基于PSO-LSSVM的城市时用水量预测

Hourly Urban Water Consumption Prediction Based on PSO-LSSVM

为解决传统最小二乘支持向量机(LSSVM)采用交叉验证确定参数耗时较长和粒子群(Particle Swarm Optimization,PSO)优化算法早熟收敛的问题,提出一种基于种群活性PSO算法优化LSSVM参数的方法。利用群活性加速度作为多样性测度,当群活性加速下降时,对粒子的位置和速度分别执行进化和变异操作来改进标准PSO算法,然后分析上海市时用水量序列特点及其影响因素,选取影响程度较大的主要因素,将其作为预测模型的输入变量,建立时用水量预测模型;最后采用改进的PSO算法优化LSSVM参数来预测上海市时用水量。实例分析表明,对比文中其他3种模型输入变量组合,选取的预测模型输入变量能够更有效地提高预测精度;与传统LSSVM方法相比,提出的基于改进PSO-LSSVM的时用水量预测方法计算速度更快,预测精度更高。

The traditional least squares support vector machine （ LSSVM ） is time-consuming with cross-validation parameter selection, and particle swarm optimization（PSO） algorithm is with the risk of premature convergence. To deal with these defects, an improved PSO with swarm activity and LSSVM method is proposed. The method uses swarm activity as diversity index. When swarm activity is quick- ened to descend, evolution or mutation operation is added to modify the positions or velocities of particles to improve standard PSO algo- rithm. Based on the analysis results of characteristics of hourly water consumption series in Shanghai and influencing factors, main fac- tors are selected as model inputs and corresponding prediction model is built. Then the improved PSO is utilized to optimize the parame- ters of LSSVM to predict hourly water consumption. Case study shows that, in contrary to other three combinations of model inputs, this paper＇ s inputs are more effective to improve the prediction accuracy, and improved PSO-LSSVM prediction method has less computa- tional time and better estimating performance than the traditional LSSVM method.