CN118277959A - Pig house temperature prediction method based on resonance sparse Transformer network - Google Patents

Abstract

The invention belongs to the field of agricultural information processing, in particular to a pig house temperature prediction method fused with a resonance sparse transform network, which utilizes a resonance sparse decomposition method to decompose temperature sequence data of each temperature acquisition measuring point to obtain a low-frequency temperature trend sequence and a high-frequency fluctuation sequence of each temperature acquisition measuring point; predicting each low-frequency temperature trend sequence to obtain a low-frequency temperature prediction sequence of each temperature acquisition measuring point; predicting each high-frequency fluctuation sequence to obtain a high-frequency temperature prediction sequence of each temperature acquisition measuring point; and summing and calculating the low-frequency temperature prediction sequence and the high-frequency temperature prediction sequence to obtain the temperature prediction data of each final temperature acquisition measuring point, and further regulating and controlling the pigsty environment. The method considers the characteristics of low-frequency trend and high-frequency oscillation of the intensive pig house temperature sequence data, can effectively capture the long-term dependency relationship in the time sequence data, and has low calculation complexity and high prediction precision.

Description

Pig house temperature prediction method based on resonance sparse Transformer network

Technical Field

The invention belongs to the technical field of agricultural engineering and information processing, and in particular relates to a pig house temperature prediction method integrating a resonance sparse Transformer network.

Background technique

In the intensive, large-scale and IoT-monitored pig farming process, the temperature of the pig house has an important impact on the survival rate of pig farming, pork growth and metabolism, disease transmission (such as diarrhea, respiratory diseases, pneumonia, fungal reproduction, etc.). Therefore, accurately predicting the temperature of the pig house, timely regulating the pig house environment according to the temperature fluctuation trend, and formulating early warning and epidemic prevention plans are of great significance for reducing energy consumption, reducing feed consumption, improving breeding survival rate and disease prevention in intensive pig farming.

At present, intensive pig houses are closed microclimate environments. Affected by the internal and external environments, the internal temperature of the pig houses fluctuates in a small range throughout the day, and the data distribution has characteristics such as sharp peaks, thick tails, and multiple coupling. Traditional model-driven prediction methods and data-driven prediction methods, such as time series prediction methods and neural network prediction methods, are affected by the application climate scenarios, data sample size, and model parameters. The model stability and prediction accuracy do not meet the actual engineering application requirements.

Summary of the invention

The purpose of an embodiment of the present invention is to provide a pig house temperature prediction method integrating a resonant sparse Transformer network, aiming to solve the problem that traditional prediction methods are affected by application climate scenarios, data sample size and model parameters, and the model stability and prediction accuracy do not meet the requirements of practical engineering applications.

To achieve the above object, the present invention provides the following technical solutions.

Specifically, the present invention provides a method for predicting pig house temperature by integrating a resonance sparse Transformer network, the method comprising the following steps:

Step S1: setting a plurality of temperature collection points in the pig house, and picking up the temperature sequence data of each temperature collection point;

Step S2: Decomposing the temperature series data of each temperature collection point by using the resonance sparse decomposition method to obtain a low-frequency temperature trend series and a high-frequency fluctuation series of each temperature collection point;

Step S3: using the Transformer network model method to predict each low-frequency temperature trend sequence, and obtain a low-frequency temperature prediction sequence for each temperature collection point;

Step S4: using a convolutional neural network bidirectional long short-term memory network CNN-BiLSTM model to predict each high-frequency fluctuation sequence, and obtain a high-frequency temperature prediction sequence for each temperature collection point;

Step S5: The low-frequency temperature prediction sequence and the high-frequency temperature prediction sequence are summed and calculated to obtain the final temperature prediction data of each temperature collection point, and the error between the predicted time series and the actual time series is calculated at the same time, and the hyperparameters of the prediction model are adjusted in real time;

Step S6: According to the temperature prediction data, the pig house environment is regulated.

Furthermore, step S2 specifically includes:

Step S21: The temperature series data of each measuring point in the pig house collected by the temperature sensor is expressed as:

(1);

In formula (1), is the observed data, ; , are the trend sequence component of the low-frequency oscillation characteristic to be estimated and the harmonic component of the high-frequency oscillation characteristic;

Step S22: Signal Components , Using overcomplete wavelet basis Overcomplete wavelet basis The matching is characterized by: , (2);

In formula (2), and Respectively The wavelet transform coefficients of

Step S23: The constructed resonance sparse decomposition objective function is expressed as:

(3);

In formula (3), and is the regularization parameter;

Step S24: Apply the split augmented Lagrangian shrinkage algorithm to solve the minimum value of the constructed resonance sparse decomposition objective function, and iterate and update multiple times and , get the updated wavelet transform coefficients and , the trend sequence component of the low-frequency oscillation characteristics to be estimated and the harmonic component of the high-frequency oscillation characteristics are:

(4).

Furthermore, step S3 specifically includes:

Step S31: The Transformer model is composed of a model encoder and a model decoder; wherein the model encoder includes vector position encoding, multi-head self-attention mechanism, residual connection and network layer normalization processing and feedforward neural network; the model decoder includes mask multi-head self-attention mechanism, feedforward neural network, and fully connected layer;

Step S32: In the model encoder, vector position encoding is used to add label information to each position in the input sequence X to distinguish different positions and orders of the input sequence;

Step S33: In the model decoder, in order to predict that the data of a certain step will not be connected with the future data, a mask multi-head self-attention mechanism is constructed by creating a mask matrix to construct the future position, so that the attention score of the future position is set to infinitesimal, so that the current element is only connected with the historical elements, ensuring that the model can only rely on historical elements to predict the value of future elements.

Further, in step S32:

Step S321: Use the linear transformation of sin and cos functions to provide model position information. The specific operation is:

(5);

In formula (5), is the position of the input sequence, such as =0, 1, 2, ..., N; is the sequence dimension; and Indicates the parity of the sequence dimension; is the size of the embedding space dimension;

Step S322: In the model encoder, the multi-head self-attention mechanism uses multiple parallel self-attention mechanisms. A single self-attention mechanism captures the information of the subspace by learning different weights. The specific operations are:

According to the vector position encoding, add the position encoded vector through three weight matrices: query matrix , the bond matrix Sum Matrix , transformed into the vector Q, vector K and vector V required by the self-attention mechanism: , , ,in, is the input vector after adding the position encoding;

Step S323: Use the dot product method to calculate the correlation score between each element in the input sequence: Score=Q*KT;

In order to stabilize the gradient during model training, the correlation scores between each element are normalized: ,in, is the dimension of vector K; through the Softmax function, the score vector between each element in the position-encoded vector is converted into a probability distribution between [0,1], and the calculation formula is:

(6);

Step S324: The multi-head attention mechanism uses multiple sets of weight matrices ( , , ), get multiple sets of required vectors Q, vectors K and vectors V, and get the outputs of multiple heads connected together to perform linear transformation matrix Z:

(7);

In formula (7), , , is the i-th vector Q, K and V; is the attention head weight matrix;

Step S325: residual connection and network layer normalization processing are as follows:

After the multi-head attention mechanism output in the previous step, the residual connection operation is performed Normalization operation with network layer ;

Step S326: The feedforward neural network is a two-layer neural network, which first performs a linear transformation, then a ReLU nonlinear transformation, and then a linear transformation, specifically:

(8);

In formula (8), The output of the previous layer , and is the weight coefficient of the feedforward neural network, and It is the bias of the feedforward neural network, and finally the residual connection method is used to connect each layer .

Further, in step S33:

The specific masked multi-head self-attention mechanism is: (9);

In formula (9), Mmask is the mask matrix, Attention weight matrix with mask;

The feedforward neural network and fully connected layer in the model decoder Decoder. The feedforward neural network performs nonlinear transformation on the upper layer results, allowing the network to capture and represent complex nonlinear relationships. The fully connected layer uses the ReLU function as the activation function.

Compared with the prior art, the pig house temperature prediction method integrating the resonance sparse Transformer network of the present invention has the following beneficial effects:

First, the pig house temperature prediction method proposed in the present invention takes into account the low-frequency trend and high-frequency oscillation characteristics of the intensive pig house temperature series data, and can accurately extract the pig house temperature fluctuation trend to prevent the prediction trend from being distorted;

Second, compared with the time series prediction method and the data-driven prediction method, the method proposed in the present invention can effectively capture the long-term dependencies in the time series data, with low computational complexity, high prediction accuracy, and fast algorithm operation speed.

In summary, the present invention overcomes the problem of inaccuracy of traditional prediction methods, especially for prediction problems under characteristics such as small-range data fluctuations, data distribution peaks and thick tails, and multiple coupling. The pig house temperature prediction method of the present invention can realize accurate prediction of intensive pig house temperature and provide a theoretical basis for fine control of pig house environment and disease prevention and control.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can be obtained based on these drawings without paying creative work.

FIG1 is a principle framework diagram of a pig house temperature prediction method integrating a resonance sparse Transformer network according to the present invention;

FIG2 is a schematic diagram of a temperature change waveform of a certain measuring point in a pig house collected according to an embodiment of the present invention;

FIG3 is a low-frequency component diagram obtained by resonant sparse decomposition according to an embodiment of the present invention;

FIG4 is a diagram of high frequency components obtained by resonant sparse decomposition according to an embodiment of the present invention;

5 is a schematic diagram of the low-frequency temperature and humidity prediction results of a certain measuring point in an intensive piggery predicted based on the Transformer network model according to an embodiment of the present invention;

6 is a schematic diagram of the high-frequency temperature prediction results of a certain measuring point in an intensive piggery predicted by a bidirectional long short-term memory network CNN-BiLSTM model based on a convolutional neural network according to an embodiment of the present invention;

FIG7 is a temperature prediction result of a certain measuring point in a pig house predicted by the pig house temperature prediction method of the present invention;

FIG8 is a temperature prediction error based on the pig house temperature prediction method of the present invention;

FIG9 is a flow chart of a method for predicting pig house temperature by integrating a resonant sparse Transformer network according to the present invention.

Detailed ways

In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

At present, intensive pig houses are closed microclimate environments. Affected by the internal and external environments, the internal temperature of the pig houses fluctuates in a small range throughout the day, and the data distribution has characteristics such as sharp peaks, thick tails, and multiple coupling. Traditional model-driven prediction methods and data-driven prediction methods, such as time series prediction methods and neural network prediction methods, are affected by the application climate scenarios, data sample size, and model parameters. The model stability and prediction accuracy do not meet the actual engineering application requirements.

To solve the above problems, the present invention proposes a pig house temperature prediction method that integrates a resonant sparse Transformer network. The method takes into account the low-frequency trend and high-frequency oscillation characteristics of intensive pig house temperature series data, and can accurately extract the temperature fluctuation trend of the pig house to prevent the prediction trend from being distorted. Compared with the time series prediction method and the data-driven prediction method, the method proposed in the present invention can effectively capture the long-term dependencies in the time series data, with low computational complexity, high prediction accuracy, and fast algorithm running speed.

The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

As shown in FIG. 1 and FIG. 9 , the pig house temperature prediction method provided by the present invention by integrating the resonance sparse Transformer network comprises the following steps:

Step S1: setting a plurality of temperature collection points in the pig house, and picking up the temperature sequence data of each temperature collection point;

Step S2: Decomposing the temperature series data of each temperature collection point by using the resonance sparse decomposition method to obtain a low-frequency temperature trend series and a high-frequency fluctuation series of each temperature collection point;

Step S3: using the Transformer network model method to predict each low-frequency temperature trend sequence, and obtain a low-frequency temperature prediction sequence for each temperature collection point;

Step S4: using a convolutional neural network bidirectional long short-term memory network CNN-BiLSTM model to predict each high-frequency fluctuation sequence, and obtain a high-frequency temperature prediction sequence for each temperature collection point;

Step S5: The low-frequency temperature prediction sequence and the high-frequency temperature prediction sequence are summed and calculated to obtain the final temperature prediction data of each temperature collection point, and the error between the predicted time series and the actual time series is calculated at the same time, and the hyperparameters of the prediction model are adjusted in real time;

Step S6: According to the temperature prediction data, the pig house environment is regulated.

Furthermore, step S2 specifically includes:

Step S21: The temperature series data of each measuring point in the pig house collected by the temperature sensor is expressed as:

(1);

In formula (1), is the observed data, ; , are the trend sequence component of the low-frequency oscillation characteristic to be estimated and the harmonic component of the high-frequency oscillation characteristic;

Step S22: Signal Components , Using overcomplete wavelet basis Overcomplete wavelet basis The matching is characterized by:

, (2);

In formula (2), and Respectively The wavelet transform coefficients of

Step S23: The constructed resonance sparse decomposition objective function is expressed as:

(3);

In formula (3), and is the regularization parameter;

Step S24: Apply the split augmented Lagrangian shrinkage algorithm to solve the minimum value of the constructed resonance sparse decomposition objective function, and iterate and update multiple times and , get the updated wavelet transform coefficients and , the trend sequence component of the low-frequency oscillation characteristics to be estimated and the harmonic component of the high-frequency oscillation characteristics are:

(4);

Furthermore, step S3 specifically includes:

Step S31: The Transformer model is composed of a model encoder Encoder and a model decoder Decoder; wherein the model encoder Encoder includes vector position encoding, multi-head self-attention mechanism, residual connection and network layer normalization processing and feedforward neural network; the model decoder Decoder includes a masked multi-head self-attention mechanism, a feedforward neural network, and a fully connected layer;

Step S32: In the model encoder, vector position encoding is used to add label information to each position in the input sequence X to distinguish different positions and orders of the input sequence;

Step S33: In the model decoder, in order to predict that the data of a certain step will not be connected with the future data, a mask multi-head self-attention mechanism is constructed by creating a mask matrix to construct the future position, so that the attention score of the future position is set to infinitesimal, so that the current element is only connected with the historical elements, ensuring that the model can only rely on historical elements to predict the value of future elements.

Further, step S32 includes:

Step S321: Use the linear transformation of sin and cos functions to provide model position information. The specific operation is:

(5);

In formula (5), is the position of the input sequence, such as =0, 1, 2, ..., N; is the sequence dimension; and Indicates the parity of the sequence dimension; is the size of the embedding space dimension;

Step S322: In the model encoder, the multi-head self-attention mechanism uses multiple parallel self-attention mechanisms. A single self-attention mechanism captures the information of the subspace by learning different weights. The specific operations are:

According to the vector position encoding, the position-encoded vector is added through three weight matrices ( , , ), that is, the query matrix , the bond matrix , value matrix Transformed into the vector Q, vector K and vector V required by the self-attention mechanism, that is , , , is the input vector after adding the position encoding;

Step S323: Use the dot product method to calculate the correlation score between each element in the input sequence: Score=Q*KT;

In order to stabilize the gradient during model training, the correlation scores between each element are normalized: ,in, is the dimension of vector K; through the Softmax function, the score vector between each element in the position-encoded vector is converted into a probability distribution between [0,1], and the calculation formula is:

(6);

Step S324: The multi-head attention mechanism uses multiple sets of weight matrices ( , , ), get multiple sets of required vectors Q, vectors K and vectors V, and get the outputs of multiple heads connected together to perform linear transformation matrix Z:

(7);

In formula (7), , , is the i-th vector Q, K and V; is the attention head weight matrix;

Step S325: residual connection and network layer normalization processing are as follows:

After the multi-head attention mechanism output in the previous step, the residual connection operation is performed Normalization operation with network layer ;

Step S326: The feedforward neural network is a two-layer neural network, which first performs a linear transformation, then a ReLU nonlinear transformation, and then a linear transformation, specifically:

(8);

In formula (8), The output of the previous layer , and is the weight coefficient of the feedforward neural network, and It is the bias of the feedforward neural network, and finally the residual connection method is used to connect each layer .

Further, in step S33:

The specific masked multi-head self-attention mechanism is: (9);

In formula (9), Mmask is the mask matrix, Attention weight matrix with mask;

The feedforward neural network and fully connected layer in the model decoder Decoder. The feedforward neural network performs nonlinear transformation on the upper layer results, so that the network can capture and represent complex nonlinear relationships. The fully connected layer uses the ReLU function as the activation function, and the calculation process is similar to that of the model encoder Encoder.

The pig house temperature prediction method proposed in the present invention takes into account the low-frequency trend and high-frequency oscillation characteristics of the intensive pig house temperature series data, and can accurately extract the pig house temperature fluctuation trend to prevent the prediction trend from being distorted;

Compared with the time series prediction method and the data-driven prediction method, the method proposed in the present invention can effectively capture the long-term dependencies in the time series data, has low computational complexity, high prediction accuracy, and fast algorithm operation speed.

In summary, the present invention overcomes the problem of inaccuracy of traditional prediction methods, especially for prediction problems under characteristics such as small-range data fluctuations, data distribution peaks and thick tails, and multiple coupling. The pig house temperature prediction method of the present invention can realize accurate prediction of intensive pig house temperature and provide a theoretical basis for fine control of pig house environment and disease prevention and control.

The present invention adopts a certain intensive pig breeding enterprise in Wuhu to pick up the temperature series data of a certain measuring point based on Ethernet and sensor fusion network technology. The temperature change sequence of a certain measuring point for about 11 months is shown in Figure 2. One temperature point is collected every day, and the data totals 333 points.

The temperature time series of a certain measuring point is decomposed by using the resonance sparse decomposition method {refer to I. W. Selesnick, Resonance-based signal decomposition: a new sparsity-enabled signal analysis method, Signal Processing, 2011, 91(12): 2793-2809.} to obtain a low-frequency temperature trend sequence and a high-frequency fluctuation sequence of the measuring point. FIG3 is a low-frequency component diagram obtained by the resonance sparse decomposition of an embodiment of the present invention, and FIG4 is a high-frequency component diagram obtained by the resonance sparse decomposition of an embodiment of the present invention. It can be seen that the low-frequency component signal reflects the temperature fluctuation trend of the pig house, and the high-frequency component signal reflects the influence of external interference factors on the fluctuation of the ambient temperature.

The Transformer network model {reference A. Vaswani, N. Shazeer, N. Parmar, J.Uszkoreit, L. Jones, A.N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need [J]. Advances in neural informationprocessing systems, 2017, 30, 1-15. } is used to predict the low-frequency temperature trend sequence of a certain measuring point, and the low-frequency temperature prediction sequence of a certain measuring point is obtained. The result is shown in Figure 5. It can be seen that the Transformer network model method can accurately track the temperature fluctuation trend of the pig house.

The bidirectional long short-term memory network (CNN-BiLSTM) model of a convolutional neural network is used to predict the high-frequency fluctuation sequence of a certain measuring point, and the high-frequency temperature prediction sequence of a certain measuring point is obtained. The result is shown in FIG6 . It can be seen that the CNN-BiLSTM model can accurately track the oscillation process of the high-frequency temperature component. Finally, the low-frequency temperature prediction sequence and the high-frequency temperature prediction sequence are summed and calculated to obtain the temperature prediction data of a certain measuring point, as shown in FIG7 . FIG8 is a temperature prediction error based on the method of the present invention according to an embodiment of the present invention. It can be seen that the prediction error range of the method proposed by the present invention fluctuates slightly, the prediction accuracy is high, and the method has good industrial application value.

Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scope, the present invention is not limited to the specific details and the illustrations shown and described here.

Claims (5)

1. The pig house temperature prediction method fused with the resonance sparse transducer network is characterized by comprising the following steps of:

step S1: setting a plurality of temperature acquisition measuring points in a pig house, and picking up temperature sequence data of each temperature acquisition measuring point;

Step S2: decomposing the temperature sequence data of each temperature acquisition measuring point by using a resonance sparse decomposition method to obtain a low-frequency temperature trend sequence and a high-frequency fluctuation sequence of each temperature acquisition measuring point;

Step S3: predicting each low-frequency temperature trend sequence by using a transducer network model method to obtain a low-frequency temperature prediction sequence of each temperature acquisition measuring point;

Step S4: predicting each high-frequency fluctuation sequence by using a bidirectional long-short-time memory network CNN-BiLSTM model of the convolutional neural network to obtain a high-frequency temperature prediction sequence of each temperature acquisition measuring point;

Step S5: summing and calculating the low-frequency temperature prediction sequence and the high-frequency temperature prediction sequence to obtain temperature prediction data of each final temperature acquisition measuring point, calculating errors of a prediction time sequence and an actual time sequence, and adjusting super parameters of a prediction model in real time;

Step S6: and regulating and controlling the pigsty environment according to the temperature prediction data.

2. The pig house temperature prediction method of the fused resonance sparse fransformer network according to claim 1, wherein the step S2 specifically comprises:

step S21: the temperature sequence data of each measuring point of the pigsty collected by the temperature sensor is expressed as follows:

（1）；

In the formula (1), the components are as follows, In order to observe the data of the object,；、The trend sequence component of the low-frequency oscillation characteristic to be estimated and the harmonic component of the high-frequency oscillation characteristic are adopted;

Step S22: signal component 、Using overcomplete wavelet basisAnd overcomplete wavelet basisThe matching is characterized by:

； （2）；

In the formula (2), the amino acid sequence of the compound, And (3) withRespectively as componentsWavelet transform coefficients of (a);

Step S23: the constructed resonance sparse decomposition objective function is expressed as:

（3）；

In the formula (3), the amino acid sequence of the compound, And (3) withIs a regularization parameter;

Step S24: solving and constructing minimum value of resonance sparse decomposition objective function by using split augmentation Lagrangian contraction algorithm, and repeatedly updating And (3) withObtaining updated wavelet transform coefficientsAnd (3) withThe trend sequence component of the low-frequency oscillation characteristic to be estimated and the harmonic component of the high-frequency oscillation characteristic are respectively:

（4）。

3. The pig house temperature prediction method of the fused resonance sparse fransformer network according to claim 2, wherein the step S3 specifically comprises:

Step S31: the transducer model consists of a model encoder Encoder and a model Decoder; the model encoder Encoder comprises vector position coding, a multi-head self-attention mechanism, residual connection, network layer normalization processing and a feedforward neural network; the model Decoder comprises a mask multi-head self-attention mechanism, a feedforward neural network and a full-connection layer;

Step S32: in the model encoder Encoder, vector position coding is used to add marker information to each position in the input sequence X, distinguishing between different positions and order of the input sequence;

step S33: in the model Decoder, a masking multi-headed self-attention mechanism is used to construct future locations by creating a masking matrix such that the attention score for the future locations is infinitely small, causing the current element to be tied only to historical elements.

4. The pig house temperature prediction method fused with the resonance sparse fransformer network according to claim 3, wherein in step S32:

Step S321: linear transformation using sin and cos functions provides model position information, specifically operating as:

（5）；

In the formula (5), the amino acid sequence of the compound, For inputting the position of a sequence, e.g.=0，1,2,…,N；Is the sequence dimension; And Representing parity of sequence dimensions; The size of the embedded space dimension;

Step S322: in model encoder Encoder, the multi-headed self-attention mechanism uses multiple parallel self-attention mechanisms, a single self-attention mechanism captures subspace information by learning different weights, specifically operating as:

adding the vector after the position coding according to the vector position coding, and passing through three weight matrixes: query matrix Key matrixSum matrixThe transformation is to the vector Q, the vector K and the vector V required by the self-attention mechanism respectively:

，， Wherein, the method comprises the steps of, wherein, For adding the input vector after the position coding;

Step S323: a dot product method is used to calculate a correlation score between each element in the input sequence: score=q×kt;

Normalizing the relevance scores among the elements: Wherein, the method comprises the steps of, wherein, Is the dimension of vector K;

Converting the score vector between each element in the position coded vector into probability distribution between [0,1] through a Softmax function, wherein the calculation formula is as follows:

（6）；

Step S324: multi-head attention mechanism uses multiple sets of weight matrices ,,Obtaining a plurality of groups of required vectors Q, vectors K and vectors V, and connecting the outputs of the plurality of heads to perform linear transformation matrix Z:

（7）；

In the formula (7), the amino acid sequence of the compound, ,,Is the ith vector Q, K and V; a weight matrix for the attention head;

Step S325: the residual connection and network layer normalization process is as follows:

after the multi-head attention mechanism is output in the last step, residual connection operation is carried out Normalizing operation with network layer；

Step S326: the feedforward neural network is a two-layer neural network, which is transformed first, then is subjected to non-linear transformation of ReLU, and then is subjected to linear transformation, specifically:

（8）；

In the formula (8), the amino acid sequence of the compound, For the output of the previous layer，And (3) withAs the weighting coefficients of the feed-forward neural network,And (3) withFor biasing the feedforward neural network, connecting all layers by residual connection mode。

5. The method for pig house temperature prediction by fusing a resonance sparse fransformer network according to claim 4, wherein in step S33:

the mask multi-head self-attention mechanism is specifically: (9)；

In equation (9), mmask is a mask matrix, Attention weight matrix with mask;

And the feedforward neural network in the model Decoder performs nonlinear conversion on the upper layer result, so that the network can capture the nonlinear relation representing the complexity, and the full connection layer uses a ReLU function as an activation function.