Abstract: In recent years, facility cultivation technology, mainly in the form of solar greenhouses, and the major breakthroughs made by neural networks in the fields of speech recognition, computer vision, and sequence classification as effective tools for realizing time series prediction, have resulted in the increasing application of neural network technology in greenhouse microclimate prediction. This paper summarizes and describes the application of different types of neural network models in the prediction of greenhouse microclimate according to the development sequence of neural networks. The optimization of the network model is introduced in detail. However, the current neural network model for greenhouse microclimate prediction has some limitations, such as that a single model input parameter cannot consider the overall environmental changes of the greenhouse, that a single model structure leads to insufficient robustness, and that model testing methods are inappropriate for deep models and difficult to deploy in the field. This paper put forward suggestions to establish a hybrid model, improve model testing methods, and optimize the network structure of deep models. This paper expects to provide a reference for further research on intelligent control of greenhouses.

Key words: facility agriculture, microclimate, neural network, model prediction, optimization algorithm, deep learning

摘要： 近年来，随着以日光温室为主要形式的设施栽培技术的兴起以及神经网络在语音识别、计算机视觉、序列分类领域取得重大突破，作为一种实现时间序列预测的有效工具，越来越多的神经网络技术开始应用在温室小气候预测技术中。按照神经网络的发展顺序对不同类型的神经网络模型在温室小气候预测中的应用进行总结叙述，分别针对前馈神经网络、循环神经网络、深度神经网络及混合神经网络的发展现状和研究者对网络模型的优化情况进行详细的介绍。在此基础上，分析当前针对温室小气候预测的神经网络模型存在模型输入参数单一而无法考虑温室整体环境变化、模型结构单一导致鲁棒性不足以及模型可靠性检验方法不合适、深度模型难以实地部署的问题，提出建立混合模型、改进模型检验方式、优化深度模型网络结构等建议。以期为面向温室智能控制的进一步研究提供参考。

关键词: 温室农业, 小气候, 神经网络, 模型预测, 优化算法, 深度学习