#	Pagina
attuale pagina	/open-h2020/projects/203417/results.html

Report

Teaser, summary, work performed and final results

Periodic Reporting for period 2 - SYSDYNET (Data-driven Modelling in Dynamic Networks)

Teaser

Summary

Our changing society is demanding smart and intelligent engineering solutions for major technological challenges, where resources and energy, as well as human resources, have to be used in the most efficient way, to arrive at sustainable and smart solutions for, e.g.,
â€¢ industrial production (smart industries; manufacturing and processing)
â€¢ energy generation and distribution (integrated generation, smart grids)
â€¢ transportation systems (smart cars, drones, planes, next-generation infrastructures)
â€¢ urban environments (smart cities, smart buildings)
One of the common characteristics in these challenges is the control and optimized operation of highly complex, large-scale, multi-physics and interacting dynamic systems. In these automated operations, computation, (internet-) communication and control are integrated in efficiently operating plug and play intelligent automation systems, that warrant flexibility, robustness, stability and performance. The result is automation systems that optimally optimally manage and control cyber-physical systems of systems.

In this development dynamic models play a key role. They serve important purposes of simulation, diagnosis, and learning/understanding the characteristics of processes in their behavior over time, and they have a paramount role in (model-based) simulation, fault detection, measurement, control and optimization. In view of the technological developments, the models will have to reflect the large-scale interconnected and networked character of the dynamic processes of study.

Models can be built either on the basis of first-principles relations or on the basis of experimental data, or a combination of both. Data-driven modelling is particularly important for (a) effectively incorporating the emergent behavior of systems, (b) quantifying and minimizing the effect of uncertainties, (c) adapting to time-varying behavior, (d) accurately estimating the parameters in first-principles models and (e) possibly avoiding the time-consuming task of first principles modelling. Therefore effective data-driven modelling tools for dynamic networks are essential ingredients for operating and controlling many of our future engineering systems. Since a comprehensive theory for data-driven modelling of (parts of) dynamic networks is lacking, the overall objective of this project is to
develop a comprehensive theory for the data-driven modelling of dynamic networks, that can address (a) the identification of dynamics and interconnection structure (topology) of local parts of the network, (b) aspects and optimal choices of sensor and actuator placements and of experiment design (c) incorporation of prior (partial) knowledge on network topology and local network dynamics and (d) the properties of identified (local) models that are relevant for model-based distributed control.

With the increasing size and complexity of dynamic networks it becomes relevant for data-driven modelling methods to be able to focus on a particular part of the network, as identification of the global network will be too costly and out of reach. We will call this {it local identification}. Identifying the local dynamic subsystems is one target, but additionally also identifying the network topology/structure is an important aspect to consider. Which node variables are (causally) connected to which other node variables, and can we detect the presence of (feedback) connections in the network? Statistical properties of the estimation procedures will typically be highly dependent on the information content of measured signals. This involves the possible choice and location of sensors and of actuators, and the possible design of dedicated excitation (probing) signals to achieve optimal model accuracy. If some modules in the network are known (through first-principles-based models or as known local controllers), this will substantially influence the modelling results, and so it has to be taken into account. Finally our

Work performed

The problem of local (or single-module) identification has been addressed, and an algorithmic solution has been developed for determining which signals need to be measured in the network on the basis of which a consistent estimate of the target module can be made. Several estimation algorithms have been studied, and their properties have been analyzed, aiming at the completion of a theory for consistent module estimation. At the same time research has been started to extend the approach to not only aim at consistent estimates but also at minimum variance estimates, reaching the Cramer Rao lower bound. This is leading to a very fundamental basic building block in network identification problems.

The global network identification problem, including topology estimation, has been addressed by introducing and analyzing the concept of network identifiability. The appropriate concepts have been defined and verifiable conditions for analyzing network identifiability have been developed, where the conditions are formulated in terms of type and location of external signals and prior knowledge of the network, and where the conditions can be verified by path-based conditions in the graph of the network. In a next step the results are extended to address identifiability questions related to parts (modules) of the network.
Additionally a new algorithm for topology estimation has been developed based on tools from Bayesian methods, connecting to recent developments in machine learning approaches to identification problems.

A scalable algorithm has been developed for estimating network dynamics that can act as a basic ingredient in network estimation problems. And for the estimation of single module dynamics a machine-learning algorithm has been developed that avoids preselecting model orders and model structures of the many auxiliary modules that need to be estimated.

On the modelling side, a research project has started to more fundamentally evaluate which modelling framework is most suitable for addressing questions of data-driven modelling in dynamic networks. The so-called module representations are compared to state-space-based representations, and relations between the two model structures have been specified. Additionally, particular structures of physical networks, induced by diffusive couplings, have been analysed and brought in the same context as the so-called module networks. This has serious impact for available identification algorithms in physically structured networks.

In a first step towards facilitating distributed model-based control, a distributed identification algorithm has been developed that applies a multi-agent type of setting to a network identification problem, where computations are being distributed and limited exchange of information between the local agents is allowed.

The developed theories are currently being applied in different application projects, including power grid modeling, distributed climate control in buildings, brain networks and leak detection in gas pipelines.

Other research projects have started to address the problems of (a) topology identification and detecting nonlinearities, (b) distributed control and estimation, (c) sampling strategies.

Final results

We expect to develop a basic set of analysis tools, methods and algorithms for solving a range of problems in data-driven modeling of dynamic networks.