Application of Machine Learning techniques for the prediction of solar radiation

Abstract

Generation of electricity from the sun is an important renewable energy source, and the number of both, solar thermal energy systems and photovoltaic system, are proliferating around the world. However, the integration of large amounts of solar production into the electricity grid poses technical challenges due to the uctuating characteristics of available solar energy sources. Solar energy output is not easily predictable in advance and varies based on both weather conditions and site speci c conditions. Such variability of solar energy resources at ground level thus raises concerns regarding how to manage and integrate output from the solar energy systems to the power grid. Given the issues above, there is increasing interest in more precise modeling and forecasting of solar power. Irradiance is a measurement of solar power and usually measures the power per unit area. Most works consider the solar irradiance forecasting at a site, which is essentially the same problem as forecasting solar power. The ability to forecast solar irradiation will enable power grid operators to be able to ensure the quality and control of solar electricity supplies and will allow them to better accommodate highly variable electricity generation in their scheduling, dispatching, and regulation of power. In particular, the possibility to forecast solar irradiance can became fundamental in making power dispatch plans, and also a useful reference for improving the control algorithms. Today it is widely acknowledged by power producers, utility companies and independent system operators that it is only through advanced forecasting, communications and control that these distributed resources can collectively provide a rm, dispatchable generation capacity to the electricity market. Di erent solar irradiance forecast time horizons are usually employed: Some of them forecast up to 24 h or even more in what it is usually termed as mid-tolong term forecast. These forecast methods are tightly linked to weather prediction methodologies and provide averages of solar irradiation on daily or weakly basis. Whilst useful for mid-term planning of energy production, such techniques do not meet the demands of modern electricity markets. In contrast to other commodities, electricity is characterized by some speci c features that are di cult to deal with such as non-storability, the simultaneousness of electricity production, transmission and consumption, and potential network congestions, among others. All these issues lead to the inconvenience of having to program electricity generation in advance. For this reason, in the electricity liberalized markets, the day-ahead market is commonly referred to as the spot market. Nevertheless, after the closure of the day-ahead market, participants are able to sell or buy the surplus or de cit electricity in more real time (Intraday and Balancing markets). Thus, during intraday trading, participants continue to netune their positions in the light of new information about their own production,consumption and also the overall system position. For renewable energy, solar in particular, to be able to compete in intraday market, it is compulsory to improve the short-term predictability of solar production a di erent time scales. For example the Spanish electricity market considers intraday energy auctions as short as four hours ahead, that is, the utility commits to produce a given amount of energy four hours ahead of the real dispatch. Other markets operate in even more stringent basis: The German European Energy Exchange (EEX) operates a continuous trading system where power in the intraday market can be traded until 1.25 hours before delivery, and the Australian electricity market uses 5-min dispatch price and a 30-min trading price. Thus, accurate short-term forecast is essential for energy market participation, both due to forward contracting and the need for a predictable, stable and smooth supply. Accurately forecasting direct normal irradiance or global horizontal irradiance in the seconds-to-minutes time-frame (usually termed as Nowcasting) ultimately enables nely-tuned dynamic operational schedules that can reduce fuel costs, increase network stability or maximize system lifetimes. The aim of this project is to apply machine learning techniques to predict solar radiation for applications in Solar energy plants. The objective is to be able to perform nowcasting (5 min to 1 hour ahead), in the solar radiation using recent past registry of available meteorological data (solar irradiance, ambient temperature, moisture, wind velocity, barometric pressure, etc). Solar irradiance forecast methods include time series [14, 15], wavelet analysis and fuzzy logic [17, 16], satellite data and sky images [18, 19] and statistical learning methods such as arti cial neural network (ANN) [20, 21]. This work will focuss on the use of ANN to predict solar radiation. The software MATLAB and its Toolbox for Arti cial Neural Networks will be used. This document can be divided into three main parts: the rst part establishes the theoretical background of arti cial neural networks focussing in multilayer perceptrons as the main tool to be used in this work; the second part explains how the raw meteorological data has been processed (normalized, classi ed and ltered) to be used for ANN training purposes; and the third and nal part presents a series of experimental evaluations to analyze the performance of the technique proposed. The experiments carried out in this work have the objective of evaluating di erent sets of inputs and ANN con gurations to analyze the best performing structure for nowcasting. The available inputs, are subsets of the following channels of information: global radiation, di use radiation, relative humidity, atmospheric pressure, temperature and the position of the sun represented by its zenith angle. All the data used for training the multilayer perceptrons is provided by the GTER, in particular from its radiological station in the roof of the laboratories building of the Escuela T ecnica Superior de Ingenier a (Superior School of Engineering) of the University of Seville, located in Seville in Cartuja's island. With this data and the usage of the software MATLAB and its Toolbox for Arti cial Neural Networks several neural networks will be created and trained as well as tested. The experiments that will be carried out in this work deal with di erent sets of inputs from the available measurement , which are global radiation, di use radiation, relative humidity, atmospheric pressure, temperature and the position of the sun represented by its zenith angle. Another characteristic to take in account is the structure of the multilayer perceptron: number of neurons and number of layers. Di erent structures are also tested in this work to achieve the optimal structure that provides best results. After the experiments were carried out the main conclusions were that, with the data available, only clear days or clear parts of one day can be predicted accurately, while cloudy days with irregular change in the solar radiation could not be predicted with good results. In the nal chapter of this work the results of the experiments are discussed. Suggestions of future lines of work to improve this one are included as well in the nal chapter.

El objetivo de este proyecto es utilizar técnicas de machine learning para realizar la predicción a corto plazo (nowcasting) de la radiación solar. En particular, la técnica usada serán redes neuronales artificiales (Artificial Neural Networks). La meta es conseguir predicciones precisas para un tiempo cercano usando información del instante presente. En este trabajo se realiza la prueba de diferentes arquitecturas de redes neuronales, así como diferentes ventanas de tiempo. Para ello, el GTER (Grupo de Trabajo de Energías Renovables) ha cedido información recogida por una estación radiológica situada en el tejado del edificio de los laboratorios de la Escuela Superior de Ingeniería de la Universidad de Sevilla, situado en Sevilla en la isla de la Cartuja. Se dispone así de varias variables radiológicas y meteorológicas: radiación global y difusa, presión atmosférica, humedad relativa y temperatura del aire. Usando la hora del día se calcula también la posición del sol como otra variable disponible.

Con estos datos y el uso del software MATLAB y su módulo (toolbox ) para redes neuronales artificiales se crean y se entrenan diferentes redes para ser después puestas a prueba. El documento principal está dividido en tres partes: el entorno teórico en el que se trabaja, el procesado de los datos y finalmente los experimentos realizados y las conclusiones. Después de realizar diferentes experimentos con diferentes combinaciones de datos de entrada, diferentes arquitecturas de redes y diferentes ventanas de predicción se puede concluir que con los datos disponibles sólo es posible predecir de manera precisa días, o partes de un día, claros mientras que en los días nublados la radiación varía de tal manera que la red es incapaz de predecir su comportamiento.