温室高效生产依赖于适宜的温室小气候环境,建立高精度的温室小气候多步预测模型对实现温室环境优化调控具有重要意义.本研究提出一种基于滚动的反向传播神经网络(Rolling Back Propagation Neural Network,R-BP)的温室小气候多步滚动预测模型.模型主要包括两个阶段:(1)建立初始的BP神经网络.采用自动编码器无监督学习方法获取初始网络参数,并利用改进的粒子群算法优化网络参数.(2)建立滚动的BP神经网络群.将前一个网络输出作为后一个网络的部分输入进行滚动训练和预测,实现温室小气候多步滚动预测.为验证R-BP模型的有效性,在阿拉伯联合酋长国阿布扎比市的自控温室和中国苏州市的非自控温室分别进行试验.验证试验表明,与传统BP神经网络相比,在阿布扎比温室试验中,采用R-BP模型预测未来6h室内温度,其平均误差降低69.9％,预测相对湿度,其平均误差降低47％;在苏州温室试验中,采用R-BP模型预测未来6h室内温度,其平均误差降低43.3％,预测相对湿度,其平均误差降低55.6％.说明R-BP模型能够较准确预测未来6h内温室小气候环境变化,可为制定温室小气候优化调控方案提供依据.%High greenhouse efficiency depends on the proper greenhouse environment.So it is of great significance to establish an accurate multi-step greenhouse microclimate prediction model to optimize the greenhouse environment control.A rolling back propagation (R-BP) neural network group model was proposed in this paper.The R-BP model mainly includes two stages:(1) Establish an initial BP neural network model,which adopts an auto-encoder (AE) to learn initial network factors and optimize the network factors by improved particle swarm algorithm;(2) Establish a rolling prediction model to realize multi step prediction of greenhouse microclimate.It used the output of the previous network as partial input of the next network.To prove the effectiveness of the R-BP model,several experiments were implemented in the Abu Dhabi auto-controlled green house and non-controlled Suzhou greenhouse.The experiments in Abu Dhabi greenhouse proved that the R-BP model achieved an average of 69.9％ error decrease in 6h temperature prediction in the greenhouse and an average of 47％ error decrease in relative humidity prediction,compared with the traditional BP neural network.In Suzhou greenhouse,the average prediction error of temperature was reduced by 43.3％ and the average prediction error of humidity was reduced by 55.6％.The experimental results prove that the R-BP model can accurately predict the change of the greenhouse microclimate for future 6 hours,to provide the basis for greenhouse microclimate control optimization.
展开▼
