ET0 prediction method based on signal decomposition and improved deep learning model

Abstract

The invention discloses an ET0 prediction method based on signal decomposition and an improved deep learning model. The ET0 prediction method comprises the following steps: step 1, collecting data; 2, missing value processing, abnormal value detection, normalization operation and interpolation are conducted on the collected data, the data obtained after interpolation are linearly mapped to be between 0 and 1 through Min-Max standardization, and normalized data are obtained; 3, calculating ET0 according to the P-M model; 4, signal decomposition is conducted on the ET0 obtained through calculation through a CEEMDAN algorithm, the signal is decomposed into a plurality of IMFs, and the IMFs are different frequency components in the signal and used for extracting data features under different time scales; step 5, combining all IMFs obtained by decomposition of the CEEMDAN algorithm with other meteorological feature vectors to form a complete input feature, and constructing a sample for the input feature by using a sliding window technology; the problem that ET0 cannot be accurately predicted under the background of global climate change and increasingly tense water resources is solved.

Description

Technical Field

[0001] The present invention belongs to the field of agricultural water management, and specifically relates to an ETO prediction method based on signal decomposition and an improved deep learning model. Background Art

[0002] With the intensification of global climate change, agricultural production is facing increasingly complex environmental challenges, especially the scarcity of water resources, which poses a serious threat to the sustainable development of agriculture. Water is one of the important factors in agricultural production, and crop evapotranspiration is a key indicator of water management. It is not only closely related to meteorological conditions such as temperature, humidity, and precipitation, but also affected by multiple factors such as soil type, crop type, and planting method. How to accurately predict crop evapotranspiration and rationally allocate water resources has become an important issue in modern agricultural production. However, the existing reference crop evapotranspiration (ET0) prediction method often has certain limitations when dealing with complex climate and environmental changes, and new technical means are urgently needed to improve it.

[0003] Traditional ET0 prediction methods are mostly based on physical models and empirical formulas. The Penman-Monteith formula, proposed by the Food and Agriculture Organization of the United Nations (FAO), is a representative example. It uses meteorological factors (such as temperature, humidity, and wind speed) and physical quantities such as radiation to derive ET0. These methods have a sound theoretical foundation and can reflect the direct impact of climate on ET0. However, because they are often based on fixed assumptions, they fail to account for complex climate fluctuations and nonlinear environmental changes, limiting the reliability of the prediction results. Furthermore, physical models typically require large amounts of accurate meteorological data, which makes their implementation difficult in regions with limited meteorological observations. Furthermore, these traditional methods are unable to effectively process large-scale, long-term, multidimensional data, resulting in insufficient accuracy and adaptability in addressing the ever-changing environment of agricultural production.

[0004] With the continuous advancement of machine learning technology, particularly the rise of deep learning methods, a growing number of researchers are exploring data-driven ET0 prediction methods. Unlike traditional methods, deep learning methods automatically learn features from large amounts of historical data, enabling them to better capture nonlinearities and complex spatiotemporal dependencies, resulting in greater adaptability and predictive power. For example, the Long Short-Term Memory (LSTM) network effectively handles long-term dependencies in time series data, addressing the limitations of traditional machine learning methods in handling time series problems. Through its unique gating structure, the LSTM can retain long-term information and effectively forget useless information, resulting in high prediction accuracy for time series problems such as ET0 prediction. Furthermore, the Transformer offers the advantage of global modeling in time series modeling, simultaneously focusing on all moments in the sequence. This addresses the LSTM's limitations in capturing long-term dependencies.

[0005] While LSTM and Transformer each have unique advantages, single models often have limitations when handling complex, nonlinear ETO prediction problems. While LSTM can capture long-term dependencies, it often faces training difficulties and slow convergence when dealing with highly complex time series data. While Transformer offers the advantage of global modeling, its complex model structure makes training demanding on both data volume and computing resources. To overcome these issues, we developed an ETO prediction method based on CEEMDAN and the Transformer-BiLSTM model. CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise), an advanced signal decomposition method, decomposes the original time series signal into multiple intrinsic mode functions (IMFs), each representing a different frequency component in the signal. This decomposition method allows for the extraction of more fine-grained time series features, helping to capture underlying patterns in ETO data. Combining CEEMDAN with signal decomposition allows Transformer and BiLSTM models to model from the perspectives of global and local temporal dependencies, respectively, thereby addressing the shortcomings of single models and improving prediction accuracy and reliability.

[0006] By combining CEEMDAN with deep learning models, we can more effectively handle complex climate data and environmental factors, providing more accurate ET0 predictions. Compared to traditional physical models and statistical regression methods, deep learning models are better able to handle nonlinear and time-varying data characteristics, resulting in higher prediction accuracy and greater adaptability. Especially in the context of global climate change and increasingly scarce water resources, accurate ET0 predictions will facilitate more rational planning of agricultural irrigation and water resource management, providing strong support for sustainable agricultural development. Summary of the Invention

[0007] To solve the above technical problems, the present invention proposes an ET0 prediction method based on signal decomposition and improved deep learning model, which solves the problem of inaccurate ET0 prediction of complex climate data in the context of global climate change and increasingly tight water resources.

[0008] The technical solutions of the present invention are as follows:

[0009] An ET0 prediction method based on signal decomposition and improved deep learning model includes the following steps:

[0010] Step 1: Collect data;

[0011] Step 2: Perform missing value processing, outlier detection, normalization, interpolation, and Min-Max standardization on the collected data to linearly map the interpolated data to between 0 and 1 to obtain normalized data;

[0012] Step 3: Calculate ET0 based on the PM model;

[0013] Step 4: Use the CEEMDAN algorithm to decompose the calculated ET0 into multiple IMFs. IMFs are different frequency components in the signal, which are used to extract data features at different time scales.

[0014] Step 5: Combine all IMFs decomposed by the CEEMDAN algorithm with other meteorological feature vectors to form a complete input feature. Use the sliding window technique to construct a sample of the input feature. The sample includes the comprehensive feature vectors of the past W moments.

[0015] Among them, the comprehensive feature vector at time t is:

[0016] X t =F1(t),F2(t),,F M (t),IMF1(t),IMF2(t),,IMF N (t)

[0017] The sample is:

[0018] X=X t-W+1 ,X t-W+2 ,...,X t

[0019] Among them, X is the input feature matrix, W is the number of comprehensive feature vectors at each moment;

[0020] Step 6: Using ET0 as the target label, combine the input features of each sample with the corresponding ET0 to form a complete training sample. The training sample is divided into a training set, a validation set, and a test set according to the ratio of 70%, 20%, and 10% respectively.

[0021] Step 7: Use the parallel computing framework Pytorch to build the Transformer-BiLSTM model;

[0022] Step 8: Train and verify the constructed Transformer-BiLSTM model;

[0023] Step 9: Export the trained Transformer-BiLSTM model to the ONNX format, optimize and accelerate the ONNX format using TensorRT, and perform real-time prediction on a hardware platform that supports GPU acceleration.

[0024] Preferably, the Min-Max normalization calculation formula in step 2 is:

[0025]

[0026] Among them, min(X) and max(X) are the minimum and maximum values of the feature respectively, X is the original data, and X norm The data are normalized.

[0027] Preferably, the PM model calculation formula in step 3 is:

[0028]

[0029] Where ET0 is the reference crop evapotranspiration, Δ is the slope of the saturated vapor pressure curve, Rn is the net surface radiation, G is the soil heat flux, γ is the dry-wet constant, and T mean is the daily average temperature, es and ea are the saturated vapor pressure value and the actual vapor pressure value, respectively.

[0030] Preferably, the CEEMDAN algorithm in step 4 includes the following sub-steps:

[0031] Step 4-1: Add white noise w with a signal-to-noise ratio of ε0 to the ET0 sequence x(t) i (t)(i=1,2,...,n), and use the following formula to obtain the data x for the first decomposition i (t)(i=1,2,...,n):

[0032] x i (t)=x(t)+ε0w i (t)

[0033] Where i is the i-th white noise added, t is a different time point, and n is the total number of times white noise is added;

[0034] Step 4-2: Use the CEEMDAN algorithm to decompose the ET0 series into a finite number of IMFs and residuals. Since the mean of white noise is zero, its influence can be eliminated by calculating the mean.

[0035] Step 4-3: Add residual r k-1 Add adaptive noise E k (w (i) ), the calculation formula is:

[0036]

[0037] Where, β k-1 is the weight coefficient of k-1 stage Gaussian white noise; is the residual of the i-th white noise in the k-1 stage;

[0038] Step 4-4: Obtain r according to the calculation formula k-1 The local mean M(r k-1 +β K-1 E k (w (i) )) and the average value, which is calculated as follows:

[0039]

[0040] Where I is the total number of EMD decompositions, r k is the residual sequence;

[0041] Step 4-5: Obtain the kth IMF according to the calculation formula The calculation formula is:

[0042]

[0043] Where, is the kth IMF component;

[0044] Step 4-6: When the residual has fewer than three extreme points and cannot be decomposed further, terminate the CEEMDAN algorithm.

[0045] Preferably, step 7 includes the following sub-steps:

[0046] Step 7-1: Input embedding to the training set, validation set, and test set to obtain the input embedding sequence;

[0047] Step 7-2: Perform position encoding on the input embedding sequence to obtain the position encoding, which is calculated as follows:

[0048]

[0049] Among them, PE is the position code; pos is the position of the element in the vector; d is the dimension of PE; 2i is the even dimension; 2i+1 is the odd dimension;

[0050] Step 7-3: Perform multi-head attention on the position encoding to obtain the query matrix Q, key matrix K and value matrix V, which are calculated as follows:

[0051]

[0052] Where X is the input matrix; W Q , W K , W V is the matrix to be trained into the model, d K is the dimension of vector V; Q, K and V are the selected feature data of ET0;

[0053] Step 7-4: Add and normalize the input X;

[0054] Step 7-5: Feed forward network for different position vectors;

[0055] Step 7-6: Multi-head attention, feedforward network, addition and normalization can be combined into an encoder structure;

[0056] Step 7-7: Decode the encoder.

[0057] Preferably, the BiLSTM model includes forward propagation and backpropagation;

[0058] The forward propagation formula is:

[0059] f t =σ(W f h t-1 +W f x t +b f )

[0060] i t =σ(W i h t-1 +W i x t +b i )

[0061] c t =f t c t-1 +i t tanh(W c h t-1 +W c x t +b c)

[0062] o t =σ(W o h t-1 +W o x t +b o )

[0063] h t =o t tanh(c t )

[0064] Among them, x t is the input signal at time t; h t is the hidden state; c t is the memory signal; f t It is the forget gate; t is the input gate; o t is the output gate; W i , W f , W o and W c are the weight matrices of the input gate, forget gate, output gate, and memory unit of the forward LSTM; b i , b f , b o and b c . are the biases of each gate respectively; tanh represents the hyperbolic tangent activation function, and σ represents the Sigmoid function;

[0065] The calculation formula for back propagation is:

[0066]

[0067] Among them, x t is the input signal at the moment; is a hidden state; It is a memory signal; It is the Gate of Oblivion; is the input gate; is the output gate; and They are the weight matrices of the input gate, forget gate, output gate, and memory unit of the forward LSTM; and are the biases for each gate respectively.

[0068] Preferably, step S8 is specifically:

[0069] An evaluation system was constructed using four indicators to assess the prediction accuracy of different models;

[0070] The four indicators are: mean absolute error MAE, mean square error MSE, root mean square error RMSE and determination coefficient R2,

[0071] The calculation formula is:

[0072]

[0073] Among them, y i is the standard value of ET0, x i is the predicted value of ET0, is the mean of the predicted values, is the mean of the standardized values, and n is the number of actual observations or predicted values.

[0074] The beneficial effects of the ETO prediction method based on signal decomposition and improved deep learning model of the present invention are as follows:

[0075] 1. The present invention uses CEEMDAN signal decomposition technology to effectively extract multi-level time series features from crop evapotranspiration data, reduce noise interference, and thus improve the accuracy of the model.

[0076] 2. This paper combines the Transformer and BiLSTM models, giving full play to the advantages of both in capturing global and local temporal dependencies, and can better process complex and nonlinear meteorological and environmental data, thereby improving prediction accuracy.

[0077] 3. The present invention uses TensorRT to optimize and accelerate the model, significantly improving the reasoning speed and computing efficiency, enabling this method to achieve fast and real-time ETO prediction in large-scale agricultural production environments.

[0078] 4. The present invention not only has high accuracy and strong stability, but also has good computing performance and adaptability, and is suitable for ETO prediction needs under various climatic and environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly explain the purpose, design ideas and innovation of the ET0 prediction method based on signal decomposition and improved deep learning model proposed in the present invention, the present invention will be described in detail with reference to the accompanying drawings and appendixes.

[0080] Figure 1 It is a technical flow chart of the present invention.

[0081] Figure 2 This is a diagram showing the ET0 prediction effects of different algorithms of the present invention at sites P1 to P4.

[0082] Figure 3 This is a diagram of the ET0 prediction error of different algorithms of the present invention at sites P1 to P4. DETAILED DESCRIPTION

[0083] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0084] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0085] like Figure 1 As shown, an ET0 prediction method based on signal decomposition and improved deep learning model includes the following steps:

[0086] Step 1: Collect data;

[0087] Step 2: Perform missing value processing, outlier detection, normalization, interpolation, and Min-Max standardization on the collected data to linearly map the interpolated data to between 0 and 1 to obtain normalized data;

[0088] Step 3: Calculate ET0 based on the PM model;

[0089] Step 4: Use the CEEMDAN algorithm to decompose the calculated ET0 into multiple IMFs. IMFs are different frequency components in the signal, which are used to extract data features at different time scales.

[0090] Step 5: Combine all IMFs decomposed by the CEEMDAN algorithm with other meteorological feature vectors to form a complete input feature. Use the sliding window technique to construct a sample of the input feature. The sample includes the comprehensive feature vectors of the past W moments.

[0091] Among them, the comprehensive feature vector at time t is:

[0092] X t =F1(t),F2(t),,F M (t),IMF1(t),IMF2(t),,IMF N (t)

[0093] The sample is:

[0094] X=X t-W+1 ,X t-W+2 ,...,X t

[0095] Among them, X is the input feature matrix, W is the number of comprehensive feature vectors at each moment;

[0096] Step 6: Using ET0 as the target label, combine the input features of each sample with the corresponding ET0 to form a complete training sample. The training sample is divided into a training set, a validation set, and a test set according to the ratio of 70%, 20%, and 10% respectively.

[0097] Step 7: Use the parallel computing framework Pytorch to build the Transformer-BiLSTM model;

[0098] Step 8: Train and verify the constructed Transformer-BiLSTM model;

[0099] Step 9: Export the trained Transformer-BiLSTM model to the ONNX format, optimize and accelerate the ONNX format using TensorRT, and perform real-time prediction on a hardware platform that supports GPU acceleration.

[0100] The Min-Max normalized calculation formula in step 2 of this embodiment is:

[0101]

[0102] Among them, min(X) and max(X) are the minimum and maximum values of the feature respectively, X is the original data, and X norm The data are normalized.

[0103] The PM model calculation formula in step 3 of this embodiment is:

[0104]

[0105] Where ET0 is the reference crop evapotranspiration, Δ is the slope of the saturated vapor pressure curve, Rn is the net surface radiation, G is the soil heat flux, γ is the dry-wet constant, and T mean is the daily average temperature, es and ea are the saturated vapor pressure value and the actual vapor pressure value, respectively.

[0106] The CEEMDAN algorithm in step 4 of this embodiment includes the following sub-steps:

[0107] Step 4-1: Add white noise w with a signal-to-noise ratio of ε0 to the ET0 sequence x(t) i (t)(i=1,2,...,n), and use the following formula to obtain the data x for the first decomposition i (t)(i=1,2,...,n):

[0108] x i (t)=x(t)+ε0w i (t)

[0109] Where i is the i-th white noise added, t is a different time point, and n is the total number of times white noise is added;

[0110] Step 4-2: Use the CEEMDAN algorithm to decompose the ET0 series into a finite number of IMFs and residuals. Since the mean of white noise is zero, its influence can be eliminated by calculating the mean.

[0111] Step 4-3: Add residual r k-1 Add adaptive noise E k (w (i) ), the calculation formula is:

[0112]

[0113] Where, β k-1 is the weight coefficient of k-1 stage Gaussian white noise; is the residual of the i-th white noise in the k-1 stage;

[0114] Step 4-4: Obtain r according to the calculation formula k-1 The local mean M(r k-1 +β k-1 E k (w (i) )) and the average value, which is calculated as follows:

[0115]

[0116] Where I is the total number of EMD decompositions, r k is the residual sequence;

[0117] Step 4-5: Obtain the kth IMF according to the calculation formula The calculation formula is:

[0118]

[0119] Where, is the kth IMF component;

[0120] Step 4-6: When the residual has fewer than three extreme points and cannot be decomposed further, terminate the CEEMDAN algorithm.

[0121] Step 7 of this embodiment includes the following sub-steps:

[0122] Step 7-1: Input embedding to the training set, validation set, and test set to obtain the input embedding sequence;

[0123] Step 7-2: Perform position encoding on the input embedding sequence to obtain the position encoding, which is calculated as follows:

[0124]

[0125] Among them, PE is the position code; pos is the position of the element in the vector; d is the dimension of PE; 2i is the even dimension; 2i+1 is the odd dimension;

[0126] Step 7-3: Perform multi-head attention on the position encoding to obtain the query matrix Q, key matrix K and value matrix V, which are calculated as follows:

[0127]

[0128] Where X is the input matrix; W Q , W K , W V is the matrix to be trained into the model, d K is the dimension of vector V; Q, K and V are the selected feature data of ET0;

[0129] Step 7-4: Add and normalize the input X;

[0130] Step 7-5: Feed forward network for different position vectors;

[0131] Step 7-6: Multi-head attention, feedforward network, addition and normalization can be combined into an encoder structure;

[0132] Step 7-7: Decode the encoder.

[0133] The BiLSTM model of this embodiment includes forward propagation and backpropagation;

[0134] The forward propagation formula is:

[0135] f t =σ(W f h t-1 +W f x t +b f )

[0136] i t =σ(W i h t-1 +W i x t +b i )

[0137] c t =f t c t-1 +i t tanh(W c h t-1 +W c x t+b c )

[0138] o t =σ(W o h t-1 +W o x t +b o )

[0139] h t =o t tanh(c t )

[0140] Among them, x t is the input signal at time t; h t is the hidden state; c t is the memory signal; f t It is the forget gate; t is the input gate; o t is the output gate; W i , W f , W o and W c are the weight matrices of the input gate, forget gate, output gate, and memory unit of the forward LSTM; b i , b f , b o and b c . are the biases of each gate respectively; tanh represents the hyperbolic tangent activation function, and σ represents the Sigmoid function;

[0141] The calculation formula for back propagation is:

[0142]

[0143] Among them, x t is the input signal at the moment; is a hidden state; It is a memory signal; It is the Gate of Oblivion; is the input gate; is the output gate; and They are the weight matrices of the input gate, forget gate, output gate, and memory unit of the forward LSTM; and are the biases for each gate respectively.

[0144] Step S8 of this embodiment is specifically as follows:

[0145] An evaluation system was constructed using four indicators to assess the prediction accuracy of different models;

[0146] The four indicators are: mean absolute error MAE, mean square error MSE, root mean square error RMSE and determination coefficient R2,

[0147] The calculation formula is:

[0148]

[0149] Among them, y i is the standard value of ET0, x i is the predicted value of ET0, is the mean of the predicted values, is the mean of the standardized values, and n is the number of actual observations or predicted values.

[0150] When this implementation plan is implemented,

[0151] S1: Data collection: Collect daily and monthly historical meteorological data from 1940 to 2024 at the meteorological station, including maximum temperature, average temperature, minimum temperature, sunshine hours, shortwave radiation, average relative humidity at a height of 2 m, average dew point temperature at a height of 2 m, and wind speed at a height of 10 m.

[0152] S2: Data preprocessing: Perform missing value processing, outlier detection, and normalization on the collected data. Use interpolation or filling methods (such as linear interpolation or spline interpolation). Finally, use Min-Max normalization to linearly map the data to between 0 and 1. The formula is as follows:

[0153]

[0154] Among them, min(X) and max(X) are the minimum and maximum values of the feature respectively, X is the original data, and X norm The data are normalized.

[0155] S3: Calculation of reference crop evapotranspiration: Use ET0 calculated by the PM model as the standard value. The expression of the PM model is as follows:

[0156]

[0157] where ET0 is the reference crop evapotranspiration (mm / d), Δ is the slope of the saturated vapor pressure curve (kPa / °C), Rn is the net surface radiation (MJ / m2·d), G is the soil heat flux (MJ / m2·d), γ is the dry-wet constant (kPa / °C), Tmean is the daily mean temperature (°C), es and ea are the saturated vapor pressure value and the actual vapor pressure value, respectively (kPa).

[0158] S4: Use CEEMDAN to perform signal decomposition on the data: Use CEEMDAN to perform signal decomposition on the calculated standard ET0, and decompose the original time series signal into multiple intrinsic mode functions (IMFs). Each IMF represents a different frequency component in the signal, and can extract data features at different time scales.

[0159] The CEEMDAN method can be described as follows:

[0160] (1) Add white noise w with a signal-to-noise ratio of ε0 to the original ET0 sequence x(t) i (t)(i=1,2,...,n), and use the following formula to obtain the data x for the first decomposition i (t)(i=1,2,...,n):

[0161] x i (t)=x(t)+ε0w i (t)

[0162] Where i represents the i-th white noise added, t represents different time points, and n represents the total number of times white noise is added.

[0163] (2) Use CEEMDAN to decompose the ET0 series into a finite number of IMFs and residuals. Since the mean of white noise is zero, its influence can be eliminated by calculating the mean.

[0164] (3) Add the residual r as follows k-1 Add adaptive noise E k (w (i) ):

[0165]

[0166] (4) Obtain r as follows k-1 The local mean M(r k-1 +β k-1 E k (w (i) )) and the average value:

[0167]

[0168] (5) Obtain the kth IMF as follows

[0169]

[0170] (6) Finally, when the residual has less than three extreme points and cannot be decomposed further, the CEEMDAN algorithm is terminated.

[0171] S5: Create a dataset: All IMFs decomposed from CEEMDAN are combined with other meteorological feature vectors to form a complete input feature. Then, a sliding window technique is used to construct a dataset, where each sample consists of the comprehensive feature vectors of the past W moments. The comprehensive feature vector at each moment t is:

[0172] X t =F1(t),F2(t),,F M (t),IMF1(t),IMF2(t),,IMF N (t)

[0173] The sample can be expressed as:

[0174] X=X t-W+1 ,X t-W+2 ,...,X t

[0175] Among them, X is the input feature matrix, which contains the comprehensive feature vectors of the past W moments.

[0176] S6: Using ET0 as the target label, combine the features of each sample with the corresponding ET0 to form a complete training sample. Finally, divide it into training set, validation set, and test set according to 70%, 20%, and 10% respectively for model training and evaluation.

[0177] S7: Transformer and BiLSTM Modeling: Transformer is a deep learning model based on the attention mechanism, mainly composed of an encoder and a decoder. The main steps are as follows:

[0178] (1) Input Embedding

[0179] Input embedding is the process of converting each element of the input sequence into a fixed-size vector by finding a pre-trained embedding matrix that maps each unique element to a point in a high-dimensional space.

[0180] (2) Positional encoding

[0181] Positional encoding provides the model with information about the position of the token in the input sequence, and its calculation formula is as follows:

[0182]

[0183] Among them, PE is the position encoding; pos is the position of the element in the vector; d is the dimension of PE; 2i is the even dimension; 2i+1 is the odd dimension.

[0184] (3) Multi-head attention

[0185] Multi-head attention consists of multiple self-attentions, which enables the model to weight information at different positions when processing the input sequence, rather than relying solely on the order of the sequence. Self-attention achieves interaction between meteorological features by calculating attention weights. The input of self-attention is represented by the matrix X, which represents the input sequence. The output of self-attention can be determined by calculating the query matrix Q, the key matrix K, and the value matrix V respectively. The formula is as follows:

[0186]

[0187]

[0188] Where X is the input matrix; W Q , W K , W V is the matrix to be trained into the model, d K is the dimension of vector V; Q, K, and V are the selected feature data of ET0. The input X is passed to different self-attentions respectively, and the output matrix is calculated. The multi-head attention concatenates the output matrices together and passes them into a linear layer to obtain the final output of the multi-head attention.

[0189] (4) Addition and normalization

[0190] Addition and normalization refer to residual connections and layer normalization. The purpose of residual connections is to mitigate the vanishing gradient problem in deep neural networks by preserving information about the original input X, which makes gradients more easily propagated. Layer normalization normalizes the summation result, aiming to make the inputs of each layer have similar statistical properties and increase stability when training deep neural networks.

[0191] (5) Feedforward Network

[0192] Feedforward networks typically consist of two layers of linear transformations, connected by an activation function such as ReLU. The first layer expands the input to a higher-dimensional space to increase the model's expressive power. The activation function introduces nonlinearity to help the model capture complex features. The second layer compresses the high-dimensional representation back to its original dimensions. This process is performed independently for each position vector, allowing the feedforward network to have the same effect across different positions in the sequence, enhancing the model's generalization capabilities.

[0193] (6) Encoder

[0194] Multi-head attention, feedforward networks, addition, and normalization can be combined into an encoder structure. A Transformer can be composed of multiple parallel encoders to map the input sequence into a high-dimensional representation space and extract relationships and features within the sequence. Through multiple attentions, the encoder can focus on different aspects of information in parallel, thereby improving the model's sensitivity to different types of information.

[0195] (7) Decoder

[0196] The decoder primarily processes the hidden representations from the encoder output, utilizing a self-attention mechanism to transform the encoded contextual information into the target sequence. Similar to the encoder, the decoder also includes an attention layer that enables the decoder to focus on already generated portions of the target sequence to ensure that the current generation is consistent with the known generation. The decoder also includes an encoder-decoder attention layer that enables the decoder to focus on the hidden representations of the encoder output, ensuring that the decoder can effectively utilize information in the input sequence and better capture the correspondence between input and output.

[0197] The BiLSTM model consists of a forward LSTM and a backward LSTM. The forward LSTM processes the input sequence in forward order, primarily processing the forward information of the sequence data. The backward LSTM processes the input sequence in reverse order, primarily processing the backward information of the sequence data. The hidden states of the forward and backward LSTMs are then alternately combined along the time series dimension to generate a new bidirectional hidden state sequence. This fully utilizes both the forward and backward information of the input sequence.

[0198] BiLSTM is primarily composed of an input layer, a forward layer, a backward layer, an activation layer, and an output layer. BiLSTM can simultaneously consider both past and future information in a sequence, helping to understand contextual relationships within the sequence and exploit correlations in the input data sequence. Therefore, compared to standard LSTM, it has more comprehensive information processing capabilities.

[0199] The specific calculation process of BiLSTM is as follows:

[0200] Forward propagation:

[0201] f t =σ(W f h t-1 +W f x t +b f )

[0202] i t =σ(W i h t-1 +W i x t +b i )

[0203] c t =f t c t-1 +i t tanh(W c h t-1 +Wc x t +b c )

[0204] o t =σ(W o h t-1 +W o x t +b o )

[0205] h t =o t tanh(c t )

[0206] Among them, x t is the input signal at time t; h t is the hidden state; c t is the memory signal; f t It is the forget gate; t is the input gate; o t is the output gate; W i , W f , W o and W c are the weight matrices of the input gate, forget gate, output gate, and memory unit of the forward LSTM; b i , b f , b o and b c . are the biases of each gate respectively; tanh represents the hyperbolic tangent activation function, and σ represents the Sigmoid function.

[0207] Backward Propagation:

[0208]

[0209] Among them, x t is the input signal at the moment; is a hidden state; It is a memory signal; It is the Gate of Oblivion; is the input gate; is the output gate; and They are the weight matrices of the input gate, forget gate, output gate, and memory unit of the forward LSTM; and are the biases for each gate respectively.

[0210] S8: Model training and validation: Train and validate the constructed Transformer-BiLSTM model. Use common loss functions such as mean squared error (MSE) to optimize the model, and evaluate the model performance through cross-validation. Define the Early Stopping mechanism to prevent overfitting. Use four evaluation indicators, namely mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and coefficient of determination (R2), to build an evaluation system to evaluate the prediction accuracy of different models. The calculation method of each statistical indicator is as follows:

[0211]

[0212]

[0213] Among them, yi is the standard value of ET0, xi is the predicted value of ET0, is the mean of the predicted values, is the mean of the standardized values, and n is the number of actual observations or predicted values.

[0214] S9: TensorRT Optimization and Deployment: Export the trained model to ONNX format and use TensorRT for optimization and acceleration. After converting the model using TensorRT, deploy it on a hardware platform that supports GPU acceleration for real-time prediction.

Claims

1. An ETO prediction method based on signal decomposition and improved deep learning model, characterized in that: The following steps are involved: Step 1: Collect data; Step 2: Perform missing value processing, outlier detection, normalization, interpolation, and Min-Max standardization on the collected data to linearly map the interpolated data to between 0 and 1 to obtain normalized data; Step 3: Calculate ET0 based on the PM model; Step 4: Use the CEEMDAN algorithm to decompose the calculated ET0 into multiple IMFs. IMFs are different frequency components in the signal, which are used to extract data features at different time scales. Step 5: Combine all IMFs decomposed by the CEEMDAN algorithm with other meteorological feature vectors to form a complete input feature. Use the sliding window technique to construct a sample of the input feature. The sample includes the comprehensive feature vectors of the past W moments. Among them, the comprehensive feature vector at time t is: X t =F1(t),F2(t),,F M (t),IMF1(t),IMF2(t),,IMF N (t) The sample is: X=X t-W+1 ,X t-W+2 ,...,X t Among them, X is the input feature matrix, W is the number of comprehensive feature vectors at each moment; Step 6: Using ET0 as the target label, combine the input features of each sample with the corresponding ET0 to form a complete training sample. The training sample is divided into a training set, a validation set, and a test set according to the ratio of 70%, 20%, and 10% respectively. Step 7: Use the parallel computing framework Pytorch to build the Transformer-BiLSTM model; Step 8: Train and verify the constructed Transformer-BiLSTM model; Step 9: Export the trained Transformer-BiLSTM model to the ONNX format, optimize and accelerate the ONNX format using TensorRT, and perform real-time prediction on a hardware platform that supports GPU acceleration.

2. The ETO prediction method based on signal decomposition and improved deep learning model according to claim 1, characterized in that: The Min-Max normalization calculation formula of step 2 is: Among them, min(X) and max(X) are the minimum and maximum values of the feature respectively, X is the original data, and X norm The data are normalized.

3. The ETO prediction method based on signal decomposition and improved deep learning model according to claim 1, characterized in that: The PM model calculation formula in step 3 is: Where ET0 is the reference crop evapotranspiration, Δ is the slope of the saturated vapor pressure curve, Rn is the net surface radiation, G is the soil heat flux, γ is the dry-wet constant, and T mean is the daily average temperature, es and ea are the saturated vapor pressure value and the actual vapor pressure value, respectively.

4. The ETO prediction method based on signal decomposition and improved deep learning model according to claim 1, characterized in that: The CEEMDAN algorithm in step 4 includes the following sub-steps: Step 4-1: Add white noise w with a signal-to-noise ratio of ε0 to the ET0 sequence x(t) i (t)(i=1,2,...,n), and use the following formula to obtain the data x for the first decomposition i (t)(i=1,2,...,n): x i (t)=x(t)+ε0w i (t) Where i is the i-th white noise added, t is a different time point, and n is the total number of times white noise is added; Step 4-2: Use the CEEMDAN algorithm to decompose the ET0 series into a finite number of IMFs and residuals. Since the mean of white noise is zero, its influence can be eliminated by calculating the mean. Step 4-3: Add residual r k-1 Add adaptive noise E k (w (i) ), the calculation formula is: Where, β k-1 is the weight coefficient of k-1 stage Gaussian white noise; is the residual of the i-th white noise in the k-1 stage; Step 4-4: Obtain r according to the calculation formula k-1 The local mean M(r k-1 +β K-1 E k (w (i) )) and the average value, which is calculated as follows: Where I is the total number of EMD decompositions, r k is the residual sequence; Step 4-5: Obtain the kth IMF according to the calculation formula The calculation formula is: Where, is the kth IMF component; Step 4-6: When the residual has fewer than three extreme points and cannot be decomposed further, terminate the CEEMDAN algorithm.

5. The ETO prediction method based on signal decomposition and improved deep learning model according to claim 1, characterized in that: The step 7 includes the following sub-steps: Step 7-1: Input embedding to the training set, validation set, and test set to obtain the input embedding sequence; Step 7-2: Perform position encoding on the input embedding sequence to obtain the position encoding, which is calculated as follows: Among them, PE is the position code; pos is the position of the element in the vector; d is the dimension of PE; 2i is the even dimension; 2i+1 is the odd dimension; Step 7-3: Perform multi-head attention on the position encoding to obtain the query matrix Q, key matrix K and value matrix V, which are calculated as follows: Where X is the input matrix; W Q , W K , W V is the matrix to be trained into the model, d K is the dimension of vector V; Q, K and V are the selected feature data of ET0; Step 7-4: Add and normalize the input X; Step 7-5: Feed forward network for different position vectors; Step 7-6: Multi-head attention, feedforward network, addition and normalization can be combined into an encoder structure; Step 7-7: Decode the encoder.

6. The ETO prediction method based on signal decomposition and improved deep learning model according to claim 1, characterized in that: The BiLSTM model includes forward propagation and back propagation; The calculation formula for the forward propagation is: f t =σ(W f h t-1 +W f x t +b f ) i t =σ(W i h t-1 +W i x t +b i ) c t =f t c t-1 +i t tanh(W c h t-1 +W c x t +b c ) o t =σ(W o h t-1 +W o x t +b o ) h t =o t fishy t ) Among them, x t is the input signal at time t; h t is the hidden state; c t is the memory signal; f t is the forget gate; i t is the input gate; o t is the output gate; W i , W f , W o and W c are the weight matrices of the input gate, forget gate, output gate, and memory unit of the forward LSTM; b i , b f , b o and b c . are the biases of each gate respectively; tanh represents the hyperbolic tangent activation function, and σ represents the Sigmoid function; The calculation formula of the back propagation is: Among them, x t is the input signal at the moment; is a hidden state; It is a memory signal; It is the Gate of Oblivion; is the input gate; is the output gate; and They are the weight matrices of the input gate, forget gate, output gate, and memory unit of the forward LSTM; and are the biases for each gate respectively.

7. The ETO prediction method based on signal decomposition and improved deep learning model according to claim 1, characterized in that: The step S8 is specifically as follows: An evaluation system was constructed using four indicators to assess the prediction accuracy of different models; The four indicators are: mean absolute error MAE, mean square error MSE, root mean square error RMSE and determination coefficient R2, The calculation formula is: Among them, y i is the standard value of ET0, x i is the predicted value of ET0, is the mean of the predicted values, is the mean of the standardized values, and n is the number of actual observations or predicted values.

Images