Energy-water-carbon chain coupling evaluation method and system based on artificial intelligence driving

Abstract

The invention relates to the technical field of integrated energy system and resource environment management, in particular to an energy-water-carbon chain coupling evaluation method and system based on artificial intelligence driving. The method comprises the following steps: collecting water, energy, carbon and scene associated data, delimiting an evaluation boundary and determining a scene type; actual nodes are screened, missing information of the actual nodes is reinforced according to different time scales, and meanwhile two types of virtual nodes are identified and constructed; on the basis of combination of a physical mechanism and an AI algorithm, the association weight of the actual node and the-water-energy-carbon core element and a virtual node influence correction coefficient are calculated; a deep learning hybrid coupling model containing a physical constraint layer is constructed and optimized; performing multi-time scale coupling evaluation and dynamic prediction based on the model; and outputting a report of the optimization scheme. According to the invention, by introducing a dimension and virtual node processing mechanism and combining a physically constrained AI model, the evaluation precision and reliability of the-water-energy-carbon four-dimensional coupling system are improved, and multi-scene decision-making requirements are met.

Description

Technical Field

[0001] This invention relates to the field of integrated energy systems and resource and environmental management technology, and more specifically to an artificial intelligence-driven method and system for evaluating the energy-water-carbon chain coupling. The energy-water-carbon chain refers to the energy-water-energy-carbon coupling chain. Background Technology

[0002] Integrated energy systems and resource and environmental management involve a "coupled assessment" of the interrelationships and overall performance of multiple subsystems such as energy, water resources, and carbon emissions. The aim is to achieve synergistic optimization and sustainable development by analyzing their interactions. However, this coupled assessment faces three main technical challenges: First, the assessment dimensions lack key efficiency indicators. Existing solutions mostly focus on "flow statistics" of energy consumption, water resource usage, and carbon emissions, without including "flow" as a dimension that reflects the essence of resource utilization efficiency. This makes it impossible to effectively identify hidden resource waste patterns such as "high flow but low efficiency". Second, the data defect handling mechanism restricts the accuracy of the assessment. In the face of the common situation of missing or abnormal data in practice, the existing solution does not distinguish between two fundamentally different data defect scenarios: "missing high-frequency essential key attributes (such as the intake temperature of a gas turbine, without which efficiency cannot be calculated)" and "abnormal or unreliable data (such as manually recorded emissions from chemical plants, which usually have large errors, for example, the error may exceed 15%)". The same set of reinforcement or removal logic is used, which makes it difficult for the repaired data to support high-precision assessment. Third, the models lack universality and physical rationality. Short-term real-time monitoring, medium-term scheduling optimization, and long-term planning and prediction rely on multiple independent tools. The data interfaces and parameter systems between models are inconsistent, and the migration costs are high. Moreover, existing data-driven models lack clear physical constraints (such as equilibrium equations and the law of conservation of mass), which can easily lead to prediction results that violate common sense about the process (such as predicting CO2 emissions lower than the theoretical lower limit calculated from the carbon content of the fuel). This results in low credibility of the assessment results and makes it difficult to use them directly for engineering decision-making.

[0003] In this invention, "Exergy" refers to the portion of energy in a system or device that can theoretically be converted into maximum useful work, and is a key thermodynamic indicator for measuring energy quality and utilization efficiency.

[0004] In this invention, a complex system that simultaneously encompasses four key dimensions—water resources, energy consumption, and carbon emissions—is referred to as a “water-energy-carbon four-dimensional coupled system.” Summary of the Invention

[0005] In view of this, this invention provides an AI-driven method and system for evaluating the coupling of energy, water, and carbon chains. It aims to improve the evaluation system, enhance data quality and evaluation accuracy, and ensure the physical rationality of the model output by introducing an evaluation dimension, distinguishing and processing two types of virtual data nodes, and constructing a deep learning hybrid model embedded with physical constraints. This achieves the evaluation and scientific decision-making of a four-dimensional coupled energy-water-energy-carbon system. This invention constructs a general evaluation framework that, through the configurability of data acquisition, node processing, model constraints, and key evaluation dimensions, can flexibly adapt to the evaluation needs of different scenarios such as energy systems, industrial processes, and urban planning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an artificial intelligence-driven method for evaluating the coupling of energy, water, and carbon chains, comprising the following steps: S1: Multi-source data acquisition and boundary delineation: Delineate the assessment boundary based on the physical equipment scope, system topology, or administrative region of the assessment object; collect thermal data, water data, energy data, carbon emission data, and scenario-related data within the boundary. The thermal data is obtained by measuring the temperature, pressure, flow rate, and composition parameters of the system or equipment and then calculating them according to thermodynamic formulas; determine the assessment scenario type, which is divided into energy system, industrial process, or urban planning. S2: Attribute node processing involves filtering and evaluating real attribute nodes that are suitable for the scenario, and reinforcing missing information in real attribute nodes according to a time scale, where the time scale is divided into short-term, medium-term, or long-term; simultaneously, it identifies and constructs two types of virtual nodes, including a first type of virtual node and a second type of virtual node, wherein: Based on the benchmark library of real attribute nodes of the same type of evaluation object, the frequency of occurrence of each attribute is counted, and the attributes that have an occurrence frequency of not less than a preset first threshold and are missing in the current evaluation object are defined as the first type of virtual node. Based on historical data statistical principles or industry standards, numerical thresholds are set. Real attribute nodes whose values ​​exceed the thresholds and have no reasonable explanation, or whose data collection scenarios are abnormal and whose data credibility coefficients are lower than the preset second thresholds, are defined as the second type of virtual nodes. Calculate the degree of influence of the first type of virtual nodes on the evaluation results, and classify the data quality of the second type of virtual nodes to obtain the corresponding quality parameters; S3: Calculation and correction of association weights. Based on the influence degree and quality parameters of the real attribute nodes and the virtual nodes obtained in step S2, the association weights of the real attribute nodes and the four-dimensional core elements of water, energy and carbon are calculated by integrating physical mechanism modeling, information entropy analysis and deep learning attention mechanism optimization, and the influence correction coefficient of the virtual nodes is determined. S4: Four-dimensional coupling mechanism modeling, taking the correlation weights and influence correction coefficients as inputs, constructing a deep learning hybrid coupling model with a physical constraint layer, training and optimizing the model parameters; S5: Multi-timescale coupling assessment and dynamic prediction. Based on the optimized coupling model and combined with the information characteristics of different time scales, it outputs the comprehensive performance assessment results and future trend prediction of the four-dimensional coupling system of carbon-water-energy-carbon. S6: Optimize decision-making and visualization output, outputting an evaluation report that includes virtual node prompts, credibility distribution and data enhancement suggestions, as well as multi-objective optimization schemes for "improvement-water saving-energy saving-carbon reduction" adapted to the scenario.

[0007] In a specific feasible implementation, step S2, which involves reinforcing the missing information of real attribute nodes according to a time scale, specifically includes: For short-term missing information, linear interpolation or data supplementation from adjacent sensors is used to reinforce it. For missing information in the medium term, an LSTM time series prediction model is used to supplement it, with historical data from the same period and scenario trend data as input; For information that has been missing for a long time, a system dynamics model combined with multi-scenario analysis is used to supplement it.

[0008] In a specific feasible implementation, in step S2: The calculation of the impact of the first type of virtual nodes on the evaluation results specifically includes: calculating the impact of a single node and calculating the comprehensive impact of multiple nodes based on the impact of each single node. The classification of data quality for the second type of virtual nodes specifically includes: classifying anomaly level and credibility level.

[0009] In a specific feasible implementation, step S3 involves calculating the correlation weights and determining the correction coefficients through a fusion optimization method, specifically including: S31: Based on the physical mechanism model of the evaluation scenario, calculate the basic association weight between each real attribute node and the four-dimensional core element; S32: Use mutual information entropy to analyze the information overlap between the real attribute nodes, and reduce the weight of nodes with high overlap. S33: Combining the attention mechanism of deep learning models, the down-adjusted weights are dynamically optimized to obtain the AI-corrected weights; S34: Based on the influence degree and quality parameters of the two types of virtual nodes obtained in step S2, determine the influence correction coefficient, correct the AI-corrected weights, and obtain the final association weights.

[0010] In one specific implementation, step S4, the construction of the deep learning hybrid coupling model including the physical constraint layer, specifically includes: The model structure includes an input layer, a CNN layer for extracting spatial correlation features, an LSTM layer for extracting temporal correlation features, a physical constraint layer, and an output layer; The physical constraint layer embeds one or more physical law constraints, which include the equilibrium equation, the water-energy conservation law, or the carbon emission calculation logic. During model training, the loss function includes a physical constraint penalty term to penalize outputs that violate physical laws.

[0011] In a specific feasible implementation, step S5, the multi-timescale coupled evaluation and dynamic prediction, specifically includes: Short-term assessment: Based primarily on measured data, supplemented by time-series forecasting models to complete short-term fluctuation data, and assess the real-time coupling coordination degree; Mid-term prediction: By integrating measured data and LSTM prediction data, the trend of coupling state changes and the critical triggering conditions of the coordinated state are output; Long-term forecasting: Combining climate models and policy scenario analysis, a multi-scenario forecasting model is constructed to output the long-term evolution trend and potential risks of the system.

[0012] In a specific feasible implementation, step S6, the optimization decision and visualization output, specifically includes: Output core evaluation indicators and their levels, including efficiency, water resource utilization rate, energy consumption intensity, carbon emission intensity, and four-dimensional coupling coordination degree; Output a data confidence distribution visualization chart, which uses different colors to distinguish high confidence data, low confidence data, and first-type virtual node blank data; Based on reinforcement learning algorithms, multiple optimization schemes are generated with the goals of improving efficiency, saving water, saving energy, and reducing carbon emissions. The implementation costs and expected benefits of each scheme are then analyzed.

[0013] Secondly, the present invention provides an artificial intelligence-driven energy-water carbon chain coupling evaluation system for implementing the aforementioned artificial intelligence-driven energy-water carbon chain coupling evaluation method, comprising: The data acquisition module is used to collect multi-source data and define the evaluation boundaries; The attribute node processing module is used to filter real attribute nodes, supplement missing information, and identify and construct two types of virtual nodes. The association calculation module is used to calculate the association weight between real attribute nodes and four-dimensional core elements, as well as the influence correction coefficient of virtual nodes. The Coupled Modeling module is used to build, train, and optimize deep learning hybrid coupled models that include a physical constraint layer; The evaluation and prediction module is used to perform coupled evaluation and dynamic prediction across multiple time scales. The decision output module is used to generate visual evaluation reports and multi-objective optimization schemes.

[0014] In one specific implementation scheme, the attribute node processing module includes: The information enhancement submodule is configured to use different algorithms to enhance missing information in the short, medium and long term. The virtual node processing submodule is configured to construct the first type of virtual nodes and calculate their impact, as well as identify the second type of virtual nodes and classify them into anomalies and credibility levels.

[0015] In one specific implementation, the coupling modeling module includes: The model building unit is configured to provide a visual interface for users to build model structures that include CNN layers, LSTM layers, and physical constraint layers. The online incremental iteration unit is configured to automatically start incremental training to update the model parameters when the amount of newly added valid data reaches a threshold. Iterative archive units are used to store relevant information for each iteration, supporting querying and analysis.

[0016] Compared with existing technologies, the AI-driven energy-water-carbon chain coupling evaluation method and system described in this invention have the following advantages: 1. A more complete evaluation system has been constructed: By introducing the key thermodynamic dimension of "flow rate", a four-dimensional coupled evaluation framework of "flow rate-water-energy-carbon" has been formed, which overcomes the limitation of existing technologies that only focus on flow statistics, and enables the system to more deeply identify and evaluate the hidden resource waste phenomenon of "high flow rate and low efficiency". 2. Improved the precision and reliability of data processing: By distinguishing between two fundamentally different types of data defects, namely "missing high-frequency required attributes" and "data anomalies / low reliability", and constructing virtual nodes for differentiated processing, combined with targeted reinforcement strategies at multiple time scales, the quality and completeness of input data have been improved, which is more conducive to the data requirements of subsequent high-precision evaluation. 3. Improved the physical rationality and reliability of the model: By constructing a deep learning hybrid model with an embedded physical constraint layer, the data-driven capability is combined with physical laws such as energy balance and conservation of matter to fundamentally avoid the risk that the model output violates basic scientific principles (such as predicting carbon emissions lower than the theoretical lower limit calculated by the carbon content of fuel), thereby improving the reliability of the evaluation results and enabling them to directly support engineering decisions. 4. Enhanced model universality and scenario compatibility: Through the "scenario-based adaptation mechanism", the same methodology framework only needs to be configured in four dimensions: data collection, node processing, model constraints, and evaluation focus. It can adapt to the evaluation needs of various scenarios such as energy systems, industrial processes, and urban planning. This solves the problem of poor model compatibility and high migration costs caused by existing solutions relying on multiple independent tools, and greatly improves the universality and deployment efficiency of the evaluation method. 5. Provides comprehensive decision support capabilities: By supporting multi-timescale assessment and forecasting in the short, medium and long term, and integrating multi-objective optimization scheme generation and multi-agent collaborative decision-making functions, the assessment results not only reflect the system status, but also provide managers and decision-makers at different levels with timely, forward-looking and operable scientific basis. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is an overall flowchart of the AI-driven energy-water-carbon chain coupling evaluation method described in this invention.

[0019] Figure 2 This is a block diagram of the architecture of the AI-driven energy-water-carbon chain coupling evaluation system described in this invention.

[0020] Figure 3 The flowcharts are for processing virtual nodes; (a) is the flowchart for constructing and calculating the impact of the first type of virtual nodes; and (b) is the flowchart for identifying, classifying, and repairing the second type of virtual nodes.

[0021] Figure 4 The diagrams are multi-timescale evaluation logic diagrams; (a) is the short-term evaluation logic diagram; (b) is the medium-term prediction logic diagram; and (c) is the long-term prediction logic diagram.

[0022] Figure 5 A visualization of the data credibility distribution.

[0023] Figure 6 A flowchart for AI model training and iteration. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0025] This invention uses the four dimensions of "water, energy, and carbon" as the core of its evaluation, sequentially going through key stages: refined data processing, intelligent correlation calculation, physical constraint modeling, multi-scale evaluation and prediction, and scenario-based decision output. Specifically, it first refines the attribute nodes, including scenario-specific data enhancement and differentiated processing of two types of virtual nodes. Then, it integrates physical mechanisms and AI algorithms to perform dynamic correlation weight calculation. Next, it constructs a deep learning hybrid coupling model embedded with physical constraints. Based on the above steps, it uses the deep learning hybrid coupling model embedded with physical constraints to perform multi-time-scale evaluation and prediction, and finally outputs a differentiated decision report.

[0026] like Figure 1 As shown, the AI-driven energy-water-carbon chain coupling assessment method of this invention can be widely applied to energy systems (such as regional microgrids and gas turbine power generation systems), industrial processes (such as 10t / h gas boilers and chemical reaction units), and urban planning (such as integrated energy systems in industrial parks and urban water supply and drainage systems). The method comprises the following six interconnected steps: Step S1, Data Acquisition and Boundary Delineation: Define the physical boundaries of the assessment object, and collect energy data (such as energy efficiency and energy loss), water data (such as water intake and discharge, and water quality), energy data (such as energy intensity and conversion efficiency), carbon emission data (such as direct and indirect emissions), and scenario-related data (such as equipment parameters, geographic information, and policy standards); the collected data form the basis for subsequent assessments. For example, scenario-related data refers to auxiliary data closely related to the operating environment, external conditions, management strategies, and inherent characteristics of the assessment object, in addition to data that directly reflects the core states of the system's energy, water, and carbon (such as consumption, emissions, and efficiency). This data is used to explain the causes of system state changes, constrain the boundary conditions of the assessment model, and support cross-scenario comparative analysis. Data collection follows the principle of "scenario-driven, on-demand configuration." Scenario-related data typically includes, but is not limited to, the following categories: Object characteristic data: Evaluate the object's inherent physical, design, or planning parameters, such as equipment rated power, capacity, design efficiency, pipeline topology, building / park area, industrial structure, etc. Environmental status data: Parameters of the natural and human environment in which the assessment object operates, such as ambient temperature, humidity, air pressure, wind speed, irradiance (meteorological data), rainfall, electricity price, fuel price, policy standards and regulatory limits, etc. Operational control data: parameters reflecting human intervention or scheduling strategies, such as production plans, load curves, equipment start-up and shutdown commands, maintenance and repair plans, and operating modes (such as base load / peak shaving). Related system data: Key data of upstream, downstream or parallel systems that exchange matter or energy with the object of assessment, such as power grid dispatch instructions, upstream water supply quality, and changes in downstream user demand; The types and quantities of scenario-related data to be collected in a specific assessment task are jointly determined by the type of assessment scenario (energy system / industrial process / urban planning) and the specific assessment objectives. For example, the system has built-in "scenario-related data configuration templates" for different scenarios, which users can add or delete based on. The core principle of selection is to choose data items that have significant physical or statistical correlations in explaining or predicting changes in the four core indicators of "energy-water-energy-carbon". The data can be selected as one or more of the system or device's input efficiency, output efficiency, efficiency loss rate, etc., and can be calculated by directly measuring relevant parameters (such as temperature and pressure) through sensors. Step S2, Attribute Node Processing: The attribute nodes refer to quantifiable objects that are directly related to the four-dimensional core element assessment of water, energy, and carbon, or that reflect the characteristics of the assessment scenario. Examples include equipment operating parameters (such as boiler pressure), resource consumption indicators (such as water intake), environmental impact parameters (such as carbon emission intensity), and scenario-related data (such as raw material prices).

[0027] The system filters real-attribute nodes that are highly relevant to the evaluation scenario from the collected data. For missing information in real-attribute nodes, different strategies are used to supplement the node information according to the time scale of the predicted demand (short-term, medium-term, and long-term). For example, interpolation is used to supplement the node information of real-attribute nodes with a short-term predicted demand time scale, time series prediction model is used to supplement the node information of real-attribute nodes with a medium-term predicted demand time scale, and scenario analysis is used to supplement the node information of real-attribute nodes with a long-term predicted demand time scale. At the same time, two types of virtual nodes are identified and constructed: The first type of virtual node targets key attributes that appear frequently in the same type of object but are currently missing. Its impact is identified and evaluated by comparison with a benchmark library. The second type of virtual node identifies attributes with abnormal values ​​or low data collection confidence by setting thresholds and classifying their abnormality and confidence levels. The construction and influence calculation of the first type of virtual nodes are as follows: Figure 3 As shown in (a), the identification, classification, and repair of the second type of virtual nodes are as follows: Figure 3 As shown in (b); for details on the specific construction, identification and calculation methods of the two types of virtual nodes, please refer to the content in "(II) Attribute Node Processing" below; Attribute node processing ensures the integrity and reliability of input data; Step S3, Calculation of correlation tightness: To dynamically quantify the association strength between attribute nodes and core elements, the calculation process generates three types of association weights: basic association weight (theoretical initial values ​​calculated based on physical mechanisms), AI-corrected weight (values ​​dynamically corrected through redundancy removal and attention mechanisms), and final association weight (the final applied value after incorporating the influence of virtual nodes). The specific calculations are as follows: First, based on the physical mechanism model of the evaluation scenario (such as energy balance and matter conservation), calculate the basic correlation weight between each real attribute node and the four-dimensional core elements of "energy-water-energy-carbon". Secondly, AI methods such as mutual information entropy are used to analyze the information overlap between nodes and redundancy correction is performed on the basic association weights; and the attention mechanism of deep learning models is combined to dynamically optimize the weights and obtain the AI-corrected weights. Finally, the influence coefficients of the two types of virtual nodes mentioned above are introduced to further modify the weights after AI correction, so as to obtain the final association weights used for model input, making them more in line with the actual data situation. For example, the calculation of the influence correction coefficient and the correction of the association weights are performed according to the following logic: Application of the influence of the first type of virtual nodes: For the first type of virtual nodes constructed in step S2, their influence (including the influence of a single node and the overall influence) has been calculated. The results are then used directly for weight correction.

[0028] Quantification of the degree of defect in the second type of virtual node: The data quality defect level coefficient is calculated based on its anomaly level and credibility level.

[0029] The abnormality levels are divided into mild (deviation ±1σ~±2σ), moderate (deviation ±2σ~±3σ), and severe (deviation >±3σ), with corresponding defect severity coefficients of 0.3, 0.6, and 0.9, respectively.

[0030] The credibility levels are divided into high (0.9-1.0), medium (0.6-0.8), low (0.3-0.5), and very low (0.1-0.2).

[0031] Final adjustment of association weights: When a first-type virtual node exists, the final association weight = AI-corrected weight × (1 - overall influence).

[0032] When a second type of virtual node exists, the final association weight = AI-corrected weight × (1 - defect degree coefficient × (1 - credibility coefficient)).

[0033] The above formula quantifies the impact of the two types of virtual nodes into specific correction coefficients, which are then applied to the calculation of association weights, thereby achieving differentiated and quantitative processing of data defects.

[0034] Step S4, Four-dimensional Coupled Modeling: Using the calculated weights and virtual node influence as input, a deep learning hybrid coupling model with a physical constraint layer is constructed. This model typically includes a convolutional neural network (CNN) layer for extracting spatial features, a long short-term memory (LSTM) layer for extracting temporal features, and a crucial physical constraint layer. This layer embeds physical laws such as the equilibrium equation and the water-energy conservation law to ensure that the model output does not violate basic scientific principles. Through training and optimization (see process...), a hybrid model is constructed. Figure 6 This yields a coupling model that can be used for evaluation; Figure 6 The online incremental learning and continuous optimization closed loop of the described deep learning hybrid coupling model after deployment is demonstrated. The process is presented in a loop, and the specific steps are as follows: Incremental data monitoring: The system continuously monitors and collects new and valid operational data; Dataset partitioning: When the amount of new data reaches a preset threshold, it is partitioned into datasets suitable for incremental training; Incremental training: Incremental learning (fine-tuning) of an already deployed model based on a new dataset, rather than training from scratch; Model training: Executing specific training algorithms to update some or all of the model's parameters; Accuracy verification: Use validation data to evaluate the performance and verify the accuracy of the updated model; Parameter optimization: Based on the validation results, further adjust and optimize the model parameters or training strategy.

[0035] After completing the above steps, the process returns to the "incremental data monitoring" stage, forming a continuous monitoring-learning-optimization-validation closed loop, which enables the model to adapt to system changes and maintain the accuracy of evaluation and prediction.

[0036] Step S5, Multi-timescale Coupled Evaluation and Dynamic Prediction: Based on the trained coupling model, differentiated assessments are conducted, including: short-term assessment focusing on real-time status, primarily using high-frequency measured data, and outputting high temporal resolution of system coupling coordination; medium-term prediction focusing on operational trends, integrating measured and predicted data, and outputting the system state change trajectory and early warning conditions; and long-term prediction focusing on strategic risks, combining macro-level scenarios such as climate and policy, and outputting the system's long-term evolution trend and potential risks. The multi-scale assessment logic is as follows: Figure 4 As shown; Step S6, optimize decision-making and visualization output: The system output includes core evaluation indicators (such as efficiency, carbon emission intensity, and coupling coordination) and data credibility distribution visualization charts (such as...). Figure 5 (As shown) and a comprehensive report on the multi-objective optimization scheme of "improvement, water conservation, energy saving, and carbon reduction" generated based on reinforcement learning algorithms. The report can support the differentiated decision-making needs of users at different levels (such as management, enterprises, and industrial parks) and ensures the traceability of the evaluation process through log recording.

[0037] The data credibility distribution is visualized in a chart format, clearly distinguishing between high-credibility data, low-credibility data, and blank data with dummy nodes. This chart and its corresponding detailed data can be exported for subsequent analysis and auditing.

[0038] The steps S1 to S6 described above constitute a complete technical closed loop from data input, intelligent processing, model building, comprehensive evaluation to optimization decision-making. This closed loop systematically solves core problems such as incomplete evaluation dimensions, poor data quality, and insufficient physical rationality of the model.

[0039] The technical details will be described one by one below.

[0040] (I) Data Acquisition and Boundary Delineation: Contextualized Data Scope and Preprocessing Based on the selected assessment scenario type (such as energy systems, industrial processes, urban planning), determine the scope of data to be collected and perform unified preprocessing to ensure data availability.

[0041] 1. Evaluation of boundary delineation methods The evaluation boundary is defined using one or a combination of the following methods: (1) Physical equipment scope method: based on the physical equipment boundary of the object being evaluated, such as boiler body, pipelines, auxiliary equipment, etc.; (2) System topology method: taking the energy flow and material flow topology of the system as the boundary, including all directly related nodes; (3) Administrative Region Law: Based on administrative divisions, it is applicable to urban planning scenarios; (4) Functional unit method: The boundary is the unit that performs a specific function, such as production workshop, process section, etc.

[0042] Once the boundaries are defined, it is necessary to identify all data collection points and evaluation elements included within those boundaries.

[0043] 2. Data Acquisition Methods The acquisition of the data includes direct calculation and indirect deduction methods: (1) Direct calculation method: Calculate the following parameters by measuring the working fluid's temperature T, pressure P, flow rate m, and composition xi: Physics equation: E_ph = m × [(h - h0) - T0 × (s - s0)] Chemistry equation: E_ch = m × Σ(xi × ei_ch) Where h and s are specific enthalpy and specific entropy, respectively, the subscript 0 indicates the environmental state, and ei_ch is the component quantification value of each component.

[0044] (2) Indirect deduction method: For systems that are difficult to measure directly, the deduction is made by inversely using energy balance and efficiency: E_loss = E_in × (1 - η_ex) Where η_ex is the system efficiency, which can be obtained from the database based on the device type.

[0045] Example: Taking a gas-fired boiler as an example, measurements need to be taken within the boundary: Fuel inlet: natural gas flow rate, temperature, pressure, composition; Steam outlet: steam temperature, pressure, and flow rate; Exhaust port: flue gas temperature, pressure, flow rate, and composition; Environmental parameters: ambient temperature and pressure.

[0046] The input ε, output ε, and ε loss are calculated using the parameters mentioned above.

[0047] The scope of data collection must match the core requirements of the evaluation scenario: In energy system scenarios (such as regional power grids), it is necessary to focus on collecting data that reflects the energy conversion and transmission process, such as the efficiency of power generation equipment, power grid transmission losses, and the output characteristics of renewable energy.

[0048] Industrial process scenarios (such as chemical and metallurgical industries): It is necessary to focus on collecting data that reflects the process equipment and reaction process, such as key equipment design parameters, conversion rates of raw materials and products, and core process status parameters.

[0049] In urban planning scenarios (such as industrial parks), it is necessary to collect data that reflects the macro-characteristics of the regional system, such as population and building density, traffic energy consumption levels, and infrastructure (water, heat, electricity) network operation data.

[0050] In addition, all scenarios must collect corresponding direct and indirect carbon emission data based on their system boundaries.

[0051] Taking a 10t / h gas-fired boiler in an industrial process scenario as an example. First, define the physical system boundary for its evaluation, which may include, for example, "the boiler body, waste heat recovery device, and water treatment system." Within this boundary, typical data to be collected includes: Energy data: such as flue gas energy loss (e.g., 7500 kJ / h) and steam energy output (e.g., 12000 kJ / h).

[0052] Water data: such as water replenishment (e.g., 10 m³ / h), sewage discharge (e.g., 0.5 m³ / h), and water quality indicators (e.g., hardness 200 mg / L).

[0053] Energy data: such as natural gas consumption (e.g., 82 m³ / h) and auxiliary power consumption (e.g., wind turbine power consumption of 25 kWh).

[0054] Carbon emission data: such as direct emissions from combustion Emissions (e.g., 1.23 t / h).

[0055] Scenario-related data: such as boiler design parameters (rated pressure 1.2MPa), fuel characteristics (natural gas calorific value 36MJ / m³), and production plans (such as quarterly load rate 80%-100%).

[0056] The collected raw data needs to be preprocessed to ensure its quality and consistency. Preprocessing typically includes: Outlier removal: Statistical methods (such as the 3σ principle) are used to identify and remove outlier data points that significantly deviate from the normal range.

[0057] Format standardization: Heterogeneous data from different interfaces and protocols (such as sensor binary streams and JSON data returned by platform APIs) are uniformly converted into standardized data formats (such as CSV) to provide standardized input for subsequent steps.

[0058] For example, taking the calculation of "flue gas heat loss" of the 10t / h gas-fired boiler as an example, the process of obtaining heat loss data is explained: First, the flue gas temperature T_pg, pressure P_pg, flow rate V_pg, and flue gas composition (such as...) are measured by sensors. The volume fraction of the exhaust gas; and simultaneously obtain the ambient temperature T0. Then, according to the engineering thermodynamic formula, the exhaust gas loss E_loss can be approximately calculated as: E_loss=V_pg×[c_p×(T_pg-T0)-T0×(c_p×ln(T_pg / T0)-R×ln(P_pg / P0))]; Where c_p is the average isobaric specific heat capacity of the flue gas, R is the gas constant of the flue gas, and P0 is the ambient pressure. The required parameters c_p and R can be obtained by weighted averaging of the flue gas composition. Other parameters (such as steam output) can be obtained by measuring parameters such as steam temperature, pressure, and flow rate using a similar principle.

[0059] (II) Attribute Node Processing: This mainly includes supplementing missing information for real attribute nodes, and identifying and constructing two types of virtual nodes; 1. Strengthening of real attribute node information Based on the time scale characteristics required for prediction and evaluation, a differentiated strategy is adopted to reinforce the missing information of real attribute nodes: Short-term scale: Since the data usually has high continuity and stability, interpolation and other methods based on the data's own change patterns can be used to enhance it.

[0060] Medium-term scale: It is necessary to consider trend factors such as periodic changes and operating load adjustments, and time series prediction models (such as Long Short-Term Memory networks LSTM) are often used for estimation.

[0061] Long-term scale: Macroeconomic scenario variables such as climate change and policy adjustments need to be included, and multi-scenario simulation and analysis can be carried out using models such as system dynamics.

[0062] For example, a Long Short-Term Memory (LSTM) time series prediction model is used to reinforce missing information in the intermediate period. A specific implementation is as follows: An LSTM model is constructed using historical data from the same period of the previous year (e.g., data from the same quarter of the previous year) for the attribute node to be reinforced (e.g., natural gas consumption of a gas boiler) and related trend data reflecting the system's operating status (e.g., production load growth rate) as input features. This model can be set to have 64 hidden layer nodes, using the Adam optimizer with an initial learning rate of 0.001, and trained using mean squared error (MSE) as the loss function, iterating approximately 200 times until convergence. After training, continuous historical data before the missing point is input into the model, allowing iterative prediction of the data sequence for the missing period with a prediction accuracy exceeding 85%.

[0063] 2. Differentiated processing of two types of virtual nodes To characterize and address data defects, this invention defines two types of virtual nodes: (1) Construction of the first type of virtual nodes: Establish a benchmark library of real attribute nodes of the same type of evaluation object, and count the occurrence frequency of each attribute; define the attributes that have an occurrence frequency of not less than the preset first threshold (e.g., 80%) and are missing in the current evaluation object as the first type of virtual nodes.

[0064] Subsequently, it is necessary to calculate the degree of influence of this virtual node on the evaluation results (single-node influence and multi-node combined influence) to quantify the evaluation uncertainty introduced by the lack of this key attribute. The specific calculation methods include: First, the impact of a single first-class virtual node is evaluated, which depends on the theoretical importance of the attribute represented by the node in the evaluation model (basic association weight) and the prevalence of the absence of this attribute in similar scenarios (absence frequency). If there are multiple first-type virtual nodes, their combined influence needs to be calculated further. Usually, a preset aggregation function is used to combine the influence of each individual point. This function should take into account the nonlinearity and superposition attenuation effect that may be caused by the absence of multiple factors.

[0065] For example, the specific calculation method can be as follows: Single node influence = Basic association weight of the missing attribute × Similar missing rate correction coefficient. The similar missing rate correction coefficient can be set according to the missing rate of the attribute among similar objects; for example: 1.2 when the missing rate is ≤30%, 1.0 when the missing rate is between 30% and 60%, and 0.8 when the missing rate is >60%.

[0066] If multiple type I virtual nodes exist, their combined multi-node influence can be calculated using the following formula: Combined Multi-Node Influence = 1 - ∏(1 - Single Node Influence × 0.5). The formula introduces a coefficient of 0.5 to reduce the cumulative effect of multiple virtual nodes and avoid overly pessimistic estimates of the overall assessment results. This coefficient is an empirical value and can be adjusted according to actual circumstances.

[0067] The single-node influence degree and the multi-node comprehensive influence degree will be used to calculate the influence correction coefficient in the subsequent step S3.

[0068] (2) Identification of the second type of virtual nodes: Based on historical data statistical principles (such as the 3σ principle, box plot method) or industry standards (such as "Energy Management System Requirements" GB / T23331, etc.), a numerical threshold is set; real attribute nodes whose values ​​exceed the threshold without reasonable explanation, or whose data collection scenario is abnormal and their data credibility coefficient is lower than the preset second threshold (which can be set to an empirical value between 0.5 and 0.8, with a default value of 0.6), are defined as the second type of virtual nodes. The data quality of the second type of virtual nodes needs to be graded according to the degree of data deviation and the credibility of the collection scenario, including the division of anomaly level and credibility level. For example, the anomaly level can be divided into mild (deviation ±1σ~±2σ), moderate (deviation ±2σ~±3σ), and severe (deviation > ±3σ). The credibility level can be divided into high (0.9-1.0), medium (0.6-0.8), low (0.3-0.5), and very low (0.1-0.2). The anomaly level and confidence level will be used in the subsequent step S3 to calculate the impact correction coefficient.

[0069] 3. Threshold setting method The "preset first threshold" for determining the first type of virtual node and the "preset second threshold" (i.e., the reliability threshold) for determining the second type of virtual node are not fixed values, but can be determined by one or more of the following methods based on data quality requirements, scenario importance, and available historical data: Method 1: Empirical threshold method (can be used as the default method) The initial threshold is set based on the experience of domain experts and industry consensus, which is suitable for scenarios where historical data is insufficient or rapid deployment is required.

[0070] Preset first threshold (frequency threshold): usually set between 70% and 90%. For example, in industrial process scenarios, 80% can be used for key operating parameters of equipment (that is, if the frequency of this attribute in the benchmark library of similar objects is ≥80%, it is considered a key missing value if it is missing); in urban planning macro indicators scenarios, 90% can be used because the data items are relatively standardized.

[0071] A preset second threshold (credibility threshold) is typically set between 0.5 and 0.8. For example, for automated sensor data streams, it can be set to 0.7; a value below this will trigger an anomaly review. For manually entered or indirectly calculated data, due to the inherently higher uncertainty, it can be set to 0.6.

[0072] Method 2: Statistical Learning Method When there is a sufficient historical dataset of similar evaluation objects, the threshold can be learned automatically through a data-driven approach.

[0073] For the first threshold: the frequency distribution of all attributes in the benchmark library can be analyzed, and the upper quartile (75th percentile) of the frequency distribution or the lower boundary of the "high-frequency attribute cluster" identified by cluster analysis can be used as the threshold.

[0074] For the second threshold: the lower quartile (25th percentile) or the upper boundary of the "low confidence sample" identified by anomaly detection algorithms (such as isolated forest) can be used as the threshold based on the confidence coefficient distribution of each attribute point in historical data.

[0075] Method 3: Dynamic Adjustment During system operation, the threshold can be dynamically optimized based on the quality feedback and application feedback from the evaluation results.

[0076] Sensitivity analysis adjustment: By analyzing the impact of small changes in the threshold on the stability of the final evaluation results (such as coupling coordination), a threshold range that makes the results insensitive to data defects is selected.

[0077] Effect feedback adjustment: If the system frequently generates too many virtual nodes due to overly strict thresholds (leading to the inability to conduct the evaluation), or if the thresholds are too lenient and let a large number of data defects pass through (leading to distorted evaluation results), the thresholds can be adjusted automatically or manually to achieve a balance between "evaluation feasibility" and "result reliability".

[0078] In practical systems, the "empirical threshold method" can be used for initialization first, and then gradually switched to the "statistical learning method" as data accumulates, while always allowing fine-tuning through the "dynamic adjustment method". The set values ​​and methods used can be recorded in a work log for traceability.

[0079] 4. Method for calculating data credibility coefficient For the identified second-type virtual nodes, their corresponding data credibility coefficients need to be calculated. These data credibility coefficients are calculated through a comprehensive evaluation across the following multiple dimensions: Reliability coefficient = α1 × sensor status score + α2 × data continuity score + α3 × logic consistency score + α4 × acquisition environment score; in: Sensor status score: Based on sensor calibration status, runtime, failure rate, etc., the value ranges from 0 to 1; Data continuity score: assessed based on data missing rate and jump frequency, with a value of 0-1; Logical consistency score: Evaluated based on the consistency with the physical logical relationship of the relevant parameters, with a value of 0-1; Environmental score: Evaluated based on environmental factors such as temperature, humidity, and electromagnetic interference, with a value of 0-1; α1, α2, α3, and α4 are weighting coefficients that satisfy α1+α2+α3+α4=1, and are all set to 0.25 by default.

[0080] The criteria for judging abnormal data acquisition scenarios include: sensor communication interruption duration exceeding the preset time (e.g., 5 minutes), data acquisition environment parameters exceeding the normal working range, and abnormal fluctuations in data acquisition frequency exceeding 20%.

[0081] Taking the energy management system of an industrial park as an example, this paper illustrates the setting and calculation of relevant thresholds for virtual nodes: 1. Setting the first threshold Monitoring data from 50 similar gas-fired boilers within the industrial park were analyzed to create a baseline database of real-attribute nodes. Frequency of occurrence of "smoke exhaust temperature": 98% "Inlet water temperature" frequency of occurrence: 92% Frequency of occurrence of "oxygen content in flue gas": 85% Frequency of occurrence of "furnace pressure": 78% If the first threshold is set to 80%, then if the current evaluation object is missing any of the attributes of "exhaust gas temperature", "inlet water temperature" and "flue gas oxygen content", it will be defined as a first-class virtual node.

[0082] 2. Setting the second threshold and calculating credibility Regarding the "natural gas consumption" data for a certain boiler: Sensor status: Operating normally, within calibration validity period, status score = 0.9 Data continuity: Missing rate 2% in the past 24 hours, 3 jumps, continuity score = 0.85 Logical consistency: Logically consistent with parameters such as steam output and operating load; consistency score = 0.95 Data collection environment: Ambient temperature 25℃ (normal range), no strong electromagnetic interference, environmental score = 0.9 The confidence coefficient is calculated as follows: 0.25 × 0.9 + 0.25 × 0.85 + 0.25 × 0.95 + 0.25 × 0.9 = 0.9 The second threshold is set to 0.6. Since 0.9 > 0.6, this data is not defined as a second type of virtual node.

[0083] If the data exceeds the historical statistical threshold (e.g., ±3σ) and there is no reasonable explanation such as operating condition adjustment, it is defined as a second type of virtual node.

[0084] For example, taking the evaluation of a 10t / h gas-fired boiler as an example, the process of handling its virtual nodes is explained as follows: Identification and Influence Calculation of Type I Virtual Nodes: By analyzing historical data of similar boilers, it was found that "inlet water temperature" is a key parameter that is commonly monitored. If the current assessment object lacks this data, it is constructed as a Type I virtual node, and its single-node influence and overall influence in the whole are calculated according to the method in (1) above.

[0085] Second type of virtual node identification, classification, and repair: Assuming that at a certain moment, the "..." If the "emissions" data deviates significantly from the normal range based on historical statistics (e.g., exceeding 3 times the standard deviation) and there is no reasonable explanation for the operating conditions, then the data point is constructed as a second type of virtual node and judged as severely abnormal with a confidence level of "low" according to the standard in (2) above. During the repair, the normal data trend of the adjacent time is first referenced for smoothing, and then a constraint correction is made according to the physical conservation relationship of "CO2 emissions = fuel consumption × carbon content" to obtain a possible value that conforms to physical laws (e.g., 1.23 t / h). The correction result can be verified by comparing with the measured data under subsequent normal operating conditions. For the construction and processing procedures of the two types of virtual nodes, please refer to [link / reference]. Figure 3 As shown.

[0086] (III) Calculation of the degree of correlation: The calculation in this step is a dynamic process that integrates physical mechanisms and data-driven approaches, and mainly includes three stages: basic weight calculation based on physical mechanisms, weight correction based on AI algorithms, and final optimization that introduces the influence of virtual nodes.

[0087] 1. Basic weight calculation The calculation of basic weights must be closely based on the inherent physical laws of the evaluation scenario. For example, in the energy system scenario, it can be based on the flow network and energy balance relationship; in the industrial process scenario, it can be based on the process reaction equation and the law of conservation of mass; and in the urban planning scenario, it can be based on the infrastructure supply and demand model. Through the corresponding physical model, the basic weights reflecting the theoretical correlation strength between nodes and core elements can be calculated.

[0088] For example, taking the calculation of the basic correlation weight between the natural gas consumption node and the core input element of the system in a gas-fired boiler scenario as an example, the specific process is as follows: Based on the composition and thermodynamic formulas of natural gas, calculate its input gas value per unit volume, e_gas; Obtain the cumulative natural gas consumption V_gas during the assessment period; Calculate the node's contribution: E_node = e_gas × V_gas; The total input E_total within the system boundary (including natural gas and all other energy sources); Then the basic association weight of this node for the 'total input' element is W_base=E_node / E_total.

[0089] 2. AI algorithm correction After obtaining the basic weights, a data-driven approach is introduced for optimization and correction. On the one hand, methods such as mutual information entropy are used to analyze the degree of information overlap between data from different nodes, and the weights of redundant nodes are adjusted reasonably to eliminate the effects of collinearity. On the other hand, the attention mechanism in deep learning models (such as CNN-LSTM) is used to dynamically identify and enhance the weight attention of feature nodes that are more critical to the current evaluation task.

[0090] 3. Correction of the impact of virtual nodes and final weight generation To reflect the impact of data defects on the evaluation, the quantification results of the two types of virtual nodes obtained in step S2 need to be integrated into the weight calculation chain.

[0091] First, let's define the "AI-corrected weights": The weights obtained by dynamic optimization based on mutual information entropy redundancy removal and attention mechanism are called "AI-corrected weights," which are the weight benchmarks before virtual node correction.

[0092] Next, integrate the effects of virtual nodes: For attributes affected by the first type of virtual nodes, the corresponding "AI-corrected weights" are reduced based on their overall impact to reflect the assessment uncertainty caused by the lack of key attributes.

[0093] For attributes affected by the second type of virtual nodes, the corresponding "AI-corrected weights" are adjusted according to their anomaly level (or defect degree coefficient) and confidence level (or confidence coefficient) to reduce the interference of abnormal or low-confidence data on the evaluation.

[0094] Through the above steps, the influence of the two types of virtual nodes is quantified into adjustments to the weights, and the final output is the "final association weight" used for coupling modeling.

[0095] For example, taking a 10t / h gas-fired boiler scenario as an example, the calculation process of the association weight is explained: Basic weight calculation: Calculated based on the thermal balance physical model of the boiler system. For example, "natural gas consumption" is the main source of steam output, and its basic weight can be estimated by (natural gas input × boiler thermal efficiency / steam output), with a calculated result of approximately 0.81, indicating a strong correlation. However, "makeup water" has a weaker direct impact on steam output, and its calculated basic weight may be very small, indicating a weak correlation. AI Algorithm Correction: After obtaining the basic weights, AI algorithms are used for correction. First, mutual information entropy analysis reveals a high degree of overlap between the data on "boiler pressure" and "steam output," indicating redundancy. Therefore, their basic weights are appropriately reduced. Second, the attention mechanism in the model dynamically identifies features more critical to the current state assessment, such as potentially increasing the weight of core input nodes like "natural gas consumption." Virtual Node Impact Correction: Finally, the impact of virtual nodes is considered. If a first-type virtual node exists (e.g., missing inlet water temperature), the AI-corrected weights associated with it are reduced based on its calculated overall impact to obtain the final association weight. If a second-type virtual node exists (e.g., abnormal...), the impact of virtual nodes is considered. For emissions, the weights after AI correction are adjusted based on the degree of data defects and credibility to obtain the final association weights, in order to reflect the uncertainty caused by the unreliability of the data.

[0096] (iv) Four-dimensional coupling modeling: Its core is to establish a deep learning hybrid coupled model that includes a physical constraint layer, capable of simultaneously learning complex patterns in data while strictly adhering to physical common sense, and optimizing it through training and iteration (see process). Figure 6 ).

[0097] 1. Model Structure Design The model receives processed attribute node features, weights, and time scale information as input. Its structure typically includes: Feature extraction layer: Convolutional neural network (CNN) is used to capture the spatial correlation features between attribute nodes, and long short-term memory network (LSTM) is used to capture the temporal evolution features of the system state.

[0098] Physical constraint layer: This is a key component of the model, embedding one or more fundamental physical constraints, such as the equilibrium equation, the law of conservation of water and energy, and the law of conservation of mass. This layer ensures that the model's forward propagation and parameter optimization processes always proceed in a direction consistent with physical laws.

[0099] Output layer: Outputs the four-dimensional coupling coordination degree of "water-energy-carbon" and the specific evaluation values ​​of each dimension.

[0100] 2. Model Training and Optimization The model learns the mapping relationships in the data through training, and its optimization process takes into account both the accuracy of data fitting and the compliance with physical laws.

[0101] Initial training: Using a historical dataset, it is divided into training, validation, and test sets. Training employs an adaptive optimization algorithm, with the loss function consisting of a prediction error term and a physical constraint violation penalty term. This forces the model to reduce prediction error while avoiding outputs that violate the embedded physical rules.

[0102] Online Incremental Iteration: To enable the model to adapt to system changes, an online update mechanism is designed. When the system accumulates a certain amount of new and valid data, incremental training can be automatically triggered. This training typically only fine-tunes some parameters of the model (such as the parameters of the LSTM layer), thereby improving the model's timeliness while maintaining its overall stability. After incremental training, the model's performance needs to be validated using new data and some historical data. Only after passing the validation can the model be updated and deployed.

[0103] During model training, the loss function consists of a data prediction error term (such as mean squared error, MSE) and a physical constraint penalty term to ensure that the model output does not violate fundamental physical laws. The physical constraint layer is implemented using a dual-path constraint mechanism: combining forward propagation constraints with loss function penalty term constraints, as detailed below: 1. Forward propagation hard constraint mechanism During the forward propagation of the model, intermediate results that violate physical laws are forcibly corrected: (1) For the loss balance constraint: if the loss value E_loss of the model output is <0, then it is forced to be set to E_loss=0; (2) For the mass conservation constraint: if the model predicts If the emissions are lower than the theoretical lower limit C_min calculated from the carbon content of the fuel, they are forcibly corrected to C_min×(1+ε), where ε is a small positive number (e.g., 0.01). (3) For energy conservation constraints: the output energy value must not exceed the input energy value. If it does, it will be scaled up to the conservation value.

[0104] 2. Soft constraint penalty term in the loss function Add a physical constraint penalty term to the loss function to penalize predictions that violate physical laws: (1) Non-negative penalty term: L_exergy=λ1×max(0,-E_loss_pred) (2) Carbon emission lower limit penalty term: L_carbon=λ2×max(0,C_min-C_pred)^2 (3) Energy conservation penalty term: L_energy=λ3×(E_in_pred-E_out_pred-E_loss_pred)^2 (4) Water balance penalty term: L_water=λ4×(W_in_pred-W_out_pred-W_loss_pred)^2 Total loss function: L_total = L_MSE + L_exergy + L_carbon + L_energy + L_water Among them, λ1, λ2, λ3, and λ4 are penalty weight coefficients, which are determined through cross-validation and range from 0.1 to 1.0.

[0105] 3. Pseudocode Example of Constraint Implementation class PhysicsConstraintLayer(nn.Module): def forward(self, x): # x represents the original output of the model, including predicted values ​​for nitrogen, water, energy, and carbon. # Apply hard constraints x['exergy_loss'] = torch.clamp(x['exergy_loss'], min=0) # Non-negative loss x['carbon_emission'] = torch.max(x['carbon_emission'], torch.tensor([carbon_min])) # Lower limit for carbon emissions # Energy Conservation Constraint energy_diff = x['energy_in'] x['energy_out']x['exergy_loss'] if energy_diff<0: x['energy_out'] = x['energy_in']x['exergy_loss'] return x 4. Training Process (1) Data preparation: Prepare a labeled training dataset {D_train} and a validation dataset {D_val}; (2) Model initialization: Construct a hybrid model containing CNN layers, LSTM layers, and physical constraint layers; (3) Constraint weight tuning: λ1-λ4 are optimized using the grid search method, with the goal of minimizing the total loss on the validation set; (4) Model training: The Adam optimizer was used with a learning rate of 0.001, a batch size of 32, and 300 training rounds. (5) Constraint verification: After each round of training, check the physical constraint satisfaction rate. If it does not meet the standard, adjust the penalty weight.

[0106] The implementation of the physical constraint layer involves two paths: hard constraint correction during forward propagation and soft constraint penalty in the loss function. These two paths work together to ensure that the model output conforms to both data patterns and physical laws.

[0107] For example, the physical constraint penalty term is designed as follows: First, physical laws are expressed as mathematical inequalities or equations that the model output should satisfy; for example, the law that “exhaust smoke loss is non-negative” can be expressed as the constraint: E_loss>=0.

[0108] Then, this constraint is transformed into a penalty term and added to the loss function.

[0109] A common approach is to use the ReLU function: L_physics = λ × ReLU(-E_loss), where E_loss is the predicted smoke emission loss value. The ReLU function outputs the absolute value of the input when the input is negative, and outputs 0 otherwise, thus penalizing only when the predicted value violates the constraints (E_loss < 0). λ is the penalty weight coefficient, used to balance prediction accuracy and compliance with physical laws. Its value can be determined by cross-validation on the validation set with the physical constraint violation rate as one of the key indicators, for example, set between 0.1 and 1.0.

[0110] For equality constraints (such as those based on fuel carbon content) The emission calculation logic can be designed with a penalty term of L_physics = λ × (predicted emissions - theoretical calculated values)^2. During training, the optimization algorithm drives the model to fit the data while minimizing the total loss function (L_total = L_MSE + L_physics), making its output tend to satisfy all embedded physical constraints.

[0111] For example, in the evaluation model of a 10t / h gas-fired boiler, the physical constraint layer can embed "non-negative flue gas emission loss", " Rules such as "emissions shall not exceed the carbon content of the fuel" apply. The initial training of the model can use historical operating data from a past period. When the boiler is modified or the operating strategy is significantly adjusted, the accumulated new data can trigger an incremental iteration, quickly updating the model's predictive ability for the current operating conditions.

[0112] (v) Multi-timescale assessment: The "energy-water-energy-carbon" four-dimensional coupled evaluation framework constructed in this invention fundamentally changes the limitation of traditional evaluations that only focus on "flow" (such as total energy consumption and total water consumption) by introducing "energy" as a key dimension reflecting energy quality and utilization efficiency. By simultaneously evaluating "energy consumption" (flow index) and "energy efficiency" (quality index), the model of this invention can automatically identify the hidden resource waste pattern of "high flow, low efficiency" in the system. For example, when the model output shows that the "energy consumption" index of a certain system is within the normal range, but the "energy efficiency" index is significantly low, the evaluation system described in this invention can determine that there is a waste phenomenon of "high flow, low efficiency" in this operating condition, and clearly indicate it in the evaluation report, thereby providing a direct decision-making basis for precise energy saving and efficiency improvement.

[0113] Based on the trained coupled model, differentiated time-scale assessments and predictions are implemented for different decision-making needs (logic can be found in [link]). Figure 4 ), and dynamically update the evaluation elements.

[0114] 1. Scaled Assessment and Prediction Strategies The assessment and forecast are designed differently based on the time span of interest: Short-term assessment (see) Figure 4 (a): Focuses on the real-time operating status and immediate performance of the system. It is usually based on high-frequency measured data to evaluate and output high-time-resolution system coupling coordination and various dimensional indicators, which is suitable for operation monitoring and rapid adjustment.

[0115] Medium-term forecast (see) Figure 4(b) focuses on the changing trends of the system over a future operating cycle. It integrates historical patterns with the current state to predict the development trajectory of the system's coupled states and identify critical conditions that may lead to system degradation. It is suitable for preventative maintenance and operation plan optimization.

[0116] Long-term forecasts (see) Figure 4 (c): Focuses on the long-term evolution and strategic risks of the system under the influence of macroeconomic factors. It requires multi-path simulations that combine various possible scenarios such as climate change, policy adjustments, and technological evolution to output the long-term evolution trend and potential risks of the system, which is suitable for strategic planning and investment decisions.

[0117] For example, taking a 10t / h gas-fired boiler scenario as an example, the output of its multi-timescale evaluation might look like this: Short-term assessment: Based on the measured operation data of a certain day, after supplementing some missing measurement points, the model assessment shows that the real-time coupling coordination degree of the system at the current moment is at a high level, indicating that the system is operating well.

[0118] Mid-term forecast: Combining recent measured data with trend forecasts, the model outputs a predicted trend of coupling coordination for each month within the next quarter. For example, under the preset condition of a monthly increase in load factor, the predicted value may show a gradual downward trend. The system can set early warnings based on this, and if the forecast shows a significant decrease, it will trigger operational optimization suggestions.

[0119] Long-term forecasts: Under specific scenarios such as "extreme drought" and "stricter carbon tax policies," model simulations show that in the long run, the system's efficiency may be under pressure, water demand may increase, while carbon emission intensity is expected to decrease. Simultaneously, the forecasts indicate a risk that the system's coupling coordination may fall into an imbalanced range during certain periods, and corresponding forward-looking regulatory recommendations are generated, such as "installing wastewater recycling devices."

[0120] 2. Dynamic iterative updates of evaluation elements To maintain the timeliness and accuracy of the evaluation model, the system establishes a regular update mechanism. Using continuously collected new measured data, key elements from previous evaluations (such as the impact assessment of virtual nodes and the model's prediction bias) are backtested and validated.

[0121] For example, in the monthly evaluation and update of the boiler system, the latest month's operating data can be used to reassess whether the impact of the "inlet water temperature" dummy node, which was previously defined due to data gaps, is reasonable. If a significant deviation is found between the calculation results based on the new data and the original evaluation value, the calculation logic or parameters for the impact of such dummy nodes can be corrected, allowing the system's evaluation capabilities to continuously improve over time.

[0122] (vi) Decision output: The aforementioned assessment and prediction results are integrated and processed to generate visual reports and optimization suggestions for decision-making, and can support collaborative applications by users at multiple levels.

[0123] 1. Multi-dimensional evaluation report and optimization plan The evaluation reports generated by the system are designed to provide comprehensive information for decision-making and typically include: Core performance indicators: Display key assessment results such as efficiency, water resource utilization rate, energy consumption intensity, carbon emission intensity, and system coupling coordination.

[0124] Data credibility analysis: through visualization charts (such as...) Figure 5 As shown, the quality of the data used for evaluation is presented intuitively, for example, by using different colors to distinguish high-confidence data, low-confidence data, and blank areas of data identified by virtual nodes.

[0125] For example, such as Figure 5 As shown, blue represents high-confidence data, red represents low-confidence data, and gray represents blank areas of data identified by virtual nodes. The distribution is illustrated with a schematic proportion (e.g., 65% high-confidence data, 20% low-confidence data, and 15% blank data with virtual nodes).

[0126] Multi-objective optimization schemes: Based on algorithms such as reinforcement learning, multiple technical and management optimization schemes are automatically generated with the goals of improving efficiency, saving water, saving energy, and reducing carbon emissions. These schemes can be tailored to different implementation cycles (such as short-term operational adjustments, medium-term technological upgrades, and long-term strategic upgrades), and include estimations and analyses of expected benefits and input costs.

[0127] For example, an evaluation report for a 10t / h gas-fired boiler might show a high efficiency and good coupling coordination. The reliability distribution chart reflects that most of the operating data is reliable, but there are some manually recorded data with lower reliability. The system may recommend optimization schemes including: short-term fine-tuning of operating parameters, medium-term installation of waste heat recovery devices, and long-term upgrading to low-NOx burners, and will provide an estimate of the expected effects and costs of each scheme.

[0128] This example clearly demonstrates the core advantage of this invention: its ability to identify hidden resource waste characterized by high flow rates and low efficiency. In the boiler case, the model does not judge whether natural gas consumption (flow rate indicator) meets standards in isolation, but rather couples it with efficiency (quality / efficiency indicator). If the system experiences a condition such as "flue gas flow rate meets standards but efficiency loss exceeds 30%", the model will output a deviation signal of "natural gas consumption is normal but efficiency is significantly low", thereby automatically locating and alerting to the hidden waste problem of excessive flue gas efficiency loss. Compared to traditional methods that only focus on flow statistics, this invention, by introducing the efficiency dimension and performing the above-mentioned coupled assessment, achieves insight into the essential efficiency of energy utilization, shifting from superficial flow management to in-depth efficiency tapping—the model analyzes natural gas consumption (flow rate indicator) and efficiency (quality indicator) in parallel, and when there is a significant deviation between the two, i.e., normal consumption but low efficiency, it can automatically identify the hidden waste condition of excessive flue gas efficiency loss.

[0129] 2. Multi-stakeholder collaborative decision support To adapt to the characteristics of integrated energy system management involving multiple stakeholders, this system supports multi-level and multi-role collaborative work. Through permission configuration, it can provide differentiated data views, report content, and functional interactions (such as viewing, commenting, and providing solution suggestions) for regulatory personnel, enterprise operation personnel, and park management personnel, thereby promoting information sharing and collaborative decision-making.

[0130] The evaluation framework and method proposed in this invention have broad scenario adaptability and can be applied to the evaluation of complex systems of different types, such as energy systems, industrial processes, and urban planning. It only requires scenario-based configuration and adjustment around the following four dimensions, without the need to reconstruct the model. Specifically: Data collection dimensions: Based on the characteristics of the scenario, expand the collection of unique core data. For example, energy systems need to focus on grid losses and renewable energy fluctuations; industrial processes need to focus on equipment parameters and raw material conversion rates; urban planning needs to focus on population density, transportation energy consumption, and infrastructure data.

[0131] Node processing dimensions: The benchmark library for identifying the first type of virtual nodes (high-frequency required missing attributes) differs across different scenarios. For example, industrial scenarios focus on process parameters, while urban scenarios focus on planning indicators.

[0132] Model constraint dimension: Embed the core physical or planning laws of the scenario into the model's physical constraint layer. For example, energy systems must follow power balance, industrial processes must follow the conservation of matter and energy, and urban planning must follow regional supply and demand balance.

[0133] Key evaluation dimensions: The time scale of the evaluation output should match the decision-making priorities and scenario management needs. For example, energy systems focus on short-term scheduling and medium-term forecasting, industrial processes focus on equipment efficiency and cost, and urban planning focuses on long-term strategy and risk.

[0134] By flexibly configuring the above dimensions, the same methodology framework can accurately serve different evaluation objectives, significantly improving the versatility and deployment efficiency of the methodology.

[0135] An example is an energy system scenario (e.g., a regional photovoltaic-storage microgrid system). 1. Data Acquisition and Boundary Delineation: The evaluation boundary is defined as a microgrid system including photovoltaic arrays, wind turbines, batteries, and critical loads. Exemplary data categories collected include: Data: Output power of each power generation unit (calculated based on output power and theoretical maximum work capacity), and power loss of distribution lines.

[0136] Water / resource data: (If the system includes water-cooled equipment) Cooling water consumption.

[0137] Energy data: Real-time output of photovoltaic / wind turbines (kW), battery charging and discharging power (kW), and load power (kW).

[0138] Carbon emission data: Indirect carbon emissions calculated based on the electricity exchanged between the microgrid and the main grid and the grid's average carbon emission factor.

[0139] Scene data: Weather forecast data (irradiance, wind speed), electricity price signals, and equipment rated parameters.

[0140] 2. Example of attribute node processing: Information reinforcement: For short-term missing power data of a certain wind turbine, linear interpolation is performed to reinforce it using data from adjacent wind turbines and historical data from the same period.

[0141] Virtual node identification: The first type of virtual node: By comparing with similar microgrid benchmark databases, it was found that "battery internal resistance" is a key health parameter that is commonly monitored. If the current system lacks this data, it will be constructed as a first type of virtual node, and its comprehensive impact will be calculated based on its physical correlation with system efficiency and scheduling reliability.

[0142] The second type of virtual node: At a certain moment, the photovoltaic output data suddenly drops to zero under clear weather conditions, and the communication status is marked abnormally. Its data credibility coefficient is judged to be lower than the threshold (such as 0.3). Therefore, the data point is identified as the second type of virtual node (severely abnormal, low credibility).

[0143] 3. Model building and physical constraint embedding: In the constructed deep learning hybrid model, the physical constraint layer embeds the following core physical laws: Node power balance constraint: ∑P_gen(t)+P_batt(t)-P_load(t)-P_loss(t)=0 (for any time t).

[0144] Energy storage SOC dynamic constraints: SOC_min≤SOC(t)≤SOC_max, and SOC(t+1)=SOC(t)+η·P_batt(t)·Δt / Capacity.

[0145] During training, the loss function will include a penalty term for predicted outputs that violate the above equation.

[0146] 4. Example of evaluation output format: After running the model, the system output includes: Key metrics: overall efficiency of microgrid systems, renewable energy penetration rate, equivalent carbon emission intensity, and source-storage-load coupling coordination degree (a scalar value between 0 and 1).

[0147] Multi-scale prediction report: Short term (future 24 hours): Outputs coupling coordination prediction curves at 15-minute intervals, and marks the periods when the predicted coordination is below the safety threshold (e.g., 0.7).

[0148] Mid-term (next week): Output the trend of the system's daily average coordination degree under different weather forecast scenarios, indicating that "under the continuous rainy scenario, the coordination degree may remain below 0.75 from the third day onwards, and the backup power plan needs to be activated."

[0149] Optimization Plans: For the predicted low-carbon coordination period, multiple optimization plans are generated, such as "adjusting energy storage charging and discharging strategies" and "suggesting the start of backup gas generators". The expected improvement effect of each plan (such as the coordination level increasing to 0.85), additional costs and changes in carbon emissions are compared in a list.

[0150] Data Acquisition: It is necessary to expand the collection of data closely related to the characteristics of power grid operation, such as power grid transmission losses, the fluctuation characteristics of renewable energy output, and the charging and discharging status of energy storage devices.

[0151] Node processing: The first type of virtual node is often related to the lack of key operating parameters of power generation equipment (such as photovoltaic panel irradiance).

[0152] Model constraints: Rules such as real-time power balance of the power grid must be embedded in the physical constraint layer to ensure that the evaluation results conform to the basic physical laws of power system operation.

[0153] Evaluation focus: More emphasis is placed on short-term (hourly / daily) real-time scheduling matching evaluation and medium-term (monthly / quarterly) output prediction and collaborative optimization.

[0154] Examples include industrial process scenarios (such as gas-fired boilers and chemical plants). Data Acquisition: It is necessary to focus on collecting data that reflects specific process equipment and reaction processes, such as equipment rated parameters, conversion rates of raw materials and products, and core process state variables (such as temperature and pressure).

[0155] Node handling: The first type of virtual node often appears when process medium parameters that are crucial to thermal efficiency calculation (such as inlet water temperature and reactant concentration) are missing.

[0156] Model constraints: Constraints such as energy balance, mass conservation, and reaction conversion rate limits of the process system must be embedded to ensure that the evaluation does not violate the principles of chemical reaction and thermodynamics.

[0157] Assessment focus: Focusing on energy efficiency analysis and optimization at the equipment / process level, and production cost control, to provide direct basis for production scheduling and technological transformation.

[0158] An example is an urban planning scenario (such as a high-tech industrial park). 1. Data Acquisition and Boundary Delineation: The assessment boundary is set as the entire geographical area of ​​the park. Exemplary data categories collected include: Energy data: The quantity and physical amount of heat, electricity, and cooling energy provided by the park's integrated energy station, and the energy consumption data of each major building.

[0159] Water data: total water intake of the park, water supply of different types (industrial and domestic), utilization rate of reclaimed water, and leakage rate of pipeline network.

[0160] Carbon emission data: Total carbon emissions for primary, secondary, and tertiary emissions within the park area, calculated based on activity level data such as energy consumption and traffic flow.

[0161] Scenario data: enterprise occupancy rate, industry type distribution, building area, green building ratio, future new project plans, and regional climate data.

[0162] 2. Example of attribute node processing: Identification of Category I Virtual Nodes: Compared with benchmark databases of similar green parks, "average thermal performance of building envelope" is a high-frequency key indicator. If the current park lacks system detection data for this, it will be constructed as a Category I virtual node, and its impact on the accuracy of building energy consumption prediction will be evaluated.

[0163] The second type of virtual node identification: The monthly water consumption data reported by a certain enterprise is abnormally high compared with its historical average and the water consumption quota of similar enterprises, and there is no reasonable explanation such as expansion. This data point is identified as a second type of virtual node and needs to be corrected or downgraded in conjunction with on-site verification.

[0164] 3. Model building and embedding of physical / planning constraints: The physical constraint layer embeds the following macroscopic laws: Resource supply and demand balance constraint: Total energy supply (t) ≥ Total predicted energy demand (t).

[0165] Infrastructure capacity constraint: Forecast peak load ≤ substation / pipeline design capacity × safety factor.

[0166] Carbon emission budget constraint: Cumulative predicted carbon emissions ≤ total carbon emission control target of the park.

[0167] 4. Example of evaluation output format: After running the model, the system output includes: Key indicators: GDP intensity per unit of the park, water resource recycling rate, carbon emission intensity reduction rate, and energy-water-carbon system coupling coordination degree.

[0168] Long-term scenario forecast report: Under three scenarios—"benchmark development," "enhanced energy conservation," and "energy transition"—the evolution path of the aforementioned core indicators of the park over the next 10 years is simulated and output.

[0169] It is clearly stated that "under the baseline scenario, the system coupling coordination will enter a downward trend after 2028 due to the rapid growth of carbon emissions; the energy transition scenario can enable the coordination to steadily improve." Strategic optimization recommendations: Generate a comprehensive package of optimization solutions, including "full coverage of distributed photovoltaic power generation + construction of smart microgrids", "near-zero discharge transformation of industrial wastewater", and "establishment of corporate carbon accounts and trading mechanisms".

[0170] The analysis points out that each plan contributes to improving different core indicators, investment estimates, and investment payback periods, in order to support management decision-making.

[0171] Data collection: Data reflecting the macro-characteristics of the regional system needs to be collected, such as population and building density distribution, traffic energy consumption level, and operation and loss data of various infrastructure (water, heat, electricity) networks.

[0172] Node processing: The first type of virtual node may involve the lack of planning indicators (such as greening irrigation quotas).

[0173] Model constraints: Macro-level constraints such as the supply and demand balance of regional resources / energy and the carrying capacity of infrastructure need to be considered.

[0174] Assessment focus: Long-term (interannual) system evolution trends, low-carbon development path planning, and benefit assessment of major infrastructure investments.

[0175] The examples given above demonstrate that by configuring data, selecting rules, and setting goals in a targeted manner across the aforementioned dimensions, the same evaluation method can accurately serve evaluation needs at different scales and types, from device level and system level to regional level, achieving a high degree of methodological universality and deployment flexibility.

[0176] This invention further provides an AI-driven energy-water carbon chain coupling assessment system for implementing the aforementioned AI-driven energy-water carbon chain coupling assessment method. The system architecture is as follows: Figure 2 As shown, the specific functions and methods of each module correspond one-to-one: the data acquisition module corresponds to step S1, the attribute node processing module corresponds to step S2, the association calculation module corresponds to step S3, the coupled modeling module corresponds to step S4, the evaluation and prediction module corresponds to step S5, the decision output module corresponds to step S6, and the log auditing module is responsible for recording and auditing the data and operations of the entire process. Figure 2 As shown, the modules work together, as described in detail below: The data acquisition module is responsible for providing standardized basic data for the evaluation. Its main functions include: 1) Multi-source interface adaptation: Supports data access from various heterogeneous data sources such as industrial sensors, urban statistical databases, and energy management platforms; 2) Data preprocessing: Cleaning the raw data, including outlier removal and format conversion based on statistical principles; 3) Boundary Delineation: Provides visualization tools to assist users in defining the physical and logical boundaries of the system being evaluated; 4) Data Management: Store and manage the collected and processed data, and support subsequent retrieval.

[0177] The attribute node processing module is responsible for improving the completeness and reliability of the data used in the evaluation. Its main functions include: 1) Node Management: Maintain an attribute node library for different scenarios and support the filtering of core real attribute nodes; 2) Information reinforcement: For data gaps at different time scales (short-term, medium-term, long-term), appropriate strategies are adopted for reinforcement, such as using interpolation methods, time series prediction models (such as LSTM), or system dynamics scenario analysis. 3) Virtual node processing: Identify and construct two types of virtual nodes—the first type of virtual nodes that are “high-frequency mandatory missing” are identified by statistical frequency, and the second type of virtual nodes that are “abnormal or low confidence” are identified by setting a threshold. The latter can be modified based on physical rules.

[0178] The correlation calculation module is responsible for quantifying the dynamic correlation between attribute nodes and the core elements of "water, energy, and carbon". Its main functions include: 1) Physical correlation modeling: Built-in physical mechanism models for different evaluation scenarios (such as flow networks and matter conservation equations) are used to calculate the theoretical basic correlation weights; 2) Data-driven optimization: Utilize algorithms such as mutual information entropy to analyze information redundancy between nodes and adjust weights, and integrate the attention mechanism of deep learning models (such as CNN-LSTM) to dynamically optimize weights; 3) Integration of the impact of virtual nodes: The impact of the two types of virtual nodes mentioned above is quantified into correction coefficients and integrated into the final association weight calculation.

[0179] The coupled modeling module is responsible for building, training, and maintaining the core evaluation model. Its main functions include: 1) Hybrid Model Construction: Provides the ability to construct deep learning hybrid models that include an input layer, a feature extraction layer (such as CNN for spatial feature extraction and LSTM for temporal feature extraction), a physical constraint layer, and an output layer, where the physical constraint layer is used to embed domain physical laws (such as equilibrium equations). 2) Model training and optimization: The model is trained using historical data and optimized using a loss function that combines a prediction error term and a physical constraint violation penalty term; 3) Online iterative update: When the new data accumulated by the system reaches a certain scale, an incremental learning process can be triggered to fine-tune the model parameters to adapt to changes, and the stability of the model after iteration can be ensured through a verification mechanism.

[0180] The assessment and prediction module is responsible for performing system state assessments and trend predictions across multiple time scales. Its main functions include: 1) Multi-scale analysis: Supports short-term real-time status assessment, medium-term operational trend prediction, and long-term multi-scenario strategic simulation; 2) Accuracy Management and Verification: Continuously verify the model evaluation and prediction results, and trigger the optimization and update of model parameters based on accuracy feedback; 3) Risk warning: Based on the prediction results and user-defined thresholds, graded warnings are issued for potential system imbalance risks; 4) Dynamic iteration of virtual nodes: Periodically use newly added measured data to re-evaluate and update parameters such as the influence of virtual nodes.

[0181] The decision output module is responsible for transforming evaluation results into decision support information. Its main functions include: 1) Comprehensive report generation: Automatically generates an evaluation report that includes core evaluation indicators (such as efficiency and coupling coordination), data quality (credibility) analysis, and multi-objective optimization suggestions; 2) Visual Interaction: The evaluation results, especially the data credibility distribution, are displayed intuitively through charts and other forms, and the results can be exported. 3) Multi-entity collaboration support: Supports configuring differentiated data views and function permissions for different roles (such as regulators and enterprise operators) to promote collaborative decision-making.

[0182] The log auditing module is responsible for ensuring the traceability and auditability of the evaluation process. Its main functions include: 1) Full-process log recording: Completely record key operations, intermediate data and final results at each stage from data input to decision output; 2) Traceability and Query: Provides a flexible query interface, supporting full-link data tracing and review of any historical evaluation case; 3) Audit support: Provide necessary tools and interfaces to assist in auditing the compliance and technical rationality of the assessment process.

[0183] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An AI-driven method for evaluating the coupling of energy, water, and carbon chains, characterized in that... Includes the following steps: S1: Multi-source data acquisition and boundary delineation: Delineate the assessment boundary based on the physical equipment scope, system topology, or administrative region of the assessment object; collect thermal data, water data, energy data, carbon emission data, and scenario-related data within the boundary. The thermal data is obtained by measuring the temperature, pressure, flow rate, and composition parameters of the system or equipment and then calculating them according to thermodynamic formulas; determine the assessment scenario type, which is divided into energy system, industrial process, or urban planning. S2: Attribute node processing involves filtering and evaluating real attribute nodes that are suitable for the scenario, and reinforcing missing information in real attribute nodes according to a time scale, where the time scale is divided into short-term, medium-term, or long-term; simultaneously, it identifies and constructs two types of virtual nodes, including a first type of virtual node and a second type of virtual node, wherein: Based on the benchmark library of real attribute nodes of the same type of evaluation object, the frequency of occurrence of each attribute is counted, and the attributes that have an occurrence frequency of not less than a preset first threshold and are missing in the current evaluation object are defined as the first type of virtual node. Based on historical data statistical principles or industry standards, numerical thresholds are set. Real attribute nodes whose values ​​exceed the thresholds and have no reasonable explanation, or whose data collection scenarios are abnormal and whose data credibility coefficients are lower than the preset second thresholds, are defined as the second type of virtual nodes. Calculate the degree of influence of the first type of virtual nodes on the evaluation results, and classify the data quality of the second type of virtual nodes to obtain the corresponding quality parameters; S3: Calculation and correction of association weights. Based on the influence degree and quality parameters of the real attribute nodes and the virtual nodes obtained in step S2, the association weights of the real attribute nodes and the four-dimensional core elements of water, energy and carbon are calculated by integrating physical mechanism modeling, information entropy analysis and deep learning attention mechanism optimization, and the influence correction coefficient of the virtual nodes is determined. S4: Four-dimensional coupling mechanism modeling, taking the correlation weights and influence correction coefficients as inputs, constructing a deep learning hybrid coupling model with a physical constraint layer, training and optimizing the model parameters; S5: Multi-timescale coupling assessment and dynamic prediction. Based on the optimized coupling model and combined with the information characteristics of different time scales, it outputs the comprehensive performance assessment results and future trend prediction of the four-dimensional coupling system of carbon-water-energy-carbon. S6: Optimize decision-making and visualization output, outputting an evaluation report that includes virtual node prompts, credibility distribution and data enhancement suggestions, as well as multi-objective optimization schemes for "improvement-water saving-energy saving-carbon reduction" adapted to the scenario.

2. The AI-driven energy-water-carbon chain coupling evaluation method according to claim 1, characterized in that, In step S2, the missing information of the real attribute nodes is reinforced according to the time scale, specifically including: For short-term missing information, linear interpolation or data supplementation from adjacent sensors is used to reinforce it. For missing information in the medium term, an LSTM time series prediction model is used to supplement it, with historical data from the same period and scenario trend data as input; For information that has been missing for a long time, a system dynamics model combined with multi-scenario analysis is used to supplement it.

3. The AI-driven energy-water-carbon chain coupling evaluation method according to claim 1, characterized in that, In step S2: The calculation of the impact of the first type of virtual nodes on the evaluation results specifically includes: calculating the impact of a single node and calculating the comprehensive impact of multiple nodes based on the impact of each single node. The classification of data quality for the second type of virtual nodes specifically includes: classifying anomaly level and credibility level.

4. The AI-driven energy-water-carbon chain coupling evaluation method according to claim 1, characterized in that, In step S3, the correlation weights are calculated and the correction coefficients are determined through fusion optimization, specifically including: S31: Based on the physical mechanism model of the evaluation scenario, calculate the basic association weight between each real attribute node and the four-dimensional core element; S32: Use mutual information entropy to analyze the information overlap between the real attribute nodes, and reduce the weight of nodes with high overlap. S33: Combining the attention mechanism of deep learning models, the down-adjusted weights are dynamically optimized to obtain the AI-corrected weights; S34: Based on the influence degree and quality parameters of the two types of virtual nodes obtained in step S2, determine the influence correction coefficient, correct the AI-corrected weights, and obtain the final association weights.

5. The AI-driven energy-water-carbon chain coupling evaluation method according to claim 1, characterized in that, In step S4, constructing a deep learning hybrid coupled model including a physical constraint layer specifically includes: The model structure includes an input layer, a CNN layer for extracting spatial correlation features, an LSTM layer for extracting temporal correlation features, a physical constraint layer, and an output layer; The physical constraint layer embeds one or more physical law constraints, which include the equilibrium equation, the water-energy conservation law, or the carbon emission calculation logic. During model training, the loss function includes a physical constraint penalty term to penalize outputs that violate physical laws.

6. The AI-driven energy-water-carbon chain coupling evaluation method according to claim 1, characterized in that, In step S5, the multi-timescale coupled evaluation and dynamic prediction specifically includes: Short-term assessment: Based primarily on measured data, supplemented by time-series forecasting models to complete short-term fluctuation data, and assess the real-time coupling coordination degree; Mid-term prediction: Integrating measured data and LSTM prediction data, outputting the trend of coupling state changes and the critical triggering conditions of the coordinated state; Long-term forecasting: Combining climate models and policy scenario analysis, a multi-scenario forecasting model is constructed to output the long-term evolution trend and potential risks of the system.

7. The AI-driven energy-water-carbon chain coupling evaluation method according to claim 1, characterized in that, In step S6, the optimization decision and visualization output specifically include: Output core evaluation indicators and their levels. The core evaluation indicators include efficiency, water resource utilization rate, energy consumption intensity, carbon emission intensity, and four-dimensional coupling coordination degree. Output a data confidence distribution visualization chart, which uses different colors to distinguish high confidence data, low confidence data, and first-type virtual node blank data; Based on reinforcement learning algorithms, multiple optimization schemes are generated with the goals of improving efficiency, saving water, saving energy, and reducing carbon emissions. The implementation costs and expected benefits of each scheme are then analyzed.

8. An AI-driven energy-water-carbon chain coupling assessment system, used to implement the AI-driven energy-water-carbon chain coupling assessment method according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect multi-source data and define the evaluation boundaries; The attribute node processing module is used to filter real attribute nodes, supplement missing information, and identify and construct two types of virtual nodes. The association calculation module is used to calculate the association weight between real attribute nodes and four-dimensional core elements, as well as the influence correction coefficient of virtual nodes. The Coupled Modeling module is used to build, train, and optimize deep learning hybrid coupled models that include a physical constraint layer; The evaluation and prediction module is used to perform coupled evaluation and dynamic prediction across multiple time scales. The decision output module is used to generate visual evaluation reports and multi-objective optimization schemes.

9. The AI-driven energy-water-carbon chain coupling evaluation system according to claim 8, characterized in that, The attribute node processing module includes: The information enhancement submodule is configured to use different algorithms to enhance missing information in the short, medium and long term. The virtual node processing submodule is configured to construct the first type of virtual nodes and calculate their impact, as well as identify the second type of virtual nodes and classify them into anomalies and credibility levels.

10. The AI-driven energy-water-carbon chain coupling evaluation system according to claim 8, characterized in that, The coupling modeling module includes: The model building unit is configured to provide a visual interface for users to build model structures that include CNN layers, LSTM layers, and physical constraint layers. The online incremental iteration unit is configured to automatically start incremental training to update the model parameters when the amount of newly added valid data reaches a threshold. Iterative archive units are used to store relevant information for each iteration, supporting querying and analysis.

Images