Nicole Ludwig

Putting Big Data analytics to work: Feature selection for forecasting electricity prices using the LASSO and random forests

Successful companies are increasingly those companies that excel in the task of extracting knowledge from data. Tapping the source of ‘Big Data’ requires powerful algorithms combined with a strong understanding of the data used. One of the key challenges in predictive analytics is the identification of relevant factors that may explain the variables of interest. In this paper, we present a case study in predictive analytics in which we focus on the selection of relevant exogenous variables. More specifically, we attempt to predict the German electricity spot prices with reference to historical prices and a deep set of weather variables. In order to choose the relevant weather stations, we use the least absolute shrinkage selection operation (LASSO) and random forests to implicitly execute a variable selection. Overall, in our case study of German weather data, we manage to improve forecasting accuracy by up to 16.9% in terms of mean average error.

A comprehensive modelling framework for demand side flexibility in smart grids

The increasing share of renewable energy generation in the electricity system comes with significant challenges, such as the volatility of renewable energy sources. To tackle those challenges, demand side management is a frequently mentioned remedy. However, measures of demand side management need a high level of flexibility to be successful. Although extensive research exists that describes, models and optimises various processes with flexible electrical demands, there is no unified notation. Additionally, most descriptions are very process-specific and cannot be generalised. In this paper, we develop a comprehensive modelling framework to mathematically describe demand side flexibility in smart grids while integrating a majority of constraints from different existing models. We provide a universally applicable modelling framework for demand side flexibility and evaluate its practicality by looking at how well Mixed-Integer Linear Program solvers are able to optimise the resulting models, if applied to artificially generated instances. From the evaluation, we derive that our model improves the performance of previous models while integrating additional flexibility characteristics.

Towards coding strategies for forecasting-based scheduling in smart grids and the energy lab 2.0

Development of the power supply system towards a more decentralized system with a growing share of renewable energies constitutes an additional complexity for its reliable, secure, and economic operation. This has a strong impact on a variety of optimization tasks, such as power plant resource scheduling, reactive power management, or the expansion of the system by additional transmission lines, power generators or storage systems. In particular, scheduling and expansion planning depend strongly on a reliable forecast of expected demands and electricity production, the latter being a demanding task for volatile sources, such as wind power plants or photovoltaic power generators (PV). For testing new approaches and strategies, the Karlsruhe Institute of Technology (KIT) develops a test bed comprising different energy grids called Energy Lab 2.0. This test bed will allow studying the effects of new tools, forecasting and scheduling techniques, and other algorithms aimed at managing a smart grid. The lab and applied forecasting techniques will be briefly introduced in the present contribution. First ideas about metaheuristic scheduling of different energy sources based on production and demand forecasts with the aim of ensuring a reliable and economic energy supply are introduced. Appropriate representations for Evolutionary Algorithms (EAs) are discussed and some experience from earlier scheduling projects for fast scheduling of many jobs to heterogeneous resources are given.

How much demand side flexibility do we need?: Analyzing where to exploit flexibility in industrial processes

We introduce a novel approach to demand side management: Instead of using flexibility that needs to be defined by a domain expert, we identify a small subset of processes of e. g. an industrial plant that would yield the largest benefit if they were time-shiftable. To find these processes we propose, implement and evaluate a framework that takes power usage time series of industrial processes as input and recommends which processes should be made flexible to optimize for several objectives as output. The technique combines and modifies a motif discovery algorithm with a scheduling algorithm based on mixed-integer programming. We show that even with small amounts of newly introduced flexibility, significant improvements can be achieved, and that the proposed algorithms are feasible for realistically sized instances. We thoroughly evaluate our approach based on real-world power demand data from a small electronics factory.

Demand Response clustering: Automatically finding optimal cluster hyper-parameter values

Time series clustering methods, such as Fuzzy C-Means (FCM) noise clustering, can be efficiently used to obtain typical price-influenced load profiles (TPILPs) through the data-driven analysis and modelling of the consumption behaviour of household electricity customers in response to price signals (Demand Response, DR). However, the analysis of load time series with cluster methods presupposes that the user has a lot of experience in selecting good cluster hyper-parameter values (e.g. number of clusters or fuzzifier). The present contribution proposes a practical method to the automatic selection of optimal hyper-parameter values for DR clustering.

Assessment of Unsupervised Standard Pattern Recognition Methods for Industrial Energy Time Series

Finding and extracting standard patterns in energy time series is very important to many real-world applications. Hence, there exists a multitude of pattern recognition algorithms with a majority of them being supervised ones. The advantage of supervision is that it can easily be checked if the algorithm is performing well or not. However, if no labels are available, an unsupervised pattern search is necessary. This search is faced with the challenge of how to measure success. Thus the question arises, when is a found pattern -- for example a motif or a mean cluster curve -- really describing the standard behaviour of a process and not just some kind of irrelevant behaviour? The present paper introduces a new method to assess two methods -- namely clustering and motif discovery -- in their quest to find standard profiles in energy time series data from industrial processes. Although both methods share the same aim, the results are incongruent. This finding has profound implications for real-world applications.

SCiBER: A new public data set of municipal building consumption

Data about the energy consumption of buildings contains valuable information which is essential for the future energy system and smart cities. However, only few researchers publish the data on which their methods and analysis is based. This lack of publicly available data sets, makes it difficult to compare strategies and results, and hinders a stronger development of the research field. Thus, this paper describes a data set of municipal energy consumption data, which is published with the objective to facilitate the comparability of research methods in the field.

Dataset accompanying "How much demand side flexibility do we need? - Analyzing where to exploit flexibility in industrial processes"

We introduce a novel approach to demand side management: Instead of using flexibility that needs to be defined by a domain expert, we identify a small subset of processes of e. g. an industrial plant that would yield the largest benefit if they were time-shiftable. To find these processes we propose, implement and evaluate a framework that takes power usage time series of industrial processes as input and recommends which processes should be made flexible to optimize for several objectives as output. The technique combines and modifies a motif discovery algorithm with a scheduling algorithm based on mixed-integer programming. We show that even with small amounts of newly introduced flexibility, significant improvements can be achieved, and that the proposed algorithms are feasible for realistically sized instances. We thoroughly evaluate our approach based on real-world power demand data from a small electronics factory.

Concept and benchmark results for Big Data energy forecasting based on Apache Spark

The present article describes a concept for the creation and application of energy forecasting models in a distributed environment. Additionally, a benchmark comparing the time required for the training and application of data-driven forecasting models on a single computer and a computing cluster is presented. This comparison is based on a simulated dataset and both R and Apache Spark are used. Furthermore, the obtained results show certain points in which the utilization of distributed computing based on Spark may be advantageous.