Method for predicting and interpreting denitrification performance and microbial abundance of single-stage PNA system

Abstract

The invention relates to the technical field of sewage / wastewater denitrification, in particular to a method for predicting and explaining denitrification performance and microbial abundance of a single-stage PNA system, which comprises the following steps: constructing an ML model of nitrogen removal rate and functional microbial abundance, and incorporating key water quality and operating parameters into the model to improve the interpretability and prediction accuracy of the model; and the influence of each input feature on the model and the optimal range of the environmental parameters are determined, so that targeted operation adjustment is realized. Quantitative analysis is carried out through the model, potential synergistic effects among the characteristics can be explored, and therefore the optimal characteristic range can be found, the wastewater treatment efficiency can be improved, and the optimal water environment can be obtained.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of wastewater denitrification technology, and in particular to a method for predicting and interpreting the denitrification performance and microbial abundance of a single-stage PNA system. BACKGROUND

[0002] Compared with the two-stage PNA (short-cut nitrification-anaerobic ammonium oxidation) process, the single-stage PNA process requires less land and has lower construction costs, making it more suitable for upgrading and retrofitting existing water plants. However, the control factors affecting the stability of the denitrification performance of the single-stage PNA, such as dissolved oxygen (DO), substrate concentration, temperature, and pH, are still challenges in its application. Therefore, it is very important to reveal the effects of environmental factors on the single-stage PNA, understand its optimal parameters, and propose operation strategies.

[0003] In the single-stage PNA, various methods are used to specifically inhibit NOB by taking advantage of the different physiological characteristics of anaerobic ammonium-oxidizing bacteria (AnAOB), ammonia-oxidizing bacteria (AOB), and nitrosation bacteria (NOB). Common strategies include adjusting the concentrations of oxygen, free ammonia (FA), and free nitrous acid (FNA). However, the adaptability of NOB to inhibitory factors remains a challenge for stable NOB inhibition. Therefore, it is crucial to develop a comprehensive strategy that uses multiple methods to synergistically inhibit NOB. Candidatus Brocadia anammoxidans and Candidatus Kuenenia stuttgartiensis Brooksia and Candidatus Brokacia are the two most common anaerobic ammonium-oxidizing microorganisms in PNA systems. Although they both have anaerobic ammonium-oxidizing ability, their optimal environmental conditions (such as oxygen, temperature, and substrate concentration) are different, and they occupy different ecological niches in anaerobic microbial communities. Therefore, using existing large-scale data to accurately predict the dominant anaerobic microbial species in a particular environment is crucial for optimizing their effectiveness in treating various nitrogen-containing wastewater and sludge inoculation. SUMMARY

[0004] ​In order to realize accurate prediction of the main anaerobic microbial species in a specific environment by using existing large-scale data, the purpose of the present disclosure is to propose a method for predicting and explaining the denitrification performance and microbial abundance of a single-stage short-cut nitrification-anammox system, in which a prediction model of nitrogen removal rate (NRR) and functional microbial abundance is constructed, key water quality and operating parameters are incorporated into the model by using artificial neural network (ANN), random forest (RF), support vector regression (SVR) and eXtreme gradient boosting (XGBoost) to improve the interpretability and prediction accuracy of the model. Then SHAP (SHapley Additive ex Planations) is applied to determine the influence of each input feature on the model and the optimal range of environmental parameters, so as to realize targeted operation adjustment.

[0005] To achieve the above technical purpose, the first aspect of the present application proposes a denitrification performance analysis method of a single-stage PNA system, which comprises the following steps: adopting a single-stage PNA process, selecting data with preset characteristics as samples, and pre-processing the samples, wherein the preset characteristics include influent ammonia nitrogen concentration, free ammonia concentration, free nitrite concentration, influent nitrogen load, running days, , pH value, temperature, hydraulic retention time, ammonia-oxidizing bacteria, nitrosobacteria, anammox bacteria, Candidatus Brocadia anammoxidans, Candidatus Kuenenia stuttgartiensis , nitrogen removal rate; constructing a denitrification performance prediction model of the single-stage PNA process for training, and inputting the influent ammonia nitrogen concentration, free ammonia concentration, free nitrite concentration, influent nitrogen load, running days, , pH value, temperature, hydraulic retention time, ammonia-oxidizing bacteria, nitrosobacteria, anammox bacteria, Candidatus Brocadia anammoxidans, Candidatus Kuenenia stuttgartiensis as input and nitrogen removal rate as output; applying the trained denitrification performance prediction model of the single-stage PNA process for inference, and using SHAP value quantitative analysis to explain the influence of different characteristics on the nitrogen removal rate.

[0006] In an embodiment of the above technical solution, the pre-processing includes converting and standardizing the units of the preset characteristics, eliminating abnormal data, and normalizing the characteristic values of the preset characteristics.

[0007] In an embodiment of the above technical solution, the denitrification performance prediction model of the single-stage PNA process is constructed by using neural network ANN, and the hyperparameters of the Bayesian optimization model are used in the training.

[0008] In an implementation form of the above technical solution, the SHAP value quantitative analysis comprises: calculating a SHAP value of a single feature and a SHAP value of two features, the former being used to explain the influence of each feature, and the latter being used to quantify the optimal range of the feature.

[0009] Based on the similar method, the second aspect of the present disclosure proposes an analysis method of functional microbial abundance of a single-stage PNA system, the method comprising the following steps: adopting a single-stage PNA process, selecting data with preset characteristics as samples, and pretreating the samples, wherein the preset characteristics include influent ammonia nitrogen concentration, effluent ammonia nitrogen concentration, effluent nitrite concentration, effluent nitrate nitrogen concentration, effluent total nitrogen concentration, free ammonia concentration, free nitrite concentration, influent nitrogen load, NRR, running days, , pH value, temperature, hydraulic retention time, anammox bacteria, ammonia oxidizing bacteria, and nitrosation bacteria, Candidatus Brocadia anammoxidans, Candidatus Kuenenia stuttgartiensis ; constructing a microbial abundance model of the single-stage PNA process for training, taking the influent ammonia nitrogen concentration, the effluent ammonia nitrogen concentration, the effluent nitrite concentration, the effluent nitrate nitrogen concentration, the effluent total nitrogen concentration, the free ammonia concentration, the free nitrite concentration, the influent nitrogen load, the NRR, the running days, , the pH value, the temperature, and the hydraulic retention time as inputs, and taking the abundance of each functional microorganism as a target output; and performing inference application on the trained microbial abundance model of the single-stage PNA process, and performing SHAP value quantitative analysis to explain the influence of different characteristics on the abundance of each functional microorganism.

[0010] In an implementation form of the above technical solution, the microbial abundance model of the single-stage PNA process is constructed by using an XGBoost model, and each target output corresponds to a model.

[0011] The present disclosure has the following technical effects: by using machine learning to construct a model to predict and elucidate the denitrification performance and functional microbial abundance in the PNA process, the application and improvement of the PNA technology in wastewater treatment are facilitated. By performing quantitative analysis on the model, the potential synergistic effect between characteristics can be explored, thereby helping to find the optimal characteristic range, improve the wastewater treatment efficiency, and obtain the optimal water environment. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0013] Figure 1Schematic diagram of the conceptual framework for NRR and functional microorganism abundance in single-stage PNA process.

[0014] Figure 2 Predicted performance of the ML model for each target: (a) NRR, (b) AnAOB, (c) AOB, (d) NOB, (e) Candidatus Brocadia anammoxidans , (f) Candidatus Kuenenia stuttgartiensis .

[0015] Figure 3 Total SHAP values of input parameters related to (a) NRR, (b) NRR, (c) NOB, (d) NOB, (e) Candidatus Brocadia anammoxidans and (f) Candidatus Kuenenia stuttgartiensis The size of the bubble reflects the size of the total SHAP value, and the dashed line represents the optimal or inhibitory range of the parameter. DETAILED DESCRIPTION

[0016] Explanation of some terms.

[0017]

[0018] The technical solutions of the present application will be described clearly and completely below in combination with the drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0019] Referring to Figure 1 A method for predicting and interpreting the denitrification performance and microbial abundance of a short-cut nitrification-anammox system, comprising the following steps: Step 1. Data collection The quantity and quality of data collected for developing the machine learning model will affect the prediction performance. For example, a literature search was conducted in Scopus and SCIE databases with “partial nitrification and anammox” and “functional microorganisms” as keywords. To ensure the reliability of model prediction and interpretation, a single-stage PNA process was used, thus eliminating the two-stage PNA process. Subsequently, 19 features, respectively , , , , , FA, FNA, NLR, NRR, running days, oxygen ( pH, temperature, hydraulic retention time (HRT), and AOB, NOB, and AnAOB (including...) Candidatus Brocadia anammoxidans and Candidatus Kuenenia stuttgartiensis The abundance of these features is used to determine whether data with these features are valid, and to exclude samples with missing features.

[0020] Step 2. Data Preprocessing The parameters collected as input features from various tabular, text, and graphical data sources are converted and standardized in units. For example, microbial abundance is expressed as a percentage (%), and NLR and NRR are expressed as... Units are as follows: HRT is in hours, temperature is in °C, and other characteristics are in mg / L.

[0021] Next, outlier (or extreme) data is removed from the sample to ensure a relatively uniform distribution of the 19 features. The criteria for removing outlier data are as follows: (1) (1) pH range: limited to 10-2000 mg / L; (2) pH range: limited to 6-9; (3) running days: less than 1000 days; (4) temperature: 10-40°C; (5) NLR: limited to 0-5 kg / L. (6) Abundance of functional microorganisms: 0%–60%.

[0022] Furthermore, to prevent over-contribution and bias, a normalization technique was employed on the input features before modeling. Using a min-max normalization method, each feature is mapped to the range [0, 1], thus proportionally adjusting the input feature values. This, in turn, improves the stability and efficiency of model training.

[0023] It should be noted that normalization can also be performed before removing extreme data.

[0024] The target variables were ultimately constructed as follows: (a) NRR, (b) AnAOB, (c) AOB, (d) NOB, (e) Candidatus Brocadia anammoxidans and (f) Candidatus Kuenenia stuttgartiensis These target variables accurately represent the denitrification performance and functional microbial abundance of a single-stage PNA process.

[0025] Based on empirical domain knowledge, these input features are divided into three groups: (1) nine water quality conditions, namely , , , , FA, FNA, NLR and NRR; (2) Five operating conditions, namely running days, pH, HRT, temperature; and (3) five functional microbial abundances, i.e., AnAOB, AOB, NOB, Candidatus Brocadia anammoxidans and Candidatus Kuenenia stuttgartiensis .

[0026] Step3. Model prediction Four different supervised ML algorithms were adopted in this study: ANN, RF, SVR, and XGBoost. ANN was used as a viable alternative to physical models. RF performs well in handling both classification and regression tasks by building a large number of decision trees. SVR is good at mapping data points to a higher dimensional space and then finding the hyperplane that best separates the data. XGBoost combines multiple weak learners into a strong one and is often used by researchers in water quality prediction studies.

[0027] The dataset was randomly split into two parts: 80% as the training set and 20% as the test set for model evaluation. 10-fold cross-validation was used in this study to minimize the overfitting of the model, with 1 fold for model validation and the remaining 9 folds for model training. This process was repeated 10 times, and the average was taken as the final performance of the model on the training dataset. Bayesian optimization can minimize the target loss of any objective function and has been widely used in hyperparameter tuning in ML models.

[0028] In one embodiment, it was found through experiments that ANN is suitable for predicting denitrification performance, and XGBoost is suitable for microbial prediction.

[0029] Step4. Optimization and evaluation During the model training process, Bayesian optimization was used to adjust and optimize the hyperparameters of the model. The Bayesian optimization process included 10 iterations, and each iteration selected a new combination of hyperparameters based on the previous evaluation results, gradually improving the performance of the model.

[0030] Finally, the test set was used to verify the generality of the model. The standard for model fitting was , root mean square error (RMSE), and mean square error (MSE). Higher , smaller RMSE, and smaller MSE indicate higher reliability. Python 3.6 was used when modeling using ANN, RF, SVM, and XGBoost, mainly using the scikit-learn and Keras software packages.

[0031] Step5. Feature analysis After the ML (Machine Learning) model is established, the SHAP value of each feature is calculated. The working principle of SHAP is to check the difference in predicted values before and after removing the feature. In addition, the interaction information between features is also considered, including all possible removed features. The reason for choosing SHAP to explain the ML model is that it can provide a detailed theoretical basis for any ML model, making the interpretation method consistent and unbiased. The marginal contribution of the 19 features to the prediction target is represented by the SHAP value. The SHAP value of a feature measures whether its contribution is positive or negative. The higher the absolute value of SHAP, the greater its contribution to the prediction of the target. In this study, the SHAP value is generated using an optimization model, and the impact of different features on NRR and functional microbial abundance is analyzed.

[0032] Step6. Quantitative analysis The quantitative analysis method used in the present application helps researchers and decision-makers to fully understand the impact of various input parameters and determine the optimal range of water quality conditions, operating parameters, and microbial community abundance. Quantitative analysis includes controllable input water quality parameters, operating conditions, and microbial parameters. Since there is interaction between input features, it is more reasonable to combine the SHAP values of two features than to use a single SHAP value.

[0033] Step7. Decision making The present application aims to develop a predictive model of nitrogen removal performance (NRR) and functional microbial abundance during single-stage PNA. Artificial neural networks (ANN), random forests (RF), support vector regression (SVR), and extreme gradient boosting (XGBoost) are used to incorporate key water quality and operating parameters into the model to improve the model's interpretability and prediction accuracy. SHAP is then applied to determine the impact of each input feature on ML and the optimal range of environmental parameters, and targeted adjustments are made to achieve targeted operational adjustments. The present application improves the nitrogen removal performance of the system by analyzing key parameters in a single-stage PNA system. By implementing a combined strategy, the activity and abundance of AOB and AnAOB are maintained while inhibiting NOB. By comparing the key parameters of Candidatus Brocadia anammoxidans and Candidatus Kuenenia stuttgartiensis different survival and inoculation environments are determined.

[0034] Based on the above-established model and the collected data, experimental analysis is performed.

[0035]

[0036] Experiment 1: In the prediction process for nitrogen removal performance (NRR), water quality conditions ( , FA, FNA, and NLR), operating conditions (running days, , pH, HRT, and temperature) and functional microbial abundance (AnAOB, AOB, NOB, Candidatus Brocadia anammoxidans and Candidatus ) as input parameters.

[0037] In the prediction process for functional microbial abundance, water quality conditions ( , , , , , NLR, and NRR) and running conditions (running days, , pH, HRT, and temperature) were selected as input features. ML models were established to predict AnAOB (including Kuenenia stuttgartiensis and Figure 2 Figure 3 ), AOB, and NOB as the target to study their abundance and role in the PNA process. The results are shown in Figure 3 , all the optimal ML models are reliable enough to provide satisfactory target prediction according to the selected input features.

[0038] Experiment 2: The present application uses SHAP values for quantitative analysis. Quantitative analysis includes controllable input water quality parameters ( , NLR, FA, and FNA), running conditions (running days, oxygen, pH, and temperature), and microbial parameters (AOB, NOB, and AnAOB). Since there is interaction between input features, it is more reasonable to combine the SHAP values of two features than to use a single SHAP value. The present application lists the optimal environmental conditions for denitrification and functional microbial abundance in a single-stage PNA system.

[0039] To improve the denitrification performance, the interactive analysis of key parameters (FA and pH) and the abundance of specific functional microorganisms (AnAOB and AOB) was carried out. As shown in Figure 3 a, it is recommended to control the pH and FA values in the range of 7.4-8.4 and 20-120 mg / L, respectively. As shown in Figure 3 b, the positive effect of AnAOB and AOB abundance on NRR mainly concentrates in the range of 20-40% and 1-5%, respectively. This indicates that the higher the abundance of AnAOB and AOB is not necessarily better, but a proper balance must be maintained, which may be due to the existence of both cooperation and competition between them.

[0040] To determine the optimal range of using multiple factor combinations to inhibit the growth of NOB, the present study conducted an interactive analysis of methods commonly used in engineering and easy to control (adjusting FA, FNA, and ). As shown in Candidatus Brocadia anammoxidans c, when When NOB is below 0.5 mg / L and FA exceeds 10 mg / L, NOB in It is inhibited under the combined action of FA. For example... Candidatus Kuenenia stuttgartiensis As shown in d, when NOB will be affected when levels are below 0.4 mg / L or when FNA levels exceed 0.2 mg / L. Co-inhibition with FNA.

[0041] In order to determine Figure 3 and Candidatus Brocadia anammoxidans The present invention addresses the optimal living conditions and appropriate inoculum conditions for sludge, in accordance with the present invention. Candidatus Kuenenia stuttgartiensis Adjustable input parameters were analyzed in d and 3d, with particular emphasis on ammonia nitrogen concentration ( and NLR). Candidatus Brocadia anammoxidans and Candidatus Kuenenia stuttgartiensis The optimal conditions differ significantly. Data indicate a significant difference in substrate concentration: when Less than 200 mg / L, NLR less than 0.2 hour, Candidatus Brocadia anammoxidans It grows vigorously, and Candidatus Kuenenia stuttgartiensis Candidatus Brocadia anammoxidans Candidatus Kuenenia stuttgartiensis Figure 2 Figure 3 Figure 3 Figure 3 Figure 3 Candidatus Brocadia anammoxidans Candidatus Kuenenia stuttgartiensis Figure 3 Candidatus Brocadia anammoxidans Candidatus Kuenenia stuttgartiensis Candidatus Brocadia anammoxidans Candidatus Kuenenia stuttgartiensis Candidatus Bro Then like Above 400 mg / L, NLR above 0.4 The environment.

[0042] Through the above description of the embodiments, those skilled in the art can clearly understand that the method of this disclosure can be implemented by means of software plus necessary general-purpose hardware, and of course, it can also be implemented by special-purpose hardware, including application-specific integrated circuits, dedicated CPUs, dedicated memory, dedicated components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or dedicated circuits. However, for the purposes of this disclosure, software program implementation is more often a preferred implementation method.

[0043] Although the embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this disclosure is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of this disclosure, and all of these are within the scope of protection of this disclosure.

Claims

1. A method for analyzing denitrification performance of a single-stage PNA system, characterized by, The method comprises the following steps: The single-stage PNA process is adopted, data with preset characteristics are selected as samples, and the samples are pretreated, the preset characteristics including influent ammonia nitrogen concentration, free ammonia concentration, free nitrite concentration, influent nitrogen load, running days, pH value, temperature, hydraulic retention time, ammonia oxidizing bacteria, nitrosifying bacteria, anaerobic ammonia oxidizing bacteria, Candidatus Brocadia, Candidatus Kuenenia nitrogen removal rate; The denitrification performance prediction model of the single-stage PNA process is trained, and the influent ammonia nitrogen concentration, free ammonia concentration, free nitrite concentration, influent nitrogen load, running days, pH value, temperature, hydraulic retention time, ammonia-oxidizing bacteria, nitrosifying bacteria, and anammox bacteria, Candidatus Brocadia, Candidatus Kuenenia as input and the nitrogen removal rate as output; The trained denitrification performance prediction model of the single-stage PNA process is applied for inference, and SHAP value quantitative analysis is used to explain the influence of different features on the nitrogen removal rate.

2. The method of claim 1, wherein, The preprocessing includes converting and standardizing the units of preset features, eliminating abnormal data, and normalizing the feature values of preset features.

3. The method of claim 1, wherein, The denitrification performance prediction model of the single-stage PNA process is constructed by using a neural network ANN, and the hyperparameters of the Bayesian optimization model are used in training.

4. The method of claim 1, wherein, The SHAP value quantitative analysis includes calculating the SHAP value of a single feature and the SHAP value of two features, the former is used to explain the influence of each feature, and the latter is used to quantify the optimal range of the feature.

5. A method of analyzing the abundance of functional microorganisms in a single-stage PNA system, characterized by, The method comprises the following steps: The single-stage PNA process is adopted, data with preset characteristics are selected as samples, and the samples are pretreated, the preset characteristics including influent ammonia nitrogen concentration, effluent ammonia nitrogen concentration, effluent nitrite concentration, effluent nitrate nitrogen concentration, effluent total nitrogen concentration, free ammonia concentration, free nitrite acid concentration, influent nitrogen load, NRR, running days, , pH value, temperature, hydraulic retention time, anaerobic ammonia oxidation bacteria, ammonia oxidation bacteria, nitrite bacteria, Candidatus Brocadia, Candidatus Kuenenia ; The microbial abundance model of the single-stage PNA process is constructed for training, and the influent ammonia nitrogen concentration, effluent ammonia nitrogen concentration, effluent nitrite concentration, effluent nitrate nitrogen concentration, effluent total nitrogen concentration, free ammonia concentration, free nitrite acid concentration, influent nitrogen load, NRR, running days, , pH value, temperature, and hydraulic retention time are taken as inputs, and the abundance of each functional microorganism is taken as a target output; The trained microbial abundance model of the single-stage PNA process is applied for inference, and SHAP value quantitative analysis is used to explain the influence of different features on the abundance of each functional microorganism.

6. The method of claim 5, wherein, The preprocessing includes converting and standardizing the units of preset features, eliminating abnormal data, and normalizing the feature values of preset features.

7. The method of claim 5, wherein, The microbial abundance model of the single-stage PNA process is constructed by using an XGBoost model, and each target output corresponds to a model.

8. The method of claim 5, wherein, The SHAP value quantitative analysis includes calculating the SHAP value of a single feature and the SHAP value of two features, the former is used to explain the influence of each feature, and the latter is used to quantify the optimal range of the feature.

Images