Publication: Predicción de energía eólica utilizando técnicas de aprendizaje automático
Loading...
Identifiers
Publication date
2015
Defense date
2015-03
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
El uso de energías renovables beneficia al medio ambiente, sustituyen la utilización de combustibles fósiles reduciendo de manera considerable el impacto medioambiental. La generación de energía eólica es una forma de generación no programable, ya que solo se produce energía cuando sopla el viento, que puede llegar a ser muy variable incluso en el corto plazo, por lo cual es difícil conocer con antelación y precisión suficiente la cantidad de energía eólica con la que podremos contar en cada momento. Esta variabilidad ocasiona más complejidad a su operación, por lo que su producción en el futuro tiene que ser estimada o prevista inevitablemente, para hacer factible el desarrollo e implantación de la energía eólica, y su integración en el sistema eléctrico.
Por ello, para este proyecto se ha tomado la decisión de estudiar y analizar diferentes métodos de regresión para la predicción de la energía eólica.
Para realizar este estudio y análisis se utiliza los datos de los siete parques eólicos de una competición de 2012. Con estos datos se estudian diferentes aproximaciones o modelos para abordar el problema de la competición, que consiste en la predicción de la generación de la energía eólica en un horizonte de 48 horas.
Para construir los diferentes modelos de predicción se ha utilizado la interfaz gráfica de la herramienta de análisis de datos Weka, y el lenguaje Java para dar formato y estructura al conjunto de datos de entrenamiento y test (asignar a las variables de entradas del modelo la variable de salida correspondiente), también se realiza una breve descripción de otras alternativas como herramientas y lenguajes de análisis y tratamiento de datos.
Para validar y evaluar los diferentes modelos utilizamos tres métodos de regresión, el método de regresión lineal, el Perceptron Multicapa y el M5P. De esta manera se aplican diferentes técnicas de aprendizaje automático tales como técnicas basadas en redes de neuronas, y árboles de regresión, para construir diferentes modelos lineales y no lineales de predicción.
En este trabajo se realiza tres alternativas de diseño de la solución, en otras palabras, tres alternativas para construir modelos de predicción. La primera alternativa consiste en generar modelos de predicción para la energía eólica basados en las variables meteorológicas que miden la velocidad y dirección del viento, la segunda alternativa consiste en generar modelos que incluyen variables de entrada con valores de instantes anteriores de la serie de la producción además de incluir el horizonte temporal (hora a la que se predice) y las variables meteorológicas utilizadas para el mejor modelo que resulte en la primera alternativa y por último la tercera alternativa consiste en generar modelos con diferentes bloques de horizontes de predicción.
Durante el estudio y análisis de cada modelo se realiza el estudio de los mejores parámetros y configuración para los métodos de regresión basados en redes de neuronas y árboles de regresión, se valida cada modelo de predicción para cada una de las granjas y finalmente se evalúa el mejor modelo con el test real o test de la competición.
Una vez estudiadas y analizadas las tres alternativas de diseño de modelos predictivos mediante diferentes técnicas de aprendizaje automático, se decide qué modelos y técnicas son más adecuados para realizar la predicción de la energía eólica para un horizonte de 48 horas, para cada una de las granjas con los datos proporcionados por una competición que se ha celebrado en el 2012.
Renewable energy use benefits the environment, it subtitutes fossil fuels reducing considerably in this way the environmental impact. The production of wind energy isn’t programmmable since it produces energy only when the wind blows, which can be highly variable even in the short term, so it is difficult to know in advance and sufficient precision the amount of wind energy that we can count with at all times. This variability brings more complexity to its operation, so its future production must inevitably be estimated or expected, to make possible the development and implementation of wind energy, and its integration into the electrical system. Therefore, to carry out this project we decided to study and analyze different regression methods for predicting wind energy. To perform this study and analysis, we used the data of seven wind farms of a competition in 2012. With these data we studied different approaches or models to address the problem of competition, which is the prediction of wind power generation in a horizon of 48 hours. To construct the different prediction models we used the graphical interface tool Weka data analysis, and Java language to give format and structure to the training data and test set (To assign the model input variables to its corresponding output variables), we also perform a brief description of alternatives such as languages and tools of analysis and data processing. To validate and evaluate the different models we used three methods of regression: linear regression method, the Multilayer Perceptron and M5P. Thus different automatic learning techniques such as techniques based on neuronal networks and regression trees are applied to construct different linear and nonlinear prediction models. In this work, we have performed three alternatives for the solution design, in other words, three alternatives to build predictive models, the first one is to generate predictive models for wind power based on meteorological variables that measure the wind speed and direction , the second alternative is to generate models including input variables with values from previous moments of the series production as well as including the time horizon (time which is predicted) and meteorological variables used for the best model that results in the first alternative and finally the third alternative consists to generate models with different block prediction horizons. During the study and analysis of each model, we perform the study of the best parameters and configuration for the regression methods based on neural networks and regression trees, we validate each prediction model for each of the farms and finally we evaluate the best model with the real test or the competition test. Once the three design alternatives predictive models studied and analysed which uses different machine learning techniques, then we decided which models and techniques are more suitable for predicting wind power for a horizon of 48 hours for each of the farms with data provided by a competition that was held in 2012.
Renewable energy use benefits the environment, it subtitutes fossil fuels reducing considerably in this way the environmental impact. The production of wind energy isn’t programmmable since it produces energy only when the wind blows, which can be highly variable even in the short term, so it is difficult to know in advance and sufficient precision the amount of wind energy that we can count with at all times. This variability brings more complexity to its operation, so its future production must inevitably be estimated or expected, to make possible the development and implementation of wind energy, and its integration into the electrical system. Therefore, to carry out this project we decided to study and analyze different regression methods for predicting wind energy. To perform this study and analysis, we used the data of seven wind farms of a competition in 2012. With these data we studied different approaches or models to address the problem of competition, which is the prediction of wind power generation in a horizon of 48 hours. To construct the different prediction models we used the graphical interface tool Weka data analysis, and Java language to give format and structure to the training data and test set (To assign the model input variables to its corresponding output variables), we also perform a brief description of alternatives such as languages and tools of analysis and data processing. To validate and evaluate the different models we used three methods of regression: linear regression method, the Multilayer Perceptron and M5P. Thus different automatic learning techniques such as techniques based on neuronal networks and regression trees are applied to construct different linear and nonlinear prediction models. In this work, we have performed three alternatives for the solution design, in other words, three alternatives to build predictive models, the first one is to generate predictive models for wind power based on meteorological variables that measure the wind speed and direction , the second alternative is to generate models including input variables with values from previous moments of the series production as well as including the time horizon (time which is predicted) and meteorological variables used for the best model that results in the first alternative and finally the third alternative consists to generate models with different block prediction horizons. During the study and analysis of each model, we perform the study of the best parameters and configuration for the regression methods based on neural networks and regression trees, we validate each prediction model for each of the farms and finally we evaluate the best model with the real test or the competition test. Once the three design alternatives predictive models studied and analysed which uses different machine learning techniques, then we decided which models and techniques are more suitable for predicting wind power for a horizon of 48 hours for each of the farms with data provided by a competition that was held in 2012.
Description
Keywords
Inteligencia artificial, Aprendizaje, Energía eólica