Particular pollutant concentration prediction method based on space-time grid and related equipment

Abstract

The invention belongs to the technical field of pollutant prediction, and relates to a space-time grid-based specific pollutant concentration prediction method and related equipment, and the method comprises the steps: obtaining multi-source time sequence data of each process unit in a target sewage treatment system; calling a process flow space-time grid data model based on the multi-source time sequence data, and calculating a process pollution contribution vector between adjacent process units; fusing the depth time sequence features extracted from the multi-source time sequence data of each process unit with the associated features of the sewage treatment system represented by the process pollution contribution vector to generate a comprehensive feature vector; inputting the comprehensive feature vector into a trained dynamic adaptive prediction model, and outputting a feature pollutant concentration prediction value of at least one target process unit in the sewage treatment system in a specified time period in the future; therefore, the problems of insufficient consideration of a process flow space structure, one-sided data feature extraction, difficulty in adapting to process dynamic change and the like in the prior art are solved.

Description

Technical Field

[0001] This invention belongs to the field of pollutant prediction technology, and relates to a method and related equipment for predicting the concentration of characteristic pollutants based on spatiotemporal grids. Background Technology

[0002] With the rapid development of emerging industries such as semiconductors and integrated circuits, biomedicine and high-end medical devices, new energy and energy storage, the treatment of characteristic pollutants in wastewater generated during their production processes is becoming increasingly prominent. These pollutants typically exhibit characteristics such as low concentration but high risk, reluctance to degrade, and difficulty in treatment. Typical examples include fluorides and nanoparticles in semiconductor wastewater, high concentrations of organic matter and residual drugs in biomedical wastewater, and heavy metal ions and recalcitrant organic matter in wastewater from the new energy industry. Failure to accurately and timely predict the concentration and trends of these characteristic pollutants will hinder the stable operation of wastewater treatment facilities and ensure effluent quality meets standards, potentially leading to environmental pollution, restricting sustainable industrial development, and threatening regional water environment security.

[0003] Currently, prediction methods for characteristic pollutant concentrations mainly fall into three categories: First, methods based on traditional statistical models, such as the Autoregressive Moving Average (ARIMA) model in time series analysis. These methods can construct prediction models based on the temporal correlation of historical data and are suitable for pollutant concentration sequences with certain regularities. Second, machine learning methods, such as Support Vector Machines (SVM), which handle the complex relationship between pollutant concentrations and various influencing factors through nonlinear mapping relationships and show good adaptability in various water quality prediction tasks. Third, deep learning methods, such as Recurrent Neural Networks (RNN) and Long Short-Term Memory Networks (LSTM), which have strong sequence modeling capabilities and can capture long-term dependencies and dynamic features in concentration changes. They are suitable for processing high-dimensional and nonlinear water quality data and can provide technical support for pollution trend analysis.

[0004] Currently, the three types of prediction methods mentioned above still have shortcomings in practical applications. Traditional statistical methods (such as ARIMA) heavily rely on the stationarity and regularity of historical data, making it difficult to effectively adapt to non-stationary pollutant concentration sequences caused by frequent adjustments to production processes and drastic fluctuations in influent. Machine learning methods (such as SVM), when dealing with multi-source complex pollution (such as the coexistence of multiple drug residues in biopharmaceutical wastewater), rely on artificial feature construction and prior knowledge, and their generalization ability in variable coupling and high-dimensional nonlinear mapping is limited. While degree learning methods (such as LSTM) can capture complex temporal dependencies, their performance depends on a large amount of high-quality labeled data. They are prone to overfitting when actual wastewater monitoring data is scarce or noisy. At the same time, the model has poor interpretability, which restricts its practical application in real-time control and risk management.

[0005] More importantly, existing prediction methods generally treat monitoring points as isolated time-series data sources, failing to fully consider the migration, transformation, and synergistic effects of pollutants along the flow path between different process units in the wastewater treatment system. This neglect of the "process flow spatial structure" prevents models from capturing the transmission effects and causal relationships of pollutants in the spatiotemporal dimensions, limiting further improvements in prediction accuracy. Simultaneously, existing methods suffer from dual shortcomings in data processing and model building: on the one hand, they fail to fully explore the deep potential connections between multi-source time-series data (such as production processes, equipment status, water quality monitoring, and environmental data) and pollutant concentrations; on the other hand, they lack effective mechanisms to adapt to the dynamic changes in data distribution caused by rapid process iterations. These problems collectively result in poor prediction accuracy and adaptability of existing models when dealing with complex operating conditions such as influent shocks and process abrupt changes. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method and related equipment for predicting the concentration of characteristic pollutants based on spatiotemporal grids, thereby overcoming the core defect of existing prediction methods that do not adequately consider the spatial structure of wastewater treatment process flow.

[0007] This invention is achieved through the following technical solution: Methods for predicting the concentration of characteristic pollutants based on spatiotemporal grids, such as Figure 1 As shown, it includes: Acquire multi-source time-series data of each process unit in the target wastewater treatment system; Based on multi-source time-series data, the spatiotemporal grid data model of the process flow is invoked to calculate the process contamination contribution vector between adjacent process units; The deep time-series features extracted from the multi-source time-series data of each process unit are fused with the correlation features of the wastewater treatment system represented by the process pollution contribution vector to generate a comprehensive feature vector. The comprehensive feature vector is input into the trained dynamic adaptive prediction model, which outputs the predicted value of the characteristic pollutant concentration of at least one target process unit in the wastewater treatment system within a specified future time period. The construction process of the spatiotemporal grid data model of the process flow is as follows: Based on the target wastewater treatment process, the continuous wastewater treatment process is discretized into a one-dimensional chain grid structure. Each grid cell in the chain grid structure represents a process unit. At each sampling time, a data vector containing the concentration of characteristic pollutants, process parameters, and hydraulic parameters is filled into each process unit to construct a spatiotemporal grid data model of the process flow.

[0008] In some embodiments, the multi-source time-series data includes monitoring data of characteristic pollutants, process parameter data associated with the pollutants, and environmental and conversion condition data. The preprocessing of the acquired multi-source time-series data includes: The collected multi-source time-series data is cleaned, outlier removed, and missing value interpolated and filled. Normalize the process parameter data associated with pollutants; The monitoring data of the characteristic pollutants are standardized. The monitoring data of the characteristic pollutants include the concentrations of fluorides and nanoparticles in the semiconductor workshop, the chemical oxygen demand and residual drug concentrations in the biopharmaceutical workshop, and the heavy metal ion concentrations in the new energy workshop. The process parameter data associated with the contaminants include the hydrofluoric acid flow rate in semiconductor lithography, the amount of antibiotics added in biopharmaceutical fermentation, and the electrolyte concentration in new energy battery coating. The environmental and conversion condition data include pH value, temperature, dissolved oxygen, and industrial process adjustment records.

[0009] In some embodiments, the pollution contribution vector between adjacent process units is calculated as follows: For each pair of adjacent upstream and downstream process units based on the spatiotemporal grid data model of the process flow, the process contamination contribution vector of the upstream process unit to the downstream process unit is calculated.

[0010] This includes calculating the process contamination contribution vector of the upstream unit to the downstream unit. i→j The specific formula is: i→j (t) = =

[0011] Where t is the sampling time; C is the concentration of the characteristic pollutant; The formula for calculating the normalized concentration contribution is as follows: , The value is a preset minimum positive number, where pH is the acidity or alkalinity, T is the temperature, and ORP is the redox potential.

[0012] In some embodiments, the deep time-series features extracted from the multi-source time-series data of each process unit are specifically as follows: The multi-source time series data of each process unit are centered, and a covariance matrix is ​​constructed based on the centered multi-source time series data. Then, the covariance matrix is ​​decomposed into eigenvalues ​​and eigenvectors. The cumulative variance contribution rate is calculated based on the eigenvalues, and the minimum number of principal components m is selected based on the cumulative variance contribution rate and a preset threshold. The centered multi-source time series data is projected onto the eigenvectors of the selected m minimum principal component directions to generate the principal component score matrix Y. Based on the principal component score matrix Y, the linear and nonlinear relationships between its eigenvectors and pollutant concentrations are analyzed using the maximum information coefficient (MIC). Features are then ranked according to their MIC values ​​to identify key variables with strong correlation to the prediction of characteristic pollutants. The selected feature data for each process unit are then output. The selected feature data of each process unit are input into a network combining a deep belief network and a gated recurrent unit. After unsupervised deep feature learning and long-term temporal dependency capture in the sequence, the deep temporal features of each process unit are output.

[0013] In some embodiments, a comprehensive feature vector is generated, specifically as follows: In the process flow spatiotemporal grid data model, the calculated process contamination contribution vectors between all adjacent units are flattened to form the system correlation features. The comprehensive feature vector is generated by concatenating the depth-series features of each process unit with the system-related feature vector.

[0014] In some embodiments, the dynamic adaptive prediction model is a hybrid architecture model of Temporal Convolutional Network (TCN) and Transformer. The TCN is used to perform dilated convolution on the input data to extract multi-scale local temporal features; The Transformer receives the output of the TCN and processes the global dependencies in the composite feature vector through its multi-head self-attention mechanism.

[0015] In some embodiments, the dynamic adaptive prediction model has dynamic adaptive capabilities, achieved through a reinforcement learning-based dynamic parameter adjustment mechanism, specifically including: Define state space It should at least include the prediction error index of the current dynamic adaptive prediction model and the pollutant concentration change rate. and the rate of change of process parameters ; Define action space , as an instruction to adjust the hyperparameters of the dynamic adaptive prediction model; Based on reward function To evaluate the effectiveness of the action, the reward function is configured to encourage improved prediction accuracy and penalize model fluctuations and drastic parameter adjustments. When the preset triggering conditions are met, based on the current state space Select Action Space Adjust the hyperparameters of the dynamic adaptive prediction model to optimize its dynamic adaptive capability. The triggering conditions include at least one of the following: Pollutant concentration change rate Exceeding the first threshold; Process parameter change rate Exceeding the second threshold; The prediction error index exceeded the historical benchmark value for several consecutive periods.

[0016] In some embodiments, an optimization step of the dynamic adaptive prediction model is also included: According to the evaluation index of the dynamic adaptive prediction model, when the prediction accuracy does not meet the preset requirements, the importance of each feature vector in the comprehensive feature vector is re-evaluated, and the structural parameters of the dynamic adaptive prediction model or the fusion strategy is adjusted accordingly. A feature pollutant concentration prediction system based on spatiotemporal grids includes: The data acquisition module acquires multi-source time-series data of each process unit in the target wastewater treatment system; The process flow spatiotemporal grid data model construction module discretizes the continuous wastewater treatment process into a one-dimensional chain grid structure based on the target wastewater treatment process. Each grid cell in the chain grid structure represents a process unit. Each process unit is filled with a data vector containing the concentration of characteristic pollutants, process parameters, and hydraulic parameters at each sampling time to construct the process flow spatiotemporal grid data model. The contribution vector calculation module calls the spatiotemporal grid data model of the process flow based on multi-source time-series data to calculate the process contamination contribution vector between adjacent process units. The feature fusion module fuses the deep time-series features extracted from the multi-source time-series data of each process unit with the associated features of the wastewater treatment system represented by the process pollution contribution vector to generate a comprehensive feature vector. The pollutant concentration prediction module inputs the comprehensive feature vector into a trained dynamic adaptive prediction model and outputs the predicted value of the characteristic pollutant concentration of at least one target process unit in the wastewater treatment system within a specified future time period.

[0017] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the steps of the method for predicting the concentration of characteristic pollutants in a spatiotemporal grid as described above.

[0018] Compared with the prior art, the present invention has the following beneficial technical effects: This invention aims to provide a method for predicting the concentration of characteristic pollutants in emerging industries based on a spatiotemporal grid dynamic model. Addressing the shortcomings of existing technologies, such as insufficient consideration of the spatial structure of the process flow, one-sided data feature extraction, and difficulty in adapting models to dynamic process changes, this invention constructs a spatiotemporal grid data model of the process flow and calculates the process pollution contribution vector. This effectively models the spatiotemporal transfer and transformation process of characteristic pollutants between various process units, thus overcoming the deficiencies of traditional methods in considering the spatial structure of wastewater treatment processes and achieving accurate characterization of the pollutant migration and transformation trajectory along the process. By deeply fusing the unit depth time-series features extracted from deep networks with the system correlation features represented by the process pollution contribution vector, the generated comprehensive feature vector fully and adequately represents the spatiotemporal state of the system, solving the problem of one-sided feature extraction from multi-source heterogeneous data and laying a data foundation for high-precision prediction. Based on this, the established dynamic adaptive prediction model can adjust parameters in real time according to system state and performance feedback, significantly improving the model's adaptability to dynamic changes such as process fluctuations and production adjustments, as well as its long-term prediction stability. Ultimately, this method can not only accurately predict future pollutant concentration changes, but also characterize their migration and transformation paths in the treatment process, thereby enabling precise optimization of wastewater treatment process parameters (such as reagent dosing and equipment start-up and shutdown) to avoid undertreatment or waste of resources; at the same time, it can provide early warning of the risk of concentration exceeding the standard, providing a reliable guarantee for stable and compliant discharge.

[0019] Furthermore, this invention innovatively introduces a spatiotemporal grid data model of process flow and a process pollution contribution vector, treating the wastewater treatment system as a whole with spatial topology and dynamic interactions. This enables the model not only to learn the temporal variation patterns of pollutant concentrations but also to accurately characterize their spatial migration and transformation paths within the process flow, fundamentally improving the physical meaning and accuracy of the predictions, especially demonstrating stronger robustness during process fluctuations.

[0020] Furthermore, by plotting pollutant concentration change curves over a future period using predictive models and providing confidence intervals, the system can provide operators with intuitive early warnings, greatly shortening the response time from problem discovery to action, and enhancing the ability of emerging industry wastewater treatment systems to cope with sudden pollution load shocks from characteristic pollutants.

[0021] Furthermore, the dynamic adaptive model of this invention can intelligently adjust model parameters based on real-time data and prediction errors, ensuring the long-term effectiveness and stability of model prediction performance in the context of rapid iteration of industrial processes and frequent changes in data distribution. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the feature pollutant concentration prediction based on spatiotemporal grids according to the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0025] This invention aims to overcome the core deficiency of existing prediction methods in failing to adequately consider the "spatial structure of process flow" in wastewater treatment. To this end, the proposed technical solution first collects multi-source data on production processes, equipment status, water quality, and the environment through a sensor array. After preprocessing and feature filtering based on principal component analysis (PCA) and mutual information, an innovative "process flow spatiotemporal grid data model" is introduced to discretize the process, and a "process pollution contribution vector" is calculated to quantify the migration and transformation of pollutants between units. Based on this, a deep feature fusion model combining a deep belief network (DBN) and a gated recurrent unit (GRU) with an attention mechanism is used. Finally, a dynamic adaptive model based on a Transformer architecture combined with a temporal convolutional network (TCN) is employed for concentration prediction, ultimately outputting the concentration curves and confidence intervals of characteristic pollutants for future time periods.

[0026] Methods for predicting the concentration of characteristic pollutants based on spatiotemporal grids, such as Figure 1 As shown, it includes: Acquire multi-source time-series data of each process unit in the target wastewater treatment system; Based on multi-source time-series data, the spatiotemporal grid data model of the process flow is invoked to calculate the process contamination contribution vector between adjacent process units; The deep time-series features extracted from the multi-source time-series data of each process unit are fused with the correlation features of the wastewater treatment system represented by the process pollution contribution vector to generate a comprehensive feature vector. The comprehensive feature vector is input into the trained dynamic adaptive prediction model, which outputs the predicted value of the characteristic pollutant concentration of at least one target process unit in the wastewater treatment system within a specified future time period. The construction process of the spatiotemporal grid data model of the process flow is as follows: Based on the target wastewater treatment process, the continuous wastewater treatment process is discretized into a one-dimensional chain grid structure. Each grid cell in the chain grid structure represents a process unit. At each sampling time, a data vector containing the concentration of characteristic pollutants, process parameters, and hydraulic parameters is filled into each process unit to construct a spatiotemporal grid data model of the process flow.

[0027] This invention addresses wastewater treatment processes in emerging industries by constructing a spatiotemporal grid data model of the process flow. This model discretizes the continuous treatment process into a sequence of units with spatial topological relationships. The definition and calculation method of the process pollution contribution vector are used to quantify the dynamic impact of upstream process units on pollutant concentrations and environmental parameter changes in downstream units. The overall architecture and implementation process of a characteristic pollutant concentration prediction system for emerging industries, based on multi-source data fusion and a dynamic adaptive model, are also presented. This method not only achieves an accurate description of pollutant migration and transformation along the process but also enables the optimization of process parameters through back-calculation of predicted values. This improves prediction accuracy and provides a basis for decision-making in reducing characteristic pollutant emissions at the source.

[0028] In some embodiments, the multi-source time-series data includes monitoring data of characteristic pollutants, process parameter data associated with the pollutants, and environmental and conversion condition data. The preprocessing of the acquired multi-source time-series data includes: The collected multi-source time-series data is cleaned, outlier removed, and missing value interpolated and filled. Normalize the process parameter data associated with pollutants; The monitoring data of the characteristic pollutants are standardized. The monitoring data of the characteristic pollutants include the concentrations of fluorides and nanoparticles in the semiconductor workshop, the chemical oxygen demand and residual drug concentrations in the biopharmaceutical workshop, and the heavy metal ion concentrations in the new energy workshop. The process parameter data associated with the contaminants include the hydrofluoric acid flow rate in semiconductor lithography, the amount of antibiotics added in biopharmaceutical fermentation, and the electrolyte concentration in new energy battery coating. The environmental and conversion condition data include pH value, temperature, dissolved oxygen, and industrial process adjustment records.

[0029] In some embodiments, the pollution contribution vector between adjacent process units is calculated as follows: For each pair of adjacent upstream and downstream process units based on the spatiotemporal grid data model of the process flow, the process contamination contribution vector of the upstream process unit to the downstream process unit is calculated.

[0030] This includes calculating the process contamination contribution vector of the upstream unit to the downstream unit. i→j The specific formula is: i→j (t) = =

[0031] Where t is the sampling time; C is the concentration of the characteristic pollutant; The formula for calculating the normalized concentration contribution is as follows: , The value is a preset minimum positive number, where pH is the acidity or alkalinity, T is the temperature, and ORP is the redox potential.

[0032] In some embodiments, the deep time-series features extracted from the multi-source time-series data of each process unit are specifically as follows: The multi-source time series data of each process unit are centered, and a covariance matrix is ​​constructed based on the centered multi-source time series data. Then, the covariance matrix is ​​decomposed into eigenvalues ​​and eigenvectors. The cumulative variance contribution rate is calculated based on the eigenvalues, and the minimum number of principal components m is selected based on the cumulative variance contribution rate and a preset threshold. The centered multi-source time series data is projected onto the eigenvectors of the selected m minimum principal component directions to generate the principal component score matrix Y. Based on the principal component score matrix Y, the linear and nonlinear relationships between its eigenvectors and pollutant concentrations are analyzed using the maximum information coefficient (MIC). The features are sorted according to the MIC values ​​to screen out key variables that are strongly correlated with the prediction of characteristic pollutants, and the feature data of each process unit after screening are output. The selected feature data of each process unit are input into a network combining a deep belief network and a gated recurrent unit. After unsupervised deep feature learning and long-term temporal dependency capture in the sequence, the deep temporal features of each process unit are output.

[0033] In some embodiments, a comprehensive feature vector is generated, specifically as follows: In the process flow spatiotemporal grid data model, the calculated process contamination contribution vectors between all adjacent units are flattened to form the system correlation features. The comprehensive feature vector is generated by concatenating the depth-series features of each process unit with the system-related feature vector.

[0034] In some embodiments, the dynamic adaptive prediction model is a hybrid architecture model of Temporal Convolutional Network (TCN) and Transformer; The TCN is used to perform dilated convolution on the input data to extract multi-scale local temporal features; The Transformer receives the output of the TCN and processes the global dependencies in the composite feature vector through its multi-head self-attention mechanism.

[0035] In some embodiments, the dynamic adaptive prediction model has dynamic adaptive capabilities, achieved through a reinforcement learning-based dynamic parameter adjustment mechanism, specifically including: Define state space It should at least include the prediction error index of the current dynamic adaptive prediction model and the pollutant concentration change rate. and the rate of change of process parameters ; Define action space , as an instruction to adjust the hyperparameters of the dynamic adaptive prediction model; Based on reward function To evaluate the effectiveness of the action, the reward function is configured to encourage improved prediction accuracy and penalize model fluctuations and drastic parameter adjustments. When the preset triggering conditions are met, based on the current state space Select Action Space Adjust the hyperparameters of the dynamic adaptive prediction model to optimize its dynamic adaptive capability. The triggering conditions include at least one of the following: Pollutant concentration change rate Exceeding the first threshold; Process parameter change rate Exceeding the second threshold; The prediction error index exceeded the historical benchmark value for several consecutive periods.

[0036] In some embodiments, an optimization step of the dynamic adaptive prediction model is also included: According to the evaluation index of the dynamic adaptive prediction model, when the prediction accuracy does not meet the preset requirements, the importance of each feature vector in the comprehensive feature vector is re-evaluated, and the structural parameters of the dynamic adaptive prediction model or the fusion strategy is adjusted accordingly. A feature pollutant concentration prediction system based on spatiotemporal grids includes: The data acquisition module acquires multi-source time-series data of each process unit in the target wastewater treatment system; The process flow spatiotemporal grid data model construction module discretizes the continuous wastewater treatment process into a one-dimensional chain grid structure based on the target wastewater treatment process. Each grid cell in the chain grid structure represents a process unit. At each sampling time, each process unit is filled with a data vector containing the concentration of characteristic pollutants, process parameters, and hydraulic parameters to construct the process flow spatiotemporal grid data model.

[0037] The contribution vector calculation module calls the spatiotemporal grid data model of the process flow based on multi-source time-series data to calculate the process contamination contribution vector between adjacent process units. The feature fusion module fuses the deep time-series features extracted from the multi-source time-series data of each process unit with the associated features of the wastewater treatment system represented by the process pollution contribution vector to generate a comprehensive feature vector. The pollutant concentration prediction module inputs the comprehensive feature vector into a trained dynamic adaptive prediction model and outputs the predicted value of the characteristic pollutant concentration of at least one target process unit in the wastewater treatment system within a specified future time period. A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the steps of the method for predicting the concentration of characteristic pollutants in a spatiotemporal grid as described above.

[0038] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1 This invention aims to provide a method for predicting the concentration of characteristic pollutants in emerging industries based on a spatiotemporal grid dynamic model. Addressing the shortcomings of existing technologies, such as insufficient consideration of the spatial structure of process flow, incomplete data feature extraction, and difficulty in adapting models to dynamic process changes, this invention constructs a spatiotemporal grid data model of process flow and a pollution contribution vector, deeply integrating multi-source time-series data including production parameters, water quality monitoring, and environmental conditions, and establishing a dynamic adaptive prediction model. This method can not only accurately predict future pollutant concentration changes but also characterize their migration and transformation paths in the treatment process, thereby enabling precise optimization of wastewater treatment process parameters (such as reagent dosing and equipment start-up and shutdown) to avoid undertreatment or resource waste. Simultaneously, it provides early warning of the risk of concentration exceeding standards, offering a reliable guarantee for stable and compliant discharge.

[0040] To achieve the above objectives, this invention provides a method for predicting the concentration of characteristic pollutants based on a spatiotemporal grid dynamic model, the flowchart of which is shown below. Figure 1 As shown, the specific steps are as follows: Step 1: Multi-source time-series data acquisition. Details are as follows: (1) Specific monitoring data of characteristic pollutants: including the concentration of fluoride and nanoparticles in the semiconductor workshop, the chemical oxygen demand and residual drug concentration in the biopharmaceutical workshop, and the heavy metal ion concentration in the new energy workshop; the above parameters are obtained by deploying a dedicated sensor array, including the fluoride ion selective electrode in the semiconductor workshop (detection range 0.1-100mg / L, accuracy ±2%), the high performance liquid chromatography (HPLC) online monitoring of residual drugs in the biopharmaceutical workshop (detection limit 0.01mg / L), and the heavy metal electrochemical sensor in the new energy workshop (simultaneous detection of nickel / cobalt / manganese, response time <30s), with a data acquisition frequency of 1-5 minutes / time.

[0041] (2) Process parameter data related to pollutants: Process parameters directly related to the generation of characteristic pollutants are collected through the Industrial Internet of Things, such as the flow rate of hydrofluoric acid (L / h) in semiconductor lithography, the amount of antibiotics added in biopharmaceutical fermentation (mg / L), and the concentration of electrolyte in new energy battery coating (%). The collection frequency is synchronized with the sensor data.

[0042] (3) Environmental and transformation condition data: Key environmental parameters affecting the form of pollutants, including pH value (accuracy ±0.01), temperature (±0.1℃), dissolved oxygen (±0.1mg / L), and industrial process adjustment records (such as new drug production and battery formula updates).

[0043] Step 2: Preprocessing of multi-source time series data. Details are as follows: A Python script was used to clean the collected multi-source time-series data, removing outliers and duplicates, and interpolation was used to fill in missing values. Process parameters were standardized using Min-Max, and characteristic pollutant concentrations were standardized using Z-Score.

[0044] Step 3: Construct a spatiotemporal grid data model of the process flow and process contamination contribution vectors. This step is crucial for improving the model's understanding of the process. Details are as follows: (1) Constructing a spatiotemporal grid data model of the process flow: Based on the typical wastewater treatment process of the target industry, the entire continuous wastewater treatment process is discretized into a one-dimensional chain-like grid structure, where each grid cell represents a specific process unit. Examples are shown in Table 1 below: Table 1. Grid Construction of Representative Wastewater Treatment Processes in Three Emerging Industries

[0045] Each arrow (→) represents a grid cell. For example, in the semiconductor industry, the entire grid consists of [conditioning tank, reaction tank, coagulation tank, flocculation tank, sedimentation tank, neutralization tank]. Each grid cell (process unit) is filled with a data vector at each sampling time (e.g., every 5 minutes), including but not limited to: Concentration of characteristic pollutants (such as fluoride concentration, COD value, nickel ion concentration); Key process parameters (such as pH value, temperature, ORP (oxidation-reduction potential), and dosage); Hydraulic parameters (such as influent flow rate and residence time within the unit); Thus, a continuous and complex physicochemical process is transformed into a structured spatiotemporal data cube that can be processed by algorithms, resulting in a spatiotemporal grid data model of the process flow. (2) Calculation of process pollution contribution vector Based on the spatiotemporal grid data model of process flow, for each pair of adjacent upstream and downstream process units (denoted as U), i U j ), calculate upstream unit U i For downstream unit U j Pollution contribution vector i→j (t), this vector represents the amount of pollutants and information moving from U within the time interval [t, t+1]. i To U j The vector calculation formula is as follows: It quantifies and slices the dynamic process of the process flow by considering the transmission and transformation effects. i→j (t) = =

[0046] Where t is the sampling time; C is the concentration of characteristic pollutants (the core indicator characterizing pollution load). The formula for calculating the normalized concentration contribution is as follows: , It is a very small positive number, to prevent upstream concentration (t) being zero leads to a calculation error. The core component quantifies the impact efficiency of the upstream unit's pollutant concentration at time t on the concentration change of the downstream unit at the next time step. Positive values ​​indicate an increase in downstream concentration, which may be related to upstream transport; negative values ​​indicate a decrease in downstream concentration, which may be related to reaction removal. pH represents acidity / alkalinity, T represents temperature, and ORP represents redox potential. i represents the position index of the upstream process unit in the spatiotemporal grid chain structure; j represents the position index of the downstream process unit in the spatiotemporal grid chain structure.

[0047] Step 4: Data Feature Extraction and Fusion (1) Deep temporal feature extraction: 1. Perform principal component analysis (PCA) on multi-source time series data. The specific steps are as follows: First, the standardized process parameter data is centered so that the mean of each parameter is zero, and the covariance matrix of the centered data is calculated to reveal the correlation between different process parameters. Then, the covariance matrix is ​​decomposed into eigenvalues ​​and eigenvectors. The magnitude of the eigenvalues ​​represents the amount of original data information (variance) carried by each principal component, while the eigenvectors define the direction of the principal components. Subsequently, the cumulative variance contribution rate is calculated based on the eigenvalues, and the minimum number m of principal components whose cumulative contribution rate reaches a predetermined threshold (≥85%) is selected, thereby achieving dimensionality reduction.

[0048] Finally, the centered data is projected onto the eigenvectors of the selected m principal component directions to generate the final principal component score matrix Y, completing the transformation from high-dimensional space to low-dimensional space.

[0049] 2. Data feature screening: Based on the principal component score matrix Y, the maximum information coefficient (MIC) is used to analyze the linear and nonlinear relationships between each feature vector and the pollutant concentration. The features are sorted according to the MIC value to screen out key variables that are strongly correlated with the prediction of feature pollutants, effectively remove redundant features, and obtain the screened multi-source time series data features. The multi-source time-series data features (concentration, pH, ORP, flow rate, etc.) of each process unit, after preprocessing and feature screening, are input into a network combining a Deep Belief Network (DBN) and a Gated Recurrent Unit (GRU). The DBN performs unsupervised deep feature learning, while the GRU captures long-term temporal dependencies in the sequence, ultimately outputting the deep temporal features of each unit. The calculation formula for the GRU unit is as follows:

[0050]

[0051]

[0052]

[0053] Among them, z t To update the gate, r t To reset the door, h t Let x be the hidden state at the current moment. t For the current input, W z W r W is the weight matrix.

[0054] (2) Integration of related features: Then, all (n-1) process pollution contribution vectors calculated in the third step are flattened to form a complete associated feature vector of the wastewater treatment system.

[0055] (3) Feature splicing and fusion: The deep temporal features of each unit are concatenated with the associated feature vector of the entire system to generate the final comprehensive feature vector. This comprehensive feature vector contains information on both the internal state evolution of the unit and the interactions between units, providing a more comprehensive data foundation for subsequent predictions.

[0056] Step 5: Building a Dynamic Adaptive Model (1) Architecture selection for the dynamic adaptive model: A Transformer-based architecture is selected as the main body of the prediction model. Its multi-head self-attention mechanism can capture global dependencies in multi-source time-series data in parallel, effectively handling the nonlinear changes in the concentration of characteristic pollutants. The formula for calculating the attention weight is as follows:

[0057] Where Q, K, and V are the query, key, and value matrices, respectively, and d k is the dimension of the key vector.

[0058] Based on the Transformer, a Temporal Convolutional Network (TCN) is introduced to enhance the model's ability to extract local temporal features. The TCN uses dilated convolution to capture dependencies at different time scales, and its output serves as the input to the Transformer.

[0059] (2) Dynamic parameter adjustment of the dynamic adaptive model: A dynamic parameter adjustment mechanism based on reinforcement learning is designed, using prediction error (e.g., RMSE) and model stability as reward functions. Through interaction with the environment (i.e., real-time changing multi-source time-series data), the hyperparameters (e.g., learning rate, number of hidden layers, etc.) of the model are intelligently adjusted. When a sudden change in the concentration of a characteristic pollutant is detected or a significant adjustment in the production process is detected, the model can automatically trigger the parameter optimization process and quickly adapt to the new data distribution. The specific steps are as follows: 1. The intelligent agent is mainly responsible for the hyperparameter adjustment strategy of the decision model. The input is the current data distribution and model state, and the output is the hyperparameter adjustment action.

[0060] 2. Environmental simulation consists of real-time multi-source time-series data streams (including process parameters, pollutant concentrations, and equipment status) and the model's current prediction performance (such as errors and stability indicators).

[0061] 3. State Space The calculation formula is as follows:

[0062] in, Let be the root mean square error of the model prediction at time t. The variance of the last 10 predicted values ​​(measures stability; the smaller the value, the more stable the value). The rate of change of pollutant concentration between time t and time t-1 ( ), The overall variation range of process parameters at time t (such as the comprehensive coefficient of variation of hydrofluoric acid flow rate, etching temperature, etc.) The hyperparameters (learning rate) of the current model Number of neurons in the GRU hidden layer TCN expansion coefficient wait); 4. Movement space The calculation formula for (hyperparameter adjustment dimension) is as follows:

[0063] in, (Learning rate fine-tuning) (Adjustment of the number of neurons in the GRU hidden layer) (TCN expansion coefficient adjustment, must be a power of 2).

[0064] (3) Reward function design: Reward function To maximize prediction accuracy (reduce RMSE) and minimize parameter fluctuations (maintain stability), the calculation formula is as follows:

[0065] in, , For industrial scenarios, the weights are adjusted dynamically. The first parameter is to give a positive reward if the current RMSE is lower than the previous time step, encouraging accuracy improvement; the second is to penalize high variance (model instability); and the third is to penalize excessive parameter adjustments (to avoid model oscillation).

[0066] Parameter optimization is triggered when any of the following conditions are met (optimization metrics are derived from relevant industry big data): 1. Sudden changes in pollutant concentration: >0.2 (change rate exceeding 20%, such as a sudden increase in semiconductor fluoride concentration).

[0067] 2. Significant process adjustments: >0.15 (Comprehensive changes in process parameters exceeding 15%, such as material change in biopharmaceutical fermenters).

[0068] 3. Decrease in model accuracy: over 5 consecutive time periods ( (1.2 times the historical best value) (4) Optimize the process: 1. The agent learns the policy through the Proximal Policy Optimization (PPO) algorithm. The strategy is updated stably while ensuring exploratory nature. (The agent is a reinforcement learning decision-making model based on the PPO algorithm framework. Essentially, it is a deep neural network that learns and optimizes the strategy through continuous interaction with the characteristic pollutant concentration monitoring and prediction system.) 2. After each trigger, perform a 50-step parameter search on the sliding window data (nearly 1000 records) and select the action with the highest reward. Update the model hyperparameters.

[0069] 3. After updating, freeze the parameters for 5 cycles (e.g., 30 minutes) to avoid frequent adjustments. Once the model stabilizes, re-enter the monitoring state.

[0070] (5) Model training: The model is trained using historical data from the past three years. The Adam (Adaptive Moments Estimation) optimization algorithm is employed to minimize the Huber loss function between predicted and actual values. This loss function is insensitive to outliers and is suitable for predicting characteristic pollutant concentrations. The learning rate is reduced when the RMSE (Real-Time Sequence of Errors) increases. An online learning mechanism is implemented to update model parameters daily with incremental new data. When the characteristic pollutant concentration fluctuates by more than 20%, a reinforcement learning parameter adjustment process is automatically triggered.

[0071] Step 6: Model Evaluation and Optimization The comprehensive feature vector is input into the trained model, which outputs the predicted pollutant concentration values ​​for each process unit over a future period. Root mean square error (RMSE) and coefficient of determination (R²) are used as core evaluation metrics, and the model's prediction accuracy for each process unit (spatial dimension) is also examined.

[0072] (1) Core evaluation indicators: The root mean square error (RMSE) and the coefficient of determination (R²) are used to evaluate the model's predicted data, and the calculation formulas are as follows: 1. Root Mean Square Error (RMSE): Measures the overall error magnitude and is sensitive to large errors.

[0073]

[0074] 2. R² Coefficient of Determination: Measures the model’s ability to explain the variation in the data (the closer to 1, the better).

[0075]

[0076] in, This refers to the "predicted value" (model output value) of the characteristic pollutant concentration. This refers to the "true value" (actual monitoring value) of the characteristic pollutant concentration.

[0077] 3. Spatial Prediction Accuracy Evaluation: In addition to the overall accuracy, the accuracy of the model's concentration prediction for each process unit is evaluated, the RMSE of each unit is calculated, and the predictive performance of the model on the spatial structure of the process flow is analyzed.

[0078] Based on the evaluation results, the predictive performance of the model under different industries and pollutants is analyzed, and data fusion strategies and model parameters are optimized for weak links. The RMSE of the prediction results is calculated daily. If the RMSE exceeds a specific industry threshold (e.g., semiconductors: 0.3, biomedicine: 0.5, new energy: 0.4), an optimization process is initiated. Feature importance is reassessed, and the number of Transformer model layers and attention heads is optimized by adjusting attention weight allocation. If necessary, the number of TCN dilated convolutional layers can be increased to capture longer temporal dependencies. The model is comprehensively evaluated and retrained monthly to ensure continuous improvement in prediction accuracy. The model's prediction output includes not only the concentration of the effluent unit but also the concentration distribution of characteristic pollutants along the process flow, thus depicting the migration and transformation trajectory of characteristic pollutants in the spatiotemporal grid. This provides accurate and forward-looking decision support for process parameter optimization (such as precision dosing) and early warning of exceedance risks.

[0079] One embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the above-described dynamic scene image deblurring method; or, when the processor executes the computer program, it implements the functions of each module in the above-described dynamic scene image deblurring system.

[0080] The computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention.

[0081] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0082] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0083] The memory can be used to store the computer program and / or module. The processor implements various functions of the terminal device by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory.

[0084] If the modules / units integrated in the terminal device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0085] Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-described dynamic scene image deblurring method. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate form.

[0086] The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0087] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0088] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the accompanying drawings and the above description. However, any modifications, alterations, or variations made by those skilled in the art without departing from the scope of the present invention, utilizing the disclosed technical content, are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. A method for predicting the concentration of characteristic pollutants based on spatiotemporal grids, characterized in that, include: Acquire multi-source time-series data of each process unit in the target wastewater treatment system; Based on multi-source time-series data, the spatiotemporal grid data model of the process flow is invoked to calculate the process contamination contribution vector between adjacent process units; The deep time-series features extracted from the multi-source time-series data of each process unit are fused with the correlation features of the wastewater treatment system represented by the process pollution contribution vector to generate a comprehensive feature vector. The comprehensive feature vector is input into the trained dynamic adaptive prediction model, which outputs the predicted value of the characteristic pollutant concentration of at least one target process unit in the wastewater treatment system within a specified future time period. The construction process of the spatiotemporal grid data model of the process flow is as follows: Based on the target wastewater treatment process, the continuous wastewater treatment process is discretized into a one-dimensional chain grid structure. Each grid cell in the chain grid structure represents a process unit. At each sampling time, a data vector containing the concentration of characteristic pollutants, process parameters, and hydraulic parameters is filled into each process unit to construct a spatiotemporal grid data model of the process flow.

2. The method for predicting characteristic pollutant concentrations based on spatiotemporal grids according to claim 1, characterized in that, The multi-source time-series data includes monitoring data of characteristic pollutants, process parameter data associated with the pollutants, and environmental and conversion condition data. The preprocessing of the acquired multi-source time-series data includes: The collected multi-source time-series data is cleaned, outlier removed, and missing value interpolated and filled. Normalize the process parameter data associated with pollutants; The monitoring data of the characteristic pollutants are standardized. The monitoring data of the characteristic pollutants include the concentrations of fluorides and nanoparticles in the semiconductor workshop, the chemical oxygen demand and residual drug concentrations in the biopharmaceutical workshop, and the heavy metal ion concentrations in the new energy workshop. The process parameter data associated with the contaminants include the hydrofluoric acid flow rate in semiconductor lithography, the amount of antibiotics added in biopharmaceutical fermentation, and the electrolyte concentration in new energy battery coating. The environmental and conversion condition data include pH value, temperature, dissolved oxygen, and industrial process adjustment records.

3. The method for predicting characteristic pollutant concentrations based on spatiotemporal grids according to claim 1, characterized in that, The pollution contribution vector between adjacent process units is calculated as follows: For each pair of adjacent upstream and downstream process units based on the spatiotemporal grid data model of process flow, calculate the process contamination contribution vector of the upstream process unit to the downstream process unit. This includes calculating the process contamination contribution vector of the upstream unit to the downstream unit. i→j The specific formula is: i→j （t）= = Where t is the sampling time; C is the concentration of the characteristic pollutant; The formula for calculating the normalized concentration contribution is as follows: , The value is a preset minimum positive number, where pH is the acidity or alkalinity, T is the temperature, and ORP is the redox potential.

4. The method for predicting characteristic pollutant concentrations based on spatiotemporal grids according to claim 1, characterized in that, The deep temporal features extracted from the multi-source time-series data of each process unit are as follows: The multi-source time series data of each process unit are centered, and a covariance matrix is ​​constructed based on the centered multi-source time series data. Then, the covariance matrix is ​​decomposed into eigenvalues ​​and eigenvectors. The cumulative variance contribution rate is calculated based on the eigenvalues, and the minimum number of principal components m is selected based on the cumulative variance contribution rate and a preset threshold. The centered multi-source time series data is projected onto the eigenvectors of the selected m minimum principal component directions to generate the principal component score matrix Y. Based on the principal component score matrix Y, the linear and nonlinear relationships between its eigenvectors and pollutant concentrations are analyzed using the maximum information coefficient (MIC). The features are sorted according to the MIC value to screen out key variables that are strongly correlated with the prediction of characteristic pollutants, and the feature data of each process unit after screening are output. The selected feature data of each process unit are input into a network combining a deep belief network and a gated recurrent unit. After unsupervised deep feature learning and long-term temporal dependency capture in the sequence, the deep temporal features of each process unit are output.

5. The method for predicting characteristic pollutant concentrations based on spatiotemporal grids according to claim 1, characterized in that, The comprehensive feature vector is generated as follows: In the process flow spatiotemporal grid data model, the calculated process contamination contribution vectors between all adjacent units are flattened to form the system correlation features. The comprehensive feature vector is generated by concatenating the depth-series features of each process unit with the system-related feature vector.

6. The method for predicting characteristic pollutant concentrations based on spatiotemporal grids according to claim 1, characterized in that, The dynamic adaptive prediction model is a hybrid architecture model of TCN and Transformer; The TCN is used to perform dilated convolution on the input data to extract multi-scale local temporal features; The Transformer receives the output of the TCN and processes the global dependencies in the composite feature vector through its multi-head self-attention mechanism.

7. The method for predicting characteristic pollutant concentrations based on spatiotemporal grids according to claim 1, characterized in that, The dynamic adaptive prediction model has dynamic adaptive capabilities, which are achieved through a dynamic parameter adjustment mechanism based on reinforcement learning, specifically including: Define state space It should at least include the prediction error index of the current dynamic adaptive prediction model and the pollutant concentration change rate. and the rate of change of process parameters ; Define action space , as an instruction to adjust the hyperparameters of the dynamic adaptive prediction model; Based on reward function To evaluate the effectiveness of the action, the reward function is configured to encourage improved prediction accuracy and penalize model fluctuations and drastic parameter adjustments. When the preset triggering conditions are met, based on the current state space Select Action Space Adjust the hyperparameters of the dynamic adaptive prediction model to optimize its dynamic adaptive capability. The triggering conditions include at least one of the following: Pollutant concentration change rate Exceeding the first threshold; Process parameter change rate Exceeding the second threshold; The prediction error index exceeded the historical benchmark value for multiple consecutive periods.

8. The method for predicting characteristic pollutant concentrations based on spatiotemporal grids according to claim 1, characterized in that, It also includes optimization steps for the dynamic adaptive prediction model: Based on the evaluation metrics of the dynamic adaptive prediction model, when the prediction accuracy does not meet the preset requirements, the importance of each feature vector in the comprehensive feature vector is re-evaluated, and the structural parameters of the dynamic adaptive prediction model or the fusion strategy is adjusted accordingly.

9. A characteristic pollutant concentration prediction system based on spatiotemporal grids, characterized in that, include: The data acquisition module acquires multi-source time-series data of each process unit in the target wastewater treatment system; The process flow spatiotemporal grid data model construction module discretizes the continuous wastewater treatment process into a one-dimensional chain grid structure based on the target wastewater treatment process. Each grid cell in the chain grid structure represents a process unit. Each process unit is filled with a data vector containing the concentration of characteristic pollutants, process parameters, and hydraulic parameters at each sampling time to construct the process flow spatiotemporal grid data model. The contribution vector calculation module calls the spatiotemporal grid data model of the process flow based on multi-source time-series data to calculate the process contamination contribution vector between adjacent process units. The feature fusion module fuses the deep time-series features extracted from the multi-source time-series data of each process unit with the associated features of the wastewater treatment system represented by the process pollution contribution vector to generate a comprehensive feature vector. The pollutant concentration prediction module inputs the comprehensive feature vector into a trained dynamic adaptive prediction model and outputs the predicted value of the characteristic pollutant concentration of at least one target process unit in the wastewater treatment system within a specified future time period.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the characteristic pollutant concentration prediction method based on a spatiotemporal grid as described in any one of claims 1-8.

Images