Thermoelectric decoupling comprehensive evaluation method based on self-evolution learning

Abstract

The invention discloses a thermoelectric decoupling comprehensive evaluation method based on self-evolution learning, and the method comprises the steps: building a thermocouple transformation evaluation index system and an index calculation model, and calculating the comprehensive weight of a secondary index in the evaluation index system; constructing historical environment state vectors based on historical multi-source information, and clustering the historical environment state vectors to obtain a plurality of scene clusters; at a new evaluation moment, acquiring a real-time environment state vector at the moment, determining corresponding scene clusters, and for each scene cluster, dynamically correcting the comprehensive weight through a self-evolution mechanism constrained by a rule base and driven by deviation; and obtaining secondary index quantitative data of each thermoelectric decoupling transformation scheme through the index calculation model, and calculating a comprehensive score of each thermocouple transformation scheme in combination with the corrected comprehensive weight of each secondary index.

Description

Technical Field

[0001] This invention relates to the field of scheme evaluation technology, and in particular to a comprehensive evaluation method for thermo-electric decoupling based on self-evolutionary learning. Background Technology

[0002] With the advancement of the "dual carbon" target, the power system urgently needs to improve its flexibility to absorb a high proportion of renewable energy. Cogeneration units, as the main source of heat in northern regions, suffer from severely limited peak-shaving capacity due to their "heat-driven power generation" operation mode. Cogeneration decoupling retrofitting, which decouples heat and electricity output through technical means, is a key approach to improving unit flexibility. Currently, mainstream retrofitting solutions include near-zero output from the low-pressure cylinder, bypass heating, waste heat recovery via heat pumps, thermal storage tanks, and electric boilers.

[0003] However, existing evaluation systems for the flexibility retrofitting of decoupled thermal power systems have the following problems: ① Static evaluation: Existing evaluations rely heavily on static indicator systems (such as financial indicators, technical performance, and environmental benefits), failing to consider dynamic factors such as unit operating status, policy adjustments, and market electricity prices. ② Fixed weights: Traditional comprehensive evaluation methods (such as AHP, entropy weight method, and TOPSIS) rely on fixed weights, making them unable to adapt to the differentiated needs of different regions, unit types, and operating scenarios. ③ Lack of learning ability: Existing methods cannot autonomously learn and optimize evaluation rules from historical retrofitting cases, resulting in evaluation results lagging behind technological development and practical feedback. ④ Conflicting multiple objectives: There are trade-offs among objectives such as economy, flexibility, and environmental protection, and traditional methods struggle to dynamically balance the priorities of each objective. Summary of the Invention

[0004] In view of the above-mentioned prior art, the present invention provides a comprehensive evaluation method for thermo-electric decoupling based on self-evolutionary learning, which mainly solves the technical problems existing in the above-mentioned background art.

[0005] To achieve the above objectives, the technical solution of this invention is implemented as follows: This invention discloses a comprehensive evaluation method for thermoelectric decoupling based on self-evolutionary learning, the aforementioned comprehensive evaluation method comprising: Establish an evaluation index system and index calculation model for thermocouple modification, and calculate the comprehensive weight of the secondary indicators in the evaluation index system; Historical environment state vectors are constructed based on historical multi-source information, and multiple scene clusters are obtained by clustering the historical environment state vectors. At the new evaluation moment, the real-time environmental state vector at that moment is obtained, and the corresponding scene cluster is determined. For each scene cluster, the comprehensive weight is dynamically corrected through a self-evolution mechanism driven by rule base constraints and bias. The quantitative data of the secondary indicators of each thermocouple decoupling modification scheme are obtained through the index calculation model, and the comprehensive score of each thermocouple modification scheme is calculated by combining the comprehensive weight of each secondary indicator.

[0006] Optionally, the thermocouple retrofit evaluation index system includes several primary and secondary indicators. The primary indicators include financial indicators, thermo-electric decoupling retrofit indicators, energy consumption indicators, environmental protection indicators, and macroeconomic policy indicators. The financial indicators include three secondary indicators: net present value, internal rate of return, and dynamic payback period. The thermoelectric decoupling retrofit indicators include four secondary indicators: increased peak-shaving capacity, increased maximum heating capacity, technical safety, and technical applicability. The environmental protection indicators include four secondary indicators: sulfur dioxide, carbon dioxide, nitrogen oxides, and particulate matter emission reduction. The macroeconomic policy indicators include two secondary indicators: policy support and promotion of the development of the thermal power industry.

[0007] Optionally, the subjective weights of the secondary indicators are calculated using the analytic hierarchy process (AHP), and the objective weights of the secondary indicators are calculated using the entropy weight method. The subjective and objective weights are then multiplied, summed, and normalized to obtain the comprehensive weights.

[0008] Optionally, a historical environmental state vector can be constructed based on historical multi-source information, including: Collect historical multi-source information, including historical policy and market information, historical unit operation data, and historical project feedback information; Natural language processing technology is used to analyze historical policy documents and quantify them into a policy intensity index. Historical market prices and historical unit operation data are normalized. Deviation statistics are performed on the predicted values ​​and actual values ​​in historical project feedback information to obtain standardized multi-source historical data. Based on the standardized multi-source historical data, multiple historical environment state vectors are constructed according to a preset time dimension or project cycle dimension.

[0009] Optionally, the historical environment state vectors are clustered to obtain multiple scene clusters, including: The K-means unsupervised clustering algorithm is used to perform cluster analysis on all historical environmental state vectors. The range of the number of clusters is preset, and the cluster groups are determined by calculating the Euclidean distance between vectors. Calculate the center vector of each cluster group. The center vector is the mean vector of each component of all historical environmental state vectors in the corresponding cluster group. Each cluster group corresponds to a scene cluster, and each scene cluster represents a group of operating scenes with similar environmental characteristics.

[0010] Optionally, at a new evaluation moment, the real-time environment state vector at that moment is obtained, and the corresponding scene cluster is determined, including: calculating the Euclidean distance between the real-time environment state vector and the center vector of each scene cluster, and selecting the scene cluster corresponding to the center vector of the scene cluster with the smallest Euclidean distance as the matching scene cluster at that evaluation moment.

[0011] Optionally, for each scene cluster, the comprehensive weights are dynamically adjusted through a rule base constraint and bias-driven self-evolution mechanism, including: Configure a weight correction vector and an environment sensitivity coefficient matrix for each scene cluster, and construct the local loss function corresponding to the scene cluster; The weight correction vector and the environmental sensitivity coefficient matrix are iteratively updated based on the local loss function. The iterative update of the weight correction vector and the environmental sensitivity is completed after the decrease of the local loss function is less than a preset threshold or the preset number of iterations is reached. The final update result is normalized to obtain the final corrected result of the weight correction vector and the environmental sensitivity coefficient matrix.

[0012] Optionally, the rule base contains multiple constraint rules, which are used to limit the initial values ​​and ranges of the weight correction vector and environmental sensitivity coefficient matrix corresponding to each scene cluster. The constraint rules are set in a preset priority order.

[0013] Optionally, construct the local loss function corresponding to the scene cluster, specifically including: Determine the set of implemented solutions corresponding to the target scenario cluster, wherein the implemented solutions are thermal-electric decoupling retrofit projects that belong to the scenario cluster and have obtained actual operating data; Calculate the multidimensional deviation vector for each implemented solution and convert the multidimensional deviation vector into a comprehensive deviation scalar. The multidimensional deviation vector includes instantaneous deviation, statistical characteristic deviation, and temporal morphological deviation. Assign a project importance weight to each implemented project, which is determined based on project size, runtime, and scenario matching degree; A local loss function is constructed based on the comprehensive deviation scalar and the project importance weights.

[0014] Optionally, the method further includes, after the implementation of the scheme, collecting the actual operation data of all implemented schemes and the data calculated by the index calculation model, calculating the corresponding global multidimensional deviation vector, and combining the global multidimensional deviation vector to update the physical mechanism parameters, weight generation parameters and rule threshold parameters of the index calculation model through a hierarchical parameter optimization algorithm.

[0015] The beneficial effects of this invention are as follows: by introducing a self-evolutionary learning mechanism, relying on a multi-source signal sensing module to capture external dynamic signals such as policies, markets, and unit operation in real time, and combining a preset rule base to dynamically correct and optimize the indicator weights, it can quickly respond to changes in the external environment, realize the dynamic adaptation of the evaluation model, greatly improve the real-time performance and accuracy of the evaluation results, effectively make up for the environmental adaptation defects of existing static models, and significantly improve the timeliness and accuracy of decision-making. By recording the evaluation and prediction results of each project and the actual operation data after commissioning, and by quantitatively calculating the multidimensional deviation between the two, the deviation information is fed back to the self-evolutionary learning mechanism for iterative optimization of the core parameters of the weight adjustment rules and indicator calculation model. With the continuous accumulation of historical cases, the model can gradually correct the initial cognitive bias, continuously conform to the actual operation law of thermal-electric decoupling transformation, realize the iterative evolution of its own capabilities, fundamentally reduce long-term decision-making risks, and improve the reliability of scheme selection. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning in the embodiments of this application; Figure 2 This is a technical logic diagram of the thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning in the embodiments of this application. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0019] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0020] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0021] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0022] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0023] Please refer to the attached document. Figure 1 This application provides a comprehensive evaluation method for thermoelectric decoupling based on self-evolutionary learning. The aforementioned comprehensive evaluation method includes: S1. Establish an evaluation index system and index calculation model for thermocouple modification, and calculate the comprehensive weight of the secondary indicators in the evaluation index system. Specifically, the primary and secondary indicators in the thermocouple modification evaluation index system constructed in this application are shown in Table 1 below: Table 1

[0024] The financial indicators include three secondary indicators: Net Present Value (NPV), Internal Rate of Return (IRR), and Payback Period. NPV and IRR are positive indicators; higher values ​​indicate better economic performance. Payback Period is a negative indicator; lower values ​​are better. NPV is a crucial indicator for measuring a project's profitability and is generally calculated based on the minimum return on investment. A NPV greater than zero indicates that the project investment is profitable and acceptable; a higher NPV indicates better investment returns. The calculation model for NPV is as follows:

[0025] In the formula, It is the cash inflow value; It is the value of cash outflow; t is the net cash flow in year t; n is the number of years in the calculation period; i is the set discount rate.

[0026] The internal rate of return (IRR) is the discount rate at which the total present value of annual cash flows equals the total present value of cash inflows over the entire project's calculation period. It reflects the profitability of the project investment process; a higher IRR indicates better returns. When the IRR is greater than or equal to the benchmark rate of return, the project is considered feasible. The financial IRR is generally approximated using a linear interpolation formula, and its calculation model is as follows:

[0027] In the formula, This is the net cash flow in year t; It is the internal rate of return.

[0028] The dynamic payback period refers to the time from project investment to profitability, taking into account the time value of money. Its calculation model is as follows:

[0029] When the result is less than the benchmark investment payback period, it indicates that the project or plan can recover the investment within the required time and the investment is feasible; otherwise, the project or plan is not feasible.

[0030] The thermoelectric decoupling index includes four indicators: increased peak-shaving capacity, increased maximum heating capacity, technical safety, and technical applicability. Increased peak-shaving capacity refers to the extent to which the unit's load regulation range expands under the same heating capacity. Both increased peak-shaving capacity and maximum increased heating capacity are important indicators for measuring the effectiveness of thermoelectric decoupling; the larger the value, the better the thermoelectric decoupling effect. Technical safety and applicability are qualitative indicators, representing the impact of the thermoelectric decoupling scheme on the safe operation of the unit and its applicability to different regions and different types of units, respectively.

[0031] Energy consumption indicators include three aspects: total new power generation, change in coal consumption per unit of power generation, and change in plant power consumption. All three are traditional energy consumption evaluation indicators for thermal power plant units. A higher total new power generation, lower coal consumption per unit of power generation, and greater savings in plant power consumption indicate lower operating costs and higher returns for the unit. The changes in coal consumption per unit of power generation and changes in plant power consumption referred to in this article represent the degree of savings between the modified operating conditions and the original operating conditions; these are positive indicators. The above indicators are calculated using the following model: A) Coal consumption calculation model for pure condensing units: Since the total coal consumption for power generation in a pure condensing steam turbine unit is related to the efficiency of the boiler, pipelines, generator, and turbine, the formula can be described as follows:

[0032] Where P represents the power generation capacity.

[0033] By consulting the unit's heat balance diagram, the boiler efficiency and turbine absolute internal efficiency under pure condensing conditions can be obtained. Substituting these into the above formula, the unit coal consumption for power generation at each load point under pure condensing conditions can be obtained.

[0034] The coal consumption q of a pure condensing unit can be expressed as a quadratic function of the power generation P. By establishing a linear model, the unit coal consumption of the unit under different operating conditions can be calculated, and the energy consumption changes of different thermal-electric decoupling retrofit schemes can be analyzed.

[0035] In the formula, Coal consumption per unit of electricity generation; This represents the marginal coal consumption rate under pure condensing conditions. Let P be the no-load coal consumption; P be the power generation capacity. The formula for unit power generation coal consumption under pure condensing conditions can be obtained through linear regression.

[0036] B) Coal Consumption Calculation Model for Cogeneration Units: Since steam from heating units typically first enters the turbine to perform work, and the resulting low-quality steam then enters the heating network system for reheating. Because the heat of the heating steam is essentially not lost within the turbine, it can be understood that the heating steam comes directly from the boiler, and the coal consumption for heating mainly depends on the boiler efficiency. The method for calculating the comprehensive coal consumption of cogeneration based on the ratio of power generation to heating power is called the heat method. Combining this with the formula for the total coal consumption of a pure condensing turbine unit, the total coal consumption of a cogeneration unit can be described as:

[0037] Therefore, the unit coal consumption of a combined heat and power unit can be further described as:

[0038]

[0039]

[0040] In the formula, This refers to the unit coal consumption of a combined heat and power (CHP) unit. Coal consumption per unit of power generation Coal consumption per heating unit; The share of power generation in a combined heat and power (CHP) unit. This refers to the heating portion of the combined heat and power (CHP) unit.

[0041] C) Coal Consumption Calculation Model for Bypass Heating Retrofit: Compared to conventional cogeneration units, the modified cogeneration unit with bypass heating adds a back-pressure turbine. The efficiency and unit coal consumption of the back-pressure turbine differ from those of pure condensing generator units and cogeneration units. Based on the heat method calculation principle, the comprehensive heat consumption of the modified cogeneration unit with bypass heating can be considered to consist of three parts: coal consumption for power generation by the pure condensing unit, coal consumption for power generation by the back-pressure turbine, and coal consumption for heating. The total coal consumption calculation formula is as follows:

[0042] Therefore, the unit coal consumption of the unit after bypass modification can be described as:

[0043] Environmental protection indicators include four categories: sulfur dioxide, carbon dioxide, nitrogen oxides, and particulate matter emissions. These are the most important emission monitoring indicators for thermal power plants and are of great significance in evaluating the energy conservation and emission reduction effects of thermal power plants. All four indicators are positive indicators; the higher the value, the better the emission reduction effect. According to the Environmental Statistics Handbook, the emissions of sulfur dioxide and nitrogen oxides can be calculated based on material measurements.

[0044]

[0045] In the formula, B is the coal consumption; F is the conversion rate of sulfur in coal to sulfur dioxide (0.9 for thermal power plant boilers, 0.85 for industrial boilers and kilns, and 0.8 for commercial furnace hoods); S is the total sulfur content in coal; NSO2 is the desulfurization efficiency; and B is the amount of nitrogen converted from fuel to fuel. - NO conversion rate; n - nitrogen content in fuel; Vy - amount of flue gas generated from fuel; CNOX - concentration of temperature-dependent NO generated during combustion.

[0046] Formula for calculating smoke and dust generation:

[0047] In the formula, M1 is the amount of flue gas generated by the boiler, B is the fuel consumption under the maximum continuous output condition of the boiler, W(Aar) is the ash content of the fuel as received; Qnet,ar is the lower heating value of the fuel as received; q4 is the heat loss due to incomplete combustion of solids; and afh is the fly ash coefficient.

[0048] Macroeconomic policy indicators include two secondary indicators: policy support and promotion of the thermal power industry. Policy support indicates the level of policy support from national and local governments; the stronger the policy support, the better the potential economic benefits of technological upgrading projects. The performance of these two indicators is primarily quantified through expert evaluation, with higher values ​​indicating better results.

[0049] In some optional implementations, after determining the primary and secondary indicators in the thermocouple modification evaluation index system, for each secondary indicator, the subjective weight of the secondary indicator is calculated using the analytic hierarchy process (AHP), and the objective weight of the secondary indicator is calculated using the entropy weight method. The subjective weight and the objective weight are multiplied, summed, and normalized to obtain the comprehensive weight.

[0050] When using the Analytic Hierarchy Process (AHP) to calculate the subjective weights of the secondary indicators, since the application scenarios, units, and value ranges of different indicators vary, it is necessary to preprocess the collected raw indicator data to eliminate the influence of numerical characteristics on subsequent weight calculations and evaluation results in order to establish a unified reference system.

[0051] Assuming there are m evaluation objects and n evaluation indicators, the original data matrix X is defined as follows:

[0052] For indicators containing negative or zero values ​​in the original data matrix X, interval processing is performed to preserve the magnitude relationship between parameters while avoiding interference from outlier values. The processing formula is as follows:

[0053] For positive indicators such as net present value and increased peak-shaving capacity, the larger the value of the positive indicator, the higher the performance of the thermal-electric decoupling flexibility retrofit technology. To standardize the positive indicators, fixed maximum and minimum values ​​are selected to eliminate the influence of actual indicator values. Therefore, the positive indicators are processed using the following formula:

[0054] For inverse metrics like dynamic recovery cycles, the processed data needs to still conform to the evaluation logic of "the larger the value, the better." Therefore, the following formula is used for forward transformation:

[0055] Finally, normalization is performed, and the normalization expression is:

[0056] Convert all indicator data into relative proportion form to obtain a standardized indicator matrix. .

[0057] After the data preprocessing described above, the comprehensive weight of each secondary indicator is calculated using a combination of the analytic hierarchy process (AHP) and the entropy weight method, employing both subjective and objective weighting. For subjective weight calculation, the AHP was used. Multiple experts from the fields of thermal power, environmental protection, and economics were invited to conduct pairwise comparisons of indicators within the same level using a 1-9 scale. If indicator i was considered "slightly more important" than indicator j, it was assigned a value of 3; if it was considered "strongly more important," it was assigned a value of 7; otherwise, it was assigned a value of 1 / 3 or 1 / 7 of the reciprocal. This process was used to construct the judgment matrix A, whose expression is:

[0058] This application uses the 1-9 scale method as shown in Table 2 below: Table 2

[0059] The scale 1 indicates equal importance, 3 indicates slightly important, 5 indicates relatively important, 7 indicates very important, 9 indicates extremely important, and 2, 4, 6, and 8 are the median values ​​for adjacent judgments.

[0060] Calculate the largest eigenvalue of judgment matrix A And its corresponding feature vector. Then, the feature vector is normalized to obtain the subjective weight vector Ws=(w s1 ,w s2 ,...,w sn ).

[0061] Objective weighting is calculated using the entropy weighting method. Based on standardized indicator data, the weight of the i-th evaluation object under the j-th indicator is calculated using the following formula:

[0062] The information entropy of the j-th indicator is then calculated using the following formula:

[0063] Finally, the objective weight vector is derived using the following formula:

[0064] The objective weight vector is Wo=(w o1 ,w o2 ,...,w on ).

[0065] The subjective weight vector and the objective weight vector are multiplied element-wise, and then summed and normalized to obtain the comprehensive weight:

[0066] S2. Construct historical environment state vectors based on historical multi-source information, and cluster the historical environment state vectors to obtain multiple scene clusters; The historical multi-source information includes historical policy and market information, historical unit operation data and historical project feedback information. All of the above historical multi-source information is extracted through a preset data interface. The historical policy and market information covers policy documents related to thermal power decoupling issued in previous years, including the National Energy Administration's compensation policy on flexibility retrofitting, as well as historical market price data such as carbon emission trading prices and peak-shaving ancillary service clearing prices. Historical unit operation data comes from the plant-level monitoring information system (SIS), distributed control system (DCS), and historical storage data of equipment ledgers. Specifically, it includes unit operation parameters such as generator active power, main steam flow, heating extraction steam flow, reheat steam temperature, real-time power generation coal consumption, and plant power consumption. Historical project feedback information comes from a case library built by the system. This case library fully records the evaluation results of each implemented thermal-electricity decoupling retrofit project, i.e., the predicted values ​​of each secondary indicator, as well as the actual operating data after the project was put into operation, i.e., the real values. After completing the collection of historical multi-source information, various types of information are standardized. For historical policy documents, natural language processing (NLP) technology is used to parse the text content, extract key information such as core constraints, subsidy levels, and assessment standards, and transform them into a calculable policy intensity index through preset quantitative rules. For example, a carbon price threshold is set, and when the historical carbon emission price rises above the threshold, the environmental protection-related component in the corresponding policy intensity index is adjusted upwards accordingly. For historical market price data and historical unit operation data, a unified normalization method is adopted to eliminate the difference in dimensions between different data. For historical project feedback information, the deviation between the predicted and actual values ​​of each secondary indicator is calculated and statistically analyzed. When enough implemented project cases are accumulated in the case library, the system will further analyze the deviation pattern of the model prediction.

[0067] Based on the standardized multi-source historical data, the data is divided according to a preset time dimension or project cycle dimension. The preset time dimension can be selected as annual, quarterly, or monthly, and the project cycle dimension can be selected as the entire life cycle of a single thermal-electric decoupling retrofit project. For each time interval or project cycle, all standardized multi-source historical data under that dimension are integrated to construct a unified historical environmental state vector with multiple dimensions.

[0068] In some implementations, historical environment state vectors are clustered to obtain multiple scene clusters, including: The K-means unsupervised clustering algorithm is used to perform cluster analysis on all constructed historical environmental state vectors. First, the range of values ​​for the number of clusters is preset based on the total number of historical environmental state vectors and the actual evaluation requirements. During the cluster analysis, the similarity between vectors is measured by calculating the Euclidean distance between any two historical environmental state vectors. Vectors with close Euclidean distances and similar environmental characteristics are divided into the same cluster. After clustering is completed, the center vector of each cluster is calculated. The center vector is the mean vector of each component of all historical environmental state vectors in the corresponding cluster. Each cluster corresponds to a scenario cluster. Each scenario cluster represents a group of decoupled thermal and power operation scenarios with similar environmental characteristics, such as "low carbon price + general policy + medium unit thermal and power load", "high carbon price + strong environmental protection policy + high unit peak shaving pressure", and "winter heating peak + large electricity price fluctuation + high supply pressure".

[0069] Preferably, at a new evaluation time, multi-source information such as policy, market, and unit operation is extracted from a preset data interface. After processing through the aforementioned steps, standardized information is obtained and uniformly constructed into a real-time environmental state vector for that evaluation time t. The scene attribution for the new evaluation time t is determined by the nearest neighbor center principle, that is, the Euclidean distance between the real-time environmental state vector and the center vector of each scene cluster is calculated. By comparing the magnitude of the Euclidean distance, the scene cluster corresponding to the center vector of the scene cluster with the smallest Euclidean distance is selected as the matching scene cluster for that evaluation time.

[0070] S3. At the new evaluation moment, obtain the real-time environmental state vector at that moment and determine the corresponding scene cluster. For each scene cluster, dynamically correct the comprehensive weight through the self-evolution mechanism driven by rule base constraints and deviation. Specifically, the adjustment of the comprehensive weights is based on a self-evolutionary adjustment mechanism for different scenario clusters. First, a set of scenario-specific weight adjustment factors are matched for each scenario cluster, including: indicator layer weight adjustment vector. and environmental sensitivity coefficient matrix The environmental sensitivity coefficient matrix Used to characterize the sensitivity of the weights of each evaluation index to each component of the environmental state.

[0071] In a given scenario and environmental conditions In this case, the present invention corrects the comprehensive weights using the following formula:

[0072]

[0073] In the formula, Let be the comprehensive weight of the j-th indicator, and let this comprehensive weight be the basic weight; For matrix The j-th row represents the scene The sensitivity of the j-th indicator to environmental conditions; Indicates to Normalization is performed so that the sum of the weights of all indicators is 1.

[0074] In an optional implementation, based on scenario-based weight generation, a two-layer mechanism of "rule base constraint + bias-based continuous optimization" is introduced to achieve continuous evolution of weights over time. Specifically, the aforementioned weight correction factors are constrained by a preset rule base. This constraint does not directly assign values ​​to the final weights but rather acts on { , The range and initial value of}. The rule base contains multiple constraint rules in the form of "IF [condition] THEN [action]", for example: Rule 1: If the IF carbon emission price continues to rise by more than 20% in the relevant scenario In the middle, environmental protection-related indicators will be included The lower limit is increased by 10% to ensure that the weight of environmental protection is not weakened; Rule 2: If the power grid issues an early warning of supply pressure under extreme weather conditions, then the sensitivity coefficients corresponding to peak shaving and heating security indicators will be used. The element limit is increased by 15%.

[0075] Rule 3: If historical cases show that the actual safety performance of a certain technical route (such as near-zero output of a low-pressure cylinder) is consistently better than the prediction, then the upper limit of the allowable value of the technical safety index of the scheme will be increased accordingly.

[0076] It should be noted that the above rules are merely exemplary rules, and actual rules may include more rules, and the rules can be added or reduced by those skilled in the art according to the actual situation.

[0077] When multiple rules conflict, this invention merges constraints according to the priority of "safety > environmental friendliness > economy" to obtain a feasible parameter domain, which is a weight correction vector. and environmental sensitivity coefficient matrix Boundaries during the iterative update process.

[0078] Based on the rule base constraints, a local loss function corresponding to the scene cluster is constructed to achieve bias-driven parameter optimization. The local objective function of the constructed scene cluster is as follows:

[0079] In the formula, Weighting the importance of different projects To integrate the bias scalar, the following iterative process is performed in each round of self-evolutionary update:

[0080]

[0081] , The learning rate; The feasible domain is determined by the rule base; This is a projection operator used to ensure that the updated parameters still satisfy the rule constraints and physical constraints (such as non-negative weights, and the weights of important indicators not being lower than the lower limit).

[0082] Based on the aforementioned local loss function, the weight adjustment vector is... and environmental sensitivity coefficient matrix Perform iterative updates and set the learning rate. , The parameters are continuously adjusted using gradient descent. During the iteration process, the updated parameters are projected onto the feasible parameter domain defined by the rule base using a projection operator, ensuring that the parameters meet the rule constraints and physical constraints (such as non-negative weights and weights of important indicators not lower than the lower limit). When the decrease in the local loss function is less than a preset threshold or the preset number of iterations is reached, the iteration stops and the final weight correction vector is obtained. and environmental sensitivity coefficient matrix .

[0083] The final weight correction vector and environmental sensitivity coefficient matrix Substituting these values ​​into the aforementioned expression for correcting the overall weight yields the corrected result for the overall weight.

[0084] In some optional implementations, a local loss function corresponding to the scene cluster is constructed, specifically including: Determine the set of implemented solutions corresponding to the target scenario cluster. For a specific target scenario cluster, select implemented thermal-electric decoupling retrofit projects belonging to that scenario cluster from the constructed high-fidelity structured case library. The selection criteria are: the environmental state vector at the time of project implementation belongs to the scenario cluster after cluster analysis, and the project has accumulated sufficient actual operating data (including measured data corresponding to evaluation indicators such as actual power generation coal consumption, emission reduction, financial benefits, and peak shaving performance) after commissioning, thus forming a set of implemented solutions specific to that scenario cluster.

[0085] Calculate the multidimensional deviation vector for each implemented solution and convert the multidimensional deviation vector into a comprehensive deviation scalar. The multidimensional deviation vector includes instantaneous deviation, statistical characteristic deviation, and temporal morphological deviation. Each implemented project is assigned a project importance weight, determined based on project size, runtime, and scenario matching degree. Project size is quantified by the amount of investment in the renovation, the installed capacity of the heating unit, or the heating area; the larger the size, the higher the weight. Runtime is calculated based on the actual number of months the project has been running stably; the longer the runtime, the more valuable the accumulated data, and the higher the weight. Scenario matching degree is obtained by normalizing the Euclidean distance between the environmental state vector at the time of project implementation and the center vector of the target scenario cluster; the smaller the distance, the higher the matching degree, and the higher the weight. The importance weight of each implemented project is obtained through a linear weighted formula.

[0086] Furthermore, the aforementioned multidimensional deviation vector is compressed into a scalar comprehensive deviation, which integrates the overall deviation degree of each indicator of the project. Then, using the comprehensive deviation scalar of all projects within the scene cluster as the core, a weighted sum is obtained to obtain the local loss function in the previous formula.

[0087] S4. Obtain the quantitative data of the secondary indicators of each thermocouple decoupling modification scheme through the indicator calculation model, and calculate the comprehensive score of each thermocouple modification scheme by combining the comprehensive weight of each secondary indicator after correction.

[0088] Specifically, in the calculation process of the fuzzy comprehensive evaluation method, in order to unify the evaluation standards and effectively handle the fuzziness and subjectivity in the evaluation, a four-level evaluation grade is set to form a comment set, such as V={Excellent, Good, Average, Poor}. In order to realize the quantitative calculation of comments, each evaluation grade is assigned a corresponding median score: Excellent (95 points), Good (85 points), Average (70 points), Poor (50 points). The higher the score, the better the comprehensive benefit of the scheme.

[0089] Furthermore, a membership matrix R is constructed to characterize the degree to which each scheme belongs to different evaluation levels under each secondary indicator. The matrix has dimensions m×n×4, where m is the number of schemes to be evaluated, n is the number of secondary indicators, and 4 is the number of evaluation levels. Different calculation methods are used for different indicator types. For example, for quantitative indicators such as net present value, sulfur dioxide emission reduction, and increased peak-shaving capacity, a trapezoidal membership function is used to calculate the membership degree based on the distribution characteristics and actual value range of the indicators. For inverse indicators (such as dynamic recovery cycle), the membership function is set in reverse. For qualitative indicators such as technical safety and policy support, multiple experts from the fields of thermal power, environmental protection, economics, and engineering are invited to score the qualitative indicators of each scheme based on the connotation of the indicators and the actual performance of the schemes (the scoring range corresponds to "excellent, good, average, and poor"). (Level 4) Analyze the expert scores, divide the number of votes for a certain level by the total number of experts to obtain the membership degree of the scheme under that qualitative indicator. Combine the membership degree calculation results of all quantitative and qualitative indicators to form a complete membership degree matrix.

[0090] Furthermore, the modified comprehensive weights and membership matrix are subjected to fuzzy synthesis to reflect the weight contribution of each indicator to the evaluation result. That is, for each evaluation level, the dynamic weights of each indicator are multiplied by the membership degree of the scheme under the corresponding level of that indicator and then summed to obtain the comprehensive membership vector B of the scheme belonging to the comment set V.

[0091] The weighted average method is used to transform the comprehensive membership vector B of each scheme into a quantitative comprehensive score. All the thermal power decoupling retrofit schemes to be evaluated are sorted in descending order. The scheme with the highest score is the scheme with the best comprehensive benefits at the current evaluation time, providing a direct decision-making basis for the selection of retrofit schemes for thermal power enterprises.

[0092] In some preferred embodiments, the method further includes, after the comprehensive evaluation results are output, providing them to decision-makers as a direct basis for project selection; on the other hand, the results (including input data, weights used, and output scores) are fully recorded in the historical case library, and after the implementation of the scheme, the model parameters are continuously corrected using project operation feedback to achieve continuous optimization of the evaluation model.

[0093] Specifically, after the comprehensive evaluation results are output, on the one hand, the comprehensive scores, ranking results, and core evaluation criteria of each thermal-electric decoupling retrofit scheme are provided to decision-makers as a direct basis for project selection; on the other hand, the complete information of this evaluation (including the original input data of each scheme, standardized data, dynamic weight vectors, membership degrees of each indicator, comprehensive scores, and final ranking results) is completely recorded in the historical case database, forming a structured case data object, providing a traceability basis for subsequent deviation calculation and parameter optimization. After the selected scheme is implemented and operates stably for a preset period (e.g., 6 months), the closed-loop optimization process is initiated. First, the actual operating data of the scheme is collected, with the collection scope completely corresponding to the evaluation indicator system. At the same time, historical information such as the evaluation input data, indicator calculation model parameters, and dynamic weight vectors corresponding to the scheme are retrieved from the historical case database as predicted values, establishing a one-to-one correspondence between "historical evaluation data and measured operating data".

[0094] Based on the collected actual operation data and retrieved historical evaluation data, a global multidimensional deviation vector is calculated, including instantaneous deviation, statistical characteristic deviation, and time series morphological deviation. The instantaneous deviation is calculated as follows:

[0095]

[0096]

[0097]

[0098] In the above formula, The point-by-point relative deviation between the predicted and actual values, Here is the direction-sensitive penalty function. As a time risk weight, To avoid tiny constants with a denominator of 0, This is used to reflect risk preferences for errors in different directions, such as imposing stricter penalties on the error of "overestimating returns" for economic indicators. For unit load rate, The coefficient is the index coefficient, which determines the sensitivity of the index to the load level.

[0099] This invention calculates the deviation of statistical values ​​over the entire lifecycle for statistical indicators (such as Internal Rate of Return (IRR), Net Present Value (NPV), and total CO2 emission reductions). It considers not only mean deviation but also volatility deviation, and assigns different weights to "overly optimistic" and "overly conservative" approaches. Therefore, the expression for calculating the statistical characteristic deviation vector is as follows:

[0100]

[0101]

[0102] In the formula, Used to strengthen penalties for "overestimating returns"; The volatility of actual computation; Predicted volatility; , For predictive statistics; This is the actual statistical measure; This type of bias is a scalar quantity, used on the one hand to measure the business value of the model, and on the other hand to reveal the model's systematic bias. It represents the model's ability to predict the overall benefits throughout the project cycle. It directly relates to the correctness of investment decisions. Excessive bias means the model severely underestimates the project's profitability, potentially leading to the rejection of excellent solutions. If the same statistical characteristic bias consistently shows positive or negative across multiple projects, it indicates a systematic bias in the model. This may be due to a generally optimistic or conservative tendency in the initial parameter settings, cost estimation formulas, or revenue calculation logic, requiring fundamental correction. The optimization objective is to bring this type of bias close to zero, ensuring that the evaluation results have high reference value for investors.

[0103] For calculating time-series characteristic deviations, for time-correlated indicators such as peak-shaving capacity, the actual and predicted load curves are extracted, and the similarity between the curves is calculated as the morphological deviation. The DTW algorithm is used to calculate the morphological difference between the two curves. A high time-series morphological deviation indicates that the model's evaluation of indicators such as "peak-shaving capacity" is distorted. The current weight allocation overestimates this "side-effect" peak-shaving capacity. Therefore, the system will reduce the weight of the "new peak-shaving capacity" indicator, while correspondingly increasing the weights of related indicators such as "technical security" and "changes in power generation coal consumption" to balance decisions and avoid selecting this "high-cost" peak-shaving scheme again. This achieves precise and correlated weight adjustment.

[0104] For newly added peak-shaving capacity, load-time curve, and other shape-sensitive indicators, this invention adopts a combination of "full curve shape + critical operating point error". Taking the load curve as an example, let the predicted curve be yi,jt, and the actual curve be yi,jt. The morphological distance is obtained using the DTW algorithm:

[0105] Simultaneously, select a set K of several key operating points (such as minimum load points, maximum load points, and typical peak-valley transition points), and calculate the errors at these points:

[0106] The morphological deviation is then:

[0107] Based on the aforementioned instantaneous bias, statistical characteristic bias, and time series morphological bias, a 3-dimensional bias vector is first constructed for each item-indicator (i, j):

[0108] Then utilize the index-level deviation compression vector Compress it into a scalar composite bias:

[0109] In the formula, It can be preset according to the indicator attributes (such as focusing on the time series deviation for the operation indicator and focusing on the morphological deviation for the peak shaving indicator).

[0110] In a further embodiment, the method further includes updating the physical mechanism parameters, weight generation parameters, and rule threshold parameters of the index calculation model by combining the global multidimensional deviation vector with a hierarchical parameter optimization algorithm.

[0111] Specifically, the model parameters in the aforementioned indicator calculation model are first grouped, and the model parameter vector is defined as follows:

[0112] in, For parameters related to the thermal power unit mechanism model, such as boiler efficiency correction coefficients and coal consumption model coefficients, etc. The parameters for the self-evolving weight generation module, including those for various scenarios. wait; These are threshold-type parameters in the rule base, such as the carbon price threshold Tc and the extreme weather warning intensity threshold.

[0113] Overall deviation at the project-indicator level Based on this, the following hierarchical loss function is constructed in this embodiment:

[0114] In the formula, , , These are sets of indicators for economy, flexibility, and environmental friendliness, respectively. , , The inter-level weights of the three types of objectives can reflect the decision-makers' preferences for different objectives; This represents the basic importance weight of the j-th evaluation indicator within its category, used to characterize the relative contribution of different indicators to the overall deviation; Indicator-level nonlinear mapping is used to limit the impact of extreme biases; A vector of prior parameters set by the experts; The regularization coefficient is used to prevent the model from deviating significantly from physical experience. Furthermore, to avoid unreasonable physical adjustments caused by "purely data-driven" approaches, this invention introduces a mechanistic model sensitivity matrix to weight the gradient. Let the j-th index calculation function output by the mechanistic model be... Then, its sensitivity to physical parameters can be approximated by finite difference:

[0115] Construct parameter-level sensitivity weights:

[0116] In the formula, Step size; The larger the value, the more sensitive the output index j is to the physical parameter p.

[0117] Calculate the gradient of the overall loss function with respect to the parameters. Then, a sensitivity-weighted gradient is applied to the physical parameter part:

[0118] This ensures that physical parameters that are highly sensitive to a certain indicator are updated first, while the update magnitude of parameters that have minimal impact on the model output is automatically suppressed, reducing the risk of invalid parameter tuning and physical inconsistencies.

[0119] Based on the construction of sensitivity-weighted gradients, this invention employs a grouped alternating optimization algorithm, with a fixed... In this case, stochastic gradient descent is used to generate parameters for the weights. Update

[0120] In the formula, Make the feasible domain given by the rule base; This is the learning rate.

[0121] In fixed In this case, the physical parameters are updated using a quasi-Newton method with sensitivity weighting. To ensure that the deviation between the mechanistic model and the measured data is minimized, the rule threshold parameter is used. Discrete, threshold-type parameters (such as carbon price thresholds and extreme weather warning thresholds) are not suitable for gradient descent. Instead, particle swarm optimization is used to find the combination of thresholds that maximizes the reduction in function loss within a finite candidate set. This process is repeated iteratively until the decrease in the hierarchical loss function is less than a preset threshold or the maximum number of iterations is reached. To accommodate the continuous addition of new projects, this invention employs an incremental update strategy: whenever a new batch of projects is added to the case library, a small batch of samples consisting of the latest cases is used to perform a short-term iteration on the parameters, allowing the model to gradually absorb the latest operational information and achieve true online self-evolution.

[0122] In some preferred embodiments of this application, a structured data object is created and stored for each project case that has been evaluated and put into operation. This object first fully records the input context at the time of the decision, that is, the system state vector at the moment of decision, which includes not only the real-time market environment, such as the average electricity price and carbon price sequence of the key period before the decision, but also the quantitative policy intensity index generated by parsing the policy text through natural language processing technology, as well as the initial state of the unit itself at the moment of decision.

[0123] Building upon this foundation, the case library precisely records the decision-making actions taken by the system, including the final dynamic weight vector adopted, the comprehensive evaluation score, and the resulting scheme ranking and final selection. Most importantly, the case objects contain detailed output and feedback information. Ultimately, this complete data chain, rich in context, decisions, results, and biases, is structured and stored in the case library, forming a knowledge base that continuously expands over time. This provides a high-quality, traceable supervised learning dataset for subsequent model parameter optimization, thereby substantially supporting the system's closed-loop self-evolutionary capability.

[0124] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A comprehensive evaluation method for thermoelectric decoupling based on self-evolutionary learning, characterized in that, The aforementioned comprehensive evaluation methods include: Establish an evaluation index system and index calculation model for thermocouple modification, and calculate the comprehensive weight of the secondary indicators in the evaluation index system; Historical environment state vectors are constructed based on historical multi-source information, and multiple scene clusters are obtained by clustering the historical environment state vectors. At the new evaluation moment, the real-time environmental state vector at that moment is obtained, and the corresponding scene cluster is determined. For each scene cluster, the comprehensive weight is dynamically corrected through a self-evolution mechanism driven by rule base constraints and bias. The quantitative data of the secondary indicators of each thermocouple decoupling modification scheme are obtained through the index calculation model, and the comprehensive score of each thermocouple modification scheme is calculated by combining the comprehensive weight of each secondary indicator.

2. The thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning according to claim 1, characterized in that, The thermocouple retrofit evaluation index system includes several primary and secondary indicators. The primary indicators include financial indicators, thermo-electric decoupling retrofit indicators, energy consumption indicators, environmental protection indicators, and macroeconomic policy indicators. The financial indicators include three secondary indicators: net present value, internal rate of return, and dynamic payback period. The thermoelectric decoupling retrofit indicators include four secondary indicators: increased peak-shaving capacity, increased maximum heating capacity, technical safety, and technical applicability. The environmental protection indicators include four secondary indicators: sulfur dioxide, carbon dioxide, nitrogen oxides, and particulate matter emission reduction. The macroeconomic policy indicators include two secondary indicators: policy support and promotion of the development of the thermal power industry.

3. The thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning according to claim 2, characterized in that, The subjective weights of the secondary indicators are calculated using the analytic hierarchy process (AHP), and the objective weights of the secondary indicators are calculated using the entropy weight method. The subjective and objective weights are multiplied, summed, and normalized to obtain the comprehensive weights.

4. The thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning according to claim 3, characterized in that, Constructing historical environment state vectors based on historical multi-source information includes: Collect historical multi-source information, including historical policy and market information, historical unit operation data, and historical project feedback information; Natural language processing technology is used to analyze historical policy documents and quantify them into a policy intensity index. Historical market prices and historical unit operation data are normalized. Deviation statistics are performed on the predicted values ​​and actual values ​​in historical project feedback information to obtain standardized multi-source historical data. Based on the standardized multi-source historical data, multiple historical environment state vectors are constructed according to a preset time dimension or project cycle dimension.

5. The thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning according to claim 4, characterized in that, Clustering the historical environment state vectors yields multiple scene clusters, including: The K-means unsupervised clustering algorithm is used to perform cluster analysis on all historical environmental state vectors. The range of the number of clusters is preset, and the cluster groups are determined by calculating the Euclidean distance between vectors. Calculate the center vector of each cluster group. The center vector is the mean vector of each component of all historical environmental state vectors in the corresponding cluster group. Each cluster group corresponds to a scene cluster, and each scene cluster represents a group of operating scenes with similar environmental characteristics.

6. The thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning according to claim 5, characterized in that, At a new evaluation moment, the real-time environment state vector at that moment is obtained, and the corresponding scene cluster is determined, including: calculating the Euclidean distance between the real-time environment state vector and the center vector of each scene cluster, and selecting the scene cluster corresponding to the center vector of the scene cluster with the smallest Euclidean distance as the matching scene cluster at that evaluation moment.

7. The thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning according to claim 5, characterized in that, For each scene cluster, the comprehensive weights are dynamically adjusted through a self-evolutionary mechanism driven by rule base constraints and biases, including: Configure a weight correction vector and an environment sensitivity coefficient matrix for each scene cluster, and construct the local loss function corresponding to the scene cluster; The weight correction vector and the environmental sensitivity coefficient matrix are iteratively updated based on the local loss function. The iterative update of the weight correction vector and the environmental sensitivity is completed after the decrease of the local loss function is less than a preset threshold or the preset number of iterations is reached. The final update result is normalized to obtain the final corrected result of the weight correction vector and the environmental sensitivity coefficient matrix.

8. The thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning according to claim 7, characterized in that, The rule base contains multiple constraint rules, which are used to limit the initial values ​​and ranges of the weight correction vector and environmental sensitivity coefficient matrix corresponding to each scene cluster. The constraint rules are set in a preset priority order.

9. The thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning according to claim 7, characterized in that, Constructing the local loss function corresponding to the scene cluster includes: Determine the set of implemented solutions corresponding to the target scenario cluster, wherein the implemented solutions are thermal-electric decoupling retrofit projects that belong to the scenario cluster and have obtained actual operating data; Calculate the multidimensional deviation vector for each implemented solution and convert the multidimensional deviation vector into a comprehensive deviation scalar. The multidimensional deviation vector includes instantaneous deviation, statistical characteristic deviation, and temporal morphological deviation. Assign a project importance weight to each implemented project, which is determined based on project size, runtime, and scenario matching degree; A local loss function is constructed based on the comprehensive deviation scalar and the project importance weights.

10. The thermoelectric decoupling comprehensive evaluation method based on self-evolutionary learning according to claim 1, characterized in that, The method further includes, after the implementation of the scheme, collecting the actual operation data of all implemented schemes and the data calculated by the index calculation model, calculating the corresponding global multidimensional deviation vector, and combining the global multidimensional deviation vector to update the physical mechanism parameters, weight generation parameters and rule threshold parameters of the index calculation model through a hierarchical parameter optimization algorithm.

Images