Modified Back-Propagation Neural Networks And Its Application Research On Prediction Problems

Prediction is the premise of plan-making, providing a basis for decision-making. Study on prediction is a key activity for human survival and development. As the most accurate and scientific method to describe uncertain things, quantitative prediction, based on statistical data, employs statistical methods, mathematical models and algorithms to carry out determination for scales, trends, speeds, etc. of things in future development. Quantitative prediction with great advantages of objective, refinement and standardization, plays a vital rule in human civilization advance, economic and social development, disease and disaster prevention, exploration in unknown world and so forth.Back-Propagation neural network, which adjust connection weights in accordance with error gradient descent rule, is one of multilayer feed-forward neural networks. As an important tool on prediction research, BP neural network is capable of nonlinear mapping, self-organizing, error feedback adjustment, generalization and fault tolerance, and thus its application field is broad and wide. However, inherent defects of basic BP neural network, such as being trapped at local minimum easily, long training time, slow convergence depending on parameters excessively and so on, seriously affect its application effect. Modified BP networks aiming at the defects mentioned above are usually prefered choices among neural networks to solve real-world prediction problems, as in comparison with basic BP network, modified BP networks have higher training effeciency and sample fitting rate, which correspond to prediction effeciency and accuracy in prediction problems.On the basis of proposing two innovative improved Back-Propagation neural networks, this paper is devoted to study hybrid quantitative prediction models and methods with modified BP neural networks as the main tool, and focus on solving rainfall prediction problem and population prediction problem effectively. The specific work of this paper is illustrated as follows:1. The emergence, principle, algorithm description, advantages and disadvantages of BP neural network were summed up and generalized. In addition, an overview of many modified methods aiming at basic BP algorithm's shortcomings was given, which laid a solid foundation for BP neural network's applications in quantitative prediction research. On the basis of analyzing and recalling overview of prediction research and quantitative prediction research, regression prediction belonging to causal relationship prediction and moving average prediction, exponential smoothing prediction, which are two typical time-series prediction methods, were discussed.2. Two innovative modified methods of BP network were put forward. Modified BP network with guarantee factor (GF-MBP) carried out sub-treatment to derivative function of sigmoid function by introducing guarantee factor, which theoretically ensured sufficient weight adjustment. Experimental results of GF-MBP for three benchmark problems showed that when guarantee factor was approximately equal to 0.9, GF-MBP network outperformed BP network with momentum in training convergence and success rate. Modified BP network based on second order momentum (SOM-MBP) took auxiliary rule of enhancing (weakening) weight change scope with second order momentum into account on the basis of exerting advantages of first order momentum, discussed the quantitative relation between First and Second Order Momentum Factor, and borrowed ideas from the successful experience of Vogl algorithm which adjusted weight change range dynamically. Experimental results of SOM-MBP for three benchmark problems indicated that SOM-MBP network had fast convergence and strong power of searching globally and locally in weights space.3. Applying Modified BP network to rainfall prediction problem: Above all, dimension of original meteorological data was reduced by using principal component analysis; then inherent laws contained in meteorological data were learned by training GF-MBP network; and thus PCA-BP rainfall prediction model was established. Prediction results showed that PCA-BP model based on GF-MBP algorithm outperformed similar models based on basic BP algorithm and BP algorithm with momentum in both training effeciency and prediction accuracy. Meanwhile, PCA-BP model was also superior to multi-variable linear and non-linear regression prediction methods in prediction accuracy.4. Applying Grey Verhulst-BP model (GV-BP) to population prediction: According to population increasing tendency, GV-BP model was innovatively presented. GV-BP model not only maintained Grey Verhulst model's S-type characteristics, but also originally integrated Grey Verhulst model with SOM-MBP network and made full use of certain information. It brought BP network's powerful abilities of self-organizing, error feedback adjustment and SOM-MBP network's comparative advantages of fast convergence, global search in weights space into full play, and carried out minute adjustment to developmental coefficient and grey usage with the purpose to improve prediction precision and adaptivity. Applied to predict China's population, GV-BP model had higher prediction precision and stronger adaptability than II-Moving Average Prediction and II-Exponent Smoothing Prediction in both fitting known data and short, medium and long-term prediction. In comparison with other Grey Models, GV-BP also had obvious comparative advantages in prediction effectiveness. GV-BP model had important theoretical and practical value for S-type time-series prediction study.