Solar greenhouse temperature and humidity models play an important role in its structure design and control. Solar greenhouses are multi-input and multi-output (MIMO) systems, and they are highly nonlinear and strongly coupled systems that are largely influenced by the outside weather (such as wind speed, outside temperature and humidity) and many other practical constraints (such as blowing and moistening cycle). Therefore, solar greenhouse temperature and humidity models are difficult to establish by mechanism analysis methods. Due to its ability to approximate complex nonlinear mapping directly from the input samples, neural network can provide models for many kinds of natural and artificial phenomena that are difficult to handle using classical parametric techniques. Among many kinds of neural networks, extreme learning machine (ELM) for single-hidden layer feed forward neural networks has been studied more thoroughly. But there are limitations existing in ELM such as fixed hidden-layer activation function and overfitting when minimizing training error. In order to achieve comprehensive control of temperature and humidity in the solar greenhouse and improve prediction accuracy, an improved ELM based on orthonormal basis function is proposed in the paper. First, it determines the number of the nodes in hidden layer by using empirical mode decomposition (EMD); second, on the basis of statistical learning theory combined with the empirical risk and structural risk, it takes minimal value of the sum of the minimum output weight and the minimum error; third, it identifies the greenhouse microclimate environmental factors. The prediction model of temperature and humidity is established by the improved ELM. The proposed method is tested in the solar greenhouse of Vegetable and Fruit Research Institution of Chinese Academy of Agricultural Sciences, which is located in 40°07′N, 116°09′E. According to the characteristics of solar greenhouse environment, the inputs of ELM are temperature and humidity outside solar greenhouse, light and wind speed, and the outputs of ELM are temperature and humidity inside solar greenhouse. Root mean square error and model validity are used as index to measure the generalization ability and the accuracy of model. According to the results of EMD for signal of temperature and humidity inside solar greenhouse, the number of nodes for the traditional ELM and the improved ELM is 9. This paper adopts sigmoidal function as the excitation function of the traditional ELM. The improved ELM is based on orthonormal basis function, and its excitation function coefficient of nerve cells in hidden layer is 30 and 10, respectively. Compared to the traditional ELM, the prediction results of the improved ELM show that temperature error and humidity error are reduced by 2℃ and 5% respectively, root mean square error of temperature is reduced by 0.4758℃ and that of humidity is reduced by 0.6857 percent, and model validity of temperature and humidity are improved by 0.0384 and 0.0314 respectively, so the improved ELM is effective, and it has certain reference value for intelligent control of the solar greenhouse microclimate.%日光温室温湿度模型是其结构设计与控制的重要基础，因日光温室系统具有大惯性、强耦合、非线性等特性，采用机理分析法，难以建立其准确的数学模型，导致日光温室控制效果差。神经网络建模能更加灵活地得到日光温室系统的参数，但传统的极限学习机（extreme learning machine，ELM）存在隐含层神经元激励函数固定，只考虑经验风险（即训练误差最小化），而导致过拟合等问题。为了实现对日光温室内温湿度环境因子的综合控制，需要进一步提高日光温室环境因子的预测精度，该文将基于正交基函数的改进型极限学习机对日光温室环境因子进行辨识，并利用经验模态分解（empirical mode decomposition，EMD）方法确定网络隐含层节点数，建立了日光温室温湿度环境因子预测模型。利用所建立的模型对日光温室内的温度和湿度等环境因子进行预测结果表明：温度模型有效性为0.9434，湿度模型有效性为0.9208，实测值与预测值的拟合关系比较理想，说明基于正交基函数的改进型极限学习机对日光温室进行系统辨识是可行的，且对日光温室智能控制的发展有一定的参考价值。
展开▼
