Adaptive regulation of super-heated steam temperature: a case study in an industrial boiler

Abstract

This paper describes the application of the MUSMAR predictive adaptive controller to the regulation of super heated steam temperature in a commercial boiler. The boiler considered produces $\text{150}\text{t}\text{/}\text{h}$ of steam at maximum load, used both for electric energy production in a turbine and industrial use. The combination of predictive and adaptive techniques, relying on multiple models redundantly estimated, allows a continuous adjustment of the controller tuning for tracking plant dynamics variations. This paper describes experiments actually performed on the plant with adaptive predictive control, in particular in the presence of load changes. A reduction of steam temperature fluctuations with respect to an optimized cascade of PI controllers is observed.

Introduction

In rehabilitation projects for large-scale power plant units it is natural to consider the inclusion of advanced control strategies as a mean for improving control performance (Soares, Gonçalves, Silva & Lemos, 1997). In this respect, the regulation of super heated steam temperature in boilers provides a relevant problem both from a research and economic point of view. From the standpoint of research in the application of adaptive predictive control, the plant considered for this case study has a number of features which render its control nontrivial. These include nonlinearities, time-varying dynamics, fast disturbances and, since the real plant is distributed parameter, unmodeled dynamics. Time-varying dynamics stems from two types of causes. Load variations induce changes in the dynamics of the plant. Furthermore, plant dynamics may also change due to unpredictable causes such as dust deposition in the furnace walls by flue gas, thereby affecting heat transfer coefficients.

The economic advantages of steam temperature regulation are recognized (Mann, 1992). The following argument, inspired in a classic example concerning paper manufacturing (Åstrom, 1970), provides part of the justification of economic improvement. Fig. 1 explains (part of ) the relationship between the quality of the regulation achieved by the controller and the economic performance of a thermoelectric group. This figure depicts the probability density functions (pdf ) of super heated steam temperature, under stationary conditions, for two different situations, marked A and B.

The economic performance of the turbine is proportional to the steam temperature. The higher the temperature of the steam leaving the final superheater (T_vsato in Fig. 1), the higher the economic performance. There is however a limit, T_max, imposed in the steam temperature due to the allowable operating conditions. If T_vsato exceeds T_max, damage can result. Since the regulation is not perfect, in order to prevent T_vsato from exceeding T_max, the set point of that variable is chosen at a value low enough such that the maximum of T_vsato is below T_max with an acceptable risk. This risk is defined as the probability of the steam temperature exceeding T_max and is given, for each situation A and B, by the area under the curve of the corresponding pdf, to the right of T_max. For the risk to be the same when the pdf is more spread out, the pdf mean value must be at a smaller value.

The two situations shown in Fig. 1, and labeled A and B, are obtained with different controllers. In situation A, regulation is more perfect than in situation B, with a correspondingly reduced dispersion of T_vsato. Thus, the set point in situation A, $\text{T}{}_{\mathrm{A}}{}^{\mathrm{\ast }}$, can be made higher with respect to the set point in situation B, $\text{T}{}_{\mathrm{B}}{}^{\mathrm{\ast }}$, without increasing the risk of violating the upper bound T_max i.e. the area under the corresponding pdf to the right of T_max is the same in A as in B. Thus, there is an increase of performance in situation A with respect to situation B, proportional to the difference of set points, $\text{T}{}_{\mathrm{A}}{}^{\mathrm{\ast }}\text{−T}{}_{\mathrm{B}}{}^{\mathrm{\ast }}$. It is remarked that, since performance is proportional to the temperature (i.e. since these variables are related by a straight line), under stationary conditions, the average performance only depends on the average of the temperature (assuming its pdf to be symmetric).

To sum up, a better controller reduces the dispersion of the steam temperature, allowing the use of higher set-point values. This provides increased average efficiency of the turbine.

The reduction of the variability of T_vsato is also beneficial from two other aspects, both related to economic performance. By reducing temperature fluctuations, mechanical stress causing microcracks is diminished, thereby increasing the useful life of the plant and decreasing maintenance costs. This may in itself be a motivation for control, as in the concept of life-extending control (Kallappa, Holmes & Ray, 1997). Furthermore, since the subsystems of a boiler are highly interactive, steam temperature fluctuations induce deviations of excess oxygen from its optimal value, thereby degrading boiler performance.

Due to the difficulties in performing experiments in industrial plants, experimental studies as the one described here are seldom found in the literature. In Gibbs, Weber and Porter (1991) a simulation study on the application of nonlinear model-based predictive (nonadaptive) control to fossil power plants is reported. The model used is a reduced-order nonlinear model relying on first principles that captures the dominant static and dynamic characteristics of the plant. For tackling mismatches with respect to the true plant structure, parameters are estimated off-line by prediction error methods or nonlinear least squares. This model is then used in a Kalman filter to estimate process states in real time. Estimated states are in turn used for prediction, enabling the computation of the optimal multivariable control sequence. This refers to the overall plant variables and not only to super-heated steam temperature. The results of this simulation study are reported as very encouraging, the only serious disadvantage being the effort required to develop a simplified model of a specific plant.

Other simulation studies concerning the application of predictive and/or adaptive control to boilers are described in Yang and Hogg (1992), Manayathara, Bentsman, Pellegrinetti and Blauwkamp (1994), Oda, Toyoda and Nakamura (1995), Prasad, Swidenbank and Hogg (1998a). In particular Prasad et al. (1988a) uses a nonlinear long-range predictive controller based on a neural network model to control main steam temperature and pressure and to reheat steam temperature under several operating conditions. Several simulation studies have been performed concerning variables other than steam temperature and using different control strategies. In Prasad, Swidenbank and Hogg (1998b) it is shown how to configure a local model network-based multivariable long-range predictive control strategy for thermal power plants using a physically based model.

In Fessl and Jarkovský (1988) an LQ self-tuning controller minimizing a multi-step quadratic control performance index was applied to replace the main PI in a cascade structure for super-heated steam temperature control on a 200 MW coal-fueled power plant. When connected to the actual plant, favorable control results with respect to the standard analog PI controller were obtained, but some problems remained open such as the influence of the inner loop dynamics, nonlinearities and the best choice of the sampling interval.

In Unbehauen, Keuchel and KOcaarslan (1991) an adaptive model reference controller for enthalpy and electrical power has been developed and applied to a 750 MW once-through boiler. Experimental results are reported to exhibit the advantages of adaptive control under all operating conditions, which reflects in increased efficiency and uniform steam quality. The adaptive controller was implemented in parallel to the conventional controller, its output being limited if it differs by more than 10% in magnitude from the signals of the conventional one.

The difficulty of maintaining boiler efficiency in a thermal power station within a wide range of situations, due in particular to load changes, is the motivation of Pérez, Perez, Cerezo and Catediano (1997) for considering the application of adaptive predictive control to a 214 MW unit. The predictive–adaptive algorithm used is not described, results being presented for a number of relevant variables. The performance achieved is compared with Electric Power Research Institute (EPRI) ideal control, almost all these specifications being satisfied.

In Matsumura, Ogata, Fujii and Shioya (1998) an adaptive controller was applied to the main and reheat steam temperatures. This resulted in reduced temperature excursions caused by load changes, especially in the reheat steam temperature.

In this paper, a predictive adaptive strategy is considered for the regulation of the super-heated steam temperature in an industrial boiler with drum. The boiler considered nominally produces $\text{150}\text{t}\text{/}\text{h}$ of steam, used both for electric energy production in a turbine and industrial use. Due to unpredictable changing demand from industrial users, there can be fast fluctuations in steam flow, which can be as low as $\text{100}\text{t}\text{/}\text{h}$. This, as well as other factors, such as the way in which the variables defining combustion conditions are operated, induce changes in both static and dynamic plant behavior. Adaptive control was then considered as an alternative to circumvent the problems referred to in Gibbs et al. (1991). This provides a controller which not only tackles plant changes, but may also be used in different plants, with minor reconfiguration.

The predictive adaptive multivariable multistep adaptive regulator (MUSMAR) control algorithm (Greco, Menga, Mosca & Zappa, 1984), incorporating feed-forward terms (Lemos, 1991; Coito, Lemos, Silva & Mosca, 1997), was selected. This controller is based on a number of separately estimated predictive models. In the presence of plant/model mismatches, such as the situation found here, the redundancy thereby introduced proves important for achieving a correct control action (Mosca, Zappa & Lemos, 1989, Coito et al., 1997). This multiple model approach is a distinctive feature with respect to other approaches to similar problems, relying on the adaptation of a single model from which others are then obtained.

It is stressed that, although the algorithm relies on a set of multiple predictive models, these are not used one at a time, e.g. depending on the operating conditions, or as in Morse (1996). Instead, these models are combined together at each sampling time to yield the controller gains. Adaptation results from the continuous identification of the model parameters. This scheme has a number of interesting properties.

In Greco et al. (1984) it was shown that MUSMAR is equivalent to a bank of parallel self-tuners, each one tuned to a different value of plant delay and with different weights. If the actual plant delay is bigger than the delay assumed for a given self-tuning channel, then the corresponding weight will be zero. Insensivity to uncertainty in plant delay is thus achieved to some degree. Clearly, the predictive horizon should be bigger than the plant input/output transport delay.

In Mosca et al. (1989) (see also Coito et al., 1997) for an analysis of a simplified situation involving feedforward action) it was shown that, if the prediction horizon is high enough, in the neighborhood of a cost extremum, the controller gains will be updated in a direction close to that opposite of the gradient of the cost modified by the inverse of the Hessian matrix. As a result, it was shown that the only possible convergence points of the algorithm are the local minima of the steady-state quadratic cost constrained to the chosen controller structure. Extensive simulation results show that the algorithm is actually able to find these minima, even in the presence of unmodeled plant dynamics.

Furthermore, when the gains converge, it was shown in Mosca and Zappa (1989) that, even for ARMAX plants, the set of models on which MUSMAR relies, correctly predicts the plant behavior in least-squares sense, provided they are separately estimated.

It is not the aim of this paper to make performance comparisons with other types of adaptive controllers. In this respect, a comprehensive discussion concerning structural aspects is found in Mosca (1995). It is apparent from the previous remarks that MUSMAR is a tool for finding local minima of the linear stochastic cost. Considering the main motivation here, i.e. the modelling of the fluctuations of superheated steam temperature from a stochastic point of view, their variance being directly related with economic performance, MUSMAR naturally appears as an appropriate tool.

Of course in a commercial/final product, aspects other than control gain tuning ought to be considered. These include global control for tackling loop interactions, fast controller reconfiguration in response to unexpected faults or command signal handling to tackle constraints. However, the problem addressed in this paper is considered important enough to merit attention by itself. Also from a practical point of view, the importance of having a tool available for retuning the controller “at the touch of a button” was recognized by the personnel operating the plant.

The main contribution of the paper is the experimental application in a real industrial site, with the emphasis on controller gain tuning and its relation with plant performance. The paper describes experiments actually performed on the plant, including a comparison with the standard control system, the response to load reduction and fast load changes and a study of the dependence of input and output variances on the weight on the valve variance in the cost functional. The set of experiments performed show substantial improvements with respect to the performance obtained with standard controllers obtained by heuristic methods. It also shows the importance of including feedforward from accessible disturbances (air and steam flow).