Leiprecht, Stefan
Refine
Document Type
- conference proceeding (article) (3)
- Article (2)
- Handout (1)
- Preprint (1)
- Report (1)
Language
- English (6)
- German (1)
- Multiple languages (1)
Publication reviewed
- begutachtet (6)
- nicht begutachtet (2)
Keywords
- Fernwärme (6)
- Lastprognose (6)
- Maschinelles Lernen (6)
- deepDHC (6)
- Machine Learning (2)
- Deep Learning (1)
- Dispatch optimisation (1)
- District Heating (1)
- KWK-Flex (1)
- Leitfaden (1)
Das Wissen um den zukünftigen Wärmebedarf gewinnt, bei komplexer werdenden Fernwärmesystemen mit volatilen und heterogenen Erzeugungsanlagen, immer weiter an Bedeutung. Denn zur Vermeidung von ineffizienten und unwirtschaftlichen Betriebszuständen muss für eine passende Einsatzoptimierung eine möglichst präzise und zuverlässige Wärmelastprognose vorliegen. Maschinelle Lernverfahren, steigende Datengrundlagen und Verfügbarkeit von ausreichender Rechenleistung bergen hierbei erhebliches Verbesserungspotential. Das von der Hochschule für angewandte Wissenschaften Kempten durchgeführte Forschungsvorhaben „DeepDHC - Untersuchung und Weiterentwicklung modernster maschineller Lernverfahren für die hochgenaue Lastprognose in Fernwärmenetzen“ (FKZ: 03EN3017) befasste sich mit der Performance unterschiedlicher maschineller Lernverfahren zur Wärmelastprognose in Fernwärmenetzen. Unter Berücksichtigung unterschiedlicher Fernwärmenetztopologien, der Einbindung von Smart Meter Daten und der automatisierten Berücksichtigung von Veränderungen im Fernwärmenetz wurden hierbei besonders relevante Fragestellungen aufgegriffen und bearbeitet.
Decarbonisation of heat generation has become a priority for district heating network operators. In order to avoid the use of fossil-fired boilers, operators need to know peaks in heat demand in advance. Accurate thermal load forecasting is playing an increasingly important role in this respect. This paper presents the final results of the research project “deepDHC” (deep learning for district heating and cooling) funded by the German Federal Ministry for Economic Affairs and Climate Action (BMWK). The three-year project focused on systematically benchmarking thermal load forecasts for district heating networks, based on state-of-the-art machine learning methods. The analysis covers a variety of machine learning techniques, such as neural networks – including latest deep learning methods – (e.g. LSTM, TFT, ESN, RC), decision trees (random forests, adaptive boosting, XGB) and statistical methods (SARIMAX). In addition, the impact of combining methods by so-called “stacking” was investigated. Training and validation of the machine learning algorithms was based on historical operating data from the district heating network for the city of Ulm in Germany, in combination with historical weather data, and weather forecasts. Thermal load forecasts – typically for three days ahead – are presented and compared against one another. An automatic tuning routine was developed as part of the project, which enables regular re-training of the machine learning algorithms based on the latest operating data from the heating network. Furthermore, a web interface for real-time forecasting was developed and implemented at the power station.
Zeitreihenprognosen können eine leistungsfähige Methode zur Vorhersage des Wärmebedarfs eines Fernwärmenetzes darstellen. Durch die Analyse historischer Daten über den Energieverbrauch und andere relevante Faktoren wie Wetterdaten kann ein Neuronales Netz Muster und Trends erkennen, die ihm helfen, genaue Vorhersagen über den künftigen Bedarf zu treffen. Auf diese Weise kann der Netzbetreiber die Ressourcennutzung der Anlage optimieren, Kosten einsparen und Emissionen reduzieren und sicherstellen, dass die Wärme effizient und kostengünstig an die Kunden geliefert wird. Darüber hinaus können genaue Prognosen dem Betreiber bei der Planung von Wartungs- und Reparaturarbeiten helfen und sicherstellen, dass das Netz auch in Zeiten hoher Nachfrage zuverlässig und funktionsfähig bleibt. Insgesamt kann die Zeitreihenprognose dazu beitragen, dass Fernwärmenetze effizienter arbeiten und Geld sparen, während sie den Kunden einen guten Servicebieten.
Eine mögliche Deep Learning Methode, die für die Zeitreihenprognose verwendet werden kann, ist das LSTM. Ein LSTM-Netzwerk (Long Short-Term Memory) ist ein Rekurrentes Neuronales Netzwerk (RNN), das sich besonders für die Verarbeitung und Vorhersage sequenzieller Daten eignet. Es verwendet eine komplexe Architektur von Zellen, die sich Werte über lange Zeiträume hinwegmerken können, so dass es Abhängigkeiten und Muster in sequentiellen Daten erfassen kann. In diesem kurzen Leitfaden wird anhand eines Beispiels gezeigt, wie thermische Last- und Wetterdatenverwendet werden können, um ein LSTM-Modell zu trainieren. Dieses Modell wird dann gespeichert und von der Festplatte geladen, um Lastvorhersagen für einen anderen Zeitraum zu machen. Darüber hinaus wird gezeigt, wie die Fehler der Vorhersagen berechnet werden können, wenn der tatsächliche Lastbedarf bekannt ist, wodurch ein kurzer Überblick über alle grundlegenden Komponentengegeben wird, die für die Erstellung eines Modells zur Vorhersage des Fernwärmebedarfs erforderlich sind.
Dieser Leitfaden wurde von der Hochschule Kempten im Rahmen des Forschungsprojekts "deepDHC" erstellt.
In the last decade autonomous driving has evolved from a science fictional dream to an everyday reality. With the advance of more and more companies bringing their versions of self-driving cars on the street it is just a matter of time before the majority of transportation will be in the hand of computers. But with the deadly car accident involving a self-driving Uber car back in 2018 there is also the question about how reliable autonomous driving really is and how we can validate and test the safety of this new road user 1. An uprising approach towards creating robust and adaptable neural networks is called domain randomization. This paper explores the possibility of using this method to create training data with driving simulations. It will propose a list of important criteria and factors affecting the selection of a fitting simulation. Furthermore it will present a track generator which is able to create useful tracks and export them to a format which can be used by several common simulations used in the field of autonomous driving research.
Temporal Fusion Transformer for thermal load prediction in district heating and cooling networks
(2022)
Accurate forecasting of thermal loads is a critical factor for operating district heating and cooling networks economically,efficiently and with minimized emissions. If thermal loads are known with high accuracy in advance, use of renewable energiescan be maximized, and fossil generation, in particular in peaking units, can be avoided. Machine learning has already provento be an efficient tool for time series forecasting in this context. One recent advancement in machine learning is the "TemporalFusion Transformer" (TFT), which shows especially good results in the area of time series forecasting. This paper examinesthe performance of TFT in the concrete context of thermal load forecasting for district heating and cooling networks. First,a brief summary of differences between TFT and other machine learning methods is given. Secondly, it is described how themethod can be adopted to train a machine learning model for thermal load forecasting. The data to train and evaluate the neuralnetwork is based on 8 years of hourly operating data made available from the district heating network of the city of Ulm inGermany. The presented technique is used to produce 72 hours of heating load forecasts for three different district heating gridsin the city of Ulm. The results are compared to forecasts of other machine learning methods that have been previously madeas part of the publicly funded research project "deepDHC", in order to evaluate if TFT is an improvement to further reduceforecasting uncertainties.
Efficient operation of district heating networks requires a precise forecasting of the thermal loads and an optimised dispatch strategy for the available generation and storage portfolio. This paper presents a holistic modelling and optimisation approach: first, detailed process modelling and optimisation of power plants and thermal storages; second, a numerical model for dispatch optimisation; and third, machine-learning-based load forecasting. The work is based on operating data from the district heating network of the city of Ulm in Germany. The paper presents the modelling, validation and simulation results of stationary and instationary process simulation for a biomass-fired combined heat and power plant. The analysis identifies a potential to integrate additional renewable power by “power-to-heat” technologies into different parts of the process. The economic benefit is quantified by mixed-integer linear programming optimisation applied to the district heating network. In order to allow for real-time dispatch optimisation, a machine-learning-based thermal load forecasting method was developed and evaluated, based on a 72-h forecast horizon. In addition, the economic impact of prediction uncertainties is analysed with the numerical dispatch optimisation tool.
Precise forecasting of thermal loads is a critical factor for economic and efficient operation of district heating and cooling networks. If thermal loads are known with high accuracy in advance, use of renewable energies can be maximized, and – in combination with thermal storage units – fossil generation, in particular in peaking units, can be avoided. Machine learning has proven to be a powerful tool for time series forecasting, and has demonstrated significant advancements in recent years. This paper presents the scientific methodology and first results of the publicly funded research project “deepDHC”, which aims at a broad benchmarking of traditional and advanced machine learning methods for thermal load forecasting in district heating and cooling applications. The analysis covers autoregressive forecasting approaches, decision trees such as “adaptive boosting”, but also latest “deep learning” techniques such as the “long short-term memory” (LSTM) neural network. This work is based on data from the district heating network of the city of Ulm in Germany. First, different performance metrics for evaluating forecasting qualities are introduced. Second, approaches for data screening and results of a linear and non-linear correlation analysis are presented. Third, the machine learning tuning process is described. For thermal load forecasting, weather data are key input parameters. This work uses hourly weather forecasts from weather models provided by the German meteorological service. These weather data are updated automatically, and have been statistically corrected in order to represent very accurate point forecasts for up to ten days ahead. In addition, a user-friendly web interface has been developed for use by the district heating network operator. The performance of different machine-learning algorithms is compared based on 72 h heating load forecasts.
Decarbonisation of district heating networks requires a heat generation and storage portfolio that allows for maximised integration of renewable energies. However, it is also essential to have a precise forecast of the thermal loads to be expected in the network and – based on this forecast – an optimised dispatch strategy in order to best match the available generation and storage portfolio with the actual heat demands in the grid. This paper presents a holistic approach that combines modelling and optimisation activities related to these three aspects: first, detailed process modelling and optimisation of power plants and thermal storages; second, a numerical model for dispatch optimisation; and third, machine-learning-based load forecasting. This work, which was performed as part of the publicly funded research projects “KWKflex” and “deepDHC”, was based on operating data from the district heating network of the city of Ulm in Germany. The paper presents the modelling, validation and simulation results of a stationary and instationary process simulation for a 58 MW thermal biomass-fired combined heat and power plant. The analysis identifies a potential to integrate additional renewable power of up to 17 MW thermal power by “power-to-heat” technologies into different parts of the process. The economic benefit is quantified with a mixed-integer linear programming dispatch optimisation model of the district heating network. In order to allow for real-time optimisation of the power plant and thermal energy storage dispatch, a machine-learning-based thermal load forecasting method was developed. The performance of different machine learning algorithms, including decision trees and deep learning techniques, is compared based on a 72-hour forecast horizon. In addition, the economic impact of uncertainties in thermal load prediction is analysed with the numerical dispatch optimisation tool.