Water quality prediction system based on VMD-DBO

Abstract

The invention discloses a VMD-DBO-based water quality prediction system, which relates to the technical field of water quality prediction and comprises a water quality data acquisition unit, a feature selection unit, a network parameter setting unit and a model prediction unit. By constructing the sample data and decomposing the time sequence of the original sewage quality indexes, the actual effluent chemical oxygen demand, effluent total nitrogen and effluent total phosphorus corresponding to the water quality index data are obtained, the instability of the data can be reduced, the prediction effect of the model is enhanced, and the prediction accuracy of the model is improved. Meanwhile, all water quality indexes are associated with a time sequence, so that the prediction effect of a subsequent model can be more accurate, Transform and an LSTM neural network model are fused, the problem of long and short term dependence in the sequence can be better solved, and meanwhile, the method has the advantages of improving the parallel processing capability and improving the model robustness, and is suitable for large-scale popularization and application. And the system optimizes network parameters, so that the model has good adaptability and precision.

Description

Technical Field

[0001] The present invention relates to the technical field of water quality prediction, and specifically to a water quality prediction system based on VMD-DBO. Background Art

[0002] The water quality prediction system uses network models and statistical analysis to monitor and predict the time series data of the sewage effluent water quality in real time, providing a scientific basis for water resource management and pollution prevention and control. In the invention patent with the application number 202210108885.1, "A water quality prediction method based on parameter migration" is disclosed. According to the water quality information of the target monitoring station in the water area and the water quality information of multiple adjacent monitoring stations, a water quality prediction model of the target monitoring station based on RVFL and multiple water quality prediction models of adjacent monitoring stations based on RVFL are respectively established, and the models are trained using the water quality training set of the target monitoring station and the water quality training set of the adjacent monitoring stations; according to the model parameters of the trained target monitoring station and adjacent monitoring stations, parameter migration is carried out between the models to obtain the migrated water quality prediction model of the target monitoring station; the migrated water quality prediction model of the target monitoring station is used for water quality prediction, and weighted averaging is carried out to obtain the final water quality prediction result. This method applies parameter migration to water quality prediction, migrates the shared parameters in the water quality prediction models of the target monitoring station and adjacent monitoring stations, effectively utilizes the non-linear correlation between the water quality information of the target monitoring station and adjacent monitoring stations, and improves the water quality prediction accuracy.

[0003] The above-mentioned prior art solves problems such as the inability to effectively utilize the non-linear correlation between the water quality information of the target monitoring station and adjacent monitoring stations. However, during operation, it is often difficult to accurately capture the subtle details of water quality changes, unable to effectively analyze the non-linear structure of large-scale water quality time series, and the prediction ability of a single model is very limited, making it difficult to provide relatively accurate prediction results. At the same time, the network parameters of the model are not optimized, resulting in the model not having good adaptability during operation. Summary of the Invention

[0004] The purpose of the present invention is to provide a water quality prediction system based on VMD-DBO to solve the problems raised in the above background art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A water quality prediction system based on VMD-DBO, including a model prediction unit; A water quality data acquisition unit, which acquires the water quality index data of the sewage treatment plant at different time periods, constructs multiple sample data according to the water quality index data at different time periods, and stores them in the original data set; A feature selection unit, which takes different water quality indicators as the initial feature vectors, analyzes them using a correlation determination algorithm to obtain the correlation coefficients of the feature vectors, selects the final feature vectors according to the correlation coefficients, sets the number of modes to be decomposed, the initial mode components and the center frequencies, constructs an unconstrained variational model, statistically analyzes the frequency domain representation corresponding to the initial center frequencies, analyzes the initial center frequencies, the frequency domain representations and the initial mode components in different modes to obtain new center frequencies and mode components, transmits them to the unconstrained variational model, and repeats the operation until the optimal mode components and center frequencies are determined, and analyzes the chemical oxygen demand, total nitrogen and total phosphorus in the effluent under different modes according to the output current mode components and center frequencies; A network parameter setting unit, which determines the network parameters in the current neural network model, constructs the corresponding dung beetle populations according to each network parameter, initializes the population size, the maximum number of iterations and the upper and lower bounds of the search space, calculates the fitness value according to the difference between the actual value of the sample data and the predicted value output by the current neural network model after obtaining the actual value and the predicted value, iteratively updates the positions of the corresponding dung beetles to obtain the best fitness value and the best position at the current stage, and determines the optimal network parameters according to the dung beetle individuals with the best fitness value and the best position in different populations.

[0006] Preferably, the water quality data acquisition unit includes a water quality acquisition module and a sample generation module. The water quality acquisition module acquires the water quality indicator data of the sewage treatment plant at different time periods, where the water quality data includes the influent flow rate, influent chemical oxygen demand, influent total nitrogen, influent total phosphorus, influent ammonia nitrogen, suspended solids and pH value. The sample generation module constructs a plurality of sample data using the water quality indicator data at different time periods, stores them in the original data set, divides the original data set into a training set, a validation set and a test set according to 7:1:2, and processes each sample data in the training set, the validation set and the test set using a sample analysis algorithm. The sample analysis algorithm is specifically: where, represents the minimum value of the variable in the sample, represents the maximum value of the variable in the sample; is the processed sample data.

[0007] Preferably, the feature selection unit includes a correlation coefficient calculation module and a feature vector determination module. The correlation coefficient calculation module uses the influent flow rate, influent chemical oxygen demand, influent total nitrogen, influent total phosphorus, influent ammonia nitrogen, suspended solids, and pH value as the initial feature vectors, and analyzes the initial feature vectors using a correlation determination algorithm to obtain the correlation coefficients of different initial feature vectors. The feature vector determination module selects the initial feature vectors according to the correlation coefficients, and uses the influent ammonia nitrogen, influent total nitrogen, influent total phosphorus, influent chemical oxygen demand, and pH value as the final feature vectors. The specific correlation determination algorithm is as follows: Wherein, represents the difference in rank between two feature vectors and , represents the length of each feature vector, represents the correlation coefficient between two feature vectors and , and represent the feature vector numbers, ranges from . When represents complete negative correlation, while represents complete positive correlation, 0 indicates no correlation between feature vectors.

[0008] Preferably, the feature selection unit further includes a model construction module and an initial parameter analysis module. After the model construction module extracts the sample data and recording time from the original dataset according to the feature vectors, it sets the number of modes to be decomposed, the corresponding initial mode components in different modes, and the initial center frequency , introduces a penalty term coefficient and a Lagrange multiplier, thereby constructing an unconstrained variational model. The initial parameter analysis module calculates the frequency domain representation corresponding to the initial center frequency , and then analyzes the initial center frequency , and the corresponding initial mode components in different modes to obtain a new center frequency , where . Combining and the initial mode components to obtain a new mode component , where , and represent parameters.

[0009] Preferably, the feature selection unit further includes a parameter output module and an index extraction module. The parameter output module transmits the new modal component and the central frequency to the unconstrained variational model to determine whether the convergence condition is satisfied. If it is satisfied, the current modal component and the central frequency are output. If not, the new modal component and the central frequency are recalculated. The index extraction module analyzes the chemical oxygen demand, total nitrogen, and total phosphorus in the effluent under different modes based on the output current modal component and central frequency.

[0010] Preferably, the network parameter setting unit includes a population construction module and a parameter determination module. The population construction module determines the network parameters in the current Transformer-LSTM neural network model, where the network parameters include the heads and keys of the attention mechanism, the number of neurons, the Dropout rate, and the learning rate. A corresponding dung beetle population is constructed according to each network parameter. The parameter determination module classifies the dung beetle population into rolling behavior, reproduction behavior, foraging behavior, and stealing behavior, and initializes the population size, the maximum number of iterations, the upper and lower bounds of the search space, the dimension of the solution vector, and the percentage of the producer population size in the total population size.

[0011] Preferably, the network parameter setting unit further includes a fitness calculation module and a network parameter analysis module. After the fitness calculation module obtains the actual chemical oxygen demand, total nitrogen, and total phosphorus in the effluent corresponding to the sample data and the predicted values output by the current neural network model, it calculates the initial fitness value according to the difference between the actual value and the predicted value. The network parameter analysis module updates the corresponding dung beetle positions for different individuals in the population to obtain the best fitness value and the best position at the current stage. Until the iteration ends, the dung beetle individuals with the best fitness value and the best position are output, and the optimal network parameters are determined according to the dung beetle individuals with the best fitness value and the best position in different populations.

[0012] Preferably, the model prediction unit includes a parameter setting module and a predicted value determination module. The parameter setting module transmits the optimal network parameters to the Transformer-LSTM neural network model, where the number of heads of the attention mechanism is 4, the key is 8, the time step is 3, the number of neurons in the LSTM hidden layer is 16, the number of LSTM stacking layers is 2, the minimum batch size is 328, and the maximum number of training times is 100. The predicted value determination module transmits the sample data in the training set to the neural network model for analysis, thereby completing the training of the network model. After determining the predicted values of ammonia nitrogen, total phosphorus, and total nitrogen in the effluent according to the sample data in the validation set and the test set, the error of the prediction result is calculated, and four evaluation indexes, namely the mean absolute percentage error, the root mean square error, the mean absolute error, and the coefficient of determination, are selected to evaluate the prediction result.

[0013] Compared with the prior art, the beneficial effects of the present invention are: The present invention constructs sample data and decomposes the time series of the original sewage water quality indicators, thereby obtaining the actual chemical oxygen demand, total nitrogen, and total phosphorus in the effluent corresponding to the water quality indicator data. This method can reduce the instability of the data, enhance the prediction effect of the model, and at the same time associate all water quality indicators with the time series, making the prediction effect of the subsequent model more accurate. The Transformer and LSTM neural network models are fused. Since the LSTM network model is good at capturing long-term dependence relationships and is suitable for dealing with long-distance dependence problems in time series data, while the Transformer can consider all positions in the sequence and helps capture global dependence relationships. Combining the two can better handle the long-term and short-term dependence problems in the sequence, and at the same time has the advantages of improving parallel processing ability and model robustness. And the system optimizes the network parameters, making the model have good adaptability and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 FIG. is a schematic diagram of the overall system flow provided by an embodiment of the present invention; Figure 2 FIG. is an internal module block diagram of the water quality data acquisition unit provided by an embodiment of the present invention; Figure 3 FIG. is an internal module block diagram of the network parameter setting unit provided by an embodiment of the present invention; Figure 4 FIG. is a sewage effluent prediction framework based on Transformer-LSTM provided by an embodiment of the present invention.

[0015] In the figure: 1. Water quality data acquisition unit; 101. Water quality acquisition module; 102. Sample generation module; 2. Feature selection unit; 201. Correlation coefficient calculation module; 202. Feature vector determination module; 203. Model construction module; 204. Initial parameter analysis module; 205. Parameter output module; 206. Index extraction module; 3. Network parameter setting unit; 301. Population construction module; 302. Parameter determination module; 303. Fitness calculation module; 304. Network parameter analysis module; 4. Model prediction unit; 401. Parameter setting module; 402. Predicted value determination module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0017] Please refer toFigure 1 - Figure 4 , the present invention provides a technical solution: a water quality prediction system based on VMD-DBO, including a model prediction unit 4; A water quality data acquisition unit 1, the water quality data acquisition unit 1 obtains water quality index data of a sewage treatment plant at different time periods, constructs a plurality of sample data according to the water quality index data at different time periods, and stores them in the original data set; A feature selection unit 2, the feature selection unit 2 uses different water quality indicators as initial feature vectors, analyzes them using a correlation determination algorithm to obtain the correlation coefficients of the feature vectors, screens out the final feature vectors according to the correlation coefficients, sets the number of modes to be decomposed, initial mode components and center frequencies, and constructs an unconstrained variational model. After statistically analyzing the frequency domain representation corresponding to the initial center frequency, analyzes the initial center frequency, frequency domain representation and initial mode components under different modes to obtain new center frequencies and mode components, and transmits them to the unconstrained variational model, and repeats the operation until the best mode components and center frequencies are determined, and analyzes the effluent chemical oxygen demand, effluent total nitrogen and effluent total phosphorus under different modes according to the output current mode components and center frequencies; A network parameter setting unit 3, the network parameter setting unit 3 determines the network parameters in the current neural network model, constructs corresponding dung beetle populations according to each network parameter, initializes the population size, maximum number of iterations and upper and lower bounds of the search space, obtains the actual values of the sample data and the predicted values output by the current neural network model, calculates the fitness value according to the difference between the actual value and the predicted value, iteratively updates the corresponding dung beetle positions, obtains the best fitness value and best position at the current stage, and determines the optimal network parameters according to the dung beetle individuals with the best fitness values and best positions in different populations.

[0018] The water quality data acquisition unit 1 includes a water quality acquisition module 101 and a sample generation module 102. The water quality acquisition module 101 obtains water quality index data of a sewage treatment plant at different time periods, where the water quality data includes influent flow rate, influent chemical oxygen demand, influent total nitrogen, influent total phosphorus, influent ammonia nitrogen, suspended solids and pH value. The sample generation module 102 constructs a plurality of sample data using the water quality index data at different time periods, stores them in the original data set, divides the original data set into a training set, a validation set and a test set according to 7:1:2, and processes each sample data in the training set, validation set and test set using a sample analysis algorithm. The sample analysis algorithm is specifically: Among them, represents the minimum value of the variable in the sample, represents the maximum value of the variable in the sample; is the processed sample data; The feature selection unit 2 includes a correlation coefficient calculation module 201 and a feature vector determination module 202. The correlation coefficient calculation module 201 takes the influent flow rate, influent chemical oxygen demand, influent total nitrogen, influent total phosphorus, influent ammonia nitrogen, suspended solids, and pH value as the initial feature vectors, and uses a correlation determination algorithm to analyze the initial feature vectors to obtain the correlation coefficients of different initial feature vectors. The feature vector determination module 202 selects the initial feature vectors according to the correlation coefficients, and takes the influent ammonia nitrogen, influent total nitrogen, influent total phosphorus, influent chemical oxygen demand, and pH value as the final feature vectors. The specific correlation determination algorithm is as follows: Wherein, represents the difference in rank between two feature vectors and is the rank difference, represents the length of each feature vector, represents two feature vectors and is the correlation coefficient between them, and represent the feature vector numbers, ranges from When represents a perfect negative correlation, while represents a perfect positive correlation, 0 indicates no correlation between the feature vectors; The feature selection unit 2 further includes a model construction module 203 and an initial parameter analysis module 204. After the model construction module 203 extracts the sample data and recording time from the original dataset according to the feature vectors, it sets the number of modes to be decomposed, the corresponding initial mode components in different modes, and the initial center frequencies , introduces a penalty term coefficient and a Lagrange multiplier, thereby constructing an unconstrained variational model. The initial parameter analysis module 204 statistically analyzes the frequency domain representation corresponding to the initial center frequency , and then analyzes the initial center frequency , and the corresponding initial mode components in different modes to obtain a new center frequency , where , combines and the initial mode components to obtain a new mode component , where , and represent parameters; The feature selection unit 2 further includes a parameter output module 205 and an index extraction module 206. The parameter output module 205 transmits the new modal component and the central frequency to the unconstrained variational model to determine whether the convergence condition is satisfied. If it is satisfied, the current modal component and the central frequency are output. If not, the new modal component and the central frequency are recalculated. The index extraction module 206 analyzes the chemical oxygen demand, total nitrogen, and total phosphorus in the effluent under different modes according to the output current modal component and central frequency. The unconstrained variational model is specifically: where, represents the th modal component generated by data decomposition, represents the central frequency corresponding to the th modal component, represents the Lagrange multiplier, represents the penalty term coefficient, represents the charging function, represents the imaginary unit, represents the time unit, represents the convolution operation symbol, represents the time series, represents the set function, represents the time function, represents the natural constant, represents the unconstrained variational function, represents the total number of modes, represents the modal number; The network parameter setting unit 3 includes a population construction module 301 and a parameter determination module 302. The population construction module 301 determines the network parameters in the current Transformer-LSTM neural network model, where the network parameters include the heads and keys of the attention mechanism, the number of neurons, the Dropout rate, and the learning rate. A corresponding dung beetle population is constructed according to each network parameter. The parameter determination module 302 classifies the dung beetle population into rolling behavior, breeding behavior, foraging behavior, and stealing behavior, and initializes the population size, the maximum number of iterations, the upper and lower bounds of the search space, the dimension of the solution vector, and the percentage of the producer population size in the total population size; The network parameter setting unit 3 further includes a fitness calculation module 303 and a network parameter analysis module 304. After the fitness calculation module 303 obtains the actual chemical oxygen demand of the effluent, total nitrogen of the effluent, total phosphorus of the effluent corresponding to the sample data and the predicted values output by the current neural network model, it calculates the initial fitness value according to the difference between the actual value and the predicted value. The network parameter analysis module 304 updates the corresponding dung beetle positions of different individuals in the population to obtain the best fitness value and the best position at the current stage. Until the iteration ends, it outputs the dung beetle individuals with the best fitness value and the best position, and determines the optimal network parameters according to the dung beetle individuals with the best fitness value and the best position in different populations; The model prediction unit 4 includes a parameter setting module 401 and a predicted value determination module 402. The parameter setting module 401 transmits the optimal network parameters to the Transformer-LSTM neural network model, where the number of heads of the attention mechanism is 4, the key is 8, the time step is 3, the number of neurons in the LSTM hidden layer is 16, the number of LSTM stacking layers is 2, the minimum batch size is 328, and the maximum number of training times is 100. The predicted value determination module 402 transmits the sample data in the training set to the neural network model for analysis, thereby completing the training of the network model. After determining the predicted values of ammonia nitrogen, total phosphorus, and total nitrogen at the next moment according to the sample data in the validation set and the test set, it calculates the error of the prediction result, and selects four evaluation indicators, namely mean absolute percentage error, root mean square error, mean absolute error, and coefficient of determination, to evaluate the prediction result.

[0019] Working principle: In the present invention, the water quality acquisition module 101 in the water quality data acquisition unit 1 acquires water quality index data of a sewage treatment plant at different time periods. The sample generation module 102 constructs multiple sample data. The correlation coefficient calculation module 201 in the feature selection unit 2 determines the correlation coefficients of different initial feature vectors. The feature vector determination module 202 selects the initial feature vectors according to the correlation coefficients. The model construction module 203 initializes the modal components and central frequencies in different modes. The initial parameter analysis module 204 iteratively updates the modal components and central frequencies. The parameter output module 205 outputs the current modal components and central frequencies. The index extraction module 206 analyzes the chemical oxygen demand, total nitrogen, and total phosphorus in the effluent under different modes. The population construction module 301 in the network parameter setting unit 3 constructs corresponding dung beetle populations according to each network parameter. After the parameter determination module 302 initializes the population size and the maximum number of iterations, the fitness calculation module 303 calculates the initial fitness value according to the difference between the actual value and the predicted value of the sample data. The network parameter analysis module 304 determines the optimal network parameters by using the dung beetle individuals with the best fitness values and the best positions in different populations. The parameter setting module 401 in the model prediction unit 4 transmits the optimal network parameters to the neural network model. The predicted value determination module 402 selects four evaluation indexes, namely, mean absolute percentage error, root mean square error, mean absolute error, and coefficient of determination, to evaluate the prediction results.

[0020] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0021] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A water quality prediction system based on VMD-DBO, including a model prediction unit (4), characterized in that: A water quality data acquisition unit (1), the water quality data acquisition unit (1) obtains water quality index data of a sewage treatment plant at different time periods, constructs a plurality of sample data according to the water quality index data at different time periods, and stores them in the original data set; A feature selection unit (2), the feature selection unit (2) takes different water quality indicators as the initial feature vectors, analyzes them using a correlation determination algorithm to obtain the correlation coefficients of the feature vectors, screens out the final feature vectors according to the correlation coefficients, sets the number of modes to be decomposed, the initial mode components and the center frequency, and constructs an unconstrained variational model. After statistically analyzing the frequency domain representation corresponding to the initial center frequency, analyzes the initial center frequency, the frequency domain representation and the initial mode components under different modes to obtain a new center frequency and mode components, and transmits them to the unconstrained variational model, and repeats the operation until the best mode components and center frequency are determined, and analyzes the effluent chemical oxygen demand, effluent total nitrogen and effluent total phosphorus under different modes according to the output current mode components and center frequency; A network parameter setting unit (3), the network parameter setting unit (3) determines the network parameters in the current neural network model, constructs corresponding dung beetle populations according to each network parameter, initializes the population size, the maximum number of iterations and the upper and lower bounds of the search space. After obtaining the actual value of the sample data and the predicted value output by the current neural network model, calculates the fitness value according to the difference between the actual value and the predicted value, iteratively updates the corresponding dung beetle positions, obtains the best fitness value and the best position at the current stage, and determines the optimal network parameters according to the dung beetle individuals with the best fitness value and the best position in different populations.

2. The water quality prediction system based on VMD-DBO according to claim 1, wherein: The water quality data acquisition unit (1) includes a water quality acquisition module (101) and a sample generation module (102), the water quality acquisition module (101) obtains water quality index data of a sewage treatment plant at different time periods, where the water quality data includes influent flow rate, influent chemical oxygen demand, influent total nitrogen, influent total phosphorus, influent ammonia nitrogen, suspended solids and pH value, the sample generation module (102) constructs a plurality of sample data using the water quality index data at different time periods, stores them in the original data set, divides the original data set into a training set, a validation set and a test set according to 7:1:2, and processes each sample data in the training set, the validation set and the test set using a sample analysis algorithm.

3. The water quality prediction system based on VMD-DBO according to claim 1, characterized in that: The feature selection unit (2) includes a correlation coefficient calculation module (201) and a feature vector determination module (202). The correlation coefficient calculation module (201) takes the influent flow rate, influent chemical oxygen demand, influent total nitrogen, influent total phosphorus, influent ammonia nitrogen, suspended solids, and pH value as the initial feature vector, analyzes the initial feature vector using a correlation determination algorithm to obtain the correlation coefficients of different initial feature vectors. The feature vector determination module (202) selects the initial feature vector according to the correlation coefficient, and takes the influent ammonia nitrogen, influent total nitrogen, influent total phosphorus, influent chemical oxygen demand, and pH value as the final feature vector.

4. The water quality prediction system based on VMD-DBO according to claim 3, characterized in that: The feature selection unit (2) further includes a model construction module (203) and an initial parameter analysis module (204). After the model construction module (203) extracts the sample data and recording time in the original dataset according to the feature vector, it sets the number of modes to be decomposed , the corresponding initial modal components in different modes and the initial center frequency . By introducing the penalty term coefficient and Lagrange multiplier, an unconstrained variational model is constructed. The initial parameter analysis module (204) statistically analyzes the frequency domain representation corresponding to the initial center frequency , and then analyzes the initial center frequency , and the corresponding initial modal components in different modes to obtain a new center frequency , where . By combining and the initial modal components , a new modal component is obtained, where . and represent parameters.

5. The water quality prediction system based on VMD-DBO according to claim 4, characterized in that: The feature selection unit (2) further includes a parameter output module (205) and an index extraction module (206). The parameter output module (205) transmits the new modal component and the center frequency to the unconstrained variational model to determine whether the convergence condition is satisfied. If it is satisfied, the current modal component and the center frequency are output. If not, the new modal component and the center frequency are recalculated. The index extraction module (206) analyzes the effluent chemical oxygen demand, effluent total nitrogen, and effluent total phosphorus under different modes according to the output current modal component and center frequency.

6. The water quality prediction system based on VMD-DBO according to claim 1, characterized in that: The network parameter setting unit (3) includes a population construction module (301) and a parameter determination module (302). The population construction module (301) determines the network parameters in the current Transformer-LSTM neural network model. The network parameters include the number of heads and keys of the attention mechanism, the number of neurons, the Dropout rate, and the learning rate. A corresponding dung beetle population is constructed according to each network parameter. The parameter determination module (302) divides the dung beetle population into rolling ball behavior, reproduction behavior, foraging behavior, and stealing behavior, and initializes the population size, the maximum number of iterations, the upper and lower bounds of the search space, the dimension of the solution vector, and the percentage of the producer population size in the total population size.

7. The water quality prediction system based on VMD-DBO according to claim 6, characterized in that: The network parameter setting unit (3) further includes a fitness calculation module (303) and a network parameter analysis module (304). After the fitness calculation module (303) obtains the actual effluent chemical oxygen demand, effluent total nitrogen, effluent total phosphorus corresponding to the sample data and the predicted values output by the current neural network model, it calculates the initial fitness value according to the difference between the actual value and the predicted value. The network parameter analysis module (304) updates the corresponding dung beetle positions of different individuals in the population to obtain the best fitness value and the best position at the current stage. Until the iteration ends, the dung beetle individual with the best fitness value and the best position is output, and the optimal network parameters are determined according to the dung beetle individuals with the best fitness value and the best position in different populations.

8. The water quality prediction system based on VMD-DBO according to claim 1, characterized in that: The model prediction unit (4) includes a parameter setting module (401) and a predicted value determination module (402). The parameter setting module (401) transmits optimal network parameters to the Transformer-LSTM neural network model, where the number of heads of the attention mechanism is 4, the number of keys is 8, the number of time steps is 3, the number of neurons in the LSTM hidden layer is 16, the number of LSTM stacking layers is 2, the minimum batch size is 328, and the maximum number of training times is 100. The predicted value determination module (402) transmits the sample data in the training set to the neural network model for analysis, thereby completing the training of the network model. After determining the predicted values of ammonia nitrogen, total phosphorus, and total nitrogen at the time of water discharge according to the sample data in the validation set and the test set, an error calculation is performed on the prediction results, and four evaluation indicators, namely mean absolute percentage error, root mean square error, mean absolute error, and coefficient of determination, are selected to evaluate the prediction results.

Images