Wastewater biological treatment process effluent prediction method and system based on multi-modal learning

Abstract

The invention discloses a wastewater biological treatment process effluent prediction method and system based on multi-modal learning. The method comprises the following steps: firstly, collecting multi-modal data of a wastewater biological treatment process; secondly, respectively encoding different modal data based on a feature extraction technology; dividing the feature vector into low-dimension data and high-dimension data, and combining the low-dimension data and the dimension-reduced high-dimension data to respectively establish a reference effluent prediction model; and finally, performing effluent prediction on the wastewater biological treatment process based on a multi-modal decision fusion strategy and the prediction model. The system comprises a multi-modal data collection module, a multi-modal data representation module, a reference model construction module and a multi-modal decision fusion module. According to the method, multiple data modes in the wastewater biological treatment process are combined, the effluent prediction model is constructed based on the multi-mode learning fusion strategy, the accuracy and robustness of the effluent prediction model are remarkably improved, and intelligent development of the wastewater treatment process is facilitated.

Description

Technical Field

[0001] The present invention relates to a method for predicting effluent from a wastewater biological treatment process, and in particular to a method and system for predicting effluent from a wastewater biological treatment process based on multimodal learning. Background Art

[0002] Biological wastewater treatment processes are widely used by sewage treatment plants to efficiently treat industrial and domestic wastewater to ensure that it meets discharge standards. However, biological treatment methods still face many technical challenges in effluent quality prediction and process control. At present, a large number of key water quality and ecological risk parameters are involved in the sewage treatment process, some of which can be measured in real time by online sensors, while other parameters rely on manual detection, which is usually time-consuming. This delay reduces the timeliness of data and increases operating costs, which poses significant challenges to the real-time control of sewage treatment and the prediction of effluent quality.

[0003] At present, machine learning and deep learning technologies are widely used in effluent prediction in the field of sewage treatment. Patent CN112149799A discloses a water quality parameter prediction model based on a joint neural network, which realizes the effective prediction of the sludge volume index by combining numerical data (pH, chemical oxygen demand COD, dissolved oxygen concentration DO and total nitrogen TN). Patent application US17472433 discloses an intelligent detection method for effluent total nitrogen based on multi-objective optimization-fuzzy neural network, which combines numerical data (dosage, redox potential in the middle of the anaerobic tank, orthophosphate at the end of the aerobic tank, pH in the inlet tank, ammonia nitrogen in the inlet tank, nitrate nitrogen at the end of the anoxic tank and chemical oxygen demand in the primary sedimentation tank), and effectively predicts the total nitrogen concentration in the effluent water body. Patent CN118260574A discloses a method for predicting the effluent water quality of industrial sewage based on deep learning, which combines the inlet water quality data indicators and sludge concentration data, and successfully realizes the effective prediction of effluent chemical oxygen demand, ammonia nitrogen, total phosphorus and total nitrogen concentrations.

[0004] Although the above methods have achieved certain results in the effluent prediction of sewage treatment plants, most of them are based on numerical feature data and ignore the potential information of other modal data. Relying only on a single data modality makes it difficult for existing methods to fully capture the potential relationships and interactions between various types of data when dealing with complex water quality prediction and ecological risk monitoring problems. Therefore, it not only limits the adaptability of the prediction model to multi-factor, nonlinear dynamic changes, but also makes the prediction results lack accuracy and robustness when facing complex and changeable application environments. Patent application CN117273540A discloses a method for predicting the effluent quality of a sewage treatment system, which fuses the multimodal data of the sewage liquid level image and the water quality sensor to generate a weighted multimodal feature vector, and inputs the weighted feature vector into the CNN-LSTM network to obtain the prediction result of the effluent quality of sewage treatment. However, the patent application also only integrates two types of information, measured data and liquid level image, and the prediction accuracy obtained is still greatly limited. In general, the effluent prediction method of wastewater biological treatment process based on multimodal learning strategy is still relatively limited, and the accuracy and robustness of the prediction cannot meet the actual application requirements. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a method and system that can meet the requirements of accuracy and robustness of effluent prediction of wastewater biological treatment processes in complex application scenarios.

[0006] Technical solution: The water output prediction method of the present invention comprises the following steps:

[0007] (1) Collecting multimodal data of the wastewater biological treatment process; the multimodal data includes effluent ecological risk data;

[0008] (2) Encode different modal data separately based on feature extraction technology to obtain feature vectors;

[0009] (3) dividing the feature vector into low-dimensional data and high-dimensional data; merging the low-dimensional data based on a multimodal feature fusion strategy, and building a benchmark water discharge prediction model 1 in combination with the water discharge ecological risk data; and reducing the dimension of the high-dimensional data based on feature selection, and building a benchmark water discharge prediction model 2 in combination with the water discharge ecological risk data;

[0010] (4) Based on the multimodal decision fusion strategy and the prediction model, the effluent of the wastewater biological treatment process is predicted.

[0011] In step (1), the effluent ecological risk data include effluent discharge content of total nitrogen, ammonia nitrogen, nitrate nitrogen, nitrite nitrogen, total phosphorus, heavy metals, characteristic pollutants, micropollutants and effluent biological toxicity.

[0012] In step (2), the data collected in step (1) is preprocessed and then feature extraction is performed.

[0013] In step (3), the selected feature set is used as input, and the water ecological risk data is used as the prediction target to establish the benchmark water prediction model. Specifically, the first benchmark water prediction model is constructed based on machine learning and deep learning methods, using the feature vector after feature fusion as input, the water ecological risk data as the prediction target, and the second benchmark water prediction model is constructed based on machine learning and deep learning methods, using the feature vector after feature dimensionality reduction as input, the water ecological risk data as the prediction target, and the machine learning and deep learning methods.

[0014] The present invention combines multiple data modalities in the process of biological treatment of wastewater, solves the problem of feature imbalance between different modalities through the feature fusion strategy of multimodal learning, and fully explores the potential correlation and complementary characteristics of each modal data to construct a benchmark model for effluent prediction of biological treatment process. Finally, combined with the decision fusion strategy of multimodal learning, the accuracy and robustness of effluent prediction of wastewater biological treatment process are effectively improved.

[0015] Preferably, in step (1), the multimodal data includes numerical, image and text modal data. For example: 1) Basic water quality data of inlet and outlet water: chemical oxygen demand, biochemical oxygen demand, suspended solids, total nitrogen, ammonia nitrogen, nitrate nitrogen, nitrite nitrogen, total phosphorus, animal and plant oils, petroleum, anionic surfactants, total organic carbon, dissolved oxygen, chromaticity, temperature, pH value, total bacteria count, fecal coliform count, heavy metal content, characteristic pollutant content and micropollutant content, etc.; 2) Water quality fingerprint data: three-dimensional fluorescence spectrum, Raman spectrum and ultraviolet-visible spectrum, etc.; 3) Sludge performance data: sludge volume index, sludge settling ratio, sludge moisture content, mixed liquor suspended solids concentration, mixed liquor volatile suspended solids concentration, sludge specific resistance and sludge dewatering performance, etc.; 4) Biological treatment process information: biological treatment process used by sewage treatment plants.

[0016] Preferably, the numerical modal data is preprocessed, including missing value interpolation, outlier detection and processing, deduplication and redundancy checking, and normalization and standardization processing. The value obtained after the numerical data is preprocessed is used as a new feature value.

[0017] Preferably, the image-type modal data is preprocessed, including denoising, baseline correction, smoothing and spectral normalization. The image-type modal data includes spectral data. The image-type modal data is converted into a numerical feature vector after preprocessing, with wavelength, Raman shift, etc. as input features and corresponding fluorescence intensity or absorbance as corresponding feature values.

[0018] Preferably, the text-based modal data is a wastewater biological treatment process method; the method for preprocessing it is to divide the wastewater biological treatment process method into three types of characteristic ponds according to the dissolved oxygen in the biological treatment unit pond, including an anaerobic pond, an anoxic pond, and an aerobic pond.

[0019] Preferably, the division criteria for the characteristic ponds are as follows: the dissolved oxygen DO in the anaerobic pond ≤ 0.2 mg / L; the anoxic pond 0.2 mg / L < DO ≤ 0.5 mg / L; the aerobic pond DO > 0.5 mg / L.

[0020] Preferably, the wastewater biological treatment process method is encoded according to the number, combination, or connection type of the characteristic ponds.

[0021] The specific encoding method is as follows:

[0022] 1) According to the number of characteristic ponds, the wastewater treatment process method is encoded as [a, b, c], which respectively represent the numbers of the three types of characteristic ponds.

[0023] 2) According to the combination of characteristic ponds, the encoding of the wastewater treatment process method includes: encoding the anaerobic pond as [1, 0, 0], the anoxic pond as [0, 1, 0], and the aerobic pond as [0, 0, 1] based on the one-hot algorithm; representing the wastewater treatment process with "n characteristic ponds", and encoding the process method as a 3×n-dimensional feature vector in sequence according to the combination order of the characteristic ponds. When the characteristic pond combination of the process method is less than n, the blank pond [0, 0, 0] is used for supplementation.

[0024] 3) According to the connection type of characteristic ponds, the wastewater treatment process method is encoded as [d, e, f, g, h, i], which respectively represent the numbers of six characteristic pond connection types (anaerobic pond connected to anoxic pond, anaerobic pond connected to aerobic pond, anoxic pond connected to anaerobic pond, anoxic pond connected to aerobic pond, aerobic pond connected to anaerobic pond, aerobic pond connected to anoxic pond).

[0025] Preferably, in step (3), the performance of the prediction model is evaluated based on cross-validation experiments. When establishing a benchmark effluent prediction model, different machine learning and deep learning algorithms can be selected, and by evaluating the performance of the obtained models, the benchmark effluent prediction model with the optimal prediction performance can be selected. The evaluation indicators can be MAE, MSE, R 2 , and the calculation method is:

[0026]

[0027] Among them, y i represents the true value, represents the predicted value, represents the mean value of the samples.

[0028] Preferably, a weight is assigned to the output of each benchmark model according to the performance, and the final prediction result is calculated by weighted average. The calculation method is:

[0029]

[0030] Among them, Y represents the final predicted value, w j represents the weight of the benchmark model, Represents the predicted value of the baseline model. The sum of weights w is 1.

[0031] The water output prediction system of the present invention comprises a multimodal data collection module, a multimodal data characterization module, a benchmark model construction module and a multimodal decision fusion module.

[0032] Preferably, the multimodal data collection module collects multimodal data involved in the wastewater biological treatment process, including basic inlet and outlet water quality, water quality fingerprint, sludge performance, biological treatment process information and effluent ecological risk data, covering data in different modalities such as numerical, image and text types.

[0033] Preferably, the multimodal data characterization module encodes and maps different modal data in wastewater biological treatment based on feature extraction technology, converting the original data into an efficient feature representation that is more suitable for model processing, so as to better capture the potential information in different modal data.

[0034] Preferably, the benchmark model construction module divides the low-dimensional and high-dimensional data, merges the low-dimensional data based on a multimodal feature fusion method, performs feature dimensionality reduction on the high-dimensional data based on a sequential forward feature selection algorithm, and constructs different benchmark prediction models respectively.

[0035] Preferably, the multimodal decision fusion module fuses the results of multiple benchmark prediction models based on a multimodal decision fusion strategy, further optimizes the decision-making process, and ultimately achieves a more accurate prediction of sewage effluent water quality and risks.

[0036] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. The effluent prediction method provided by the present invention adopts a multimodal fusion strategy, which can more comprehensively capture the complex dynamic relationships in the wastewater biological treatment process, significantly improve the prediction accuracy and robustness of effluent water quality and risks, and provide more reliable decision-making data support for the intelligent optimization of sewage treatment processes; 2. The effluent prediction system provided by the present invention is highly scalable, and can flexibly adjust data types and prediction targets according to the prediction and control requirements of different wastewater biological treatments, thereby meeting the actual needs of different effluent water quality and risk assessment and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the water output prediction method of the present invention. DETAILED DESCRIPTION

[0038] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0039] Example 1: A method for predicting effluent from a wastewater biological treatment process based on multimodal learning

[0040] The prediction method comprises the following steps:

[0041] (1) Collect multimodal data of wastewater biological treatment process.

[0042] The basic water quality, water quality fingerprint, biological treatment process information and effluent ecological risk data of 71 sewage treatment plants that use wastewater biological treatment technology in Jiangsu, Anhui, Jiangxi and other provinces and cities were collected.

[0043] Among them, the basic water quality characteristics of inlet and outlet water include chemical oxygen demand, total nitrogen, ammonia nitrogen, total phosphorus, temperature, and pH value; the water quality fingerprint characteristics include three-dimensional fluorescence data; the biological treatment process information is the biological treatment process used by the sewage treatment plant, and the ecological risk data is biological toxicity data.

[0044] (2) Based on feature extraction technology, different modal data are encoded separately to obtain feature vectors.

[0045] The collected multimodal data are preprocessed separately, and features are extracted from the preprocessed data of different modalities.

[0046] The preprocessing and feature extraction steps for the collected numerical modal data are missing value interpolation, outlier detection and processing, deduplication and redundancy check, and normalization and standardization. The value obtained after the numerical data is preprocessed is used as the new feature value.

[0047] The preprocessing and feature extraction steps for the collected image-based modal data (three-dimensional fluorescence data) are denoising, baseline correction, smoothing, and spectral normalization. The spectral data obtained after the preprocessing steps are converted into numerical feature vectors, where the excitation wavelength-emission wavelength is used as the input feature and the corresponding fluorescence intensity is used as the corresponding feature value.

[0048] The steps of preprocessing and feature extraction for the collected text-based modal data (process method) are as follows: First, divide the collected wastewater biological treatment process methods into three types of characteristic pools: anaerobic pool (DO ≤ 0.2 mg / L), anoxic pool (0.2 mg / L < DO ≤ 0.5 mg / L), and aerobic pool (DO > 0.5 mg / L) according to the different dissolved oxygen DO of the biological treatment unit pools; then encode the wastewater biological treatment process methods according to three encoding methods: the number, combination, and connection type of different characteristic pools. The specific encoding method for the wastewater biological treatment process includes:

[0049] (A) Encode the wastewater treatment process method according to the number of different characteristic pools as [a, b, c], which respectively represent the numbers of the three types of characteristic pools.

[0050] (B) Encoding the wastewater treatment process method according to the combination of different characteristic pools includes: encoding the anaerobic pool as [1, 0, 0], the anoxic pool as [0, 1, 0], and the aerobic pool as [0, 0, 1] based on the one-hot algorithm; characterizing the wastewater treatment process with "four characteristic pools", and encoding the process method as a 12-dimensional feature vector in sequence according to the combination order of the characteristic pools. When the combination of characteristic pools of the process method is less than 4, use the blank pool [0, 0, 0] to supplement.

[0051] (C) Encode the wastewater treatment process method according to the connection type of different characteristic pools as [d, e, f, g, h, i], which respectively represent the numbers of six characteristic pool connection types (anaerobic pool connected to anoxic pool, anaerobic pool connected to aerobic pool, anoxic pool connected to anaerobic pool, anoxic pool connected to aerobic pool, aerobic pool connected to anaerobic pool, aerobic pool connected to anoxic pool).

[0052] For example, the anaerobic-anoxic-aerobic process can be respectively characterized as A: [1, 1, 1], B: [0, 0, 1, 0, 1, 0, 1, 0, 0, 0, 0, 0], and C: [1, 0, 0, 1, 0, 0] according to three different encoding methods.

[0053] (3) Divide the feature vectors into low-dimensional data and high-dimensional data; for the low-dimensional data, merge it based on the multi-modal feature fusion strategy; for the high-dimensional data, perform dimensionality reduction based on feature selection. Based on the selected feature set as the input, establish a benchmark effluent prediction model with the effluent ecological risk data as the prediction target.

[0054] (3.1) Divide the feature vectors encoded by the multi-modal data into low-dimensional data (basic water quality of influent and effluent, biological treatment process characteristics) and high-dimensional data (three-dimensional fluorescence data).

[0055] (3.2) Merge the low-dimensional data based on the multi-modal feature fusion strategy, and construct a benchmark effluent prediction model 1 in combination with the effluent ecological risk data. Specifically as follows:

[0056] Based on the multimodal feature fusion strategy, low-dimensional data are merged, the feature vector after feature fusion is used as input, and the effluent ecological risk data (biological toxicity) is used as the prediction target. A benchmark effluent prediction model is constructed based on different machine learning and deep learning methods (support vector regression, random forest regression and neural network regression algorithm).

[0057] The prediction performance of the established benchmark prediction model was evaluated using cross-validation experiments, using MAE, MSE and R 2 The three evaluation indicators are evaluated, and the results are shown in Table 1. The prediction model based on the support vector regression algorithm has the best evaluation performance. Therefore, this prediction model is selected as the benchmark prediction model 1.

[0058] Table 1 Evaluation index results of prediction model 1 based on different algorithms

[0059]

[0060] (3.3) Based on feature selection technology, high-dimensional data is reduced in dimension, and the second benchmark water discharge prediction model is constructed in combination with water discharge ecological risk data. The details are as follows:

[0061] The three-dimensional fluorescence data were processed for feature dimensionality reduction based on the sequential forward feature selection algorithm. The feature vector after feature dimensionality reduction was used as input, and the effluent ecological risk data (biological toxicity) was used as the prediction target. A benchmark effluent prediction model was constructed based on different machine learning and deep learning methods (support vector regression, random forest regression and neural network regression algorithm).

[0062] The prediction performance of the established benchmark prediction model was evaluated using cross-validation experiments, using MAE, MSE and R 2 The three evaluation indicators are evaluated, and the results are shown in Table 2. The prediction model based on the neural network regression algorithm has the best evaluation performance. Therefore, this prediction model is selected as the benchmark prediction model 2.

[0063] Table 2 Evaluation index results of prediction model 2 based on different algorithms

[0064]

[0065] (4) Based on the multimodal decision fusion strategy and the prediction model, the effluent of the wastewater biological treatment process is predicted.

[0066] According to the prediction performance index R 2 Assign a weight w to the output of each baseline model.

[0067] in, The final prediction result is calculated by weighted average.

[0068] For example, when the chemical oxygen demand is 70 mg / L, the total phosphorus is 2 mg / L, the total nitrogen is 20 mg / L, the ammonia nitrogen is 17 mg / L, the temperature is 25°C, and the pH value is 6.8, the actual effluent toxicity is 0.77. The effluent toxicity obtained using the benchmark prediction model 1 is 0.73, and the effluent toxicity obtained using the benchmark prediction model 2 is 0.82. Based on the multimodal decision fusion strategy, the final prediction result is Y = 0.5069 × 0.73 + 0.4931 × 0.82 = 0.7744.

[0069] The final prediction results based on the multimodal learning strategy are calculated using MAE, MSE and R 2 The three evaluation indicators were evaluated, and the results were MAE = 0.0564, MSE = 0.0673, R 2 =0.9036.

[0070] It can be seen from the results that the effluent prediction model constructed based on the multimodal learning fusion strategy shows significant advantages in the sewage treatment process, and the prediction results have good accuracy and robustness.

Claims

1. A method for predicting effluent from a wastewater biological treatment process based on multimodal learning, characterized in that: It includes the following steps: (1) Collect multi-modal data of the wastewater biological treatment process; the multi-modal data includes effluent ecological risk data; (2) Encode different modal data respectively based on feature extraction technology to obtain feature vectors; (3) Divide the feature vectors into low-dimensional data and high-dimensional data; for the low-dimensional data, merge them based on the multi-modal feature fusion strategy, and combine the effluent ecological risk data to construct a benchmark effluent prediction model one; for the high-dimensional data, perform dimensionality reduction based on feature selection, and combine the effluent ecological risk data to construct a benchmark effluent prediction model two; (4) Perform effluent prediction on the wastewater biological treatment process based on the multi-modal decision fusion strategy and the prediction model.

2. The water output prediction method according to claim 1, characterized in that: In step (1), the multi-modal data includes numerical, image, and text modal data.

3. The water output prediction method according to claim 2, characterized in that: Preprocess the numerical modal data, including missing value imputation, outlier detection and processing, duplicate and redundancy checking, and normalization and standardization processing.

4. The water output prediction method according to claim 2, characterized in that: Preprocess the image modal data, including noise removal, baseline correction, smoothing, and spectral normalization processing.

5. The water output prediction method according to claim 2, characterized in that: The text modal data is the wastewater biological treatment process method; the method for preprocessing it is to divide the wastewater biological treatment process method into three types of feature pools according to the dissolved oxygen in the biological treatment unit pool, including anaerobic pool, anoxic pool, and aerobic pool.

6. The water output prediction method according to claim 5, characterized in that: The division standard of the feature pool is: the dissolved oxygen DO of the anaerobic pool ≤ 0.2 mg / L; the anoxic pool 0.2 mg / L < DO ≤ 0.5 mg / L; the aerobic pool DO > 0.5 mg / L.

7. The water output prediction method according to claim 5, characterized in that: Encode the wastewater biological treatment process method according to the number, combination, or connection type of the feature pools.

8. The water output prediction method according to claim 1, characterized in that: In step (3), evaluate the performance of the prediction model based on cross-validation experiments.

9. The water discharge prediction method according to claim 8, characterized in that: Assign weights to the outputs of each benchmark model according to the performance, and the final prediction result is calculated by the weighted average method.

10. An effluent prediction system according to the effluent prediction method described in claim 1, including a multi-modal data collection module, a multi-modal data characterization module, a benchmark model construction module, and a multi-modal decision fusion module.

Images