Thesis PREDICCIÓN DE POTENCIA A CORTO PLAZO PARA PLANTAS FOTOVOLTAICAS UTILIZANDO REDES NEURONALES ARTIFICIALES
Date
2017
Authors
CARRASCO CATALÁN, NICOLÁS ALEJANDRO
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
La predicción de potencia fotovoltaica es importante para planear la operación del sistema eléctrico.En nuestro país hasta ahora no hay un método para predicción de variables fotovoltaicas. La empresaPhineal SpA, en sus continuos avances dentro de la energía solar, intenta estudiar el problema delpronóstico para ofrecer una herramienta de predicción solar.En este estudio se ha predicho la potencia fotovoltaica de dos maneras: 1) utilizando aproximaciónde superficies mediante polinomios y 2) usando redes neuronales artificiales autoregresivas convariables exógenas (NARX). El objetivo es predecir valores horarios en un día entero. Para ambosmétodos, de forma previa se requiere la predicción de la irradiancia solar, entre otras variables meteorológicascomo la temperatura y la humedad relativa, todas predichas mediante redes neuronalesartificiales (RNA). Los datos para alimentar las RNA se recogieron desde una planta fotovoltaicaque posee una estación de medición meteorológica que capta: potencia fotovoltaica de salida de laplanta (AC), temperatura ambiente y humedad relativa (desde 02/01/2015 hasta 31/12/2015). Paraevaluar el poder de generalización de las RNA se prueban sucesivas redes, variando el tamaño deestas, probando con un conjunto de datos para validación. Las redes a escoger para predicción sonlas que tienen el menor error RMSE en el set de datos. Se prueba el poder de predicción de las RNAen dos condiciones distintas: días soleados y días nublados, pronosticando cinco días para cadasituación. Para evaluar los pronósticos, se utilizan estadísticos recomendados por el Departamentode Energía y el Laboratorio de Energías Renovables de los Estados Unidos.Los resultados indican que las Redes Neuronales pueden captar la tendencia de los datos y porconsiguiente entregar una predicción adecuada en forma, pero a veces con errores considerables.
Forecasting of solar power is important for the operation of the electrical system. In our country,until now, there isn’t a method for prediction of solar variables. Phineal SpA, in their continuousadvancements within the solar energy, tries to study the forecasting problem to offer a solar predictiontool.In this study, the solar power has been predicted in two ways: 1) using polynomial surface fittingand 2) using NARX Neural Networks. The objective is to predict hourly values in an entire day.Both methods require the prediction of solar irradiance, among other meteorological variables suchas temperature and relative humidity. All predictions have been made with Neural Networks. Thedata to feed the networks has been chosen from a photovoltaic plant that possess a meteorologicalstation that measures: solar electric output power (AC), ambient temperature and relative humidity(data from 02/01/2015 to 31/12/2015). To evaluate the generalization of the Neural Networks,by varying the size of the networks, several configurations have been tested with a validationdataset. The final networks to choose are those who have a minimal RMSE error in the dataset.The predictions are made in two conditions: clear and cloudy days, forecasting five days for eachsituation. To evaluate the predictions, statistical tools recommended by the Department of Energyand the National Renewable Energy Laboratory from USA are used.The results indicate that the Neural Networks are able to capture the trend in the data and thereforedeliver an adequate prediction in form but sometimes with considerable errors.
Forecasting of solar power is important for the operation of the electrical system. In our country,until now, there isn’t a method for prediction of solar variables. Phineal SpA, in their continuousadvancements within the solar energy, tries to study the forecasting problem to offer a solar predictiontool.In this study, the solar power has been predicted in two ways: 1) using polynomial surface fittingand 2) using NARX Neural Networks. The objective is to predict hourly values in an entire day.Both methods require the prediction of solar irradiance, among other meteorological variables suchas temperature and relative humidity. All predictions have been made with Neural Networks. Thedata to feed the networks has been chosen from a photovoltaic plant that possess a meteorologicalstation that measures: solar electric output power (AC), ambient temperature and relative humidity(data from 02/01/2015 to 31/12/2015). To evaluate the generalization of the Neural Networks,by varying the size of the networks, several configurations have been tested with a validationdataset. The final networks to choose are those who have a minimal RMSE error in the dataset.The predictions are made in two conditions: clear and cloudy days, forecasting five days for eachsituation. To evaluate the predictions, statistical tools recommended by the Department of Energyand the National Renewable Energy Laboratory from USA are used.The results indicate that the Neural Networks are able to capture the trend in the data and thereforedeliver an adequate prediction in form but sometimes with considerable errors.
Description
Catalogado desde la version PDF de la tesis.
Keywords
POTENCIA FOTOVOLTAICA , PREDICCION DE VARIABLES , REDES NEURONALES , RNA