By Béla Lipták
Globally, more than 80 million barrels of crude oil are refined daily. The amount of energy used for distillation is approximately 8% of the total energy used in the industrial sector of the United States. American refineries spend 50 to 60 percent of their operating costs (i.e., excluding capital costs and depreciation) on energy, in Venezuela, Russia and the Middle East that number is even higher, about 70%. Compared to refineries, the much more advanced chemical industry spends only 30 to 40%. This difference shows the savings that can be obtained just by implementing the state-of-the-art controls and optimization described in this series.
A model-based, multivariable control system requires the development of dynamic models based upon fractionator testing and data collection. These dynamic models predict future fractionator performance based on the memorized past performance. Model-based multivariable control can be valuable for towers that are complex, subject to many constraints and experience severe interactions.
In the case of well-understood processes, such as distillation, the material and heat balance relationships are combined into this overall dynamic model. Once a process model has been established, the controller is simply the inverse of that model. In this sense, the PID controller is a linear inverse model of a single loop. A simple internal model-based controller (IMC) is shown at the bottom of Figure 1. It has the same structure as the Smith predictor described in the first part of this series of articles, which consists of a first-order system with dead time in combination with a PI controller. (See Part 1, Control, November, 2006, p. 16)
In this case, the two towers have two products and an impurity stream. The objective is to control the composition of both products and therefore, when an adjustment is made in response to an upset in the composition of one, that manipulation tends to cause a disturbance in the operation of the other composition loop. In this example, the MVC controls the two product compositions, while keeping the operation within the constraints of the process equipment capacities and makes its corrections while considering the dead time introduced by the stripper.
In this process, the feed flow rate is the main disturbance variable. The steam to the first column and the temperature at the top of that column are the manipulated variables. A constraint variable is the internal reflux flow, which is calculated from the tower temperatures and flows.
Dynamic Matrix Control (DMC) is another MVC technique using a set of linear differential equations to describe the process. The DMC method obtains its data from the response of the column to upsets, and after the model is so obtained, calculates the required manipulations utilizing an inverse model. Coefficients for the linear equations describing the process dynamics are determined by actual testing. During these tests, manipulated and load variables are perturbed and their dynamic responses are observed. This identification procedure is time-consuming and requires substantial local expertise.
Artificial Neural Networks (ANN)
Figure 2 shows a three-layer, back-propagation ANN, which predicts the manipulated steam and reflux flows of a column. The process model is stored by the way its processing elements (nodes) are connected and by the importance that is assigned to each node (weight). The ANN is trained by example and, therefore, it contains the adaptive mechanism for learning from examples. During the training of these networks, the weights are adjusted until the output of the ANN matches that of the real process. Naturally, when process conditions change, the network requires retraining. The hidden layers help the network to generalize and even to memorize.
In the single-input/single-output (SISO) configuration, the ANN network builds an internal nonlinear model relating the controlled and manipulated variables. It builds this model by learning or training, based on a data set of past measurements and process responses. This makes the neural controller more useful and more robust than the standard PID.
Because the neural network paradigm can accommodate multiple inputs and outputs, an entire fractionator model can be built into a single controller. The neural controller can be thought of in the same terms as model-based control algorithms, whereby the neural network is used to obtain the inverse of the process model. As shown on the top right of Figure 2, the back-propagation network can be trained to behave as an inverse model of the process, with load and controlled variables being input and output vectors.
To build such a model, all inputs and outputs must be normalized based upon expected minimum and maximum values and presented to the network during training. By using such historical data, the network is trained and a nonlinear internal model is created. The networks ability to do the prediction of the dynamics of the fractionator improves as more data become available for training. Thus, the neural controller is a type of nonlinear, multivariable, model-based control algorithm. The difference is,that, instead of creating the nonlinear process model with explicit equations, the neural controller builds its own process model based on the actual operation of the tower.