Sewage treatment effluent quality prediction method based on frequency domain adaptive division and two-channel network

Abstract

The invention discloses a sewage treatment effluent quality prediction method based on frequency domain adaptive division and a two-channel network, and belongs to the technical field of water quality prediction. The method comprises the following steps: firstly, carrying out global multi-scale decomposition on an effluent quality sequence by utilizing ICEEMDAN; secondly, introducing a sample entropy as a feature criterion, and adaptively reconstructing a modal component into a low-frequency trend sequence and a high-frequency non-stationary sequence by using a K-means algorithm; respectively carrying out differentiated screening on the influence of inlet water quality parameters for the high-frequency sequence and the low-frequency sequence by adopting a maximum mutual information coefficient; and finally, constructing a dual-channel parallel model: capturing a long-term evolution rule by adopting a CNN-BiLSTM combined network in a low-frequency channel, introducing an Autoformer network in a high-frequency channel, and deeply mining transient impact characteristics by utilizing an embedded sequence decomposition module and an autocorrelation mechanism. According to the method, decoupling and differential modeling of the multi-scale frequency domain features are effectively realized, and the prediction precision of the water quality peak value under the complex working condition is improved.

Description

Technical Field

[0001] This invention relates to the field of water quality prediction technology, specifically to a method for predicting the effluent quality of wastewater treatment based on frequency domain adaptive partitioning and a dual-channel network. Background Technology

[0002] Urban wastewater treatment is a critical line of defense for ensuring water environment safety. Achieving accurate and forward-looking prediction of effluent quality (such as key indicators like COD and ammonia nitrogen) not only provides a scientific and forward-looking basis for adjusting blower aeration volume and precisely adding carbon sources, but also serves as core technological support for reducing process costs and increasing efficiency, ensuring all-weather compliance with emission standards. This has significant practical implications for promoting the construction of smart water management.

[0003] However, wastewater treatment is essentially a complex biochemical process characterized by strong nonlinearity, large time-varying lag, and high noise. Although data-driven methods based on deep learning (such as LSTM and CNN) have been gradually applied to this field, they still have significant limitations when facing complex operating conditions: First, they lack frequency adaptability. Existing methods often adopt a "one-size-fits-all" homogeneous modeling approach for different frequency components after decomposition, making it difficult to balance the stability of low-frequency trends with the sensitivity to high-frequency fluctuations. Second, they have weak resistance to shock loads. Traditional point-to-point attention mechanisms are unable to capture high-frequency transient changes caused by industrial illegal discharges or storm runoff, often resulting in model prediction lag or distortion at water quality peaks, failing to meet the emergency early warning needs in practical engineering. Summary of the Invention

[0004] 1. The technical problem that the invention aims to solve To overcome the limitations of existing technologies, such as the inability of a single model to simultaneously predict stationary trends and transient impacts, and the simplistic signal reconstruction strategies of traditional decomposition methods, this invention proposes a wastewater treatment effluent quality prediction method based on frequency domain adaptive partitioning and a dual-channel network. This invention goes beyond simple decomposition of the original water quality time series; instead, it innovatively designs a frequency domain adaptive partitioning strategy based on sample entropy clustering to decouple non-stationary sequences into low-frequency trend components and high-frequency fluctuation components. Furthermore, it constructs CNN-BiLSTM and Autoformer networks for differentiated collaborative modeling based on different frequency characteristics. This method aims to efficiently separate effective impact features from random noise, solving the problem of lag in effluent peak prediction under conditions of sudden changes in influent load, thereby significantly improving the accuracy, robustness, and physical interpretability of urban wastewater treatment plant effluent quality prediction.

[0005] 2. Technical Solution To achieve the above objectives, the technical solution provided by the present invention is as follows: The present invention provides a wastewater treatment effluent quality prediction method based on frequency domain adaptive partitioning and a dual-channel network, comprising the following steps: Step S1: Obtain historical data of the wastewater treatment process, including influent water quality parameters and effluent water quality parameters, and preprocess the historical data to obtain a preprocessed water quality test sample dataset; Step S2: Use the ICEEMDAN algorithm to perform a global multi-scale decomposition on the preprocessed historical effluent water quality time series, and calculate the sample entropy values ​​of all intrinsic mode components (IMFs) obtained after the decomposition. Step S3: Using the sample entropy value as the feature input, the K-means clustering algorithm is used to adaptively divide all intrinsic mode components into two categories: high-frequency components and low-frequency components. The high-frequency components are superimposed and reconstructed into a high-frequency non-stationary sequence, and the low-frequency components and residual components are superimposed and reconstructed into a low-frequency trend sequence. Step S4: Using the maximum mutual information coefficient analysis method, calculate the correlation between each influent water quality parameter and the low-frequency trend sequence and the high-frequency non-stationary sequence, set the correlation threshold, screen out the strongly correlated influent parameters to form a low-frequency feature set and a high-frequency feature set, and divide the feature set into a training set and a test set according to the proportion. Step S5: Construct a frequency domain adaptive dual-channel prediction model to perform parallel predictions on the low-frequency trend sequence and the high-frequency non-stationary sequence, respectively. The low-frequency channel uses a CNN-BiLSTM combined network to output the predicted low-frequency components, and the high-frequency channel uses an Autoformer network to output the predicted high-frequency components. Finally, the predicted low-frequency components and the predicted high-frequency components are linearly superimposed at the corresponding time points, and the superposition result is inversely normalized to obtain the final predicted water quality of the wastewater treatment effluent.

[0006] 3. Beneficial effects Compared with existing known technologies, the technical solution provided by this invention has the following significant advantages: (1) This invention proposes a frequency domain adaptive partitioning strategy based on sample entropy, which effectively achieves precise decoupling of signal physical characteristics. Unlike traditional methods that blindly perform secondary decomposition on historical effluent water quality sequences, this invention uses entropy as a criterion to physically separate the long-term stable trend and transient shock fluctuations implicit in the original sequence. This design not only reduces the complexity of processing mixed modal data with a single model, but also avoids signal distortion that may be caused by over-decomposition, laying the foundation for high-precision differentiated modeling from the data source.

[0007] (2) The dual-channel differentiated collaborative modeling architecture constructed in this invention significantly solves the problem of peak prediction failure under inflow impact conditions. For the low-frequency channel, CNN-BiLSTM is used to ensure the stability of basic trend prediction; for the high-frequency channel, the Autoformer model is introduced, using its unique autocorrelation mechanism to replace the traditional point-to-point attention. This design can penetrate high-frequency noise and uncover the periodic dependence of subprocesses hidden under the chaotic waveform, so as to respond in a timely manner and accurately predict the peak water flow when the inflow load changes abruptly, significantly improving the model's anti-interference ability and robustness.

[0008] (3) This invention enhances the interpretability and generalization ability of the model through multi-source information fusion. In the Autoformer high-frequency prediction channel, not only historical effluent data is used, but also historical influent water quality parameters are used as covariate inputs. This multivariate modeling method effectively simulates the hydraulic retention time lag effect in the wastewater treatment process, and uses the prior mutation information at the influent end to correct the prediction trajectory at the effluent end, so that the model has excellent generalization performance under different seasons and different operating conditions. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the process for predicting the effluent quality of wastewater treatment based on frequency domain adaptive partitioning and dual-channel network according to the present invention. Detailed Implementation

[0010] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments.

[0011] Example 1 Combination Figure 1 This embodiment of a wastewater treatment effluent quality prediction method based on frequency domain adaptive partitioning and dual-channel network includes the following steps: Step S1: Obtain historical data of the wastewater treatment process, including influent water quality parameters and effluent water quality parameters. Preprocess the historical data to obtain a preprocessed water quality test sample dataset. The preprocessing process includes outlier detection, missing value filling and data normalization. In this embodiment, multi-source water quality data is collected by a sensor system deployed at the wastewater treatment site as historical data, and is divided into two categories according to the purpose of the data: The first category is model input feature parameters (covariates), which select influent and process indicators that have a priori indicative effect on effluent water quality, specifically including: wastewater influent flow rate, influent dissolved oxygen concentration, influent ammonia nitrogen concentration, influent total nitrogen concentration, and influent total phosphorus concentration; The second category is model prediction target parameters (target variables), which require the model to output historical and future effluent indicators, specifically including: effluent dissolved oxygen concentration, effluent ammonia nitrogen concentration, effluent total nitrogen concentration, and effluent total phosphorus concentration.

[0012] The outlier detection specifically involves identifying and deleting negative values, extremely large outliers, and non-numerical data from historical water quality monitoring data.

[0013] The missing value filling method involves using linear interpolation to fill in the missing values ​​generated by monitoring and the missing values ​​generated after deleting outliers, that is, using the average of the data from the previous time step and the next time step to replace the missing data at this time step.

[0014] The data normalization process described above specifically involves normalizing the historical water quality monitoring data after outlier detection and missing value imputation, using max-min normalization to scale all data to the range [0,1]. The normalization formula is as follows:

[0015] in, This is the i-th input data; The minimum value in the input data; The maximum value in the input data; n is the total number of data points; This is the normalized data.

[0016] Step S2: The improved Complete Ensemble Empirical Mode Decomposition (ICEEMDAN) algorithm is used to perform a global multi-scale decomposition on the preprocessed historical effluent water quality time series, and the sample entropy values ​​of all IMF components obtained after the first decomposition are calculated; specifically including: Step S2.1: Use the ICEEMDAN algorithm to perform a decomposition on the pretreated historical effluent water quality time series. Improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is a time-frequency analysis method with adaptive characteristics. This algorithm decomposes the effluent water quality time series into a set of IMF components with frequencies ranging from high to low, each IMF component reflecting the water quality data characteristics of a different frequency range.

[0017] Compared to traditional empirical mode decomposition and ensemble empirical mode decomposition, ICEEMDAN more effectively solves the mode aliasing problem by adding non-Gaussian white noise, further reducing the impact of residual noise.

[0018] The steps ICEEMDAN takes to decompose the historical effluent water quality time series are as follows: 1) Let x For historical outflow water quality time series, specific noise was included. Add tox middle:

[0019] In the formula: This is the effluent water quality time series after adding the i-th group of special noise; The expected signal-to-noise ratio for the first decomposition iteration; The k-th order modal components produced by empirical mode decomposition; This represents the i-th group of white noise used to aid in decomposition; 2) Calculate the residuals of the first decomposition. And using effluent water quality time series x Subtract residuals Obtain intrinsic mode components :

[0020]

[0021] In the formula: M(·) is the operator for generating the local mean of the signal; I is the number of white noise groups; 3) Continue adding white noise signal and calculate the second residual. Find the second eigenmode components of the historical wastewater treatment effluent quality time series. .

[0022]

[0023]

[0024] 4) Similarly to step 3), calculate the k-th residual and intrinsic mode components.

[0025]

[0026]

[0027] 5) Repeat step 4) until the maximum number of iterations is reached or the residual can no longer be decomposed.

[0028] Step S2.2: Calculate the sample entropy values ​​of all IMF components after one decomposition: Sample entropy (SE) is an algorithm for measuring the complexity of time series data. Compared to approximate entropy, sample entropy offers better consistency and its calculation is independent of data length. Sample entropy is used to measure the IMF complexity of all intrinsic mode components (IMFs) obtained from ICEEMDAN decomposition. Generally, a higher sample entropy value indicates higher sequence complexity, and vice versa.

[0029] The sample entropy values ​​of each IMF component obtained from ICEEMDAN decomposition are calculated according to the following formula:

[0030] In the formula: Samplen(·) is the function for calculating sample entropy (SE); N is the length of the modal component; , These represent the probabilities of the sequence matching m points and m+1 points under a tolerance of r, respectively. In practical scenarios, m is the dimension, typically taken as 2; r is the tolerance, usually taken as 10%~25% of the calculated standard deviation (SD). In this embodiment, r is taken as 20% of the SD.

[0031] Step S3: Using the sample entropy values ​​of each intrinsic mode component obtained in Step S2 as feature input, the K-means clustering algorithm is used to adaptively divide all intrinsic mode components into two categories: high-frequency components and low-frequency components. The components that belong to the high-frequency component category are superimposed to reconstruct a high-frequency non-stationary sequence, and the components that belong to the low-frequency component category are superimposed with the residual components obtained in Step S2 to reconstruct a low-frequency trend sequence. Specifically, this includes: Step S3.1: Use K The MESS algorithm clusters the IMF components based on the sample entropy values ​​obtained in step S2, dividing them into two categories: high-frequency components and low-frequency components. 1) Constructing the feature dataset: Extract the sample entropy values ​​of each intrinsic mode component obtained from the previous steps to form a one-dimensional feature set. , which serves as the feature input for the K-means clustering algorithm.

[0032] 2) Initialize cluster centers: Based on the requirement of dividing the signal into high-frequency components and low-frequency components, the number of clusters K is set to 2. Two sample entropy values ​​are randomly selected from the feature set E as the initial cluster centers.

[0033] 3) Distance Calculation and Classification: Compare the values ​​of two cluster centers. Since a larger sample entropy value indicates higher sequence complexity, the cluster center with the larger value is designated as the center of the "high-frequency component class," and the cluster center with the smaller value is designated as the center of the "low-frequency component class." Calculate the Euclidean distance between the entropy value of each sample in the feature set and these two cluster centers. Following the principle of minimizing distance, if the sample entropy value of a certain IMF component is closer to the center of the "high-frequency component class," it is marked as a high-frequency component; otherwise, it is marked as a low-frequency component.

[0034] 4) Update cluster centers: Calculate the arithmetic mean of the entropy values ​​of all samples in the current high-frequency component class and low-frequency component class respectively, and use these two averages as the new cluster centers.

[0035] 5) Iterative convergence: Repeat steps 3) and 4) until the change in cluster centers is less than the preset convergence threshold or the maximum number of iterations is reached, thus completing the final attribute division of all IMF components.

[0036] Step S3.2: Reconstruct the IMF components by overlaying them based on the clustering results. The specific reconstruction method is as follows: High-frequency non-stationary sequence reconstruction: All intrinsic mode components classified into high-frequency component classes are linearly added in the time domain to obtain a high-frequency non-stationary sequence, which is then used as the input to the Autoformer model; Low-frequency trend sequence reconstruction: All intrinsic mode components classified into low-frequency component classes and the residual components obtained from step S2 decomposition are linearly added in the time domain to obtain the low-frequency trend sequence, which is used as the input of the CNN-BiLSTM model.

[0037] Step S4: Using the maximum mutual information coefficient analysis method, calculate the influent water quality parameters (corresponding to feature set X) from step S1 and the low-frequency trend sequence reconstructed in step S3. High-frequency non-stationary sequences Correlation between them; set a correlation threshold for low-frequency trend sequences. Strongly correlated influent parameters were selected to form a low-frequency feature set; for high-frequency non-stationary sequences... Strongly correlated influent parameters are selected to form a high-frequency feature set, thereby achieving differentiated screening of influencing factors. Finally, the selected low-frequency feature set and high-frequency feature set are strictly divided into training set and test set in 8:2 ratio according to the time sequence.

[0038] To accurately identify the driving factors of different frequency components, this step introduces the Maximum Information Coefficient (MIC) for nonlinear correlation analysis. Compared to the Pearson coefficient, which only measures linear relationships, the MIC can effectively capture the time-varying, nonlinear, and complex functional dependencies between influent characteristics and effluent response. The specific process is as follows: S4.1: Calculate the correlation: Calculate the MIC values ​​between each influent water quality parameter in step S1 and the low-frequency trend series and high-frequency non-stationary series reconstructed in step S3, respectively. The formula for calculating MIC is defined as:

[0039] In the formula, For mutual information, , B represents the number of grid divisions, where B is the upper limit for the number of grids.

[0040] S4.2: Constructing the Feature Set Set a low-frequency correlation threshold to filter out For influent parameters exceeding the threshold, a high-frequency correlation threshold is set to filter out [parameters / items]. Inflow parameters exceeding a set threshold. The selected low-frequency feature set and high-frequency feature set are strictly divided into training and test sets in an 8:2 ratio according to chronological order, to construct standard input data samples for model training and validation.

[0041] Step S5: Construct a frequency domain adaptive dual-channel prediction model to perform parallel predictions on the low-frequency trend sequence and the high-frequency non-stationary sequence respectively: For the low-frequency trend sequence, input it and the low-frequency feature set obtained in step S4 into a CNN-BiLSTM combined network, extract local features using convolutional layers, extract long-term evolution trends using a bidirectional long short-term memory network, and output low-frequency component prediction values; For the high-frequency non-stationary sequence, input it and the high-frequency feature set obtained in step S4 into an Autoformer network, separate residual trends using the sequence decomposition module inside the Autoformer network, calculate the periodic similarity of the sequence using an autocorrelation mechanism and aggregate sub-process information, and output high-frequency component prediction values; The low-frequency component prediction values ​​and the high-frequency component prediction values ​​are linearly superimposed at corresponding time points, and the superposition result is inversely normalized to obtain the final wastewater treatment effluent quality prediction result.

[0042] Based on the two types of differential sequences obtained from the clustering reconstruction in step S3: low-frequency trend sequences With high-frequency non-stationary sequences Based on the influent water quality parameters obtained in step S1, this step constructs a parallel dual-channel prediction model targeting different frequency characteristics to capture the long-term evolution and transient impact characteristics of water quality changes. The specific implementation process is as follows: S5.1: Low-Frequency Channel: CNN-BiLSTM trend prediction for low-frequency trend sequences. This channel aims to predict low-frequency trend sequences that characterize the basic biochemical reaction patterns in wastewater treatment processes. This sequence reflects a stable baseline variation influenced by hydraulic retention time and microbial activity.

[0043] 1) Constructing a multidimensional input feature matrix: The low-frequency trend sequence obtained from step S3 is selected as the main feature, and the low-frequency feature set obtained from step S4 (containing influent water quality parameters and biochemical process parameters strongly correlated with the low-frequency sequence) is used as auxiliary features. The main and auxiliary features are concatenated along the channel dimension to construct a multidimensional input feature matrix reflecting the long-term evolution relationship of "influent load - biochemical process - effluent standard". Each row of this matrix represents a time step, and each column represents a feature variable.

[0044] 2) Local biochemical feature extraction based on CNN: The above feature matrix is ​​processed using a one-dimensional convolutional layer (1D-CNN). To address potential minute observation noise and non-reactive jitter in the sensor data, sliding filtering and feature extraction are performed along the time axis using convolutional kernels. Assume the input to the l-th convolutional layer is... Then the output of the k-th convolution kernel The calculation formula is:

[0045] in, This represents the convolution operation. As weight, For bias, As the activation function, this step smooths out the observation noise in the sensor data through sliding window calculation.

[0046] 3) Hysteresis Modeling Based on BiLSTM: Features extracted by CNN are input into a bidirectional long short-term memory network. Due to the significant hysteresis (current effluent state is influenced by past influent) and continuity of wastewater biochemical reactions, unidirectional LSTMs easily overlook future contextual information (reverse dependence). BiLSTM thoroughly captures the evolutionary patterns through the hidden layer states in both forward and reverse directions. The hidden layer states of BiLSTM at time t are shown. From forward state and backward state It is pieced together:

[0047]

[0048]

[0049] in, and These are the weighting coefficients.

[0050] 4) Low-frequency component prediction output: The final hidden layer state output by BiLSTM. The input is a fully connected layer, which uses a linear transformation to map the high-dimensional feature space to a one-dimensional output space, thus obtaining the predicted value of the low-frequency trend sequence at the next time step. The calculation formula is:

[0051] In the formula, This is the weight matrix of the fully connected layer. This is a bias term. S5.2 High-frequency channel: High-frequency impact prediction based on Autoformer network. This step uses Autoformer network to predict high-frequency non-stationary sequences that reflect inflow impact and load changes.

[0052] 1) Sequence Decomposition Module: This module serves as the fundamental computational unit of the Autoformer network, permeating every layer of the encoder and decoder. Its function is to dynamically decouple the hidden layer features, separating them into "trend terms (…)". ")" and "periodic terms ( )".

[0053]

[0054]

[0055] In the formula, It's a pooling operation that extracts the smoothed trend portion of the sequence. This indicates that the input sequence is padded.

[0056] By subtracting the moving average from the original sequence to obtain a smoothed trend, the model is forced to focus only on the trend. This purely fluctuating component allows for a more sensitive capture of instantaneous changes caused by the impact of incoming water.

[0057] 2) Encoder processing: The encoder is used to analyze historical high-frequency sequences and extract impact mode features from historical operating conditions.

[0058] Input data: Select high-frequency non-stationary sequences from the past L time steps. and auxiliary features as encoder input .

[0059] Processing flow and formulas: Assuming the encoder consists of N stacked layers, the operation logic of the l-th layer is as follows: Autocorrelation calculation: First, the input is processed using the autocorrelation mechanism.

[0060] Residual connection and decomposition: Add the relevant output to the original input and feed it into the sequence decomposition module.

[0061] In the formula, This represents the input features of the l-th layer encoder. This indicates the autocorrelation mechanism operation. This represents the sequence decomposition operation, indicating the trend component that is separated and ignored in this step.

[0062] Feedforward Networks and Decomposition: Transforming Intermediate Results The nonlinear representation is enhanced by feeding it into a feedforward network before further decomposition.

[0063] In the formula, This represents a feedforward neural network layer, used to perform nonlinear transformations on features to enhance the model's ability to express complex biochemical reactions.

[0064] Output: The encoder's final output feature map .

[0065] 3) Decoder Prediction Process: The decoder package adopts a trend and periodicity fusion structure, combining intra-layer and inter-layer autocorrelation mechanisms to enhance trend modeling and periodic feature capture, respectively. During the decoding process, the model extracts latent trends from the hidden states and optimizes the prediction output layer by layer. The decoding calculation process of the l-th layer (out of M layers) is as follows:

[0066] In the formula, This represents the input to the l-th layer decoder;

[0067]

[0068] The output of layer l is .

[0069] 4) Final prediction output: The output of the last layer of the decoder This is the predicted high-frequency fluctuation component. Adding it to the average trend component obtained based on the decoder input projection yields the final predicted high-frequency component value:

[0070] The predicted values ​​of the low-frequency component and the high-frequency component are linearly superimposed at the corresponding time steps to obtain the superimposed prediction sequence. Then, using the parameters from the normalization process in step S1, the superimposed prediction sequence is inversely normalized to obtain the final predicted value of the wastewater treatment effluent with physical dimensions.

[0071] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention, and are not actually limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar methods and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for predicting wastewater treatment effluent quality based on frequency domain adaptive partitioning and a dual-channel network, characterized in that, Includes the following steps: Step S1: Obtain historical data of the wastewater treatment process, including influent water quality parameters and effluent water quality parameters, and preprocess the historical data to obtain a preprocessed water quality test sample dataset; Step S2: Use the ICEEMDAN algorithm to perform a global multi-scale decomposition on the preprocessed historical effluent water quality time series, and calculate the sample entropy values ​​of all intrinsic mode components (IMFs) obtained after the decomposition. Step S3: Using the sample entropy value as the feature input, the K-means clustering algorithm is used to adaptively divide all intrinsic mode components into two categories: high-frequency components and low-frequency components. The high-frequency components are superimposed and reconstructed into a high-frequency non-stationary sequence, and the low-frequency components and residual components are superimposed and reconstructed into a low-frequency trend sequence. Step S4: Using the maximum mutual information coefficient analysis method, calculate the correlation between each influent water quality parameter and the low-frequency trend sequence and the high-frequency non-stationary sequence, set the correlation threshold, screen out the strongly correlated influent parameters to form a low-frequency feature set and a high-frequency feature set, and divide the feature set into a training set and a test set according to the proportion. Step S5: Construct a frequency domain adaptive dual-channel prediction model to perform parallel predictions on the low-frequency trend sequence and the high-frequency non-stationary sequence, respectively. The low-frequency channel uses a CNN-BiLSTM combined network to output the predicted low-frequency components, and the high-frequency channel uses an Autoformer network to output the predicted high-frequency components. Finally, the predicted low-frequency components and the predicted high-frequency components are linearly superimposed at the corresponding time points, and the superposition result is inversely normalized to obtain the final predicted water quality of the wastewater treatment effluent.

2. The wastewater treatment effluent quality prediction method based on frequency domain adaptive partitioning and dual-channel network according to claim 1, characterized in that, The preprocessing in step S1 includes outlier detection, missing value imputation, and data normalization. The outlier detection is used to identify and remove negative values, extremely large outliers, and non-numeric data from historical data. The missing value filling is performed by using linear interpolation to fill in the missing values; The data normalization process uses max-min normalization to scale all data to the range [0,1].

3. The wastewater treatment effluent quality prediction method based on frequency domain adaptive partitioning and dual-channel network according to claim 1, characterized in that, In step S1, the influent water quality parameters are used as model input feature covariates, including: wastewater influent flow rate, influent dissolved oxygen concentration, influent ammonia nitrogen concentration, influent total nitrogen concentration, and influent total phosphorus concentration; the effluent water quality parameters are used as target variables for model prediction, including: effluent dissolved oxygen concentration, effluent ammonia nitrogen concentration, effluent total nitrogen concentration, and effluent total phosphorus concentration.

4. A method for predicting wastewater treatment effluent quality based on frequency domain adaptive partitioning and a dual-channel network according to any one of claims 1-3, characterized in that, The K-means clustering algorithm in step S3 includes the following sub-steps: Construct a feature dataset by combining the sample entropy values ​​of each intrinsic mode component into a one-dimensional feature set; Initialize cluster centers and set the number of clusters to 2; Calculate the Euclidean distance between the sample entropy value and the cluster center, and classify them according to the principle of minimizing the distance; Update the cluster centers and iterate until convergence, thus completing the high-frequency and low-frequency division of all intrinsic mode components.

5. The wastewater treatment effluent quality prediction method based on frequency domain adaptive partitioning and dual-channel network according to claim 4, characterized in that, In step S3, all intrinsic mode components that will be assigned to high-frequency components are linearly added in the time domain to obtain a high-frequency non-stationary sequence, which is used as the input to the Autoformer network. All intrinsic mode components and residual components that are assigned to low-frequency components are linearly added in the time domain to obtain a low-frequency trend sequence, which is used as the input to the CNN-BiLSTM combined network.

6. The wastewater treatment effluent quality prediction method based on frequency domain adaptive partitioning and dual-channel network according to claim 5, characterized in that, In step S4, the maximum mutual information coefficient analysis method is used to calculate the influent water quality parameters and the low-frequency trend sequence, respectively. High-frequency non-stationary sequences Non-linear correlation between them; setting low-frequency correlation thresholds and high-frequency correlation thresholds for low-frequency trend sequences. A low-frequency feature set was constructed by screening out strongly correlated influent water quality parameters; for high-frequency non-stationary sequences... Strongly correlated influent water quality parameters were selected to form a high-frequency feature set. The selected low-frequency feature set and high-frequency feature set were strictly divided into training set and test set in an 8:2 ratio according to the chronological order.

7. The wastewater treatment effluent quality prediction method based on frequency domain adaptive partitioning and dual-channel network according to claim 6, characterized in that, In step S5, the low-frequency trend sequence and the low-frequency feature set obtained in step S4 are input into the CNN-BiLSTM combined network. Local features are extracted using convolutional layers, and long-term evolution trends are extracted using a bidirectional long short-term memory network. The predicted values ​​of the low-frequency components are then output.

8. The wastewater treatment effluent quality prediction method based on frequency domain adaptive partitioning and dual-channel network according to claim 7, characterized in that, In step S5, the high-frequency feature set obtained in step S4 of the high-frequency non-stationary sequence is input into the Autoformer network. The residual trend is separated by the sequence decomposition module inside the Autoformer network, the periodic similarity of the sequence is calculated by the autocorrelation mechanism, and the sub-process information is aggregated to output the predicted value of the high-frequency component.

9. The wastewater treatment effluent quality prediction method based on frequency domain adaptive partitioning and dual-channel network according to claim 8, characterized in that, In step S5, the predicted values ​​of the low-frequency components and the high-frequency components are linearly superimposed at the corresponding time steps; using the parameters from the normalization process in step S1, the superimposed prediction sequence is inversely normalized to obtain the final predicted value with physical dimensions.

Images