基于SVR空调负荷预测模型的参数优化研究

Parameter Optimization of HVAC Load Forecasting Model Based on Support Vector Regression

基于数据驱动的负荷预测模型在既有建筑运行阶段的准确性和可靠性受到人们越来越广泛的认可。特别是支持向量回归机(SVR)算法,基于统计学习理论和结构风险最小化原则克服了小样本学习困难、维数灾难等缺陷,在准确有效地解决复杂工程问题方面展现出巨大潜力,越来越深入地应用于暖通空调负荷预测领域。但在实际建模过程中,SVR算法对应的输入参数及算法本身的固有参数和核函数参数的选取,也会对模型最终性能产生显著影响。在现有建筑负荷预测模型的研究中,针对建模过程中参数优化的研究较少。主要研究SVR负荷预测模型参数的优化,将统计学方法与机器学习算法相结合,结果显示:①可以采用Pearson相关系数对影响空调负荷的参数进行两两相关性分析,以解决输入参数之间存在多重共线性而导致模型估计失真或准确率较低的问题;②可采用随机森林算法计算输入参数对预测结果的贡献度,以剔除冗余参数,降低模型复杂度,并综合实际工程应用对经济成本和允许误差的考量来确定最终输入参数的个数;③以表征拟合度优劣的相关系数平方R^(2)、反映预测精度的均方根误差RMSE和平均绝对百分比误差MAPE作为评价指标,针对研究对象,网格搜索法和粒子群优化算法对应的MAPE均在7.5%以下,RMSE远小于遗传算法的对应值,R2高达0.95,参数寻优结果能使预测模型具有更高的预测精度及更强的泛化能力。标准遗传算法容易陷入局部最优解,因此迭代完成后训练误差可能仍不理想。此外,网格搜索法的寻优时间为粒子群优化算法的数倍,随着训练样本量的增加,寻优时间的差距也会不断扩增,因此当训练样本较大时,利用粒子群优化算法可以缩短空调负荷预测模型的建模时间。

The accuracy and reliability of the HVAC load forecasting model based on data-driven methods has been widely recognized in operating existing buildings. Machine learning techniques especially the Support Vector Regression( SVR) algorithm based on the Statistic Learning Theory( SLT) and Structural Risk Minimization( SRM),overcomes the disadvantages of the difficulty in learning with small samples,dimensional disasters,etc.,has shown great potential in solving complex engineering problems accurately and efficiently without actually getting involved in the mechanics of the problem. Hence,SVR algorithm has been increasingly applied in HVAC load forecasting. However,in the actual application process,the selection of the input parameters corresponding to the SVR algorithm and the inherent parameters and kernel function parameter of the algorithm itself will also have a significant impact on the final performance of the model. In the existing researches on building HVAC load forecasting models,there are few researches on parameter optimization in the modeling process. This paper focuses on the optimization of SVR load forecasting model parameters, combining statistical methods with machine learning algorithms. The results show that:① Pearson correlation coefficient can be used to analyze the pairwise correlation of parameters that affect HVAC load,in order to solve the problem of multiple collinearity between parameters,which leads to the distortion of model estimation or low accuracy of the model.②The random forest algorithm can be used to calculate the contribution of input parameters to the prediction results and provide an interpretable ranking list,so as to eliminate redundant parameters and reduce the complexity of the model,determine the number of input parameters based on the comprehensive consideration of economic costs and allowable errors in practical engineering applications.③The selection of inherent parameters and kernel function parameter of SVR also has influence on the prediction accuracy and performance. Take correlation coefficient square R^(2) which characterizes the goodness of fitting degree,Root Mean Square Error( RMSE) and Mean Absolute Percentage Error( MAPE) that reflect the prediction accuracy are taking as evaluation indexes. For the research object of this paper,the MAPE corresponding to the grid search method and the particle swarm optimization algorithm are both below 7. 5%,the RMSE is much smaller than the corresponding value of the genetic algorithm,R^(2) is as high as 0. 95,and the parameter optimization results can make the forecasting model have higher prediction accuracy and stronger generalization ability. Standard genetic algorithm is easy to fall into a local optimal solution,so the training error may still be not ideal after the iteration is completed. In addition,the optimization time of the grid search method is several times that of the particle swarm optimization algorithm. With the increase of the training sample size,the gap in the optimization time will continue to expand. Therefore,when the training samples are great,the particle swarm optimization algorithm can shorten the modeling time of HVAC load forecasting model.