Abstract: 基于粒子群优化（PSO）算法和遗传算法（GA）对支持向量机（SVM）的核函数及主要参数进行训练优化，分别建立PSO算法、GA的支持向量机模型（PSO_SVM、GA_SVM）。选用ECMWF及T639数值预报产品资料和乌东德水电站降水资料，普查最优预报因子，构建包含各种类型降水过程的训练样本和测试样本。比较分析SVM模型RBF和Sigmoid核函数优劣。尝试先分段寻找局部最优，再选择全局最优的参数优化方法。通过增大训练样本集、降低交叉验证准确率、迭代次数截断和控制惩罚系数范围的方法，提高模型的稳定性和泛化能力，防止过拟合和收敛缓慢现象。利用测试样本对SVM、PSO_SVM和GA_SVM三种方案进行对比检验，优化的GA_SVM预报效果较好且稳定。经2018年试报表明，GA_SVM逐3 h累计降水量预报TS评分在50%以上，漏报率在15%以下，与ECMWF和T639比较，该模型TS评分提高1.4%。

Key words: Based on , PSO (particle swarm optimization) and GA (genetic algorithm), kernel function and main parameters of SVM were trained and optimized. The data of ECMWF and T639 numerical prediction products and the precipitation data at Wudongde hydropower station were selected to investigate the optimal prediction factors. Training samples and test samples of various precipitation processes were constructed. The kernel functions of RBF and Sigmoid were compared and analyzed. Try to segment to find local optimal firstly, then globally optimal parameters were selected. The stability and generalization ability of the model were improved by increasing the training sample set, reducing the accuracy of cross validation, truncating the iteration times and controlling the range of penalty coefficients, and the overfitting and the slow convergence were prevented. Through the test samples, simulation verification and comparative analysis of SVM, PSO-SVM and GA-SVM were carried out, and the prediction effect of GA-SVM was good and stable. According to the experimental forecast in 2018, the TS score of 3 h accumulative precipitation level forecast of GA-SVM was above 50% and the missing rate was below 15%. Compared with ECMWF and T639, the TS score of this model increased by 1.4%.

关键词: 降水量预报, 支持向量机, 核函数, 全局最优, 乌东德水电站