Plant liquid flow prediction method based on multi-dimensional environment sequence and graph convolutional network

Abstract

The invention discloses a plant liquid flow prediction method based on a multi-dimensional environment sequence and a graph convolutional network. The method comprises the steps that 1, environment data and liquid flow data in a certain period of time are synchronously collected and preprocessed; 2, establishing a plant liquid flow prediction network based on a multi-dimensional environment sequence and a graph convolutional network, including a Fourier denoising module, a feature distillation module, a GCN module, a multi-head attention module and a prediction head, and processing the sample to obtain a liquid flow prediction value; and 3, on the basis of the minimum mean square error, obtaining an optimal plant liquid flow prediction model. According to the method, the prediction precision of the plant liquid flow can be improved to meet the requirements of practical application.

Description

Technical Field

[0001] The present invention relates to the technical field of plant physiological parameter prediction, and in particular to a plant sap flow prediction method based on multidimensional environmental sequence and graph convolutional network. Background Art

[0002] As an important physiological parameter of plants, sap flow not only reflects the absorption and utilization efficiency of water resources by plants, but also plays a key role in the soil-plant-atmosphere continuous water cycle system. Sap flow itself has significant nonlinear characteristics and has a highly complex coupling relationship with various environmental factors such as humidity, light, and temperature. However, most current research focuses on using sensor network technology to monitor changes in sap flow in real time, which makes it difficult to capture the impact of environmental factors on sap flow changes. Therefore, considering the joint impact of multiple environmental factors on sap flow is the key to achieving accurate prediction of sap flow.

[0003] In recent years, there have been many studies on prediction models, all of which have different degrees of deficiencies. For example, the prediction method based on RNN has long-term dependency problems and takes a long time to train; the prediction method based on LSTM alleviates the gradient vanishing and gradient exploding problems of RNN, but is mainly good at capturing local time dependencies; the prediction method based on TCN relies on a large number of parameter tuning, and the model performance will be significantly reduced in complex environmental data; the prediction based on Transformer performs well in capturing complex dependencies, but the computational complexity is high. All of the above problems will lead to poor accuracy of the prediction results. Summary of the invention

[0004] In order to address the shortcomings of the above-mentioned technologies, the present invention proposes a plant sap flow prediction method based on multidimensional environmental sequences and graph convolutional networks, in order to make full use of the hidden temporal characteristics of environmental factors and the dynamic influence of environmental factors on plant sap flow changes, thereby improving the prediction accuracy of plant sap flow to meet the needs of practical applications.

[0005] In order to achieve the following objectives, the present invention provides the following solutions:

[0006] The plant sap flow prediction method based on multidimensional environmental sequence and graph convolutional network of the present invention is characterized in that it is carried out according to the following steps:

[0007] Step 1: synchronously collect the environmental data and liquid flow data of the plant within a certain period of time and perform maximum and minimum normalization processing to obtain a normalized data sequence X;

[0008] Step 2: Set the sliding step size to s and the sliding window size to , thereby performing sliding window processing on X and obtaining the tth sample in the sliding window from time t to time t+s , and each sample includes environmental data and liquid flow data, where n represents the number of features in the sample;

[0009] make The true value of the liquid flow in is recorded as ;

[0010] Step 3: Construct a plant sap flow prediction network, including: Fourier denoising module, feature distillation module, adaptive adjacency matrix module, GCN module, multi-head attention mechanism module and prediction head, and After processing, the predicted liquid flow value is recorded as ;

[0011] Step 3.1: The Fourier denoising module uses equations (1) to (8) to Processing is performed to obtain the tth time domain denoising feature :

[0012] (1)

[0013] (2)

[0014] (3)

[0015] (4)

[0016] (5)

[0017] (6)

[0018] (7)

[0019] (8)

[0020] In formula (1) to formula (8), represents the Fourier transform function; represents the inverse Fourier transform function, represents term-by-term multiplication, express The frequency domain feature representation is: express The power spectral density of represents the threshold value, represents the power spectral density after denoising, express The frequency domain representation after denoising is: represents the global filter to be learned, represents the local filter to be learned, represents the frequency characteristics on the global scale, represents the frequency characteristics at the local scale, Represents multi-scale frequency feature representation;

[0021] Step 3.2: The feature distillation module uses formula (9) to Perform convolution calculation and get Local features ;

[0022] (9)

[0023] In formula (9), is a one-dimensional convolution, is the activation function, is the maximum pooling operation;

[0024] Step 3.3: The adaptive graph learning module processes X using formula (10) to obtain an adaptive adjacency matrix ;

[0025] (10)

[0026] In formula (10), is the activation function; is the activation function; T represents transposition;

[0027] Step 3.4: The GCN module uses formula (11) to and Perform GCN convolution calculation to get S t Aggregation features ;

[0028] (11)

[0029] In formula (11), is the linear transformation parameter, is the degree matrix;

[0030] Step 3.5: The multi-head attention mechanism module uses equations (12) to (17) to Process and obtain S t The predicted probability of liquid flow in the sliding window from time t to time t+s :

[0031] (12)

[0032] (13)

[0033] (14)

[0034] (15)

[0035] (16)

[0036] (17)

[0037] In formula (12) to formula (17), h, h is the number of attention heads; Indicates S t Aggregate features at the i-th attention head; Indicates S t The query feature vector at the i-th attention head; Indicates S t The key feature vector in the i-th attention head, Indicates S t The value vector at the i-th attention head; represents the query vector projection parameter matrix, represents the key vector projection parameter matrix, Represents the value vector projection parameter matrix; for The dimension of Indicates S t The correlation between the environmental factors and plant sap flow at the i-th attention head is, YesS t The correlation output matrix of the i-th attention head is, is the weight matrix; Indicates splicing;

[0038] Step 3.6: The prediction head uses a fully connected layer to Process and obtain S t Predicted value of plant sap flow in the sliding window from time t to time t+s ;

[0039] Step 4: Process all samples in the sliding window according to the process of step 3, and use formula (18) to construct the mean square error of the plant sap flow prediction network: ;

[0040] (18)

[0041] In formula (18), (nws) / s is the number of samples;

[0042] Step 5: Use the gradient descent method to train the plant sap flow prediction network and calculate the mean square error. To update the network parameters until The training is stopped when it tends to be stable, thereby obtaining the optimal plant sap flow prediction model for predicting plant sap flow.

[0043] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the plant sap flow prediction method, and the processor is configured to execute the program stored in the memory.

[0044] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the plant sap flow prediction method when the computer program is executed by a processor.

[0045] Compared with the prior art, the advantages of the present invention are:

[0046] 1. The present invention can focus on the data itself, identify and eliminate noise data, reduce interference while avoiding the loss of effective data features, and provide a high-quality data source for liquid flow prediction.

[0047] 2. The present invention can extract key features of data, thereby removing redundant information. This enables the network to effectively extract local features of environmental data when facing a local complex environment, thereby improving the accuracy of liquid flow prediction.

[0048] 3. The present invention can identify and strengthen attention to key factors affecting liquid flow. By allocating different attention weights, it can to a certain extent solve the problems of large data fluctuations and redundancy in the plant growth environment, as well as the complex coupling relationship between liquid flow and environmental factors, which lead to large deviations in prediction results, thereby providing more reliable support for the regulation of liquid flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a modeling flow chart of the present invention;

[0050] Figure 2 It is the prediction flow chart of the present invention. DETAILED DESCRIPTION

[0051] In this embodiment, a modeling process of a plant sap flow prediction method based on multidimensional environmental sequence and graph convolutional network is as follows: Figure 1 As shown, for the sample Processing is performed to obtain the predicted value of plant sap flow The method includes a Fourier denoising module, a feature distillation module, an adaptive adjacency matrix module, a GCN module, a multi-head attention mechanism module and a prediction head. First, the environmental data and liquid flow data of the plant are collected synchronously, and the maximum and minimum normalization processing is performed. Then the data is input into the prediction model, and the sliding window strategy is used to predict the liquid flow value in the sliding window from time t to time t+s. When the training converges, the optimal liquid flow prediction model is obtained. The prediction process of a plant liquid flow prediction method based on multidimensional environmental sequence and graph convolutional network is as follows: Figure 2 The plant is illustrated by taking bamboo as an example, and the method comprises:

[0052] Step 1: Synchronously collect the environmental data (including air humidity, air temperature, soil temperature, soil humidity, light, CO2 concentration, etc.) and sap flow data of plants in a certain period of time, and then perform maximum and minimum normalization processing to obtain the normalized bamboo forest data sequence X. Among them, sap flow is used as the model prediction label.

[0053] Step 2: Set the sliding step size to s and the sliding window size to , so as to perform sliding window processing on X and obtain the t-th bamboo forest data sample in the sliding window from time t to time t+s , each sample includes environmental data and sap flow data, where n represents the number of features in the bamboo forest data sample;

[0054] make The true value of the liquid flow in is recorded as .

[0055] Step 3: Since the data of sensor nodes are easily disturbed by the external environment, the collected data will fluctuate greatly and be easily lost. In order to ensure the integrity and reliability of sensor data, the Fourier denoising module uses equations (1) to (8) to Processing is performed to obtain the tth time domain denoising feature :

[0056] (1)

[0057] (2)

[0058] (3)

[0059] (4)

[0060] (5)

[0061] (6)

[0062] (7)

[0063] (8)

[0064] In formula (1) to formula (8), represents the Fourier transform function; represents the inverse Fourier transform function, represents term-by-term multiplication, express The frequency domain feature representation of express The power spectral density of represents the threshold value, represents the power spectral density after denoising, express The frequency domain representation after denoising is: represents a learnable global filter, represents a learnable local filter, represents the frequency characteristics of bamboo forest data at the global scale, represents the frequency characteristics of bamboo forest data at a local scale, Multi-scale frequency feature representation of bamboo forest data.

[0065] Step 4: The feature distillation module uses formula (9) to Perform convolution calculation and get Local features , :

[0066] (9)

[0067] In formula (9), is a one-dimensional convolution, is the activation function, is the maximum pooling operation, yes The number of convolution kernels, yes The convolution kernel size is yes The step length.

[0068] Step 5: The adaptive graph learning module processes X using formula (10) to obtain the adaptive adjacency matrix ;

[0069] (10)

[0070] In formula (10), is the activation function; is the activation function; T denotes transpose.

[0071] Step 6: The GCN module uses formula (11) to and Perform GCN convolution calculation to get S t Aggregation features ;

[0072] (11)

[0073] In formula (11), is the linear transformation parameter, is the degree matrix.

[0074] Step 7: The coupling relationship between liquid flow and other environmental factors is complex. However, not all environmental factors have the same importance to them. It is necessary to learn different attention weights between each environmental factor and liquid flow. The multi-head attention mechanism module uses equations (12) to (17) to Process and obtain S t The predicted probability of liquid flow in the sliding window from time t to time t+s :

[0075] (12)

[0076] (13)

[0077] (14)

[0078] (15)

[0079] (16)

[0080] (17)

[0081] In formula (12) to formula (17), h, h is the number of attention heads; Indicates S t Aggregate features at the i-th attention head; Indicates S t The query feature vector at the i-th attention head; Indicates S t The key feature vector in the i-th attention head, Indicates S t The value vector at the i-th attention head; represents the query vector projection parameter matrix, represents the key vector projection parameter matrix, Represents the value vector projection parameter matrix; for The dimension of Indicates S t The correlation between the environmental factors and the liquid flow at the i-th attention head is, YesS t The correlation output matrix of the i-th attention head is, is the weight matrix; Indicates splicing.

[0082] Step 8: The prediction head uses the fully connected layer to Process and obtain S t Liquid flow prediction value in the sliding window from time t to time t+s ;

[0083] Step 9: Process all the bamboo forest samples in the sliding window according to the process of steps 3 to 8, and then use formula (18) to construct the mean square error of the bamboo sap flow prediction network: ;

[0084] (18)

[0085] In formula (18), (nws) / s is the number of bamboo forest samples;

[0086] Step 10: Use the gradient descent method to train the bamboo sap flow prediction network and calculate the mean square error To update the network parameters until The training is stopped when it tends to be stable, thereby obtaining the optimal bamboo sap flow prediction model, which is used to predict the bamboo sap flow.

[0087] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0088] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

Claims

1. A plant sap flow prediction method based on multidimensional environmental sequence and graph convolutional network, characterized in that: The steps are as follows: Step 1: synchronously collect the environmental data and liquid flow data of the plant within a certain period of time and perform maximum and minimum normalization processing to obtain a normalized data sequence X; Step 2: Set the sliding step size to s and the sliding window size to , thereby performing sliding window processing on X and obtaining the tth sample in the sliding window from time t to time t+s , and each sample includes environmental data and liquid flow data, where n represents the number of features in the sample; make The true value of the liquid flow in is recorded as ; Step 3: Construct a plant sap flow prediction network, including: Fourier denoising module, feature distillation module, adaptive adjacency matrix module, GCN module, multi-head attention mechanism module and prediction head, and After processing, the predicted liquid flow value is recorded as ; Step 4: Process all samples in the sliding window according to the process of step 3, and use formula (18) to construct the mean square error of the plant sap flow prediction network: ; (18) In formula (18), (nws) / s is the number of samples; Step 5: Use the gradient descent method to train the plant sap flow prediction network and calculate the mean square error. To update the network parameters until The training is stopped when it tends to be stable, thereby obtaining the optimal plant sap flow prediction model for predicting plant sap flow.

2. The method for predicting plant sap flow based on multidimensional environmental sequence and graph convolutional network according to claim 1, characterized in that: The step 3 is carried out as follows: Step 3.1: The Fourier denoising module uses equations (1) to (8) to Processing is performed to obtain the tth time domain denoising feature : (1) (2) (3) (4) (5) (6) (7) (8) In formula (1) to formula (8), represents the Fourier transform function; represents the inverse Fourier transform function, represents term-by-term multiplication, express The frequency domain feature representation is: express The power spectral density of represents the threshold value, represents the power spectral density after denoising, express The frequency domain representation after denoising is: represents the global filter to be learned, represents the local filter to be learned, represents the frequency characteristics on the global scale, represents the frequency characteristics at the local scale, Represents multi-scale frequency feature representation; Step 3.2: The feature distillation module uses formula (9) to Perform convolution calculation and get Local features ; (9) In formula (9), is a one-dimensional convolution, is the activation function, is the maximum pooling operation; Step 3.3: The adaptive graph learning module processes X using formula (10) to obtain an adaptive adjacency matrix ; (10) In formula (10), is the activation function; is the activation function; T represents transposition; Step 3.4: The GCN module uses formula (11) to and Perform GCN convolution calculation to get S t Aggregation features ; (11) In formula (11), is the linear transformation parameter, is the degree matrix; Step 3.5: The multi-head attention mechanism module uses equations (12) to (17) to Process and obtain S t The predicted probability of liquid flow in the sliding window from time t to time t+s : (12) (13) (14) (15) (16) (17) In formula (12) to formula (17), h, h is the number of attention heads; Indicates S t Aggregate features at the i-th attention head; Indicates S t The query feature vector at the i-th attention head; Indicates S t The key feature vector in the i-th attention head, Indicates S t The value vector at the i-th attention head; represents the query vector projection parameter matrix, represents the key vector projection parameter matrix, Represents the value vector projection parameter matrix; for The dimension of Indicates S t The correlation between the environmental factors and plant sap flow at the ith attention head is, YesS t The correlation output matrix of the i-th attention head is, is the weight matrix; Indicates splicing; Step 3.6: The prediction head uses a fully connected layer to Process and obtain S t Predicted value of plant sap flow in the sliding window from time t to time t+s .

3. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the plant sap flow prediction method according to claim 1 or 2, and the processor is configured to execute the program stored in the memory.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the plant liquid flow prediction method according to claim 1 or 2 are executed.

Images