Method for predicting aeration air volume of sewage treatment system based on time sequence data driving

Abstract

The invention discloses a method for predicting the aeration air volume of a sewage treatment system based on time sequence data driving. The method comprises the following steps of: acquiring multi-source data; a multi-source data processing step; processing abnormal values in the multi-source data sample X (tk) by adopting an isolated forest algorithm; performing multi-scale feature extraction on the feature sequence Xclean (tk) by adopting a time window layering mode; establishing an initial static topology and updating the initial static topology; constructing an XGBoost aeration air volume prediction model, and optimizing a hyper-parameter theta of the XGBoost aeration air volume prediction model by adopting a Bayesian optimization method; and evaluating the XGBoost aeration air volume prediction model through a multi-dimensional evaluation index. The sewage treatment system aeration air volume prediction method based on time sequence data driving and the control method thereof have the advantages that the control precision of the aeration air volume can be remarkably improved, system response is more timely, the system maintenance cost is reduced, and the cross-scene adaptive capacity of the system is enhanced.

Description

Technical Field

[0001] This invention relates to an aeration volume control technology for wastewater treatment processes, and more particularly to a time-series data-driven method for predicting aeration volume in wastewater treatment systems. Background Technology

[0002] In wastewater treatment, certain methods and equipment are used to forcibly introduce air into the wastewater, allowing the wastewater to come into contact with air and become oxygenated. This agitation accelerates the transfer of oxygen from the air into the liquid, prevents suspended solids from settling, and enhances the contact between organic matter, microorganisms, and dissolved oxygen, thus oxidizing and decomposing the organic matter in the wastewater. This process of forcibly oxygenating wastewater is called aeration. Aeration control is a systematic process that adjusts parameters such as aeration rate and dissolved oxygen concentration to ensure wastewater treatment effectiveness and optimize operational efficiency.

[0003] As a core component of wastewater treatment, aeration control has evolved over a period of time, which can be roughly divided into three stages: (1) First stage: experience-driven control; early aeration control mainly relied on manual experience or fixed time settings for adjustment. Operators manually adjusted the frequency or volume of the blower based on historical experience, lacking the ability to dynamically respond to changes in water quality. This method was not only inefficient, but also prone to insufficient or excessive aeration, resulting in unstable effluent quality or energy waste. (2) Second stage: model-driven control; with the deepening of research on wastewater treatment mechanisms, mathematical modeling methods were gradually applied to aeration control, such as activated sludge model (ASM) and oxygen transfer model. These methods achieve a certain degree of automated control by establishing the physicochemical relationship between water quality parameters and aeration volume. However, due to the complexity of microbial metabolic processes and frequent system disturbances, model parameters are difficult to calibrate accurately and require frequent parameter adjustments to adapt to different working conditions. The high maintenance cost and poor versatility limit its large-scale application. (3) Third stage: Data-driven control; In recent years, with the development of the Industrial Internet of Things and big data analysis, machine learning technology has been widely introduced into the field of aeration control. By collecting historical operating data, predictive models are built to optimize aeration strategies, which has higher flexibility and adaptability. However, most solutions still have problems such as imperfect feature engineering, insufficient mining of time series information, and weak model generalization ability, especially in dealing with sudden load changes and cross-plant migration applications.

[0004] Currently, existing technologies for airflow control mainly employ the following methods.

[0005] 1. Manual / Experience-Based Control Method: This method involves manually adjusting the aeration volume based on experience. Operators judge the aeration intensity based on historical experience or simple instrument readings, resulting in strong subjectivity and a lack of scientific basis. The main drawbacks of this control method are as follows: inaccurate aeration volume adjustment easily leads to large fluctuations in DO (dissolved oxygen) and unstable treatment effects; "over-aeration" often occurs, resulting in increased energy consumption and poor economic efficiency.

[0006] 2. Constant DO Control Strategy: PID control based on dissolved oxygen feedback is a common automatic control method. It monitors the DO value using sensors and adjusts the aeration rate to maintain DO within a set range. The main drawbacks of this method are as follows: DO detection has a lag, making it difficult to respond promptly to sudden changes in influent load; the control response is slow, and control delays easily occur during drastic water quality fluctuations, affecting the stability of the effluent quality; it has a high dependence on equipment, and sensor drift or malfunction can easily lead to false control.

[0007] 3. Mathematical Model Control Method: Control methods based on activated sludge models such as ASM establish mathematical models of the wastewater biodegradation process to simulate the relationship between pollutant removal and oxygen consumption, thereby calculating the required aeration rate. However, this method has the following drawbacks: Numerous and difficult-to-calibrate model parameters, such as oxygen mass transfer coefficient, sludge concentration, and microbial growth rate, are greatly affected by factors such as water quality and temperature; in actual operation, model parameters need to be continuously adjusted according to process conditions, resulting in high maintenance costs; the process flow and structure of different wastewater treatment plants vary greatly, making it difficult to directly reuse the model, resulting in poor versatility and portability; high computational complexity and poor real-time performance make it difficult to meet the rapid response requirements of online control.

[0008] 4. Machine Learning Control Methods; In existing technologies, some wastewater treatment systems attempt to use machine learning methods such as random forests, DNNs, and LSTMs for aeration prediction and control modeling, using historical data to train models and achieve intelligent decision-making on aeration volume. Its advantages include greater flexibility and adaptability compared to traditional methods, the ability to capture nonlinear relationships, and suitability for multivariate control scenarios. However, it suffers from the following problems: Coarse data preprocessing: insufficient handling of missing and outlier values; some systems lack sliding windows or forward filling mechanisms, affecting model input quality; Insufficient utilization of temporal features: most schemes only use real-time data points for prediction, ignoring long-term trends and periodic patterns (such as diurnal load variations and weekly fluctuations), leading to large model prediction errors; Weak model generalization ability: model training is often limited to a single plant area, lacking universal features such as holiday markers and daily / weekly cycle codes, making it difficult to adapt to different regions or seasonal changes. In summary, although existing aeration control schemes have improved automation to some extent, they still generally suffer from low control accuracy, slow response, and poor adaptability. Summary of the Invention

[0009] To overcome the shortcomings of the existing technologies, this invention provides a time-series data-driven method for predicting aeration volume in wastewater treatment systems. This method aims to achieve precise matching between aeration volume and water quality requirements, effectively avoid aeration lag, and improve the adaptability and self-adaptability of wastewater treatment systems under different operating conditions.

[0010] The present invention adopts the following technical solution to solve the technical problem.

[0011] The present invention provides a time-series data-driven method for predicting aeration volume in a wastewater treatment system, comprising the following steps:

[0012] Step 1: Multi-source data acquisition; Collect multi-source data X(t) from the wastewater treatment plant; the multi-source data includes aeration equipment operating parameters, water quality indicators, and system auxiliary parameters;

[0013] Step 2: Multi-source data processing steps; the Isolation Forest algorithm is used to process multi-source data samples X(t) k The system identifies outliers in the dataset, obtains and outputs the feature sequence X after anomaly handling. clean (t) k );

[0014] Step 3: Feature extraction; The feature sequence X is extracted using a time window layering method. clean (t) k Multi-scale feature extraction is performed to obtain the final multi-scale feature vector. ;

[0015] Step 4: Dynamic Topology Graph Establishment Steps; Establish an initial static topology and update the initial static topology to obtain the updated topology G(t) k );

[0016] Step 5: Model optimization step; Construct an XGBoost aeration air volume prediction model and use the Bayesian optimization method to optimize the hyperparameter θ of the XGBoost aeration air volume prediction model;

[0017] Step 6: Model evaluation step; The XGBoost aeration air volume prediction model is evaluated using multi-dimensional evaluation indicators.

[0018] The structural features of the time-series data-driven method for predicting aeration volume in wastewater treatment systems also lie in:

[0019] Furthermore, the operating parameters of the aeration equipment include air volume, air pressure, and current.

[0020] Furthermore, the water quality indicators include chemical oxygen demand (COD), dissolved oxygen (DO), and ammonia nitrogen (NH3-N).

[0021] Furthermore, the system auxiliary parameters include influent flow rate, temperature, and pH value.

[0022] Furthermore, the processing of multi-source data includes the following steps:

[0023] Step 21: Construction of the Isolation Forest algorithm model and sample preparation;

[0024] Step 22: Calculate the anomaly score s(X(t)). k ), n);

[0025] Step 23: Identify and correct outliers;

[0026] Step 24: Output the processed feature sequence X clean (t) k ).

[0027] Furthermore, the feature extraction in step 3 includes the following steps:

[0028] Step 31: Calculation of multi-scale statistical characteristics;

[0029] Step 32: Periodic encoding;

[0030] Step 33: Obtain the final multi-scale feature vector .

[0031] Furthermore, the dynamic topology graph establishment process in step 4 includes the following steps:

[0032] Step 41: Define the relationship between feature parameters and nodes in the dynamic topology graph;

[0033] Step 42: State update of dynamic topology graph.

[0034] Furthermore, the model optimization process in step 5 includes the following steps;

[0035] Step 51: Construct an XGBoost aeration airflow prediction model;

[0036] Step 52: Use Bayesian optimization to optimize the hyperparameter θ of the XGBoost aeration airflow prediction model.

[0037] Furthermore, in step 6, the model evaluation process includes the following steps:

[0038] Step 61: Determine multi-dimensional evaluation indicators;

[0039] Step 62: Use multi-dimensional evaluation indicators to evaluate the XGBoost aeration airflow prediction model;

[0040] Step 63: Control after the XGBoost aeration airflow prediction model meets the standard;

[0041] Furthermore, in step 61, the multi-dimensional evaluation indicators include water quality compliance rate (RDO), prediction error (Ewind), and control stability.

[0042] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0043] This invention discloses a time-series data-driven method for predicting aeration volume in wastewater treatment systems, comprising:

[0044] Step 1: Multi-source data acquisition; Collect multi-source data X(t) from the wastewater treatment plant; the multi-source data includes aeration equipment operating parameters, water quality indicators, and system auxiliary parameters;

[0045] Step 2: Multi-source data processing steps; the Isolation Forest algorithm is used to process multi-source data samples X(t) k The system identifies outliers in the dataset, obtains and outputs the feature sequence X after anomaly handling. clean (t) k );

[0046] Step 3: Feature extraction; The feature sequence X is extracted using a time window layering method. clean (t) k Multi-scale feature extraction is performed to obtain the final multi-scale feature vector. ;

[0047] Step 4: Dynamic Topology Graph Establishment Steps; Establish an initial static topology and update the initial static topology to obtain the updated topology G(t) k );

[0048] Step 5: Model optimization step; Construct an XGBoost aeration air volume prediction model and use the Bayesian optimization method to optimize the hyperparameter θ of the XGBoost aeration air volume prediction model;

[0049] Step 6: Model evaluation step; The XGBoost aeration air volume prediction model is evaluated using multi-dimensional evaluation indicators.

[0050] The time-series data-driven method for predicting and controlling aeration volume in wastewater treatment systems has the following technical advantages.

[0051] 1. Significantly improved control accuracy: By fusing multi-scale time-series features and using the XGBoost model, the system upgrades from passive adjustment relying on a single indicator to active prediction based on multi-parameter collaboration, greatly reducing energy waste and water quality fluctuations caused by control deviations. This achieves precise matching between aeration volume and water quality requirements, significantly reducing energy waste and water quality fluctuations caused by control deviations compared to traditional solutions.

[0052] 2. More timely system response: By using multi-scale time windows (short-term capture of sudden changes, medium-term tracking of trends) and real-time correlation updates of dynamic topology maps, the system can quickly identify water quality fluctuation patterns. Combined with the efficient reasoning capabilities of the XGBoost model, the system responds to sudden changes in water quality more promptly, effectively avoiding the risk of exceeding standards due to lag in traditional control. Based on capturing short-term changes and long-term trends in water quality through multi-level time windows and combining real-time correlation learning of dynamic topology maps, the system's response speed to load fluctuations is significantly accelerated, effectively avoiding aeration lag problems.

[0053] 3. Reduced model maintenance costs: Full-process automation (data processing - parameter optimization - model update) replaces manual operations, reducing the labor costs and technical barriers to model maintenance; Bayesian optimization is used to achieve automatic hyperparameter tuning, replacing the manual calibration process, significantly reducing the maintenance cost of the model throughout its life cycle and reducing reliance on professional operation and maintenance personnel.

[0054] 4. Enhanced cross-scenario adaptability: Through dynamic association learning and periodic encoding, the model maintains stable performance under different sewage treatment plants and different seasonal operating conditions, and can be quickly deployed without extensive customized development; through multi-source feature fusion, periodic encoding and dynamic topology adaptive learning, the model's transfer adaptability under different sewage treatment plants and different operating conditions is significantly improved, and deployment efficiency and generalization performance are greatly improved.

[0055] The present invention provides a time-series data-driven method for predicting and controlling aeration volume in wastewater treatment systems. This method has advantages such as significantly improving the control accuracy of aeration volume, providing more timely system response, reducing system maintenance costs, and enhancing the system's cross-scenario adaptability. Attached Figure Description

[0056] Figure 1 This is a flowchart of the time-series data-driven method for predicting aeration volume in a wastewater treatment system according to the present invention.

[0057] The present invention will be further described below through specific embodiments and in conjunction with the accompanying drawings. Detailed Implementation

[0058] See Figure 1 The present invention provides a time-series data-driven method for predicting aeration volume in a wastewater treatment system, comprising the following steps:

[0059] Step 1: Multi-source data acquisition; Collect multi-source data X(t) from the wastewater treatment plant; the multi-source data includes aeration equipment operating parameters, water quality indicators, and system auxiliary parameters;

[0060] Step 2: Multi-source data processing steps; the Isolation Forest algorithm is used to process multi-source data samples X(t) kThe system identifies outliers in the dataset, obtains and outputs the feature sequence X after anomaly handling. clean (t) k );

[0061] Step 3: Feature extraction; The feature sequence X is extracted using a time window layering method. clean (t) k Multi-scale feature extraction is performed to obtain the final multi-scale feature vector. ;

[0062] From time t1 to t N Multiple source data samples at N time points. The purified feature sequences. Extract each multi-source data sample X by time window stratification clean (t) k The statistical characteristics of the data cover three scales: short-term (abrupt), medium-term (trend), and long-term (cycle); the basic control cycle is defined. Short-term window Mid-term window Long-term window N is the number of samples, and X(t) is the number of samples from multiple data sources. k ) includes m parameters x1 (t k ), x2(t k )……. x m (t) k (m) refers to the operating parameters of the aeration equipment, water quality indicators, and system auxiliary parameters.

[0063] Step 4: Dynamic Topology Graph Establishment Steps; Establish an initial static topology and update the initial static topology to obtain the updated topology G(t) k );

[0064] Step 5: Model optimization step; Construct an XGBoost aeration air volume prediction model and use the Bayesian optimization method to optimize the hyperparameter θ of the XGBoost aeration air volume prediction model;

[0065] Step 6: Model evaluation step; The XGBoost aeration air volume prediction model is evaluated using multi-dimensional evaluation indicators.

[0066] For wastewater treatment aeration control scenarios, multi-dimensional evaluation indicators were designed to verify the prediction accuracy and control stability of the XGBoost aeration airflow prediction model, ensuring that the output aeration strategy (such as airflow) meets the objectives of water quality compliance and energy consumption optimization. These multi-dimensional evaluation indicators include, but are not limited to, the water quality compliance rate R. DO Prediction error E wind And parameters such as control stability.

[0067] In practice, the operating parameters of the aeration equipment include air volume, air pressure, and current.

[0068] In practice, the water quality indicators include chemical oxygen demand (COD), dissolved oxygen (DO), and ammonia nitrogen (NH3-N).

[0069] In practice, the system auxiliary parameters include influent flow rate, temperature, and pH value.

[0070] Various data were extracted from multiple heterogeneous tables in the wastewater treatment plant database according to time intervals. These multi-source data include, but are not limited to, aeration equipment operating parameters (such as air volume, air pressure, current, etc.), water quality index parameters (such as COD, DO, NH3-N, etc.), and system auxiliary parameters (such as influent flow rate, temperature, pH value, etc.). To ensure that the air volume control output is once every 10 minutes, the multi-source data is sampled into groups of 10 minutes. The multi-source data are aggregated together, assuming that the total number of various characteristic parameters of the multi-source data is m, and the multi-source data X(t) after the aggregation of various data at time t is expressed as a one-dimensional vector as follows (1);

[0071] X(t)=[x1(t),x2(t)…….x m (t)] (1)

[0072] In formula (1), x i (t) represents the i-th feature parameter collected at time t; x i X(t) is a scalar representing the set of multi-source data related to wind volume prediction at time t; m represents the total number of feature parameters, 1≤i≤m, X(t)∈R m .

[0073] t k Multi-source data samples X(t) at time t k The set of ) is denoted as . Including from time t1 to t N Multiple data samples are obtained at N time points, with one sample obtained at each time point. N is the number of samples, and the multiple data sample X(t) is the number of data samples. k = [x1(t)] k ), x2(t k )……. x i (t) k )…….x m (t) k [ ], including characteristic parameters such as operating parameters of m aeration devices, water quality indicators and system auxiliary parameters.

[0074] In practice, the processing of multi-source data includes the following steps:

[0075] The Isolation Forest algorithm is used to process the multi-source data samples X(t) k = [x1(t)] k ), x2(t k )……. x i (t) k )…….x m (t) k The exception handling process includes the following steps:

[0076] Step 21: Construction of the Isolation Forest algorithm model and sample preparation;

[0077] Will The input is fed into the Isolation Forest algorithm model to construct an Isolation Forest composed of multiple isolated trees. Here, 1 ≤ k ≤ N, and N is the number of samples. For each multi-source data sample X(t)... k They will be processed one by one.

[0078] For each isolated iTree, recursively and randomly select feature dimensions and split values ​​to divide the data samples into subsets until a leaf node contains only one sample or the maximum depth is reached. ;in This represents a randomly selected subset of samples. The `BuildTree` function is a recursive function used to construct a binary tree based on the input data. `max_depth` is the maximum depth of the isolated tree, set to log2(n), where n is the number of samples randomly drawn from the N samples when constructing a single isolated tree `iTree`.

[0079] Step 22: Calculate the anomaly score s(X(t)). k ), n);

[0080] For t k Multi-source data samples X(t) at time t k ), calculate the mean path length of this sample across all isolated trees. Combined with the sample size n, the anomaly score s(X(t)) is calculated. k The calculation formula for n is shown in equation (2) below.

[0081] (2);

[0082] In formula (2), n is the number of samples taken from N samples for a single isolated tree; c(n) is a correction coefficient used to eliminate the influence of sample size, and H(k) is the harmonic number; where the anomaly score s(X(t)) is the harmonic score. k The closer the value of X(t) is to 1, the better the multi-source data sample X(t) is to 1. k The more likely it is to be abnormal.

[0083] Step 23: Identify and correct outliers;

[0084] Set a threshold τ for outliers. If the outlier score s(X(t) is... k If n) ≥ τ, determine the multi-source data sample X(t). k If a value is an outlier, it is replaced with the interpolated / mean value of the features from adjacent normal time points to obtain multi-source data sample X(t). k After correction, the feature vector See equation (3) below;

[0085] (3);

[0086] In formula (3), The corrected feature vector The m-th eigenvalue; m represents The total number of feature parameters.

[0087] Step 24: Output the processed feature sequence X clean (t) k ).

[0088] Finally, the feature sequence X after anomaly handling clean (t) k See equation (4) below;

[0089] (4);

[0090] In formula (4), .

[0091] Follow the steps above to From time t1 to t N All multi-source data samples at N time points are subjected to anomaly processing to obtain purified feature sequences. .

[0092] In practice, the feature extraction in step 3 includes the following steps:

[0093] Step 31: Calculation of multi-scale statistical characteristics;

[0094] For each window W, compute the feature vector. The statistical measures (mean, variance, mutation index) are concatenated into a multi-scale feature vector. .

[0095] Step 311: Short-term window

[0096] The first-order difference extremum of the feature within the window is calculated (to identify abrupt changes). The calculation formula is shown in equation (5) below.

[0097] (5);

[0098] In formula (5), the Mutate function is used to create new variables or modify existing variables in a data frame; Based on the control cycle.

[0099] Short-term eigenvector X short (t) k )for: , This represents the mean value within the window.

[0100] Step 312: Mid-term window

[0101] The linear regression slope (quantitative trend) of the features within the window is calculated using the following formula (6).

[0102] (6);

[0103] In formula (6), the trend function is used to identify trends in the data sequence; the slope function is a mathematical function used to calculate the slope.

[0104] Mid-term eigenvector X mid (t) k )for: , This represents the variance within the window.

[0105] Step 313: Long-term window

[0106] Calculate the periodic similarity of features within the window (compared with historical periodic patterns), and the calculation formula is shown in the following formula (7).

[0107] (7);

[0108] In formula (7), K represents the number of historical cycles. The correlation coefficient; Refers to the cleaned feature time series within the current window; This refers to the cleaned feature time series within the i-th historical period; the periodSim function, short for "Period Similarity," is a function that measures the similarity between the current time series and historical periodic patterns. By averaging the correlation coefficients of multiple historical periods, the overall similarity score between the current window and historical periods is obtained, which is used to capture the long-term periodic characteristics of the data.

[0109] Long-term eigenvector X long (t) k ): .

[0110] Step 32: Periodic encoding;

[0111] purified characteristic sequence There are day and night cycles and process cycles (such as aeration-sedimentation cycles), and the temporal periodicity is embedded and encoded through Fourier transform and periodicity.

[0112] Step 321: Obtain the feature vector X clean The principal period T of (t) dominant ;

[0113] For long-term windows The main period T within the feature is extracted using Fourier transform. dominant See equation (8) below;

[0114] (8);

[0115] In formula (8), argmax is a function that calculates the parameters (set) of a function. For the function y=f(x), if the result x0=argmax(f(x)), it means that when the function f(x) takes x=x0, the maximum value of the range of values ​​of f(x) is obtained. If there are multiple points (or values ​​of x) that make f(x) obtain the same maximum value, then the result of argmax(f(x)) is a set of points. That is, argmax(f(x)) is the variable point x (or the set of x) that makes f(x) obtain the maximum value. Let f be the Fourier transform function, f be the frequency, and t be the time interval from t1 to t2. N Any point in time.

[0116] Step 322: Periodic embedding encoding;

[0117] The current time t k Mapped to the main period T dominant phase And encoded as trigonometric function feature PeriodCode(t) k See equation (9) below;

[0118] (9);

[0119] In formula (9), phase The PeriodCode function is used to encode periodic features by transforming periodic patterns into learnable features for the model through trigonometric functions; the mod function is the modulo function.

[0120] Step 33: Obtain the final multi-scale feature vector .

[0121] The multi-scale statistical features are concatenated with periodic codes to obtain the final multi-scale feature vector. See equation (10) below.

[0122] (10).

[0123] In practice, the dynamic topology graph establishment process in step 4 includes the following steps:

[0124] Step 41: Define the relationship between feature parameters and nodes in the dynamic topology graph;

[0125] The characteristic parameters such as dissolved oxygen (DO), influent flow rate, and COD are set as nodes V in the dynamic topology graph, as shown in equation (11).

[0126] (11);

[0127] In formula (11), v DO For the dissolved oxygen (DO) node in the dynamic topology graph, v 流量 For the inflow node in the dynamic topology graph, v COD This is the chemical oxygen demand (COD) node in the dynamic topology graph.

[0128] The node features are integrated with basic features (real-time measured values, such as real-time DO concentration, instantaneous influent flow rate, and real-time COD value) and multi-scale feature vectors. This includes information on short-term changes, medium-term trends, and long-term cycles in water quality, obtaining the feature vector h of node i. i (t) k See equation (12) below;

[0129] (12).

[0130] In formula (12), V(t) k () refers to the nodes (dissolved oxygen, influent flow rate, COD, etc.) in the dynamic topology diagram at time t k The set of feature parameters at any given time is used to integrate the real-time state of the nodes into the feature vector in order to capture the dynamic information of each parameter of the wastewater treatment system at the current time.

[0131] Step 42: State update of dynamic topology graph.

[0132] The initial static topology of the dynamic topology graph is as follows: based on the wastewater treatment process mechanism (such as COD degradation consuming DO, and influent flow affecting DO mass transfer), the edges and weights between nodes are initialized, and the initial topology G0 = (V, E0, W0) is constructed. E0 is the initial edge set (representing the physical / chemical association of nodes); W0 is the initial edge weight (assigned based on domain knowledge and the correlation with historical data, such as setting the initial association weight of COD-DO to the theoretical consumption ratio of the process).

[0133] The dynamic weight learning process of a dynamic topology graph includes: using an attention mechanism, based on the feature vector h of node i... i (t) k Dynamically update the edge weight W. For v i v j Nodes, calculate attention coefficient α ij (t) k To measure the two nodes v i and node v j The degree of correlation between them, attention coefficient α ij (t) k The calculation is shown in the following formula (13);

[0134] (13).

[0135] In formula (13), a T is the learnable parameter; ‖ is the vector concatenation operation symbol, which updates the edge weights through the attention coefficient, reflecting the dynamic relationship between COD and DO consumption, the impact of traffic on DO, etc.; the Softmax function is a normalized exponential function that converts any real vector into a probability distribution, and converts a K-dimensional input vector into a K-dimensional output vector through exponential operation and normalization; LeakyReLU is an activation function that allows a non-zero output for an input value less than zero.

[0136] By combining a time decay factor (to avoid topological abrupt changes) to update the topology in real time, the updated topology G(t) is obtained. k The update formula is shown in equation (14) below.

[0137] (14).

[0138] In formula (14), E(t) k ) for t k The set of edges in the time-domain topology (representing the physical / chemical associations of nodes); W(t) k ) for t k Edge weights in the time-varying topology.

[0139] In practice, the model optimization process in step 5 includes the following steps;

[0140] Step 51: Construct an XGBoost aeration airflow prediction model;

[0141] Extract the association features (such as the flattened vector of the association weight matrix) of node i in the dynamic topology graph, and compare them with the feature vector h of node i. i (t) k By splicing, multi-scale features are obtained. The prediction targets are key aeration control indicators (such as target air volume and DO adjustment compensation). A prediction model for XGBoost aeration air volume is constructed, as shown in equation (15).

[0142] (15).

[0143] In formula (15), XGBoost() is the regression model function, and θ is the hyperparameter of the XGBoost aeration air volume prediction model (such as tree depth, learning rate, etc.).

[0144] Step 52: Optimize the hyperparameter θ of the XGBoost aeration airflow prediction model using Bayesian methods;

[0145] To improve the accuracy and control stability of the XGBoost aeration airflow prediction model, Bayesian optimization is used to search for and optimize the hyperparameter θ of the XGBoost aeration airflow prediction model. A loss function is defined. Considering both prediction bias and control stability; loss function See equation (16) below;

[0146] (16).

[0147] In formula (16), Let λ be the prediction error and λ be the stability penalty coefficient. This represents predictive variance.

[0148] Bayesian optimization process: First, randomly sample a small number of hyperparameters θ to train XGBoost and calculate the loss function. Then, Gaussian process regression is used to fit the hyperparameter-loss mapping to construct a surrogate model; new hyperparameters are selected through the acquisition function for iterative optimization until the loss converges, and the optimized hyperparameters θ are output.

[0149] In practice, step 6, the model evaluation process includes the following steps:

[0150] Step 61: Determine multi-dimensional evaluation indicators;

[0151] (1) Water quality compliance rate R DO : The control target range and water quality compliance rate are adjusted by the threshold according to the process requirements; dissolved oxygen (DO) is considered to be compliant if it is in the range of [1.2, 2.2], and not compliant if it is outside this range.

[0152] (2) Prediction error E wind Predicted value of aeration volume Q pred (t) k ) and the actual optimal value Q opt (t) k The formula for calculating the average deviation (threshold custom) is shown in the following formula (17);

[0153] (17)

[0154] In formula (17), N is the number of samples.

[0155] (3) Control stability; use the fluctuation range S of the aeration air volume Q at continuous time. 波动 The threshold for fluctuation amplitude, used as a parameter to measure the control stability of the XGBoost aeration airflow prediction model, can be defined according to actual operating conditions. Fluctuation amplitude S 波动 The calculation formula is shown in the following formula (18);

[0156] (18).

[0157] In formula (18), Q(t k ) for t k Aeration air volume at any given time, Q(t) k-1 ) for t k The aeration rate at any given time, where N is the number of samples.

[0158] Step 62: Use multi-dimensional evaluation indicators to evaluate the XGBoost aeration airflow prediction model;

[0159] In practice, both offline and online evaluation methods were used to evaluate the XGBoost aeration air volume prediction model.

[0160] (1) Offline evaluation; if the training indicators are within the defined range, then enter online control; if any indicator is not satisfied, the XGBoost aeration air volume prediction model is deemed unqualified, and the training phase is returned. The abnormal sample weights are increased, the dynamic topology learning rate is adjusted, or the multi-scale window is expanded, and the XGBoost aeration air volume prediction model is retrained.

[0161] (2) Online evaluation: After the XGBoost aeration air volume prediction model is deployed, the operating indicators are monitored in real time and evaluated every hour. If the dissolved oxygen (DO) control and fan fluctuations are within acceptable range, the operation continues. Otherwise, the XGBoost aeration air volume prediction model is incrementally trained and the control strategy is switched back to other modes. When the target is met again, the system is switched back to intelligent control.

[0162] Step 63: Control after the XGBoost aeration airflow prediction model meets the standard;

[0163] After the XGBoost aeration volume prediction model passed the evaluation, the final output aeration volume Q was... final (t) k It needs to undergo safety boundary verification and is limited to the physical range of the equipment, as shown in the following formula (19);

[0164] (19).

[0165] In formula (19), Q pred (t) k Q represents the predicted air volume output by the XGBoost aeration air volume prediction model; min and Q max These are the minimum and maximum air volume that the fan can actually output.

[0166] In specific implementation, the multi-dimensional evaluation indicators in step 61 include water quality compliance rate (RDO), prediction error (Ewind), and control stability.

[0167] The aeration volume prediction method for wastewater treatment systems based on time-series data has the following technical features.

[0168] 1. Multi-scale-periodic synergistic feature engineering: This innovative approach utilizes statistical feature extraction strategies across different time windows, combining periodic coding to capture short-term patterns and long-term cycles of water quality changes, providing multi-dimensional feature support for accurate prediction. It breaks through the reliance on a single dissolved oxygen (DO) index, integrating multi-scale statistical features and periodic coding to capture short-term abrupt changes and long-term cyclical patterns in water quality, shortening system response time, improving prediction accuracy in special scenarios, and achieving a leap from "passive adjustment" to "active prediction."

[0169] 2. Dynamic topology association learning mechanism: Parameters such as DO, COD, and influent flow rate are modeled as graph nodes. Based on multi-scale features, the association weights between nodes (such as the consumption relationship between COD and DO) are dynamically learned, and the topology structure is updated in real time. This breaks through the limitations of single-factor dependence, optimizes the aeration control accuracy, improves the control robustness under complex working conditions, and enhances the stability when water quality changes abruptly.

[0170] 3. Isolated Forest-Topology-XGBoost Collaborative Framework: This framework uses isolated forests to purify anomalous data and dynamically mines parameter correlations through topology mining. XGBoost enables nonlinear mapping, and the three-level collaboration enhances the model's robustness against industrial noise data, adapting to complex field environments. It couples ensemble learning and probabilistic optimization algorithms to automatically learn the nonlinear relationship between parameters and water quality, achieving global optimization that eliminates manual calibration, reduces model maintenance costs, and improves parameter iteration efficiency.

[0171] 4. Closed-loop evaluation and adaptive update architecture: Construct a closed-loop mechanism of "offline training - online evaluation - real-time fine-tuning", combining physical constraints of the device and multi-index judgment to ensure the long-term stable operation of the model and solve the problem of insufficient generalization ability of traditional machine learning models.

[0172] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0173] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A time-series data-driven method for predicting aeration volume in wastewater treatment systems, characterized in that, Includes the following steps: Step 1: Multi-source data acquisition; Collect multi-source data X(t) from the wastewater treatment plant; the multi-source data includes aeration equipment operating parameters, water quality indicators, and system auxiliary parameters; Step 2: Multi-source data processing steps; the Isolation Forest algorithm is used to process multi-source data samples X(t) k The system identifies outliers in the dataset, obtains and outputs the feature sequence X after anomaly handling. clean (t) k ); Step 3: Feature extraction; The feature sequence X is extracted using a time window layering method. clean (t) k Multi-scale feature extraction is performed to obtain the final multi-scale feature vector. ; Step 4: Dynamic Topology Graph Establishment Steps; Establish an initial static topology and update the initial static topology to obtain the updated topology G(t) k ); Step 5: Model optimization step; Construct an XGBoost aeration air volume prediction model and use the Bayesian optimization method to optimize the hyperparameter θ of the XGBoost aeration air volume prediction model; Step 6: Model evaluation step; The XGBoost aeration air volume prediction model is evaluated using multi-dimensional evaluation indicators.

2. The method for predicting aeration volume in a wastewater treatment system based on time-series data as described in claim 1, characterized in that, The operating parameters of the aeration equipment include air volume, air pressure, and current.

3. The method for predicting aeration volume in a wastewater treatment system based on time-series data as described in claim 1, characterized in that, The water quality indicators include chemical oxygen demand (COD), dissolved oxygen (DO), and ammonia nitrogen (NH3-N).

4. The method for predicting aeration volume in a wastewater treatment system based on time-series data as described in claim 1, characterized in that, The system auxiliary parameters include influent flow rate, temperature, and pH value.

5. The method for predicting aeration volume in a wastewater treatment system based on time-series data as described in claim 1, characterized in that, The multi-source data processing procedure in step 2 includes the following steps: Step 21: Construction of the Isolation Forest algorithm model and sample preparation; Step 22: Calculate the anomaly score s(X(t)). k ), n); Step 23: Identify and correct outliers; Step 24: Output the processed feature sequence X clean (t) k ).

6. The method for predicting aeration volume in a wastewater treatment system based on time-series data as described in claim 1, characterized in that, The feature extraction in step 3 includes the following steps: Step 31: Calculation of multi-scale statistical characteristics; Step 32: Periodic encoding; Step 33: Obtain the final multi-scale feature vector .

7. The method for predicting aeration volume in a wastewater treatment system based on time-series data as described in claim 1, characterized in that, The dynamic topology graph creation process in step 4 includes the following steps: Step 41: Define the relationship between feature parameters and nodes in the dynamic topology graph; Step 42: State update of dynamic topology graph.

8. The method for predicting aeration volume in a wastewater treatment system based on time-series data as described in claim 1, characterized in that, Step 5, the model optimization process, includes the following steps; Step 51: Construct an XGBoost aeration airflow prediction model; Step 52: Use Bayesian optimization to optimize the hyperparameter θ of the XGBoost aeration airflow prediction model.

9. The method for predicting aeration volume in a wastewater treatment system based on time-series data as described in claim 1, characterized in that, In step 6, the model evaluation process includes the following steps: Step 61: Determine multi-dimensional evaluation indicators; Step 62: Use multi-dimensional evaluation indicators to evaluate the XGBoost aeration airflow prediction model; Step 63: Control after the XGBoost aeration air volume prediction model meets the standard.

10. The method for predicting aeration volume in a wastewater treatment system based on time-series data as described in claim 1, characterized in that, In step 61, the multi-dimensional evaluation indicators include water quality compliance rate (RDO), prediction error (Ewind), and control stability.

Images