Le 12/05/2022 par paterijk :
Bonjour,
Je vous propose ci-dessous un sujet de thèse en "optimisation" énergétique à IMT Atlantique (campus de Nantes). N'hésitez pas à diffuser auprès de personnes qui pourraient être intéressées.
Cordialement,
Patrick Meyer
----
Heating networks are made up of nodes and branches (two pipes in parallel) in which the fluid circulates. These networks can be as complex as meshed structures. At each node (energy delivery substation or production nodes), the flows are characterized by physical variables (temperatures, flow rate, etc.) whose evolution is usually monitored at time steps of the order of ten minutes, over periods of up to a full year. Similarly, each substation is associated with a time profile of energy demand resulting from complex time parameters (occupancy profile of the connected buildings, profile of internal equipment, associated intermittences, etc.). These two characteristics of the nodes make their detailed physical modeling difficult, in particular when it comes to modeling real networks with the aim of changing existing structures (extension, distribution modifications, integration of new production units, etc.) based on supervisory measures. A simplification strategy consists in simplifying the topology of the network by aggregating a well selected number of nodes. The reduction must however preserve the overall operation of the network and its compatibility with the measured data.
In this work, we propose to study the contributions of artificial intelligence (AI) in the definition of a methodology for the topological simplification of District Heating networks. The work consists in defining a method for simplifying the structure of the network by aggregating a set of nodes and branches to limit the complexity to be modeled. The difficulty lies on the one hand in the relevant choice of nodes and branches to be aggregated, and on the other hand in the need to keep the values ??of the physical variables at the interface of the aggregated area, in order for the simplified network to be equivalent to the real network. This problem poses two major difficulties which make a systematic physics-based modeling impossible without penalizing simplifications and the use of classical optimization methods incompatible with acceptable computation times (non-linearities, large number of combinations of topology reductions to be tested, etc.). Furthermore, the characteristics of the aggregated nodes must be determined on the basis of the time series of all the nodes they represent, and this for each of the variables (temperatures, flow rates, demand, storage states, etc.). To overcome these difficulty, we propose to study the contributions of artificial intelligence to:
Identify the most relevant areas of the network to be aggregated on the basis of predefined simplification parameters (number of nodes to aggregate, inclusion or exclusion of energy systems in the areas to be aggregated, etc.)
Identify the number of relevant nodes to represent the area to be aggregated, depending on the complexity of this area
Determine the time series of the physical variables of this (these) aggregated node(s) so that their dynamic behavior is equivalent to the corresponding part of the real network. In particular, the values??of these physical variables at the boundaries of the zone will be constraints of the procedure.
The generated temporal profiles will be integrated into the physical models of heating networks and their relevance for an analysis of the complete network will be realized. In particular, the results obtained outside the aggregation zone of the simplified network will be compared to the solutions of the complete problem. A follow-up in terms of calculation performance of the proposed methods will be carried out to qualify and quantify their relevance.
With this work, we aim to provide a robust and replicable methodology of for topology reduction of District Heating based on AI (topic barely addressed in the literature). The genericity of the approach will be replicable for other energy networks (gas, electricity) and will contribute to the analysis of Multi Energy Systems. The works is based on an interdisciplinary approach which combines physics of systems and artificial intelligence expertise. The challenge is situated in managing the complexity to be modeled while maintaining a high level of explainability of the results given by AI. The problem of simplifying the representation of district heating (for the purpose of modeling and optimization) is a very recent and sustained activity in several international laboratories specialized in these energy systems.
The main objective of this work is to propose a robust and replicable methodology based on explainable machine learning. AI techniques will be used in three steps in the work: first the identification of the areas of the network to be simplified, based on time series and characteristics of the original areas and nodes (geospatial, physical, socio-economic features, etc.), second the proposal of a new network topology, while guaranteeing a similar physical behavior of the entire network, and, third the determination of the equivalent time series of the characteristic parameters of the nodes
The objective being to establish a simplification methodology, the support network will be a fully defined network (structure, demand, systems, etc.) in order to control all the characteristics. It will be modeled in an approach totally based on physical models (Gemellus platform of the DSEE Dept.). This physical model will make it possible to build the data necessary for the different phases of the AI, thus simulating a history of measurements and characteristics. It will also make it possible to check the relevance of the simplifications proposed by comparison with the physically simulated data. It is planned to explore different machine learning or optimization algorithms to meet these goals. Since the District Heating network can be represented by a graph, graph clustering algorithms will be investigated (one of the key techniques to understand the structures and data in a graph). Regarding the determination of the time series of the simplified structure, machine learning methods such as LSTM networks, which are neural networks that make predictions based on previous time periods, will be used. More explainable algorithms will be studied, such as random forests or gradient boosting, which allow temporal information to be included through a set of delays that are added to the input, so that data is represented at different timesteps. In the end, this work must lead to a clear methodology relevant to the subject, as well as to the analysis of their contribution for complex physical models
Literature review on AI methods for time series aggregation and clustering of District Heating networks nodes
Selection of a set of AI methods for the identification of the candidate areas for aggregation
Identification of the characteristics of the nodes to be aggregated (features)
Selection of a set of AI methods to test for the aggregation phase
Application on a test network (with constitution of the data necessary for the analysis of the performance of the methodology)
Scaling up and testing the method on a complete network.
Comparison with the results of a complete model
Comparison with other network simplification solutions (based on physical models) and analysis of the performance of the proposed method
The candidate must hold a master degree in energy with an open mind for data driven models. He/she can also have a master degree in computer science (or equivalent), with however an ability to integrate the required energy concepts. The candidate should have skills in the use and development of machine learning algorithms. Ideally, he or she should have solid knowledge of Python. He or she should also have an advanced level in English, and a basic knowledge of French.
To apply for this position, please send your resume (CV), motivation letter, and some references to Bruno Lacarrière (bruno.lacarriere@imt-atlantique.fr) and Patrick Meyer (patrick.meyer@imt-atlantique.fr).