Publication: Predicción demanda eléctrica española. Implementación de redes neuronales recurrentes en Python
Loading...
Official URL
Full text at PDC
Publication Date
2018-09-01
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Este proyecto de fin de máster está realizado de manera conjunta a la empresa Innova-tsn. El objetivo principal de este proyecto es la predicción de la demanda eléctrica nacional mediante el desarrollo de modelos de redes neuronales recurrentes. Las redes neuronales están tomando gran importancia en temas de inteligencia artificial, como pueden ser modelos de clasificación y predicción. Por ello el trabajo tratará de explicar de manera teórica qué es una red neuronal y cómo funciona.
Una vez tengamos una primera visión de las redes nos centraremos en las recurrentes, las cuales son más sofisticadas por lo que nos permitirán obtener mejores resultados. Dentro de estas redes recurrentes nos centraremos en las redes de Elman, Jordan, LSTM y GRU.
Después de una primera parte teórica nos centramos en el estudio de la serie a tratar. Para ello se hace un análisis descriptivo de la demanda eléctrica nacional previo a la implementación en Python de los modelos de redes elegidos. También se estudiará el efecto de algunas variables explicativas en nuestra predicción.
Por último se presentarán los algoritmos realizados y se compararán predicciones realizadas sin tener un modelo, mediante modelos ARIMA/SARIMA y los obtenidos con las redes neuronales con el objetivo de ver si estas redes mejoran o no los modelos que actualmente manejamos. Esta comparación se realizará mediante un backtesting durante un año. Además se tratará de predecir también una demanda a nivel horario mediante dos estrategias.
This dissertation has been developed in collaboration with Innova-tsn consulting. The main goal of the proyect is the prediction of the national electrical demand through the development of recurrent neural network models. Neural networks are acquiring importance in artificial intelligence, as they can be used for classification and prediction. Because of that, the dissertation initially provides an explanation of what a neural network is and how it works. Once we have a first view on neural networks, we focus on recurrent neural networks, which are more sofisticated and they allow us to achieve better results. More specifically, within these recurrent networks we study Elman and Jordan networks and LSTM and GRU cells. After a first theoric part we start the study of the time serie. To do that, we make a descriptive analysis of the national electrical demand prior to the implementation of the network models chosen in Python. We also study the effect of some input variables in our prediction. Finally, we present the developed algorithms and we compare the predictions obtained without a model, with ARIMA/SARIMA models, and those obtained with neural networks, to understand if they actually improve on the results. This comparison is done by backtesting over a year. Moreover, we also predict the hourly national electrical demand with two different strategies and compare their performance.
This dissertation has been developed in collaboration with Innova-tsn consulting. The main goal of the proyect is the prediction of the national electrical demand through the development of recurrent neural network models. Neural networks are acquiring importance in artificial intelligence, as they can be used for classification and prediction. Because of that, the dissertation initially provides an explanation of what a neural network is and how it works. Once we have a first view on neural networks, we focus on recurrent neural networks, which are more sofisticated and they allow us to achieve better results. More specifically, within these recurrent networks we study Elman and Jordan networks and LSTM and GRU cells. After a first theoric part we start the study of the time serie. To do that, we make a descriptive analysis of the national electrical demand prior to the implementation of the network models chosen in Python. We also study the effect of some input variables in our prediction. Finally, we present the developed algorithms and we compare the predictions obtained without a model, with ARIMA/SARIMA models, and those obtained with neural networks, to understand if they actually improve on the results. This comparison is done by backtesting over a year. Moreover, we also predict the hourly national electrical demand with two different strategies and compare their performance.
Description
Proyecto fin de máster realizado de manera conjunta a la empresa Innova-tsn.
UCM subjects
Unesco subjects
Keywords
Citation
[1] ARMANDO FANDANGO, Python Data Analysis- Second Edition, Packt Publishing, Birmingham,
marzo 2017.
[2] AURELIÉN GÉRON, Hands-On Machine Learning with Scikit-Learn and TensorFlow: Concepts,
Tools, and Techniques to Build Intelligent Systems, O’Reilly Media, Sebastopol, marzo 2017.
[3] FRANÇOIS CHOLLET, Deep learning with Python, http://www.deeplearningitalia.com/
wp-content/uploads/2017/12/Dropbox_Chollet.pdf Manning Publications, Nueva York.
[4] IAN GOODFELLOW, YOSHUA BENGIO, AARON COURVILLE, Deep Learning, http://www.deeplearningbook.org, MIT Press, 2016.
[5] N. D LEWIS, Deep time series forecasting with Python, diciembre, 2016.
[6] KERAS, Keras: The Python Deep Learning library, https://keras.io/.
[7] KAGGLE, Kaggle, https://www.kaggle.com/.
[8] DANIEL VÉLEZ SERRANO, CARLOS RIVERO, MARÍA DEL CARMEN PARDO, Material de la
asignatura Estadística Aplicada y Minería de Datos, Universidad Complutense de Madrid.
[9] ELENA ROSA PÉREZ, FEDERICO LIBERATORE, Material de la asignatura Técnicas de Predicción,
Universidad Complutense de Madrid.