BDA.16 Towards Prescriptive Analytics in Cyber-Physical Systems

More and more of our physical world today is being monitored and controlled by so-called cyber-physical systems (CPSs). These are compositions of networked autonomous cyber and physical agents such as sensors, actuators, computational elements, and humans in the loop. Today, CPSs are still relatively small-scale and very limited compared to CPSs to be witnessed in the future. Future CPSs are expected to be far more complex, large-scale, wide-spread, and mission-critical, and found in a variety of domains such as transportation, medicine, manufacturing, and energy, where they will bring many advantages such as the increased efficiency, sustainability, reliability, and security. To unleash their full potential, CPSs need to be equipped with, among other features, the support for automated planning and control, where computing agents collaboratively and continuously plan and control their actions in an intelligent and well-coordinated manner to secure and optimize a physical process, e.g., electricity flow in the power grid. In today’s CPSs, the control is typically automated, but the planning is solely performed by humans. Unfortunately, it is intractable and infeasible for humans to plan every action in a future CPS due to the complexity, scale, and volatility of a physical process. Due to these properties, the control and planning has to be continuous and automated in future CPSs. Humans may only analyse and tweak the system’s operation using the set of tools supporting prescriptive analytics that allows them (1) to make predictions, (2) to get the suggestions of the most prominent set of actions (decisions) to be taken, and (3) to analyse the implications as if such actions were taken. This thesis considers the planning and control in the context of a large-scale multiagent CPS. Based on the smart-grid use-case, it presents a so-called PrescriptiveCPS – which is (the conceptual model of) a multi-agent, multi-role, and multi-level CPS automatically and continuously taking and realizing decisions in near real-time and providing (human) users prescriptive analytics tools to analyse and manage the performance of the underlying physical system (or process). Acknowledging the complexity of CPSs, this thesis provides contributions at the following three levels of scale: (1) the level of a (full) PrescriptiveCPS, (2) the level of a single PrescriptiveCPS agent, and (3) the level of a component of a CPS agent software system. At the CPS level, the contributions include the definition of PrescriptiveCPS,
according to which it is the system of interacting physical and cyber (sub-)systems. Here, the cyber system consists of hierarchically organized inter-connected agents, collectively managing instances of so-called flexibility, decision, and prescription models, which are short-lived, focus on the future, and represent a capability, an (user’s) intention, and actions to change the behaviour (state) of a physical system, respectively. At the agent level, the contributions include the three-layer architecture of an agent software system, integrating the number of components specially designed or enhanced to support the functionality of PrescriptiveCPS. At the component level, the most of the thesis contribution is provided. The contributions include the description, design, and experimental evaluation of (1) a unified multi-dimensional schema for storing flexibility and prescription models (and related data), (2) techniques to incrementally aggregate flexibility model instances and disaggregate prescription model instances, (3) a database management system (DBMS) with built-in optimization problem solving capability allowing to formulate optimization problems using SQL-like queries and to solve them “inside a database”, (4) a real-time data management architecture for processing instances of flexibility and prescription models under (soft or hard) timing constraints, and (5) a graphical user interface (GUI) to visually analyse the flexibility and prescription model instances. Additionally, the thesis discusses and exemplifies (but provides no evaluations of) (1) domain-specific and in-DBMS generic forecasting techniques allowing to forecast instances of flexibility models based on historical data, and (2) powerful ways to analyse past, current, and future based on so-called hypothetical what-if scenarios and flexibility and prescription model instances stored in a database. Most of the contributions at this level are based on the smart-grid use-case. In summary, the thesis provides (1) the model of a CPS with planning capabilities,
(2) the design and experimental evaluation of prescriptive analytics techniques allowing to effectively forecast, aggregate, disaggregate, visualize, and analyse complex models of the physical world, and (3) the use-case from the energy domain, showing how the introduced concepts are applicable in the real world. We believe that all this contribution makes a significant step towards developing planning-capable CPSs in the future.