Publication: Métodos de predicción de la generación agragada de energía eólica
Loading...
Identifiers
Publication date
2010
Defense date
2010-11-12
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
El fomento y desarrollo de las energías renovables, y como consecuencia la alta penetración eólica en los sistemas eléctricos, ha hecho necesario el uso de herramientas de predicción, para conocer con antelación suficiente la cantidad de energía de origen eólico que será inyectada en la red y poder coordinar el resto de fuentes de generación. En ocasiones existe un especial interés en la predicción de generación agregada de varios parques en una región, ya sea por parte del operador del sistema eléctrico, o por los agentes que gestionan la energía que generan varios parques. Por otro lado, suele haber un suavizado espacial que reduce el error de las predicciones agregadas frente al error de predicción de los parques individuales. Estos motivos hacen de la predicción agregada un modo muy atractivo de predicción eólica. En esta tesis se propone un método de predicción de la generación agregada de energía eólica en una región, mediante la búsqueda de similitudes entre el vector de viento previsto en algunas coordenadas de la región y otros vectores históricos de predicción de viento. El modelo propuesto se basa en modelos locales de suavizado de Media Ponderada y de Regresión Local Ponderada. Estos modelos ajustan de forma no paramétrica un modelo de predicción para cada punto de entrada al problema, que en nuestro caso será un vector de viento previsto. Para hacer el suavizado de los datos, se proponen y comparan distintos procedimientos para tener en cuenta la medida de distancias entre vectores, el método de selección de datos cercanos, y las funciones de ponderación por distancia y antigüedad de los datos seleccionados. Mediante un procedimiento de estimación adaptativa de parámetros, se conseguirá que el modelo de predicción se adecue a la evolución temporal de la compleja relación no lineal existente entre el viento previsto en la región y la potencia eólica total generada. Las predicciones de potencia obtenidas con los modelos de Media Ponderada y Regresión Local Ponderada, se combinan para aprovechar las ventajas que ofrecen ambas aproximaciones en función de la complejidad del problema de predicción en cada momento. El modelo propuesto se validará comparando sus predicciones con la agregación de predicciones que se obtienen con Sipreólico, una herramienta que calcula predicciones para cada uno de los parques en la región y después las suma. Por último, para dar mayor valor a las predicciones agregadas, se propone un modelo de predicción probabilista condicionada a los vectores de viento previsto mediante estimación de densidades con Kernel, para el cual se utilizan las medidas de distancia y funciones de ponderación propuestas para el modelo de predicción de potencia.---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The promotion and development of renewable energy resources, and as a result, the high wind power penetration in the power systems, makes necessary the use of forecasting tools, in order to know in advance the amount of wind energy that will be injected into the network and to coordinate all the other generation sources. Sometimes, there is a particular interest in predicting the aggregated wind power in a region, either by the system operator or other practitioners who manage the energy generated by several wind farms. On the other hand, there is usually a spatial smoothing effect that reduces the prediction error when making aggregated wind power predictions, instead of predicting individual wind farms. These reasons make the aggregated wind power prediction models very appealing. This thesis proposes a method to make aggregated wind power predictions in a region, by seeking similarities between the next wind speed forecasting vector and a set of historical vectors of wind speed in some coordinates of the region. The proposed model is based on Weighted Average and Locally Weighted Regression smoothing models. These models fit a non-parametric prediction model for each query point, which in our case is a vector containing the wind speed forecasts for a set of coordinates in the region. In order to smooth the data, different smoothing procedures are proposed and compared. These procedures must take into account the measure of distances between wind speed vectors, the nearby data selection method, and the distance and age weighting functions for the selected data. An adaptive estimation procedure of the parameters will let the model to fit the timevarying evolution of the complex nonlinear relationship between the expected wind speed in the coordinates and the total wind power generated in the region. The wind power predictions obtained with the alternative models, based on Weighted Average and Locally Weighted Regression, are combined in order to exploit the advantages of both approaches depending on the complexity of the prediction problem in every moment. The proposed model is validated by comparing its predictions with the aggregation of predictions obtained with Sipreólico, a wind power prediction tool that makes predictions for each of the wind farms in the region and adds them. Finally, more value is added to the aggregated model by proposing a method to make probabilistic forecasts conditioned to the expected wind speed vectors using Kernel density estimation. This approach uses the measures of distance and the weighting functions proposed for the wind power prediction model.
The promotion and development of renewable energy resources, and as a result, the high wind power penetration in the power systems, makes necessary the use of forecasting tools, in order to know in advance the amount of wind energy that will be injected into the network and to coordinate all the other generation sources. Sometimes, there is a particular interest in predicting the aggregated wind power in a region, either by the system operator or other practitioners who manage the energy generated by several wind farms. On the other hand, there is usually a spatial smoothing effect that reduces the prediction error when making aggregated wind power predictions, instead of predicting individual wind farms. These reasons make the aggregated wind power prediction models very appealing. This thesis proposes a method to make aggregated wind power predictions in a region, by seeking similarities between the next wind speed forecasting vector and a set of historical vectors of wind speed in some coordinates of the region. The proposed model is based on Weighted Average and Locally Weighted Regression smoothing models. These models fit a non-parametric prediction model for each query point, which in our case is a vector containing the wind speed forecasts for a set of coordinates in the region. In order to smooth the data, different smoothing procedures are proposed and compared. These procedures must take into account the measure of distances between wind speed vectors, the nearby data selection method, and the distance and age weighting functions for the selected data. An adaptive estimation procedure of the parameters will let the model to fit the timevarying evolution of the complex nonlinear relationship between the expected wind speed in the coordinates and the total wind power generated in the region. The wind power predictions obtained with the alternative models, based on Weighted Average and Locally Weighted Regression, are combined in order to exploit the advantages of both approaches depending on the complexity of the prediction problem in every moment. The proposed model is validated by comparing its predictions with the aggregation of predictions obtained with Sipreólico, a wind power prediction tool that makes predictions for each of the wind farms in the region and adds them. Finally, more value is added to the aggregated model by proposing a method to make probabilistic forecasts conditioned to the expected wind speed vectors using Kernel density estimation. This approach uses the measures of distance and the weighting functions proposed for the wind power prediction model.
Description
Keywords
Energía eólica, Predicción de generación de energía