By Richard Vesel, ABB
Multi-variable model-based control is a technology known to provide superior performance over traditional single-input-single-output control strategies. Originally developed for petroleum refineries and power plants, it has become common in process industries over its 30 years in applied practice. One of the first such applications was for a power plant in France. But only recently has it found its way to power plant control and optimization.
This slower progress in the power sector can be partially explained by higher performance requirements. The dynamic behavior of power plant components is usually much faster than that found in industrial processes and requires computing power that until recently was either not available or not cost-effective. As compared to the benefits, the very high installed cost (for example, hardware, software and engineering) was also prohibitive. Relatively low energy costs and lack of strict environmental requirements also played a role.
Advanced control technologies have taken many years to work their way from academia and first-use industrial trials into mainstream applications and use. For utility applications of multivariable combustion control technologies, there have been three overlapping generations of advanced controls which are currently in use.
The first generation used neural networks to build the multi-variable model. The second generation was the linear multivariable model-predictive controls. The third-generation relies on state-space models. ABB has further refined this approach through the addition of Kalman filters, which are used to estimate the states of dynamic systems from a series of what might be “noisy” and incomplete process variable data. This technique has provided two advantages. First, regular historical dynamic process data is used to automatically build the state-space models. No open-loop step testing is required, provided that the historical data contains sufficiently dynamic process behavior. Second, the dynamic performance of the state-space model is markedly improved over the neural net models and linear multivariable models when process disturbances occur.
The rapidly increasing performance of computers has now made it possible to apply this highly advanced multivariable model predictive control (MPC) to power plant applications. Complex MPC solutions implemented at industrial power plants focus on coordinated control and optimization of multiple boilers, fuels, turbines, steam headers and power flows to and from the grid. These solutions have already shown a range of benefits such as improved plant stability and availability and lower overall energy costs.
The most common application of MPC for utility power plants has been combustion optimization. More recently, MPC-based solutions have been deployed in other areas, such as main and reheat temperature control and boiler-turbine coordination.
Defining Model Predictive Control
MPC is a common name for control technology using dynamic process models representing the relationship between independent variables (model inputs) and dependent variables (model outputs). Inputs include manipulated variables (MV) and disturbance feed-forward variables (FF). The outputs are called controlled variables (CV). The models predict future outputs based on past and future moves of manipulated variables, and the past values of the feed-forward variables.
A multi-variable MPC has inherent knowledge of the dynamic behavior and all the interactions between the process quantities and it controls all the variables at the same time. This is different from traditional control where each controller has one input and one output. With the many constraints and complex interactions involved in power plant control, multi-variable MPC is well suited to provide advanced control and optimization in the power industry.
In the past, various third-party software packages were used to implement MPC-based solutions for process industries and power plants. Firsthand experience with these packages revealed some shortcomings. One survey of industrial model predictive control technology considered the following typical deficiencies as serious:
- There are limitations in the choice of model types available. The commonly available impulse and step response models can be applied for inherently stable processes only and they handle integrating processes poorly.
- The controllers work poorly in presence of significant measurement noise or unmeasured disturbances.
- Model identification relies on open loop step testing, and only single-input/single- output (SISO) models can be identified.
As it became clear that the process control community had not taken advantage of the advances in modeling technology, ABB began work on a new MPC-based system. The new product, Predict&Control (P&C) was designed to address shortcomings of the existing solutions. It is based on new technology that replaces the typical MPC collection of single input/single output (SISO) step response models with a true multiple-input/multiple-output (MIMO) state space model.
ABB’s approach is made possible by a new algorithm that identifies accurate state space models from plant test data. The ability to identify MIMO models from a single set of closed-loop tests reduces the required testing time and greatly simplifies the modeling task.
The state space modeling approach permits P&C to use a Kalman filter for state estimation as part of the feedback control algorithm. The Kalman filter is a mathematical technique originally developed for trajectory estimation of spacecraft. One advantage to this particular controller technology is its ability to dynamically handle process noise, process variability and process drift over time. Older fixed-model technologies tout their ability to automatically relearn over time, but this is unnecessary with the P&C control technology.
The structure used by the P&C controller means that at each cycle time, the controller reads actual process variable values and uses the process inputs (u) and outputs (y) to estimate the current process state (Xˆ ), input disturbances (w) and output disturbances (h). This approach is different from standard MPC packages that can estimate the output disturbances only and it leads to better estimates of state (Xˆ ) and better control of y.
Predict&Control has two main parts: the offline Engineering Tool and the Runtime Controller. The Engineering Tool is used for application configuration, data pre-processing, model building, dynamic model identification, controller tuning, off-line simulation and analysis. The online portion consists of the runtime control server that executes the applications configured with the off-line tools. The runtime control server connects to any underlying distributed digital control system (DCS) by means of a standard OPC server.
The implementation steps for a model predictive control-based combustion optimization system include scope definition, basic configuration, tuning of base controls, dynamic testing, model identification, controller configuration, offline simulation and preliminary tuning, online controller commissioning and closed-loop testing and fine tuning.
In a power plant application, there are many possible model inputs (manipulated and disturbance feed-forward variables) and model outputs (controlled, constraint and additional state estimation process variables). Selecting the model scope depends on the project objectives, plant configuration and the specific local economic factors.
Implementation of an online realtime MPC-based COS requires developing process models through the use of advanced data capture and model generation tools. Implementation engineers perform the data capture activities during normal plant operations and during a few unit step tests within the normal range of plant operations. Model development is then done off-line at an engineering workstation, before downloading the model into the server which runs the actual MPC COS.
The server typically is connected to the plant DCS through either OPC or DCS specific communications interfaces. The preference is to use OPC with open control systems. A rack-mountable server is usually physically co-located with the DCS hardware, so no separate facility space is required.
The COS can be placed on and off line bumplessly during the model tuning process, so there is little to no impact on normal operations. Once the COS has been tuned to meet the plant’s stated objectives, it may be placed online for regular realtime control.
Multivariable controllers do not have an advisory mode and it is not a real-world use of an MPC to be used in this fashion. That would be like having a proportional integral derivative (PID) loop set in manual and “suggesting” to the operator what its individual control moves should be. In such a scenario, one could see operator soon becoming exhausted or overwhelmed by the minute-by-minute demands of controlling a few active loops. Imagine the exhaustion of an operator trying to follow realtime “advice” from an MPC controller that is simultaneously manipulating 15 to 20 individual setpoints or biases. It is clearly not a realistic situation to run an MPC-based COS in an advisory open-loop fashion.
Advanced MPC COS systems have established a track record for improving plant operations. Each of these improvements has an economic value associated with it. The remainder of this article will focus particularly on the value of NOX reduction, heat rate improvements, their particular sources as well as their values and some postulation about potential additional value associated with CO2 reduction.
Implementation and Use at Colstrip
An example of heat rate improvements through P&C controls included reheat spray valve demands as constraint variables, with a suitable target of preventing control output saturation. Because of the reduced variability through MPC advanced process control, the target could be lowered, which in turn produced higher overall unit efficiency.
The system was able to reduce reheat spray flow rates to one half of the original. The corresponding heat rate improvement was 0.36 percent, providing approximately 25,000 MWh of additional power each year with no increase in fuel consumption. At the same time, NOX emissions were reduced by 12 percent.
Finally, O2 bias and balances were implemented as a user configurable COS values. The balance was incorporated as part of the MPC control, whereas the O2 level bias merely reflected basic control system functionality. During operation, O2 imbalance was eliminated. Testing the O2 bias level indicated that O2 could be biased downwards by 15 percent and the COS would still maintain NOX, O2 imbalance and reheat spray flow reductions at their desired levels. While operations has not implemented long-term O2 reduction targets, the system demonstrated the ability to perform at target reduction levels.
One of the main tools for evaluating the operating cost-reductions which can result from a COS system is a proprietary COS evaluation spreadsheet. It is simple to use and allows an operations manager or performance engineer to quickly evaluate the system for typical operating cost reductions. This evaluation tool implements separate value calculations.
First, there is an evaluation of several boiler-related heat rate improvements which are typically affected by a COS. Second, there is a calculation for the value associated with NOX reduction. This is an annualized calculation and is based on the unit’s estimated cost to control NOX. While Colstrip is not currently under a NOX management mandate, the plant is anticipating the eventual tightening of NOX limits at the state and/or federal level; hence NOX control was central to the COS implementation.
COS systems have been implemented all four Colstrip Units, with the last unit completed in December 2008. The four COS’s were ordered as part of the four-unit control system upgrade program, where each unit has replaced its original complement of analog controls with a modern third-generation DCS. The COS’s are an integral part of the upgrades, but were designed to be a coordinating supervisory layer on top of the basic single-loop PID controls. In this way, the plant could operate with or without the COS online.
It typically takes 12 to 24 months for an operator staff to fully trust the DCS to do its job, especially under conditions of plant upset, such as loss of a boiler feedpump. Therefore, it is unrealistic to expect operations to initially place a high degree of faith in an MPC control scheme which essentially predicts the behavior of the process, before single loop controls are even getting the first inklings of process changes through their error signals.
During this time of familiarization and trust building, it has been found to be best to leave the COS off line. Additionally, the COS models will also need to be updated due to the installation of low-NOX burners on Units 3 & 4, which occurred after the COS models were built.
The nature of an MPC COS is such that operators need to let it exercise its supervisory control of the process, without intervention due to some sense of distrust of where the COS is taking the process. Only then will the plant be able to consistently reap the full benefits and short term payback available from the COS system. In the meantime, plant management can be confident that they have yet another layer of sophistication in their control equipment, which can be switched on as operational comfort increases and as environmental compliance becomes ever more critical.
Author: Richard Vesel is product manager, eBoP Energy Efficiency Power Generation, North America for ABB.