中文摘要：

城市建成区面积预测是城市研究的一个核心问题，其与城市经济社会之间表现为一种复杂的非线性关系，传统的方法模型难以精确预测。作为一种较新的人工神经网络模型，RBF神经网络能以任意精度全局逼近任意非线性关系，表现出了极强的处理复杂非线性系统的能力。以合肥市建成区面积预测为例，构建了基于RBF网络的预测模型，作为对比，同时用BP神经网络、一元线性回归和多元线性回归模型进行了预测。预测结果的综合分析表明，在预测精度上，RBF网络〉BP网络〉多元线性回归模型〉一元线性回归模型。研究显示，RBF网络能为城市建成区面积预测提供一种新思路和新方法，进而可为城市土地利用及其规划制定提供科学的决策依据。

英文摘要：

Prediction of urban built-up area is a core issue in urban studies. There is always a complex nonlinear relationship between urban built-up area and urban economy and society. It is difficult to accurately predict urban built-up area by using traditional methods and models such as linear regression, time series analysis,gray system theory and BP neural network. As a new artificial neural network model, RBF neural network has some advantages of fast learning, easily getting in the local minimum and approximating any arbitrary accuracy of the global non-linear relationship. Therefore,RBF neural network can overcome some shortcomings of BP neural network and show a ability to handle complex nonlinear system. Currently, RBF neural network is one of the most accepted prediction methods. Taking the prediction of built-up area in Hefei City from 1997 to 2007 as a research sample, this paper established a prediction model based on RBF neural network from five impact indexes including GDP, financial income, total fixed asset investment, non-agricultural population and average wage of workers. As a comparison, this paper also used BP neural network, simple linear regression （SLR） model and multiple linear regression （MLR） model to predict. The results indicate that the means of prediction residuals and relative errors of RBF neural network were only 0. 4027 km2 and 0. 29 % which were the minimum values in the prediction results of the four models; the means of prediction residuals of BP neural network,SLR model and MLR model were 3. 5794 km2, 6. 8531 km2 and 3. 6668 km2 respectively; the means of prediction relative errors of BP neural network, SLR model and MLR model were 2.08%, 4.57% and 2.38% respectively. The residuals of RBF neural network in each year were the smallest except for 1999; the residuals of SLR model in each year were the largest except for 1999,2004 and 2007. From 2004 to 2005, the built-up area of Hefei was mutated, which led to large errors 〉8 km2） predicted by BP neural network,SLR model a