毕业生知识库

我国设施农业规模已居世界前列，但现代化的设施农业起步比较晚，大部分的温室缺乏科学有效的管理和控制方法。农产品价格和温室环境分别是农业社会物理信息系统（Cyber-Physics-Social-System，CPSS）中温室管理与控制的社会信息和物理信息。通过对设施蔬菜价格趋势的预测，农户可以科学地安排作物的种植和采收。而通过对温室内环境的精确控制，设施蔬菜的生长过程和状态可以被较为精确地控制。两相结合可以使得温室种植获得更高的经济效益，降低风险。因此，本文以提高温室生产管理质量和环境控制精度为目的，对设施蔬菜价格预测和温室内部环境控制进行研究。

  本文的工作主要包括以下几个方面：

      (1) 利用网络爬虫技术采集北京地区设施蔬菜的价格数据，提出了一种基于循环神经网络的设施蔬菜价格预测方法，并与传统时间序列方法ARIMA模型和经典机器学习算法BP神经网络进行了对比。本文采用上述三种方法对北京地区设施蔬菜价格进行月平均、周平均和日平均价格预测。结果表明，该方法具有更高的精度。网络爬虫技术扩展了数据规模，使得价格预测结果更有效，对温室的生产管理有很重要的指导意义。

      (2) 针对实际温室中实验成本高、建模困难的问题，本文提出了一种基于lightGBM算法的温室环境预测方法，基于第二届“Autonomous Greenhouses International Challenge”挑战赛上获取的温室环境及相关控制设备开关状态数据，建立了温室内部温度和二氧化碳浓度预测模型，并通过模型收敛性实验验证了该方法的收敛周期和程度。结果表明，该方法可有效拟合温室环境模型，且具有良好的收敛性，可以应用于实际温室的环境控制。该方法实验成本较低、实验周期较短。

      (3) 基于构建的数据驱动的温室环境模型，结合平行控制理论，提出了一种基于ACP理论的温室控制方法，并以通风口控制为例进行了温室环境控制计算实验，在不同条件下验证该方法的可行性。结果表明，该方法可以有效优化温室通风口的控制策略，实现温度控制的目的，并具有一定的泛化能力。

  最后，对本文进行了总结，并指出了需要进一步开展的研究工作。

The scale of facility agriculture in China has been among the top in the world, but the modern facility agriculture started relatively late so that most of the greenhouse lack of scientific and effective management and control methods. The prices of agricultural products and the greenhouse environment are respectively the social and physical information of the greenhouse management and control in the agricultural Cyber-Physics-Social-System (CPSS). By forecasting the trend of the price of facilities vegetables, farmers can arrange planting and harvesting cycles scientifically. Through the precise control of the greenhouse environment, the growth process and state of agricultural products can be more accurately controlled. By combining both, China's large-scale greenhouse industry and the greenhouse planting work of individual farmers can obtain higher economic benefits and reduce risks. Therefore, in order to improve the quality of greenhouse production management and the accuracy of environmental control, this paper studied the price forecast of agricultural products and the internal environmental control of greenhouse.

The work of this paper mainly includes the following aspects:

(1) This paper used the crawler technology to collect the price data of facility vegetable products in Beijing area, and proposed a method of agricultural product price prediction based on recurrent neural network, which was compared with the traditional time series method, ARIMA model, and the classical machine learning algorithm, BP neural network. These three methods are used to forecast the monthly average, weekly average and daily average prices of agricultural products in Beijing. The results showed that recurrent neural network has higher precision. Due to the expansion of the data scale brought by the web crawler, the deep learning method brought very effective price prediction results, which was of great help to the production management of greenhouse.

(2) As the cost for actual greenhouse experiments is high and the model of actual greenhouse is difficult to built, this paper put forward a greenhouse environment prediction method based on lightGBM algorithm. Then the greenhouse temperature and CO₂ concentration prediction model was established based on the data obtained in the 2^nd "Autonomous Greenhouses International Challenge", and the convergence period and degree of the method are verified by model convergence experiment. The results showed that the method can effectively fit the greenhouse data with good convergence, and can be applied to the actual greenhouse. The experimental cost is low, the experimental period is short.

(3) Based on the constructed data-driven greenhouse environment model and the parallel control theory, this paper proposed a greenhouse control method based on ACP theory, and the ventilation control was taken as an example to carry out the greenhouse environment control calculation experiment, so as to verify the feasibility of this method under different conditions. The results showed that the method can effectively optimize the control strategy of greenhouse ventilation and achieve the purpose of temperature control. The method also has a certain generalization ability.

Finally, the conclusions and future work are given.