Algal community structure consistency prediction method based on hierarchical deep learning

Abstract

This invention discloses a method for consistent forecasting of algal community structure based on hierarchical deep learning, belonging to the field of intelligent water ecology prediction technology. The method includes the following steps: S1, constructing an original time-series dataset containing biomass of different algal groups, total algal biomass, water quality factors, and hydrological and meteorological factors, and performing data preprocessing and standardization; S2, generating basic forecast results for biomass of different algal groups and total algal biomass using an LSTM-KAN basic forecasting model; S3, embedding an end-to-end hierarchical coordination module based on dynamic weight learning into the LSTM-KAN basic forecasting model to perform consistency correction on the basic forecast results for different algal groups and total algal biomass; S4, performing parameter optimization, model training, and performance evaluation, and outputting consistent forecast results that satisfy the hierarchical summation relationship between biomass of different algal groups and total algal biomass. This invention improves forecast accuracy and can provide core technical support for water ecology prediction and management.

Description

Technical Field

[0001] This invention belongs to the field of intelligent prediction technology for aquatic ecosystems, specifically involving a method for predicting the consistency of algal community structure based on hierarchical deep learning. Background Technology

[0002] Algae are the core primary producers in aquatic ecosystems, and their community structure (i.e., the biomass composition of various algal communities) directly reflects the trophic status of water bodies and the stability of ecosystems. In practical water environment management, total algal biomass forecasting is used to assess the overall risk of eutrophication in water bodies, while biomass forecasting of various algal communities supports the identification of dominant algae and the determination of algal bloom types; both are indispensable.

[0003] Algal community structure changes are driven by multiple factors, including water quality, hydrology, and meteorology, exhibiting strong nonlinearity, time lag, seasonality, and abrupt changes. Traditional statistical methods (such as multiple linear regression and time series analysis) and shallow machine learning methods (such as support vector machines and random forests) are insufficient to fully characterize the complex multi-factor response relationships. In recent years, deep learning models, with their powerful time series modeling capabilities, have been widely used in algal biomass forecasting, significantly improving forecast accuracy.

[0004] However, existing methods generally suffer from two core flaws: First, they often employ independent modeling, forecasting total algal biomass and biomass of various populations separately, completely ignoring the hierarchical summation constraint between the two. This leads to numerical contradictions in the forecast results, such as "the sum of biomass of various populations is not equal to total algal biomass," making reasonable ecological interpretation impossible. Second, the few hierarchical coordination methods used employ a two-stage model of "predict first, then correct," where a basic forecast is generated by an independent model first, and then consistency correction is performed by an external algorithm. This approach not only severs the integration of hierarchical structural information with the forecasting process but also fails to adaptively coordinate the forecast accuracy at different levels under different ecological conditions, often sacrificing some forecast accuracy for consistency.

[0005] Therefore, there is an urgent need to develop an algal community structure forecasting method that can simultaneously ensure forecast accuracy and hierarchical consistency. This method should deeply integrate hierarchical structure constraints into the training process of deep learning models to achieve end-to-end consistent forecasting and provide reliable technical support for the refined management of aquatic ecological environment. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a hierarchical deep learning-based method for predicting the consistency of algal community structure. This method captures the complex nonlinear characteristics of algal biomass changes using an LSTM-KAN basic prediction model, while embedding an end-to-end hierarchical coordination module based on dynamic weight learning. During model training, it automatically satisfies the hierarchical summation constraint between total algal and biomass of various groups, achieving a synergistic improvement in accuracy and consistency. This solves the problem of inconsistent numerical relationships and ecological interpretations caused by independent modeling of various algal groups in existing algal biomass predictions, and can provide core technical support for eutrophication risk identification, algal bloom early warning, and refined water ecology management.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A hierarchical deep learning-based method for predicting the consistency of algal community structure includes the following steps: S1. Construct a raw time series dataset containing biomass of different algal groups, total algal biomass, water quality factors and hydro-meteorological factors, and perform data preprocessing and standardization. S2. Basic forecast results of biomass of different algal groups and total algal biomass were generated using the LSTM-KAN basic forecast model. S3. An end-to-end hierarchical coordination module based on dynamic weight learning is embedded in the LSTM-KAN basic forecast model to make consistent corrections to the basic forecast results of different algal groups and total algal biomass. S4. Perform parameter optimization, model training and performance evaluation, and output consistent prediction results that satisfy the hierarchical summation relationship between biomass of different algal groups and total algal biomass.

[0008] Preferably, the specific process of step S1 is as follows: S11. Collect biomass data of different algal groups, total algal biomass data, water quality factor data and hydro-meteorological factor data of the target water body during the continuous monitoring period, and construct the original time series dataset; S12. Preprocess the raw time series data, including time series alignment, outlier identification, missing value imputation, and feature selection. S13. Set the total algal biomass as the top-level node and the biomass of different algal groups as the bottom-level nodes. The hierarchical summation relationship is expressed as follows: In the formula, Let be the total algal biomass at time point t; t is the time number; A is the identifier of the total algal community at the top layer; c is the algal community number. The total number of algal groups; For the first Each time point Taxonomic biomass; This serves as an identifier for the c-th type of basal algae group; S14. Perform Z-score standardization on the preprocessed data. The calculation formula is as follows: In the formula, For the t-th time point The standardized values ​​of each variable; Number the variables; For the t-th time point The original values ​​of each variable; For the first The mean of each variable in the training set; For the first The standard deviation of each variable in the training set.

[0009] Preferably, the specific process of step S2 is as follows: S21. Based on standardized time-series data, the biomass of different algal groups, total algal biomass, water quality factors, and hydrological and meteorological factors within historical periods are used as model inputs, and the biomass of different algal groups and total algal biomass at future times are used as model outputs. The basic forecast results are as follows: In the formula, h is the prediction step size; These are model parameters; This is the historical input sequence corresponding to the t-th time point; This is the LSTM-KAN basic forecast model; This provides the basic forecast results for the biomass of different algal groups and total algal biomass at the next time point t+h, i.e. In the formula, The basic forecast result of total algal biomass at the next time point t+h; For the time point t+h in the future Basic biomass forecast results for different algal groups; This is the transpose of the matrix; S22. The time-dependent features of algal biomass changes are extracted using the coding layer of an LSTM network to capture the nonlinear, time-delay, and seasonal characteristics of algal community structure changes. The state update process is represented as follows: , , , , , In the formula, Let be the standardized input vector at time point t; Use the Sigmoid activation function; Represents element-wise product; , , , For bias terms; , , , This is the weight matrix; Let be the hidden state of the LSTM network at time t. Let be the hidden state of the LSTM network at time t-1. Let be the candidate cell state at time point t; Let be the cell state at time point t; This represents the cell state at time point t-1. , , These are the forget gate, input gate, and output gate, respectively. S23. The hidden state output by the coding layer of the LSTM network. The input is a KAN network, which uses a learnable nonlinear function to characterize the mapping relationship between hidden features and target biomass. The output mapping process is as follows: In the formula, is the dimension of the hidden features in the LSTM network; The number of the hidden feature; For the first One hidden feature; For the connection in the KAN network The hidden feature and the first A learnable nonlinear function of output variables; For the future t+h time point The basic forecast results for each output variable, when When, corresponding to the total algal biomass; when At that time, respectively corresponding to Algal biomass; S24. Utilize the LSTM-KAN basic forecast model to output the basic forecast results of biomass for different algal groups and total algal biomass. .

[0010] Preferably, the specific process of step S3 is as follows: S31. Based on the hierarchical summation relationship between the biomass of different algal groups and the total algal biomass, construct a hierarchical summation matrix for hierarchical coordination calculation: In the formula, It is a hierarchical summation matrix; Represents a set of real number elements OK, Column matrix space; The first row of the matrix represents the total number of algal groups at the bottom layer. The first row of the matrix is ​​used to represent the aggregation relationship between the total algal biomass and the biomass of each algal group. The remaining rows are used to retain the bottom layer node information corresponding to the biomass of different algal groups. S32. After the basic forecast results are denormalized to restore them to the original biomass scale, they are input into a dynamic weight generation network constructed by a multilayer perceptron and jointly optimized with the LSTM-KAN basic forecast model. The coordinated weights of the basic biomass forecast results at different levels are adaptively learned under the current forecast state, as shown in: , In the formula, For parameters The dynamic weight generation network; Softplus is a positive activation function; For the dynamic coordination weight vector at the (t+h)th future time point; The dynamic coordination weights corresponding to the basic forecast results of total algal biomass; For the time point t+h in the future Dynamic coordination weights corresponding to the basic forecast results of biomass of different algal groups; S33. Based on the dynamic coordination weights, construct the dynamic coordination weight matrix for the (t+h)th future time point. The calculation formula is as follows: In the formula, This is the dynamic coordination weight matrix for the future time point t+h, used to characterize the differences in coordination weights between different levels of basic forecast results under the current forecast state; These are positive constants used to ensure the numerical stability of matrix inversion and model training processes; S34, Based on hierarchical summation matrix Dynamic Coordination Weight Matrix and basic forecast results End-to-end hierarchical coordination calculations are performed to coordinate the basic forecast results of biomass for different algal groups and total algal biomass. The calculation formula is as follows: , , , In the formula, The results of the coordinated forecast of total algal biomass at the next time point t+h; For the time point t+h in the future Prediction results of biomass coordination among different algal groups; This is the coordinated biomass forecast vector for different algal groups at the bottom layer at the (t+h)th time point in the future. This is the coordinated, fully consistent forecast vector for the (t+h)th time point in the future. For hierarchical summation matrix The transpose of the matrix; For dynamically coordinated weight matrix The inverse matrix.

[0011] Preferably, the specific process of step S4 is as follows: S41. The time series sample dataset, which has been preprocessed and standardized in step S1, is divided into training set, validation set and test set in 7:2:1 ratio according to time order, and historical input sequence samples and their corresponding future output target vectors are constructed based on sliding window. S42. The hyperparameters of the joint model consisting of the LSTM-KAN base prediction model and the end-to-end hierarchical coordination module are optimized using a grid search method. The optimal hyperparameter combination is determined based on minimizing the validation set loss. The calculation formula is as follows: In the formula, The optimal combination of hyperparameters; The combination of hyperparameters to be searched; For the hyperparameter search space; This represents the validation set loss under the corresponding hyperparameter combination; In hyperparameter space Searching for the objective function Minimize the set of hyperparameters ; S43. The joint model is trained end-to-end according to the optimal hyperparameter combination. Under the hierarchical consistency constraint, the optimization objective is to minimize the error between the coordinated forecast results and the actual observations. The parameters of the LSTM-KAN basic forecast model and the dynamic weight generation network are updated synchronously through gradient backpropagation. The trained joint model is used for forecasting on the test set and outputs a full-level consistent forecast result that satisfies the hierarchical summation constraint of total algal biomass and biomass of different algal groups. S44. Evaluate model performance in terms of both forecast accuracy and hierarchical consistency: Forecast accuracy is evaluated using mean absolute error (MAE) and Nash efficiency coefficient (NSE). By calculating the MAE and NSE of the basic forecast and the coordinated forecast respectively, the changes in forecast accuracy before and after hierarchical coordination are compared. Hierarchical consistency is evaluated by comparing the summation difference between the total algal biomass forecasts before and after coordination and the biomass forecasts for different algal groups. The calculation formula is as follows: , In the formula, This represents the number of test samples; Number the test samples; For the first Basic forecast results of total algal biomass before coordination of individual test samples; For the first Before the coordination of individual test samples Basic forecast results of algal biomass; For the first Coordinated forecast results of total algal biomass after coordination of individual test samples; For the first After coordination of individual test samples Predicted values ​​for biomass coordination among algal groups; The summation difference between the base forecast results and the total algal biomass before harmonization and the biomass of different algal groups; Based on the basic forecast results and the summation difference between the total algal biomass and the biomass of different algal groups after harmonization, under the condition of satisfying hierarchical summation constraints, =0; S45. Output the consistent forecast results of total algal biomass and biomass of different algal groups at future time.

[0012] Preferably, in step S42, the hyperparameters of the LSTM-KAN basic prediction model include the historical input window length, hidden layer dimension, number of network layers, learning rate, batch size, and regularization coefficient; the hyperparameters of the end-to-end hierarchical coordination module include the hidden layer dimension and number of network layers of the dynamic weight generation network.

[0013] By adopting the above technical solution, the present invention has the following beneficial effects: 1. This invention embeds an end-to-end hierarchical coordination module into a deep learning prediction model, ensuring that the output results strictly satisfy the summation constraint of total algal biomass and biomass of various populations, thus avoiding numerical contradictions and inconsistent ecological interpretations caused by independent modeling.

[0014] 2. This invention employs the LSTM-KAN basic forecast model, which is a hybrid deep learning model used to generate basic forecast results of algal biomass. It fully leverages the temporal feature extraction advantages of LSTM (Long Short-Term Memory Network) and the nonlinear mapping capability of KAN (Kolmogorov-Arnold Network). Compared with traditional deep learning models, it can more accurately simulate the complex changes in algal community structure and improve forecast accuracy.

[0015] 3. This invention uses a dynamic weight generation network to automatically adjust and coordinate weights based on the differences in forecast accuracy at different forecast times. While ensuring consistency, it maximizes the preservation of the accuracy of the basic forecast, solving the problem of severe accuracy loss in traditional two-stage methods and enabling adaptive coordination.

[0016] 4. This invention can simultaneously output consistent forecast results of total algal biomass and biomass of various algal communities, which can directly support the decision-making of eutrophication risk classification, algal bloom type early warning and algal community structure regulation, and provide core technical means for the refined management of aquatic ecological environment, and has important application value. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] like Figure 1 As shown, a method for predicting the consistency of algal community structure based on hierarchical deep learning includes the following steps: S1. Construct a raw time series dataset containing biomass of different algal groups, total algal biomass, water quality factors and hydro-meteorological factors, and perform data preprocessing and standardization. The specific process of step S1 is as follows: S11. Collect biomass data of different algal groups, total algal biomass data, water quality factor data and hydro-meteorological factor data of the target water body during the continuous monitoring period, and construct the original time series dataset; S12. Preprocess the raw time series data, including time series alignment, outlier identification, missing value imputation, and feature selection. S13. Set the total algal biomass as the top-level node and the biomass of different algal groups as the bottom-level nodes. The hierarchical summation relationship is expressed as follows: In the formula, Let be the total algal biomass at time point t; t is the time number; A is the identifier of the total algal community at the top layer; c is the algal community number. The total number of algal groups; For the first Each time point Taxonomic biomass; This serves as an identifier for the c-th type of basal algae group; S14. Perform Z-score standardization on the preprocessed data. The calculation formula is as follows: In the formula, For the t-th time point The standardized values ​​of each variable; Number the variables; For the t-th time point The original values ​​of each variable; For the first The mean of each variable in the training set; For the first The standard deviation of each variable in the training set; S2. Generate basic forecast results of biomass of different algal groups and total algal biomass through the LSTM-KAN basic forecast model; where LSTM is a long short-term memory network, KAN is a Kolmogorov-Arnold network, and the LSTM-KAN basic forecast model is a hybrid deep learning model used to generate basic forecast results of algal biomass. The specific process of step S2 is as follows: S21. Based on standardized time-series data, the biomass of different algal groups, total algal biomass, water quality factors, and hydrological and meteorological factors within historical periods are used as model inputs, and the biomass of different algal groups and total algal biomass at future times are used as model outputs. The basic forecast results are as follows: In the formula, h is the prediction step size; These are model parameters; This is the historical input sequence corresponding to the t-th time point; This is the LSTM-KAN basic forecast model; This provides the basic forecast results for the biomass of different algal groups and total algal biomass at the next time point t+h, i.e. In the formula, The basic forecast result of total algal biomass at the next time point t+h; For the time point t+h in the future Basic biomass forecast results for different algal groups; This is the transpose of the matrix; S22. The time-dependent features of algal biomass changes are extracted using the coding layer of an LSTM network to capture the nonlinear, time-delay, and seasonal characteristics of algal community structure changes. The state update process is represented as follows: , , , , , In the formula, Let be the standardized input vector at time point t; Use the Sigmoid activation function; Represents element-wise product; , , , For bias terms; , , , This is the weight matrix; Let be the hidden state of the LSTM network at time t. Let be the hidden state of the LSTM network at time t-1. Let be the candidate cell state at time point t; Let be the cell state at time point t; This represents the cell state at time point t-1. , , These are the forget gate, input gate, and output gate, respectively. S23. The hidden state output by the coding layer of the LSTM network. The input is a KAN network, which uses a learnable nonlinear function to characterize the mapping relationship between hidden features and target biomass. The output mapping process is as follows: In the formula, is the dimension of the hidden features in the LSTM network; The number of the hidden feature; For the first One hidden feature; For the connection in the KAN network The hidden feature and the first A learnable nonlinear function of output variables; For the future t+h time point The basic forecast results for each output variable, when When, corresponding to the total algal biomass; when At that time, respectively corresponding to Algal biomass; S24. Utilize the LSTM-KAN basic forecast model to output the basic forecast results of biomass for different algal groups and total algal biomass. ; S3. An end-to-end hierarchical coordination module based on dynamic weight learning is embedded in the LSTM-KAN basic forecast model to make consistent corrections to the basic forecast results of different algal groups and total algal biomass. The specific process of step S3 is as follows: S31. Based on the hierarchical summation relationship between the biomass of different algal groups and the total algal biomass, construct a hierarchical summation matrix for hierarchical coordination calculation: In the formula, It is a hierarchical summation matrix; Represents a set of real number elements OK, Column matrix space; The first row of the matrix represents the total number of algal groups at the bottom layer. The first row of the matrix is ​​used to represent the aggregation relationship between the total algal biomass and the biomass of each algal group. The remaining rows are used to retain the bottom layer node information corresponding to the biomass of different algal groups. S32. After the basic forecast results are denormalized to restore them to the original biomass scale, they are input into a dynamic weight generation network constructed by a multilayer perceptron and jointly optimized with the LSTM-KAN basic forecast model. The coordinated weights of the basic biomass forecast results at different levels are adaptively learned under the current forecast state, as shown in: , In the formula, For parameters The dynamic weight generation network; Softplus is a positive activation function; For the dynamic coordination weight vector at the (t+h)th future time point; The dynamic coordination weights corresponding to the basic forecast results of total algal biomass; For the time point t+h in the future Dynamic coordination weights corresponding to the basic forecast results of biomass of different algal groups; S33. Based on the dynamic coordination weights, construct the dynamic coordination weight matrix for the (t+h)th future time point. The calculation formula is as follows: In the formula, This is the dynamic coordination weight matrix for the future time point t+h, used to characterize the differences in coordination weights between different levels of basic forecast results under the current forecast state; These are positive constants used to ensure the numerical stability of matrix inversion and model training processes; S34, Based on hierarchical summation matrix Dynamic Coordination Weight Matrix and basic forecast results End-to-end hierarchical coordination calculations are performed to coordinate the basic forecast results of biomass for different algal groups and total algal biomass. The calculation formula is as follows: , , , In the formula, The results of the coordinated forecast of total algal biomass at the next time point t+h; For the time point t+h in the future Prediction results of biomass coordination among different algal groups; This is the coordinated biomass forecast vector for different algal groups at the bottom layer at the (t+h)th time point in the future. This is the coordinated, fully consistent forecast vector for the (t+h)th time point in the future. For hierarchical summation matrix The transpose of the matrix; For dynamically coordinated weight matrix The inverse matrix; S4. Perform parameter optimization, model training and performance evaluation, and output consistent prediction results that satisfy the hierarchical summation relationship between biomass of different algal groups and total algal biomass; The specific process of step S4 is as follows: S41. The time series sample dataset, which has been preprocessed and standardized in step S1, is divided into training set, validation set and test set in 7:2:1 ratio according to time order, and historical input sequence samples and their corresponding future output target vectors are constructed based on sliding window. S42. The hyperparameters of the joint model consisting of the LSTM-KAN base prediction model and the end-to-end hierarchical coordination module are optimized using a grid search method. The optimal hyperparameter combination is determined based on minimizing the validation set loss. The calculation formula is as follows: In the formula, The optimal combination of hyperparameters; The combination of hyperparameters to be searched; For the hyperparameter search space; This represents the validation set loss under the corresponding hyperparameter combination; In hyperparameter space Searching for the objective function Minimize the set of hyperparameters ; S43. The joint model is trained end-to-end according to the optimal hyperparameter combination. Under the hierarchical consistency constraint, the optimization objective is to minimize the error between the coordinated forecast results and the actual observations. The parameters of the LSTM-KAN basic forecast model and the dynamic weight generation network are updated synchronously through gradient backpropagation. The trained joint model is used for forecasting on the test set and outputs a full-level consistent forecast result that satisfies the hierarchical summation constraint of total algal biomass and biomass of different algal groups. S44. Evaluate model performance in terms of both forecast accuracy and hierarchical consistency: Forecast accuracy is evaluated using mean absolute error (MAE) and Nash efficiency coefficient (NSE). By calculating the MAE and NSE of the basic forecast and the coordinated forecast respectively, the changes in forecast accuracy before and after hierarchical coordination are compared. Hierarchical consistency is evaluated by comparing the summation difference between the total algal biomass forecasts before and after coordination and the biomass forecasts for different algal groups. The calculation formula is as follows: , In the formula, This represents the number of test samples; Number the test samples; For the first Basic forecast results of total algal biomass before coordination of individual test samples; For the first Before the coordination of individual test samples Basic forecast results of algal biomass; For the first Coordinated forecast results of total algal biomass after coordination of individual test samples; For the first After coordination of individual test samples Predicted values ​​for biomass coordination among algal groups; The summation difference between the base forecast results and the total algal biomass before harmonization and the biomass of different algal groups; Based on the basic forecast results and the summation difference between the total algal biomass and the biomass of different algal groups after harmonization, under the condition of satisfying hierarchical summation constraints, =0; S45. Output the consistent forecast results of total algal biomass and biomass of different algal groups at future time.

[0020] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting algal community structure consistency based on hierarchical deep learning, characterized in that, Includes the following steps: S1. Construct a raw time series dataset containing biomass of different algal groups, total algal biomass, water quality factors and hydro-meteorological factors, and perform data preprocessing and standardization. S2. Basic forecast results of biomass of different algal groups and total algal biomass were generated using the LSTM-KAN basic forecast model. S3. An end-to-end hierarchical coordination module based on dynamic weight learning is embedded in the LSTM-KAN basic forecast model to make consistent corrections to the basic forecast results of different algal groups and total algal biomass. S4. Perform parameter optimization, model training and performance evaluation, and output consistent prediction results that satisfy the hierarchical summation relationship between biomass of different algal groups and total algal biomass.

2. The algal community structure consistency prediction method based on hierarchical deep learning according to claim 1, wherein, The specific process of step S1 is as follows: S11. Collect biomass data of different algal groups, total algal biomass data, water quality factor data and hydro-meteorological factor data of the target water body during the continuous monitoring period, and construct the original time series dataset; S12. Preprocess the raw time series data, including time series alignment, outlier identification, missing value imputation, and feature selection. S13, set the total algal biomass as the top layer node, and set the different algal group biomasses as the bottom layer nodes, and the hierarchical addition relationship is expressed as: , wherein, is the total algal biomass at the tth time point; t is the time number; A is the identification of the top layer total algal group; c is the number of algal groups; is the total number of algal groups; is the biomass of the cth bottom layer algal group at the tth time point; is the biomass of the cth bottom layer algal group at the tth time point; is the biomass of the cth bottom layer algal group at the tth time point; is the identification of the cth bottom layer algal group; S14. Perform Z-score standardization on the preprocessed data. The calculation formula is as follows: In the formula, For the t-th time point The standardized values ​​of each variable; Number the variables; For the t-th time point The original values ​​of each variable; For the first The mean of each variable in the training set; For the first The standard deviation of each variable in the training set.

3. The algal community structure consistency prediction method based on hierarchical deep learning as described in claim 2, characterized in that, The specific process of step S2 is as follows: S21. Based on standardized time-series data, the biomass of different algal groups, total algal biomass, water quality factors, and hydrological and meteorological factors within historical periods are used as model inputs, and the biomass of different algal groups and total algal biomass at future times are used as model outputs. The basic forecast results are as follows: In the formula, h is the prediction step size; These are model parameters; This is the historical input sequence corresponding to the t-th time point; This is the LSTM-KAN basic forecast model; This provides the basic forecast results for the biomass of different algal groups and total algal biomass at the next time point t+h, i.e. In the formula, The basic forecast result of total algal biomass at the next time point t+h; For the time point t+h in the future Basic biomass forecast results for different algal groups; This is the transpose of the matrix; S22. The time-dependent features of algal biomass changes are extracted using the coding layer of an LSTM network to capture the nonlinear, time-delay, and seasonal characteristics of algal community structure changes. The state update process is represented as follows: , , , , , In the formula, Let be the standardized input vector at time point t; Use the Sigmoid activation function; Represents element-wise product; , , , For bias terms; , , , This is the weight matrix; Let be the hidden state of the LSTM network at time t. Let be the hidden state of the LSTM network at time t-1. Let be the candidate cell state at time point t; Let be the cell state at time point t; This represents the cell state at time point t-1. , , These are the forget gate, input gate, and output gate, respectively. S23. The hidden state output by the coding layer of the LSTM network. The input is a KAN network, which uses a learnable nonlinear function to characterize the mapping relationship between hidden features and target biomass. The output mapping process is as follows: In the formula, is the dimension of the hidden features in the LSTM network; The number of the hidden feature; For the first One hidden feature; For the connection in the KAN network The hidden feature and the first A learnable nonlinear function of output variables; For the future t+h time point The basic forecast results for each output variable, when When, corresponding to the total algal biomass; when At that time, respectively corresponding to Algal biomass; S24. Utilize the LSTM-KAN basic forecast model to output the basic forecast results of biomass for different algal groups and total algal biomass. .

4. The algal community structure consistency prediction method based on hierarchical deep learning as described in claim 3, characterized in that, The specific process of step S3 is as follows: S31. Based on the hierarchical summation relationship between the biomass of different algal groups and the total algal biomass, construct a hierarchical summation matrix for hierarchical coordination calculation: In the formula, It is a hierarchical summation matrix; Represents a set of real number elements OK, Column matrix space; The first row of the matrix represents the total number of algal groups at the bottom layer. The first row of the matrix is ​​used to represent the aggregation relationship between the total algal biomass and the biomass of each algal group. The remaining rows are used to retain the bottom layer node information corresponding to the biomass of different algal groups. S32. After the basic forecast results are denormalized to restore them to the original biomass scale, they are input into a dynamic weight generation network constructed by a multilayer perceptron and jointly optimized with the LSTM-KAN basic forecast model. The coordinated weights of the basic biomass forecast results at different levels are adaptively learned under the current forecast state, as shown in: , In the formula, For parameters Dynamic weight generation network; Softplus is a positive activation function; For the dynamic coordination weight vector at the (t+h)th future time point; The dynamic coordination weights corresponding to the basic forecast results of total algal biomass; For the time point t+h in the future Dynamic coordination weights corresponding to the basic forecast results of biomass of different algal groups; S33. Based on the dynamic coordination weights, construct the dynamic coordination weight matrix for the (t+h)th future time point. The calculation formula is as follows: In the formula, This is the dynamic coordination weight matrix for the future time point t+h, used to characterize the differences in coordination weights between different levels of basic forecast results under the current forecast state; These are positive constants used to ensure the numerical stability of matrix inversion and model training processes; S34, Based on hierarchical summation matrix Dynamic Coordination Weight Matrix and basic forecast results End-to-end hierarchical coordination calculations are performed to coordinate the basic forecast results of biomass for different algal groups and total algal biomass. The calculation formula is as follows: , , , In the formula, The results of the coordinated forecast of total algal biomass at the next time point t+h; For the time point t+h in the future Prediction results of biomass coordination among different algal groups; This is the coordinated biomass forecast vector for different algal groups at the bottom layer at the (t+h)th time point in the future. This is the coordinated, fully consistent forecast vector for the (t+h)th time point in the future. For hierarchical summation matrix The transpose of the matrix; For dynamically coordinated weight matrix The inverse matrix.

5. The algal community structure consistency prediction method based on hierarchical deep learning as described in claim 4, characterized in that, The specific process of step S4 is as follows: S41. The time series sample dataset, which has been preprocessed and standardized in step S1, is divided into training set, validation set and test set in 7:2:1 ratio according to time order, and historical input sequence samples and their corresponding future output target vectors are constructed based on sliding window. S42. The hyperparameters of the joint model consisting of the LSTM-KAN base prediction model and the end-to-end hierarchical coordination module are optimized using a grid search method. The optimal hyperparameter combination is determined based on minimizing the validation set loss. The calculation formula is as follows: In the formula, The optimal combination of hyperparameters; The combination of hyperparameters to be searched; For the hyperparameter search space; This represents the validation set loss under the corresponding hyperparameter combination; In hyperparameter space Searching for the objective function Minimize the set of hyperparameters ; S43. The joint model is trained end-to-end according to the optimal hyperparameter combination. Under the hierarchical consistency constraint, the optimization objective is to minimize the error between the coordinated forecast results and the actual observations. The parameters of the LSTM-KAN basic forecast model and the dynamic weight generation network are updated synchronously through gradient backpropagation. The trained joint model is used for forecasting on the test set and outputs a full-level consistent forecast result that satisfies the hierarchical summation constraint of total algal biomass and biomass of different algal groups. S44. Evaluate model performance in terms of both forecast accuracy and hierarchical consistency: Forecast accuracy is evaluated using mean absolute error (MAE) and Nash efficiency coefficient (NSE). By calculating the MAE and NSE of the basic forecast and the coordinated forecast respectively, the changes in forecast accuracy before and after hierarchical coordination are compared. Hierarchical consistency is evaluated by comparing the summation difference between the total algal biomass forecasts before and after coordination and the biomass forecasts for different algal groups. The calculation formula is as follows: , In the formula, This represents the number of test samples; Number the test samples; For the first Basic forecast results of total algal biomass before coordination of individual test samples; For the first Before the coordination of individual test samples Basic forecast results of algal biomass; For the first Coordinated forecast results of total algal biomass after coordination of individual test samples; For the first After coordination of individual test samples Predicted values ​​for biomass coordination among algal groups; The summation difference between the base forecast results and the total algal biomass before harmonization and the biomass of different algal groups; Based on the basic forecast results and the summation difference between the total algal biomass and the biomass of different algal groups after harmonization, under the condition of satisfying hierarchical summation constraints, =0; S45. Output the consistent forecast results of total algal biomass and biomass of different algal groups at future time.

6. The algal community structure consistency prediction method based on hierarchical deep learning as described in claim 5, characterized in that: In step S42, the hyperparameters of the LSTM-KAN base prediction model include the historical input window length, hidden layer dimension, number of network layers, learning rate, batch size, and regularization coefficient; the hyperparameters of the end-to-end hierarchical coordination module include the hidden layer dimension and number of network layers of the dynamic weight generation network.

Images