Chlorophyll-a concentration multi-step prediction method across component interaction attention

Abstract

The application discloses a chlorophyll-a concentration multi-step prediction method based on cross-component interaction attention, and relates to the technical field of marine environment monitoring and prediction, and comprises the following steps: S01, acquiring an original chlorophyll-a concentration time sequence, performing multi-domain feature extraction on the original chlorophyll-a concentration time sequence, wherein the multi-domain feature extraction comprises extracting global periodicity features through Fourier analysis and extracting low-frequency approximation components and high-frequency detail components through wavelet decomposition; S02, inputting the global periodicity features, the low-frequency approximation components and the high-frequency detail components into a pre-constructed hierarchical feature selection gate unit, generating a dynamic weight vector through the gate unit, and adaptively weighting and screening the three kinds of features. The application realizes the essence transcendence of the disclosed simple parallel decomposition model, significantly improves the accuracy and robustness of multi-step prediction, and provides reliable technical support for early warning and accurate management of harmful algal blooms.

Description

Technical Field

[0001] This invention relates to the field of computer processing technology, and in particular to a multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention. Background Technology

[0002] Chlorophyll a (Chla) concentration is a core indicator for assessing the degree of eutrophication in water bodies and for early warning of harmful algal blooms. Achieving high-frequency, multi-step accurate prediction of Chla concentration is of vital importance for water resource management and ecological risk prevention and control.

[0003] Currently, time series prediction techniques for Chla concentration can be mainly divided into three categories. The first category is traditional linear models, such as ARIMA and its seasonal variant SARIMA. While these models can handle stationary linear sequences, the generation and evolution of algal blooms exhibit typical nonlinear and nonstationary characteristics, making it difficult for linear models to capture their complex dynamics. The second category is traditional machine learning methods, such as Random Forest (RF) and Support Vector Machine (SVM). These methods can partially alleviate the nonlinearity problem, but they generally lack long-term dependency memory mechanisms, making it difficult to effectively model the long-term evolutionary patterns of Chla concentration. The third category is deep learning models that have emerged in recent years, especially methods based on the Transformer architecture (such as KEDformer and ChloroFormer). Leveraging the advantages of self-attention mechanisms in capturing global dependencies, these methods have made progress in some time series prediction tasks.

[0004] However, the aforementioned existing technologies, especially the spatiotemporal fusion Transformer (STF-Transformer) and ChloroFormer models which are closest to this application, still have the following significant technical shortcomings when dealing with Chla concentration prediction: Existing technologies either employ only Fourier analysis for global spectral decomposition, neglecting local time-frequency information; or they use simple decomposition strategies that fail to simultaneously obtain global periodicity and local abrupt change features. More importantly, existing models lack hierarchical adaptive feature selection mechanisms, failing to dynamically assign weights to extracted multi-scale features. This results in redundant features and noise being treated equally, severely limiting the model's generalization ability and predictive robustness.

[0005] Secondly, existing decomposition prediction models (such as decomposing a sequence into low-frequency trends and high-frequency details) typically employ independent parallel strategies when predicting each component. This results in the prediction of high-frequency detail components failing to obtain macroscopic correction signals from low-frequency trend components, while low-frequency components cannot utilize high-frequency information for local refinement. This information isolation causes high-frequency prediction errors to accumulate rapidly during the recursive multi-step process, while macroscopic trend predictions also become inaccurate due to the lack of local feedback.

[0006] Furthermore, existing technologies generally employ a recursive strategy in multi-step prediction, where the prediction result of each step serves as the input for the next. Due to the lack of an effective cross-component interactive correction mechanism, small errors in previous steps are amplified in subsequent steps. This is particularly serious when predicting long sequences (e.g., more than 20 steps) of Chla concentration sequences with strong nonlinearity and abruptness, leading to a significant deviation of the final prediction result from the true value.

[0007] In summary, existing technologies for Chla concentration prediction generally suffer from low prediction accuracy and poor robustness due to incomplete feature extraction, component information isolation, and accumulation of recursive prediction errors, making it difficult to meet the high requirements for near-real-time early warning of harmful algal blooms. Summary of the Invention

[0008] To address the aforementioned technical problems, the present invention employs a multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention, comprising the following steps: S01. Obtain the original chlorophyll a concentration time series, and perform multi-domain feature extraction on the original chlorophyll a concentration time series. The multi-domain feature extraction includes extracting global periodic features through Fourier analysis, and extracting low-frequency approximate components and high-frequency detail components through wavelet decomposition. S02. Input the global periodic features, low-frequency approximation components and high-frequency detail components into a pre-constructed hierarchical feature selection gating unit, generate a dynamic weight vector through the gating unit, and adaptively weight and filter the three features to output an enhanced feature set. S03. The enhanced feature set is input into the pre-constructed core interactive prediction network. The network adopts a recursive multi-step prediction strategy and utilizes its internal cross-component interactive attention mechanism to use the prediction output of the low-frequency component as a correction signal to dynamically and interactively correct the prediction process of the high-frequency component, thereby generating low-frequency prediction results and high-frequency prediction results. S04. The low-frequency prediction results and high-frequency prediction results are inversely normalized and then reconstructed by inverse wavelet transform to obtain the final multi-step prediction results of chlorophyll a concentration.

[0009] Preferably, the step S01 of extracting low-frequency approximate components and high-frequency detail components by wavelet decomposition includes: using Daubechies wavelet as the basis function to perform single-layer decomposition on the original chlorophyll a concentration time series, decoupling the original series into low-frequency approximate components that reflect the overall trend and high-frequency detail components that capture sudden fluctuations.

[0010] Preferably, the structure of the hierarchical feature selection gating unit in step S02 includes: a first fully connected layer, a Sigmoid activation function layer, and a second fully connected layer. The generation of the dynamic weight vector includes: concatenating the global periodic features, low-frequency approximation components, and high-frequency detail components, and then sequentially passing them through the first fully connected layer, the Sigmoid activation function layer, and the second fully connected layer to generate a dynamic weight vector γ with a numerical range of 0 to 1 and consistent with the original multi-domain feature dimension.

[0011] Preferably, in step S02, adaptive weighting and filtering are performed, including: multiplying the dynamic weight vector γ element-wise with the original multi-domain feature vector composed of global periodic features, low-frequency approximation components and high-frequency detail components to achieve dynamic enhancement or suppression of different feature dimensions, thereby eliminating redundant information and generating the enhanced feature set.

[0012] Preferably, the core interaction prediction network in step S03 is a convolution-attention hybrid architecture based on Transformer, in which one-dimensional convolutional blocks for extracting local temporal dependencies are integrated in both the encoder and decoder, and the output of the one-dimensional convolutional blocks is fed into a multi-head attention mechanism to capture global dependencies. The feedforward neural network layers in the encoder and decoder use a gated linear unit structure instead of the traditional ReLU activation function. The gated linear unit structure consists of two parallel linear transformation layers and a sigmoid gate. Its output is the element-wise product of the two linear transformation results, one of which is activated by the sigmoid function. The multi-head attention mechanism is configured with different numbers of attention heads, with some attention heads set to 5 heads to capture long-term trends in chlorophyll a concentration sequences, and others set to 10 heads to capture high-frequency detail fluctuations in sequences.

[0013] Preferably, the cross-component interactive attention mechanism in step S03 is specifically implemented in the decoder module. Its interactive correction steps include: using the hidden state output by the branch predicting low-frequency components as the key matrix and value matrix in the attention mechanism, using the hidden state output by the branch predicting high-frequency detail components as the query matrix in the attention mechanism, obtaining the attention weight by calculating the similarity between the query matrix and the key matrix, and using the weight to perform a weighted summation on the value matrix, thereby realizing the unidirectional dynamic correction of the high-frequency detail prediction by the low-frequency macro trend information.

[0014] Preferably, the cross-component interactive attention mechanism also integrates a dynamic gating function, which calculates a dynamic correction factor based on the characteristic state of the current time step. The dynamic correction factor is used to control the degree to which low-frequency information influences high-frequency information.

[0015] Preferably, the step of reconstructing by inverse wavelet transform in step S04 specifically includes: superimposing the low-frequency prediction results after inverse normalization with the high-frequency prediction results to recover the complete chlorophyll a concentration time series at the original scale, and using root mean square error, mean absolute error and coefficient of determination to perform stepwise performance evaluation on the multi-step prediction results.

[0016] The present invention has at least the following beneficial effects: By adding a hierarchical feature selection gating unit and a cross-component interactive attention mechanism, the structure is completely different from traditional and publicly available simple parallel decomposition models, demonstrating a high degree of structural innovation.

[0017] The most complete time-frequency features were obtained through dual-domain analysis of Fourier and wavelet. At the same time, the key problem of error accumulation in recursive multi-step prediction was effectively solved through cross-component interactive correction, thereby significantly improving the accuracy of multi-step prediction tasks and the robustness across datasets.

[0018] The hierarchical feature selection gating unit can dynamically select the optimal features, effectively suppressing the interference of background noise and redundant information, and greatly enhancing the model's adaptive learning ability on complex Chla data. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 Here is a flowchart of a multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention, provided in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the internal structure of the interactive prediction network provided in Embodiment 1 of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0023] Example 1

[0024] This embodiment provides a multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention. The method includes the following steps: Figure 1 As shown: S01. Obtain the original chlorophyll a concentration time series and perform multi-domain feature extraction on the original chlorophyll a concentration time series. Multi-domain feature extraction includes extracting global periodic features through Fourier analysis and extracting low-frequency approximate components and high-frequency detail components through wavelet decomposition. Furthermore, the steps described above for extracting low-frequency approximate components and high-frequency detail components through wavelet decomposition include: using Daubechies wavelet as the basis function to perform single-layer decomposition on the original chlorophyll a concentration time series, decoupling the original series into low-frequency approximate components that reflect the overall trend and high-frequency detail components that capture sudden fluctuations.

[0025] Specifically, a dual-domain analysis strategy combining Fourier analysis and wavelet decomposition was employed. Fourier analysis, by transforming the time-domain signal to the frequency domain, reveals macroscopic periodic patterns in the Chla concentration sequence from a global perspective, such as dominant periodic components caused by seasonal changes, diurnal variations, or tidal patterns. Simultaneously, wavelet decomposition, as a time-frequency localization analysis tool, decouples the original nonlinear, non-stationary Chla concentration sequence into two components with clear physical meaning: the low-frequency approximation component reflects the overall evolution trend and dominant period of the sequence, while the high-frequency detail component captures rapid fluctuations and random noise caused by sudden events such as algal blooms and torrential rain.

[0026] Within the aforementioned multi-domain feature extraction framework, Daubechies wavelets (specifically db4 wavelets) are used as basis functions, and a single-level decomposition operation is performed on the original Chla concentration time series. This choice has clear technical considerations: the Daubechies wavelet family has good regularity and tight support characteristics, effectively balancing time resolution and frequency resolution, and is particularly suitable for processing non-stationary environmental time series with transient change characteristics. In the specific implementation process, the operating principle of this wavelet decomposition step is as follows: Let the original Chla concentration time series be x[n]. Using the selected db4 wavelet basis function, its inner product with the scaling function (low-pass filter) and the wavelet function (high-pass filter) is calculated respectively. The output of the low-pass filter corresponds to the low-frequency approximation component A1, which is essentially a coarse outline of the original sequence after smoothing, preserving the main trend and long-term dependencies of the sequence; the output of the high-pass filter corresponds to the high-frequency detail component D1, which is essentially the difference between the original sequence and the low-frequency approximation, containing local details, edge information, and random noise in the sequence. Because this step uses a single-level decomposition instead of a multi-level decomposition, it preserves sufficient temporal resolution and avoids information loss or computational redundancy caused by over-decomposition.

[0027] S02. Input the global periodic features, low-frequency approximation components and high-frequency detail components into the pre-constructed hierarchical feature selection gating unit. Generate a dynamic weight vector through the gating unit, and perform adaptive weighting and filtering on the three features to output an enhanced feature set.

[0028] Furthermore, the structure of the aforementioned hierarchical feature selection gating unit includes: a first fully connected layer, a Sigmoid activation function layer, and a second fully connected layer; The process of generating a dynamic weight vector involves concatenating global periodic features, low-frequency approximation components, and high-frequency detail components, and then sequentially passing them through a first fully connected layer, a Sigmoid activation function layer, and a second fully connected layer to generate a dynamic weight vector γ with a numerical range of 0 to 1 and consistent with the dimensions of the original multi-domain features.

[0029] Secondly, adaptive weighting and filtering are performed, including: multiplying the dynamic weight vector γ element-wise with the original multi-domain feature vector composed of global periodic features, low-frequency approximation components and high-frequency detail components to achieve dynamic enhancement or suppression of different feature dimensions, thereby eliminating redundant information and generating an enhanced feature set.

[0030] Specifically, through a learnable neural network subnetwork, the importance weight of each feature dimension is dynamically calculated, and the original features are weighted and enhanced or suppressed accordingly, thereby eliminating redundant information and noise and generating an enhanced feature set with higher information concentration. First, the three types of features obtained in step S01—global periodic feature vector, low-frequency approximate component feature vector, and high-frequency detail component feature vector—are concatenated along the feature dimension to form an original multi-domain feature vector X that contains information on macro-periods, overall trends, and local mutations.

[0031] Next, the concatenated feature vector X is input into the gating unit. This gating unit adopts a cascaded structure of "fully connected layer - Sigmoid activation function - fully connected layer". The first fully connected layer performs nonlinear fusion and dimensional projection on the input high-dimensional features to extract the interaction information between the features. Subsequently, the Sigmoid activation function compresses the output of the fully connected layer to the [0, 1] interval, generating an original attention weight vector. The second fully connected layer then maps this weight vector back to the same dimension as the original multi-domain feature vector X, finally obtaining the dynamic weight vector γ. It is worth noting that, due to the characteristics of the Sigmoid function, each element value in γ is between 0 and 1, representing the degree to which the corresponding feature dimension is preserved or enhanced—the closer the value is to 1, the more important the feature is; the closer it is to 0, the more the feature should be suppressed.

[0032] Finally, an adaptive weighted filtering operation is performed: the dynamic weight vector γ is multiplied element-wise with the original multi-domain feature vector X (i.e., Hadamard product) to obtain the enhanced feature set X' = γ ⊙ X. This operation allows the model to dynamically adjust its weights at each time step based on the actual contribution of different feature dimensions. For example, during a relatively stable winter period when chlorophyll a concentration is high, the corresponding weights of low-frequency approximation components (reflecting the overall trend) and global periodic features (reflecting the annual rhythm) may be close to 0.9, while the weights of high-frequency detail components (reflecting sudden disturbances) may be reduced to 0.2, allowing the model to focus more on macroscopic evolution. Conversely, during the peak of algal blooms in summer, the weights corresponding to high-frequency detail components will significantly increase, enabling the model to keenly capture short-term mutations. Through this element-wise weighting method, redundant or irrelevant features are effectively suppressed, and key features are highlighted and enhanced, thereby significantly improving the input quality and generalization ability of the subsequent core interactive prediction network.

[0033] S03. The enhanced feature set is input into the pre-constructed core interactive prediction network. The network adopts a recursive multi-step prediction strategy and utilizes its internal cross-component interactive attention mechanism to use the prediction output of the low-frequency component as a correction signal to dynamically and interactively correct the prediction process of the high-frequency component, generating low-frequency prediction results and high-frequency prediction results.

[0034] Furthermore, in the above embodiments, the core interaction prediction network is a convolution-attention hybrid architecture based on Transformer improvement. Both its encoder and decoder integrate one-dimensional convolutional blocks for extracting local temporal dependencies. The output of the one-dimensional convolutional blocks is fed into a multi-head attention mechanism to capture global dependencies. The feedforward neural network layers in the encoder and decoder use a gated linear unit structure instead of the traditional ReLU activation function. The gated linear unit structure consists of two parallel linear transformation layers and a sigmoid gate. Its output is the element-wise product of the two linear transformation results, one of which is activated by the sigmoid function. The multi-head attention mechanism is configured with different numbers of attention heads, with some attention heads set to 5 heads to capture long-term trends in chlorophyll a concentration sequences, and others set to 10 heads to capture high-frequency detail fluctuations in the sequences.

[0035] Specifically, for the input feature sequence, a one-dimensional convolutional block first performs a sliding window convolution operation along the time dimension, with the kernel size set to, for example, 3 and a stride of 1, to extract local temporal patterns within each time step and its immediate and neighboring regions. The output of this convolutional block preserves the local structural information of the original sequence and passes these feature maps rich in local dependencies to the subsequent multi-head attention layer. The multi-head attention layer then captures the long-distance dependencies between any two time steps in the sequence by calculating the similarity between the query matrix (Q), the key matrix (K), and the value matrix (V). Through this hybrid architecture of "local convolution first, global attention later," the network can sensitively capture sudden fluctuations in algal concentration within a short period (e.g., within a few hours) and stably associate long-term periodic patterns across seasons or months.

[0036] For an input feature vector x, it is first projected into a high-dimensional space through two parallel linear transformation layers (with weight matrices W and V, respectively), resulting in two intermediate feature vectors (xW+b) and (xV+c). One branch is activated by the Sigmoid function, outputting a gated signal with values ​​between 0 and 1; the other branch remains unchanged. Finally, the output of the GLU is the element-wise product of the two branch vectors, i.e. ; Where x is the input feature vector, W and V are the weight matrices of the two linear transformation layers, and b and c are the bias terms. It is the Sigmoid activation function. This is for element-wise multiplication.

[0037] GLU controls the information flow through a gating mechanism, achieving a more powerful nonlinear feature mapping capability. This gating mechanism allows the network to adaptively determine which information can continue to propagate and which should be suppressed based on the dynamic state of the input features. Compared to ReLU's hard truncation (negative values ​​are directly set to zero), GLU provides smoother and more flexible nonlinear control capabilities, making it particularly suitable for handling non-stationary signals with sudden spikes and drastic fluctuations in Chla sequences, significantly improving the model's ability to represent complex distributions.

[0038] This isolated decoder contains two parallel computational branches: a low-frequency prediction branch (L-Stream) and a high-frequency prediction branch (D-Stream). Each branch independently maintains a set of query (Q), key (K), and value (V) matrices. The key to this invention is that in each layer of attention computation in the D-Stream, its query matrix (QD) remains unchanged, but the key and value matrices no longer use their own KD and VD; instead, they borrow the key matrix (KL) and value matrix (VL) from the L-Stream branch. Mathematically, the output of this interactive attention can be expressed as: .

[0039] S04. Perform inverse normalization on the low-frequency and high-frequency prediction results, and then reconstruct them through inverse wavelet transform to obtain the final multi-step prediction results of chlorophyll a concentration.

[0040] Furthermore, the reconstruction steps via inverse wavelet transform specifically include: superimposing the low-frequency prediction results after inverse normalization with the high-frequency prediction results to recover the complete chlorophyll a concentration time series at the original scale, and using root mean square error, mean absolute error, and coefficient of determination to perform stepwise performance evaluation on the multi-step prediction results.

[0041] Specifically, inverse normalization is performed. Before model training, the original Chla concentration data is typically normalized (e.g., using the RobustScaler method, which scales based on the median and interquartile range) to eliminate the influence of dimensions and accelerate model convergence. Therefore, both the low-frequency and high-frequency predictions output by the core interactive prediction network are within the normalized numerical space. Inverse normalization is the inverse operation of this process: the system saves the normalization parameters used during training (such as mean, standard deviation, or the median and interquartile range of RobustScaler), and then substitutes each predicted value output by the network into the inverse transformation formula, mapping it from the normalized space back to the original Chla concentration numerical space. For example, if the Z-score method is used for normalization, the inverse normalization formula is: Original value = Predicted value × Original standard deviation + Original mean.

[0042] Secondly, inverse wavelet transform is performed for reconstruction. Wavelet decomposition (e.g., using the Daubechies wavelet db4 as the basis function) was employed in the feature extraction stage to decouple the original Chla sequence into low-frequency approximate components (reflecting the overall trend) and high-frequency detail components (reflecting local fluctuations). The inverse wavelet transform is the precise inverse process of this decomposition. The low-frequency and high-frequency predicted components, after inverse normalization, are used as inputs and passed layer by layer through wavelet reconstruction filters (including upsampling and convolution operations). The two components are then re-aligned and superimposed, ultimately synthesizing a complete Chla concentration prediction sequence with the same length and scale as the original time series.

[0043] Example 2

[0044] Based on the above embodiment one, the cross-component interactive attention mechanism in this example specifically runs in the decoder module. Its interactive correction steps include: using the hidden state output by the branch predicting low-frequency components as the key matrix and value matrix in the attention mechanism, using the hidden state output by the branch predicting high-frequency detail components as the query matrix in the attention mechanism, obtaining the attention weight by calculating the similarity between the query matrix and the key matrix, and using the weight to perform a weighted summation on the value matrix, thereby realizing the unidirectional dynamic correction of the high-frequency detail prediction by the low-frequency macro trend information.

[0045] Secondly, the cross-component interactive attention mechanism also integrates a dynamic gating function. The dynamic gating function calculates a dynamic correction factor based on the characteristic state of the current time step. The dynamic correction factor is used to control the degree to which low-frequency information affects high-frequency information.

[0046] Specifically, the hidden states output by the branch predicting low-frequency components (i.e., trend information) are set as the "key" (K matrix) and "value" (V matrix) in the attention mechanism; simultaneously, the hidden states output by the branch predicting high-frequency detail components (i.e., fluctuation information) are set as the "query" (Q matrix) in the attention mechanism. From an operational perspective, this design essentially uses the global, smooth macroscopic evolutionary information provided by the low-frequency branch as a reference benchmark to guide and constrain the attention focusing process of the high-frequency branch on local sudden fluctuations. In the calculation process, a similarity matrix is ​​first obtained through the dot product of the Q matrix and the K matrix. This similarity represents the correlation strength between each time step of the current high-frequency detail and each time step in the low-frequency macroscopic sequence. Then, the similarity is normalized into attention weights using the Softmax function. Finally, the V matrix is ​​weighted and summed using these attention weights. The resulting high-frequency detail representation incorporates low-frequency macroscopic correction information.

[0047] To further enhance the flexibility and effectiveness of this interaction, and to prevent low-frequency information from excessively smoothing or "overcorrecting" high-frequency fluctuations, this mechanism also integrates a dynamic gating function. The function operates by not applying the same intensity of correction at all time steps and in all states. Instead, based on the fusion characteristics of multi-dimensional information such as the input features and hidden states at the current time step, a "dynamic correction factor" between 0 and 1 is dynamically calculated through a learnable gating network (typically composed of a fully connected layer connected to a sigmoid activation function). This factor controls the degree to which low-frequency information influences high-frequency information.

[0048] Example 3

[0049] This invention provides a non-transitory computer-readable storage medium storing at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the following steps: The original chlorophyll a concentration time series was obtained, and multi-domain feature extraction was performed on the original chlorophyll a concentration time series. Multi-domain feature extraction included extracting global periodic features through Fourier analysis, and extracting low-frequency approximate components and high-frequency detail components through wavelet decomposition. The global periodic features, low-frequency approximation components, and high-frequency detail components are input into a pre-constructed hierarchical feature selection gating unit. The gating unit generates a dynamic weight vector and adaptively weights and filters the three features to output an enhanced feature set. The enhanced feature set is input into the pre-built core interactive prediction network. The network adopts a recursive multi-step prediction strategy and utilizes its internal cross-component interactive attention mechanism to dynamically and interactively correct the prediction process of the high-frequency components by using the prediction output of the low-frequency components as a correction signal, thereby generating low-frequency prediction results and high-frequency prediction results. The low-frequency and high-frequency prediction results are inversely normalized and then reconstructed using inverse wavelet transform to obtain the final multi-step prediction results of chlorophyll a concentration.

[0050] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0051] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0052] Example 4

[0053] This invention provides an electronic device, including a processor and a memory, wherein the memory stores at least one instruction or at least one program segment, and the at least one instruction or the at least one program segment is loaded and executed by the processor to implement the following steps: The original chlorophyll a concentration time series was obtained, and multi-domain feature extraction was performed on the original chlorophyll a concentration time series. Multi-domain feature extraction included extracting global periodic features through Fourier analysis, and extracting low-frequency approximate components and high-frequency detail components through wavelet decomposition. The global periodic features, low-frequency approximation components, and high-frequency detail components are input into a pre-constructed hierarchical feature selection gating unit. The gating unit generates a dynamic weight vector and adaptively weights and filters the three features to output an enhanced feature set. The enhanced feature set is input into the pre-built core interactive prediction network. The network adopts a recursive multi-step prediction strategy and utilizes its internal cross-component interactive attention mechanism to dynamically and interactively correct the prediction process of the high-frequency components by using the prediction output of the low-frequency components as a correction signal, thereby generating low-frequency prediction results and high-frequency prediction results. The low-frequency and high-frequency prediction results are inversely normalized and then reconstructed using inverse wavelet transform to obtain the final multi-step prediction results of chlorophyll a concentration.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention, characterized in that, Includes the following steps: S01. Obtain the original chlorophyll a concentration time series, and perform multi-domain feature extraction on the original chlorophyll a concentration time series. The multi-domain feature extraction includes extracting global periodic features through Fourier analysis, and extracting low-frequency approximate components and high-frequency detail components through wavelet decomposition. S02. Input the global periodic features, low-frequency approximation components and high-frequency detail components into a pre-constructed hierarchical feature selection gating unit, generate a dynamic weight vector through the gating unit, and adaptively weight and filter the three features to output an enhanced feature set. S03. The enhanced feature set is input into the pre-constructed core interactive prediction network. The network adopts a recursive multi-step prediction strategy and utilizes its internal cross-component interactive attention mechanism to use the prediction output of the low-frequency component as a correction signal to dynamically and interactively correct the prediction process of the high-frequency component, thereby generating low-frequency prediction results and high-frequency prediction results. S04. The low-frequency prediction results and high-frequency prediction results are inversely normalized and then reconstructed by inverse wavelet transform to obtain the final multi-step prediction results of chlorophyll a concentration.

2. The multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention as described in claim 1, characterized in that, The step S01, which involves extracting low-frequency approximate components and high-frequency detail components through wavelet decomposition, includes: using Daubechies wavelet as the basis function to perform single-layer decomposition on the original chlorophyll a concentration time series, thereby decoupling the original series into low-frequency approximate components that reflect the overall trend and high-frequency detail components that capture sudden fluctuations.

3. The multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention as described in claim 1, characterized in that, The structure of the hierarchical feature selection gating unit in step S02 includes: a first fully connected layer, a Sigmoid activation function layer, and a second fully connected layer. The generation of the dynamic weight vector includes: concatenating the global periodic features, low-frequency approximation components, and high-frequency detail components, and then sequentially passing them through the first fully connected layer, the Sigmoid activation function layer, and the second fully connected layer to generate a dynamic weight vector γ with a numerical range of 0 to 1 and consistent with the original multi-domain feature dimension.

4. The multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention as described in claim 3, characterized in that, In step S02, adaptive weighting and filtering are performed, including: multiplying the dynamic weight vector γ element-wise with the original multi-domain feature vector composed of global periodic features, low-frequency approximation components and high-frequency detail components to achieve dynamic enhancement or suppression of different feature dimensions, thereby eliminating redundant information and generating the enhanced feature set.

5. The multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention as described in claim 1, characterized in that, The core interaction prediction network in step S03 is a hybrid convolution-attention architecture based on Transformer. Both its encoder and decoder integrate one-dimensional convolutional blocks for extracting local temporal dependencies. The output of the one-dimensional convolutional blocks is fed into a multi-head attention mechanism to capture global dependencies. The feedforward neural network layers in the encoder and decoder use a gated linear unit structure instead of the traditional ReLU activation function. The gated linear unit structure consists of two parallel linear transformation layers and a sigmoid gate. Its output is the element-wise product of the two linear transformation results, one of which is activated by the sigmoid function. The multi-head attention mechanism is configured with different numbers of attention heads, with some attention heads set to 5 heads to capture long-term trends in chlorophyll a concentration sequences, and others set to 10 heads to capture high-frequency detail fluctuations in sequences.

6. The multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention as described in claim 5, characterized in that, The cross-component interactive attention mechanism described in step S03 specifically operates in the decoder module. Its interactive correction steps include: using the hidden state output by the branch predicting low-frequency components as the key matrix and value matrix in the attention mechanism, using the hidden state output by the branch predicting high-frequency detail components as the query matrix in the attention mechanism, obtaining the attention weight by calculating the similarity between the query matrix and the key matrix, and using the weight to perform a weighted summation on the value matrix, thereby realizing the unidirectional dynamic correction of high-frequency detail prediction by low-frequency macro trend information.

7. The multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention as described in claim 6, characterized in that, The cross-component interactive attention mechanism also integrates a dynamic gating function, which calculates a dynamic correction factor based on the characteristic state of the current time step. The dynamic correction factor is used to control the degree to which low-frequency information influences high-frequency information.

8. The multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention as described in claim 1, characterized in that... The step of reconstructing by inverse wavelet transform in step S04 specifically includes: superimposing the low-frequency prediction results after inverse normalization with the high-frequency prediction results to recover the complete chlorophyll a concentration time series at the original scale, and using root mean square error, mean absolute error and coefficient of determination to perform stepwise performance evaluation on the multi-step prediction results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the multi-step prediction method for chlorophyll a concentration based on cross-component interactive attention as described in any one of claims 1 to 8.

Images