Efficient Management of the Electric Energy Consumption by Means of the Classification, Prediction and Clustering of Time Series

Objective:

In this project we plan to develop computationally efficient transversal methodologies and algorithms that allow to cluster, classify, and predict complex data such as the electric energy consumption. The problems approached in this project are complementary and they will allow to improve the efficiency in the management of electric energy. Clustering and classification of time series will be approached using elastic dissimilarity measures, which have been selected due to their flexibility for dealing with time axis distortions. In order to deal with both problems, we will develop dissimilarity elastic measures between time series adapted for online scenarios. Using these measures we will analyze the time series in the dissimilarity space, which is given in terms of a pairwise dissimilarity matrix, and we will cluster the time series into profiles using efficient variants of the Lloyd algorithm. Besides, we will develop efficient learning algorithms of probabilistic graphical models for dealing with high-dimensional domains. These algorithms will be used for approaching the supervised classification problem and the early classification of time series. Finally, we will develop efficient techniques for learning probability distributions by minimizing general divergences. The use of divergences adapted to the underlying decision problem will produce probability distributions that allow to take near optimal decisions. In addition, we will develop sequential and distributed techniques to ensure an efficient learning by means of the continuous update of distributions. Such techniques are especially important for dealing with time series and data with complex interrelations. The methodological tools and the algorithms proposed will be used in order to develop efficient techniques for energy management, including the analysis of electric energy consumers, the early classification of new consumers in consumption profile classes, and the probabilistic prediction of the energy (net) load. We will start by representing the electric consumption time series in the dissimilarity space and by clustering them in electric consumption profiles. Next, we will select a subset of prototype time series by means of multiobjective feature selection techniques for characterizing the different profiles. Then, we will construct classifiers based on probabilistic graphical models for performing early classification of time series in the profiles previously identified. Finally, we will develop probabilistic techniques for the prediction of electric consumption adapted to energy management. These techniques will improve the energy management by means of the use of divergences adapted to the specific repercussion of the errors in the energy prediction.