Facilitating and Enhancing the Performance of Model Selection for Energy Time Series Forecasting in Cluster Computing Environments

Applying Machine Learning (ML) manually to a given problem setting is a tedious and time-consuming process which brings many challenges with it, especially in the context of Big Data. In such a context, gaining insightful information, finding patterns, and extracting knowledge from large datasets are quite complex tasks. Additionally, the configurations of the underlying Big Data infrastructure introduce more complexity for configuring and running ML tasks. With the growing interest in ML the last few years, particularly people without extensive ML expertise have a high demand for frameworks assisting people in applying the right ML algorithm to their problem setting. This is especially true in the field of smart energy system applications where more and more ML algorithms are used e.g. for time series forecasting. Generally, two groups of non-expert users are distinguished to perform energy time series forecasting. The first one includes the users who are familiar with statistics and ML but are not able to write the necessary programming code for training and evaluating ML models using the well-known trial-and-error approach. Such an approach is time consuming and wastes resources for constructing multiple models. ... mehrThe second group is even more inexperienced in programming and not knowledgeable in statistics and ML but wants to apply given ML solutions to their problem settings.

The goal of this thesis is to scientifically explore, in the context of more concrete use cases in the energy domain, how such non-expert users can be optimally supported in creating and performing ML tasks in practice on cluster computing environments. To support the first group of non-expert users, an easy-to-use modular extendable microservice-based ML solution for instrumenting and evaluating ML algorithms on top of a Big Data technology stack is conceptualized and evaluated. Our proposed solution facilitates applying trial-and-error approach by hiding the low level complexities from the users and introduces the best conditions to efficiently perform ML tasks in cluster computing environments.

To support the second group of non-expert users, the first solution is extended to realize meta learning approaches for automated model selection. We evaluate how meta learning technology can be efficiently applied to the problem space of data analytics for smart energy systems to assist energy system experts which are not data analytics experts in applying the right ML algorithms to their data analytics problems. To enhance the predictive performance of meta learning, an efficient characterization of energy time series datasets is required. To this end, Descriptive Statistics Time based Meta Features (DSTMF), a new kind of meta features, is designed to accurately capture the deep characteristics of energy time series datasets. We find that DSTMF outperforms the other state-of-the-art meta feature sets introduced in the literature to characterize energy time series datasets in terms of the accuracy of meta learning models and the time needed to extract them. Further enhancement in the predictive performance of the meta learning classification model is achieved by training the meta learner on new efficient meta examples. To this end, we proposed two new approaches to generate new energy time series datasets to be used as training meta examples by the meta learner depending on the type of time series dataset (i.e. generation or energy consumption time series). We find that extending the original training sets with new meta examples generated by our approaches outperformed the case in which the original is extended by new simulated energy time series datasets.