Aluno: Rafael Thomazi Gonzalez
Orientador: Prof. Dr. Claudio Dante Augusto Couto Barone
Título: Using Deep Learning and Evolutionary Algorithms for Time Series Forecasting
Linha de Pesquisa: Aprendizado de Máquina, Representação de Conhecimento e Raciocínio
Local: Prédio 43412 – Sala 215 (sala de videoconferência) do Instituto de Informática.
Prof. Dr. Claudio Fernando Resin Geyer (UFRGS)
Profª. Drª. Mara Abel (UFRGS)
Prof. Dr. Rodrigo Coelho Barros (PUCRS)
Presidente da Banca: Prof. Dr. Claudio Dante Augusto Couto Barone
Abstract: Time series analysis is widely used in fields such as business, economics, finance, science, and engineering. One of the main purposes of time series data analysis is to use past observations from the data to forecast future values. Moreover, time series data analysis allows you to represent the data in a form that can convey changes over time. Many different time series forecasting algorithms have been explored in machine learning and statistics literature. More recently, deep neural networks have been increasingly used, since they can be trained in such a way that they are effective at representing many kinds of data, including raw and featurized data. This thesis aims to assess the performance of Deep Learning algorithms optimized by an Evolutionary Algorithm in predicting different time series. First, a description of the selected Deep Learning algorithms will be presented, namely Stacked Autoencoder (SAE), Stacked Denoising Autoencoder (SDAE) and Long Short-Term Memory Networks (LSTM). The Feedforward Multilayer Perceptron (MLP) network is used frequently in time series prediction, and thus it is used as baseline to compare these Deep Learning models. Given the complexity of these models, their hyperparameters are optimized by an Evolutionary Algorithm called Covariance Matrix Adaptation Evolution Strategy (CMAES). The strengths and drawbacks of CMAES are also highlighted in order to explain why it is considered as state-of-the-art and one of the most powerful Evolutionary Algorithms for real-valued optimization. In order to demonstrate the performance of the proposed approach on forecasting time series, experiments are performed using three different datasets. Two of them are artificial data generated by the Mackey-Glass and Lorenz System equations. The third one includes real data of hourly energy demand. Throughout the analysis of the results, it was found that some models, such as LSTM and MLP, perform better on data presenting some degree of seasonality; while models with pre-processing layers (i.e. SAE and SDAE) have difficulties learning the time structure of the data. Problems containing time series data behave similar to many other machine learning problems such that there is no master algorithm that is the best for all problems. Therefore, this study supports the effectiveness of deep learning models for usage on time series forecasting problems, as well as for usage of CMAES for hyperparameters optimization.
Keywords: Deep Learning. Evolutionary Algorithm. Time series forecasting.