A multi-source data fusion and intelligent modeling optimization method for a water treatment system
By integrating multi-source data and using intelligent modeling, and leveraging large language models and machine learning techniques, the problems of data dispersion and unstructured processes in electrochemical water treatment have been solved. This has enabled automated data integration and systematic optimization, provided efficient parameter optimization strategies, and improved the design and operation efficiency of electrochemical water treatment systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2026-04-01
- Publication Date
- 2026-07-03
AI Technical Summary
In practical applications, electrochemical water treatment technology faces challenges due to the complex and nonlinear coupling effects of multidimensional parameters such as electrode materials, operating conditions, and water quality characteristics, making performance analysis difficult. Existing research data is scattered and unstructured, leading to difficulties in modeling and optimization relying on empirical trial and error, resulting in low efficiency.
By employing multi-source data fusion and intelligent modeling methods, key parameter information is extracted through a large language model to construct a structured database. Machine learning models such as gradient boosting decision trees are used for prediction, and SHAP analysis is combined to generate optimization strategies, thereby achieving automated and standardized data integration and systematic optimization.
It enables efficient and rapid optimization of electrochemical water treatment systems, provides high-quality datasets and accurate parameter optimization suggestions, reduces trial-and-error costs, reveals nonlinear laws that are difficult to detect with traditional experience, and improves design and optimization efficiency.
Smart Images
Figure CN122333976A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of water treatment system optimization technology, and relates to an intelligent modeling and optimization method for electrochemical water treatment systems, specifically an intelligent modeling and optimization method for electrochemical water treatment systems based on multi-source data fusion and machine learning. Background Technology
[0002] Electrochemical water treatment technology has shown great potential in the removal of recalcitrant organic pollutants due to its high efficiency and cleanliness. However, the practical application and scale-up of this technology face severe challenges. The core issue lies in the complex and nonlinear coupling of its performance with multiple dimensions of parameters, such as electrode materials, operating conditions, and water quality characteristics, and the underlying mechanism has not yet been fully elucidated.
[0003] Currently, research on electrochemical water treatment has generated a massive amount of literature data, but this knowledge is highly fragmented. Key experimental parameters are scattered across tens of thousands of papers in unstructured and varied formats, making it difficult to systematically integrate and analyze this data. This data barrier severely restricts researchers' ability to uncover potential patterns and build reliable predictive models from existing research. Existing system optimization methods largely rely on researchers' experience and repeated trial and error, a time-consuming and laborious process that makes it difficult to guarantee finding the globally optimal solution.
[0004] In recent years, although machine learning technology has been applied in process industry optimization, progress in intelligent modeling and optimization research in the field of electrochemical water treatment has been slow due to the lack of high-quality, standardized, machine-readable datasets. Traditional data processing methods rely on manual extraction, which is inefficient and error-prone, and cannot meet the needs of large-scale knowledge discovery. Therefore, there is an urgent need for a systematic approach that can automatically integrate scattered knowledge, construct high-quality datasets, and perform intelligent modeling and optimization based on these datasets, in order to promote the paradigm shift of electrochemical water treatment from experience-driven to data-driven and prediction-driven approaches. Summary of the Invention
[0005] To overcome the shortcomings of existing electrochemical water treatment methods, such as difficulty in modeling and reliance on trial and error due to scattered and non-standardized data, this invention provides a multi-source data fusion and intelligent modeling optimization method for water treatment systems.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] A method for multi-source data fusion and intelligent modeling optimization for water treatment systems includes the following steps:
[0008] Step 1: Obtain multi-source unstructured text data related to the target water treatment system, including:
[0009] The multi-source unstructured text data includes one or more of the following: scientific research papers, patent documents, and technical reports;
[0010] Step 2: Based on a pre-defined prompting strategy that includes domain knowledge constraints, the large language model is invoked to extract key parameter information from the unstructured text data, whereby:
[0011] The prompting strategy adopts a hierarchical structure, specifically including a role definition layer, a task description layer, a rule guidance layer, and an output format constraint layer;
[0012] The role definition layer is used to define the large language model as an expert in the field of electrochemical water treatment. This is not a simple label, but a fundamental way to guide the model's thinking. It enables the model to prioritize the use of professional knowledge in the field of electrochemistry when analyzing literature, such as understanding the accurate meaning and context of terms like "BDD anode", "current density", and "reaction rate constant", thereby avoiding misinterpretation or generalization that may occur with general models.
[0013] The task description layer is used to explicitly specify the set of target parameters to be extracted from the literature. This layer precisely defines the extraction target, which concretizes the abstract target of "extracting information from the literature" into a clear list of 14 target parameters. The target parameters include pollutant category, pollutant concentration, anode material, anode area, cathode material, cathode area, electrolyte, electrolyte concentration, current density, pH, temperature, electrode spacing, reaction rate constant, and solution volume. This provides the model with a clear and executable checklist to ensure that the scope of the extraction work is clear and no key information is missed.
[0014] The rule guidance layer provides domain rules to ensure consistency and accuracy in data extraction. This layer is crucial for ensuring data quality and consistency. It addresses the unavoidable heterogeneity in literature reports by embedding 18 detailed instructions and domain logic. Specific functions include: missing value handling: requiring unmentioned parameters to be explicitly marked as null; standardization: mandating unit conversions (e.g., unifying all concentrations to mg / L). -1 (or mM), and use standard chemical names (such as Na2SO4 instead of “sodium sulfate”); logical inference: allow the model to make inferences under reasonable circumstances (e.g., extrapolate “room temperature” as 25°C); filtering: exclude data from non-electrochemical processes or involving non-specific contaminants; these rules transform raw, unstructured text descriptions into clean, uniform, machine-readable data that can be directly used for downstream analysis;
[0015] The output format constraint layer is used to force the large language model to output results in a predetermined structured data format and to standardize the numerical values and units. This layer strictly defines the delivery format of the results. It forces the model to output in a preset JSON format, where each key in the JSON corresponds to a parameter defined in the "task description layer".
[0016] The key parameter information includes the electrode material properties, operating conditions, pollutant characteristics, and reaction kinetic parameters of the water treatment system;
[0017] Step 3: Standardize the extracted key parameter information and construct a structured database;
[0018] Step 4: Based on a structured database, train a machine learning model to predict the performance indicators of the water treatment system, specifically its key performance metrics.
[0019] The machine learning model is a gradient boosting decision tree model, a random forest model, a support vector regression model, or an artificial neural network model.
[0020] The performance index is the reaction rate constant of pollutants during electrochemical water treatment;
[0021] Step 5: Using model interpretability analysis, quantify the impact and interaction of key parameter information on performance indicators, including:
[0022] The model interpretability analysis method is the SHAP analysis method;
[0023] Step 6: Based on the degree of influence and interaction, generate parameter optimization strategies for the target water treatment system, wherein:
[0024] The parameter optimization strategy includes recommended value ranges for key operating parameters and suggestions for collaborative optimization among parameters.
[0025] After generating the parameter optimization strategy, the above method also includes step 7: applying the optimization strategy to an actual water treatment system for verification, and using the verification results as new data to supplement the structured database to iteratively update the machine learning model.
[0026] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.
[0027] Compared with the prior art, the present invention has the following advantages:
[0028] 1. This invention realizes the automated and standardized integration of scattered knowledge in the field of electrochemical water treatment, solves the data barrier problem, and provides a high-quality benchmark dataset for subsequent research.
[0029] 2. By constructing a high-precision machine learning prediction model, this invention can quickly evaluate the system performance under different parameter combinations, greatly reducing the time and resource consumption required by traditional trial and error methods.
[0030] 3. This invention utilizes model interpretability analysis to quantitatively reveal the contribution and interaction of various parameters in complex systems, discovering nonlinear laws that are difficult to detect through traditional empirical analysis, and providing a new perspective for mechanism research.
[0031] 4. The method of the present invention has high versatility and scalability. Its technical framework is not only applicable to electrochemical advanced oxidation processes, but can also be applied to the modeling and optimization of other types of water treatment processes (such as adsorption, membrane filtration, and biological treatment) after appropriate adjustments, and has broad application prospects.
[0032] 5. This invention realizes a closed loop from scattered data to system knowledge, which can significantly improve the design and optimization efficiency of electrochemical water treatment systems, reduce trial and error costs, and solve the problem of difficulty in systematic modeling and parameter optimization in the field of electrochemical water treatment due to the scattered experimental data and inconsistent formats. Attached Figure Description
[0033] Figure 1 A schematic diagram of the overall methodology for multi-source data fusion and intelligent modeling optimization for water treatment systems;
[0034] Figure 2 The fitting performance of four machine learning models;
[0035] Figure 3 This is a single-factor variable dependency plot of pollutant concentration and current density. Detailed Implementation
[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.
[0037] This invention provides a multi-source data fusion and intelligent modeling optimization method for water treatment systems, capable of automatically fusing multi-source data, constructing predictive models, and intelligently generating optimization strategies. This method transforms unstructured literature knowledge into practically applicable optimization strategies through a closed-loop workflow. For example... Figure 1 As shown, the specific steps are as follows:
[0038] Step 1: Collect relevant literature on the target water treatment system (such as electrochemical advanced oxidation treatment of antibiotic wastewater) from public databases or local libraries as raw unstructured data sources.
[0039] Step 2: Design and apply a hierarchical hinting engineering framework that integrates domain knowledge to drive a large language model to accurately extract key parameters from literature. This framework ensures the accuracy and consistency of extraction by defining model roles (e.g., electrochemical experts), specifying extraction tasks (e.g., 14 target parameters: pollutant category, pollutant concentration, anode material, anode area, cathode material, cathode area, electrolyte, electrolyte concentration, current density, pH, temperature, electrode spacing, reaction rate constant, and solution volume), setting constraint rules (e.g., unit normalization), and enforcing structured output (e.g., JSON format). For pollutants, molecular fingerprints (e.g., MACCS keys) or quantum chemical descriptors are generated using computational chemistry methods, transforming molecular structure information into numerical features.
[0040] Step 3: The extracted raw data needs to be processed by a standardization module to unify terminology, numerical ranges, and units of measurement, eliminate differences in expression between different documents, and ultimately construct a high-quality structured database. This database integrates multi-dimensional information such as electrode materials, operating variables, contaminant properties, and performance indicators.
[0041] Step 4: Construct and train the intelligent prediction model. Machine learning algorithms, especially gradient boosting decision tree algorithms (such as XGBoost and LightGBM), are selected as the core modeling tools. The aforementioned mixed feature matrix is used as input, and key performance indicators (such as the reaction rate constant k for pollutant degradation) are used as output targets. The dataset is divided into training, validation, and test sets. The model is trained on the training set, and hyperparameters are tuned using the validation set to prevent overfitting. Finally, the model's prediction accuracy is evaluated on the test set (e.g., R-value). 2 (RMSE). This step enables the model to learn complex nonlinear mappings from the fused, rich data.
[0042] Step 5: Perform model interpretability analysis. Interpretation analysis techniques such as SHAP and partial dependency graphs are used to analyze the trained prediction model. SHAP analysis calculates a SHAP value for each feature of each sample, quantifying the contribution of that feature to the prediction result for that sample relative to the baseline (the average prediction value of all samples). By analyzing the distribution of feature SHAP values across all samples, the key drivers with the greatest impact on performance metrics can be identified globally. Partial dependency graphs can show the average trend of model prediction output as a feature value changes, and are particularly suitable for revealing nonlinear relationships (such as initial promotion followed by inhibition) and interaction effects between features.
[0043] Step 6: Generate Optimization Decision Recommendations. Leveraging the high-precision predictive capabilities of the integrated model and the mechanistic insights provided by interpretable analysis, virtual "parameter scans" or optimization algorithms (such as Bayesian optimization) are performed to find the optimal solution for specific water treatment objectives (e.g., maximizing degradation rate, minimizing energy consumption). For example, based on the core variables (e.g., pollutant concentration, current density) and their interaction patterns identified by SHAP analysis, an optimal current density operating range can be recommended; based on the threshold effect revealed by partial dependency plots, parameters can be avoided in performance-degrading regions. The final output includes an optimized solution containing recommended electrode types, operating parameter combinations, and reactor adjustment suggestions.
[0044] Example 1: Intelligent modeling and optimization of wastewater from electrochemical degradation of antibiotics:
[0045] This embodiment uses the electrochemical advanced oxidation degradation of antibiotics in water as an application scenario to specifically implement the method described in this invention.
[0046] The first step was data acquisition and preprocessing. Using keywords such as "electrochemical advanced oxidation" and "antibiotic removal," a search was conducted in the Web of Science Core Collection, downloading 245 full-text (PDF) articles. All PDF files were converted to Markdown format, preserving text structure and tables, using a document parsing tool. Subsequently, a rule-based paragraph filter was applied to remove irrelevant sections such as introductions, acknowledgments, and references from each article, resulting in concise text data.
[0047] The second step is key parameter extraction. A hierarchical prompting framework is designed to guide the large language model (GPT-4 model is used in this embodiment) in information extraction. The prompts include: (a) role definition: "You are an electrochemical water treatment expert"; (b) task description: "Please extract the following 14 parameters from the provided text..."; (c) rule guidance: "If a parameter is not explicitly mentioned in the text, the output will be null; pay attention to consistent units, for example, all concentration units should be consistent to mg / L..."; (d) output format constraints: "Please output in strict JSON format, with the following keys: pollutant, concentration, anode_material...". The prompts and preprocessed text are input into the model together for batch data extraction.
[0048] The third step is data standardization and database construction. This involves post-processing the raw JSON data extracted from the model. This includes: removing "Ti / SnO" from the material description... 2 -Sb”, SnO 2-Sb / Ti”, etc., are uniformly replaced by “Ti / SnO”. 2 -Sb”; Convert concentration values “50 mg / L” and “0.05 g / L” to the numerical value “50” and the unit “mg / L”; Convert different expressions of the reaction rate constant k (e.g., “0.05 min”) to the numerical value “50” and the unit “mg / L”; -1 “k = 3.0 h” -1 ) Unified as "min -1 The values are expressed in units of "". Ultimately, a structured database containing 2199 records is constructed, with each record covering fields such as pollutant, concentration, anode material, cathode material, current density, pH, and rate constant.
[0049] The fourth step is model training and optimization. The logarithm of the rate constant k (log k) is used as the prediction target (performance metric). Fourteen features, including pollutant concentration, current density, anode area, and pH value, are selected as inputs. The dataset is randomly divided into training and test sets in an 8:2 ratio. XGBoost, Random Forest (RF), Support Vector Regression (SVR), and Artificial Neural Network (ANN) models are trained using the training set, and hyperparameters are tuned using grid search or Bayesian optimization. Evaluation on the test set shows that the XGBoost model performs best, with a coefficient of determination R0. 2 The root mean square error (RMSE) reached 0.85, which is the lowest. Figure 2 ).
[0050] Step 5: Model Interpretation and Policy Generation. SHAP analysis is performed on the trained optimal XGBoost model. The analysis results show that pollutant concentration and current density are the two factors with the greatest impact on log k. Further, a SHAP dependency plot of pollutant concentration is plotted. Figure 3 The marginal effect of performance improvement is most significant in the current density range of 5-50 mA / cm². Based on these findings, the resulting optimization strategy report recommends that, for target antibiotic wastewater, the current density should be preferentially increased to 5-50 mA / cm². 2 Within this range, a significant improvement in processing efficiency is expected.
[0051] Example 2, Extended Application:
[0052] To demonstrate the versatility of this method, the data source and target can be changed. For example, for the treatment of dye-containing wastewater, relevant electrochemical literature can be collected, and steps one through three can be repeated to construct a dye wastewater treatment database. In step four, the prediction target can be replaced with decolorization rate or total organic carbon removal rate. Similarly, by training a machine learning model and performing SHAP analysis, key operating parameters (such as electrolyte type, initial pH, etc.) and their optimization ranges affecting dye decolorization efficiency can be identified, thereby providing targeted guidance for optimizing the electrochemical treatment process of this type of wastewater. This embodiment shows that the method framework of this invention is transferable and can be applied to water treatment system optimization scenarios for different pollutants by changing the domain data source.
Claims
1. A multi-source data fusion and intelligent modeling optimization method for a water treatment system, characterized in that The method includes the following steps: Step 1: Obtain multi-source unstructured text data related to the target water treatment system; Step 2: Based on a preset prompting strategy that includes domain knowledge constraints, call the large language model to extract key parameter information from unstructured text data; Step 3: Standardize the extracted key parameter information and construct a structured database; Step 4: Based on the structured database, train a machine learning model to predict the performance indicators of the water treatment system, in order to predict the key performance indicators of the water treatment system. Step 5: Use model interpretability analysis to quantify the impact and interaction of key parameter information on performance indicators; Step 6: Generate parameter optimization strategies for the target water treatment system based on the degree of influence and interaction.
2. The multi-source data fusion and intelligent modeling optimization method for water treatment system of claim 1, wherein The multi-source unstructured text data includes one or more of the following: scientific research papers, patent documents, and technical reports.
3. The multi-source data fusion and intelligent modeling optimization method for water treatment systems according to claim 1, characterized in that... The prompting strategy adopts a hierarchical structure, specifically including a role definition layer, a task description layer, a rule guidance layer, and an output format constraint layer. Specifically: the role definition layer defines the large language model as an expert in the field of electrochemical water treatment; the task description layer explicitly specifies the set of target parameters to be extracted from the literature, including pollutant category, pollutant concentration, anode material, anode area, cathode material, cathode area, electrolyte, electrolyte concentration, current density, pH, temperature, electrode spacing, reaction rate constant, and solution volume; the rule guidance layer provides domain rules to ensure consistency and accuracy in extraction; and the output format constraint layer mandates that the large language model output results in a predetermined structured data format and standardizes the numerical values and units.
4. The multi-source data fusion and intelligent modeling optimization method for water treatment systems according to claim 1, characterized in that... The key parameter information includes the electrode material properties, operating conditions, pollutant characteristics, and reaction kinetic parameters of the water treatment system.
5. The multi-source data fusion and intelligent modeling optimization method for water treatment systems according to claim 1, characterized in that... The machine learning model is a gradient boosting decision tree model, a random forest model, a support vector regression model, or an artificial neural network model.
6. The multi-source data fusion and intelligent modeling optimization method for water treatment systems according to claim 1, characterized in that... The performance index is the reaction rate constant of pollutants during electrochemical water treatment.
7. The multi-source data fusion and intelligent modeling optimization method for water treatment systems according to claim 1, characterized in that... The model interpretability analysis method is the SHAP analysis method.
8. The multi-source data fusion and intelligent modeling optimization method for water treatment systems according to claim 1, characterized in that... The parameter optimization strategy includes recommended value ranges for key operating parameters and suggestions for collaborative optimization among parameters.
9. The multi-source data fusion and intelligent modeling optimization method for water treatment systems according to claim 1, characterized in that... The method further includes step 7: applying the optimization strategy to an actual water treatment system for verification, and using the verification results as new data to supplement the structured database to iteratively update the machine learning model.
10. A computer-readable storage medium, characterized in that... The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-9.