基于传感器布设优化与天气预报的温室夜间温度Transformer – BiLSTM预测模型

Transformer – BiLSTM Greenhouse Nighttime Temperature Prediction Model Based on Optimized Sensor Deployment and Weather Forecasting

摘要:

为解决日光温室夜间温度预测中空间异质性高、成本约束和缺乏前瞻信息(如未来天气预报)等问题,本文提出一种融合传感器布设优化与天气预报的Transformer – BiLSTM混合预测模型。首先,通过构建基于粒子群优化BP神经网络的多传感器数据集成框架,并结合熵权法对传感器数据质量及物理空间分区代表性进行评估,将监测点从12个优化至4个,在显著降低成本的同时,最大限度保留了温室整体温度场的关键代表性信息。基于此优化监测方案,构建了引入天气预报作为关键输入特征的Transformer – BiLSTM混合预测模型。该模型融合BiLSTM捕获双向时序依赖与Transformer的全局注意力机制,增强对复杂环境动态的建模能力。基于乌鲁木齐日光温室数据的试验结果表明,本文预测模型性能显著优于BiLSTM、Transformer等基线模型,决定系数R2达到0. 98,平均绝对误差MAE为0. 30℃,均方根误差RMSE为0. 43℃,表明模型具有较高的预测精度。在不同天气条件下的泛化性验证表明,该模型在多种气象条件下均表现出良好的鲁棒性和预测精度,R2普遍高于0. 96,尤其在阴天和雨天条件下精度最高。

Abstract:

Aiming to address challenges such as high spatial heterogeneity, cost constraints, and lack of future weather data in solar greenhouse nighttime temperature prediction, a Transformer – BiLSTM hybrid model integrating optimized sensor deployment and weather forecasts was proposed. Firstly, by constructing a multi-sensor data integration framework based on a particle swarm optimization-backpropagation (PSO – BP) neural network, and then applying the entropy weight method to evaluate data quality and spatial representativeness across physical zones, an optimized sensor deployment (reducing monitoring points from 12 to 4) was achieved. This significantly cut costs while preserving key representative information of the overall greenhouse temperature field. Building on this, a Transformer – BiLSTM hybrid prediction model incorporating weather forecasts was developed. The model fused BiLSTM’s bidirectional temporal dependency capture with Transformer’s global attention mechanism for enhanced modeling of complex environmental dynamics. Experimental results from a Urumqi solar greenhouse demonstrated superior performance over baseline models (BiLSTM, Transformer), achieving an R2 of 0. 98, MAE of 0. 30℃, and RMSE of 0. 43℃. The model also showed robust generalization across diverse weather conditions, with R2 values consistently above 0. 96, particularly excelling under cloudy and rainy conditions.