Fuzzy model based predictive control of chemical processes.

The past few years have witnessed a rapid growth in the use of fuzzy logic
controllers for the control of processes which are complex and ill-defined. These
control systems, inspired by the approximate reasoning capabilities of humans
under conditions of uncertainty and imprecision, consist of linguistic 'if-then' rules
which depend on fuzzy set theory for representation and evaluation using
computers. Even though the fuzzy rules can be built from purely heuristic
knowledge such as a human operator's control strategy, a number of difficulties
face the designer of such systems. For any reasonably complex chemical process,
the number of rules required to ensure adequate control in all operating regions
may be extremely large. Eliciting all of these rules and ensuring their consistency
and completeness can be a daunting task.
An alternative to modelling the operator's response is to model the process
and then to incorporate the process model into some sort of model-based control
scheme. The concept of Model Based Predictive Control (MB PC) has been
heralded as one of the most significant control developments in recent years. It is
now widely used in the chemical and petrochemical industry and it continues to
attract a considerable amount of research. Its popularity can be attributed to its
many remarkable features and its open methodology. The wide range of choice of
model structures, prediction horizon and optimisation criteria allows the control
designer to easily tailor MBPC to his application. Features sought from such
controllers include better performance, ease of tuning, greater robustness, ability
to handle process constraints, dead time compensation and the ability to control
nonminimum phase and open loop unstable processes. The concept of MBPC is
not restricted to single-input single-output (SISO) processes. Feedforward action
can be introduced easily for compensation of measurable disturbances and the use
of state-space model formulation allows the approach to be generalised easily to
multi-input multi-output (MIMO) systems. Although many different MBPC schemes have emerged, linear process models derived from input-output data are
often used either explicitly to predict future process behaviour and/or implicitly to
calculate the control action even though many chemical processes exhibit
nonlinear process behaviour. It is well-recognised that the inherent nonlinearity of
many chemical processes presents a challenging control problem, especially where
quality and/or economic performance are important demands.
In this thesis, MBPC is incorporated into a nonlinear fuzzy modelling
framework. Even though a control algorithm based on a 1-step ahead predictive
control strategy has initially been examined, subsequent studies focus on
determining the optimal controller output using a long-range predictive control
strategy. The fuzzy modelling method proposed by Takagi and Sugeno has been
used throughout the thesis. This modelling method uses fuzzy inference to
combine the outputs of a number of auto-regressive linear sub-models to construct
an overall nonlinear process model. The method provides a more compact model
(hence requiring less computations) than fuzzy modelling methods using relational
arrays. It also provides an improvement in modelling accuracy and effectively
overcomes the problems arising from incomplete models that characterise
relational fuzzy models.
Difficulties in using traditional cost function and optimisation techniques
with fuzzy models have led other researchers to use numerical search techniques
for determining the controller output. The emphasis in this thesis has been on
computationally efficient analytically derived control algorithms. The performance
of the proposed control system is examined using simulations of the liquid level in
a tank, a continuous stirred tank reactor (CSTR) system, a binary distillation
column and a forced circulation evaporator system. The results demonstrate the
ability of the proposed system to outperform more traditional control systems. The
results also show that inspite of the greatly reduced computational requirement of
our proposed controller, it is possible to equal or better the performance of some
of the other fuzzy model based control systems that have been proposed in the
literature.
It is also shown in this thesis that the proposed control algorithm can be
easily extended to address the requirements of time-varying processes and
processes requiring compensation for disturbance inputs and dead times. The
application of the control system to multivariable processes and the ability to
incorporate explicit constraints in the optimisation process are also demonstrated.