A method and system for dynamic modeling and capacity optimization analysis of the operating characteristics of district heating units

Abstract

This invention provides a method and system for dynamic modeling and capacity optimization analysis of the operating characteristics of district heating units. The method includes: collecting multi-source data and performing dual preprocessing (physical and statistical); constructing a dynamic characteristic model incorporating energy efficiency characteristics, operating response, and loss decay using XGBoost, LSTM, and random forest, and deeply embedding thermodynamic physical boundaries and physical information loss functions into the model; generating a multi-dimensional capacity profile of the unit based on the dynamic model, mapping it to the hydraulic and thermodynamic time-varying boundary constraints of a "unit-grid-load" collaborative simulation model; adaptively adjusting the optimization objective weights according to the supply and demand tension, generating a multi-objective optimal scheduling strategy through a MILP solver, and achieving online self-correction through a closed-loop feedback mechanism. This invention overcomes the limitations of traditional static models, achieving dual-driven analysis of data and physical mechanisms, and improving the energy efficiency, economy, and operational safety of the heating system.

Description

Technical Field

[0001] This invention belongs to the field of energy management and smart heating technology, and in particular relates to a method and system for dynamic modeling and capacity optimization analysis of the operating characteristics of district heating units. Background Technology

[0002] District heating is a crucial component of urban energy infrastructure, and the operational characteristics and capabilities of its core equipment (such as thermal power units, peak-shaving boilers, and heat pumps) directly affect the energy efficiency, economy, and reliability of the heating system. In actual production and operation, existing technologies exhibit the following prominent problems: 1. Static Modeling of Characteristics: Existing technologies mostly use design conditions or historical average data to build static models, which cannot reflect the dynamic and nonlinear operating characteristics of the unit under complex variable conditions, equipment aging, and environmental factors (such as temperature and wind speed).

[0003] 2. Limited capacity assessment: The assessment of unit capacity is often limited to rated heat supply or maximum output, lacking quantitative analysis and comprehensive evaluation of multi-dimensional capabilities such as "peak shaving capacity", "energy efficiency degradation characteristics", "fault tolerance margin" and "multi-energy complementarity potential".

[0004] 3. Isolated system analysis: Existing systems typically analyze individual units independently, lacking the ability to perform integrated "unit-network-load" analysis from the perspective of the entire heating network, considering factors such as multi-heat source coordination, hydraulic characteristics of the heating network, and changes in user-side load.

[0005] 4. Lagging decision support: Operational strategies rely heavily on human experience and lack predictive optimization and adaptive control suggestions based on real-time characteristics and capacity assessments, making it difficult to achieve accurate load allocation, economic scheduling, and risk warning.

[0006] Therefore, there is an urgent need to develop a method and system for analyzing the operating characteristics and capabilities of district heating units that can achieve dynamic modeling, multi-dimensional evaluation, system coordination, and intelligent optimization. Summary of the Invention

[0007] To address the aforementioned technical problems, the purpose of this invention is to provide a system and method for dynamic modeling and capacity optimization analysis of the operating characteristics of district heating units. This invention overcomes the limitations of traditional black-box artificial intelligence models, deeply integrating thermodynamic and physical mechanisms, enabling dynamic characterization of unit capacity and generation of system-level globally optimal scheduling instructions.

[0008] The first aspect of this invention discloses a method for dynamic modeling and capacity optimization analysis of the operating characteristics of district heating units, the method comprising: Step S1: Real-time acquisition of operating parameters, environmental parameters, and external data of related systems of the heating unit; Step S2: Perform data preprocessing on the collected data, including outlier removal, missing value imputation, and feature engineering; Step S3: Using the preprocessed data, construct and update the dynamic characteristic model of the heating unit online. The dynamic characteristic model includes an energy efficiency characteristic model, an operating condition response model, and a loss and attenuation model. Thermodynamic physical boundaries and mechanism features are introduced into the model construction for correction. Step S4: Based on the output of the dynamic characteristic model, calculate and generate a multi-dimensional dynamic capability profile of the unit in real time; Step S5: Integrate the hydraulic and thermal model of the heating network with the load prediction model, and transform the multi-dimensional dynamic capacity profile into time-varying boundary constraints of the network model to perform "machine-network-load" collaborative simulation analysis. Step S6: Based on the co-simulation results and the system supply and demand status, construct and solve a multi-objective optimization model, generate an optimized operation strategy, and issue it for execution; Step S7: Continuously monitor the prediction deviation after the strategy is executed. When the deviation exceeds the threshold, trigger the incremental learning of the model to achieve online closed-loop correction.

[0009] Preferably, in step S2, the outlier removal adopts a dual judgment combining physical rules and statistical rules, wherein the physical rules include a thermodynamic consistency judgment based on the minimum / maximum allowable supply and return water temperature difference, and a sensor inaccuracy judgment based on load-flow correlation. In the feature engineering, when extracting time-series features from the operating condition response model, an adaptive time-step sequence construction method is used, with the input time step L. t The length of the time step Lt is dynamically adjusted according to the thermal inertia characteristics of the unit and the measured outdoor temperature. The lower the outdoor temperature, the longer the time step Lt.

[0010] Preferably, in step S3, the energy efficiency characteristic model adopts the XGBoost algorithm, and its output is provided with a physical boundary constraint layer. The physical boundary constraint layer calculates the theoretical upper limit efficiency under the current cold and heat source temperature conditions based on the Carnot theorem, and forces the original XGBoost prediction values ​​to be projected and constrained within the physical feasible region determined by the Carnot upper limit multiplied by the technical reachability coefficient. When the number of times the original model prediction values ​​exceed the physical boundary reaches a threshold, the bias correction amount is calculated using the exponential weighted moving average, triggering the online recalibration of the model.

[0011] Preferably, in step S3, the operating condition response model adopts an LSTM network. The training loss function of the LSTM network not only includes a prediction error term, but also adds a physical information regularization loss. The physical information regularization loss includes an energy conservation residual term constructed based on the unit's metal / working fluid heat storage change rate and input-output heat balance equation, as well as a ramp physical constraint term calculated based on the turbine thermal stress model. The LSTM network introduces a working condition-aware attention mechanism, which incorporates a working condition change sensitivity factor that is proportional to the rate of change of control commands when calculating attention weights, thereby actively amplifying the model attention weights during varying working conditions such as load increases and decreases.

[0012] Preferably, in step S3, the loss and attenuation model adopts the random forest algorithm to quantify the deviation of equipment performance degradation; The performance degradation deviation is calculated using a thermodynamic back-calculation method based on multi-sensor data fusion: real-time heat supply is calculated based on the supply and return water temperature difference and flow rate, and real-time input heat is calculated based on fuel flow rate and lower heating value. The ratio of the two is the real-time actual efficiency, and then the performance deviation from the design efficiency is obtained. When constructing decision trees, the random forest adopts a prior splitting strategy based on thermodynamic failure mechanisms, assigning priority to physical cumulative features, including low-cycle thermal fatigue, high-temperature creep, and oxygen corrosion, in the candidate feature pool; and adopts an incremental pruning strategy, which, as the data time span increases, performs depth-limited pruning on deep branches trained on old data and reduces the voting weight of the tree.

[0013] Preferably, in step S4, the multi-dimensional dynamic capability profile covers an indicator system encompassing five dimensions: steady-state supply capability, dynamic adjustment capability, economic efficiency capability, reliability and security capability, and collaborative and mutual support capability.

[0014] Preferably, in step S5, the specific mapping logic for converting the dynamic capability profile into time-varying boundary constraints of the pipeline network model includes: The minimum / maximum thermal output of the unit under the current operating conditions is mapped to the steady-state supply capacity constraint of the heat source node flow boundary; The maximum load increase / decrease rate of the unit extracted from the operating condition response model is mapped to the dynamic adjustment capability constraint of the upper limit of the time change rate of the flow of the heat source node. When the failure risk index output by the loss and attenuation model exceeds the warning threshold, the maximum allowable output constraint of the unit is dynamically reduced according to the safety derating factor.

[0015] Preferably, in step S6, the multi-objective optimization model comprehensively considers three objectives: economic cost, carbon emissions, and equipment wear and tear. The weight coefficients of each objective are not fixed constants, but are adaptively and dynamically mapped and adjusted based on the current system supply and demand tension index: when the heating system faces supply and demand tension, the weight of the economic cost objective is automatically reduced, while the weights of the carbon emissions and equipment wear and tear objectives are increased. The upper and lower limits of unit output and the ramp rate constraints in the multi-objective optimization model are time-varying parameters, which are dynamically updated by the dynamic characteristic model according to the real-time operating conditions. The solution process of the multi-objective optimization model is as follows: First, the nonlinear characteristic curve and the heat loss function of the pipeline network in the model are piecewise linearly approximated and transformed into a MILP problem. The Pareto front solution set is obtained by using the branch and bound method combined with the interior point method. Finally, the Pareto compromise solution with the largest comprehensive membership degree is selected based on the fuzzy membership degree evaluation.

[0016] Preferably, in step S7, the closed-loop correction includes: Define the cumulative prediction deviation within the rolling time window. When the cumulative deviation is greater than a fixed percentage of the rated output, incremental learning is triggered. Incremental learning adopts a strategy of constructing an experience playback buffer that is sensitive to operating conditions. The sampling probability is positively correlated with the historical prediction error and whether the operating condition is variable. Meanwhile, based on the measured pressure and flow data of the SCADA system, Kalman filtering is used to perform posterior state estimation correction on the pipeline network model, and the resistance characteristic coefficients of the pipeline network segments are updated online using the corrected state parameters.

[0017] The second aspect of this invention discloses a dynamic modeling and capacity optimization analysis system for the operating characteristics of a district heating unit, wherein the system employs the method described in any one of the first aspects, and the system comprises: The data acquisition layer is used to deploy sensor networks and external data interfaces to collect unit and environmental data. The data preprocessing layer is used to perform both physical and statistical anomaly removal and spatiotemporal feature engineering. The dynamic characteristic model layer is used to construct and update the dynamic characteristic model of the heating unit online using preprocessed data. The dynamic characteristic model includes an energy efficiency characteristic model, an operating condition response model, and a loss and attenuation model. The multi-dimensional capability profiling layer is used to transform the model output into a multi-dimensional capability radar chart and feature indicators. The "machine-network-load" collaborative simulation layer is used to perform iterative solutions to the hydraulic and thermal states of the pipeline network, including unit boundary mapping. A multi-objective optimization decision layer is used to generate an optimized runtime scheduling strategy with adaptive weights. The visualization and command delivery layer is used for strategy comparison, human-computer interaction control, and triggering closed-loop correction.

[0018] The beneficial effects of this invention are as follows: 1. Accurate and dynamic model (dual-driven by data and physics): Instead of using pure "black box" machine learning, it innovatively embeds the Carnot theorem boundary and energy conservation residuals into XGBoost and LSTM networks. This not only has the nonlinear expressive power of deep learning, but also does not violate the limits of thermodynamic physics, which greatly improves the modeling accuracy and physical interpretability under varying operating conditions and equipment aging conditions.

[0019] 2. Systematized Capability Assessment: A capability assessment indicator system covering five dimensions—steady-state, dynamic, economic, safety, and collaboration—has been constructed, providing comprehensive quantitative basis for unit role positioning and function activation.

[0020] 3. Systematized Analysis Perspective and Innovative Mapping: A collaborative simulation of "machine-network-load" was established, and the capability profile was directly "translated" into the boundary mathematical constraints of the pipeline network solution (such as mapping the maximum ramp rate to the upper limit of the time change rate of node flow). This solved the problem of hydraulic imbalance in the pipeline network caused by local heat source optimization and achieved true global optimization.

[0021] 4. Intelligent decision support and closed-loop correction: The weights of multi-objective optimization are no longer rigid set values, but adaptively fluctuate according to outdoor weather and supply and demand tension; combined with the incremental learning mechanism of experience playback, the system has the self-evolution ability of "becoming more accurate with use", which greatly improves the level of refined management of heating. Attached Figure Description

[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to an embodiment of the present invention; Figure 2 This is an overall flowchart of a method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to an embodiment of the present invention.

[0024] Figure 3 This is the overall architecture diagram of the regional heating unit operation characteristic dynamic modeling and capacity optimization analysis system of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0026] The first aspect of this invention discloses a method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit.

[0027] Example 1:

[0028] Figure 1 This is a flowchart of a method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes: Step S1: Real-time acquisition of operating parameters, environmental parameters, and external data of related systems of the heating unit; Step S2: Perform data preprocessing on the collected data, including outlier removal, missing value imputation, and feature engineering; In step S2, the outlier removal adopts a dual judgment combining physical rules and statistical rules. The physical rules include a thermodynamic consistency judgment based on the minimum / maximum allowable supply and return water temperature difference, and a sensor misalignment judgment based on load-flow correlation. In the feature engineering, when extracting time-series features from the operating condition response model, an adaptive time-step sequence construction method is used, with the input time step L. t The length of the time step Lt is dynamically adjusted according to the thermal inertia characteristics of the unit and the measured outdoor temperature. The lower the outdoor temperature, the longer the time step Lt.

[0029] Step S3: Using the preprocessed data, construct and update the dynamic characteristic model of the heating unit online. The dynamic characteristic model includes an energy efficiency characteristic model, an operating condition response model, and a loss and attenuation model. Thermodynamic physical boundaries and mechanism features are introduced into the model construction for correction. In step S3, the energy efficiency characteristic model adopts the XGBoost algorithm, and its output has a physical boundary constraint layer. The physical boundary constraint layer calculates the theoretical upper limit efficiency under the current cold and heat source temperature conditions based on the Carnot theorem, and forces the original XGBoost prediction values ​​to be projected and constrained within the physical feasible region determined by the Carnot upper limit multiplied by the technical reachability coefficient. When the number of times the original model prediction values ​​exceed the physical boundary reaches a threshold, the bias correction amount is calculated using the exponential weighted moving average, triggering the online recalibration of the model.

[0030] The operating condition response model adopts an LSTM network. The training loss function of the LSTM network not only includes the prediction error term, but also adds a physical information regularization loss. The physical information regularization loss includes an energy conservation residual term constructed based on the unit's metal / working fluid heat storage change rate and the input-output heat balance equation, as well as a ramp physical constraint term calculated based on the turbine thermal stress model. The LSTM network introduces a working condition-aware attention mechanism, which incorporates a working condition change sensitivity factor that is proportional to the rate of change of control commands when calculating attention weights, thereby actively amplifying the model attention weights during varying working conditions such as load increases and decreases.

[0031] The loss and attenuation model uses the random forest algorithm to quantify the deviation of equipment performance degradation; The performance degradation deviation is calculated using a thermodynamic back-calculation method based on multi-sensor data fusion: real-time heat supply is calculated based on the supply and return water temperature difference and flow rate, and real-time input heat is calculated based on fuel flow rate and lower heating value. The ratio of the two is the real-time actual efficiency, and then the performance deviation from the design efficiency is obtained. When constructing decision trees, the random forest adopts a prior splitting strategy based on thermodynamic failure mechanisms, assigning priority to physical cumulative features, including low-cycle thermal fatigue, high-temperature creep, and oxygen corrosion, in the candidate feature pool; and adopts an incremental pruning strategy, which, as the data time span increases, performs depth-limited pruning on deep branches trained on old data and reduces the voting weight of the tree.

[0032] Step S4: Based on the output of the dynamic characteristic model, calculate and generate a multi-dimensional dynamic capability profile of the unit in real time; In step S4, the multi-dimensional dynamic capability profile encompasses an indicator system covering five dimensions: steady-state supply capability, dynamic adjustment capability, economic efficiency capability, reliability and security capability, and collaborative and mutual support capability.

[0033] Step S5: Integrate the hydraulic and thermal model of the heating network with the load prediction model, and transform the multi-dimensional dynamic capacity profile into time-varying boundary constraints of the network model to perform "machine-network-load" collaborative simulation analysis. In step S5, the specific mapping logic for converting the dynamic capability profile into time-varying boundary constraints of the pipeline network model includes: The minimum / maximum thermal output of the unit under the current operating conditions is mapped to the steady-state supply capacity constraint of the heat source node flow boundary; The maximum load increase / decrease rate of the unit extracted from the operating condition response model is mapped to the dynamic adjustment capability constraint of the upper limit of the time change rate of the flow of the heat source node. When the failure risk index output by the loss and attenuation model exceeds the warning threshold, the maximum allowable output constraint of the unit is dynamically reduced according to the safety derating factor.

[0034] Step S6: Based on the co-simulation results and the system supply and demand status, construct and solve a multi-objective optimization model, generate an optimized operation strategy, and issue it for execution; In step S6, the multi-objective optimization model comprehensively considers three objectives: economic cost, carbon emissions, and equipment wear and tear. The weight coefficients of each objective are not fixed constants, but are adaptively and dynamically mapped and adjusted based on the current system supply and demand tension index: when the heating system faces supply and demand tension, the weight of the economic cost objective is automatically reduced, while the weights of the carbon emissions and equipment wear and tear objectives are increased. The upper and lower limits of unit output and the ramp rate constraints in the multi-objective optimization model are time-varying parameters, which are dynamically updated by the dynamic characteristic model according to the real-time operating conditions. The solution process of the multi-objective optimization model is as follows: First, the nonlinear characteristic curve and the heat loss function of the pipeline network in the model are piecewise linearly approximated and transformed into a MILP problem. The Pareto front solution set is obtained by using the branch and bound method combined with the interior point method. Finally, the Pareto compromise solution with the largest comprehensive membership degree is selected based on the fuzzy membership degree evaluation.

[0035] Step S7: Continuously monitor the prediction deviation after the strategy is executed. When the deviation exceeds the threshold, trigger the incremental learning of the model to achieve online closed-loop correction.

[0036] In step S7, the closed-loop correction includes: Define the cumulative prediction deviation within the rolling time window. When the cumulative deviation is greater than a fixed percentage of the rated output, incremental learning is triggered. Incremental learning adopts a strategy of constructing an experience playback buffer that is sensitive to operating conditions. The sampling probability is positively correlated with the historical prediction error and whether the operating condition is variable. Meanwhile, based on the measured pressure and flow data of the SCADA system, Kalman filtering is used to perform posterior state estimation correction on the pipeline network model, and the resistance characteristic coefficients of the pipeline network segments are updated online using the corrected state parameters.

[0037] Example 2:

[0038] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, such as... Figure 2 As shown, Figure 2 This demonstrates the complete closed-loop algorithm logic, from step S1 (network deployment), step S2 (data processing), to step S4 (capability profiling), then step S5 (collaborative simulation), step S6 (multi-objective optimization decision-making), and finally step S7 (closed-loop self-correction mechanism (execution effect monitoring and triggering incremental learning)). This embodiment is applied to a city's district heating system, which includes multiple thermal power units, gas-fired peak-shaving boilers, and thermal storage tanks.

[0039] Step S1: Data Acquisition. Real-time acquisition of operating parameters, environmental parameters, and external data from related systems for the heating units. Specifically, Pt100 platinum resistance temperature sensors, pressure transmitters, electromagnetic flow meters, and mass flow meters are deployed at key nodes of each unit, with a sampling frequency of 1Hz. Simultaneously, outdoor temperature and peak / valley electricity prices are obtained through data bus connections to the meteorological station and power grid interfaces.

[0040] Step S2: Data preprocessing: Perform data preprocessing on the collected data, including outlier removal, missing value imputation, and feature engineering; Specifically, a dual anomaly detection mechanism is implemented. For the acquired data vector, not only is conventional statistical extremum removal performed, but a thermodynamic consistency criterion is also introduced: when the pipeline flow rate G... At 0°C, the supply and return water temperature difference ΔT must meet the requirement of ΔT. min ≤ΔT≤ΔT max (The lower limit is used to prevent sensor zero drift), and any violations will result in rejection. For rejected data, linear interpolation or similar operating condition matching interpolation will be used depending on the time period.

[0041] Specifically, in constructing the time series sequence for the LSTM model, this embodiment uses an adaptive time step size L. t : The physical meaning is: outdoor temperature T amb,t At lower temperatures, the thermal inertia of the generating units and pipelines is large, resulting in a slow system response and requiring longer historical sequences to capture dynamic characteristics; conversely, shorter sequences are sufficient, which significantly optimizes the computing power and prediction accuracy of LSTM.

[0042] The specific methods for outlier removal are as follows: (1) Outlier removal method based on physical constraints A dual anomaly detection mechanism combining physical and statistical rules is employed. For any sampling time... The collected raw data vector The following decision logic will be executed: Rule 1: Determination of Thermodynamic Consistency For heating units, when the flow rate At that time, the temperature difference between supply and return water Must meet:

[0043] in: Minimum allowable temperature difference, with a value of (A value below this is considered a sensor zero-drift fault). Maximum permissible temperature difference, taken as the unit's design rated temperature difference. ; like If the value exceeds the above range, then the data vector at that moment... Data marked as physical violations will be removed.

[0044] Rule 2: Determining the correlation between load and flow Calculate the estimated thermal output at the current moment. Compared with the measured flow rate The ratio:

[0045] in: Estimated thermal output derived from fuel quantity ; The specific heat capacity of water at constant pressure is taken as... ; Rated thermal efficiency of the unit (constant); like If the deviation exceeds 30%, it is determined that the sensor is inaccurate and the data point is discarded.

[0046] (2) Categorization and filling strategy For data that has been removed or is originally missing, the time span of the missing data should be considered. A differentiated filling strategy is adopted, as shown in Table 1: Table 1

[0047] (3) Timing window construction method for LSTM operating condition response model The key technical feature of this invention is: the input time step of the LSTM. It is not a fixed value, but rather an adaptive adjustment based on the thermal inertia characteristics of the unit.

[0048] Input sequence construction formula: for The task of predicting time steps involves constructing the input sequence. :

[0049] Among them, time step Determined by the following adaptive function:

[0050] Symbol explanation: : Reference time step, value (Corresponding to 60 minutes); Thermal inertia sensitivity coefficient, range of values The calibration is based on the type of generating unit; Outdoor temperature under design conditions (e.g.) ); : Continuously measure outdoor temperature; Reference temperature (e.g.) ); : Rounding function.

[0051] Physical meaning explanation: When the outdoor temperature At lower temperatures, the thermal inertia of the unit and piping network increases, and the system's response to external changes slows down, thus requiring a longer historical sequence. In order to accurately capture its dynamic characteristics, the thermal inertia is small (increased) during the early and late cold periods, and short sequences can meet the accuracy requirements.

[0052] Input feature vector Composition:

[0053] Symbol explanation: Unit control command signal (load setpoint, unit: MW); Current actual heat output (unit: MW); Outdoor ambient temperature (unit: ); Water supply pressure (unit: MPa); Return water temperature (unit: ); The pressure difference between the supply and return water in the pipeline network (unit: MPa) reflects the changes in the resistance characteristics of the pipeline network.

[0054] Output mode: This invention employs a rolling multi-step prediction mode, predicting and outputting the future in a single step. Thermal output trajectory at each time step:

[0055] Among the prediction steps (Corresponding to the next 15 minutes), time resolution .

[0056] Furthermore, the LSTM network of this invention differs from the standard textbook structure by introducing a Physics-InformedGate mechanism, which embeds prior thermodynamic knowledge into the network structure.

[0057] The network hierarchy is shown in Table 2: Table 2

[0058] Design of the physical correction layer (key technical features): Traditional LSTM output This may violate energy conservation or the physical limits of the unit. This invention adds a physical correction layer before the output layer to perform the following constraint projection: ; Symbol explanation: : Raw predictions from the LSTM network; : Limiting function, Constraints in interval Inside; : The minimum physical output at the current ambient temperature (determined by the unit characteristic curve); : Maximum physical output at the current ambient temperature (determined by the unit characteristic curve).

[0059] Step S3: Dynamic Characteristic Model Construction and Physical Fusion: Using the preprocessed data, a dynamic characteristic model of the heating unit is constructed and updated online. The dynamic characteristic model includes an energy efficiency characteristic model, an operating condition response model, and a loss and attenuation model. Thermodynamic physical boundaries and mechanism features are introduced into the model construction for correction. This is the core micro-innovation of the invention: integrating thermodynamic principles into AI algorithms. (1) Energy Efficiency Characteristic Model (XGBoost): Predicts thermal efficiency under given load and environment. A physical boundary constraint layer is provided at the output. The theoretical upper limit is calculated using Carnot's theorem: .

[0060] XGBoost raw output Constrained Inside (C) tech (This is the technically attainable coefficient). If the parameters continuously exceed the limit, the deviation is calculated using an exponentially weighted moving average, triggering online model recalibration.

[0061] (2) Load Response Model (LSTM): Predicts the rate of load increase / decrease of the unit. Standard LSTM is based on pure error training. This invention adds a Physical Information Regularized Loss: the total loss function not only includes the prediction error, but also adds a penalty term for the energy conservation residual (including the rate of change of the unit's metal wall and boiler water heat storage) and the thermal stress ramp-up limit. In addition, a load-aware attention mechanism is added before the LSTM output layer. This mechanism introduces a load-sensitive factor that is proportional to the rate of change of the load command, and actively amplifies the attention weight when the load increases or decreases significantly, focusing on the characteristics of the changing load.

[0062] (3) Loss and Degradation Model (Random Forest): Quantifying Equipment Aging Deviation. Real-time calculation of actual efficiency deviation Δη based on a thermodynamic inverse algorithm. t When constructing a random forest, the feature pool splitting priority is restricted, prioritizing splits based on physical mechanisms such as low-cycle thermal fatigue (number of start-stop cycles) and high-temperature creep (equivalent running hours parameter). Combined with an incremental pruning strategy, the branch depth of trees generated from old data is limited, and voting weights are reduced.

[0063] To clarify The computational logic employs a thermodynamic inverse calculation method based on multi-sensor data fusion, as detailed below: Step 1: Calculate real-time heating supply

[0064]

[0065] Symbol explanation: Specific heat capacity of water at constant pressure, value ; The density of water, with various values. ; The volumetric flow rate of the primary water supply network is collected by an electromagnetic flow meter (unit: ); Water supply temperature, collected by a Pt100 platinum resistance temperature sensor (unit: ); Return water temperature, acquired by a Pt100 platinum resistance temperature sensor (unit: ); Unit conversion factor (kW→MW); Step 2: Calculate real-time input heat

[0066]

[0067] Symbol explanation: Instantaneous fuel consumption is collected by a mass flow meter; Coal-fired power units: unit -Gas turbine unit: unit ; : Lower heating value of fuel (laboratory test record value); Standard coal: ; natural gas: ; Time unit conversion (h→s) to make the output unit MW; Step 3: Calculate the real-time actual thermal efficiency

[0068]

[0069] Step 4: Calculate performance deviation

[0070]

[0071] Symbol explanation: Rated thermal efficiency (constant) under unit design operating conditions ); The real-time actual thermal efficiency calculated by the above steps; : Performance deviation value, a positive value indicates performance degradation, and a negative value indicates operation beyond expectations.

[0072] The list of data source sensor configurations is shown in Table 3: Table 3

[0073] Step S4: Generate a dynamic capability profile of the unit: Based on the output of the dynamic characteristic model, calculate and generate a multi-dimensional dynamic capability profile of the unit in real time; Based on the output of the three models, multi-dimensional indicators are extracted in real time: (1) Steady-state supply capacity: rated output and time-varying adjustable output upper and lower limits.

[0074] (2) Dynamic adjustment capability: the real-time maximum ramp rate constrained by thermal stress and heat storage.

[0075] (3) Economic / security / cooperation capabilities, etc.

[0076] The connection logic between the "dynamic capability profile" and the pipeline model is the core connection mechanism that distinguishes this invention from existing technologies. The unit capacity profile is transformed into time-varying boundary constraints for solving the pipeline network model. The specific method is as follows: Linkage Mechanism 1: Transmission of Steady-State Supply Capacity Constraints unit steady-state supply capacity Directly mapped to the flow boundary of the heat source node:

[0077]

[0078] Symbol explanation: :unit The minimum / maximum thermal output (unit: MW) under the current operating conditions is output in real time by the energy efficiency characteristic model; Water supply temperature setpoint (unit: ); Design value for return water temperature (unit: ).

[0079] Linkage Mechanism 2: Transmission of Dynamic Adjustment Capability Constraints unit Maximum load increase / decrease rate and (Unit: MW / min), extracted from the LSTM operating condition response model and transformed into a time-varying rate constraint for the flow rate at the heat source node:

[0080] In discrete-time simulations, this constraint is expressed as:

[0081] in: ; ; Simulation time step (unit: seconds), value .

[0082] Related Mechanism 3: Transmission of Reliability and Security Capability Constraints When the output unit Failure Risk Index When the threshold (e.g., 0.7) is reached, the system automatically adjusts its maximum output limit to:

[0083] Symbol explanation: Failure risk index, output by the loss and attenuation model, with a value range of... ; Risk warning threshold, value ; Safety derating factor, value ; : The maximum allowable output after considering safety constraints.

[0084] Step S5 "Machine-Network-Load" Co-simulation: Integrate the hydraulic and thermal model of the heating network with the load prediction model, transform the multi-dimensional dynamic capacity profile into time-varying boundary constraints of the network model, and perform "machine-network-load" co-simulation analysis; A steady-state hydraulic-quasi-dynamic thermodynamic model of the pipe network is constructed using graph theory and Kirchhoff's laws. The innovation lies in the correlation mapping mechanism between capacity and boundary conditions. The dynamic capabilities of the generating units are mapped to the boundary conditions of the network equations. For example, the real-time minimum / maximum steady-state thermal output of unit ii is converted into nodal flow constraints. and The real-time maximum load increase rate extracted by LSTM is converted into a node flow time change rate constraint. In the Newton-Raphson method for solving the hydraulic balance of the entire network in an iterative process, the above physical and dynamic constraints are forcibly satisfied.

[0085] The specific methods for collaborative simulation analysis of "machine-network-load" are as follows: The pipeline network model used is a steady-state hydraulic-quasi-dynamic thermodynamic coupling model based on graph theory and Kirchhoff's laws, which is an "electric-thermal analogy".

[0086] (1) Graph representation of the pipeline network Abstracting the heating network as a directed graph ,in: : A set of nodes, including heat source nodes Load nodes Intermediate nodes ; Pipe segment collection, each pipe segment Indicates from node Flow to Node ; Define the following matrix: Node-pipe segment association matrix; : Basic loop matrix; (2) Node flow balance equation (Kirchhoff's first law) For any node in the pipeline network The algebraic sum of the inflows is zero:

[0087] Matrix form:

[0088] Symbol explanation: Pipeline segment mass flow vector (unit: ); : No. The mass flow rate of the pipe segment; a positive value indicates that the flow direction is consistent with the reference direction. Inflow node A collection of pipe segments; Outflow node A collection of pipe segments; :node Net outflow mass flow rate (unit: ); Heat source node: (Injecting water into the pipeline network); Load nodes: (Water is drawn from the pipeline network); Intermediate nodes: .

[0089] (3) Loop pressure balance equation (Kirchhoff's second law) For any independent loop in the pipeline network, the algebraic sum of the pressure drops is zero:

[0090] Detailed breakdown of each item: a) Frictional resistance pressure drop

[0091] For the pipe section The frictional resistance pressure drop is calculated using the Darcy-Weisbach formula:

[0092] Symbol explanation: Friction coefficient, calculated using the Colbrook formula:

[0093] Pipe section length (unit: m); Pipe section inner diameter (unit: m); Density of water (unit: ); Flow velocity inside the pipe (unit: m / s); Mass flow rate (unit: kg / s); Pipeline section resistance characteristic coefficient (unit: ), Calculation formula:

[0094] Absolute roughness of pipe wall (unit: m).

[0095] Reynolds number, ,in This refers to dynamic viscosity.

[0096] b) Potential energy difference pressure drop

[0097]

[0098] Symbol explanation: Gravitational acceleration, values .

[0099] Elevation of the end point of the pipe section (unit: m).

[0100] : Elevation of the starting point of the pipe section (unit: m).

[0101] c) Pump head

[0102] For a pipe section equipped with a water pump, its head characteristic is fitted by a quadratic curve:

[0103] Symbol explanation: Pump speed ratio (actual speed / rated speed), value range ; : Pump characteristic fitting coefficient (obtained from factory test data).

[0104] (4) Thermodynamic model – nodal temperature mixing equation Hot water mixes at the nodes of the pipe network, according to the law of conservation of energy:

[0105] Symbol explanation: Pipe section Terminal outlet water temperature (unit: ); :node The temperature after mixing (unit: ).

[0106] (5) Thermal model – pipe section temperature drop equation Considering heat dissipation from the pipe wall to the soil, the formula for calculating the temperature at the end of the pipe section is as follows:

[0107] Symbol explanation: Pipe section Terminal water temperature (unit: ); Pipe section Starting water temperature (unit: ); Soil temperature at the depth of pipeline burial (unit: ), estimated based on a seasonal lag model of outdoor temperature; Heat transfer coefficient per unit length of pipe section (unit: The thickness of the insulation layer is determined by its material and thickness.

[0108] Specific heat capacity of water at constant pressure (unit: ).

[0109] Step S6 Multi-objective optimization decision: Based on the co-simulation results and the system supply and demand status, construct and solve a multi-objective optimization model, generate an optimized operation strategy, and issue it for execution; Step S6 specifically includes: Step S61, multi-objective optimization decision-making, and step S62, generation of optimization operation strategy, visualization display, and instruction issuance, are detailed as follows: A comprehensive optimization MILP model is established, encompassing economic costs, carbon emissions, and equipment depreciation. Unlike conventional algorithms, this embodiment adaptively adjusts the weights based on supply and demand tension.

[0110] Defining supply and demand tension When S(t)→1 (heating is tight, such as in extremely cold weather), the system automatically and significantly reduces the weight of economic efficiency and increases the weight of equipment loss (safety) and energy efficiency, prioritizing supply; when S(t) is small, it tilts towards economic efficiency.

[0111] The nonlinear efficiency curve and the heat loss of the pipeline network are approximated by piecewise linearization. The solution is obtained by using the branch and bound method and the original-dual interior point method. The Pareto front is output. Finally, a compromise scheduling scheme is selected based on fuzzy membership degree.

[0112] This invention establishes a three-objective weighted comprehensive optimization model, which is then transformed into a single-objective function using a linear weighting method:

[0113] Symbol explanation: : Comprehensive optimization of the objective function value; Economic cost target (unit: yuan); Carbon emission targets (unit: kgCO2); Equipment loss target (dimensionless). Weighting coefficients, satisfying The operators set the schedule according to their current scheduling preferences. Sub-objective 1: Economic cost The formula is as follows:

[0114] The detailed explanation of the symbols is shown in Table 4: Table 4

[0115] Sub-goal 2: Carbon emissions

[0116]

[0117] The detailed explanation of the symbols is shown in Table 5: Table 5

[0118] Sub-objective 3: Equipment loss

[0119]

[0120] The detailed explanation of the symbols is shown in Table 6: Table 6

[0121] Complete mathematical expression of constraints Constraint 1: Steady-state supply capacity constraint (time-varying upper and lower limits)

[0122] Symbol explanation: Current efficiency Minimum technical output (unit: MW); Current efficiency Maximum technical output (unit: MW); Unit start-up and shutdown status variables. ; in, and It is not a fixed constant, but a time-varying parameter output by the energy efficiency model in step (2), which reflects the impact of ambient temperature and performance degradation on the unit output.

[0123] Constraint 2: Dynamic Climb Rate Constraint

[0124] Symbol explanation: : The real-time maximum load increase rate (unit: MW / time period) extracted from the LSTM model in step (2); : The real-time maximum load reduction rate (unit: MW / period) extracted from the LSTM model in step (2); :unit During the period The feature vector of the running state; The gradeability is not a fixed value, but rather changes dynamically with the operating conditions (temperature, pressure, and operating time).

[0125] Constraint 3: Unit Start-up and Shutdown Logic Constraints

[0126]

[0127] Constraint 4: Minimum Start-Stop Time Constraint

[0128]

[0129] Symbol explanation: :unit Minimum continuous running time (number of time periods); :unit The minimum continuous downtime (number of time periods).

[0130] Constraint 5: Operational Constraints of Thermal Storage Devices The state transition equation for the thermal storage tank:

[0131] Symbol explanation: Time period Heat storage capacity of the final heat storage tank (unit: MWh); Thermal storage efficiency (heat charging), value ; Heat release efficiency, value ; Time period The thermal power (unit: MW); Time period Heat release power (unit: MW).

[0132] Thermal storage tank capacity constraints:

[0133] Charge and discharge heat power constraints:

[0134]

[0135] Symbol explanation: : Pattern indicator variable, 0 = heat charging mode, 1 = heat dissipation mode; Maximum charging power (unit: MW); Maximum heat dissipation power (unit: MW).

[0136] Constraint 6: Overall Network Thermal Power Balance Constraint

[0137] Symbol explanation: Total number of heat load nodes; : No. Each load node during the time period The heat demand (in MW) is provided by the load forecasting model; Pipeline heat loss (unit: MW), calculated in real time by the pipeline thermal model; It's not a fixed ratio, but rather varies with the flow rate. A dynamically changing nonlinear function.

[0138] Constraint 7: Hydraulic safety constraints of the pipeline network The pressure at each node in the entire network is ensured by the output of the network hydraulic model:

[0139] Symbol explanation: :node During the period Pressure (unit: MPa); Minimum allowable pressure to prevent vaporization (generally) (gauge pressure); Maximum permissible pressure, determined by the pressure rating of the pipeline and equipment.

[0140] The optimization problem described above belongs to Mixed-Integer Nonlinear Programming (MINLP), and its characteristics are: Includes discrete variables wait; Includes continuous variables wait; Constraints containing nonlinear terms and .

[0141] The step-by-step solution strategy adopted in this invention: Step 1: Piecewise linearization of the characteristic curve The nonlinear efficiency curve output by the energy efficiency model Perform piecewise linear approximation:

[0142]

[0143]

[0144] in For the number of segments, These are the coordinates of the segmented breakpoints.

[0145] Similarly, for the heat loss function of the pipeline network Perform linearization to transform the original MINLP problem into a Mixed-Integer Linear Programming (MILP).

[0146] Step 2: Solve using the branch and bound method For the transformed MILP problem, solve it using the branch and bound method (Branch and Bound): Algorithm: Framework for solving using the branch and bound method; Input: Linearized MILP model; Output: Optimal decision variable values.

[0147] 1. Solve the linear programming relaxation (LP Relaxation) of the original problem to obtain the lower bound LB; 2. If the LP solution is an integer feasible solution, obtain the optimal solution of the original problem and terminate; 3. Otherwise, select a non-integer variable \(x_j\) for branching: Create sub-problem 1: \(x_j\leq floor(x_j)\); Create sub-problem 2: \(x_j\geq ceil(x_j)\).

[0148] 4. Recursively solve each sub-problem and update the global upper bound UB and lower bound LB; 5. Pruning rules: If the LP objective value of the sub-problem \(\geq\) UB, prune; If the sub-problem is infeasible, prune; If the sub-problem obtains an integer solution and the objective value < UB, update UB; 6. Repeat until all nodes are explored or the convergence tolerance is reached.

[0149] Step 3: Solve the continuous sub-problem using the interior point method When solving the LP relaxation at each branch node, use the Primal-Dual Interior Point Method:

[0150] where is the barrier parameter, which gradually decreases to 0 with iteration.

[0151] Step 4: Receding horizon optimization To cope with the uncertainty of load forecasting, adopt the Receding Horizon Optimization strategy: Optimization window: (96 time periods); Execution window: (4 time periods); The status is re-collected every hour, the prediction is updated, and the solution is re-optimized.

[0152] Based on the above, the present invention is able to: 1) The operating condition response model uses an LSTM network with an adaptive time step, and its input sequence length is... With outdoor temperature Dynamic adjustments are made to meet the prediction accuracy requirements under different thermal inertia conditions.

[0153] 2) Performance deviation The calculation employs a thermodynamic inverse method based on multi-sensor data fusion, specifically calculating the heating capacity based on the supply and return water temperature difference and flow rate. Calculate the input heat based on fuel flow rate and lower heating value. The ratio of the two is the actual real-time efficiency. .

[0154] 3) The "machine-grid-load" collaborative simulation transforms the dynamic regulation capability of the generating unit into a time-rate-of-change constraint for the network model, specifically by converting the maximum load increase rate... Mapped to the upper limit of the rate of change of flow at the heat source node This constraint is then enforced during the hydraulic solution iteration.

[0155] 4) Upper and lower limits of unit output constraints in multi-objective optimization models and gradeability constraint All parameters are time-varying and are updated in real time by the dynamic characteristic model based on real-time operating conditions (ambient temperature, performance degradation, pipeline back pressure), rather than using fixed nameplate values.

[0156] Step S7 closed-loop self-correction mechanism: continuously monitor the prediction deviation after the strategy is executed, and trigger incremental learning of the model to achieve online closed-loop correction when the deviation exceeds the threshold.

[0157] Optimize the issuance of scheduling instructions to the DCS system. The system continuously accumulates monitoring and prediction deviations; when the accumulated deviation within the rolling time window... At that time, the model is triggered to update incrementally online. Incremental learning uses weighted sampling with a forgetting factor (increasing the sampling weight of samples with inaccurate predictions under varying operating conditions) to achieve experience playback. The pipeline resistance coefficient is also corrected by feedback through Kalman filtering based on SCADA measured data.

[0158] The specific methods of the closed-loop feedback and online self-correction mechanism are as follows: 1) Incremental model learning triggered by prediction bias After the instruction is issued in step S7, the system continuously monitors the execution effect of the instruction. When the prediction deviation exceeds the threshold, the online incremental update of the model is triggered.

[0159] Deviation monitoring indicators: Define a scrolling time window Cumulative prediction bias within:

[0160] Triggering conditions: when hour( Take the rated output of the unit This triggers incremental learning.

[0161] Incremental learning strategy – weighted sampling in the experience replay buffer: Construct a condition-sensitive experience replay buffer Store the most recent Samples. During incremental training, the sampling probability... It is proportional to the prediction bias of the sample and the specificity of the operating conditions:

[0162] Symbol explanation: :sample Historical prediction error : Variable operating condition sampling amplification factor, value

[0163] Variable operating condition indication function Incremental update rules: For the XGBoost model, incremental boosting with a forgetting factor is used:

[0164] Symbol explanation: Model predictions before update New regression tree trained on new samples Forgetting factor, values Controlling the speed of integration of new and old knowledge 2) State feedback correction of pipeline network model After the instruction is issued in step S7, the actual response of the pipeline network hydraulic-thermal model may deviate from the simulation prediction. This invention establishes a state estimation feedback correction based on SCADA measured data: Correction equation:

[0165] Symbol explanation: The results given by co-simulation Time-prior state estimation (node ​​pressure, flow); Posterior state estimation after fusing SCADA measured data; SCADA system in The vector of measured values ​​at time points; The observation matrix maps the state space to the observation space. Kalman gain matrix, calculation formula:

[0166] Symbol explanation: Prior estimation error covariance matrix; The noise covariance matrix is ​​measured and calibrated based on sensor accuracy. Corrected model parameter updates: Corrected state Used to update the resistance characteristic coefficient of the pipeline network :

[0167] Symbol explanation: Pipe section before update Drag coefficient; Learning rate, values ; : SCADA measured voltage drop; Model calculation of pressure drop; Actual flow rate: Measured flow rate.

[0168] Implementation scenario effect: Using the above method, during a certain period when the outdoor temperature was -5℃, the system assessed that the economic efficiency range of Unit 1 was 60-90%, indicating high costs for peak-shaving boilers and sufficient thermal storage capacity. An adaptive optimization strategy was implemented: during off-peak hours at night, Unit 1 was prioritized to operate at high load and store thermal energy in the storage tanks; during peak days, the stored thermal energy was released to avoid frequent startups of the peak-shaving boilers. After implementation, benchmarking with instrument data showed that compared to traditional manual experience-based scheduling strategies, this system reduced total heating operating costs by 8.5%, carbon emissions by 4.2%, and effectively mitigated the risk of hydraulic imbalance at the far end of the pipeline network, achieving the intended purpose of the invention.

[0169] The second aspect of this invention discloses a system for dynamic modeling and capacity optimization analysis of the operating characteristics of district heating units. For example... Figure 3 As shown, Figure 3 The system demonstrates a bottom-up, six-layer logical hierarchy: a data acquisition layer, a data preprocessing layer, a dynamic characteristic model layer, a multi-dimensional capability profiling layer, a "machine-network-load" collaborative simulation layer, a multi-objective optimization decision-making layer, and a closed-loop feedback structure for visualization and command issuance. The system 100 includes: Data acquisition layer 101 is used to deploy sensor networks and external data interfaces to collect unit and environmental data; Data preprocessing layer 102 is used to perform physical and statistical dual anomaly removal and spatiotemporal feature engineering; The dynamic characteristic model layer 103 is used to construct and update the dynamic characteristic model of the heating unit online using the preprocessed data. The dynamic characteristic model includes an energy efficiency characteristic model, an operating condition response model, and a loss and attenuation model. Multidimensional capability profile layer 104 is used to transform the model output into a multidimensional capability radar chart and feature indicators; The "machine-network-load" collaborative simulation layer 105 is used to perform iterative solutions to the hydraulic and thermal states of the pipeline network, including unit boundary mapping. A multi-objective optimization decision layer 106 is used to generate an optimized runtime scheduling strategy with adaptive weights; The visualization and command issuance layer 107 is used for strategy comparison, human-computer interaction control, and triggering closed-loop correction.

[0170] Please note that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

[0171] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic modeling and capacity optimization analysis of the operating characteristics of district heating units, characterized in that, The method includes: Step S1: Real-time acquisition of operating parameters, environmental parameters, and external data of related systems of the heating unit; Step S2: Perform data preprocessing on the collected data, including outlier removal, missing value imputation, and feature engineering; Step S3: Using the preprocessed data, construct and update the dynamic characteristic model of the heating unit online. The dynamic characteristic model includes an energy efficiency characteristic model, an operating condition response model, and a loss and attenuation model. Thermodynamic physical boundaries and mechanism features are introduced into the model construction for correction. Step S4: Based on the output of the dynamic characteristic model, calculate and generate a multi-dimensional dynamic capability profile of the unit in real time; Step S5: Integrate the hydraulic and thermal model of the heating network with the load prediction model, and transform the multi-dimensional dynamic capacity profile into time-varying boundary constraints of the network model to perform "machine-network-load" collaborative simulation analysis. Step S6: Based on the co-simulation results and the system supply and demand status, construct and solve a multi-objective optimization model, generate an optimized operation strategy, and issue it for execution; Step S7: Continuously monitor the prediction deviation after the strategy is executed. When the deviation exceeds the threshold, trigger the incremental learning of the model to achieve online closed-loop correction.

2. The method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to claim 1, characterized in that, In step S2, the outlier removal adopts a dual judgment combining physical rules and statistical rules. The physical rules include a thermodynamic consistency judgment based on the minimum / maximum allowable supply and return water temperature difference, and a sensor inaccuracy judgment based on load-flow correlation. In the feature engineering, when extracting time-series features from the operating condition response model, an adaptive time-step sequence construction method is used, with the input time step L. t The length of the time step Lt is dynamically adjusted according to the thermal inertia characteristics of the unit and the measured outdoor temperature. The lower the outdoor temperature, the longer the time step Lt.

3. The method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to claim 1, characterized in that, In step S3, the energy efficiency characteristic model adopts the XGBoost algorithm, and its output has a physical boundary constraint layer. The physical boundary constraint layer calculates the theoretical upper limit efficiency under the current cold and heat source temperature conditions based on the Carnot theorem, and forces the original XGBoost prediction values ​​to be projected and constrained within the physical feasible region determined by the Carnot upper limit multiplied by the technical reachability coefficient. When the number of times the original model prediction values ​​exceed the physical boundary reaches a threshold, the bias correction amount is calculated using the exponential weighted moving average, triggering the online recalibration of the model.

4. The method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to claim 1, characterized in that, In step S3, the operating condition response model adopts an LSTM network. The training loss function of the LSTM network not only includes the prediction error term, but also adds a physical information regularization loss. The physical information regularization loss includes an energy conservation residual term constructed based on the unit metal / working fluid heat storage change rate and the input-output heat balance equation, as well as a ramp physical constraint term calculated based on the turbine thermal stress model. The LSTM network introduces a working condition-aware attention mechanism, which incorporates a working condition change sensitivity factor that is proportional to the rate of change of control commands when calculating attention weights, thereby actively amplifying the model attention weights during varying working conditions such as load increases and decreases.

5. The method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to claim 1, characterized in that, In step S3, the loss and attenuation model uses the random forest algorithm to quantify the deviation of equipment performance degradation. The performance degradation deviation is calculated using a thermodynamic back-calculation method based on multi-sensor data fusion: real-time heat supply is calculated based on the supply and return water temperature difference and flow rate, and real-time input heat is calculated based on fuel flow rate and lower heating value. The ratio of the two is the real-time actual efficiency, and then the performance deviation from the design efficiency is obtained. When constructing decision trees, the random forest adopts a prior splitting strategy based on thermodynamic failure mechanisms, assigning priority to physical cumulative features, including low-cycle thermal fatigue, high-temperature creep, and oxygen corrosion, in the candidate feature pool; and adopts an incremental pruning strategy, which, as the data time span increases, performs depth-limited pruning on deep branches trained on old data and reduces the voting weight of the tree.

6. The method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to claim 1, characterized in that, In step S4, the multi-dimensional dynamic capability profile encompasses an indicator system covering five dimensions: steady-state supply capability, dynamic adjustment capability, economic efficiency capability, reliability and security capability, and collaborative and mutual support capability.

7. The method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to claim 6, characterized in that, In step S5, the specific mapping logic for converting the dynamic capability profile into time-varying boundary constraints of the pipeline network model includes: The minimum / maximum thermal output of the unit under the current operating conditions is mapped to the steady-state supply capacity constraint of the heat source node flow boundary; The maximum load increase / decrease rate of the unit extracted from the operating condition response model is mapped to the dynamic adjustment capability constraint of the upper limit of the time change rate of the flow of the heat source node. When the failure risk index output by the loss and attenuation model exceeds the warning threshold, the maximum allowable output constraint of the unit is dynamically reduced according to the safety derating factor.

8. The method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to claim 1, characterized in that, In step S6, the multi-objective optimization model comprehensively considers three objectives: economic cost, carbon emissions, and equipment wear and tear. The weight coefficients of each objective are not fixed constants, but are adaptively and dynamically mapped and adjusted based on the current system supply and demand tension index: when the heating system faces supply and demand tension, the weight of the economic cost objective is automatically reduced, while the weights of the carbon emissions and equipment wear and tear objectives are increased. The upper and lower limits of unit output and the ramp rate constraints in the multi-objective optimization model are time-varying parameters, which are dynamically updated by the dynamic characteristic model according to the real-time operating conditions. The solution process of the multi-objective optimization model is as follows: First, the nonlinear characteristic curve and the heat loss function of the pipeline network in the model are piecewise linearly approximated and transformed into a MILP problem. The Pareto front solution set is obtained by using the branch and bound method combined with the interior point method. Finally, the Pareto compromise solution with the largest comprehensive membership degree is selected based on the fuzzy membership degree evaluation.

9. The method for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit according to claim 1, characterized in that, In step S7, the closed-loop correction includes: Define the cumulative prediction deviation within the rolling time window. When the cumulative deviation is greater than a fixed percentage of the rated output, incremental learning is triggered. Incremental learning adopts a strategy of constructing an experience playback buffer that is sensitive to operating conditions. The sampling probability is positively correlated with the historical prediction error and whether the operating condition is variable. Meanwhile, based on the measured pressure and flow data of the SCADA system, Kalman filtering is used to perform posterior state estimation correction on the pipeline network model, and the resistance characteristic coefficients of the pipeline network segments are updated online using the corrected state parameters.

10. A system for dynamic modeling and capacity optimization analysis of the operating characteristics of a district heating unit, characterized in that, The system includes: The data acquisition layer is used to deploy sensor networks and external data interfaces to collect unit and environmental data. The data preprocessing layer is used to perform both physical and statistical anomaly removal and spatiotemporal feature engineering. The dynamic characteristic model layer is used to construct and update the dynamic characteristic model of the heating unit online using preprocessed data. The dynamic characteristic model includes an energy efficiency characteristic model, an operating condition response model, and a loss and attenuation model. The multi-dimensional capability profiling layer is used to transform the model output into a multi-dimensional capability radar chart and feature indicators. The "machine-network-load" collaborative simulation layer is used to perform iterative solutions to the hydraulic and thermal states of the pipeline network, including unit boundary mapping; A multi-objective optimization decision layer is used to generate an optimized runtime scheduling strategy with adaptive weights. The visualization and command delivery layer is used for strategy comparison, human-computer interaction control, and triggering closed-loop correction.

Images