| If the temperature is too high,it will destroy the biological enzyme activity of grain or denature the protein;if the temperature is too low,it will cause damage to the grain organization and eventually affect the quality of grain.Therefore,it is particularly important to detect the temperature of grain pile and deal with it in time.Therefore,this paper puts forward the research of neural network model in predicting the temperature of grain pile.The results show that the SSA-BiGRU-MLP neural network has a good effect on the prediction of grain pile temperature.In this paper,LSTM,BiLSTM(Bidirectional LSTM),GRU,BiGRU(bidirectional GRU)deep learning neural network models are used to predict the temperature of the bottom layer,middle layer and surface layer of grain pile respectively.The predicted results show that: For the temperature prediction of the bottom,middle and surface layers of the grain pile,the bidirectional GRU prediction error can be controlled within the range of 0.3989,1.0432 and 0.4478,respectively,which is smaller than that of LSTM,BiLSTM and GRU neural network.It is verified that BiGRU’s prediction accuracy is better than LSTM,BiLSTM and GRU neural network.In order to further improve the prediction accuracy of bidirectional GRU,multi-layer perceptron MLP is used to optimize the convergence rate and overfitting of bidirectional GRU,and SSA singular spectrum analysis is used to denoise the input data.The main approach is to first use the KNN algorithm to process the missing value of the acquired data,and then use the SSA singular spectrum analysis to construct the trajectory matrix of the collected temperature time series data of the bottom,middle and surface layers of the grain pile.Then,the trajectory matrix is decomposed by singular value,grouped and reconstructed.The main purpose is to remove the grain pile bottom containing noise.The temperature data of the middle layer and the surface layer,as well as the four dimensions of outer temperature,outer humidity,chamber temperature and chamber humidity that affect the three layers,are then input the multi-variable data of noise removal into the neural network optimized by MLP for training and prediction.The results show that: The error of SSA-BiGRU-MLP neural network can be controlled within the range of 0.1492,0.1296 and 0.111 respectively in the temperature prediction of the bottom,middle and surface layers of grain pile,and the error is smaller than that of SSA-GRU-MLP and SSA-BiLSTM-MLP.SSA-LSTM-MLP,BiGRU,so the prediction effect of SSA-BiGRU-MLP neural network is better.To further verify the prediction effect between SSA-BiGRU-MLP and SSA-BiGRU and BiGRU-MLP and BiGRU,relevant ablation experiments were performed.The results show that: For the temperature prediction of the bottom,middle and surface layers of grain pile,the error of SSA-BiGRU prediction can be controlled within 0.1926,0.6173 and0.4062 respectively,and the error of BiGRU-MLP prediction can be controlled within0.3265 and 0.6901 respectively.Within 0.2329,so SSA-BiGRU and BiGRU-MLP have better prediction effect than BiGRU,but the prediction accuracy of SSA-BiGRU-MLP is higher than that of SSA-BiGRU and BiGRU-MLP,so SSA-BiGRU-MLP has the highest prediction accuracy.By the neural network experiment of grain reactor temperature prediction,it is verified that the ratio of SSA-BiGRU-MLP to SSA-BiLSTM-MLP,SSA-LSTM-MLP,SSA-GRU-MLP,SSA-BiGRU,BiGRU-MLP,BiGRU,GRU,LSTM,BiLSTM has better prediction accuracy,so it can be applied to the actual storage process of grain.It can predict the temperature of grain pile more accurately,so as to achieve accurate adjustment of the temperature of grain pile.In this way,the quality of grain can be guaranteed in the storage process,which is of great significance for the protection of grain quality. |
