Development of an evolving neural network approach in Unit Commitment of power system

Abstract

The Unit Commitment (UC) problem involves determining a start-up and shutdown schedule for the generating units over a period of time to meet the forecasted load demand at minimum cost. Intensive research efforts have been made over the past years to achieve a near-optimal solution of the UC problem in an efficient manner. The literature displays a wide range of UC methods with various degrees of accuracy and efficiency to handle difficult constraints and heuristics. The priority list based methods are simple and fast, but highly heuristic. Other approaches are based on dynamic programming and branch-and-bound methods. These methods are general and flexible, but often prone to the curse of dimensionality. More recently, methods such as Benders decomposition and Lagrangian relaxation have been applied to solve the UC problem. The Lagrangian method is efficient and well suited for application on large-scale systems. However, under some circumstances this method requires additional heuristics to improve convergence. In recent years, Artificial Neural Network (ANN) has also been applied to solve the UC problem. However the technique has a number of limitations. In addition, a training algorithm based on the steepest decent method, such as the back-error-propagation, can be very slow in the vicinity of the final convergence. In the absence of data noise, additional learning takes place in a multilayered perceptron only if new data is introduced that the neural network improperly classifies. The closer the representation comes to the concept, the smaller the chance that this happens. This is a characteristic of the least squares and the steepest-descent techniques. An alternative to the back-error-propagation is the genetic algorithm (GA) which is an evolutionary algorithm based on the concept of natural selection and genetics. The paradigms of GA were first published by Holland in 1962. During the last few years, GA research has been substantially expanded on Unit Commitment problem. As the genetic algorithm can prevent the local minimum when guided by the parallel evolution, the learning stagnation in neural network can thus be avoided. In this project we will utilize the GA algorithm to evolve the weight and the interconnection of the neural network and then using the recognition capability of neural network integrated with the trial and error methodology of dynamic programming to solve the unit commitment problem. Firstly, by using the recognition capability of evolving neural network, a pre-schedule of the commitment schedule is produced by the existing scheduling information for a random load profile. Secondly, the pre-schedule is passed to the dynamic programming algorithm to calculate a proper commitment schedule. Different approaches of using ENN-DP have been proposed to solve the thermal unit commitment. Performance comparisons of the proposed algorithm with conventional Dynamic Programming approach and Neural Network approach based on its running time and operating cost are given on several cases. From the results shown, ENN-DP has a greater operating time saving than NN-DP. Among the two training options of GA, connection has a better performance. Among the three generation's selection of GA, Roulette Wheel has the best result. Generally speaking, ENN-DP has a shorter operating time than ANN-DP and DP due to better initiation of weights or optimized number of connections.