Warehouse Grain Pile Grain Temperature Prediction Method and Device

Abstract

The present invention discloses a method for predicting the grain temperature in a storage grain pile, including: obtaining historical meteorological data and temperature data at each position point in the grain pile; inputting the obtained historical meteorological data and temperature data at each position point in the grain pile into a neural network prediction model to obtain the temperature data at each position point in the grain pile at a future moment; wherein, the neural network prediction model includes a first branch and a second branch, the first branch includes an input layer, an LSTM network, and a Permute layer, the second branch includes an input layer, an Inception module, and a spatial attention module, the outputs of the first branch and the second branch are both input into a Concat layer, the Concat layer is connected to a first fully connected layer, and the first fully connected layer is connected to an output layer. The present invention also provides a device for predicting the grain temperature in a storage grain pile. The present invention combines the Inception module, the temporal attention module, and the spatial attention module to obtain a neural network prediction model, which can accurately predict the grain temperature at each position point in the grain pile.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning. More specifically, the present invention relates to a method and device for predicting the grain temperature in a storage grain pile. Background Art

[0002] The grain temperature in the grain pile is related to the safety of grain storage. If there is a local temperature rise, it will accelerate the reproduction and growth of microorganisms, resulting in grain mildew. Therefore, in order to ensure the quality of stored grain, it is necessary to design a grain temperature prediction model for accurately predicting the grain temperature at each position in the grain pile. Summary of the Invention

[0003] An object of the present invention is to provide a method and device for predicting the grain temperature in a storage grain pile, which combines an Inception module, a temporal attention module, and a spatial attention module to obtain a neural network prediction model, and can accurately predict the grain temperature at each position point in the grain pile.

[0004] In order to achieve these and other advantages of the present invention, according to one aspect of the present invention, there is provided a method for predicting the grain temperature in a storage grain pile, including: obtaining historical meteorological data and temperature data at each position point in the grain pile; inputting the obtained historical meteorological data and temperature data at each position point in the grain pile into a neural network prediction model to obtain temperature data at each position point in the grain pile at a future moment; wherein, the neural network prediction model includes a first branch and a second branch, the first branch includes an input layer, an LSTM network, and a Permute layer, the second branch includes an input layer, an Inception module, and a spatial attention module, and the outputs of the first branch and the second branch are both input into a Concat layer, the Concat layer is connected to a first fully connected layer, and the first fully connected layer is connected to an output layer.

[0005] Further, the first branch includes an input layer, a first Permute layer, an LSTM network, a second Permute layer, a second fully connected layer, a third Permute layer, a Merge layer, and a first Flatten layer connected in sequence, wherein the input of the Merge layer further includes the output from the LSTM network.

[0006] Further, the second branch includes an input layer, a plurality of Inception modules, and a second Flatten layer connected in sequence, wherein a plurality of max pooling layers and a spatial attention module are inserted between the plurality of Inception modules.

[0007] Further, the meteorological data includes air temperature, air pressure, relative humidity, ground temperature at 0 cm, sunshine duration, and precipitation.

[0008] Further, select the meteorological data within the historical period and the temperature data at each position point in the grain pile, and use 80% of the data to establish a training set and 20% of the data to establish a test set.

[0009] According to another invention of the present invention, there is also provided a device for predicting the temperature of stored grain in a grain pile, including: a data acquisition module for acquiring historical meteorological data and the temperature data at each position point in the grain pile; a prediction module for inputting the acquired historical meteorological data and the temperature data at each position point in the grain pile into a neural network prediction model to obtain the temperature data at each position point in the grain pile at a future moment; wherein, the neural network prediction model includes a first branch and a second branch, the first branch includes an input layer, an LSTM network, and a Permute layer, the second branch includes an input layer, an Inception module, and a spatial attention module, and the outputs of the first branch and the second branch are both input into a Concat layer, the Concat layer is connected to a first fully connected layer, and the first fully connected layer is connected to an output layer.

[0010] Further, the first branch includes an input layer, a first Permute layer, an LSTM network, a second Permute layer, a second fully connected layer, a third Permute layer, a Merge layer, and a first Flatten layer connected in sequence, wherein the input of the Merge layer further includes the output from the LSTM network.

[0011] Further, the second branch includes an input layer, a plurality of Inception modules, and a second Flatten layer connected in sequence, wherein a plurality of max pooling layers and a spatial attention module are inserted between the plurality of Inception modules.

[0012] Further, the meteorological data includes air temperature, air pressure, relative humidity, ground temperature at 0 cm, sunshine duration, and precipitation.

[0013] Further, select the meteorological data within the historical period and the temperature data at each position point in the grain pile, and use 80% of the data to establish a training set and 20% of the data to establish a test set.

[0014] The present invention has at least the following beneficial effects:

[0015] The present invention not only considers external meteorological factors but also considers the influence of heat transfer inside the grain pile on grain temperature prediction; the present invention combines the inception module, the temporal attention module, and the spatial attention module to obtain a neural network prediction model, which can accurately predict the grain temperature at each position point in the grain pile, especially at the side boundary position points.

[0016] Other advantages, objectives, and features of the present invention will be partially reflected by the following description and partially understood by those skilled in the art through the research and practice of the present invention. Description of the Drawings

[0017] Figure 1 It is a structural diagram of the neural network prediction model of the present invention;

[0018] Figure 2 Distribution map of the positions of temperature sensors in the grain pile;

[0019] Figure 3 It is the prediction result of the grain temperature point at the first layer, second column of the first row on the boundary side;

[0020] Figure 4 It is the prediction result of the grain temperature point at the third layer, fifth row of the first column on the boundary side;

[0021] Figure 5 It is the prediction result of the grain temperature point at the fourth layer, sixth row of the fifth column on the boundary side;

[0022] Figure 6 It is the prediction result of the grain temperature point at the second layer, third column of the tenth row on the boundary side.

[0023] Figure 7 Ten features with relatively large attention scores when predicting the grain temperature point at the first layer, second column of the first row on the side boundary;

[0024] Figure 8 Ten features with relatively large attention scores when predicting the grain temperature point at the third layer, fifth row of the first column on the side boundary of the grain pile;

[0025] Figure 9 Ten features with relatively large attention scores when predicting the grain temperature point at the fourth layer, sixth row of the fifth column on the side boundary of the grain pile;

[0026] Figure 10 Ten features with relatively large attention scores when predicting the grain temperature point at the second layer, third column of the tenth row on the side boundary of the grain pile;

[0027] Figure 11 Attention scores of 30 time steps when predicting the grain temperature point at the first layer, second column of the first row on the side boundary of the grain pile;

[0028] Figure 12 Attention scores of 30 time steps when predicting the grain temperature point at the third layer, fifth row of the first column on the side boundary of the grain pile;

[0029] Figure 13Attention scores for 30 time steps during the prediction of the grain temperature point at the fourth layer, sixth row, and fifth column of the side boundary of the grain pile.

[0030] Figure 14 Attention scores for 30 time steps during the prediction of the grain temperature point at the second layer, third column, and tenth row of the side boundary of the grain pile. Detailed implementation manner

[0031] The following further elaborates on the present invention with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description in the specification.

[0032] It should be understood that terms such as "having", "including", and "comprising" used herein do not exclude the presence or addition of one or more other elements or their combinations.

[0033] Embodiments of the present application provide a method for predicting the grain temperature in a storage grain pile, including:

[0034] S1. Obtain historical meteorological data and temperature data at each position point in the grain pile; the meteorological data are data such as air temperature, sunlight, and humidity that affect the grain temperature; the temperature data are collected by temperature sensors arranged in the grain pile; the position of the side boundary of the grain pile is relatively close to the wall of the granary, and the temperature points at the side boundary of the grain pile are affected by meteorological factors. Moreover, during the heat transfer process of the grain, the temperature data recorded by each sensor may be related to the temperatures of other points inside the grain pile. Therefore, this embodiment comprehensively considers the influence of meteorological factors and internal heat transfer in the grain pile on the prediction of the grain temperature.

[0035] S2. Input the obtained historical meteorological data and the temperature data at each position point in the grain pile into the neural network prediction model to obtain the temperature data at each position point in the grain pile at a future moment. Among them, the neural network prediction model includes a first branch and a second branch. The first branch includes an input layer, an LSTM network, and a Permute layer. The second branch includes an input layer, an Inception module, and a spatial attention module. The outputs of the first branch and the second branch are both input into a Concat layer, the Concat layer is connected to a first fully connected layer, and the first fully connected layer is connected to an output layer. The first branch constitutes an LSTM time attention module, and a spatial attention module is inserted into the second branch. Here, the LSTM network, Permute layer, Inception module, and spatial attention module are all basic LSTM network, Permute layer, Inception module, and spatial attention module. The Inception module includes 4 branches, which are composed of convolutional layers with convolutional kernel sizes of 1×1, 3×3, and 5×5 and a max pooling layer. After the outputs of the 4 branches are concatenated, they pass through a convolutional layer with a convolutional kernel size of 1×1. The spatial attention module includes two branches, which are composed of an average pooling layer and a max pooling layer. After the outputs of the two branches are concatenated, the output then passes through a convolutional operation to reduce the features to one channel, and then through a sigmoid activation function to finally generate spatial attention features. In this embodiment, the inception module, time attention module, and spatial attention module are combined to obtain a neural network prediction model, including a spatial attention module and a time attention module, which can better consider the influence of the temperature at key position points and time steps on the grain temperature, improve the accuracy of grain temperature prediction, especially improve the prediction accuracy of the side boundary position points.

[0036] In some other embodiments, the first branch includes an input layer, a first Permute layer, an LSTM network, a second Permute layer, a second fully connected layer, a third Permute layer, a Merge layer, and a first Flatten layer connected in sequence. Among them, the input of the Merge layer also includes the output from the LSTM network, that is, the Merge layer fuses the output of the LSTM and the output of the third Permute layer and then inputs it into the first Flatten layer. Figure 1The left half in it is the first branch. First, after the dimension transformation of the grain temperature time series through the Permute layer, it is fed into the LSTM network. The obtained output can be regarded as the features of each time node. Then, after the dimension transformation, it is passed to the fully connected layer. The reason for the dimension transformation here is that if it is directly passed to the fully connected layer, the last dimension is the feature dimension. At this time, the features of each step are separated. When directly performing the fully connected operation, there is no feature exchange between each step of the obtained attention weights, and the result will naturally be inaccurate. Therefore, first, the step dimension of the LSTM output is transferred to the last dimension, and then passed through the fully connected layer to obtain the weights of the attention mechanism based on the features of each step.

[0037] In some other embodiments, the second branch includes an input layer, a plurality of Inception modules, and a second Flatten layer connected in sequence, and a plurality of max pooling layers and a spatial attention module are inserted between the plurality of Inception modules; Figure 1 The right half in it is the second branch. After passing through 10 Inception modules, 1 spatial attention module, and 4 max pooling layers, it is flattened by the second Flatten layer, and is concatenated with the output of the first Flatten layer of the first branch and input into the Concat layer, and then output after passing through a first fully connected layer.

[0038] In some other embodiments, the meteorological data includes air temperature, air pressure, relative humidity, 0 cm ground temperature, sunshine duration, and precipitation. The influence of these meteorological factors on the grain temperature is relatively obvious.

[0039] In some other embodiments, the meteorological data and the temperature data of each position point in the grain pile within a historical period, such as 300 - 1500 days, are selected. 80% of the data is used to establish a training set, and 20% of the data is used to establish a test set.

[0040] The embodiments of the present application also provide a device for predicting the grain temperature in a storage grain pile, including: a data acquisition module for obtaining historical meteorological data and the temperature data of each position point in the grain pile; the temperature data is collected by setting temperature sensors in the grain pile; the position of the side boundary of the grain pile is relatively close to the wall of the granary. The temperature points at the side boundary of the grain pile are affected by meteorological factors, and during the heat transfer process of the grain, the temperature data recorded by each sensor may be related to the temperature of other points inside the grain pile. Therefore, this embodiment comprehensively considers the influence of meteorological factors and internal heat transfer in the grain pile on grain temperature prediction;

[0041] A prediction module for inputting the obtained historical meteorological data and the temperature data at each position point in the grain pile into a neural network prediction model to obtain the temperature data at each position point in the grain pile at a future moment; wherein, the neural network prediction model includes a first branch and a second branch, the first branch includes an input layer, an LSTM network, and a Permute layer, the second branch includes an input layer, an Inception module, and a spatial attention module, the outputs of the first branch and the second branch are both input into a Concat layer, the Concat layer is connected to a first fully connected layer, and the first fully connected layer is connected to an output layer; the first branch forms an LSTM time attention module, and a spatial attention module is inserted into the second branch; the LSTM network, the Permute layer, the Inception module, and the spatial attention module here are all basic LSTM networks, Permute layers, Inception modules, and spatial attention modules; the Inception module includes 4 branches, which are composed of convolutional layers with convolutional kernel sizes of 1×1, 3×3, and 5×5 and a max pooling layer. After the outputs of the 4 branches are concatenated, they pass through a convolutional layer with a convolutional kernel size of 1×1; the spatial attention module includes two branches, which are composed of an average pooling layer and a max pooling layer. After the outputs of the two branches are concatenated, the output then undergoes a convolutional operation with a 7×7 convolutional kernel to reduce the features to one channel, and then passes through a sigmoid activation function, and finally generates spatial attention features; in this embodiment, Inception and LSTM are combined to obtain a neural network prediction model, including a spatial attention module and a time attention module, which can better consider the influence of the temperature at key position points and the time step on the grain temperature, improve the accuracy of grain temperature prediction, especially improve the prediction accuracy of the side boundary position points.

[0042] In some other embodiments, the first branch includes an input layer, a first Permute layer, an LSTM network, a second Permute layer, a second fully connected layer, a third Permute layer, a Merge layer, and a first Flatten layer connected in sequence, wherein the input of the Merge layer further includes the output from the LSTM network; Figure 1 The left half part is the first branch. First, after the grain temperature time series undergoes a dimensional transformation through the Permute layer, it is passed into the LSTM network. The obtained output can be regarded as the features of each time node. Then, after another dimensional transformation, it is passed to the fully connected layer. The reason for the dimensional transformation here is that if it is directly passed to the fully connected layer, the last dimension is the feature dimension. At this time, the features of each step are separated. When directly performing a fully connected operation, there is no feature exchange between each step of the attention weights obtained, and the result is naturally inaccurate. Therefore, first, the step dimension of the LSTM output is transferred to the last dimension, and then passed through the fully connected layer to obtain the weights of the attention mechanism based on the features of each step.

[0043] In some other embodiments, the second branch includes an input layer, a plurality of Inception modules, and a second Flatten layer connected in sequence, wherein a plurality of max pooling layers and a spatial attention module are inserted between the plurality of Inception modules; Figure 1 The right half of [the figure] is the second branch, which is flattened by the second Flatten layer after passing through a plurality of Inception modules and a plurality of max pooling layers, and is output together with the output of the first Flatten layer of the first branch and input into the Concat layer for splicing, and then output after passing through a first fully connected layer.

[0044] In some other embodiments, the meteorological data includes air temperature, air pressure, relative humidity, 0 cm ground temperature, sunshine duration, and precipitation, and the influence of these meteorological factors on the grain temperature is relatively obvious.

[0045] In some other embodiments, meteorological data and temperature data at each position point in the grain pile within a historical period, such as 300 to 1500 days, are selected, and 80% of the data is used to establish a training set, and 20% of the data is used to establish a test set.

[0046] The following is illustrated by specific embodiments.

[0047] Embodiment 1:

[0048] For the Zhumadian Direct Depot, the input is 8 meteorological factors (air temperature, air pressure, relative humidity, 0 cm ground temperature, sunshine duration, precipitation) and the temperature time series recorded by 200 temperature sensors in the grain pile, which is a total of 208-dimensional feature vectors, and the selected time step is 30. Select data with a time series length of 1265 days, and select 80% of the historical meteorological data and grain temperature data to train Figure 1 the neural network prediction model shown, and 20% of the data is used for testing. For comparison, the neural network prediction model is also trained in the same way using only 8 meteorological factors and the temperature time series recorded by 200 temperature sensors.

[0049] Figure 2 The position distribution of the temperature sensors in the grain pile is given. The first value in the brackets represents the row number in the grain pile, the second value represents the column number in the grain pile, and the third value represents the layer number in the grain pile. For example, (1,1,1) represents the point at the first row, first column, and first layer in the grain pile, and (2,1,1) represents the point at the second row, first column, and first layer in the grain pile.

[0050] Figure 3The prediction results of the grain temperature points at the first layer of the second column in the first row on the side of the boundary are shown. The root mean square error between the predicted value (Proposed model) and the true value (observation) is 0.4879. The prediction accuracy is 25% higher than that of predicting solely using meteorological factors and 31% higher than that of predicting solely using the temperature time series recorded by the sensor network inside the grain pile. Figure 4 The prediction results of the grain temperature points at the third layer of the fifth row in the first column on the side of the boundary are shown. The root mean square error between the predicted value and the true value is 0.525. The prediction accuracy is 0.3% higher than that of predicting solely using meteorological factors and 12% higher than that of predicting solely using the temperature time series recorded by the sensor network inside the grain pile. Figure 5 The prediction results of the grain temperature points at the fourth layer of the sixth row in the fifth column on the side of the boundary are shown. The root mean square error between the predicted value and the true value is 0.2227. The prediction accuracy is 47% higher than that of predicting solely using meteorological factors and 23% higher than that of predicting solely using the temperature time series recorded by the sensor network inside the grain pile. Figure 6 The prediction results of the grain temperature points at the second layer of the third column in the tenth row on the side of the boundary are shown. The root mean square error between the predicted value and the true value is 0.7014. The prediction accuracy is 12% higher than that of predicting solely using meteorological factors and 17% higher than that of predicting solely using the temperature time series recorded by the sensor network inside the grain pile. The results show that when using both meteorological factors and all the temperature time series recorded by the sensor network in the grain pile as the features of the temperature points on the side boundary of the grain pile to predict the temperature points at different positions on the side boundary of the grain pile, the change trends of the predicted value and the true value are consistent, and the prediction accuracy is improved compared with the previous two cases, indicating that using comprehensive features can more effectively express the prediction target and verifying the rationality of the design of the neural network prediction model in the embodiment.

[0051] During the model training process, it is found that the points with relatively large attention scores obtained by the spatial attention module are also some temperature points distributed around the target point positions, and the attention scores of meteorological factors are relatively small. The results show that when using meteorological factors and all the temperature time series recorded by the sensor network in the grain pile to predict the temperature points on the side boundary of the grain pile, the temperature time series in the grain pile plays a more important role.

[0052] Figure 7 Ten features with relatively large attention scores are shown when predicting the grain temperature points at the first layer of the second column in the first row on the side boundary. Figure 8 Ten features with relatively large attention scores are shown when predicting the grain temperature points at the third layer of the fifth row in the first column on the side boundary of the grain pile. Figure 9Shows the 10 features with relatively large attention scores when predicting the grain temperature point at the fourth layer, sixth row, and fifth column of the side boundary of the grain heap. Figure 10 Shows the 10 features with relatively large attention scores when predicting the grain temperature point at the second layer, third column, and tenth row of the side boundary of the grain heap. These features are all temperature points located in the same row, same column, or same layer adjacent to the prediction point. The results show that for Zhumadian Directly-administered Depot, when using the mixed features of meteorological factors and the grain heap sensor network to predict the temperature at the side boundary of the grain heap, the spatial attention module not only focuses on the temperature points closer to the target point.

[0053] Figure 11 Are the attention scores for 30 time steps when predicting the grain temperature point at the first layer, second column, and first row of the side boundary of the grain heap. The attention module pays more attention to the features at time steps 16 and 29. Figure 12 Are the attention scores for 30 time steps when predicting the grain temperature point at the third layer, fifth row, and first column of the side boundary of the grain heap. The attention module pays more attention to the features at time steps 16 and 30. Figure 13 Are the attention scores for 30 time steps when predicting the grain temperature point at the fourth layer, sixth row, and fifth column of the side boundary of the grain heap. The attention module pays more attention to the features at time steps 15, 27, and 29. Figure 14 Are the attention scores for 30 time steps when predicting the grain temperature point at the second layer, third column, and tenth row of the side boundary of the grain heap. The attention module pays more attention to the features at time steps 17 and 30. The results show that when using the mixed features of meteorological factors and the grain heap sensor network to predict the temperature of different positions at the side boundary of the grain heap, the time attention module pays more attention to the features of the time steps at the middle time points and the features of the last few time steps. This is the result of the combined action of the meteorological factors and the temperature time series recorded by the grain heap sensor network. The time attention module pays attention to the features of the previous time steps, probably because the change of grain temperature with meteorological factors has a delay. When the meteorological factors change, the change of grain temperature cannot be immediately manifested, but it takes a certain period of time to show its change law. This is related to the poor thermal conductivity of the grain, and the difference in the change of the grain temperature point at the position of the sensor in the grain heap in time is not large. Therefore, when predicting the temperature of the side boundary point of the grain heap at this moment, the time attention module includes both the features of a certain previous moment and the features of the time steps closer to this moment.

[0054] Example 2:

[0055] Collect data with a time series length of 423 days from the grain heap of Kunming Directly-administered Depot. Select 80% of the historical meteorological data and grain temperature data for training, and 20% of the data for testing, and establish a prediction model (refer toFigure 1 ) The input of the model is 8 meteorological factors and the time series recorded by 225 temperature sensors in the grain pile, which is a total of 233-dimensional feature vectors. The selected time step is 30. For comparison, the neural network prediction model is also trained separately with 8 meteorological factors and the temperature time series recorded by 225 temperature sensors in the same way.

[0056] The results show that for the prediction result of the grain temperature point at the second layer, first column, first row of the side boundary, the root mean square error between the predicted value and the true value is 1.509. For the prediction result of the grain temperature point at the third layer, fifth column, first row of the side boundary, the root mean square error between the predicted value and the true value is 1.251. For the prediction result of the grain temperature point at the fourth layer, second column, ninth row of the side boundary, the root mean square error between the predicted value and the true value is 1.281. For the prediction result of the grain temperature point at the fifth layer, fifth column, eighth row of the side boundary, the root mean square error between the predicted value and the true value is 1.098. That is, for the temperature points at different positions on the side boundary of the grain pile, the change trends of the predicted values and the true values are consistent.

[0057] The number of devices and the processing scale described here are used to simplify the description of the present invention. The application, modification, and variation of the method and device for predicting the grain temperature in the storage grain pile of the present invention are obvious to those skilled in the art.

[0058] Although the embodiments of the present invention have been disclosed as above, it is not limited to the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily achieved. Therefore, without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to the specific details and the illustrated and described examples here.

Claims

1. A method for predicting the grain temperature of a stored grain pile, characterized in that, Including: Obtain historical meteorological data and temperature data at each position point in the grain pile; Input the obtained historical meteorological data and temperature data at each position point in the grain pile into the neural network prediction model to obtain the temperature data at each position point in the grain pile at a future moment; Among them, the neural network prediction model includes a first branch and a second branch. The first branch includes an input layer, a first Permute layer, an LSTM network, a second Permute layer, a second fully connected layer, a third Permute layer, a Merge layer, and a first Flatten layer connected in sequence. The input of the Merge layer also includes the output from the LSTM network; The second branch includes an input layer, multiple Inception modules, and a second Flatten layer connected in sequence. Multiple max pooling layers and a spatial attention module are inserted between the multiple Inception modules; The outputs of the first branch and the second branch are both input into the Concat layer. The Concat layer is connected to the first fully connected layer, and the first fully connected layer is connected to the output layer; The meteorological data includes air temperature, air pressure, relative humidity, 0 cm ground temperature, sunshine duration, and precipitation.

2. The method for predicting the grain temperature of a stored grain pile according to claim 1, wherein Select the meteorological data and temperature data at each position point in the grain pile within the historical interval, and use 80% of the data to establish a training set and 20% of the data to establish a test set.

3. Grain temperature prediction device for stored grain piles, characterized in that, Including: A data acquisition module for obtaining historical meteorological data and temperature data at each position point in the grain pile; A prediction module for inputting the obtained historical meteorological data and temperature data at each position point in the grain pile into the neural network prediction model to obtain the temperature data at each position point in the grain pile at a future moment; Among them, the neural network prediction model includes a first branch and a second branch. The first branch includes an input layer, a first Permute layer, an LSTM network, a second Permute layer, a second fully connected layer, a third Permute layer, a Merge layer, and a first Flatten layer connected in sequence. The input of the Merge layer also includes the output from the LSTM network; the second branch includes an input layer, multiple Inception modules, and a second Flatten layer connected in sequence. Multiple max pooling layers and a spatial attention module are inserted between the multiple Inception modules; The outputs of the first branch and the second branch are both input into the Concat layer. The Concat layer is connected to the first fully connected layer, and the first fully connected layer is connected to the output layer; The meteorological data includes air temperature, air pressure, relative humidity, 0 cm ground temperature, sunshine duration, and precipitation.

4. The grain temperature prediction device for stored grain stacks according to claim 3, wherein, Select the meteorological data and temperature data at each position point in the grain pile within the historical interval, and use 80% of the data to establish a training set and 20% of the data to establish a test set.

Images