[MUSIC] Hello, in this sequence I will show you methods used for forecasting. It is an essential element of the energy system to allow better penetration of highly variable renewable energies entering supply demand balance at all times and anticipating tense situations for the network. You will learn the difference between statistical and numerical methods. Their most suitable field of application for forecasting and their limit, forecasting tools are very dependent on forecast time horizons and applications. The table presents the modeling needs for two types of activities in the energy sector, energy production and decision support. Forecasting needs for energy production are available over a wide variety of time scales needs can range from minutes to our for adjusting production, at this time scales. The benchmark to beat in terms of forecasting is persistence, the techniques deployed to do better than persistence or the so called statistical methods fed by the most recent measurements available. For time horizons of a few hours forecasting needs relate to intraday market exchanges, the balancing of supply and demand and the allocation of means of production. This forecasting time horizons up to six hours, the forecast relies on a combination of statistical and numerical methods used for weather forecasting, from a few hours, two to three days. The needs relate to exchanges on the so called day ahead market, the estimation of the demand and the allocation of production means, in this context only numerical methods are applicable. Nevertheless systematic biases can appear which must be corrected by statistical methods called de biasing. Finally, beyond a few days numerical methods become more difficult to use, probabilistic approaches allow information to be extracted on monthly or seasonal time horizon. The benchmark to beat at this time scales is the climatology, at multi decennial timescales. We use numerical methods in a scenario framework to simulate possible changes in climate with key questions of great importance for the energy sector. What will be the renewable resource available? What impact will global warming have energy consumption? What should be the energy mix to limit global warming to do your scene? Finally, sometimes combined with numerical techniques, statistical techniques are widely used continuously for the detection of production ramps requiring immediate actions. They are also used for forecasting demand, which cannot be modeled by mathematical or physical equations at the heart of numerical techniques, as seen before. We can therefore identify two main families of forecasting methods, statistical methods and numerical methods. In some cases, hybridization of physical and statistical models may be necessary to derive the greatest benefit from each model. Statistical models are based on establishing statistical relationships between input and output data without integrating the physics of the underlying processes. Statistical methods can be divided into two large hounds, on the one hand, the so called supervised methods requiring a learning data set to calibrate the statistical models. On the other hand, so called unsupervised methods, which unlike supervised methods do not require learning, we can divide the supervised methods into two categories. Classification methods and regression methods, which includes for example, linear parametric or non perimeter rick regressions with random forests and non linear regressions with for example, neural networks. Unsupervised techniques can be used to solve, among other things, data clustering, for example with K means algorithm or your are kicking clustering. They can be also used for the estimation of distribution density or the reduction of dimension, with for example, the principal component analysis. In the following, I will illustrate the results of statistical forecasting methods by comparing parametric and non parametric linear regression techniques. Linear regression by auto regressive methods and non linear regression by neural networks. In a parametric linear regression model we denote by why the random variable to explain like the wind speed and X. The exogenous explanatory variable which can be a measurement or assimilated data, for example, the temperature of the pressure. The linear model amounts to assuming that on average the expectation of Y is an a fine function of X, writing the model implicitly assumes a prior notion of causality between Y and X. Random forests are part of machine learning techniques, the basis of calculation is based on learning by decision tree. This learning method designates a method based on the use of decision trees as a predictive model. In these three structures, the leaves represent the values of the variable to be explained and the branches correspond to combinations of explanatory variables that lead to these values. The Arma models, auto regressive and moving average models are the main models of time series, given time Syria variable Y of T. The Arma model is a tool for understanding and possibly predicting the future values of this series, in the context of renewable energy production. Y can be for example the wind speed or the solar radiation on the surface, the model is composed of two parts on the one hand and also aggressive part AR. Which links the value of the variable Y at time T to its value at earlier times, on the other hand, a moving average part MA. It can also include exogenous variables such as measurement or simulated data of temperature or pressure, an example of non leader regression is the artificial neural network. It is a system whose designed was originally schematically inspired by the functioning of biological reinsurance and which subsequently approached statistical methods. Neural networks are generally optimized by probabilistic learning methods, they are placed both in the family of statistical applications and in the family of artificial intelligence methods. This figure shows the forecast of the wind speed from 10 minutes to about three hours on a wind farm of a producer in the west of France, the forecast to beat is persistent. In the figure the ECMWF forecast is the numerical forecast from the European weather forecast center. Other simulations corresponding to statistical methods include only measurements prior to the forecast, these are the Arma model and the ANN neural network. The linear regression LR and the random forest RF only contain measurements taken during the first hour of forecast, beyond one hour. These two models combined measurements before the forecast and variables from the mm WF numerical forecast, we see that for images forecast time horizon less than one hour. The statistical methods are particularly suitable since they all improve persistence, this is not the case for numerical forecasting, the second family of models is the numerical models. Numerical models are based on mathematical representation of certain physical processes. These processes can generally be described by a set of mathematical equations. For example, the Navy Stock equations cover in the movement of fluids and therefore geophysical fluids such as water in the ocean and in the atmosphere. In meteorological or climate modeling, it is the set of conservation equations of momentum energy, water that are sold by the numerical model. These equations are then simplified by placing themselves within the framework of certain approximations. For example, the business approximation widely used for geophysical fleets, these simplified equations called primitive equations are the basis of the model. It is not necessary to implement this equation on a computer, for this establish an artificial three dimensional mesh of the environment that we want to model here the atmosphere and the ocean. We virtually cut the atmosphere ocean into mesh several kilometers on each size. The size of the mesh will condition the computing time, we sold the equations inside each box in the resolution of the primitive equations. Apart translates the exchanges between boxes and is resolved in an explicit way by dis criticizing the partial differential equations on the meshes for the numerical resolution. This part is identified in blue, some phenomena are not explicitly resolved and must be taken into account in the conservation equations as a source or sink term and are circled in red. This source or sink terms are called physical parameterization and their development is a field of research in its own in climate sciences. Parameterization is a method for replacing processes that are too small or complex to be physically represented in a numerical model by a simplified process. Example include speed of full of raindrops, convective clouds, simplifications of atmospheric radiative transfer, cloud micro physics, turbulence, surface atmosphere exchanges. The number of processes to be parameterized are numerous, we can count about 20. Since the advent in the 1970s of numerical modeling as a tool for analyzing and forecasting, weather and climate models have evolved considerably. First, their special resolution is very strongly refined whatever the applications of the model in weather forecast or climate projection mode, with the horizontal resolutions of several 100 kilometers for the first IPCC report. The climate models currently have a resolution of the order of or less than 100 kilometers, for weather forecasting the horizontal resolution is around 10 kilometers on a global scale. The parameterize physical processes have multiplied and become more complex, very rudimentary. In the 70s, couple models for weather and climate integrate a very large number of processes controlling the functioning of the art system. But whatever the period of their development, these models all work the same way, we program in computer language, the mathematical model, the complexity of which has been constantly increasing. Then we start the temporal integration, that is to say that at each time step, delta T, the equations are sold by integrating the explicit part and the para motorist part. Inside each box variables such as temperature, precipitation, humidity, wind speed and direction are determined at each time step. At the end of the process we test the model against field observations in order to continuously improve it, modeling can be performed on a global scale. In this case the whole atmosphere is meshed, modeling can also be carried out on a regional scale in order to be able to apply a finer special resolution without prohibitive numerical cost. And thus obtain a more realistic representation of atmospheric phenomena, in this case a sub domain is defined with a finer mesh. The simulations at the global scale is then used to be able to provide the regional model with lateral boundary conditions at each time step. This figure is similar to that previously shown, this time, it shows the forecast of the wind speed on the wind farm from 10 minutes to about 11 hours. The forecast to beat is always persistence, in the figure the ECMWF numerical forecast is this time the most suitable forecasting method for forecast time horizons beyond 23 hours. To conclude them, the tools differ according to forecasting needs for very short term weather forecasting less than six hours. The hypothesis of persistence or affection are often preferred and are generally based on statistical methods, for forecasting needs from a few hours to a few days. Numerical methods are the most suitable, this is also the case for climate projections. We are talking here more about projection than forecast because on climate scales their system has no memory and therefore the predictable time horizons are largely exceeded. It is less a forecast than a scenario based on possible trajectories of greenhouse gas emissions. Finally, as intermediate scales of the week to a few months, the use of climatology is the most classic, for numerical methods depending on the time scales and for different applications. The implementation configuration is different, for numerical forecast from a few hours to a few days. The strong constraint comes from the initial conditions of the model, for forecasts from a week to a few months, it is the boundary conditions which become dimensioning. Finally, for climate scales, it is the anthropic scenarios themselves which constrain the trajectory of the projections, in this sequence. We have reviewed the methods used for geophysical forecasting in connection with energy issues. You have learned about the different forecasting methods, their relevance to forecasting horizons and their most appropriate field of application in the energy field. I thank you for your attention. [MUSIC]