Research And Application Of Soil Moisture Prediction Method Based On LSTM Neural Network

Soil is an important environment for plant growth,and soil moisture forecasting is important for increasing crop yields and production.Soil moisture is a data with the property of time series and non-linearity.The ease,timeliness and accuracy of soil moisture monitoring has long been limited by poor observation methods and equipment,to the extent that quantitative prediction of soil moisture has become a challenge.In response to the problems of high installation cost,easy damage and high packet loss rate of existing fixed-end sensors in the process of soil moisture monitoring,and the problems of untimely prediction,low prediction accuracy and difficulty of human adjustment in the practical application of traditional soil moisture prediction methods.This paper designs and implements a soil moisture prediction system based on an improved particle swarm algorithm optimized LSTM neural network,and the highlights of the paper’s research are described as follows:(1)A comprehensive discussion of the current status and main methods of soil moisture prediction,the advantages and disadvantages of traditional and intelligent prediction methods in soil moisture prediction,and an analysis of the advantages of neural networks in soil moisture prediction.(2)The collected data were subjected to data cleaning,kriging interpolation was applied to fill in the vacant values in the data,and grey relational analysis(GRA)was used to correlate the different meteorological data with soil moisture conditions.Meteorological data with a high degree of correlation were selected as the input parameters of the prediction model,and soil relative humidity was used as the output parameter to build a three-class neural network prediction model with BP,Elman and LSTM.(3)Based on the 2019 meteorological data of Haidian Park and soil moisture data,a soil moisture prediction model with six types of meteorological data: average temperature,average humidity,wind speed,ground temperature,rainfall and sunshine hours as input vectors and soil relative humidity as output vector was constructed,and the prediction accuracy of the several models established was experimentally analysed.(4)The inertia weights and learning factors of the particle swarm algorithm are improved and a multi-particle population approach is used to effectively avoid the problem of particles easily falling into local extremes.The improved particle swarm algorithm is analysed against the Differential Evolution algorithm(DE),the Artificial Bee Colony algorithm ABC and the Group Search algorithm(GSO),proving that the accuracy and stability of the improved PSOLSTM model have been greatly improved.The improved particle swarm algorithm is used to automatically optimize the hidden layer unit number,iteration count and dropout coefficient of the LSTM neural network,which can address the problem of difficult and large errors in human parameter tuning effectively.Through the analysis of the experimental results,the prediction error of the optimised LSTM neural network model based on the improved particle swarm search algorithm is smaller than several other prediction models,indicating that this improved model has a strong practical application and can make accurate predictions of soil moisture.