Sparse Perception And Prediction Of Crop Growth Environment Information Based On GA And LSTM

With the rapid development of the Internet of Things and big data technology,smart agriculture has become the inevitable trend of future development.Smart agriculture realizes automatic collection,remote transmission and intelligent management of information through the Agricultural Internet of Things(AIoT).In recent years,although AIoT has achieved many key breakthroughs and been widely used,there are still many problems,including limited energy of sensing nodes,incomplete sensing data,insufficient accuracy and unexplainable of time series prediction models.In order to solve these problems above,on the basis of previous research of sparse perception and time series prediction,this paper first focus on the sparse sampling in energy-constrained systems to reduce the total energy consumption of the network under the premise of ensuring data integrity,and combined with low-rank matrix completion and deep neural network to build a multi-level data reconstruction model to accurately reconstruct the missing data generated by sparse sampling.Then,aiming at the problems of difficult fitting and inexplicability of complex nonlinear data in time series prediction,a deep neural network is constructed by combining time series decomposition and physical information constraints to predict the multi-step future of the complex nonlinear environmental data collected in AIoT.The main tasks and innovations in this paper can be summarized as follows:(1)Sparse perception method based on genetic algorithm and multi-level data reconstruction model.Based on the special genetic coding method of genetic algorithm,the sparse sampling is described as a non-convex NP-hard problem with multiple constraints.To ensure the correct selection of redundant sensor nodes,the low-rank matrix completion module is embedded in the adaptive genetic algorithm to verify the integrity of the feature information in the sensing data,and a novel fitness function is designed to optimize the balance between the low energy consumption of the overall network and the high accuracy of the sensing data,reducing the possibility of convergence to the local optimal solution in non-convex optimization problems.Due to the large amount of missing data caused by sparse sampling,the low-rank matrix completion module is used as the preliminary reconstruction,and then in fully consideration of the spatio-temporal correlation in sensor data,the Seq2 Seq model combines Bi-LSTM and attention mechanisms to further optimize the reconstruction of missing data to ensure completeness and accuracy of sensor data.(2)Time series prediction model combining time series decomposition and physical information constraints.The progressive time series decomposition module is added to the traditional LSTM,empowers the model with progressive decomposition capacities,gradually separates the long-term trend information and season-cyclical information,and allows our model to alternately decompose and fit the intermediate results during the forecasting procedure.At the same time,in order to improve the convergence speed and interpretability of the time series prediction model,some interpretable physical information constraints are added to the final output layer as inductive bias,which speeds up the optimization in the feasible solution space and prevents the model from falling into local optimization or over-fitting.Applying the proposed model to the Seq2 Seq structure,the structure ablation experiments and performance comparison experiments on large-scale real-world datasets show that the addition of these two modules has significantly improved the existing methods,and also perform well in robustness,accuracy and generalization.