基于PSO的SVR模型在多联机功耗预测上的应用

Optimized Support Vector Regression Model Based on Particle Swarm Optimization for Energy Consumption Prediction of a Variable Refrigerant Flow System

支持向量回归(SVR)模型在多联机系统功耗预测稳定性和精度上存在不足,本文引入粒子群优化(PSO)算法,对SVR预测模型的惩罚系数C和核参数γ进行最优求解,来改善模型预测性能。在制冷剂充注量为95. 75%工况下,对多联机组进行运行实验,并对实验数据进行预处理。基于PSO算法建立PSO-SVR模型,对多联机功耗进行预测,并与SVR模型的预测结果和理论公式计算结果进行对比。结果表明:SVR、PSO-SVR、理论公式计算法总体预测误差分别为1. 43%、1. 08%和1. 57%,均方根误差RMSE分别为105. 36、88. 79、91. 37 W,参数寻优结果为惩罚系数C=10 000和核参数γ=4. 275。粒子群优化算法的引入显著提高了SVR模型的预测精度和稳定性;相较于理论公式计算法,PSO-SVR精度更高,且需要测量的参数数目明显减少,在降低了测量系统复杂性同时更具经济适用性。

Energy consumption prediction analysis has important significance in energy management,operation strategy optimization,control optimization,etc. For variable refrigerant flow(VRF) systems,the pure support vector regression(SVR) prediction model has insufficient stability and prediction accuracy. By introducing the particle swarm optimization(PSO) algorithm,this study optimizes the selection of punishment coefficient C and kernel parameter γ for a pure SVR prediction model and then compares the prediction results of the PSO-SVR model,pure SVR model,and theoretical formula. The results show that the overall prediction errors for SVR,PSO-SVR and theoretical formula are 1.43%,1.08% and 1.57%,and the root mean square error are 105.36 W,88.79 W,and 91.37 W respectively. By solving for the best punishment coefficient C and kernel parameter γ equal to 10 000 and 4.275,the PSO can significantly improve the performance and stability of the pure SVR prediction model. In addition,it demonstrated better results than those of the formula calculation method with less variables to be measured. It is reasonable to state that the PSO-SVR model is a convenient and economic means to solve such problems.