Today's successful power electronic products and systems are the result of engineering and management skills brought together to achieve application-specific objectives. Examples include the design of high-performance motor drives and complex system power supplies. Success — especially if it's to be sustained — requires that a coordinated balance of power and control system design theory, simulation and verification techniques, manufacturing, and test methods be continuously and effectively managed. This endeavor is a challenge to execute on time and within budget.
A power electronic building-block approach has long been envisioned as a means to simplify the design and application of power electronics technology. However, progress has been slow. Differences exist between unique applications, and while these share similar technology and principles, they differ enough in their detailed implementation to make any efforts of standardization ineffective.
Few power systems designed today reach volumes where sustainable economy of scale can be realized before obsolescence takes its toll. This is a result of the relative difficulty to reuse engineering work, compared to the reuse enjoyed in the software industry. Technological change coupled with intense competition promotes a never-ending cycle of obsolescence.
Is a standard building-block approach that can survive the continuing trauma of change possible? The answer is yes, but it requires a technology foundation that is based on fundamental, unchanging principles. This technology must produce performance benefits unattainable with traditional methods, while simplifying the entire product process from initial concept to beyond delivery. Achieving these goals requires development of a generic control architecture for power systems. To fully understand the benefits of such an architecture, we must review seven fundamental principles of traditional power electronics control.
- The Switch-Mode Principle
The chopper circuit of Fig. 1 illustrates the simple concept of turning a switch off-and-on periodically with adjustable duty cycle as a means of emulating a linear regulator. The switch-mode technique controls the output voltage by switching at high frequency to reduce the size and weight of the power supply. This is possible because the ideal switch dissipates no energy in the on or off states.
Under the assumption of continuous current, the output voltage is linear with respect to duty cycle, D, as shown in Fig. 1.
With linearity as a primary assumption in the modulation strategy, new power circuits and control systems were quickly built around the switch-mode principle, and stabilized for gain and phase margin according to widely accepted traditional methods of linear control theory. Another example is the popular 3-phase voltage source inverter used to drive induction motors.
Note that regardless of the power circuit type, for every N switches in the power circuit up to 2N switch states can be used. All specific control and modulation function implementations can be thought of as one of an infinite number of possible controlled sequencing solutions used to achieve the system objectives.
No matter what the power circuit structure is, there's a minimum allowable gating command sequence time interval. This is because the system gate-signal propagation delays and the power device characteristics can't support a change sequence that lasts less than hswmin without damaging or destroying the device.
In this article, variables such as hswmin employ the concept of discrete time step size and will be referred to often. Therefore, we will use the nomenclature hx where the units of h are seconds, and the subscript x is a description of the type of time step to which we are referring.
Traditional control systems encounter limit cycle problems when operating at low duty cycle because the controller doesn't maintain linear behavior in this region due to the lock up of the switch at hswmin pulse width. Therefore, the generic controller must incorporate hswmin as a parameter and must account for the nonlinear effects produced by the switch locking on or off for hswmin time as part of a uniform switch-mode-process.
- Control Systems and Objectives
Traditional closed-loop feedback systems were conceived using the switch-mode principle, allowing regulation of current, voltage or other variables. Traditional methods assume the transfer function of the load and the “tuning” of the analog controller determine the steady-state and transient behavior of the system. Tuning of the controller is determined by selection of a controller transfer function, and tuning parameters are designed to meet and balance applicable steady-state and transient characteristics according to “the performance specification.”
Fig. 2 illustrates the transfer function-shaping objective that is the foundation of traditional closed-loop control theory. However, the process of design is complex, and sensitive to changes that may occur in the dynamic specification, the load transfer function or the power circuit hswmin characteristics. Any change can trigger the need for revised stability analysis. Therefore, a generic controller must simplify the tuning process and eliminate or minimize additional work traditionally required to retune the performance when specifications change.
The need for more comprehensive understanding of nonlinear effects resulting from the modulator, power circuit and other circuits is critical to success. Discrete time simulators use initial conditions of state variables (the current in an inductor or voltage across a capacitor), and the applicable circuit node and differential relations of a specific circuit for verifying time-domain responses of the system before they are built. These simulators used complex numerical algorithms to solve the differential equations in small discrete increments of time, or hsim.
However, in the area of power electronics where the action of switching causes abrupt changes of current or voltage in small intervals of time, the early simulators had difficulty achieving convergence. Convergence is the process involving the numerical stability of reaching a stable and unique solution at each time step of the state variables in the system.
Improvements have been made with larger word-length computers (higher numerical dynamic range and precision) as well as multistep algorithms, which selectively make smaller hsim steps when high rates of change are detected. These multistep approaches, however, make simulation of complex, multiswitch systems slow. Even today, on some of the fastest workstations, a 30-sec real-time run-up of an induction motor drive operating at a 50-kHz switching frequency may take several hours to simulate.
The simplest form of discrete-time simulation techniques is based on the simple relations iterated in real time: I = C dV/dt for capacitors, and V = L di/dt for inductors where dt is replaced with hsim, and dV and di are replaced with differences determined by past and present samples of the voltage and current.
Simulation is critical for design and verification. Therefore, the ideal generic control architecture would incorporate a virtual real time simulation capability within it to improve and verify performance, eliminating the time and overhead of working with offline simulation tools.
- Digital Control and Digital Filters
With the introduction of low-cost microcontrollers and DSPs, digital filters and control became popular in motor drives. These components allowed complex analog controllers to be converted to digital format for newer, more versatile products.
In this discipline, present and past inputs [u(k), u(k-1)] and outputs [y(k), y(k-1)] of regularly sampled signals are multiplied and added with suitably chosen sets of coefficients Knum and Kden to produce the desired new output y(k) at each time step we'll call hc. The values of the coefficients determine the behavior, such as low pass, high pass or bandpass filter, PID control. Two classes of digital filters exist: infinite impulse response (IIR) and finite impulse response (FIR).
IIR filters use feedback samples y(k) to calculate the output. IIR filters can be unstable because they use feedback but are extremely efficient. FIR filters are always stable because they only use inputs, but are less efficient. The following equation shows the basic structure used to implement virtually all digital filters.
The unconditional stability characteristics of FIR structures are of particular interest in the control system discussed in this article.
Therefore, the ideal generic control architecture should utilize FIR concepts in its implementation. This computational structure has been optimized for speed, promotes stability and is shared with the communications industry, where they continue to drive the technology to cost efficient-levels as a result of mass-market opportunity.
- Physical Power Systems and Scalability
The need to scale a motor drive design from watts to megawatts requires intelligent reuse of engineering. Switch technologies such as IGBT, FET, BJT and GTO have unique values of hswmin that vary according to the specific technology, power level and gate-drive train design characteristics that exist in any given embodiment.
Traditionally, scalability of power has been managed with a lot of work. It's important to recognize variations of hswmin between large GTO motor drives compared to smaller IGBT motor drives can be as great as 60 µs to 1µs. This often leads to the design and management of new and different designs, based on different modulation strategies.
Thus, the pursuit of a generic architecture must support the ability to maintain these variations in power circuit personality, without requiring a complete change in control philosophy as a result of these differences.
- Physical Computational Systems and Scalability
Suppose a specific control system uses a DSP version A as the computational engine running hc microseconds in a certain application. The same control architecture should be able to accommodate operation with an improved DSP version B that exhibits 0.5*hc processing capability. At the same time, the architecture should be able to selectively exploit the improvement in processing capability to deliver higher performance. Or, the improvement may be leveraged to operate two motor drives simultaneously in version B but with version A quality performance.
Since rapid advances in computational capabilities will continue, the ideal generic control architecture must allow these advances to be readily employed to achieve cost versus quality tradeoffs at any time.
- Power-Quality Control
It's common knowledge that the switch-mode principle applied to a motor drive or a power supply can improve power quality as the switching frequency is increased. Of course, higher switching frequencies eventually bring other problems, such as EMI, thermal losses in inductors, capacitors, switches and proximity effects.
The relationship between switching frequency, power quality, size, weight, cost and performance isn't easily determined or managed with traditional design. If the power quality or transient response specification is changed in the midst of the design process, complete rework must often be done. In some applications, lower power or performance quality may be acceptable where higher cost is not.
Therefore, a generic control architecture must provide the opportunity to trade cost for quality throughout the design and after delivery. This capability isn't one that traditional controllers can easily provide.
In summary, to be both generic and universal, the control architecture must apply to any power circuit, any load, any computational engine and any combination of the three. It must be able to accept any of today's electronic devices and work with their existing limitations, yet be adaptable to the inevitable evolution of those devices in the future. It must be capable of effectively dealing with all of the limitations of the traditional control systems as outlined above. It must be practical and cost effective.
Generic Cognitive Predictive Control Architecture
In the past, traditional control depended on frequency-domain engineering techniques to establish performance. In contrast, the Generic Cognitive Predictive Control Architecture depends on time-domain synthesis techniques to achieve the desired goals.
Most problems with traditional control are solved by recognizing that the actuator, “the switch”, is a discrete-time device. As such, viewing the switch state as a time slice hswmin allows the mathematics to be integrated with well-established simulation and discrete-time control methods. This frees us to look at the switching sequence challenge in a different way.
While all switching processes are inherently the same, it's clear there are an infinite number of possible sequences. Thus, our challenge degrades to finding and tapping into the optimal sequence that meets the desired goals, subject to the constraints of the system.
All traditional modulators and control systems do the same thing, except they only tap into one possible set of solutions early in the design, and are then forever limited. This is why reuse and reapplication with traditional power electronic control systems has been difficult; hence, this has deterred the evolution of a generic control standard.
The generic control architecture described here has no such limits. It will always do the best that can be done given the applicable constraints of the system. For example, all control architectures are fundamentally constrained to the hswmin principle. This narrows the realm of possible solutions. A power circuit and a load are selected, and the addition of performance objectives and control engine capability further reduces the viable switching solutions.
Therefore, the primary goal of the generic control architecture is to sift through the remaining available sequences in real time to select the best applicable sequence under the conditions that exist at each and every time step.
Numerical techniques beyond the scope of this article have been developed that synchronize and amalgamate the discrete-time equations of digital filters (hc), discrete-time-simulators (hsim) and the power switch minimum gate drive train time concept (hswmin) into one comprehensive, yet elegantly simple algorithm. This algorithm exhibits exceptional simplicity, simulation convergence characteristics and allows faster than real-time simulation to be achieved.
The architecture shown in Fig. 3 is cognizant because it is “aware” of the specific nature of the load, the power circuit and the computational engine. Fig. 4 shows the Generic Cognitive Predictive Control Architecture. It uses an embedded simulator and a predictive engine to build in real time the ideal switching sequence given the performance goals (e.g., controlled torque ripple and current ripple) and constraints (e.g., hswmin, hsim and hc) From here on, we'll refer to the architecture as CogniSim.
The control is executed in every hint microseconds or nanoseconds. The hint parameter is derived from hsim, hswmin, and hc primary time slice constraints.
Every hint time step, CogniSim samples and does A/D conversion on a set of real signals, such as current or voltage in the load model, and the dc bus voltage. These samples are impressed to the embedded simulation to lock state variables with the actual load circuit. In some cases all signals are synthesized. In these cases, no A/D system is required.
The present state contemplator runs the embedded simulation to infer the remaining unmeasured state variables. Then, it determines according to a configurable set of criterion whether or not a change to a next state should be contemplated. If a change of state is determined worth considering, the next state contemplator builds a set of fictitious virtual simulations and seeds each of them with the state-variable values present in the embedded simulation.
All relevant “trial-run” virtual simulations are executed for a single virtual hint time step. The number of virtual models depends on the unique nature of every power circuit and the unique state transition constraints of the CogniSim controller. Figs. 5 and 6 show how the chopper and motor drive are configured within CogniSim.
Following this, the control evaluates the outcome of the set of virtual simulations and decides which next state to choose. Often, the same state — that is no change at all — may be selected because it may actually produce better results than the alternatives.
CogniSim switches only when necessary. Thus, it's able to account for hswmin lockup, and it typically reduces the average switching frequency by 30% compared to traditional modulation techniques in applications, such as sinusoidal motor drives where the reference goals are periodic in time.
In applications where the reference is non-periodic like all-digital audio, active suspension or camera stabilization systems, CogniSim can reduce the switching frequency by up to 80%, because it exploits the random signal for opportunities to improve quality, reduce switching frequency and to fit the dynamic trajectories to the desired reference trajectories in the best way possible. This leads to meaningful savings in heatsinks, EMI suppression and battery consumption.
Because CogniSim by definition simulates the dynamic transient response every hint cycle, it typically produces the ideal transient response possible under all dynamic conditions, and yet maintains steady-state performance.
Fig. 7a shows operation of the chopper of Fig. 1 configured with goals that tell CogniSim to stay as close to the reference signal as possible but remain above it. CogniSim shuts off completely when the action of switching does nothing to contribute to a better trajectory. Fig. 7b shows large step square wave response. Note there are no filters or tuning in CogniSim. The model is the tuning. Fig. 7c shows CogniSim's outstanding dc bus ripple rejection.
Since hint is a variable, CogniSim can be ported into slow or fast DSPs or FPGA hardware, and the attainable quality is predictable because CogniSim is predictive in nature.
If a switch fails short, CogniSim automatically adapts to achieve a reduced level of control, but at least continue to control to the best ability possible. CogniSim just removes the nonworking states from the viable list of alternatives, and does its best with the remaining available states.
In motor drives and other inverter applications, such as UPSs, the interlock period when two transistors are off introduces ambiguity as to whether the pole voltage is high or low.
In these cases, the pole voltage is determined entirely by the load. By using the “direction” of the current during the interlock time, CogniSim reveals another outstanding exploitive capability: its ability to detect and reflect the effect of mechanical load disturbances. This is the key to outstanding sensorless control.
CogniSim can produce extremely accurate stator flux signals at zero speed, so shaft rotation imposed from an external user can be detected without any other sensor except the dc bus voltage and 2-phase currents. Furthermore, rotating restarts are possible where synchronization to a pre-spinning motor is done quickly and reliably without transient, and torque control to zero speed is possible.
In the case of the aforementioned inductor motor drive, the state diagram may show eight possible next states. There is a 1-in-8 chance of choosing the best next state without CogniSim having any information. The role of the CogniSim controller is to simply remove the uncertainty of choosing the wrong next state.
CogniSim reveals the switch-mode principle for what it really is: a nonlinear, controlled instability. This is in contrast to the traditional approach of linearizing the modulation process for the purposes of embedding it in a traditional feedback control system, only to find that the assumptions at some point are invalid.
In the motor drive application, with CogniSim's ability to detect bipolar power flow as a result of optimal torque control, CogniSim can be used to sense abnormalities in mechanical systems, since vibration is reflected in the measured current signal trajectories. This process is called fingerprinting and offers a valuable opportunity to detect problems in, for example, aerospace applications before a catastrophic mechanical failure occurs.
Practical, Flexible and Configurable
A CogniSim controller is practical because it can be implemented in DSP, FPGA or ASIC technology. The fixed structure of CogniSim assures ease of reuse, and opportunity for optimization and integration.
The variable structure of all power electronic applications can be managed by configuring CogniSim with unique quality criterion, available switch states, and transitions, load circuit relationships and the capability of the processor used.
Because these elements can be represented in a standard format, they can be downloaded into a CogniSim controller to produce the desired system behavior. CogniSim's object-oriented nature and these key configurable elements form the basis of a standard that assures reuse and organized flexibility in present and future designs.
Today, we are at a crossroads in history where digital signal processing is becoming incredibly powerful and cost effective. The ability to “think” faster than a power system can toggle the switches is a reality. However, this power only has value if it can be exploited effectively. The generic control architecture described in the article offers the means to achieve this.
For more information on this article, CIRCLE 342 on Reader Service Card