Autoadaptives prädiktives Modell zur Quantifizierung von Gleichzeitigkeitsfeffekten in Lastverteilungen urbaner Energiesysteme

• Auto-adaptive predictive model for quantifying simultaneity effects within load distributions of urban energy systems

Koschwitz, Daniel; van Treeck, Christoph Alban (Thesis advisor); Nytsch-Geusen, Christoph (Thesis advisor)

Aachen (2020)
Dissertation / PhD Thesis

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2020

Abstract

The present work describes an auto-adaptive predictive model for quantifying simultaneity effects within load distributions of urban energy systems named AMSA (Auto-adaptive Model for Simultaneity Analysis). The mathematical structure of AMSA is based on Machine Learning techniques in combination with methods related to energy data analysis enabling the integration in intelligent interconnected energy information systems. AMSA is implemented in Matlab using a modular design including different functional components. Thus, individual modifications of interface functions and modular extensions do not influence the numerical calculation core. Previous simultaneity factor quantification for economically optimized dimensioning of energy generation plants and networks is based on experience values and empirical analysis focusing on historical measurement data. Conventional derivations of characteristic curves using measurement databases imply a decreasing simultaneity factor in combination with increasing building group sizes, which isused for replanning energy supply systems. In order to optimize the latter, it is necessary to know historical and current as well as scenario based future energy demand of single buildings and building groups taking uncertainty into account. Moreover, similarity analysis concerning building load profiles is required to recognise simultaneously recurring patterns in order to identify energy network systems. In this context, the developed model serves for knowledge gain from varying complex databases to identify decision ranges for long- and medium-term planning regarding the development of districts and urban regions. Furthermore, on operational level, it improves short-term load management strategies concerning energy supply. The derivation of suitable methods for component model development is based on appropriate scientific literature research and analysis regarding method categories with their specific characteristics as well as findings from application-based studies. An ensemble model serves to calculate future load conditions, consisting oft wore current neural networks of various depth and two support vector machine configurations using different kernels. According to input data characteristics, a building-specific suitable prediction method is automatically chosen based on an error evaluation during the method testing period. In order to analyse influences of varying weather conditions and building retrofits on simultaneity effects within urban load distributions, mathematical-statistical methods for load profile modifications are introduced. For pattern recognition of load distributions, self organizing maps combined with learning vector quantification are used, which is assigned to competitive neural networks category. In this regard, initially, buildings are automatically grouped according to their characteristic load profile. Secondly, the building group is identified, which provides the most correlated load profile compared to the district load profile. Subsequently, simultaneity factors and peak load shares of buildings an building groups are calculated. In the final part of this work, monitoring data of building-specific heating and cooling consumption of a research campus serve for model demonstration. In this context, besides statistical analysis of the database, results of preliminary studies concerning component models are presented. These include clustering of thermal load profiles, a detailed comparison concerning prediction models used as well as long-term heating load predictions based on retrofit scenarios. Furthermore, conventionally derived curves for simultaneity factors enable classification and evaluation of the results using AMSA. Taking the corporate model level into consideration, various model-related advantages may be shown: Regarding short-term load management, less generating capacities have to be provided. Regarding similarity-based grouping, key buildings and the building group may be identified, where temporal peak load shifting contains the largest share of the expected peak load on district level. This information enables predictive loading and unloading strategies for storage systems in terms of demand side management. With respect to longer-term planning periods, it is shown that simultaneity factors combined with similarity-based building classification suit for the identification of possible energy network systems. Future weather and retrofit scenario analysis demonstrates external influences on the results of simultaneity analysis, which emphasizes the demand for an auto-adaptive predictive model. The scope of future research work should focus on the validation of AMSA taking detailed information bases into account