最佳化廣義迴歸類神經網路於預測模型之研究

The Study of Predictive Models Based on the Optimal General Regression Neural Networks

邵時俊(Shih-Chun Shao)

指導教授 ： 陳文輝

國立臺北科技大學/機電學院/自動化科技研究所/碩士(2011年)

https://doi.org/10.6841/NTUT.2011.00189

摘要

在許多的工程應用中皆會利用預測模型來解決實際上的問題，因此如何根據實際工程應用上所累積的歷史數據資料，來建構一個客觀且具有高精確性的預測模型為本論文研究之重點。本論文使用廣義迴歸類神經網路作為預測模型之核心演算法，廣義迴歸類神經網路學習速度快，僅需決定平滑參數即可決定整體網路之性能，利用少量的歷史數據資料便可達到高度的預測精準性。最後，本論文針對平滑參數利用交叉驗證法、基因演算法以及粒子群演算法去計算出最佳平滑參數，進而分別建立三種不同的預測模型，如此便能確保本論文所提出之預測模型有著最好的預測推論能力。
本論文首先驗證並分析此三種預測模型的預測推論能力，再根據預測模型之特性，分別針對兩種不同的研究案例做應用；第一個研究案例為資訊末端設備之資料校正前處理應用，第二個研究案例為H.264/AVC視頻通訊之錯誤隱藏。在第一個研究案例中，本論文將利用所提出之預測模型與傳統模糊演算法、倒傳遞類神經網路做誤差校正能力的分析比較；在第二個研究案例中，本論文將所提出之預測模型加入空間域錯誤隱藏演算法，改善傳統僅使用時間域錯誤隱藏演算法的缺點。由上述兩種不同研究案例之模擬實驗結果顯示，本論文所提出之預測模型無論輸入資料為何種型態，皆能有效地解決問題，亦有較高的實際工程應用價值。

關鍵字

預測模型 ； 廣義迴歸類神經網路 ； 最佳化演算法 ； 資料前處理 ； 錯誤隱藏

並列摘要

In engineering applications, the predictive models are always adopted to solve the actual problems. Therefore, the aim of this thesis is to study how to build up a high accuracy predictive model according to the historical data in engineering applications. Hence, general regression neural networks are applied as the core algorithm of predictive models in this thesis. It is because after we choose the spread constant, the features of the whole general regression neural networks can be determined. Therefore, it has a higher-speed learning ability than other neural networks. Also, it achieves the high accuracy of prediction with few historical sample data. Finally, this thesis adopts cross-validation method, genetic algorithms as well as particle swarm optimization to find out the best spread constant to build up three different predictive models so as to make sure they have the best prediction inferences in this thesis.
First, this thesis will examine and analyze the prediction inferences of these three predictive models. Then, they will be applied to two different cases according to their features. One is the application of the data pre-processing of remote terminal units, and the other is the H.264/AVC error concealment in video communication. In the first case, this thesis will apply the proposed predictive models compared with fuzzy algorithms and back-propagation neural networks to conduct the analyzing comparison of error calibrating. In the second case, the algorithm of spatial error concealment will be added into the predictive model to improve the shortcomings of merely adopting traditional temporal error concealment. According to the experimental results of two cases, the predictive models in this thesis can solve the problems effectively no matter which type of data is imported. Furthermore, they also have a high valuation in engineering applications.

並列關鍵字

Predictive Model ； General Regression Neural Networks ； Optimal Algorithms ； Data Pre-processing ； Error Concealment

參考文獻

[1] 許哲強，台灣區域電力負載預測分析系統之建立與應用研究，博士論文，國立成功大學，資源工程學系碩博士班，台南，2002。

[4] 李穎，類神經網路應用於國道客運班車旅行時間預測模式之研究，博士論文，國立成功大學，交通管理學系碩博士班，台南，2002。

[6] G. Rigatos, P. Siano, and A. Piccolo, “Neural Network-Based Approach For Early Detection of Cascading Events in Electric Power Systems,” IET Generation on Transmission & Distribution, vol. 3, no. 7, July 2009, pp. 650-665.

[7] D. F. Specht, “Probabilistic Neural Networks for Classification, Mapping, or Associative Memory,” IEEE International Conference on Neural Networks, 24-27 July 1988.

[8] D. F. Specht, “A General Regression Neural Network,” IEEE Transactions on Neural Networks, vol. 2, no. 6, Nov. 1991, pp. 568-576.