4. The application to the control and optimization of processes for energy production
The main social requirements about the management of energy plants are focussed on the maximization of the energy efficiency and minimization of environment impact (particularly as regards the reduction of NOx and CO emissions). In this context the process control assumes an importance very relevant in respect to the past, especially for the combustion plants where the pollutants emissions are strictly related to the modality of the process management. Also the complexity of this function is widely increased because of it has to take into account many targets, like economic management, low environmental impact, plant stability and design constraints, energy efficiency. These aspects constitute a strong stimulus to develop more advanced strategies for the process control and optimization.
Today, the methodologies for advanced control and optimization (expert systems, ARM or neural predictors, process on-line simulators) are surely useful for a wide fraction of the industrial requirements, but they have serious limitations in many real field applications. These limitations could be summarized in the following items:
Rarely these requirements are fulfilled for industrial combustion plants like waste incinerators or chambers for gas turbines. One of the most serious problems for some innovative methods based on learning (like neural or fuzzy control, [17]) is that they are based on fixed optimization rules and do not take into account the evolution of the plant during its life (i.e. not controlled variables or constraints). The learning phase is generally difficult for data lacking and the development activities for process optimization require deep knowledge of the specific process. Generally, these methodologies are not extensible to other processes.
New advances in the contexts of chaos theory [31,32] (particularly in the analysis of real chaotic systems and complexity), stimulate the research in this direction in order to explore new approaches to the process control. In particular, nonlinear dynamics allows the possibility to describe the state of the process (and therefore the related performance like efficiency, emissions, etc.) on the basis of the characterization of its dynamics. The recent developments in the field of nonlinear data analysis (dynamics invariants), make it more sensible and robust [30]). The complexity description opens the possibility of the development of a continuos learning during the plant life and the continuous redefinition of the optimization strategy. In spite all these promising scientific developments, at present only few studies have been done in order to apply them to the plant optimization and control [35].
We would emphasize that the term optimization here is utilized in the sense of a continuous on-line adaptation of the management of an existing plant. The goal is to drive the process towards the optimal compromise between the management targets deriving the rules directly from the measurements.
The ideas proposed in this paper are aimed to developing a new approach to the optimization and control of complex processes for energy production/consumption. This methodology is based on evolutionary optimization and it started from some successful experiences in the dynamic characterization (for diagnostics and control) for at least two industrial applications (oil field diagnostics and combustion dynamic characterization) [33, 34, 36]. Furthermore an optimization study has shown very interesting features of artificial life environment with respect to more classical genetics techniques.
The basic features of the proposed approach are:
The essence of this approach could be synthesized by the following sentence: "not control rules but autonomous structures able to generate optimized-control rules".
The main processes we are looking for application of the evolutionary control in the context of combustion plants are:
The basic idea for the evolutionary control
The basic idea consists in the reversal of the concept of the expert systems (ES). In the construction of the ES, the knowledge of the operators is verbally transferred to the ES builder. In our proposal, the process knowledge is not verbally transferred, but it is developed directly by the system through the measurements observation. The driving process is the dynamic building of a model on the basis of the observation of the effects that the regulation actions (acted by the operators or any other existing control systems) have on the plant performance.
The real implementation of this idea consists in the realization of a system which receives measurements from the plant and activates an elaborate process based on the two following main steps (see the scheme of fig. 1 for a resume of the main components).
Fig. 1: Scheme of the evolutionary control
The Dynamic State Identification
The plant is monitored with process measurements (values of process variables averaged on the time interval which defines the period of the plant monitoring) and dynamic measurements (sensors with dynamic response following the process dynamic fluctuation). The dynamic measurements are elaborated on the time interval and the chaos invariants are computed. These discriminants describe the plant state. The plant state is identified by the dynamic behavior which is determined by the dynamical system and the parameters values at which the plant is operating.
In 1996 Annunziato and Abarbanel [33] outlined a new methodology for the classification problems based on the attractor morphology using few discriminant parameters. This methodology has been successfully applied to the identification of the multiphase flow regime in oil production plants [33], to the characterization of combustion chambers for gas turbines, to the characterization of combustibles pollutants in conventional chambers [36], and finally to the identification of working state of waste incinerators [34].
(Movie) An example of 3D attractor of a real flame in a combustion chamber.
The basic idea is to describe the attractor morphology that changes strongly depending on the operative conditions of the process. The non linear dynamics allows a higher sensibility in respect to the linear statistics (i.e. FFT). Through the computation a series of "moments of inertia" for the attractor, extending in order and dimensions, we build a series of shape descriptors, named dynamic moments. These moments are computed via which specify certain points or axes or planes with respect to which the distances to every point of the attractor are computed. For a detailed description of the dynamic discriminants see the above mentioned references.
For the mentioned applications the results obtained in the classification were better than that ones corresponding to the FFT analysis or to the more classical chaotic invariants (fractal dimension, lyapunov coefficients [30]).
The alife environment for control and progressive optimization
An artificial environment is created in parallel to the plant capturing information from the measurements. In this artificial environment we place the live individuals which represent the experimented/observed plant conditions. Furthermore, some other individuals are generated through a genetic reproduction mechanism. Each individual is defined by its genotype which includes the plant state description (derived by the measurements), the process performance (computed on the basis of the measurements), and the regulations state. An evolutionary mechanism selects continuously the plant conditions (individuals) which corresponds to the best performances. An aging mechanism is provided in order to take into account the plant real evolution.
In respect to the environment presented for the TSP optimization we use short filaments instead point in order to increase the probability ot meetings between the individuals. During an evolution cycle, the filament can growth on the head and decrease on the tail. We developed several models for the probability of reducing the filament length. Some models are related to a most probable length or to a fitness of the individual or finally to the filament age. The reduction of the filament length frees space and the life is continuously sustained.
(Movie) Two movie examples of the artificial society dynamics.
An evolutionary mechanism selects continuously the plant conditions (individuals) which correspond to the best performances. This mechanism is based on the emulation of the natural selection. The interaction model is constituted by the competition on the basis of the individual fitness (plant performances). In this way the individual with the highest fitness can survive and reproduce. The reproduction is asexual and the sons have small mutations on the regulation state.
Because the individual has an "expected life-time", very old individuals can die according to a probabilistic model. This aging mechanism is very important to warrant the possibility to lose memory of very old solutions and follow the plant evolution. This mechanism takes into effect the aging of the optimization models due to the changes of not monitored variables (i.e. combustibles in the waste incinerators or modifications introduced during the life of the plant).
The selection mechanisms warrant the selection of the individuals (plant configuration) which have produced the highest fitness (best plant performances). In this sense the Alife environment is a good evaluator of the regulation actions of the operators mixing all the operators actions, judging the single control action in terms of positive or negative effects on plant performances and building the "optimal operator".
The distinguishing feature of the proposed system is that the mechanism of the mutations introduces new regulation configurations never visited before. Therefore the environment has the possibility to generate and evaluate plant configurations completely new with respect to ones explored by the operators. Fig.2 summarize the optimization and control modules.
Fig. 2: The Artificial Life Environment
The plant performances and the process control
For each time period, the process variables are processed in order to compute the performance of the plant (the fitness function) in the current measurement time frame using a fuzzy logic approach. The evaluation of the quality of a plant configuration is made through membership functions applied to the process measurements (pollutants, efficiency, design constraints). The fusion of these functions is obtained through fuzzy operators and it represents the process performance (individual fitness).
In order to evaluate the plant performance in configurations (individuals) never visited before we have developed a "performance map" model. This model is able to evaluate the differences in the process performances induced by a control action (a change in the regulation state) starting from a specific dynamic state.
The dynamic invariants, the regulations actions, and the performance evaluation continuously update a performance map built using a neural network (based on a Radial Basis Function approach). This map gives the possibility to estimate the performance differences induced by regulation actions. Compatibly with the statistical accuracy reached by the performance map, the best individual is taken at each time period as the system suggestion for the regulation actions. The suggestion is sent to filters (rule based) which take into account the compatibility of the suggested regulation actions with the design constraints or stability constraints.
At the beginning, the system is not able to give suggestion but it only learns from the plant measurements. The artificial environment starts to become active and gives its suggestions when the performance map is quite filled. After each cycle of measurements/suggestions the performance map is updated (continuos learning), and new individuals are inserted in the artificial environment. In this way the system follows the plant not-monitored changes and drives the evolution towards better performance.
The real flames and the flame simulator
We have carried out several recordings in real plants. In this movie we show an example of the monitoring of the flames in a waste incinerator of the Ferrara (Italy - AGEA) town. The video camera records the flames image at 800 frame/sec. In the movie a speed/resolution reduction have carried out and a pseudocolor table has been applied in order to enhance the flame temperature distribution contrast.
(Movie) Real flames in an industrial incinerator (800 frames/sec, pseudocolors)
In parallel to the real processes experimentation, we have substituted the real process with a software simulator in order to study the best strategy for the control/optimization performed by the alife environment. The simulator is based on a mathematical model used for the flame front modeling: the Kuramoto-Sivashinsky model [34]. The goal is to obtain a software demonstration of the optimization features of the alife environment.
(Movie) example of simulated flames with the Kuramoto-Sivashinsky Model.
The model, includes regulation parameters which influence the flame dynamics (the dynamic state).
For each configuration we compute a process performance on the basis of a model simulating the pollutants emissions and the energy efficiency. In addition to the regulation parameters we have included some disturbance parameters which represent the not controlled variables or the process aging.
Two experiments are in phase of realization:
This specific research is partially supported by MURST (Italian Minister of University, Research and Technological Development) for the research in Italy and DOE (Department of Energy) for the research in USA.
The approach is general (not problem-dependent) and can be applied to different fields of the energy production (we are launching several other EU projects for oil industry and combustion processes), such as in the telecommunications, in the applications for the internet network and in the transportation control.
