基于改进粒子群优化的LSTM土壤温湿度预测模型研究

Research on LSTM Soil Temperature and Humidity Prediction Model Based on Optimized Particle Swarm Optimization Attention Mechanism

摘要: 随着全球气候变化加剧与粮食安全需求增长，农业生产向精细化、智能化转型成为必然趋势。土壤温湿度作为直接影响作物生长、水资源利用及温室气体排放的关键环境变量，其精准预测是实施智慧农业管理决策的核心依据。但其面临多因素耦合、非线性强、时空异质性显著等挑战。为提高预测准确性以支持灌溉优化、产量预估与灾害预警，文章聚焦深度学习技术在土壤环境预测中的应用，融合智能优化算法对土壤温湿度预测方法展开深入研究，针对传统长短期记忆网络超参数依赖经验设定、模型性能受限的问题，提出一种基于改进粒子群优化算法的LSTM超参数自适应整定方法(OPT-PSO-LSTM)。首先，构建包含气象、时间、土壤属性等多源特征的东北地区专用数据集并进行标准化处理；其次，利用粒子群算法的全局寻优能力，以验证集损失最小化为目标，对LSTM的关键超参数进行自动化搜索与优化。实验结果表明，该方法在寻优效率与效果上均表现优异。具体而言，OPT-PSO-LSTM获得最佳验证损失0.1638，显著优于网格搜索(0.1704)、随机搜索(0.1741)、遗传算法(0.1646)和粒子群算法(0.1693)，其收敛过程也快速稳定。优化后的模型在测试集上取得土壤温度预测R² = 0.8304、土壤湿度预测R² = 0.7565的可靠性能，有效提升预测精度，为构建适用于区域农业气象的鲁棒预测模型奠定自适应的参数基础。

Abstract: With the intensification of global climate change and the increasing demand for food security, the transformation of agricultural production towards precision and intelligence has become an inevitable trend. Soil temperature and humidity, as key environmental variables that directly affect crop growth, water resource utilization, and greenhouse gas emissions, require precise prediction to support the implementation of smart agricultural management decisions. However, the precise prediction of soil temperature and humidity faces challenges such as multi-factor coupling, strong nonlinearity, and significant spatial heterogeneity. To improve the prediction accuracy and support irrigation optimization, yield estimation, and disaster warning, this paper focuses on the application of deep learning technology in soil environment prediction, integrates intelligent optimization algorithms to conduct in-depth research on the prediction method of soil temperature and humidity, and addresses the problems of traditional long short-term memory networks’ reliance on empirical settings for hyperparameters and limited model performance by proposing an LSTM hyperparameter adaptive tuning method based on improved particle swarm optimization algorithm (OPT-PSO-LSTM). First, a dedicated dataset for Northeast China, containing multiple source features such as meteorology, time, and soil properties, is constructed and standardized. Secondly, using the global optimization ability of the particle swarm algorithm, with the goal of minimizing the loss in the validation set, the key hyperparameters of LSTM are automatically searched and optimized. Experimental results show that this method performs exceptionally well in terms of optimization efficiency and effectiveness. Specifically, OPT-PSO-LSTM achieves the best validation loss of 0.1638, significantly outperforming grid search (0.1704), random search (0.1741), genetic algorithm (0.1646), and particle swarm algorithm (0.1693). Its convergence process is also fast and stable. The optimized model achieves reliable performance on the test set, with a soil temperature prediction R² of 0.8304 and a soil humidity prediction R² of 0.7565, effectively improving the prediction accuracy and laying an adaptive parameter foundation for building a robust predictive model applicable to regional agricultural meteorology.