Application of genetic-based neural networks to thermal unit commitment

Shyh Jier Huang, Ching Lien Huang

Research output: Contribution to journalArticle

Abstract

In this paper, a new approach of genetic algorithms based neural network that post-processed by dynamic programming (GANN-DP) is proposed to solve unit commitment problems. The advantages of genetic optimization are applied to initialize neural networks. This is followed by the dynamic programming for processing those uncertain states. The motivation of embedding genetic algorithms in the proposed hybrids are threefolds: (1) The network learning stagnation can be avoided, (2) Although an additional computation is required for the genetic optimization, the overall computation time in GANN-DP system can be reduced due to a better initiation of weights, (3) Uncertain states in the neural network outputs can be reduced. The proposed method has been tested on the commitment of 43 thermal units from Taiwan Power System data. Table 1 shows the unit commitment results by the proposed method and the other three methods including simulated annealing (SA), Lagrangian relaxation (LR) and neural network-dynamic programming (NN-DP) approaches. The last three methods were written to be used as benchmarks against the proposed method. The computation has been performed on an IBM PC-486, 33 MHz computer. Three cases were tested and evaluated based on their operational costs respectively. In case #1 the load profile is included in the training data. In case #2 and #3, we randomly deviate the input load demands at two and four different hours individually. In this way, we can test the generalization of the proposed method to various operating conditions. Figure 1 depicts the computation time performance of test methods. We run the simulations ten times and calculate their averaged computation time of ten runs. The operational cost obtained by the proposed approach was found to be the lowest of these four methods from the above test cases. Although the LR method is comparatively efficient from the computational viewpoint, the commitment cost is not the lowest. Besides, the tuning of control parameters is a time-consuming task in LR applications. In some test cases, the operational cost by SA method is economically acceptable. However, the computation time of SA is much longer than other three methods. For the NN-DP method performed without the aid of genetic computations, its computation time was Table 1. The operational costs Operational Cost (M$) SA LR NN-DPJ GANN-DP Case 1 193.87 196.47 192.19 192.18 Case 2 195.15 197.22 194.28 193.11 Case 3 194.23 197.45 193.96 193.30 (Figure Presented) Figure 1. The computation time of methods found to be longer than that of the GANN-DP approach. Such computation inefficiency will be more significant in a larger scale power system. These comparison results reveal that although GANN-DP needs an additional genetic process, the overall computational performance and commitment schedule is superior to other three methods.

Original languageEnglish
Number of pages1
JournalIEEE Power Engineering Review
Volume17
Issue number5
Publication statusPublished - 1997 Dec 1

    Fingerprint

All Science Journal Classification (ASJC) codes

  • Electrical and Electronic Engineering

Cite this