摘要

摘

要

降水预测问题是一个非常复杂、重要的研究课题。一种事物从过去发展到现在再发

展到未来，其发展总存在某些内在的规律，国内外的一些专家和学者们积极的进行探索， 掌握事物发展的固有规律，把预测问题置于科学的基础之上，引导事物朝着预期的方向 迅速发展。精准的预测是正确决策的前提，随着科技的不断发展，几乎人类活动的各个 领域都能够提供大量的统计数据，因此在统计资料的基础上，建立了预测科学。

在对人工神经网络技术中的 BP 神经网络的学习和研究的基础上，针对 BP 神经网 络收敛速度慢、易于陷入局部最小等缺点，考虑到粒子群优化算法(Particle Swarm Optimization, PSO)和万有引力搜索算法(Gravitational Search Algorithm, GSA)的全局 性和快速性，分别采用 PSO、GSA 优化 BPNN 的初始权值和阈值，提出 PSO-BPNN 和 GSA-BPNN 降水预测方法。利用某省份 A、B、C 和 D 四个气象观测站的降水资料进行 预测实验。结果表明：两种方法较常规 BP 神经网络具有更快的收敛速度快，且预测误 差较小，同时结果也表明 GSA-BPNN 较 PSO-BPNN 降水预测方法准确率更高。

在介绍两种预测方法的基础上，以 A、B、C 和 D 四个气象观测站的年最高降水资 料为例，把这 52 年资料划分成雨涝、偏涝、正常、偏旱、干旱 5 种状态，采用均值-标 准差分级法进行旱涝指标值分级，并对旱涝状态进行预测和分析，结果表明，预测结果 与实际情况相吻合。

关键字：智能预测；BP 神经网络；粒子群优化算法；万有引力搜索算法

Abstract

Rainfall prediction is a very complicated and essential issue. The development of one thing from the past to the present and then develop into the future, there are always some internal rules in its development, some domestic and foreign experts and scholars explore and master the natural order of things actively and based on the scientific, the prediction guide things develop towards the expected direction rapidly. So accurate forecasting is the premise of correct decision-making. With the development of science and technology, almost all areas of human activity are able to provide a large number of statistical data. Therefore, establish a scientific prediction on the basis of statistical data.

BP neural network has a slow convergence and is easy to fall into local minimum of the through the artificial neural network learning. Considering the global and fast characteristic in the Particle Swarm Optimization (PSO) and Gravitational Search Algorithm (GSA), the PSO-BPNN and GSA-BPNN precipitation prediction methods are proposed which optimize the BPNN initial weights and thresholds using the PSO and GSA respectively. We predict the experiment using the precipitation data of the A, B, C and D weather stations. The results show that the two methods have faster convergence