CAJ | 학술논문

为提高温室环境控制系统的有效性，针对作物生长量的变化与环境因子的变化存在时间尺度不统一的问题，该文基于集合经验模态分解（ensemble empirical mode decomposition，EEMD）与Elman神经网络建模，提出一种线椒株高生长量预测方法。以8819线椒为试验对象，分别对线椒株高及其环境因子进行 EEMD 分解，对各尺度下的时间序列建立EEMD-Elman预测模型。结果表明：应用EEMD-Elman神经网络建立线椒株高生长量预测模型，模型预测值与实测值的平均绝对误差为1.69 cm，相关决定系数为0.996，标准误为1.104，模型预测结果与实测值呈极显著性相关。研究结果可以解决作物生长变化与环境变化时间尺度不统一的问题，为温室环境控制系统的控制目标的优化提供有效参数。
위제고온실배경공제계통적유효성，침대작물생장량적변화여배경인자적변화존재시간척도불통일적문제，해문기우집합경험모태분해（ensemble empirical mode decomposition，EEMD）여Elman신경망락건모，제출일충선초주고생장량예측방법。이8819선초위시험대상，분별대선초주고급기배경인자진행 EEMD 분해，대각척도하적시간서렬건립EEMD-Elman예측모형。결과표명：응용EEMD-Elman신경망락건립선초주고생장량예측모형，모형예측치여실측치적평균절대오차위1.69 cm，상관결정계수위0.996，표준오위1.104，모형예측결과여실측치정겁현저성상관。연구결과가이해결작물생장변화여배경변화시간척도불통일적문제，위온실배경공제계통적공제목표적우화제공유효삼수。
In order to improve the performance of greenhouse control system and figure out the multi-time scale variable problem between plant growth and environment elements in greenhouses, the ensemble empirical mode decomposition (EEMD) and Elman neural network were used to predict cayenne pepper height in this paper. Taking the cayenne pepper 8819 and its plant height and environment elements (temperature, relative humidity, total inner radiation) as research object, plant height and environment elements were decomposed by the method of EEMD. The cayenne pepper plants sampled were irrigated with the nutrient solution with electricity conductivity of 2.0 and pHvalue of 6.5 from Shandong Agricultural University. Five intrinsic mode functions (imfs) were obtained by the EEMD, named imf1, imf2, imf3, imf4 and imf5. Oscillations of cayenne pepper plant height and environment elements at different frequency were shown by imf1, imf2, imf3 and imf4, while variation trend of cayenne pepper plant height and environment elements were shown in imf5. All imfs were reconstructed by reconstruction of the EEMD with approximate values of 1, 0 and 0 respectively for correlation coefficient, standard error and mean absolute error to original time series. The imfs obtained by EEMD decomposition were used to build plant height prediction model at different time-scale frequencies based on EEMD method and Elman neural network. All the plants of cayenne pepper 8819 sampled were divided into 2 sets, which were training set and testing set. Average values of sampled plants and environment elements were used in neural network building. In this paper, the environment elements (temperature, relative humidity and inner total radiation) formed the input layer of EEMD-Elman network, and the prediction of cayenne pepper plant height was the output layer of EEMD-Elman network. Double-layer feedback structure method was used in this paper, in which there were 10 nodes in the first layer and 3 nodes in the second layer. The function ‘transig’ was used as transfer function in the feedback layer, while the function ‘purelin’ was used in the output layer. Sample data were divided into training set, validation set and testing set in proportion of 0.7:0.15:0.15. Five sub-neural networks were established by imf1, imf2, imf3, imf4 and imf5 of samples’ plant height and environment elements. Final predicted value of sampled plant height was reconstructed by EEMD reconstruction, using the results of 5 sub-neural networks. Results of EEMD-Elman prediction model showed the mean absolute error for plant height was 1.69 cm, with the correlation coefficient of 0.996 and the standard error of 1.104, which meant the prediction value was significantly correlated to the real value. Otherwise, 2 different prediction models were built by the method of EEMD-BP (back propagation) neural network and Elman neural network. Results of EEMD-BP prediction model showed the mean absolute error of 5.40 cm for plant height, with the correlation coefficient of 0.812 and the standard error of 7.012, while results of Elman model showed the mean absolute error of 8.87 cm, with the correlation coefficient of 0.908 and the standard error of 5.032. The results of 3 different prediction models were compared. The prediction of EEMD-Elman neural network was the best among the 3 neural networks. The results of the models with EEMD were better than that of the Elman model without EEMD. So, EEMD could decompose time series into different time scales according to its own features without the disturbance of noises and singular wave signals. Details of original time series were demonstrated decently by EEMD. Fluctuation of original time series could be explained better at different time scales by EEMD. Precision of prediction could be improved by Elman neural network with double-layer feedback structure construction. In conclusion, the combination of EEMD and Elman neural network can be used to figure out the issue of the prediction of multi-time scale between plant growth and environment variation in greenhouses, in order to provide effective references for control objectives optimization in greenhouse control system.