月降水量的深度学习-协同进化支持向量机短期气候混合预测模型研究

项目针对月降水量变化具有显著的非线性、复杂性和不确定性，预报建模样本量较小，预报建模需要挖掘、提取多重影响因子高维数据集有效预报信息等问题，以广西夏季(6-8月)、89站的逐月月平均降水量为预报研究对象，拟采用支持向量机(SVM)方法以及增加不同的长短期记忆(LSTM)网络模块和特征融合层对深度学习长短期记忆网络进行改进，构建可以逐层学习、深入挖掘深层次数据特征信息、更适用于小样本和高维数据集的月平均降水量短期气候混合预测模型。并针对决定支持向量机性能最重要的核函数类型选择、核函数参数匹配，以及最优特征训练子集等问题，尝试采用粒子群算法、模拟退火算法和遗传算法等混合智能计算技术，进行支持向量机的核函数、惩罚系数、不敏感损失函数及最优特征训练子集的并行协同进化算法和协同进化策略研究。建立一种基于深度学习协同进化支持向量机的月平均降水量短期气候混合预测预报建模理论方法和预报模型。

There is significant nonlinearity, complexity and uncertainty for the monthly precipitation change, besides, the sample size of forecast modeling is small, and it is necessary to mine and extract effective forecast information from high-dimensional data sets with multiple influencing factors. In this project, the monthly average precipitation of 89 stations over Guangxi in summer (June-August) is taken as the forecast object. The support vector machine (SVM) method, different Long Short-Term Memory (LSTM) network modules and feature fusion layers are employed to improve the LSTM network of deep learning. By means of layer by layer learning and deep mining of data feature information, this short-term climate mixed prediction model for monthly average precipitation is more suitable for small samples and high-dimensional data sets. Aiming at the problems of kernel function type selection, kernel function parameter matching, and optimal feature training subset that determine the performance of SVM, the hybrid intelligent computing technology such as particle swarm optimization algorithm, simulated annealing algorithm and genetic algorithm has been used for the research on parallel cooperation algorithms and coevolution strategies for kernel function, penalty coefficient, insensitive loss function and optimal feature training subset of SVM. Therefore, a mixed forecast modeling theory method and forecast model for monthly average precipitation of short-term climate has been developed based on deep learning and co-evolutionary support vector machine.