A Chlorophyll a Concentration Prediction Method Based on Signal Denoising and Hybrid Expert Model

Abstract

This invention discloses a method for predicting chlorophyll a concentration based on signal denoising and a hybrid expert model, comprising the following steps: S1, integrating historical meteorological data, water quality data, and chlorophyll a dataset of the target water body, and imputing missing values; S2, performing signal denoising on the imputed multi-source historical time-series data, filtering out random noise and short-term fluctuations, and extracting long-term trends and periodic components that reflect the intrinsic variation law of chlorophyll a concentration, to obtain denoised time-series data; S3, constructing multiple encoder models based on time-series blocks as expert models, used to collaboratively extract time-series dependency features at different levels from the denoised time-series data, and performing time-series prediction based on the extracted time-series dependency features; S4, training a hybrid expert model, and dynamically integrating the time-series prediction results of one or more expert models through a gating network layer to obtain the final predicted value of chlorophyll a concentration.

Description

Technical Field

[0001] This invention belongs to the field of water quality prediction and environmental protection technology, specifically involving a method for predicting chlorophyll a concentration based on signal denoising and hybrid expert models. Background Technology

[0002] Accurate prediction of chlorophyll a concentration in water bodies is crucial for eutrophication monitoring, algal bloom early warning, and water resource protection and management. Precise and advanced prediction of chlorophyll a concentration can provide key scientific evidence for implementing effective prevention and control measures, thereby ensuring drinking water safety and the health of aquatic ecosystems.

[0003] Currently, methods for predicting chlorophyll a concentration can be broadly categorized into two types: mechanism-driven models and data-driven models. Mechanism-driven models are built upon hydrodynamic and ecological principles, attempting to accurately describe the physical, chemical, and biological processes between algal growth and environmental factors such as light, water temperature, and nutrients. On the other hand, with the development of big data technology, data-driven models, especially machine learning models, have gradually become a research hotspot in this field. These models can automatically learn the complex nonlinear relationships between environmental factors and chlorophyll a concentration from historical monitoring data, without requiring complex prior mechanistic knowledge, and exhibit strong fitting capabilities and flexibility in various scenarios.

[0004] However, existing machine learning prediction methods still face significant challenges. First, water quality monitoring data is typically characterized by high noise and non-stationarity, stemming from sources such as sensor errors, environmental disturbances, and irregular human activities. This noise severely interferes with the model's learning of real-world patterns, leading to unstable prediction results. Second, the dynamic changes in chlorophyll a are driven by multiple factors, exhibiting complex intrinsic patterns that may change dynamically with seasons and weather conditions. A single prediction model often struggles to maintain optimal performance across all scenarios, exhibiting insufficient generalization ability. Specifically, traditional single models are easily affected by data noise when facing complex and variable aquatic environments, or are unable to simultaneously capture changes at different time scales, resulting in a bottleneck in prediction accuracy.

[0005] Therefore, effectively extracting stable temporal features from water quality data with high noise and non-stationarity, and constructing prediction models that can adapt to different data patterns, has become a key challenge in improving the accuracy and reliability of chlorophyll a concentration prediction. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a chlorophyll a concentration prediction method based on signal denoising and a hybrid expert model. This method is stable, reliable, and highly adaptable, effectively overcoming interference from data noise and pattern variations, significantly improving prediction accuracy, and possessing promising prospects for widespread application.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for predicting chlorophyll a concentration based on signal denoising and a hybrid expert model includes the following steps:

[0009] S1. Integrate historical meteorological data, water quality data, and chlorophyll a dataset of the target water body, and impute missing values;

[0010] S2. Perform signal denoising on the interpolated multi-source historical time series data to filter out random noise and short-term fluctuations, and extract long-term trend and periodic components that can reflect the intrinsic variation law of chlorophyll a concentration to obtain denoised time series data.

[0011] S3. Construct multiple encoder models based on time-series blocks as expert models to collaboratively extract time-series dependency features at different levels from the denoised time-series data, and perform time series prediction based on the extracted time-series dependency features.

[0012] S4. Train a hybrid expert model and dynamically integrate the time series prediction results of one or more expert models through a gated network layer to obtain the final predicted value of chlorophyll a concentration.

[0013] Preferably, in step S1, the meteorological data includes air temperature, rainfall, and wind speed; the water quality data includes pH value, dissolved oxygen, turbidity, total nitrogen concentration, and total phosphorus concentration; and the chlorophyll a dataset refers to historical chlorophyll a concentration monitoring data of the target water body.

[0014] Preferably, in step S1, the missing value imputation adopts different strategies according to the number of consecutive missing values: for data segments with a number of consecutive missing values ​​less than or equal to a set value, linear interpolation is used for filling; for data segments with a number of consecutive missing values ​​greater than the set value, they are retained as missing and no imputation processing is performed.

[0015] Preferably, in step S2, the signal denoising process is wavelet transform denoising, and the specific process includes:

[0016] S21. Select the wavelet basis function and the number of decomposition levels, and perform wavelet decomposition on each variable of the interpolated multi-source historical time series data to obtain high-frequency detail coefficients and low-frequency approximation coefficients.

[0017] S22. A soft thresholding function is used to filter noise in high-frequency detail coefficients. Wavelet coefficients with an absolute value less than or equal to the Sqtwolog threshold are considered noise and set to zero. Simultaneously, the amplitude of coefficients with an absolute value greater than the Sqtwolog threshold is reduced. The formula for calculating the soft thresholding function is: ,in, This is the output of the soft thresholding function; These are the high-frequency detail coefficients to be processed; For a sign function, when When the value is greater than 0, the value is 1. When < 0, the value is -1. When = 0, the value is 0; This is a function to find the maximum value. The Sqtwolog threshold is calculated using the following formula: ,in, The standard deviation of the high-frequency detail coefficients; The signal length;

[0018] S23. The high-frequency detail coefficients after thresholding and the unprocessed low-frequency approximation coefficients are used as inputs to perform wavelet inverse transform to reconstruct the complete time series data of each variable; then the complete time series data reconstructed from all variables are reintegrated to obtain a denoised time series dataset consistent with the structure of multi-source historical time series data.

[0019] Preferably, in step S3, the specific process of constructing multiple encoder models based on time-series blocks includes:

[0020] S31. Set the data segment length and moving step size to divide the denoised multivariate time series data into a data segment sequence. The expression is: ,in, for Time of the first The original input sequence of parameters; for Time of the first The sequence of data segments after being divided by parameters; To segment the data; The length of the data segment sequence; The length of the original input sequence;

[0021] S32. Linearly map the data segment sequence to an embedding vector, and add learnable positional encoding to preserve its temporal information. The expression is: ,in, The mapped embedding vector; It is a linear mapping matrix; For learnable location encoding;

[0022] S33. Construct a Transformer-based encoder layer to perform deep feature extraction on the embedded vector sequence, capture the temporal dependencies within and between variables, and map the features output by the encoder to the predicted values ​​through the projection layer.

[0023] S34. Set multiple sets of different data segment lengths and movement step sizes, and construct multiple encoder models based on time-series block division respectively.

[0024] Preferably, in step S4, the specific process of training the hybrid expert model is as follows:

[0025] S41. Construct a gated network layer, taking the denoised time-series data as input, and outputting the weights of the multiple time-block-based encoder models constructed in step S3 corresponding to the current input data. The expression is: ， in, These are the weight vectors for each expert model; This is the currently input denoised time-series data; It is a normalized exponential function; This represents the trainable weight matrix of the gated network; is the trainable bias vector of the gated network;

[0026] S42. Based on the weights generated by the gated network layer, the preliminary chlorophyll a concentration prediction results of each expert model are weighted and summed to obtain the chlorophyll a concentration prediction value of the hybrid expert model.

[0027] S43. The root mean square error between the predicted and actual values ​​is used as the loss function to iteratively optimize the parameters of the gated network layer. The calculation formula is as follows: ,in, The root mean square error between the predicted and the true values; The number of samples; For the first Observations of a sample; For the hybrid expert model to the first The predicted value for each sample.

[0028] After adopting the above technical solution, the present invention has the following beneficial effects: The chlorophyll a concentration prediction method based on signal denoising and hybrid expert models provided by the present invention effectively filters noise interference in water quality data through wavelet denoising technology, extracts stable temporal features from the water quality data, and solves the prediction instability problem caused by data quality issues in traditional models. Simultaneously, by constructing multiple temporal block encoder models with different parameter settings and combining them with the dynamic integration mechanism of the hybrid expert model, it can adaptively capture temporal dependency features under different data modes, overcoming the deficiency of insufficient generalization ability of a single model. This method significantly improves the accuracy and robustness of chlorophyll a concentration prediction, providing an innovative and efficient solution for eutrophication monitoring, algal bloom early warning, and water resource protection management in complex water environments, and has broad prospects for widespread application. Attached Figure Description

[0029] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0030] 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.

[0031] like Figure 1 As shown, the chlorophyll a concentration prediction method based on signal denoising and a hybrid expert model is characterized by the following steps:

[0032] S1. Integrate historical meteorological data, water quality data, and chlorophyll a dataset of the target water body, and impute missing values;

[0033] In step S1, the meteorological data includes air temperature, rainfall, and wind speed; the water quality data includes pH value, dissolved oxygen, turbidity, total nitrogen concentration, and total phosphorus concentration; the chlorophyll a dataset refers to historical chlorophyll a concentration monitoring data of the target water body.

[0034] In step S1, the missing value imputation adopts different strategies according to the number of consecutive missing values: for data segments with a number of consecutive missing values ​​less than or equal to a set value, linear interpolation is used for filling; for data segments with a number of consecutive missing values ​​greater than the set value, they are left as missing without imputation.

[0035] S2. Perform signal denoising on the interpolated multi-source historical time series data to filter out random noise and short-term fluctuations, and extract long-term trend and periodic components that can reflect the intrinsic variation law of chlorophyll a concentration to obtain denoised time series data.

[0036] In step S2, the signal denoising process is wavelet transform denoising, and the specific process includes:

[0037] S21. Select the wavelet basis function and the number of decomposition levels, and perform wavelet decomposition on each variable of the interpolated multi-source historical time series data to obtain high-frequency detail coefficients and low-frequency approximation coefficients.

[0038] S22. A soft thresholding function is used to filter noise in high-frequency detail coefficients. Wavelet coefficients with an absolute value less than or equal to the Sqtwolog threshold are considered noise and set to zero. Simultaneously, the amplitude of coefficients with an absolute value greater than the Sqtwolog threshold is reduced. The formula for calculating the soft thresholding function is: ,in, This is the output of the soft thresholding function; These are the high-frequency detail coefficients to be processed; For a sign function, when When the value is greater than 0, the value is 1. When < 0, the value is -1. When = 0, the value is 0; This is a function to find the maximum value. The Sqtwolog threshold is calculated using the following formula: ,in, The standard deviation of the high-frequency detail coefficients; The signal length;

[0039] S23. The high-frequency detail coefficients after thresholding and the unprocessed low-frequency approximation coefficients are used as inputs to perform inverse wavelet transform to reconstruct the complete time series data of each variable; then the complete time series data reconstructed from all variables are reintegrated to obtain a denoised time series dataset that is consistent with the structure of multi-source historical time series data.

[0040] S3. Construct multiple encoder models based on time-series blocks as expert models to collaboratively extract time-series dependency features at different levels from the denoised time-series data, and perform time series prediction based on the extracted time-series dependency features.

[0041] In step S3, the specific process of constructing multiple encoder models based on time-series blocks includes:

[0042] S31. Set the data segment length and moving step size to divide the denoised multivariate time series data into a data segment sequence. The expression is: ,in, for Time of the first The original input sequence of parameters; for Time of the first The sequence of data segments after being divided by parameters; To segment the data; The length of the data segment sequence; The length of the original input sequence;

[0043] S32. Linearly map the data segment sequence to an embedding vector, and add learnable positional encoding to preserve its temporal information. The expression is: ,in, The mapped embedding vector; It is a linear mapping matrix; For learnable location encoding;

[0044] S33. Construct a Transformer-based encoder layer to perform deep feature extraction on the embedded vector sequence, capture the temporal dependencies within and between variables, and map the features output by the encoder to the predicted values ​​through the projection layer.

[0045] S34. Set multiple sets of different data segment lengths and movement step sizes, and construct multiple encoder models based on time-series block division respectively;

[0046] S4. Train a hybrid expert model and dynamically integrate the time series prediction results of one or more expert models through a gated network layer to obtain the final predicted value of chlorophyll a concentration.

[0047] In step S4, the specific process of training the hybrid expert model is as follows:

[0048] S41. Construct a gated network layer, taking the denoised time-series data as input, and outputting the weights of the multiple time-block-based encoder models constructed in step S3 corresponding to the current input data. The expression is: ， in, These are the weight vectors for each expert model; This is the currently input denoised time-series data; It is a normalized exponential function; This represents the trainable weight matrix of the gated network; is the trainable bias vector of the gated network;

[0049] S42. Based on the weights generated by the gated network layer, the preliminary chlorophyll a concentration prediction results of each expert model are weighted and summed to obtain the chlorophyll a concentration prediction value of the hybrid expert model.

[0050] S43. The root mean square error between the predicted and actual values ​​is used as the loss function to iteratively optimize the parameters of the gated network layer. The calculation formula is as follows: ,in, The root mean square error between the predicted and the true values; The number of samples; For the first Observations of a sample; For the hybrid expert model to the first The predicted value for each sample.

[0051] 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 chlorophyll a concentration based on signal denoising and a hybrid expert model, characterized in that, Includes the following steps: S1. Integrate historical meteorological data, water quality data, and chlorophyll a dataset of the target water body, and impute missing values; S2. Perform signal denoising on the interpolated multi-source historical time series data to filter out random noise and short-term fluctuations, and extract long-term trend and periodic components that can reflect the intrinsic variation law of chlorophyll a concentration to obtain denoised time series data. In step S2, the signal denoising process is wavelet transform denoising, and the specific process includes: S21. Select the wavelet basis function and the number of decomposition levels, and perform wavelet decomposition on each variable of the interpolated multi-source historical time series data to obtain high-frequency detail coefficients and low-frequency approximation coefficients. S22. A soft thresholding function is used to filter noise in high-frequency detail coefficients. Wavelet coefficients with an absolute value less than or equal to the Sqtwolog threshold are considered noise and set to zero. Simultaneously, the amplitude of coefficients with an absolute value greater than the Sqtwolog threshold is reduced. The formula for calculating the soft thresholding function is: ,in, This is the output of the soft thresholding function; These are the high-frequency detail coefficients to be processed; For a sign function, when When the value is greater than 0, the value is 1. When < 0, the value is -1. When = 0, the value is 0; This is a function to find the maximum value. The Sqtwolog threshold is calculated using the following formula: ,in, The standard deviation of the high-frequency detail coefficients; The signal length; S23. The high-frequency detail coefficients after thresholding and the unprocessed low-frequency approximation coefficients are used as inputs to perform inverse wavelet transform to reconstruct the complete time series data of each variable; then the complete time series data reconstructed from all variables are reintegrated to obtain a denoised time series dataset that is consistent with the structure of multi-source historical time series data. S3. Construct multiple encoder models based on time-series blocks as expert models to collaboratively extract time-series dependency features at different levels from the denoised time-series data, and perform time series prediction based on the extracted time-series dependency features. S4. Train a hybrid expert model and dynamically integrate the time series prediction results of one or more expert models through a gated network layer to obtain the final predicted value of chlorophyll a concentration.

2. The chlorophyll a concentration prediction method based on signal denoising and hybrid expert models as described in claim 1, characterized in that: In step S1, the meteorological data includes air temperature, rainfall, and wind speed; the water quality data includes pH value, dissolved oxygen, turbidity, total nitrogen concentration, and total phosphorus concentration; and the chlorophyll a dataset refers to historical chlorophyll a concentration monitoring data of the target water body.

3. The chlorophyll a concentration prediction method based on signal denoising and hybrid expert models as described in claim 1, characterized in that, In step S1, the missing value imputation adopts different strategies according to the number of consecutive missing values: for data segments with a number of consecutive missing values ​​less than or equal to a set value, linear interpolation is used for filling; for data segments with a number of consecutive missing values ​​greater than the set value, they are left as missing without imputation processing.

4. The chlorophyll a concentration prediction method based on signal denoising and hybrid expert models as described in claim 1, characterized in that, In step S3, the specific process of constructing multiple encoder models based on time-series blocks includes: S31. Set the data segment length and moving step size to divide the denoised multivariate time series data into a data segment sequence. The expression is: ,in, for Time of the first The original input sequence of parameters; for Time of the first The sequence of data segments after being divided by parameters; To segment the data; The length of the data segment sequence; The length of the original input sequence; S32. Linearly map the data segment sequence to an embedding vector, and add learnable positional encoding to preserve its temporal information. The expression is: ,in, The mapped embedding vector; It is a linear mapping matrix; For learnable location encoding; S33. Construct a Transformer-based encoder layer to perform deep feature extraction on the embedded vector sequence, capture the temporal dependencies within and between variables, and map the features output by the encoder to the predicted values ​​through the projection layer. S34. Set multiple sets of different data segment lengths and movement step sizes, and construct multiple encoder models based on time-series block division respectively.

5. The chlorophyll a concentration prediction method based on signal denoising and hybrid expert models as described in claim 1, characterized in that, In step S4, the specific process of training the hybrid expert model is as follows: S41. Construct a gated network layer, taking the denoised time-series data as input, and outputting the weights of the multiple time-block-based encoder models constructed in step S3 corresponding to the current input data. The expression is: ， in, These are the weight vectors for each expert model; This is the currently input denoised time-series data; It is a normalized exponential function; This represents the trainable weight matrix of the gated network; is the trainable bias vector of the gated network; S42. Based on the weights generated by the gated network layer, the preliminary chlorophyll a concentration prediction results of each expert model are weighted and summed to obtain the chlorophyll a concentration prediction value of the hybrid expert model. S43. The root mean square error between the predicted and actual values ​​is used as the loss function to iteratively optimize the parameters of the gated network layer. The calculation formula is as follows: ,in, The root mean square error between the predicted and the true values; The number of samples; For the first Observations of a sample; For the hybrid expert model to the first The predicted value for each sample.

Images