COMBINING HUMAN INTELLIGENCE AND OPTIMIZATION

0 downloads 0 Views 278KB Size Report
Optimization techniques and expert systems are commonly used in power ... system and optimization techniques, we proposed a new approach based on the ...
COMBINING HUMAN INTELLIGENCE AND OPTIMIZATION TECHNIQUES IN POWER DISTRIBUTION SYSTEM PLANNING Yixin Yu*

Gang Duan**

ZhaoYang Dong***

* School of Electrical Automation & Energy Engineering, Tianjin University, Tianjin 300072, China ([email protected]) ** Department of Electrical Engineering, Tsinghua University, Beijing 100084, China ([email protected]) *** School of Information Technology and Electrical Engineering, The Uniersity of Queensland, QLD 4072, Australia ([email protected]) Abstract Power systems are large scale systems with huge number of interacting components and dimensions. Power system planning problems are highly nonlinear, dynamic, complex, and constrained optimization problems. Optimization techniques and expert systems are commonly used in power system planning, however, there are always some factors which can not be quantified, modeled, or even expressed by conventional expert systems; and the large scale and nonlinearity often mean poor performance with conventional optimization techniques. In order to overcome these problems with conventional expert system and optimization techniques, we proposed a new approach based on the power of human intelligence and cognitive genetic algorithms. The collaborative decision making approach was realized by a cognitive feedback mechanism, which explores the strength of human knowledge approach in decision making and genetic algorithms approach. This method is also capable of overcoming the weakness of each individual approaches when they are applied to decision making process alone. The effectiveness of the proposed method was tested in a distribution system planning problem.

1. INTRODUCTION Power distribution planning problems are multi-objective and multi-constraint optimization problems. However, the results from most conventional optimization techniques including deterministic [1] and un-deterministic [2] techniques, or mathematical and heuristic techniques (including expert system)[3], usually can not be applied directly to engineering practice for power distribution system planning until being modified by human experts. This is a result of the fact that there are always some factors which can not be quantified, modeled, or even expressed by expert systems. These factors include vague ideas and sub-consciousness of human experts, policies of governments or companies, and some preferences which are impossible to be clearly expressed. There are optimization algorithms which do can handle large scale optimization problems. These algorithms include genetic algorithms (GAs) and simulation annealing (SA), however, the computational costs are still very high. In this paper, we proposed an efficient decision making approach aimed to overcome these problems with conventional optimization algorithms. The proposed approach combines mathematical algorithms with human intelligence, and is able to achieve effective global optimization in complex engineering planning and optimization problems.

The paper is organized as following. First, the advantages and disadvantages of human intelligence in decision making were analyzed; subsequently, we concluded that humans can join decision making progresses by cognitive feedback. Based on human-cognitive feedback and genetic algorithms, we proposed the cognitive genetic algorithm. Through the human-machine interactive process, the algorithm can clarify and extract human experts’ cognition that is difficult to be expressed by computer programs. Such cognition includes human expert’s sub-consciousness, wills and vague ideas. Finally, a practical power distribution system planning problem is studied with the proposed algorithm. It showed that this practical decision method with cognitive feedback, cognitive genetic algorithm and mathematical algorithms is efficient and effective for power distribution system planning. 2. THE ROLE OF HUMANS IN OPTIMIZATION As mentioned in Section 1, NP hard problem and complex engineering problems cannot be effectively solved by computers alone. Humans have the most advanced intelligence in the world. Some of the knowledge is out of the computers’ capability, for example, thinking in images, intuitive, experiences, common knowledge, adaptability, emotion, social values, unlimited creativity, thinking in multi-dimension and integrity. These properties of humans’ intelligence are the disadvantage of computers so far. Of

course, compared with computers’ intelligence, human intelligence also has the following inferiority: (1) Humans can handle only limited amounts of information at the same time; (2) Humans tend to avoid complex processes that require many calculations; (3) Humans can not project problem elements very accurately in time and space; and (4) Human memory is often not reliable; (5) Humans often do not have enough time to perform complex processes involving a series of elementary reasoning steps. Apparently, hybridizing human’s intelligence with computer’s intelligence is a good way to solve complex engineering problems, such as power system planning problems. Human intelligence can join computer optimization process indirectly or directly. For example, in expert system based methods, human experience and learning capability are stored in the knowledge database. Information stored in this database can be used to guide optimization or decision making process via inference machine. This is the in-direct joining of human expertise into optimization process, because the whole process is realized by computers after human intelligence has been stored in computers. However, some human intelligence, as mentioned earlier, cannot be stored and realized by computers, so direct joining of human intelligence in optimization process is also necessary. This paper will discuss how to realize the direct joining. 3. COGNITIVE FEEDBACK The participation of humans in optimization can solve problems that mathematical optimization methods or computers cannot solve. However, because the some of the optimization problems are very complex and the capability and knowledge of planners or experts are limited, there are usually limitations to human cognitive decisions. Moreover, planners or experts usually cannot overcome the limitations, or even errors, in their decisions by themselves. Studies [4] have shown that cognitive feedback can improve both the process and the quality of decision making. Cognitive feedback is an important concept in cognitive science [5], which refers to the research of intelligence and intelligence systems [6]. Intelligence is closely related to adaptivity − in terms of problem solving, learning, and evolution. As intelligence is bounded by limitations of human short-term memory and processing capacity, a good intelligence system is the one that can adapt to new context as it evolves with its users over time. Therefore, cognitive feedback can be defined as information provided to decision-makers to offer them a better understanding of decision processes. This information includes relations in the decision environment, relations perceived by the decision-maker about that environment, and relations between the environment and the decision-maker. This feedback is able to help the decision-maker for better understanding of the task structure, his/her own cognitive systems, and the mapping and matching between the two

[7]. Consequently, human capability of modeling and problem solving can be improved by cognitive feedback. How to make good use of cognitive feedback in computer-based decision support system is the key point of our research. In a practical decision support system based on cognitive feedback, there should be at least two decision-makers or experts in order ensure objectiveness in decision process from human expertise. Opinions from different decision-makers not only discover the different aspects of the planning problem, but also consolidate their common understanding. In such a system, planners and experts are not only decision-makers, but also learners. According to the different objectives participating in cognitive feedback, cognitive feedback is classified as individual feedback, interpersonal feedback and collective feedback [4]. In this paper, we propose a new feedback that is feedback between humans and mathematical algorithms, as detailed in the sequel with emphasizes on power distribution planning problems. (1) Feedback between humans and mathematical algorithms: A great amount of numerical calculation is involved in power distribution system planning. In the beginning phase, planners can only give a qualitative description or a very simple scheme of the expected optimization solution. No detailed information can be presented by planners. After the calculation of mathematical optimization methods, more detailed schemes can be presented by computers. Inspired or guided by these more detailed schemes, planners can give further instructions to computer programs or directly revise the scheme according to their new cognition to the planning scheme. The feedback is executed iteratively, such that more detailed and reasonable planning schemes can be obtained successively. It is obvious that the schemes cannot be obtained by planners or by mathematical algorithms alone, if there is no such feedback. (2) Individual feedback: Through observing the effect of decisions of themselves, through feedback, planners can have deeper understanding of their decisions. For example, the cognitive feedback can provide the planners with continuing consistency analyses of the performance of the decision-making processes and outcomes. Then planners can control or revise their decisions processes. (3) Interpersonal feedback: By interaction with other planners, they become observers of their own thinking. As indicated in [8], learning the models of others compels rethinking of one’s own model, either by adaptation (i.e. revision of existing model) or generation (i.e. creation of a new model). (4) Collective feedback: This feedback aims to foster team learning to help building a shared vision. A shared vision is the one that each member can identify with and is committed to. Team learning starts with dialogue, during which members stop

making assumptions and engage in a collective thinking mode [8]. Multiple individual models bring multiple perspectives to bear, and more important, bring a wholistic view. To facilitate the cognitive feedback, information technology and multimedia technology should be explored in the development of the decision support system for planning problems. 4. COGNITIVE GENETIC ALGORITHM To make good use of the advantages of human intelligence and mathematical algorithms while avoiding their disadvantages, we proposed cognitive genetic algorithm. Using this algorithm, feedback between human and mathematical algorithms can be realized and practical optimal solution can be obtained. Its theoretical foundation is as follows: Humans often make decisions according to their intuitive and experience, however, they can not always explain clearly the reason of doing so, and can not extract the factors which determine their decisions. Therefore, these capabilities cannot be realized by computers, and they cannot be used to guide optimization processes directly. Genetic algorithms [9] have the capability to find the implicit laws in the compounds of genes automatically based on the mechanism of natural evolution. They can be used to guide optimization processes. In summarize, cognitive genetic algorithm can be defined as an iterative process between humans and computers, in which genetic algorithm can find the laws of humans’ cognition to things. The key characteristics of cognitive genetic algorithm are as following: (1) The method of encoding must be able to generate all of the possible solutions. (2) All the solutions represented by genes should be presented to planners or experts vividly so that they can think and make judgements easily. For example, tables, graphs, diagrams, or figures often can be used to show a (temporary) solution. (3) Planners or experts can rank the solutions (individuals) generated by genetic algorithm by their preference. Then, they can transmit their viewpoints to genetic algorithm by giving different fitness to different individuals, so that genetic algorithm can perform genetic operation. The process is performed iteratively, such that the algorithm can clarify and extract the humans’ cognition step by step and guide the optimization process through. Finally, optimal solutions that satisfie human experts can be obtained. Human cognition includes vague ideas and subconsciousness of human experts, policies of governments or companies, and some preference that cannot be expressed clearly. These factors cannot be quantified, modeled, or even expressed by expert systems properly, and they are difficult to be expressed by computer programs, but they are important to guide the planners in decision making process. Cognitive genetic algorithms are the way to make good use of these

resources of human intelligence and computational power. Apparently, the disadvantage of cognitive genetic algorithm is that the optimal result may change with different planners or experts, that is to say, more subjective factors are involved in the decision process. Although this cannot be totally avoided, it can be limited reasonably by the following methods to avoid unreasonable cognitive interfere. (1) Any objectives or constraints should be considered first if they can be realized by feasible mathematical optimization methods. Only when this is not possible, should expert system based decision methods be explored. Optimization by humans’ cognition should be the last resort. (2) Cognitive feedback, including feedback between human and mathematical algorithms, interpersonal feedback and collective feedback, can affect the cognition of planners and experts, and therefore revise or limit their unreasonable or unfeasible opinions. 5.

POWER DISTRIBUTION SYSTEMS PLANNING BASED ON COGNITIVE FEEDBACK AND COGNITIVE GENETIC ALGORITHM By hybridizing cognitive feedback, cognitive genetic algorithm and mathematical algorithm, a practical decision method for power distribution and transmission systems planning is proposed as following: Step 1: After an introduction to a specific distribution systems planning problem to the involved experts, a questionnaire should be given to them as well. They shall be asked to list as many factors or variables as possible that may be relevant to the appropriate planning decision making. Usually the following factors are considered: deficiency of present system, security and reliability, capacity margin, ease of operation, operation and maintain cost, construction cost, environmental impact, impact on regional and national economic growth, and policy of company or government. Every expert should complete his/her questionnaire individually to avoid being impacted by other experts, especially those with authority. This ensures comprehensive and objective information be obtained. Step 2: Collect the response, cluster them based on their similarities, and conclude the main factors and main considerations. Step 3: Each expert is given a questionnaire again and asked to re-evaluate the above selected main factors and main considerations, for example, giving the relative weight of these factors and their interaction. If an expert does not agree on the selected main factors and main considerations, he/she should give the detailed explanation. In this phase, individual judgement is still necessary, and interpersonal feedback will happen. Step 4: All experts are invited to join in a face-to-face discussion. Differences of opinions shall be analyzed. Then some opinions are abandoned, or compromises are made.

Finally, agreements about the objectives, constraints, their weights and interactions are obtained. Collective feedback will happen in this step. Step 5: Select different optimization methods for different kinds of objectives and constraints. Mathematical optimization including deterministic and undeterministic approaches are the first selection if the objectives and constraints can be expressed by mathematical models. These objectives and constraints usually include operation and construction cost, voltage constraints, capacity constraints, radiality constraints, power balance, and line type constraints. Mathematical optimizations with bearable computing burden had been found. The second selection is expert system based method. Objectives and constraints, such as special type requirement of elements and special route requirement, which cannot be realized by mathematical methods, usually can be added in a rule base. They can then be realized in optimization process by expert system (e.g. “if-then” type of reasoning) which can be executed by computers automatically. The remaining objectives and constraints such as those listed in section 1 are remembered or learned by decision-makers or planners, and they will reflect these objectives and constraints in optimization process by cognitive genetic algorithm. It should be noted that in this step, experts’ cognition capability should be exploited to reduce the scale of the optimization problem. For example, usually by experience, intuitive judgements and simple estimates experts can limit the types of transformers and lines to several most possible ones, limit the service area of feeders and transformer substations to a reasonable level, and abandon the impossible right of ways and candidate elements. Step 6: Design the encoding algorithm, and produce the first generation of individuals randomly. It is necessary that all feasible solutions can be expressed by the encoding algorithm. Step 7: Divide the entire individual group of the generation into several subgroups. Step 8: Take the individuals in the subgroups as initial solution, and employ appropriate mathematical optimization methods or mathematical optimization methods hybridizing with expert system based methods to get the corresponding local optimal solutions. Different methods can be used in different sub-groups. For large-scale NP hard problems, different initial solutions will lead to different local optimal solutions with great possibilities. Then these local optimal solutions are taken as the individuals of the generation instead of the initial ones. In our distribution planning method, the employed mathematical optimization method is a genetic algorithm, which is called sub-genetic algorithm to differentiate it from the cognitive genetic algorithm. Niche technology is used in the sub-genetic algorithm, so that it can find several different local optimal solutions. To further diversity the individuals of the generation of cognitive

genetic algorithm, we stop the sub-genetic algorithm before it converges. Mathematical methods join the optimization process in this step. Without this step, the optimization will realized totally by human experts or expert system, the optimization process will be very slow, and the solutions are not rigorous from the mathematical viewpoint. Step 9: Fitness evaluation includes three steps that is mathematical evaluation, expert-system evaluation and cognitive evaluation. In fact, mathematical evaluation is completed in step 8 where the basic electrical and economical information, such as nodes voltages, branch currents, construction cost, and operation cost, have been calculated. In expert-system evaluation phase, all electrical and economical information of the system are weighted by expert system or artificial neuron network that is trained by rules or experiences of experts. Then the raw fitness of all solutions can been obtained. In cognitive evaluation phase, experts or planners reevaluate the fitness of all the solutions according to their experiences, intuitive, intelligence and other cognitive capability with emphasis on those decisions variables which cannot been considered in the last two evaluation phases. When planners reevaluate the fitness, they have observed the effects of all the previous evaluations including mathematical evaluation, expert-system evaluation and cognitive evaluation. If these results are very different from their expectation, they will revise the cognition to the planning problem in their mind. For example, the weights of some decision variables may be changed. All these changes will be reflected on the new fitness. Then feedback between human and mathematical algorithms and individual feedback take place. If there are more than one planner, then interpersonal feedback and collective feedback also take place. Step 10: Judge the stopping criteria, which can be whether the “optimal” solution has not been improved for specific number of generations or the maximum generation number has been reached. If it is satisfied, then stop the optimization, otherwise continue to step 11. Step 11: Selection and reproduction [9]. This step obeys the principle of "survival of the fittest". Roulette wheel method is used to realize the reproduction. The probability of an individual being selected and reproduced equal to its fitness divided by the total fitness of the population. It is possible to lose the best individual by the roulette method, therefore, an elist method is used the same time to ensure that at least one of the best individuals is reproduced to the next generation. Step 12: Crossover [9]. The key of any genetic algorithm is the re-combination of genes realized by crossover. Two-point crossover is recommended, because it is good at producing offspring inheriting good characteristics of both their parents. Parents: a1a2 | a3 a4 a5 | a6 a7 b1b2 | b3b4b5 | b6b7

Offspring: a1a2 | b3b4b5 | a6 a7 b1b2 | a3a4 a5 | b6b7 . Apparently, during the later period of gentle evolution, the above crossover cannot create new individuals largely because most individuals are similar, and a premature convergence may occur. In our proposed genetic algorithms, when half of the individuals have the same fitness, inversion crossover is used to create new individuals with new combinations of genes largely. The following is an example: Parents: a1a2 | a3a4 a5 | a6 a7 b1b2 | b3b4b5 | b6b7 Offspring: a1a2 | b5b4b3 | a6 a7 b1b2 | a5 a4 a3 | b6b7 . Although the inversion crossover can diversify the population greatly, it also has great possibilities of destroying the good combinations of genes. A much lower crossover probability, for example 0.5, is used. Step 13: Mutate [9], and then go back to step 7. The main reason for using mutation operator is to keep the diversification of the population to avoid premature convergence. It is also favorable to carry out fine local searching. 6. A POWER DISTRIBUTION SYSTEM PLANNING EXAMPLE Up to date, there is no mathematical global optimization methods can be used for power distribution system planning with tie feeders being considered. Consequently, networks in normal conditions and tie feeders in emergency conditions are optimized separately in most existing planning processes. However, because the tie feeders can not be considered in optimization of the network in normal conditions by conventional optimization planning methods, some better solution may be lost, and even worse some important feeders can not be tied with the other substations. Fig. 1 is a result of conventional mathematical optimization planning method. It can be seen that the loads of feeder trees C3, C4, C8 can not be transferred to other substation s when substation C totally fails. This is because there are no rights of way for these tie feeders being left after the other feeder trees of substation C are tied. However, if human experts join in the optimization planning of the network of normal condition by cognitive genetic algorithm, then the rights of way of tie feeders in emergence condition can be considered. For example, we can give higher fitness to individuals (schemes) that are easy to realize loads transfer in emergency or let some candidate feeders are selected forcedly to get desired network structures which have more rights of way for tie feeders. Although at the beginning of optimization planning, human experts cannot tell exactly where such tie feeders may be, with the iterative searching process going on, they will have a clearer idea about it. Figure 2 gives such a more reasonable result. Although its total cost (tie feeders’ cost are not included) and operational cost in normal condition are a little higher than

the previous one, its security is improved greatly because each feeders trees can be tied with another substation. Compared with Figure 1, we can see how feeder trees C3, C4, C8 found their rights of way for tie feeders. Apparently, the computers’ intelligence is not wise enough to do this work by itself at present, and humans’ intelligence is indispensable.

7. CONCLUSIONS The intelligence of humans and computers can overcome the disadvantages of each other. Hybridizing human’s intelligence with computer capability (including mathematical optimization methods and expert system et. al.) is a good way to solve complex engineering decision making problems. Such problems, at present, cannot be solved effectively by computers alone. Humans can join the decision making progresses by four types of cognitive feedback. They are feedback between human and mathematical algorithms, individual feedback, interpersonal feedback and collective feedback. Based on cognitive feedback and genetic algorithms, we proposed a cognitive genetic algorithm in this paper. Through an iterative human-machine interactive process, the algorithm can clarify and extract human experts’ cognition that is difficult to be expressed by computer programs. The proposed algorithm is tested with a power distribution system planning problem. The effectiveness of the proposed algorithm was shown by comparing the planning result with conventional mathematical optimization techniques. 8. ACKNOWLEDGMENTS This work has been supported by the National Natural Science Foundation of China (Granted No. 59877017). 9. [1]

[2]

[3]

[4]

[5] [6] [7] [8] [9]

REFERENCES Suresh K. Khator, Lawrence C. Leung, Power Distribution Planning: A Review of Models and Issues, IEEE Trans. on Power Systems, 10 (3) (1997) 1151-1159. Ignacio J. Ramírez-Rosado, José L. Bernal-Agustí, Genetic Algorithms Applied to the Design of Large Power Distribution Systems, IEEE Trans. on Power Systems, 13 (2) (1998) 696-703. Gunther Brauner, Manfred Zobel, Knowledge Based Planning of Distribution Networks, IEEE Trans. on Power Systems, 9 (2) (1994) 942-948. Tung Bui, Claudia Loebbecke, Supporting Cognitive feedback Using System Dynamics: a Demand Model of the Global System of Mobile Telecommunication, Decision Support System, 17 (1996) 83-98. Jens Rasmussen, et. al. , Cognitive Systems Engineering , John Wiley & Sons Inc., New York, 1994. Simon, H., Models of Thought, Yale University, 1979. Hammond, K. and D.A. Summers, Cognitive Control, Psychological Review, 1 (1972) 58-67. Senge, P.M., The Fifth Disciplines, The Art and Practice of Learning Organization, New York, 1990. Goldberg D E, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley, 1989.

Fig.1. Power Distribution System Planning only by Conventional Mathematical Methods

Fig.2 Power Distribution System Planning with Tie Feeders Being Considered by Cognitive Genetic Algorithm