Research On Soil Moisture Prediction Model Based On Improved GRU Recurrent Neural Network

Soil moisture plays a key role in land water cycle and energy cycle,and has certain influence on weather and climate change.In terms of agricultural production,crop moisture can be judged according to soil moisture,timely and effective understanding of soil moisture information,and reasonable formulation of farmland irrigation strategies,which can save water resources while ensuring grain output,maximize the utilization value of agricultural water resources and achieve sustainable development.Therefore,the accuracy of soil moisture prediction has important research value for agriculture,climate and many other fields.In recent years,with the rapid development of computer technology and neural network,many scholars use deep learning algorithms to establish soil moisture prediction models,so as to continuously explore the accuracy of soil moisture prediction.However,due to the numerous influencing factors of soil moisture and the limited universal applicability of the model,how to further improve the accuracy of soil moisture prediction still needs to be further explored.In this study,a small number of influencing factors,namely atmospheric temperature,atmospheric humidity,rainfall and soil moisture data,were used to predict soil moisture.GRU recurrent neural network with time series characteristics is selected for experimental research,and the intelligent gull optimization algorithm and traditional Bayesian optimization algorithm are used to optimize the superparameters of GRU recurrent neural network,and the prediction and comparison of soil moisture at different times in the future are carried out.In the process of experiment,the optimization ability of the above two optimization algorithms on the superparameters is explored.At the same time,explore the performance of the two models for forecasting time extension.Its main research contents include:(1)The collected data of atmospheric temperature,atmospheric humidity,rainfall and soil moisture were used as influencing factors for the prediction of soil moisture for experimental exploration.Through comparative experiments,it was finally determined that the influence factor data of the first three days,namely 72 hours,were used as input variables,and the data of a total of 18 time periods,namely 72 variables,was used as input variables.The prediction effect of the model was better.(2)In this study,considering that the studied data has a certain time series,and by comparing the advantages and disadvantages of various recurrent neural networks,the GRU recurrent neural network with simple structure and fast running speed is selected as the main network of the model to be built while ensuring the prediction accuracy.Some of the superparameters of the neural network play an important role in the prediction effect of the model,and at present,there is no single combination of superparameters that can be applied to all the prediction models,so it is necessary to constantly try the optimal value when setting the parameters,so as to achieve the best prediction effect.Accordingly,this paper selects gull optimization algorithm and Bayesian optimization algorithm to achieve the optimal superparameter setting of GRU recurrent neural network in this study,including the number of hidden layer nodes,initial learning rate and L2 regularization coefficient,and establishes SOA-GRU and BOA-GRU soil moisture prediction models respectively.(3)In this study,the above two prediction models were used to predict soil moisture in the next 12 hours,24 hours,36 hours and 48 hours respectively.The prediction effect of the two models was not only compared horizontally,but also the influence of the prediction time on the prediction effect was compared longitudinally with the experimental results of multiple time steps.The final experiment showed that the SOA-GRU model had better prediction effect on soil moisture,and the best prediction evaluation index data were MAPE(12h)=4.4120%,R~2(12h)=0.94605,RMSE(12h)=1.9998.By comparing the prediction results of multiple time steps longitudinally,it was found that the prediction accuracy of the two models showed a gradual decline with the extension of the prediction time,but the decline rate of the prediction accuracy of the SOA-GRU model was more gentle than that of the BOA-GRU model.Combined with this research content,the SOA-GRU prediction model proposed in this paper has the feasibility of application in soil moisture prediction,which provides a reference direction for future research in this field.It is also proved that the optimization effect of intelligent gull optimization algorithm on soil moisture prediction model is better than that of traditional Bayesian optimization algorithm.