Final Report Summary - COGENT (Control for Green Energy-Efficient Technologies)
In order to address the aggressive environmental targets on energy efficiency, energy generation from renewable sources, as well as the reduction of greenhouse gas emissions set worldwide, future power grids will increasingly have to make use of distributed uncertain generation, storage and loads for balancing the grid. Electric vehicles (EVs) offer an important new potential in this context, not only by reducing CO2 emissions, but most importantly by providing a large-scale storage capacity as well as a deferrable load that can be used to ensure grid stability during peak times and utilize intermittent renewable energy.
The main challenges for utilizing EVs and their optimal control in the power grid are 1) their main use as a means of transportation, making the connections of EVs to the grid highly variable and dependent on driving behavior, 2) the uncertainty in total grid load and power, 3) the distributed nature of the problem, requiring the control of a large number of dynamical sub-systems in a network, 4) the fast dynamics, in turn requiring fast sampling times. The COGENT project develops innovative control methodologies in order to address these new challenges by exploiting the opportunities offered by modern computing, sensor and communication technologies providing increasing computational power and real-time data. It develops new theoretical frameworks and practical tools building on model predictive control (MPC), a paradigm ideally suited for incorporating forward looking information in a safety-critical decision making process, which is combined with powerful modeling techniques from machine learning. The main goal is to provide high-performance methodologies with guarantees for application to the safety-critical smart grid system.
Four main objectives and activities have been identified and pursued in the course of the project:
1. Learning-based identification methods are investigated with respect to their integration in an MPC control formulation. Due to their flexibility as well as computational properties, Gaussian processes (GP) were identified as a suitable means to model complex system behaviors and characterize the corresponding model uncertainty. Stochastic MPC formulations for incorporating GP models are developed, with a particular focus on the prediction properties of the model and the tractability of the resulting optimization problem.
2. Safety guarantees for learning-based control methods are developed, where the main focus is on combining control theoretic with learning techniques. Constraint satisfaction is a core requirement for ensuring safety. We have proposed a methodology based on reachability analysis that can ensure satisfaction of safety constraints for any online learning method by resorting to a safety controller when the learning-based controller would risk safety. Key aspects addressed are the online learning of models for performing the reachability analysis, as well as their validation.
3. Techniques for the practical implementation of (learning-based) predictive controllers are derived addressing the fact that optimization-based controllers have to be computed in real-time environments with limited available computation time, or using distributed platforms. We are developing the theory to certify online optimization, i.e. to pre-determine the worst case computation time to achieve a given solution accuracy, and derive distributed predictive control techniques with enhanced performances that are able to adjust to online network changes and that can additionally deal with computation and communication limitations.
4. Optimal control methods for the energy management of electric vehicles (EVs) are developed, where the main focus is on using EVs for grid services. In this context, grid-aware charging strategies are derived that can deal with varying connections of EVs as well as safety constraints of the network. The ability to infer a user’s utility is investigated as a new and central aspect for successfully utilizing EVs that are personal loads for grid balancing while maintaining user satisfaction, with the goal of tailoring the controller to the user.
The main results achieved in these activities advancing the state of the art are summarized in the following:
* Provided new learning-based modeling and control approach using Gaussian processes to predict time-periodic errors that are then compensated by means of a predictive controller, which was demonstrated for high precision control of a telescope mount.
* Proposed plug and play distributed MPC method that enables safe topology changes, i.e. subsystems joining or leaving the network, during closed-loop control by introducing a novel online feasibility assessment and a transition phase.
* Proposed fast algorithm for solving MPC problems that competes with or even outperforms some of the best available methods while offering superior theoretical properties. Bounds on the computation time were derived that allow for providing guarantees for real-time implementation.
* Provided new conditions for guaranteed convergence of distributed MPC under local computation and communication errors, addressing a key practical limitation of iterative methods.
* Developed a framework for guaranteeing safety for any online learning method by a novel technique based on reachability analysis and online model validation that improves performance over available methods by learning safety boundaries online.
* Provided first plug & play MPC scheme for integrating electric vehicle charging with voltage control in distribution grids, providing optimal charge plans under varying connections and grid constraints.
* Proposed first ideas towards a personalized control scheme for electric loads that infers a user’s utility function online to tailor the performance of an MPC controller.
Next steps will develop an extended EV study, consolidating the results of the COGENT project and demonstrating the benefits of learning-based and distributed predictive control for their optimal energy management. The results will form the basis for a broader study of human effects in a network of dynamical systems in future research.
The COGENT project provides a set of comprehensive methods ranging from the derivation of prediction models from data, control methods to practical implementation techniques. Its main outcomes provide important steps towards the use of high performance control for safety-critical applications beyond the considered application to EV control in the smart grid. Learning-based control strategies that provide a systematic and quick translation from data and design specifications to the controller enhance control performance and reduce development times, a key aspect in any industrial development process. Real-time and distributed control techniques are key for enabling the envisioned capabilities in large-scale networks with agile components and fast dynamical characteristics, allowing to address a broader range of applications. The application of these new methodologies to the optimal control of EVs and the safety-critical power system provides important contributions towards increasing the role of automatic control in complex power networks and the overall goal of a smart grid providing green and efficient energy generation, distribution and management.
The main challenges for utilizing EVs and their optimal control in the power grid are 1) their main use as a means of transportation, making the connections of EVs to the grid highly variable and dependent on driving behavior, 2) the uncertainty in total grid load and power, 3) the distributed nature of the problem, requiring the control of a large number of dynamical sub-systems in a network, 4) the fast dynamics, in turn requiring fast sampling times. The COGENT project develops innovative control methodologies in order to address these new challenges by exploiting the opportunities offered by modern computing, sensor and communication technologies providing increasing computational power and real-time data. It develops new theoretical frameworks and practical tools building on model predictive control (MPC), a paradigm ideally suited for incorporating forward looking information in a safety-critical decision making process, which is combined with powerful modeling techniques from machine learning. The main goal is to provide high-performance methodologies with guarantees for application to the safety-critical smart grid system.
Four main objectives and activities have been identified and pursued in the course of the project:
1. Learning-based identification methods are investigated with respect to their integration in an MPC control formulation. Due to their flexibility as well as computational properties, Gaussian processes (GP) were identified as a suitable means to model complex system behaviors and characterize the corresponding model uncertainty. Stochastic MPC formulations for incorporating GP models are developed, with a particular focus on the prediction properties of the model and the tractability of the resulting optimization problem.
2. Safety guarantees for learning-based control methods are developed, where the main focus is on combining control theoretic with learning techniques. Constraint satisfaction is a core requirement for ensuring safety. We have proposed a methodology based on reachability analysis that can ensure satisfaction of safety constraints for any online learning method by resorting to a safety controller when the learning-based controller would risk safety. Key aspects addressed are the online learning of models for performing the reachability analysis, as well as their validation.
3. Techniques for the practical implementation of (learning-based) predictive controllers are derived addressing the fact that optimization-based controllers have to be computed in real-time environments with limited available computation time, or using distributed platforms. We are developing the theory to certify online optimization, i.e. to pre-determine the worst case computation time to achieve a given solution accuracy, and derive distributed predictive control techniques with enhanced performances that are able to adjust to online network changes and that can additionally deal with computation and communication limitations.
4. Optimal control methods for the energy management of electric vehicles (EVs) are developed, where the main focus is on using EVs for grid services. In this context, grid-aware charging strategies are derived that can deal with varying connections of EVs as well as safety constraints of the network. The ability to infer a user’s utility is investigated as a new and central aspect for successfully utilizing EVs that are personal loads for grid balancing while maintaining user satisfaction, with the goal of tailoring the controller to the user.
The main results achieved in these activities advancing the state of the art are summarized in the following:
* Provided new learning-based modeling and control approach using Gaussian processes to predict time-periodic errors that are then compensated by means of a predictive controller, which was demonstrated for high precision control of a telescope mount.
* Proposed plug and play distributed MPC method that enables safe topology changes, i.e. subsystems joining or leaving the network, during closed-loop control by introducing a novel online feasibility assessment and a transition phase.
* Proposed fast algorithm for solving MPC problems that competes with or even outperforms some of the best available methods while offering superior theoretical properties. Bounds on the computation time were derived that allow for providing guarantees for real-time implementation.
* Provided new conditions for guaranteed convergence of distributed MPC under local computation and communication errors, addressing a key practical limitation of iterative methods.
* Developed a framework for guaranteeing safety for any online learning method by a novel technique based on reachability analysis and online model validation that improves performance over available methods by learning safety boundaries online.
* Provided first plug & play MPC scheme for integrating electric vehicle charging with voltage control in distribution grids, providing optimal charge plans under varying connections and grid constraints.
* Proposed first ideas towards a personalized control scheme for electric loads that infers a user’s utility function online to tailor the performance of an MPC controller.
Next steps will develop an extended EV study, consolidating the results of the COGENT project and demonstrating the benefits of learning-based and distributed predictive control for their optimal energy management. The results will form the basis for a broader study of human effects in a network of dynamical systems in future research.
The COGENT project provides a set of comprehensive methods ranging from the derivation of prediction models from data, control methods to practical implementation techniques. Its main outcomes provide important steps towards the use of high performance control for safety-critical applications beyond the considered application to EV control in the smart grid. Learning-based control strategies that provide a systematic and quick translation from data and design specifications to the controller enhance control performance and reduce development times, a key aspect in any industrial development process. Real-time and distributed control techniques are key for enabling the envisioned capabilities in large-scale networks with agile components and fast dynamical characteristics, allowing to address a broader range of applications. The application of these new methodologies to the optimal control of EVs and the safety-critical power system provides important contributions towards increasing the role of automatic control in complex power networks and the overall goal of a smart grid providing green and efficient energy generation, distribution and management.