The contribution of solar energy generation to electrical grids is fast growing. Because the weather is intrinsically fluctuating and chaotic and solar production is obviously connected to weather conditions, the supply of solar energy is fluctuating and chaotic as well. But the electrical grid must be balanced at any time. Anticipate solar production pattern would help to control the solar energy penetration into the network while keeping it balanced. Several methods have been proposed so as to anticipate the solar production, including methods grounded on time series analysis, on NWP (Numerical weather forecast) from specialised providers, on satellite images, or on sky images. These types of methods are known to be efficient at various temporal horizons. In the present paper, we show a new set of methods that takes advantage of a plurality of types of sources so as to frequently provide forecasts of solar power plants. Our sources include several NWP providers, satellite images, sky images and real time power/irradiance monitoring. Learning procedures are included to account for the plant specificities but our method is able to perform forecasts before receiving any historical data from the plant. The main types of solar power plants are taken into account including classical photovoltaic, tracked photovoltaic, concentrated photovoltaic and thermodynamic solar plants. In our experience (our developments started in 2008), testing the forecasting algorithms on one or several plants is not enough. Indeed, many behaviours or particular features can be observed according to plant, or to climate. Consequently, we tested on as much as possible plants located in various sites around the world (more than 150 plants on 3 continents). Especially, we claim that the solar energy forecasting is critical for tropical Islands. Indeed is such cases, on the one hand, a lot of solar energy may be injected in weak grid and on the other hand, the weather could varying fastly and specific climates could be observed. For that reason, the CEA and the Steadysun company develop algorithms with a special care of Mayotte island case in collaboration with EDM (Electricity De Mayotte, the grid operator) and SUNZIL (main solar plants operator in Mayotte).
展开▼