The unit commitment (UC) problem determines both the hourly start-up and shut-down schedule and the real power outputs of the generating units over a time period of a day or a week. The resultant schedule should supply the power demands at the lowest possible cost while adhering to various constraints of the system and the individual units. With the advent of the stochastic global search algorithms, the simulated annealing (SA) and the genetic algorithm (GA) have been applied to the conventional UC problem in an attempt to obtain the global or more near global optimal solution. The results are encouraging, but such approaches give rise to the problem of long execution time in practical applications, due to • The sequential search properties of the SA and the GA methods executed on traditional computing machines despite the implicit parallelism owned by the stochastic search algorithms, and • The operations (such as crossover and mutation in the GA) directly on the decision variables (i.e., the hourly on/off scheduling of the units) to randomly create the new trial solutions, a fact which is prone to violating the constraints. The rapid development and decreasing price of multiprocessor machines, like the transputer networks, recently have shown the promise to solve the UC problem with high performance and low costs by using the stochastic search algorithms, in particular for the GA. Complying with this trend, the paper develops and experiments with two different topologies of parallel GA (PGA) to enhance the practicality of the GA computing speed. The structures of master-slave and dual-direction ring are implemented on a distributed-memory transputer network. In addition, to overcome the difficulty in handling the constraints of the UC problem, we first propose a technique to embed the minimum up- and down-time constraints in the binary representation inherently needed in the GA. The other constraints are dealt with by integrating appropriate penalty factors, which vary with the amount of corresponding constraint violation, into the cost function to consider the constraint violations. In this way, the constraints of the UC problem can be more easily handled in the GA. Testing on a 38-unit, 24-hour UC problem in the Taipower system, numerical results have been used to demonstrate the power of the PGA in searching the global or near global optimal solution to the UC problem. In particular, the parallel processing topology of dual-direction ring was shown to be able to achieve a near linear reduction in computation time when compared with its sequential form. For reason of comparison, the SA approach and the latest Lagrangian Relaxation (LR) method were used for performance evaluation of the PGA. Although still consuming longer execution time than the LR method, the solution obtained using the proposed PGA approach has lower total cost than that of the SA and the LR methods. On the other hand, the execution time taken for the PGA is much shorter than that for the SA method. In this study, we have experienced the easiness of programming, debugging, and modifying the software of PGA. The portability of the PGA software turther makes feasible the PGA in solving diverse optimization problems of power systems within reasonable execution time.
|Number of pages||2|
|Journal||IEEE Power Engineering Review|
|Publication status||Published - 1997 Dec 1|
All Science Journal Classification (ASJC) codes
- Electrical and Electronic Engineering