Short-cut nitrification-anammox process system and intelligent prediction and regulation method thereof

Abstract

The application discloses a short-range nitrification-anaerobic ammonia oxidation process system and an intelligent prediction and regulation method thereof. In the method, multi-source operation data under different working conditions are collected to construct a machine learning data set covering water quality, operation and operation parameters; after hierarchical preprocessing, feature engineering and training / test set division, a multi-model prediction system including KNN, RF, XGBoost, GRU, LSTM and Transformer is established, and the optimal model is screened in combination with five-fold cross-validation and multi-dimensional performance indexes; further, SHAP and PDP are used to analyze the global contribution and local effect of key parameters on NRE, identify core regulation factors, and accordingly construct a scene adaptive optimization control strategy for multi-scale working condition fluctuations. The application can overcome the problems of single model, insufficient generalization ability and unclear machine explanation of the traditional method, realize precise regulation under different influent loads and reactor types, and improve the operation stability, denitrification efficiency and working condition adaptability of the PNA system.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction and control technology for wastewater treatment, specifically to a short-cut nitrification-anaerobic ammonium oxidation process system and its intelligent prediction and control method. Background Technology

[0002] Excessive nitrogen discharge into sewage and wastewater causes a series of ecological safety problems, such as eutrophication and algal blooms, endangering drinking water safety and public health. The traditional nitrification-denitrification process is widely used as the mainstream biological nitrogen removal technology, but it generally suffers from high energy consumption, high greenhouse gas emissions, low nitrogen removal efficiency, and large land area requirements. Anaerobic ammonia oxidation (ANAO) technology overturns the traditional nitrification-denitrification route, eliminating the need for organic carbon sources and intense aeration, fundamentally overcoming the bottlenecks of the traditional nitrification-denitrification process. However, the acquisition and stable supply of nitrite, the sole electron acceptor for anaerobic ammonia oxidizing bacteria, is a core drawback restricting the widespread engineering application of ANAO. The proposed short-cut nitrification-anammox process perfectly solves the substrate supply problem in anammox, thereby resolving the niche competition between anammox bacteria and heterotrophic bacteria. The short-cut nitrification-anammox coupled system not only improves nitrogen utilization and reduces energy, chemical, and sludge production, but also reduces greenhouse gas emissions and avoids the generation of N2O during denitrification, making it a key pathway to achieve "energy self-sufficiency" in wastewater treatment. However, the short-cut nitrification-anammox system is influenced by numerous factors, and its complex multi-factor coupling characteristics bring many challenges to its long-term stable operation. In practical applications, there are various limiting factors, such as the involvement of multiple biochemical reactions including partial nitrification and anammox, numerous influencing factors, and complex nonlinear relationships between influent water quality and control factors. Furthermore, anammox bacteria grow slowly and are sensitive to environmental factors such as dissolved oxygen and temperature, easily leading to process instability. To address these challenges and overcome existing problems affecting stable operation, this paper attempts to model the complex coupled process, explore potential mechanisms and control factors, and optimize the wastewater treatment process.

[0003] Traditional activated sludge process mechanistic models (ASM) are poorly adaptable to fluctuations in water quality and operating conditions, rely on idealized assumptions, and struggle to accurately characterize the dynamic nonlinear features of the complex biological nitrogen removal process in short-cut nitrification-anaerobic ammonium oxidation (LTAO) systems. In recent years, data-driven machine learning (MLM) technology has rapidly developed and gained favor among researchers in various disciplines, including environmental science. Compared to traditional activated sludge process mechanistic models, MLM does not require pre-established complex biochemical mechanisms and can autonomously uncover key correlations from actual operating data. It possesses strong nonlinear fitting and operating condition adaptive capabilities, effectively compensating for the shortcomings of traditional mechanistic models. Utilizing the data-driven characteristics of MLM to perform full-process dynamic feature analysis of multi-scale LTAO systems enables multi-objective prediction and intelligent control of the coupled system, enhancing its nitrogen removal performance and stable operation.

[0004] Current research on wastewater treatment processes based on machine learning largely focuses on end-of-pipe effluent prediction, lacking analysis of the dynamic characteristics of the entire process and targeted model optimization for short-cut nitrification-anammox processes. Furthermore, the control of short-cut nitrification-anammox processes relies heavily on human experience, leaving frontline operators without systematic and reliable control guidance. Existing technologies primarily focus on purely rule-based control, emphasizing hardware systems and lacking data-driven approaches and differentiated control requirements at different operational stages. Related research employs particle swarm optimization to construct intelligent control systems combined with short-cut denitrification-anammox processes. This method uses multi-dimensional data analysis and modeling of system operating parameters to build and optimize a predictive model for the nitrogen removal performance of the short-cut denitrification-anammox system, identifying key factors affecting system performance. However, this technology is specific to the short-cut denitrification-anammox process and cannot share characteristic control principles with the short-cut nitrification-anammox process.

[0005] In summary, no relevant technology has yet proposed key regulatory factors affecting the efficiency and stability of multi-scale short-range nitrification-anaerobic ammonium oxidation process systems, thus failing to achieve precise control and stable operation of the coupled process systems. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent predictive control method for short-cut nitrification-anammox process systems that addresses the technical problems of insufficient system denitrification efficiency and operational stability under fluctuating multi-condition conditions.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A smart control method for a short-cut nitrification-anaerobic ammonium oxidation process includes the following steps: Step S1: Collect sample point data of short-cut nitrification-anaerobic ammonium oxidation process under different operating conditions based on existing research, mainly including water quality parameters, operating parameters and control parameters, and construct a multi-source dataset suitable for machine learning; Step S2: Preprocess the dataset, implement feature engineering and input variable screening, including handling missing and outlier values, category information transformation, data standardization, correlation analysis, and dataset splitting into training and test sets in an 8:2 ratio to provide a high-quality data foundation for the model. Simultaneously, 5-fold cross-validation is adopted. Step S3: Based on the constructed dataset, machine learning and time series-based deep learning are used for training and optimization to predict the nitrogen removal performance of the short-cut nitrification-anaerobic ammonium oxidation process, and the output performance of the model on the test set is evaluated to select the optimal model. Step S4: Based on SHAP and PDP interpretability analysis, perform importance analysis on the feature parameters and screen out the key control factors that affect the efficiency and stability of the system. Step S5: Integrate prediction and uncertainty analysis, optimize and regulate the process using key control factors, and construct a whole-process optimization prediction and regulation strategy for the short-range nitrification-anaerobic ammonium oxidation process.

[0008] Preferably, in step S2, the handling of missing and outlier values ​​includes calculating the median of the corresponding feature values, selectively filling in values ​​after judging all feature values, and the judgment process is as follows: The input features in the dataset to be preprocessed are divided into categorical features and continuous numerical features. The categorical features include reactor type and sludge form, while the continuous numerical features include volume, operating days, pH, aeration ratio, DO, aeration rate, HRT, temperature, influent organic carbon, influent ammonia nitrogen, and influent nitrite. Subsequently, missing values ​​of continuous numerical features were filled with NaN using the median filling strategy, and missing values ​​of categorical features were filled with the mode filling strategy to preserve the original distribution characteristics of the data. Outlier cleaning employs a mild cleaning strategy, calculating the 1st percentile as the minimum and the 99th percentile as the maximum for each target variable, defining a reasonable data range, and filtering out 1% of extreme outliers.

[0009] Preferably, in step S2, feature encoding and normalization are performed on the supplemented dataset, including encoding and standardizing different types of features separately. The process is as follows: For numerical data, standard normalization is performed using the StandardScaler method. The process is as follows: ; in, denoted as the standardized data of the numerical features of the sample, x represents the original data of the numerical features of the sample, μ is the mean of each numerical feature data, and σ is the standard deviation of each numerical feature data; One-hot encoding is used for categorical data to convert discrete categories into numerical vectors that the model can recognize. At the same time, an unknown category ignoring strategy is set to improve the model's compatibility with unknown category data.

[0010] Preferably, in step S2, the dataset is split according to the machine learning model evaluation requirements. The preprocessed feature matrix and the target variable are randomly split into mutually exclusive datasets in an 8:2 ratio, with 80% of the data used as the training set for model training and 20% of the data used as the test set for model testing. At the same time, 5-fold cross-validation is used to improve the stability of the model.

[0011] Preferably, in step S3, the water quality parameters, operating parameters, and operational parameters in the training set are used as the training basis. Different machine learning methods are selected for training, and the most reasonable model is screened and optimized. The coefficient of determination R is used. 2 The predictive performance of the prediction models is compared by comprehensively evaluating the root mean square error (RMSE) and mean absolute error (MAE) to select the optimal prediction model.

[0012] Preferably, in step S3, the process of comparing the prediction performance of the prediction models is expressed as follows: ; ; ; in, The sample index in the test set is The actual nitrogen removal efficiency value, The sample index in the test set is The predicted nitrogen removal efficiency value is ȳ, which is the average of the actual nitrogen removal efficiency of all samples in the test set, and n is the number of samples in the test set.

[0013] Preferably, in step S4, the process of feature data importance analysis includes using Python's SHAP package to calculate the SHAP contribution value of each feature, quantifying the global impact of a single feature on the prediction result of nitrogen removal efficiency; combining single-feature and dual-feature interaction partial dependency analysis to verify the feature marginal effect; and comprehensively considering the magnitude of SHAP contribution and partial dependency response features to screen out the core key control parameters that affect process performance.

[0014] Preferably, in step S5, the process of constructing the optimal operating condition control strategy includes: selecting key process parameters as control variables, using intelligent optimization algorithms to iteratively solve for the optimal parameter combination, and forming a customized control scheme for improving nitrogen removal efficiency in wastewater treatment for multi-scale operating condition fluctuation scenarios.

[0015] This invention also provides a control system for a smart control method of short-range nitrification-anaerobic ammonium oxidation process, including a data acquisition module, a data preprocessing module, a model training and optimization module, an interpretability analysis module, and an optimization prediction and control module; The data acquisition module is used to collect sample point data of short-range nitrification-anaerobic ammonium oxidation process under different operating conditions based on existing research. The data mainly includes water quality parameters, operating parameters and operational parameters, and constructs a multi-source dataset suitable for machine learning. The data preprocessing module is used to preprocess the dataset, implement feature engineering and input variable screening, including missing value and outlier handling, category information conversion, data standardization, correlation analysis, and dataset partitioning, which splits the dataset into training and test sets in an 8:2 ratio, and simultaneously uses 5-fold cross-validation. The model training and optimization module is used to train and optimize the model based on the constructed dataset using machine learning and time series-based deep learning, predict the denitrification performance of the short-cut nitrification-anaerobic ammonium oxidation process, evaluate the output performance of the model on the test set, and select the optimal model. The interpretability analysis module is used to perform importance analysis on feature parameters based on SHAP and PDP interpretability analysis, and to screen key control factors that affect the efficiency and stability of the system. The optimization prediction and control module is used to integrate prediction and uncertainty analysis, optimize and control the process based on key control factors, and construct an optimized control strategy for the entire short-range nitrification-anaerobic ammonium oxidation process.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention integrates multi-source heterogeneous data such as sewage treatment category configuration, hydraulic conditions, and water quality parameters to construct a dataset, breaking through the limitation of insufficient representation by a single data source; at the same time, it establishes a multi-model comparison and selection system between traditional machine learning and time-series deep learning, including six mainstream algorithms such as KNN, RF, XGBoost, GRU, LSTM, and Transformer, and selects the optimal model through multi-index joint cross-validation to accurately represent the strong nonlinear correlation between sewage treatment process and nitrogen removal efficiency; (2) This invention abandons the single feature analysis method and adopts a two-dimensional interpretable analysis of SHAP global importance and PDP single / dual feature interaction local effect to quantify the global contribution of features, reveal the marginal influence and coupling effect of features, and accurately identify key control factors; it forms specific control schemes for different influent water quality fluctuations, and strengthens the adaptability and robustness of strategy working conditions. (3) This invention preserves the original valid sample information to the greatest extent. The dataset integrates multidimensional datasets such as water quality parameters, operational parameters, and running parameters. The quality of the input data is ensured through hierarchical preprocessing of numerical / categorical features. The optimal machine learning prediction models KNN, RF, and XGBoost are selected by comparing and selecting mainstream algorithms from multiple spectrums. 2 The system achieves a nitrogen removal efficiency of over 0.9 and maintains an MAE of around 5. It accurately identifies key control factors such as temperature, influent ammonia nitrogen, DO, and HRT, and significantly improves nitrogen removal efficiency after intelligent optimization and control. This effectively mitigates operational instability caused by large fluctuations in water quality and environmental parameters under different operating conditions, while also optimizing process chemical consumption and energy costs. It solves the technical shortcomings of traditional solutions, such as limited model types, easy loss of effective information during data preprocessing, weak model interpretability, and poor universality of control strategies. It possesses comprehensive technical advantages, including high data fidelity, strong reliability of multi-model selection, high prediction accuracy, complete and interpretable mechanisms, and excellent adaptive control under operating conditions.

[0017] To address the challenges of maintaining stable operation of short-cut nitrification-anammox processes under multi-scale conditions, and the lack of clarity regarding coupling parameters and key control factors, this invention provides an optimization and intelligent control model for the short-cut nitrification-anammox process. This model can predict the nitrogen removal performance of short-cut nitrification-anammox systems under multiple operating conditions, identify the optimal integrated parameters of the process, and automatically and effectively extract key features characterizing the process state from multi-dimensional, nonlinear, and strongly coupled time-series data of the short-cut nitrification-anammox process. It then constructs a process optimization control strategy to achieve precise control and operation of the process. Attached Figure Description

[0018] Figure 1 A schematic diagram of the system construction process of an intelligent control method for a short-cut nitrification-anaerobic ammonium oxidation process is provided for embodiments of the present invention; Figure 2 A comparison chart of predicted and actual NRE values ​​of the model output features provided for embodiments of the present invention; Figure 3 A heatmap showing the correlation between key features provided for embodiments of the present invention; Figure 4 A SHAP feature importance ranking diagram provided for embodiments of the present invention; Figure 5 A PDP feature marginal effect diagram provided for embodiments of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0020] Combination Figures 1 to 5 The construction of an intelligent control model for a short-cut nitrification-anaerobic ammonium oxidation process in this embodiment includes the following steps: Step S1: Dataset Construction We obtained publicly available multi-source data on short-range nitrification-anaerobic ammonium oxidation processes and constructed a dataset suitable for machine learning through data fusion and corresponding preprocessing.

[0021] PlotDigitizer was used to extract data from the short-cut nitrification-anaerobic ammonium oxidation process system. The system data consisted of data from all sample points under different scale conditions, including operating parameters, control parameters, and water quality parameters.

[0022] The operating parameters include reactor type, sludge type, number of operating days, and temperature.

[0023] Operating parameters include reactor volume, pH, aeration ratio, DO, aeration rate, and HRT.

[0024] Water quality parameters include influent ammonia nitrogen concentration, influent nitrite concentration, and influent organic carbon concentration.

[0025] The above process was used to obtain systematic data on short-range nitrification-anaerobic ammonium oxidation processes under various operating conditions. The data were then merged according to the time dimension of the samples to construct a matrix containing thousands of sample points, which served as a multi-source dataset for subsequent machine learning.

[0026] Step S2: Dataset Preprocessing The data preprocessing process includes five core steps: handling missing and outlier values, class information encoding and conversion, data feature standardization, feature correlation analysis, and dataset partitioning. These steps aim to improve data quality and the effectiveness of model training, as detailed below: (1) Handling outliers and missing values Before model training, water quality parameters (influent ammonia nitrogen concentration, influent nitrite concentration, and influent organic carbon concentration), operating parameters (reactor volume, pH, aeration ratio, DO, aeration rate, and HRT), and operational parameters (reactor type, sludge type, operating days, and temperature) were used as the core feature set. Effluent ammonia nitrogen concentration, effluent nitrite concentration, effluent nitrate concentration, and NRE were used as target variables. A mild truncation strategy was adopted for outlier handling of the target variable columns: the 1st percentile (q1) and 99th percentile (q3) of each target variable column were calculated, and extreme outliers less than q1 or greater than q3 were removed, retaining the remaining samples to preserve the original data distribution characteristics to the greatest extent possible. For missing values ​​in the feature set, median imputation was used for numerical features, and mode imputation was used for categorical features to avoid model training interruptions due to missing values, while ensuring that the imputed values ​​closely matched the data's distribution pattern.

[0027] (2) Category information encoding conversion Since the "reactor category" and "sludge form" in the feature dataset are non-numerical categorical features, they cannot be directly input into the machine learning model for training. Therefore, a categorical feature encoding pipeline was constructed based on Pandas and Scikit-learn tools: First, missing values ​​of the categorical features were filled in using the mode imputation strategy. Then, one-hot encoding was used to convert the two types of features into numerical values, generating an independent binary column (0 or 1) for each category dimension. The column value is 1 when the sample matches the corresponding category, and 0 otherwise. After encoding, a feature name mapping table is automatically generated, integrating the encoded features such as "reactor category_XX" and "sludge form_XX" with the original numerical features to form a unified numerical feature matrix, ensuring that the model can directly identify and calculate them.

[0028] (3) Data feature standardization After completing the categorical feature encoding, standardization is performed on all numerical features (including water quality parameters and operational parameters). The StandardScaler method is used to eliminate dimensional differences between different features: first, the global mean of each numerical feature is calculated ( ) and standard deviation ( Then, the eigenvalues ​​are converted into a standard distribution with a mean of 0 and a standard deviation of 1 using the following formula: ; The above, This represents the standardized data of the numerical features of the sample, where x represents the original data of the numerical features of the sample. The average value for each numerical feature data. The standard deviation of each numerical feature data.

[0029] The same standardization strategy is used for the target variables (effluent ammonia nitrogen concentration, effluent nitrite concentration, effluent nitrate concentration, NRE) to ensure scale matching between the target variables and feature variables during model training, thereby improving the model convergence speed and fitting accuracy.

[0030] (4) Feature correlation analysis Based on the Pearson correlation coefficient method, the linear correlation between numerical features was calculated. The correlation matrix of 10 core features, including influent organic carbon concentration, influent ammonia nitrogen concentration, temperature, DO, HRT, and pH, was analyzed, with a significant correlation threshold of 0.8 for the absolute value of the correlation coefficient. The distribution of correlations among features was visualized using heatmaps, and highly correlated feature pairs were selected and their correlation coefficients recorded. Considering the differences in feature sensitivity between different models such as KNN and Transformer, feature integrity was preserved to adapt to the model's feature learning capabilities, while also providing a foundation for subsequent feature importance analysis (SHAP analysis, feature importance scoring).

[0031] like Figure 3 As shown, the core features exhibit no strong collinearity, the overall features demonstrate good independence, and the model's interpretability is pure. Two moderately positively correlated feature pairs were identified: influent ammonia nitrogen concentration and HRT (r=0.51), and influent ammonia nitrogen concentration and influent organic carbon concentration (r=0.45). The remaining features showed no significant linear correlation with other process features, indicating strong independence among the core features. This aligns with the subsequent multi-model SHAP analysis, which revealed their dominant role in nitrogen removal efficiency.

[0032] (5) Dataset splitting and cross-validation A strategy combining hierarchical random splitting and five-fold cross-validation is employed: First, the preprocessed dataset is randomly split into a training set (80%) and a test set (20%) in an 8:2 ratio, with a fixed random seed (SEED=42) to ensure reproducibility of the splitting results. Then, five-fold cross-validation (KFold, n_splits=5) is introduced, further dividing the training set into five mutually exclusive subsets. The model is trained sequentially using four subsets as the training set and one subset as the validation set, with multiple rounds of validation used to evaluate model stability. The training set is used for iterative optimization of model parameters, the test set is used to evaluate the model's generalization ability, and the validation set is used to monitor the risk of overfitting during training, ensuring stable predictive performance across different data distributions.

[0033] The data preprocessing described above is a crucial preliminary step in machine learning, providing high-quality multi-source datasets for subsequent model building. The core advantages of this preprocessing workflow are: lightweight outlier handling balances data integrity with outlier removal requirements, avoiding data distribution distortion caused by excessive cleaning; one-hot encoding of categorical features eliminates data format differences, allowing categorical and numerical feature data to collaboratively participate in model training, significantly improving the prediction accuracy of multi-source models for core target variables such as NRE; data standardization unifies feature dimensions, further improving data quality, eliminating feature differences, and accelerating model training; and the combination of correlation analysis and cross-validation preserves the integrity of feature dimensions while ensuring stable and robust model performance through multiple rounds of validation.

[0034] Step S3: Prediction Model Construction Based on the preprocessed training set, multiple machine learning prediction models are constructed to form a multi-algorithm comparison and verification system. The specific process includes selecting machine learning methods, establishing multi-lineage prediction models, multi-dimensional model evaluation and selection, and optimizing models by combining feature interaction patterns.

[0035] (1) Machine learning model selection Based on the learning objectives and dataset data types, eight prediction models using traditional machine learning and deep learning are first constructed in parallel. After initial model training and comparison, this invention employs six machine learning methods, divided into two main categories: The first category is traditional machine learning methods, including three algorithms: K-Nearest Neighbors (KNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). These algorithms cover three classic machine learning frameworks: nonparametric fitting, ensemble trees, and gradient boosting, and can efficiently capture local nonlinear features and threshold effects in process data. The second category is deep learning methods, including three algorithms: Transformer, Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM). These cover two major deep learning architectures: self-attention mechanism and gated recurrent temporal modeling, which can accurately uncover high-order interaction relationships and temporal correlation features among multiple features.

[0036] (2) Establishment of multi-lineage prediction model For machine learning methods of different lineages, prediction models are built using appropriate open-source frameworks: For three traditional machine learning models—KNN, RF, and XGBoost—the open-source Python machine learning library Scikit-learn (Sklearn) is used to build and train the models. Sklearn is an open-source machine learning toolkit based on Python, containing complete machine learning algorithm implementations, data preprocessing modules, model evaluation and cross-validation tools, enabling efficient initialization, training, and tuning of traditional machine learning models. In this embodiment, Sklearn is used to initialize the parameters of the traditional model, fit the training set, and output the prediction results.

[0037] For three types of deep learning models—Transformer, GRU, and LSTM—the open-source Python deep learning framework PyTorch was used to build the network structure and train the models. PyTorch supports dynamic computation graphs and GPU-accelerated computation, adapting to complex feature interaction capture requirements. In this embodiment, a multi-head self-attention network with a 3-layer encoder was built for the Transformer model, and a gated recurrent neural network was built for the GRU and LSTM models. The AdamW optimizer was used to iteratively update the model parameters, completing the training of the multi-output prediction model.

[0038] (3) Intelligent optimization of model hyperparameters For the six types of prediction models constructed, a differentiated intelligent hyperparameter optimization strategy was adopted, combined with multi-dimensional performance evaluation to complete the selection of the optimal model. The specific process is as follows: For traditional machine learning models, RF employs a two-stage optimization scheme of "random search interval localization + fine-grained grid search for precise tuning" to iteratively optimize core hyperparameters such as the number of decision trees, maximum depth, and minimum number of split samples. XGBoost combines feature engineering for pre-screening, grid search traversal, and early stopping mechanisms to achieve optimal matching of parameters such as the number of weak learners, learning rate, and tree depth while avoiding overfitting. KNN uses a lightweight optimization strategy of "neighbor-safe search + distance-weight probing" to select the optimal number of neighbors and sample weight configuration with the prediction accuracy of the test set as the core indicator. For deep learning models, Transformer uses a five-fold cross-validation combined with a learning rate combination scheduling strategy of "linear warm-up + cosine annealing" to iteratively optimize network structure parameters such as the number of encoder layers, the number of attention heads, and the dimension of hidden layers. GRU and LSTM, two types of recurrent neural networks, respectively employ gradient pruning, layer normalization, adaptive learning rate decay, and early stopping mechanisms adapted to the characteristics of temporal modeling to effectively control the risk of overfitting while ensuring model convergence efficiency.

[0039] (4) Multidimensional model evaluation and selection The five-fold cross-validation method was used to compare and evaluate the predictive performance of each prediction model on the test set, and the optimal prediction model was selected by combining multi-dimensional evaluation indicators.

[0040] Specifically, the coefficient of determination R is used. 2 The predictive performance of a model is comprehensively evaluated using four metrics: root mean square error (RMSE), mean absolute error (MAE), and so on. Among these, the coefficient of determination (R²) is the most significant. 2 The four metrics are: Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). RMSE measures the overall deviation between predicted and actual values; a smaller value indicates higher prediction accuracy. MAE measures the average level of prediction error; a smaller value indicates stronger prediction stability. The formulas for these four metrics are as follows: ; ; ; in, The sample index in the test set is The actual nitrogen removal efficiency (NRE) value, The sample index in the test set is The nitrogen removal efficiency value predicted by the prediction model. It is the average of the actual nitrogen removal efficiency of all samples in the test set, where n is the number of samples in the test set.

[0041] Specifically, based on the comprehensive results of four evaluation indicators, the performance of each model is ranked, and the optimal prediction model is selected. Table 1 shows the prediction performance results of each prediction model on the test set. In this embodiment, the RF model test set R... 2 =0.929, the best prediction performance (e.g. Figure 2 As shown in Table 1, this invention's core prediction model system achieves the strongest multi-target simultaneous prediction capability and generalization stability.

[0042] Table 1 Performance of different prediction models

[0043] Based on the above results, this invention constructs six sets of multi-objective prediction models for short-cut nitrification-anaerobic ammonium oxidation processes, conducts multi-dimensional performance quantitative evaluation and horizontal benchmarking, and determines the RF machine learning model as the optimal model for predicting the nitrogen removal efficiency of the evaluation system in this step. The model's coefficient of determination R0 is verified on the test set. 2The accuracy can reach approximately 0.930, with a mean squared error (MAE) as low as below 5, achieving high-precision prediction under the simplest ensemble framework and demonstrating good overall prediction performance. Meanwhile, the KNN model, with a mean absolute error (MAE) of 4.643, is the model with the smallest local fitting error among the six models, and the XGBoost model has the highest R... 2 =0.917, which also has high-precision prediction. Together, the three constitute the first-tier high-precision model system for predicting process nitrogen removal efficiency. The prediction accuracy is significantly better than GRU, LSTM and Transformer recurrent neural network models. Deep learning models may be more suitable for training and learning on large sample datasets.

[0044] This solution fully integrates multi-source heterogeneous data on influent water quality components, operating parameters, reactor configuration characteristics, and sludge operating conditions of the short-cut nitrification-anaerobic ammonium oxidation process. It constructs a comprehensive predictive model benchmarking system that combines traditional machine learning and deep learning. Through five-fold cross-validation using multiple indicators such as coefficient of determination, mean absolute error, and root mean square error, the optimal model is scientifically selected. This solution simultaneously considers key control parameters identified by machine learning and the unique operating scenarios of different influent loads and reactor types. It customizes optimal control strategies for improving nitrogen removal efficiency for different operating conditions, effectively solving the problems of single-type models and poor adaptability in traditional solutions, and significantly improving the scenario adaptability and operational stability of process control strategies.

[0045] Step S4: Importance Analysis of Two-Dimensional Features of SHAP and PDP The purpose of this step is to conduct a global-local dual-dimensional analysis of the feature data, accurately identify the core control parameters affecting the short-cut nitrification-anaerobic ammonium oxidation process, reveal the mechanism and interaction law of the feature's effect on denitrification performance, and provide a quantitative basis for subsequent intelligent process control.

[0046] Based on the optimal RF prediction model selected in step S3 above, this invention employs a two-dimensional analysis method combining SHAP global feature importance analysis and PDP partial dependency local effect analysis to complete the quantitative evaluation of feature importance and the analysis of its mechanism of action. Specifically, the Shapley Additive Explanations (SHAP) method is a game theory-based model interpretability method that decomposes the prediction results of the learned model into a weighted sum of the marginal contributions of each input feature, taking into account the average contribution level of features across all feature combinations, thus achieving unbiased quantification of the global importance of features. The Partial Dependence Plot (PDP) is used to characterize the marginal influence of a single or two features on the model's prediction results, visually revealing the nonlinear response relationship between features and nitrogen removal efficiency, as well as the synergistic or antagonistic interaction effects among multiple features. This invention uses the SHAP package in Python to calculate and visualize SHAP values, and uses a self-written partial dependency analysis function to draw PDP curves and interaction heatmaps, as shown below. Figure 4 and Figure 5 As shown.

[0047] Based on the combined importance ranking of SHAP and the marginal effect analysis of PDP, the characteristics are divided into three levels of key parameters: the first level of core dominant parameters are influent ammonia nitrogen, COD, and operating parameters; the second level of key operational control parameters are DO, pH, HRT, and temperature; and the third level of important influencing parameters include sludge form and reactor type. These tiered key parameters provide a precise priority basis for the intelligent control of the short-cut nitrification-anaerobic ammonia oxidation system. In actual engineering operation, the system operating conditions can be adjusted specifically according to the importance level and effect of different characteristics to optimize and improve NRE (Nutrition Reduction).

[0048] The aforementioned key parameters provide precise priority criteria for the intelligent control of short-cut nitrification-anaerobic ammonia oxidation systems. In actual engineering operations, system operating conditions can be adjusted according to the importance level and operational patterns of different characteristics to optimize and improve nitrogen removal efficiency. For example, regarding the core parameter of operating days, the project should ensure long-term stable operation of the system, avoiding frequent start-ups and shutdowns and significant fluctuations in operating conditions, ensuring the process system smoothly overcomes the low point of the start-up period, and achieving microbial community maturation and improved nitrogen removal efficiency. For the secondary key parameters of DO, pH, HRT, and temperature, joint control should be carried out in conjunction with the system operation stages. In the initial stage of process operation, DO and pH should be strictly controlled, while in the maturation stage, nitrogen removal performance can be further enhanced by optimizing HRT and temperature. Regarding the influent ammonia nitrogen concentration, the influent load should be dynamically adjusted according to the system operation stages to ensure stable nitrogen removal capacity of the system.

[0049] Step S5: Optimize the construction of control strategies The purpose of this step is to integrate prediction and uncertainty analysis, optimize and regulate key control factors, and construct an optimized control strategy for the entire short-range nitrification-anaerobic ammonium oxidation process.

[0050] Specifically, the key control factors selected in step S4 are used as regulation variables, including core process parameters such as influent ammonia nitrogen concentration, DO, pH, HRT, and temperature. With maximizing nitrogen removal efficiency as the optimization objective, an intelligent optimization algorithm is used to iteratively solve for the optimal parameter combination. During the optimization process, the optimal prediction model constructed in step S3 is comprehensively integrated to rapidly predict and evaluate the system's nitrogen removal performance under different parameter combinations. Furthermore, uncertainty analysis is combined to assess the robustness of the regulation strategy under various operating condition fluctuation scenarios.

[0051] For multi-scale operating condition fluctuation scenarios, including influent load fluctuations, seasonal temperature changes, and operational phase transitions, customized control schemes for improving nitrogen removal efficiency in wastewater treatment are developed. For example, during the system startup phase, DO and pH are strictly controlled at low levels to ensure rapid enrichment of anaerobic ammonia-oxidizing bacteria; during the system maturity phase, nitrogen removal performance is further enhanced by optimizing HRT and temperature; and for influent load fluctuation scenarios, aeration rate and influent flow rate are dynamically adjusted to ensure stable nitrogen removal capacity of the system.

[0052] By constructing the above-mentioned optimized control strategies, intelligent closed-loop management of the short-cut nitrification-anaerobic ammonia oxidation process was realized, from data acquisition, model prediction, feature analysis to optimized control, effectively improving the denitrification efficiency and operational stability of the system under different operating conditions.

[0053] This invention constructs a multi-source, heterogeneous, multidimensional dataset integrating influent water quality parameters, operating parameters, reactor configuration characteristics, and sludge condition types for a short-cut nitrification-anaerobic ammonium oxidation process. It then builds six types of nitrogen removal efficiency prediction models covering both traditional machine learning and deep learning technologies. Through multi-dimensional performance cross-validation, the optimal RF model has a test set determination coefficient R0. 2 The accuracy can reach 0.929, and a high-precision first-tier model system supplemented by KNN and XGBoost models has been formed. By combining SHAP global feature importance analysis and PDP single / double feature marginal effect analysis, DO, pH, HRT and temperature are accurately identified as the core regulatory factors affecting the nitrogen removal performance of the process, and the nonlinear response law and synergistic interaction mechanism between each feature and nitrogen removal efficiency are fully revealed.

[0054] This invention effectively solves the problems of limited model types, insufficient generalization and adaptability to operating conditions, and unclear characteristic mechanisms in traditional short-cut nitrification-anaerobic ammonium oxidation process performance prediction methods. With the support of a large amount of multi-scale operating condition data, and based on the core control parameters identified by machine learning, it customizes exclusive denitrification efficiency improvement control schemes for actual operating scenarios with different influent loads and reactor types. This significantly improves the scenario adaptability and operational stability of the process operation strategy. It has the technical advantages of high multi-source data fusion, excellent model prediction accuracy, strong interpretability, and wide operating condition adaptability. It provides a complete technical path and data support for the intelligent operation control, denitrification performance optimization, and long-term stable operation of the short-cut nitrification-anaerobic ammonium oxidation process.

[0055] This embodiment also provides a control system for a short-range nitrification-anaerobic ammonia oxidation process intelligent control method, including a data acquisition module, a data preprocessing module, a model training and optimization module, an interpretability analysis module, and an optimization prediction and control module.

[0056] The data acquisition module is used to collect sample point data of short-cut nitrification-anaerobic ammonium oxidation process under different operating conditions based on existing research. The data mainly includes water quality parameters, operating parameters and control parameters, and constructs a multi-source dataset suitable for machine learning. The data preprocessing module is used to preprocess the dataset, implement feature engineering and input variable screening, including handling missing and outlier values, class information transformation, data standardization, correlation analysis, and dataset partitioning, which splits the dataset into training and test sets in an 8:2 ratio, and simultaneously uses 5-fold cross-validation. The model training and optimization module is used to train and optimize the model based on the constructed dataset using machine learning and time series-based deep learning, predict the nitrogen removal performance of the short-cut nitrification-anaerobic ammonium oxidation process, evaluate the output performance of the model on the test set, and select the optimal model. The interpretability analysis module is used to perform importance analysis on feature parameters based on SHAP and PDP interpretability analysis, and to screen key control factors that affect the efficiency and stability of the system. An optimized prediction and control module is used to integrate prediction and uncertainty analysis, optimize control based on key control factors, and construct an optimized control strategy for the entire short-range nitrification-anaerobic ammonium oxidation process.

[0057] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0058] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0059] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart predictive control method for a short-cut nitrification-anaerobic ammonium oxidation process, characterized in that, Includes the following steps: Step S1, Dataset Construction: Based on existing research, sample data of short-cut nitrification-anaerobic ammonium oxidation process under different operating conditions were collected, mainly including water quality parameters, operating parameters and control parameters, to construct a multi-source dataset suitable for machine learning. Step S2, Dataset Preprocessing: The dataset is preprocessed, and feature engineering and input variable screening are implemented, including handling missing and outlier values, class information transformation, data standardization, correlation analysis, and dataset splitting into training and test sets in an 8:2 ratio to provide a high-quality data foundation for the model. Simultaneously, 5-fold cross-validation is adopted. Step S3, Prediction Model Construction: Based on the constructed dataset, machine learning and time-series-based deep learning are used for training and optimization to predict the nitrogen removal performance of the short-cut nitrification-anaerobic ammonium oxidation process, and the output performance of the model on the test set is evaluated to select the optimal model. Step S4: Importance analysis of SHAP and PDP dual-dimensional features: Based on SHAP and PDP interpretability analysis, the importance of characteristic parameters is analyzed to screen key control factors that affect the efficiency and stability of the system. Step S5: Construction of Predictive Regulation Strategy: By integrating prediction and uncertainty analysis, and optimizing regulation based on key control factors, a comprehensive optimization control strategy for the short-range nitrification-anaerobic ammonium oxidation process is constructed.

2. The intelligent predictive control method for a short-cut nitrification-anaerobic ammonium oxidation process according to claim 1, characterized in that, In step S2, the handling of missing and outlier values ​​includes calculating the median of the corresponding feature values, selectively filling in missing values ​​after judging all feature values, and the judgment process is as follows: The input features in the dataset to be preprocessed are divided into categorical features and continuous numerical features. The categorical features include reactor type and sludge form, while the continuous numerical features include volume, operating days, pH, aeration ratio, DO, aeration rate, HRT, temperature, influent organic carbon, influent ammonia nitrogen, and influent nitrite. Subsequently, missing values ​​of continuous numerical features were filled with NaN using the median filling strategy, and missing values ​​of categorical features were filled with the mode filling strategy to preserve the original distribution characteristics of the data. Outlier cleaning employs a mild cleaning strategy, calculating the 1st percentile as the minimum and the 99th percentile as the maximum for each target variable, defining a reasonable data range, and filtering out 1% of extreme outliers.

3. The intelligent predictive control method for a short-cut nitrification-anaerobic ammonium oxidation process according to claim 1, characterized in that, In step S2, the supplemented dataset is subjected to feature encoding and normalization, including encoding and standardizing different types of features separately. The process is as follows: For numerical data, standard scaling is performed using the StandardScaler method. The process is as follows: ； in, denoted as the standardized data of the numerical features of the sample, x represents the original data of the numerical features of the sample, μ is the mean of each numerical feature data, and σ is the standard deviation of each numerical feature data; One-hot encoding is used for categorical data to convert discrete categories into numerical vectors that the model can recognize. At the same time, an unknown category ignoring strategy is set to improve the model's compatibility with unknown category data.

4. The intelligent predictive control method for a short-cut nitrification-anaerobic ammonium oxidation process according to claim 1, characterized in that, In step S2, the dataset is split according to the evaluation requirements of the machine learning model. The preprocessed feature matrix and the target variable are randomly split into mutually exclusive datasets in an 8:2 ratio. 80% of the data is used as the training set for model training, and 20% of the data is used as the test set for model testing. At the same time, 5-fold cross-validation is used to improve the stability of the model.

5. The intelligent predictive control method for a short-cut nitrification-anaerobic ammonium oxidation process according to claim 1, characterized in that, In step S3, the water quality parameters, operational parameters, and management parameters in the training set are used as the training basis. Different machine learning methods are selected for training, and the most suitable model is screened and optimized. The coefficient of determination R is used. 2 The predictive performance of the prediction models is compared by comprehensively evaluating the root mean square error (RMSE) and mean absolute error (MAE) to select the optimal prediction model.

6. The intelligent predictive control method for a short-cut nitrification-anaerobic ammonium oxidation process according to claim 1, characterized in that, In step S3, the process of comparing the prediction performance of the prediction models is represented as follows: ； ； ； in, The sample index in the test set is The actual nitrogen removal efficiency value, The sample index in the test set is The predicted nitrogen removal efficiency value is ȳ, which is the average of the actual nitrogen removal efficiency of all samples in the test set, and n is the number of samples in the test set.

7. The intelligent predictive control method for a short-cut nitrification-anaerobic ammonium oxidation process according to claim 1, characterized in that, In step S4, the process of feature data importance analysis includes using the Python SHAP package to calculate the SHAP contribution value of each feature, quantifying the global impact of a single feature on the prediction result of nitrogen removal efficiency; combining single-feature and dual-feature interaction partial dependency analysis to verify the feature marginal effect; and combining the magnitude of SHAP contribution and partial dependency response features to screen out the core key control parameters that affect process performance.

8. The intelligent predictive control method for a short-cut nitrification-anaerobic ammonium oxidation process according to claim 1, characterized in that, In step S5, the process of constructing the optimal operating condition control strategy includes: selecting key process parameters as control variables, using intelligent optimization algorithms to iteratively solve for the optimal parameter combination, and forming a customized control scheme for improving nitrogen removal efficiency in wastewater treatment for multi-scale operating condition fluctuation scenarios.

9. A predictive control system for an intelligent predictive control method of the short-range nitrification-anaerobic ammonium oxidation process according to any one of claims 1-8, characterized in that, It includes a data acquisition module, a data preprocessing module, a model training and optimization module, an interpretability analysis module, and an optimization, prediction, and control module; The data acquisition module is used to collect sample point data of short-range nitrification-anaerobic ammonium oxidation process under different operating conditions based on existing research. The data mainly includes water quality parameters, operating parameters and operational parameters, and constructs a multi-source dataset suitable for machine learning. The data preprocessing module is used to preprocess the dataset, implement feature engineering and input variable screening, including missing value and outlier handling, category information conversion, data standardization, correlation analysis, and dataset partitioning, which splits the dataset into training and test sets in an 8:2 ratio, and simultaneously uses 5-fold cross-validation. The model training and optimization module is used to train and optimize the model based on the constructed dataset using machine learning and time series-based deep learning, predict the denitrification performance of the short-cut nitrification-anaerobic ammonium oxidation process, evaluate the output performance of the model on the test set, and select the optimal model. The interpretability analysis module is used to perform importance analysis on feature parameters based on SHAP and PDP interpretability analysis, and to screen key control factors that affect the efficiency and stability of the system. The optimization prediction and control module is used to integrate prediction and uncertainty analysis, optimize and control the process based on key control factors, and construct an optimized control strategy for the entire short-range nitrification-anaerobic ammonium oxidation process.

Images