建筑物空调负荷预测的支持向量回归机算法研究

发布时间：2018-10-07 21:25

【摘要】：提高空调负荷预测的准确性是实现空调系统节能运行和优化控制的基础和前提条件。针对现有空调负荷预测算法精度和速度难以满足空调系统优化运行与节能控制需求的问题，本文利用支持向量回归机的强大非线性映射能力，分别对小规模训练样本和大规模训练样本条件下空调负荷预测支持向量回归机模型的在线建模、在线预测算法等方面展开了研究。论文的主要研究工作包括：（1）针对基于SVR的空调负荷预测模型参数难以确定及计算量过大的问题，本文提出了基于粒子群的空调负荷预测SVR模型参数优化算法，并建立了SVR空调负荷预测模型。仿真结果表明，本文提出的粒子群优化算法与网格搜索法、遗传算法比较，具有更快的寻优时间，，寻优时间仅为网格搜索法的7.1%~22.7%，遗传算法的22.8%~55.5%，该方法大幅度地缩短了空调负荷预测模型参数寻优时间，为空调负荷SVR预测模型提供了有效的参数优化算法。（2）针对常规离线SVR预测模型需要对模型进行重新训练，效率较差的问题，本文提出了小规模训练样本条件下建筑物空调负荷Online SVR预测算法。仿真结果表明，Online SVR预测模型在较小训练样本集下具有更优越的预测性能，但是，Online SVR预测模型受输入参数的影响较大。（3）针对当前空调负荷预测影响因素“时变”导致空调负荷预测模型不准确，影响负荷预测精度的问题，本文提出了大规模训练样本条件下基于SVR的空调负荷滚动预测算法，建立了SVR滚动预测模型。此外，算法利用当日前一小时的滚动信息，不断对模型进行修正以提高负荷预测精度。论文同时探讨了以期望误差百分比（EEP）为预测评价指标时，不同训练样本长度对神经网络和SVR算法预测精度的影响。预测结果表明，基于SVR的空调负荷滚动预测算法较常规SVR预测算法和神经网络预测算法预测精度分别提高了20.1%和19.8%，当训练样本较少时，本文提出的SVR滚动预测算法预测性能更为优越。
[Abstract]:To improve the accuracy of air conditioning load forecasting is the basis and prerequisite to realize energy saving operation and optimal control of air conditioning system. Aiming at the problem that the accuracy and speed of the existing load forecasting algorithms are difficult to meet the requirements of optimal operation and energy-saving control of air conditioning system, the powerful nonlinear mapping ability of support vector regression machine is used in this paper. The on-line modeling and on-line prediction algorithm of support vector regression model for air conditioning load forecasting under the condition of small scale training samples and large scale training samples are studied respectively. The main work of this paper is as follows: (1) aiming at the problem that the parameters of air conditioning load forecasting model based on SVR are difficult to determine and the amount of calculation is too large, a particle swarm optimization algorithm for air conditioning load forecasting SVR model parameter optimization is proposed in this paper. The load forecasting model of SVR air conditioning is established. The simulation results show that the particle swarm optimization algorithm proposed in this paper has faster searching time than the grid search algorithm and genetic algorithm. The optimization time is only 7.1and 22.722.7in the grid search method, while the genetic algorithm is 22.85.5. this method greatly shortens the optimization time of the air conditioning load forecasting model parameters, and provides an effective parameter optimization algorithm for the air conditioning load SVR forecasting model. (2) aiming at the problem that the conventional off-line SVR forecasting model needs to be retrained and its efficiency is poor, this paper proposes a Online SVR forecasting algorithm for building air conditioning load under the condition of small-scale training samples. The simulation results show that the online SVR prediction model has better prediction performance under the small training sample set, but the online SVR prediction model is greatly affected by the input parameters. (3) aiming at the problem that the influence factor of air conditioning load forecasting is "time-varying", which results in the inaccurate air conditioning load forecasting model and affecting the precision of load forecasting, this paper proposes a rolling forecasting algorithm for air conditioning load based on SVR under the condition of large-scale training sample. The rolling prediction model of SVR is established. In addition, the algorithm makes use of the rolling information of the first hour of the day and constantly modifies the model to improve the accuracy of load forecasting. At the same time, the paper discusses the influence of different training sample length on the prediction accuracy of neural network and SVR algorithm when the expected error percentage (EEP) is used as the prediction evaluation index. The prediction results show that the prediction accuracy of the air conditioning load rolling forecasting algorithm based on SVR is improved by 20.1% and 19.8%, respectively, compared with the conventional SVR forecasting algorithm and the neural network forecasting algorithm. The prediction performance of SVR rolling prediction algorithm proposed in this paper is superior.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TU831

【参考文献】