The research at IES focuses on Autonomic and Organic Computing (AOC), Technical Data Analytics (TDA), and Intelligent Technical Systems (ITS) with key application areas in energy and transportation.
Autonomic und Organic Computing
The Collaborative Interactive Learning (CIL) research group focuses on research topics related to the interaction between humans and machine learning (ML) models. Accordingly, CIL is related to the research areas Human-in-the-Loop Learning, Human-centred Learning, and Interactive Machine Learning. In particular, the goal is to lay the foundations for novel autonomous systems that can operate in dynamic, unconstrained real-world environments and are effective, understandable, and enjoyable in their interaction and collaboration with humans. CIL is also investigating how the learning process of digital systems can be enhanced through seamless, unobtrusive interaction with humans and vice versa: how autonomous interactive systems can enhance collaborative human learning. Toward the above vision, CIL explores collaborative sensing and perception, algorithms and architectures, and interaction models at both the individual and societal levels. CIL strives to create systems that provide meaningful and desirable support to humans without taking away their autonomy or even dominating and ultimately replacing them.
Our research focuses on Deep Learning (DL) techniques in combination with the following areas:
- Active Learning (AL): AL focuses on interacting with human annotators to collect an annotated dataset (e.g., images with class information) for training DL models. Since this collection is costly, AL tries to select useful data for annotation by intelligent selection.
- Uncertainty Modeling: Despite the success of DL models in many domains, their ability to model uncertainty is still severely limited. However, in areas such as AL and applications such as autonomous driving, it is critical to obtain a prediction that reflects different types of uncertainties. An appropriate example of uncertainty estimation for a two-dimensional dataset is shown in the figure below.
- Noisy Labels: Standard DL models require correctly annotated data for their successful use. However, in many applications, the annotations are often provided by multiple error-prone humans, e.g., crowd workers. Special techniques are needed to learn about the resulting erroneous annotations (so-called noisy labels).
- Explainable Artificial Intelligence (xAI): Since DL models often provide predictions that are incomprehensible to humans, the use of xAI methods to increase transparency is of utmost importance. This allows humans to better and intuitively monitor the decision-making processes of a DL model.
By combining techniques from these areas, we aim to realize CIL systems. A simple example of such a system is shown in the video below, where a robot learns a sorting task by interacting with a human.
Video: Automated Active Learning For Training a Sorting Robot
The research project Intelligent Financial Assistant (INFINA) is a joint project between the Department of Intelligent Embedded Systems (IES) at the University of Kassel and fino create GmbH, based in the Kassel Science Park. This project is being supported by Wirtschafts- und Infrastrukturbank Hessen as part of the Operational Program for the Promotion of Investments in Growth and Employment in Hessen from funds of the European Regional Development Fund (ERDF) 2014 to 2020 (IWB-EFRE program Hessen).
The object of this cooperation project is research on an automated and intelligent financial assistant based on machine learning. Thereby, the financial situation of a user shall be analyzed and possible improvement potentials identified. The technical goal of this financial assistant is to fully support users in their individual financial situation and to provide tools for making financial decisions. The sub-project of fino concentrates on technical development and on market observation. The sub-project of IES focuses on the machine learning fundamentals of the system. In this context, novel methods for the representation and comparison of time series based on discrete events will be explored as well as methods for the transfer of knowledge and the aggregation of similar user groups. Thus, the focus of the research is to test and validate different machine learning techniques in order to model a prototype. As such, it is industrial research involving aspects of basic research in cooperation between industry and university. In summary, the working areas of the IES are classification and representation of time series (using active learning), similarity assessment for time series, regression and user group comparison
The ZIM project "Entwicklung eines verteilten, zweistufigen Fraud-Detection System basierend auf künstlicher Intelligenz sowie Implementierung dieses Systems auf ARM-basierenden Hybridservern; Entwicklung einer Selektionsmethodik zur Detektion von Fraudmustern auf Basis von Stream Based Active Learning mit Konfidenz > 95% und Entwicklung einer verteilten Serverstruktur (Hybridcloud und Hybridservern) auf ARM-Basis für Echtzeitanalyse" investigates the problem of fraud detection. Machine Learning and Stream Based Active Learning are used to detect fraudulent behavior targetting telephone provider, telephone exchange and their customers. The recorded data is handled as a data stream to be able to do real time analysis using the call detail record. Stream Based Active Learning is used to adapt known fraud patterns to the fraudsters' changing behavior as well as finding new patterns. The learned classifier has to be interpretable, which means that the decisions has to be understandable, to be able to explain the decisions in case of a dispute.
The Hessen Agentur Project "VitaB - Classification of the vital Parameters for the individual vital and cognitive Determination of the state of the human being" addresses the Problem of driver state classification.
Vital Parameters, that are computed from heart data from a large Group of subjects, form the Basis for all analytical Tasks in this Project.
The goal is to discriminate between, e.g., vigilance, arousal or fatigue.
Overall, new methods in the fields of preprocessing, Feature selection, machine learning and optimization are applied to the Problem.
Eventually, a whole Demonstrator becomes available, that allows for a online driver state calssification.
This project (HA project no. 545/17-27) is financed with funds of LOEWE – Landes-Offensive zur Entwicklung Wissenschaftlich-ökonomischer Exzellenz, Förderlinie 3: KMU-Verbundvorhaben (State Offensive for the Development of Scientific and Economic Excellence).
Organische Computertechniken zur Laufzeitselbstadaption von multi-modalen Aktivitäts-Erkennungssystemen
Im DFG-Projekt "Organische Computertechniken zur Laufzeitselbstadaption von multi-modalen Aktivitäts-Erkennungssystemen" werden Techniken zur autonomen Sensoradaption entwickelt.
Angesiedelt im Feld der Aktivitätserkennung, sind Multi-Sensor-Systeme nach wie vor ein wichtiges Forschungsschema.
In diesem Bereich werden Methoden erforscht, mit denen neue Sensorquellen zur Laufzeit zu einem bestehenden System hinzugefügt werden können.
Derartige Systeme zeichnen sich zum einen durch eine höhere Robustheit aus, da Sensorausfälle kompensiert werden können (Selbstheilung).
Zum anderen kann die Leistungsfähigkeit solcher Systeme gesteigert werden, ohne dass diese zur Designzeit bekannt waren.
Intelligent Technical Systems
The KITE joint project aims to develop highly efficient motor topologies for electric drive machines. Artificial intelligence (AI)-based models will help to map the ample solution space of possible topologies and propose new highly efficient topologies. On the one hand, the project aims to substantially accelerate the development process by efficiently evaluating many possible topologies concerning various quality criteria in an early development phase. On the other hand, the design is to be supported by AI methods by specifically finding new, potentially also previously unknown optimal topology approaches for downstream, FEM-based investigations and optimisations. An optimised machine design will be prototyped to validate the procedure and prove the methodology's usefulness. The increasing market inclusion of electromobility and the growing number of units demand higher material requirements for central components in electrical machines, such as magnets. The energy efficiency of electrical machines is becoming critical both economically and ecologically, and the ability to optimise them is proving to be highly relevant to competition. Therefore, considerable efforts are being made in industry and research to optimise the topology of rotors, one of the central elements of electrical machines in terms of relevant criteria - e.g. size, shape, type and arrangement of magnets and laminations. At the same time, performance specifications (torques, speeds) and thermal, mechanical and electrical boundary conditions must be met. Due to their underlying physics, the solution space of potential topologies is enormous. Therefore, current design processes are mostly heuristic and rely heavily on expert knowledge and physical approximations. For further optimisation, physically based methods (e.g. SNOPT: Sparse Nonlinear OPTimizer) exist on the level of FEM mesh points and methods of parameter variation in combination with FEM or parameter-based models. The former theoretically covers the entire solution space (i.e. the set of all possible topologies), while the latter already restricts this space by choice of parameters. Parameter-based models are only defined for part of the solution space and show a significantly lower accuracy than FEM simulation.
FEM-based methods are also very computationally intensive, so local optima are primarily determined based on the best possible initial topologies with finite computing time. Ultimately, known topology approaches are optimised, but no new solutions are found. This problem of finding an optimal topology over the entire solution space in finite time, with high accuracy and with considerable but limited computer capacities, is addressed by the KITE project. Today, the initial topology is mainly identified from empirical values and heuristic approaches. The project's goal is an AI-based exploration of the solution space by generating promising initial topologies. In contrast, fine-tuning (exploitation of the solution space) should continue to be carried out with the existing optimisation methods.
Funding: Federal Ministry of Education and Research - BMBF
Duration: 01.10.2021 until 30.09.2023
- University of Kassel - Intelligent Embedded Systems IES
- Georg-August University Göttingen - Cellular Neurobiology
- Green Excellence GmbH
- Avacon Netz GmbH (associated partner)
- Stromnetz Hamburg GmbH (associated partner)
The transformation of the German energy system has increased the need for high-resolution consumption time series. AI enables efficient operations management and forecasting solutions and business models based on them for simplified participation of active consumers or prosumers in the integrated energy transition. Large amounts of electricity consumption data are required to train such models. Electricity consumption in households, commerce, and agriculture depends on user behavior and is undergoing significant changes. Measured load time series are subject to the GDPR, as they have a person- or company-specific behavioral characteristics.
The project aims to generate synthetic load time series for different types of consumers, which are indistinguishable from real measured data in terms of their characteristics, but at the same time ensure the greatest possible anonymization. For this purpose, different GAN methods are investigated, and neural networks are adapted and applied to generate energy time series. Self-classification methods will also be used to capture influencing variables in real measurement data and use them as a basis for GAN generation to represent the behavior of different consumers accurately. The developed methods and an example data set will be available to industry and science free of charge.
The University of Kassel is involved with five other partners from science and industry in the joint project Digital-Twin-Solar. The project will receive funding from the Federal Ministry for Economic Affairs and Energy (BMWi) from May 2020 for three years.
The future network-connected components of the energy system allow extensive data acquisition to generate the digital twins of plants and power systems in digitalization. That means these plants and power systems should be accessible for optimization using machine learning (ML). The Digital Twin Solar project deals with solutions specially tailored to the use of solar energy and electricity storage systems. The overarching goal of the sub-project for the University of Kassel / Department IES is to develop the potential of the latest ML and artificial intelligence (AI) to develop digital twins' components in anomaly detection for PV and battery inverters. And at the same time to predict the time of the anomaly.
This research includes probabilistic forecasting, transfer learning, active learning, explainable AI, generative adversarial networks, and autoencoders. The project's goal is to develop further and adapt these ML methods and algorithms to unlock their potential for application in the renewable energy sector.
The University of Kassel is involved in the joint project "AI Data Tooling" along with 17 other partners from science and industry. The Federal Ministry of Economics and Energy (Bundesministerium für Wirtschaft und Energie - BMWi) provides a total of almost €850,000 in funding for the project over three years starting from April 2020. The joint project is part of the artificial intelligence project family of the initiative "Autonomes und Vernetztes Fahren" of the German automotive industry (VDA).
Artificial intelligence (AI) and especially machine learning are the key technologies of autonomous driving. Powerful computers and algorithms learn recognition and patterns, such as the automatic detection of traffic signs, other vehicles, or pedestrians in images, radar, or laser data.
The image depicts an urban traffic situation, in which an automated vehicle approaches an intersection. The automated vehicle perceives its environment using its built-in camera. The AI seeks to recognize all road users in the camera image. Since the weather is rainy and visibility is poor, the oncoming vehicle is not detected. However, to safeguard the automated driving functions, it is necessary that an AI also reliably operates in such difficult and usually critical situations. Therefore, within the AI Data Tooling project, we aim to develop tools that can automatically identify cases that are not well covered by the AI. We add theses automatically detected cases to enlarge the dataset and to improve the AI. In this respect, a large amount of sample data is required to "train" an AI and ensure its functionality. Such an automatically acquired database can shorten development cycles and test phases considerably. In the project "AI Data Tooling," we develop methods and tools for efficiently building a database for the perception of automated driving vehicles. These new methods, for AI, are also based on AI. We test these new methods for highly-automated and intelligent data acquisition on the case study involving the detection of vulnerable road users, e.g., pedestrians.
The researchers of the Department of Intelligent Embedded Systems (Prof. Dr. Bernhard Sick) are leading the sub-project "Quality Requirements and Efficiency Potentials of Data Generation and Provision" in the joint project "AI Data Tooling" together with BMW. They are also mainly responsible for the development of methods to detect corner cases. Corner cases are rare but often critical situations in road traffic. Besides, the University of Kassel is also concerned with the highly automated annotation and labeling of data for AI using active learning methods.situations in road traffic. Besides, the University of Kassel is leading the way in automating data refinement (annotation with additional information) using active learning methods.
The DeCoInt2 project (Detecting Intention of Vulnerable Road Users based on Collective Intelligence as a Basis for Aumated Driving) is concerned with intention detection of vulnerable road users (VRUs) in automated driving using cooperative technologies. Individual mobility will still be an important aspect of future traffic and automated driving will make a key contribution. It has the potential to increase safety as well as traffic flow and to decrease environmental pollution and resource consumption. Especially in urban areas, VRUs, e.g., pedestrians and cyclists, will still play an important role in the mixed traffic of tomorrow. For an accident-free and highly efficient traffic flow with automated vehicles it is not just important to perceive VRUs but it is also essential that their intentions are detected and analyzed in a similar way as humans do it when they drive and forecast VRU trajectories. The reliable and robust perception of VRUs and their intentions with a multi-modal sensor system (e.g., video cameras, laser scanners, accelerometers and gyroscopes in mobile devices) in real-time is a big challenge. Going far beyond existing work we follow a holistic, cooperative approach to forecast movements of humans (e.g., when will a standing cyclist start to move forward) and to forecast their trajectories (e.g., will she turn left). Heterogeneous, open sets of agents (collaboratively interacting vehicles, infrastructure, and VRUs themselves, if equipped with common mobile devices) exchange information to determine individual models of their surrounding environment which allow for an accurate forecast of VRU basic movements and trajectories. Occlusions, implausibility, and inconsistencies are resolved using the collective intelligence of cooperating agents. We develop new methods by considering and combining novel signal processing and modeling techniques with machine learning based pattern recognition approaches. The cooperation of agents will be investigated on several levels including the VRU perception level, on the level of recognized trajectories, or on the level of already detected intentions. A communication strategy to exchange required information in ad hoc networks of cooperating partners will be proposed. The techniques are evaluated with real data using a research vehicle, a research intersection with public traffic, and a number of mobile devices.
The University of Kassel is involved in the joint project SALM with FLAVIA IT-Management GmbH and other associated partners from science and industry. The project will receive funding from the Federal Ministry of Education and Research (BMBF) from January 2021 for two and a half years.
Due to the current ramp-up of electromobility, the focus is not only on the actual provision of electrical energy and the further development of electric vehicles, but also on the expansion of distribution grids. The volatile load behavior of e-mobility places additional demands on a decentralized energy supply as part of the current efforts in the context of the energy transition. Simultaneous charging of all connected e-vehicles with the nominal power leads to an overload of the infrastructure due to the occurring peak loads and appears unnecessary considering the long standing times of the vehicles. For this reason, work is being done on optimizing energy distribution and regulating the load profile by deliberately shifting and controlling the charging processes. The goal is to provide a high charging density under existing network bottlenecks and taking into account fluctuating energy generation and the preferences of the users. With the help of various machine learning and artificial intelligence methods, an automatism is to be developed that can react self-adaptively to the dynamic development of e-mobility (e.g. operational status of the infrastructure, vehicle or battery technology, changes in user behavior, etc.). Furthermore, comprehensible/explainable decisions should increase the acceptance of a system by all stakeholders.
This includes, among other things, research into data-driven modelling of various components and forecasting methods. The charging behavior of the infrastructure is to be mapped and simulated on the basis of data using probabilistic and/or generative methods in order to construct a simulation that is as close to reality as possible. Based on such a simulation, reinforcement learning methods can be used to find the best possible charging strategy. From the point of view of acceptance, methods of so-called Explainable AI should be used to make the decisions of the system explainable for human actors. Furthermore, transfer learning methods can be used to abstract model knowledge from one application and transfer it to another application.
With their high pulse energies and short pulse durations free-electron lasers have revolutionized the field of X-ray experiments for over 15 years. The generation of ultrashort X-ray pulses is based on the principle of self-amplified spontaneous emission. However, the inherently random X-ray pulse shapes generated in this process impose experimental limitations on the studies in terms of time resolution and for intensity-dependent measurements. As part of the previous BMBF project SpeAR_XFEL, the project partners built a new spectrometer with angular resolution that can be used to characterize individual X-ray pulses at free-electron lasers with attosecond accuracy. In the BMBF-project TRANSALP, this angular streaking technique will be further developed from free-electron pulse characterization to a versatile instrument for ultrafast X-ray research.
A broad scientific community, ranging from materials scientists to biologists and physicians, has now discovered free-electron lasers as a research tool. Most of the interesting questions relate to biologically relevant processes such as ultrafast reaction pathways of photochemical processes in organic molecules, the role of ionization dynamics in DNA damage, or the electronic excitation of structural changes in bioproteins. TRANSALP will combine the apperature developed in SpeAR_XFEL with a LiquidJet to track ultrafast photoinduced processes on biological samples in the liquid phase. In combination with a dedicated laser system specifically tuned to the needs for angular streaking and newly developed machine learning algorithms for online analysis and error correction, the prototype of a unique instrument for time and angle resolved in-situ electron spectroscopy on organic molecules in the liquid phase will be developed at the Free-electron LASer in Hamburg (FLASH).
In the context of this projekt, the tasks of the department Intelligent Embedded Systems are:
- Further development and adaptation of existing online pulse characterization algorithms to the new data.
- Develop and evaluate new machine learning techniques for handling the incoming big data at FLASH.
- Design and implementation of new machine learning algorithms for improved scientific experimentation.
Cooperation partners in the project are:
- TU Dortmund, Zentrum für Synchrotronstrahlung (DELTA)
- Universität Kassel, Institut für Physik und CINSaT
- Deutsches Elektronen-Synchrotron DESY
- European X-Ray Free-Electron Laser Facility GmbH
Graphs in Artificial intelligence and Neural Networks
The junior research group 'GAIN - Graphs in Artificial intelligence and Neural Networks' works on Graph Neural Networks (GNNs), especially their dynamics and explainability. GNNs are one of the youngest and fastest growing fields of Machine Learning, aiming to continue the success of Deep Learning to data represented by graphs.
Currently, our focus is on the dynamics and explainability of Graph Neural Networks (GNNs). By dynamic, we mean to develop GNN-algorithms that can deal with changes in the topology of a graph or changing graph attributes. In the simplest case that might be new appearing nodes or edges or a change in an attribute describing a node.At the same time, we want our algorithms to be explainable, which means that either inherently or post-hoc our algorithms should give a reason for their predictions.
We chose this focus, because although we engage in basic algorithmic development of models, we strive to develop our models with concrete use cases in mind, e.g.in renewable energies. Our project partner is therefore the Fraunhofer Institute for Energy Economics and Energy Systems Technology (IEE). Making GNNs more dynamic will enable the use of GNN-based methods in supply structure networks and therefore help strengthen the use of renewable energies. Furthermore, explainability will increase their usability and applicability.
In almost all areas of the energy transition, the exchange and the use of information play an eminently important role in improving existing processes and establishing new processes within the changing energy landscape. The steadily advancing expansion of renewable energies, the accompanying measures and business models for making generation and consumption more flexible, and the growing number of components and stakeholders within the energy system are leading to an enormous increase in complexity and computing capacity with regard to the modeling and calculation of system states. For established processes and system components, there is usually sufficient information available to model them appropriately. Artificial intelligence (AI) techniques have been proven successful for modeling these problems as long as an adequate amount of information is available through historical measurements. The new challenge of this project is to flexibly apply the idea of machine knowledge transfer of similar systems and system states to the operation of new or changing technical infrastructures within the power system to reliably generate robust state determinations and forecasts for new and changing system components with reduced computational efforts. Dealing with missing and changing databases will be a focus here. The focus will be on creating general-purpose transfer learning models for automated knowledge transfer between individual system components. The models will recognize and use pre-estimated similarities between system components, verified during operation, to generate reliable solutions, e.g., for forecasts. The developed models will be tested with partners from science and industry based on real-world use cases and data sets from the energy domain and demonstrated within application-oriented prototypes.
Modern test benches for experimental validation provide extremely large, mostly time-based and usually heterogeneous amounts of data. The processing and, the evaluation using conventional knowledge-based methods is not possible at all, due to the size and complexity of the acquired data. For the use of AI methods, e.g. from the field of Deep Learning, this application area offers an extraordinarily interesting field of investigation still little researched and critical for the development of electrical vehicles.
At the University of Kassel, a novel, high-performance test bench for electric drive machines has recently been in use, which uses extensive measurement technology to generate very large heterogeneous data volumes of both time-based electromagnetic, electrical, acoustic, mechanical and thermal variables, as well as non-time-based, characterizing or parametric data, and also allows measurements in previously little researched borderline areas.
The BMBF joint project AIMEE addresses the processing and evaluation of heterogeneous, high-volume data sets from the test bench in order to make the applicability of innovative methods directly experienceable and interpretable for students by means of practical examples in the AI laboratory. This creates the necessary prerequisite for learners (undergraduates, doctoral students, and professionals in continuing education programs) to study and apply the various AI methods with extensive and defined data sets using practical examples. In addition, the highly interesting and rare opportunity for teaching and research opens up to create new data sets at will and to adapt the boundary conditions of data creation, i.e. measurement runs and structure, to the requirements of AI methods and method development. The close cooperation of software, hardware and application result in an excellent basis for teaching, are of great interest for the regional economy and have a high scientific connectivity.
X-ray free-electron-lasers generate laser pulses based on the principle of self-amplification of spontaneous emission. This allows for continuous tuning of the X-ray radiation wavelength and leads to the emission of a train of ultrashort, more precisely a few hundred femtoseconds long, X-ray spikes. However, the principle of self- amplification of spontaneous emission has some drawbacks, which have to be overcome to study well-defined measurements relying on precise knowledge of the activation signal. Only if parameters such as stochastically varying shot intensities, separations, durations as well as photon energies are known, the potential of ultrafast and nonlinear measurements can be fully exploited. To achieve knowledge about the underlying parameters, temporal and spectral characteristics of X-ray pulses have to be examined utilizing real-time analysis techniques.
To receive the required characteristics, the main objective during the project SpeAR_XFEL is to design and construct a novel, angle-resolved electron multi-time-of-flight spectrometer with individually variable retardation voltage and high energy resolution for XUV and X-ray-triggered ultrafast measurement. For this purpose, suitable online analysis techniques based on machine learning methods have to be developed.
Tasks of Intelligent Embedded Systems regarding this project are:
- Development of suitable analysis techniques to cope with the load of measured data.
- Consideration of correlations within individual data sets of the angle-dependent detection and therefore distill the most general information from the experiment using, e.g., machine learning techniques.
- This should result in an implementation of an online X-ray characterization tool based on deep learning methods.
Cooperation partners during this project are:
- TU Dortmund, Zentrum für Synchrotronstrahlung (DELTA)
- Universität Kassel, Institut für Physik und CINSaT
- European X-Ray Free-Electron Laser Facility GmbH
- Helmholtz Zentrum Berlin, BESSY II
In the project OCTIKT, a generic framework and process model was developed to gather, improve and safeguard the resilience in dynamic, decentralised systems. The developments are exemplarily implemented and evaluated in a concrete application scenario, i.e. in a power distribution grid.
The IES lab is primarily involved in the development of OC-based technologies for the implementation of resilience. This includes fundamental research in the fields of novelty and anomaly detection in data streams. These techniques are the foundation to implement self-awareness capabilities in industrial (cyber-physical) systems and can be used to provide concrete implementation, e.g., for digital twins.
The project was funded by the German Federal Ministry of Education and Research (BMBF) in the framework of the funding programme "IKT 2020 — Forschung für Innovationen" (ICT 2020 — Research for Innovation). The project was executed by the following partners:
Fachgebiet IES, Universität Kassel
FZI Forschungszentrum Informatik
Fraunhofer-Institut für Kurzzeitdynamik, Ernst-Mach-Institut, EMI
evohaus IRQ GmbH
Netze BW GmbH
Seven2one Informationssysteme GmbH
The project was successfully completed in January 2022.
The Energy Industry as a Lead Sector for Artificial Intelligence
As part of the Clusters 4 Future project competition, the IC4CES project was funded as a finalist for 6 months. Within the funding period, a cluster strategy was developed and a full proposal was submitted with various partners from industry, which was presented to a jury in Berlin in July 2022. Unfortunately, the IC4CES cluster was not selected in the end. The goal of IC4CES should be to make the energy industry a leading sector for the application of artificial intelligence and to position the region of North Hesse/South Lower Saxony strongly in the long term.
The goal of the project Prophesy is the simulation of time series forecasts of the expected electricity from wind energy and photovoltaics in future power supply systems in the next hours to weeks. The methods developed will be linked to the scenario tool of the Fraunhofer IEE, with which expansion scenarios of the power supply system for the next decades can be simulated and analyzed. Since forecasts play an essential role in many planning and decision-making processes for network operators and market participants, many questions regarding future power supply systems can only be answered with high accuracy by considering forecasts. These include, among other things, control power requirements, the provision of system services, the necessary network expansion, the optimal energy mix for reducing balancing energy, necessary network measures, the optimal storage dimensioning or market design.
The goal of reducing greenhouse gas emissions in Germany through the steady expansion of renewable energies and the associated increase in complexity will place the electricity grids facing significant challenges in the future.
Many of the tasks are based on probabilistic problems. Often, these can be approximated by deterministic considerations, e.g., mean and worst case. In essential tasks, however, the probabilistic problem area as a whole must be examined for strong statements. Such an investigation of the entire problem space is typically done by a Monte Carlo simulation. However, this process is very time-consuming and resource-consuming, and already in today's tasks often simplifications must be made, which limit the resilience of the results and in particular the ability to extrapolate the results.
A typical application for such probabilistic tasks in energy systems engineering is, for example, network expansion planning. The further transformation of the distribution grids into smart grids with more volatile generators, decentralized storage, and intelligent, active equipment in the electrical supply network leads to increasing uncertainty both
* In spatial planning: where do new plants originate?
* In quantity: how many new plants will there be?
* And in the schedule: What will the feed-in and demand characteristics of the systems look like concerning temporal gradients, as well as maximum and minimum values in the future?
These and similar uncertainties must each be modeled by probability distributions, thus creating all potential scenarios for the development of renewable energy.
As a rule, many of these calculations will be redundant due to the same or very similar input data. Therefore, new, efficient probabilistic methods are needed to map the entire solution space for grid expansion planning. Therefore, in the project PrIME, methods for probabilistic tasks in energy system technology are to be considered and developed in a fundamentally oriented way. The method development should be based on typical probabilistic applications of energy system technology, to ensure high practical relevance for the results of basic research. Such methods then offer excellent application potential, both in network planning and in network operation management (for example, Day-Ahead Congestion Forecast, DACF).
As part of a consortium consisting of Fraunhofer IEE, the department e²n of the University of Kassel and several associated network operators, the Department of Intelligent Embedded Systems ensures the applicability and further development of the possible methods. The methods developed are used with different network calculation types such as load flow calculations, validated, evaluated and optimized.
The Federal Ministry of Economics and Technology (BMWi) is promoting innovative technologies and processes as well as the digitization of the energy industry with the support program "Schaufenster intelligent Energie - Digital Agenda for the Energiewende" (SINTEG) in order to improve the intelligent interaction of wind energy and PV Generation, networks, consumption and storage. The aim is to demonstrate the feasibility of a climate-friendly, safe and efficient power supply in five large model regions in Germany.
C/sells, one of the projects, focuses on the sun and includes demonstration modules from Baden-Württemberg, Bavaria and Hesse. In the C/sells consortium, around 50 partners from the fields of energy services, network operators, component manufacturers, science and knowledge transfer have come together to install and demonstrate decentralized energy systems in these three federal states in the years 2017-2020.
In order to do justice to the name of the project, on the one hand technical solutions with cellular structures ("cells") are to be developed, on the other hand, new energy market participation in the energy transition is planned.
Tasks and Goals
In Hessen, project partners work primarily on the conception and model implementation of a regional flexibility market. This should be developed as a prototype and ensure the system integration of the high-performance and fluctuating supply of regenerative energies at a decentralized level. On the supply side, network and system-relevant potentials of household, commercial and industrial customers are to be identified and then prepared for use in stabilizing the grids and the electricity system. Sector coupling plays an important role here. On the demand side, in the
The role of IES in the project:
The Intelligent Embedded Systems division is responsible for predicting the condition of the grid. The power supply state gives a statement as to whether a power grid is able to fulfill the task of power distribution. This is done according to a traffic light system, the individual traffic signal phases indicate the power supply state. Green means that everything is alright, Orange indicates a need for action and red indicates that the power grid is overloaded.
Typically, the determination of the network condition is made by means of load flow calculations, a procedure which uses all components in a power grid and calculates the utilization of these components. The IES department uses deep learning methods to predict the state of the network. In the process, methods are evaluated which directly predict the network status or else only predict the input variables for the load flow calculation.
The following Hessian companies and research institutions are involved in the project:
EAM GmbH & Co. KG (mit Tochterunternehmen)
Fraunhofer-Institut für Energiewirtschaft und Energiesystemtechnik IEE (vier Abteilungen: Strom-Wärme-Systeme, Energiemanagement und Energieeffizienz, Energiewirtschaft und Systemdesign, Betrieb Verteilungsnetze)
Ramboll CUBE GmbH
Städtische Werke Netz + Service GmbH
Universität Kassel (drei Fachgebiete: Volkswirtschaftslehre mit Schwerpunkt dezentrale Energiewirtschaft, Intelligente eingebettete Systeme und Kommunikationstechnik)
Durch das Bundesministerium für Wirtschaft und Energie