Improved Informer Model for Alfalfa Soil Moisture Prediction
王静 1刘瑞 1杨松涛 1葛永琪1
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 宁夏大学 信息工程学院,宁夏 银川 750021
折叠
摘要
精准的苜蓿土壤湿度预测对于提高水资源利用率和降低智慧农业投入成本至关重要.针对传统土壤湿度预测方法在实际应用中存在预测周期短、精度低以及时空预测不足等问题,提出了一种融合快速傅里叶变换的Informer时空预测方法(Fast Fourier Transform and Spatio Temporal-Informer,FFT-ST-Informer).首先,在传统Informer模型基础上添加了独立的时空嵌入层,从而捕获各个变量之间复杂的时空相关性.然后,根据土壤墒情与环境因素的相关性分析结果,选择降雨、灌溉量为关键环境因素,并使用快速傅里叶变换,通过提取某一周期具有先验的数据序列的频谱来表示其频域特征放入模型.此外,该模型中的ProbSparse自注意机制可以集中提取时空数据的重要上下文信息.FFT-ST-Informer模型使用来自宁夏引黄灌区自采的气象和土壤数据作为输入数据.实验结果表明,FFT-ST-Informer模型性能明显优于传统模型,比LSTM模型在平均绝对误差(MAE)、均方根误差(RMSE)、相关系数(R2)等评价指标上,分别提高了 56.9%,64.4%,0.12%.
Abstract
Accurate prediction of alfalfa soil moisture is crucial for improving water resource utilization and reducing smart agriculture in-vestment costs.In view of the issues of short prediction cycle,low accuracy,and insufficient spatio-temporal prediction in traditional soil moisture prediction methods,we propose a Spatio-Temporal prediction method called FFT-ST-Informer that integrates Fast Fourier Transform.Firstly,an independent spatio-temporal embedding layer is added to the traditional Informer model to capture the complex spatio-temporal correlations between variables.Then,based on the correlation analysis between soil moisture conditions and environmental factors,rainfall and irrigation volume are selected as key environmental factors.The FFT is utilized to extract the frequency domain characteristics of data sequences with prior knowledge of a certain period,which are then incorporated into the model.In addition,the ProbSparse self-attention mechanism in this model can effectively extract important contextual information from spatio-temporal data.The FFT-ST-Informer model uses meteorological and soil data collected from the self-drained Yellow River irrigation area in Ningxia as input.Experimental results show that the performance of the FFT-ST-Informer model is significantly better than that of traditional models,with improvements of 56.9%,64.4%,and 0.12% in terms of mean absolute error(MAE),root mean square error(RMSE),and coefficient of determination(R2)evaluation metrics compared to the LSTM model.
维普
万方数据