二次聚类与神经网络结合的日光温室温度二步预测方法

发布时间：2019-02-25 12:57

【摘要】：精确预测日光温室温度是实现对温室精准调控的前提。由于温室是复杂非线性系统,受室内外众多环境因素影响,且部分因素难以准确测量和建模,因此,难以通过机理分析建立室外因素精确影响室内温度的物理模型。而现有时间序列分析、人工神经网络等仅基于数据的方法预测准确度也较低。本文提出连续时间段聚类与BP神经网络相结合的二步日光温室温度预测方法。首先,进行二次聚类,对室外温度情况相似的日进行聚类,并将全年划分为若干个类似时间段,根据连续时间段内相似日的数量进行聚类,将全年内的连续时间段归入若干类别。其次,对不同类别的时间段,分别采用BP神经网络建立室外温度、相对湿度、太阳辐射、风速和温室室内温度间的关联模型,通过数据训练,能够较为准确的根据室外环境数据预测室内温度。通过涿州实验农场2年数据试验验证,通过二次聚类,全年连续时间段可划分为3类,通过分别建立BP神经网络并分别训练,结果表明本方法预测误差仅为6.23%,与现有未分类的BP神经网络预测算法对比,本文方法有效地提高了准确度,平均误差降低5.4个百分点。
[Abstract]:Accurate prediction of solar greenhouse temperature is the premise of accurate control of greenhouse. The greenhouse is a complex nonlinear system, which is influenced by many environmental factors, and some factors are difficult to measure and model accurately. Therefore, it is difficult to establish the physical model of outdoor factors that affect indoor temperature accurately through mechanism analysis. The prediction accuracy of existing time series analysis, artificial neural networks and other data-based methods is also low. In this paper, a two-step solar greenhouse temperature prediction method based on continuous-time clustering and BP neural network is proposed. First of all, the second clustering is carried out to cluster the days with similar outdoor temperature, and the whole year is divided into several similar time periods, according to the number of similar days in the continuous time period. A continuous period of time throughout the year is classified into several categories. Secondly, the correlation models of outdoor temperature, relative humidity, solar radiation, wind speed and indoor temperature in greenhouse are established by using BP neural network for different time periods. Can accurately predict indoor temperature according to outdoor environmental data. Through 2 years data test of Zhuozhou experimental farm, through quadratic clustering, the whole year continuous time period can be divided into 3 categories, and the BP neural network is established and trained separately. The result shows that the prediction error of this method is only 6.23%. Compared with the existing unclassified BP neural network prediction algorithms, the proposed method effectively improves the accuracy and reduces the average error by 5.4%.
【作者单位】：中国农业大学信息与电气工程学院
【基金】：国家自然科学基金项目(61601471) 北京市自然科学基金项目(4164090) 中央高校基本科研业务费专项资金项目(2017QC077)
【分类号】：S625;TP183

【相似文献】