/en/solar-forecasts/ Short-term Solar Power Forecasts by meteo*swift