Future Grids Will Not Be Controllable Without Thinking Machines
The monitoring, optimization and control systems for smart grids will require computerized intelligent systems to handle the increased variability and uncertainties caused by increased penetration of intermittent renewable energy resources. What principles will govern the design of such systems and where do we find them?
The smart grid can be viewed as a digital upgrade of the existing electricity infrastructure that minimizes the cost of energy and reduces emissions. A vast amount of data is generated and must be processed, so that the pertinent information is communicated to the appropriate control centers in time for necessary decisions to be made and adaptations to take place. All that will depend on the development of computerized intelligent systems, or computational systems thinking machines (CSTMs).
Such CSTMs will have to have three basic capabilities: sense-making, decision-making and adaptation. Realization of those capabilities will depend in turn on subsystems that continuously improve their knowledge of grid dynamics and not just gather data. Using traditional methods, however, it is difficult or impossible to model, control and optimize power systems because of their nonlinearity, spatial and temporal complexity, ever-changingness, and uncertainties.
Emerging computational paradigms such as computational intelligence (CI) and adaptive critic designs (ACDs) show promise for realizing a true CSTM.
Computational intelligence, a term coined by J. C. Bezdek in 1994, is the study of adaptive mechanisms to enable or facilitate intelligent behavior in complex, uncertain and changing environments. Such mechanisms can include both nature-inspired and artificial intelligence paradigms that exhibit an ability to learn or adapt to new situations, to generalize, abstract, discover and associate. CI is the successor of artificial intelligence and the way forward in future computing. The typical paradigms of CI, pioneered by many researchers, are artificial immune systems, evolutionary computation, swarm intelligence, fuzzy systems, neural networks, and their hybrids. Applications to power systems, such as neural networks and expert systems for load forecasting, were first proposed almost two decades ago.
The adaptive critic design, proposed by Paul Werbos in 1968, is based on combined concepts of reinforcement learning and adaptive dynamic programming. It is a powerful approach for the control and optimization of complex systems. ACDs use neural networks--so-called critic and action networks--as tools to carry out optimization over time.
In combination, CI and ACD technologies can provide a smart grid with capabilities for dynamic foresight, sense-making, situational awareness, rapid adaptation, fault-tolerance and robustness. For example, given expected and actual wind power generation across geographical locations, a CSTM based on CI and ACD should be able to dynamically project and dispatch power flows in a smart grid to energy storage units in order to accommodate excess wind power, and vice versa. Under system disturbances, a CSTM should have the ability to predict the modal frequencies that would be excited based on synchrophasor measurements and should, in turn, coordinate power oscillation damping controllers to suppress critical modes.
Several research studies have reported using CI-based technologies to dynamically forecast wind and solar power, monitor voltage stability, and assess real-time stability.
At the present time, local power system oscillation damping controllers can provide satisfactory performance when local measurements supply all the information about the effect of disturbances. But if multiple adjacent areas of the power system interact, wide-area measurements have the potential to provide better stabilizing control. A wide-area control system can coordinate the actions of a number of distributed agents using synchrophasor measurements.Of particular use in this connection are cellular neural networks--scalable neural network architectures that can handle the complexities of smart power systems and capture wide-area system dynamics. Cellular neural network architecture allows for accurate system-equivalent modeling and prediction.
Another promising application is in determination of optimal power flow. It generally is based on steady-state optimization, without considering local controller and load dynamics, and its optimal solutions are obtained based on forecasts. Although optimal power flow provides optimal dispatches for the next forecasted period, any unforeseen real-time load/generation variations or post-contingency operations between two dispatches are handled by simple linear controllers or some predefined reactions with little, if any, system-wide optimization.
I proposed a conceptual framework for applying ACDs to dynamically and stochastically optimize power system operations to the U.S. National Science Foundation in 2003. My collaborators and I have recently applied this framework to achieve optimal power flow control based on wide area measurements on a simulated medium size power system.
Another promising application area will be grid integration of plug-in electric vehicles. As they grow in numbers, the effects of adding small power electronic devices to the grid will become more and more predominant. Vehicle-to-grid applications are feasible in a smart-grid framework, causing increased bidirectional power flows between vehicles and utility grids.
CI-ACD technologies can enhance the integrated electricity and transportation infrastructures by optimally scheduling and dispatching real and reactive powers, without stressing the grid or the vehicles themselves.