A method of cogeneration that fuses predictive maintenance with deep recycling

Abstract

This invention discloses a cogeneration method integrating predictive maintenance and deep waste heat recovery, belonging to the field of cogeneration technology. The method first collects real-time system operation, status monitoring, and external prediction data, and then identifies key performance parameters of the equipment online based on a physical model and adaptive estimation algorithm, generating a real-time health index (HI) and its predicted future degradation trajectory. The core lies in constructing a multi-objective real-time optimization controller, whose objective function integrates short-term operational economic benefits and long-term virtual health costs based on HI predictions, with constraints dynamically correlated with HI. The controller uses a model predictive control framework for rolling solution, outputting optimal coordinated control commands for generator output and deep waste heat recovery components such as flue gas bypass valves and variable frequency pumps. This invention achieves a deep integration of predictive maintenance and operational optimization, significantly improving the overall energy efficiency and economy of the system under all operating conditions while ensuring long-term equipment reliability.

Description

Technical Field

[0001] This invention relates to the field of cogeneration technology, and in particular to a cogeneration method that integrates predictive maintenance and deep recovery. Background Technology

[0002] Cogeneration (CHP) systems, capable of simultaneously providing electricity and useful heat, have a significantly higher overall primary energy utilization rate than traditional separate energy systems, leading to their widespread application in industrial, district energy supply, and commercial building sectors. However, in actual operation, CHP systems still face two major challenges: 1) Equipment reliability and maintenance costs: The core power equipment (such as internal combustion engines, gas turbines, and turbines) and critical thermal equipment (such as waste heat boilers and heat exchangers) of CHP systems operate under high temperature, high pressure, and variable load conditions for extended periods, leading to gradual performance degradation and potential sudden failures. Traditional maintenance strategies are mainly divided into two categories: one is periodic preventative maintenance, which involves maintenance or component replacement based on fixed intervals or operating times. Its disadvantage is the potential for "over-maintenance" (shutting down for repairs despite good equipment condition) or "under-maintenance" (developing hidden problems before the scheduled maintenance period). The other is threshold-based reactive maintenance, which intervenes when sensor readings (such as excessive temperature or vibration) trigger alarms. This usually indicates that a failure has already occurred or is about to occur, potentially leading to unplanned downtime and significant losses. Neither of these strategies can accurately predict the gradual degradation of equipment performance and the risk of sudden failures, resulting in uncertainty in system operation and high maintenance costs. 2) Energy efficiency optimization under partial load and variable operating conditions: The design and optimization of CHP systems are usually based on the rated operating point, but in actual operation, the demand for electricity and heat fluctuates frequently, and the system often operates under partial load or deviates from the design point. Under such conditions, traditional waste heat recovery strategies (such as a fixed flue gas bypass ratio and a constant heat exchanger water flow rate) are often not optimal, resulting in a large amount of medium and low grade waste heat not being fully recovered or having low recovery efficiency. Existing optimization research mostly focuses on structural improvements to the thermodynamic cycle (such as using a more complex Rankine cycle) or performance enhancements of individual components, or setpoint optimization based on static efficiency models, lacking a global energy recovery optimization mechanism that can respond in real time and dynamically to the overall system state (including equipment health status and external load demand).

[0003] Current technologies typically treat "maintenance" and "operation optimization" as two separate processes: the maintenance department plans shutdowns based on equipment status, while the operations department pursues maximum output or efficiency when equipment is available. This separation leads to two problems: First, operation optimization strategies may exacerbate wear and tear on critical equipment components for short-term efficiency (e.g., operating turbines at higher exhaust temperatures to increase short-term power generation), accelerating their degradation and increasing long-term failure risks and maintenance costs. Second, the equipment health status information obtained from predictive maintenance is only used to trigger maintenance work orders and is not fed back into daily real-time operation control variables. This fails to utilize this information to "adaptively" adjust operating modes to intelligently maintain optimal economic operation even when equipment performance has deteriorated, or proactively adopt milder operating modes to extend its remaining service life. Summary of the Invention

[0004] The purpose of this invention is to disclose a cogeneration method that integrates predictive maintenance and deep heat recovery. By dynamically optimizing the integration of equipment degradation prediction and multi-stage waste heat recovery, this method solves the problems of energy efficiency decline and sudden failures caused by component performance degradation in cogeneration systems.

[0005] The specific plan is as follows: A combined heat and power (CHP) method integrating predictive maintenance and deep recovery includes the following steps; Real-time acquisition of operating process parameters, status monitoring parameters and external prediction data of cogeneration systems; Construct physical models of key components and use real-time data to update model parameters online, generate real-time health indicators of the equipment, and predict its future trajectory. Using a comprehensive function that includes a virtual health cost model based on the prediction of future changes in health indicators and short-term operational economic benefits as the objective function, and the equipment operation boundary that is dynamically related to health indicators as the constraint, the optimal setpoint sequence of control variables for generator and waste heat recovery system is solved. The immediate instructions in the optimal setpoint sequence are sent to the underlying execution mechanism.

[0006] Furthermore, after solving for the optimal setpoint sequence of control variables for the generator and waste heat recovery system, and before issuing the immediate instructions in the optimal setpoint sequence to the underlying actuators, the method also includes a step of online correction and adjustment of the parameters and weights in the virtual health cost model based on the actual operation feedback of the cogeneration system.

[0007] Furthermore, the construction of the physical model of the key components includes constructing the physical model of the key components based on the first principles of thermodynamics and heat transfer; the online updating of model parameters using real-time data includes updating the time-varying performance parameters in the model online through parameter estimation algorithms to generate real-time health indicators characterizing the health status of the equipment; the prediction of its future trajectory includes adopting a multi-scale time-series prediction model, which integrates a long short-term memory network and an exponential smoothing state-space model. The long short-term memory network is responsible for capturing the nonlinear long-term dependence between health indicators and high-dimensional operating conditions, while the exponential smoothing state-space model is responsible for modeling the local trends and seasonal fluctuations of health indicators. Finally, a weighted fusion layer outputs a comprehensive prediction of the health indicator degradation trajectory.

[0008] Furthermore, the physical model includes a comprehensive physical model of turbocharger efficiency factors and a physical model of overall heat transfer coefficient attenuation in the cylinder liner water heat exchanger, and the parameter estimation algorithm is the extended Kalman filter algorithm. The formula for the physical model of the turbocharger efficiency comprehensive factor is: ; Where, η ＿turbo＿eff This represents the overall efficiency factor of the turbocharger; Indicates intake mass flow rate; c P，air T represents the specific heat capacity of air at constant pressure. air，in Indicates the compressor inlet air temperature; T air，out This indicates the compressor outlet air temperature; Indicates exhaust mass flow rate; c P，exh T represents the isobaric specific heat capacity of exhaust gas at average temperature. exh，in T represents the turbine inlet exhaust temperature; exh，out，isen This represents the outlet temperature of the turbine after isentropic expansion. The formula for the physical model of the overall heat transfer coefficient decay in a cylinder liner water heat exchanger is: ; U ＿fouling This indicates the overall heat transfer coefficient attenuation rate of the cylinder liner water heat exchanger. This indicates the mass flow rate of the cylinder liner water; T represents the isobaric specific heat capacity of cylinder liner water at average temperature; jw，out Indicates the outlet temperature of the cylinder liner water; T jw，in Indicates the inlet temperature of the cylinder liner water; A represents the heat exchange area; ΔT lm This represents the logarithmic mean temperature difference.

[0009] Furthermore, the turbocharger health index HI turbo The calculation formula is: HI turbo =η ＿turbo＿eff,current / η ＿turbo＿eff,baseline; Where η ＿turbo＿eff,current η is the optimal estimate of the turbocharger's overall efficiency factor, output in real time by the extended Kalman filter algorithm at the current moment; ＿turbo＿eff,baseline The nominal value of the overall efficiency factor of a turbocharger when it is in a healthy, new condition and operating under standard or typical conditions; Cylinder liner water heat exchanger health index HI jw The calculation formula is: HI jw =U current / U baseline ; Where η ＿turbo＿eff,current η is the optimal estimate of the overall heat transfer coefficient of the cylinder liner water heat exchanger, output in real time by the extended Kalman filter algorithm at the current moment; ＿turbo＿eff,baseline This refers to the design value or initial calibration value of the overall heat transfer coefficient of the cylinder liner water heat exchanger under clean and foul-free conditions.

[0010] Furthermore, the objective function is: ; Where R elec Represents electricity generation revenue; R heat For heating / cooling revenue; C fuel Indicates fuel cost; C om，var Indicates variable maintenance costs associated with operation; HI j，current This indicates the current real-time health indicator of component j; △HI j，pred (u) represents the predicted decrease in the health indicators of critical component j over a future period after strategy u is implemented; u represents the set of adjustable control commands, i.e., control variables; α j β represents the maintenance cost coefficient for component j; j λ represents the health status penalty factor; λ represents the weighting coefficient (hyperparameter) of health costs; j is the component index.

[0011] Furthermore, the control variable u=[P gen V bypass f pump1 f pump2 ..., SOC tes T set，cond ,...] T ; Where P gen V is the generator active power setpoint. bypass f is the opening degree of the flue gas bypass valve; pump1 f pump2 ... represents the frequency of the inverters for each major water pump; SOC tes The target state of charge of the thermal storage tank; T set，cond This is the setpoint for the condensation temperature.

[0012] Furthermore, the constraints include static constraints and dynamic constraints. The static constraints include: 1) upper and lower limits of generator power; 2) physical limits of equipment; 3) grid / heat network demand balance; and 4) safe operation constraints. The dynamic constraint is a function of the sub-component's real-time health indicator HI: ; where g i (u) represents the constrained value of the control variable in the sub-component; The upper limit of the design value for this control variable when the equipment is new and in good health (HI=1); Φ i (HI relevant ) is a dynamic scaling function related to relevant health indicators.

[0013] Furthermore, sequential quadratic programming and rolling time-domain solution are used to obtain the optimal setpoint sequence of control variables for the generator and waste heat recovery system. In each iteration, sequential quadratic programming constructs a quadratic programming subproblem to approximate the original problem, and updates the iteration point by solving the quadratic programming subproblem until convergence. To handle forecast uncertainty and implement feedback control, a Model Predictive Control (MPC) framework is adopted: Forecast time domain: the next hour, divided into four 15-minute intervals; Control time domain: typically equal to or shorter than the forecast time domain; In each control cycle, the optimizer solves for the optimal control sequence u within the next forecast time domain. (t), t=1, ...,N, but only the control command u for the first time period is implemented. (1); In the next cycle, based on the new measurements, predictions and optimizations are performed again, and so on, rolling forward. The MPC framework naturally incorporates forecasts of future load, electricity prices, and future health status (HI). pred .

[0014] Furthermore, based on feedback from the actual operation of the cogeneration system, the parameters and weights in the virtual health cost model are corrected and adjusted online, including parameter adjustments: Data recording: Record the control variable u corresponding to each control decision, and the actual observed ΔHI within a certain period after execution. j，observed change; Model update: Periodically update the accumulated {u k ΔHI j，observed The data pair is used to update the parameters of the virtual health cost model, where k is the index of the control variable u; And the adaptive adjustment of the weighting coefficient λ: Set a health target range: Set a healthy operating range for health indicators; Monitor health indicator trends: Calculate the moving average and trend of health indicators; If the moving average of the health indicator is below the lower limit of the operating range and the trend is one of continuous deterioration, then increase λ. If the moving average of the health indicator is higher than the upper limit of the operating range and the trend is continuously improving, then decrease λ.

[0015] Compared with the prior art, the present invention has at least one of the following technical effects: 1. In existing technologies, cogeneration maintenance departments schedule maintenance based on fixed cycles or simple thresholds, while the operation departments only pursue maximum output when equipment is available. There is a lack of information sharing and collaborative decision-making between the two departments. This often leads to "over-maintenance" or "under-maintenance," as well as short-sighted behavior that sacrifices equipment lifespan for short-term gains. This invention utilizes real-time sensor data and algorithms such as Kalman filtering to accurately estimate key performance parameters such as turbine efficiency and heat exchanger heat transfer coefficient online, quantifying them as a real-time health index (HI). More importantly, this invention deeply embeds HI into real-time operation optimization control: on the one hand, the safe operating boundaries of the equipment (such as maximum temperature and maximum load) are no longer fixed values, but rather functions dynamically bound to the current HI. When equipment health declines, the system automatically and gradually tightens operating limits, proactively reducing equipment stress and preventing failures. On the other hand, a virtual cost term based on health prediction is introduced into the optimization objective function, quantifying the "cost of wear and tear" of different operating strategies on the long-term lifespan of the equipment. This allows the optimization controller to automatically weigh short-term economic benefits against long-term equipment wear and tear at every decision, proactively selecting operating points that are more "friendly" to the equipment. Consequently, the system shifts from "repair after failure" or "regular over-maintenance" to "predictive adaptive operation based on health status," ensuring high reliability while precisely scheduling maintenance, extending overhaul intervals, and reducing total lifecycle maintenance costs. In short, this invention breaks down the decision-making barriers between maintenance and operation, achieving intelligent adaptive operation based on real-time equipment health, improving system reliability throughout its lifecycle, and reducing maintenance costs.

[0016] 2. Traditional cogeneration system optimization is often based on static models or rated operating conditions. In actual operation with frequent fluctuations in electricity and heat loads, waste heat recovery strategies (such as a fixed flue gas bypass ratio) are often not optimal, resulting in energy waste. This invention solves this problem through a health-state-driven multi-objective real-time optimization controller. This controller adopts a model predictive control (MPC) framework, considering not only the current moment but also the electricity price, load forecast, and equipment health evolution trends for the next few hours. Its decision variable u covers multiple dimensions such as generator output, flue gas bypass valve, variable frequency pump speed in each loop, and heat storage tank status, achieving refined and coordinated control of energy flow. Crucially, the optimization objective not only pursues traditional economic efficiency but also incorporates equipment health costs, giving the optimization decision a global and long-term perspective. For example, during periods of high electricity prices, even if equipment health is slightly low, the controller may moderately increase power generation priority within the limits of dynamic constraints. Conversely, during periods of low electricity prices or when equipment health is in warning, the controller may proactively reduce the power generation load, directing more waste heat from flue gas to absorption cooling or storing it in thermal storage tanks to prepare for subsequent peak demand, while allowing the equipment to "rest." This dynamic and intelligent deep energy recovery strategy ensures that the system can find its optimal energy efficiency operating point under any operating condition, realizing a technological shift from a "fixed strategy" to a "personalized dynamic strategy." This achieves global and dynamic energy efficiency optimization under complex and variable operating conditions, improving the system's overall energy utilization rate and economy. 3. Traditional optimization and maintenance system parameters often rely on experience-based settings, which, once set, remain fixed and cannot adapt to individual equipment differences, aging characteristics, and changes in the operating environment. This invention endows the system with the ability to continuously evolve: the system records every control decision... k Compared with the actual observed changes in equipment health ΔHI observed These empirical data pairs are used periodically to calibrate parameters (such as indices a, b, and coefficients K_j) in the virtual health cost model, making the model's predictions of "how operational activities affect health" increasingly accurate. Meanwhile, the core trade-off parameter—the health cost weight λ—is not a fixed value, but rather adjusts based on the overall equipment health index HI. overall The system dynamically adjusts based on the current value and trend of the data. When health deteriorates, λ is automatically increased, shifting the system to a conservative protection mode; when health is good, λ is decreased, allowing for more aggressive economic operation. Furthermore, models such as LSTM used for long-term health prediction are continuously trained and updated with the latest data. This complete closed loop of "perception-decision-execution-learning" enables the system to continuously self-optimize, increasingly conforming to the actual degradation patterns and operating environments of specific equipment. This achieves a transformation from a "static program" to a "growing intelligent agent," enhancing the universality, robustness, and long-term effectiveness of the solution. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a cogeneration method integrating predictive maintenance and deep recovery according to the present invention; Figure 2 This is an architectural diagram of a combined heat and power system that integrates predictive maintenance and deep recovery according to the present invention. Detailed Implementation

[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0020] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0021] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0023] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0025] Combined with appendix Figure 1 A combined heat and power (CHP) method integrating predictive maintenance and deep recovery is disclosed, comprising the following steps: Step 1: Acquire system operating parameters, status monitoring parameters, and external prediction data in real time.

[0026] This invention can be applied to any form of gas-fired combined heat and power (CHP) system. A typical natural gas internal combustion engine CHP system for district power supply is used as an example. This system mainly includes: a power generation unit: a natural gas internal combustion engine generator set, which outputs electrical energy to the grid or local load, while simultaneously generating high-temperature flue gas (~450°C) and cylinder liner heat (~90°C); a waste heat recovery unit: a high-temperature flue gas recovery chain: flue gas flows successively through a flue gas heat exchanger (producing high-temperature hot water or steam), a flue gas-type lithium bromide absorption chiller (cooling / heating), and finally is discharged through the chimney; a cylinder liner heat recovery chain: cylinder liner heat is used through a cylinder liner heat exchanger to preheat boiler feedwater or directly supply low-temperature hot water; optional deep recovery devices: such as organic Rankine cycle (ORC) generator sets, further utilizing the low-grade waste heat from flue gas or hot water to generate electricity; and auxiliary and control systems: various pumps, valves, frequency converters, sensors, and existing basic distributed control systems (DCS) or programmable logic controllers (PLCs).

[0027] The software system of this invention serves as the upper-level Advanced Process Control (APC) layer, deployed in an industrial server or edge computing gateway. It interacts with the underlying DCS / PLC via standard industrial communication protocols such as OPC UA and Modbus TCP, issues optimization setpoints, and receives real-time feedback data.

[0028] To achieve accurate state awareness, the parameters acquired in real time include: I. Monitoring of core power equipment (natural gas internal combustion engine): 1) Process parameters: intake manifold temperature / pressure, exhaust manifold temperature of each cylinder (infrared thermometry), turbocharger front and rear temperatures / pressure, intercooler temperature / pressure, cylinder liner water inlet / outlet temperature / pressure / flow rate, lubricating oil temperature / pressure / flow rate; 2) Condition monitoring parameters include: Vibration: Triaxial vibration acceleration sensors are installed in the engine block and turbocharger bearing housing, with a sampling frequency of not less than 10 kHz, to analyze mechanical imbalance, misalignment, and bearing wear.

[0029] Combustion analysis: Install an in-cylinder pressure sensor (at least one representative cylinder), or use a high-precision crankcase pressure sensor / vibration signal to indirectly analyze combustion pressure and assess combustion conditions, knocking, and misfire.

[0030] Online lubricating oil monitoring: Install online oil sensors to monitor the viscosity, moisture content, dielectric constant (reflecting oxidation and contamination), and metal abrasive particles (iron, copper, and aluminum content) of lubricating oil in real time.

[0031] II. Monitoring of Key Heat Recovery Equipment: 1) Flue gas heat exchanger: flue gas inlet / outlet temperature and pressure; working fluid water (or steam) inlet / outlet temperature, pressure, and flow rate; key point temperature of the tube wall (thermocouple).

[0032] 2) Absorption chiller / heater: inlet / outlet temperature / pressure / flow rate of the driving heat source (flue gas or hot water); inlet / outlet temperature / pressure / flow rate of chilled water / cooling water; solution concentration and temperature.

[0033] 3) Various water systems: inlet and outlet pressures of main pumps, motor current, frequency converter frequency; flow meters of key pipelines; multi-layer temperature sensors for thermal storage tanks.

[0034] III. External Forecast Data It mainly includes forecasts of future electricity prices, fuel prices, weather (temperature, humidity) and user heat / electricity load demand. Its function is to provide the controller with information on the future market and operating environment in order to formulate forward-looking and economically optimal scheduling and operation strategies.

[0035] Data Acquisition and Transmission: 1) Use high-precision, timestamped data acquisition cards or smart gateways to ensure time synchronization of multi-source data (error <10ms); 2) High-frequency data such as vibration and acoustic emission can be feature extracted at the edge (e.g., calculating RMS, peak value, kurtosis, and spectral characteristics) before uploading the feature values ​​to the central server to reduce bandwidth requirements; 3) Establish a unified historical real-time database (e.g., PI System, InfluxDB) to store all time-series data.

[0036] Step 2: Construct physical models of key components and use real-time data to update model parameters online, generate real-time health indicators for the equipment, and predict its future trajectory.

[0037] A physical model of key components is constructed based on the first principles of thermodynamics and heat transfer. Real-time data is used to update the time-varying performance parameters in the model online through parameter estimation algorithms to generate health indicators characterizing the health status of the equipment. At the same time, a multi-scale time-series prediction model is adopted, which integrates a long short-term memory network and an exponential smoothing state-space model. The long short-term memory network is responsible for capturing the nonlinear long-term dependence between health indicators and high-dimensional operating conditions, while the exponential smoothing state-space model is responsible for modeling the local trends and seasonal fluctuations of health indicators. Finally, a weighted fusion layer outputs a comprehensive prediction of the health indicator degradation trajectory.

[0038] Digital twins are the foundation for this invention to achieve accurate health status assessment and prediction. It is not a purely black-box, data-driven model, but rather a "grey-box" model that integrates prior physical knowledge and real-time data.

[0039] Taking natural gas internal combustion engines as an example, isentropic efficiency is a core indicator of turbocharger performance. However, directly measuring the isentropic efficiency of a turbine is very difficult. We choose two more easily observable parameters that are strongly correlated with efficiency as health proxy indicators: 1) Turbocharger efficiency composite factor (η) ＿turbo＿eff 1) Indirect calculation through measurable parameters; 2) Overall heat transfer coefficient decay rate (Ua) of cylinder liner water heat exchanger. ＿fouling ): Reflects the degree of fouling on the heat exchange surface.

[0040] Physical Model 1: The simplified physical model of the turbocharger system is constructed as follows: The theoretical power of a turbocharger comes from the exhaust energy of the engine. Based on the first law of thermodynamics and the ideal gas hypothesis, we establish the following relationship: 1) Estimation of available power of the turbine: The theoretically recoverable power P of a turbine tur，avail It is related to exhaust flow rate and enthalpy drop. A simplified formula is used: ; in, The exhaust mass flow rate (kg / s) can be estimated using fuel flow rate and air-fuel ratio. c P，exh This represents the isobaric specific heat capacity of the exhaust gas at the average temperature (kJ / (kg·K)), which can be taken as an empirical constant or looked up in a table; T exh，in This indicates the turbine inlet exhaust temperature (K), measured in K. T exh，out，isenThe outlet temperature (K) of the turbine after isentropic expansion is calculated using the following formula: ; Where, p exh，out This indicates the exhaust pressure (Pa) at the turbine outlet, measured in actual measurements. p exh，in This indicates the exhaust pressure (Pa) at the turbine inlet, as measured. γ exh This represents the specific heat capacity of the exhaust gas (absolute index), which is taken as an empirical constant.

[0041] This formula is based on the thermodynamic relationship of an isentropic process (reversible adiabatic) of an ideal gas. It assumes that the decrease (expansion) of the exhaust pressure inside the turbine will be completely converted into work done on the outside, resulting in a decrease in the internal energy of the gas and a drop in temperature. The theoretical magnitude of the drop is determined by the initial temperature, the ratio of pressure change (expansion ratio), and the physical properties of the exhaust gas itself (adiabatic index).

[0042] 2) Compressor power consumption: Power consumed by the compressor P comp Used to compress intake air: in This represents the intake air mass flow rate (kg / s), which can be measured or estimated from the air-fuel ratio and fuel flow rate. c P，air This represents the specific heat capacity of air at constant pressure (kJ / (kg·K)). T air，in This indicates the compressor inlet air temperature (K), measured. T air，out This indicates the compressor outlet air temperature (K), as measured.

[0043] 3) Turbocharger mechanical efficiency and overall efficiency factor: In practice, the power generated by the turbine is transmitted to the compressor via the shaft, resulting in mechanical losses. The overall efficiency factor η of the turbocharger is defined as follows: ＿turbo＿eff This is a dimensionless proportionality constant that comprehensively reflects the relative changes in turbine efficiency, compressor efficiency, and mechanical efficiency. Under steady-state conditions, neglecting heat loss, an approximate equilibrium exists: P tur，avail ·η ＿turbo＿eff ≈P comp ; will η ＿turbo＿eff As a distinguishing parameter, the calculation formula is: ; For a healthy new unit, under standard operating conditions, η ＿turbo＿effIt should be close to a nominal value (e.g., 0.65-0.75). This value will decrease as problems such as turbine blade fouling, bearing wear, and seal leakage occur. Therefore, η ＿turbo＿eff The ratio of the current value to its initial health baseline value constitutes the health index HI of the turbocharger subsystem. turbo The core part.

[0044] Physical Model 2: The fouling model of the cylinder liner water heat exchanger is constructed as follows: The performance degradation of a cylinder-liner water heat exchanger is mainly manifested in a decrease in the overall heat transfer coefficient U. According to the basic formula of heat transfer: Q=U·A·ΔT lm ; Where Q represents the heat exchange capacity (kW), which can be calculated using the cylinder liner water flow rate and the inlet / outlet water temperature difference: .in The mass flow rate of cylinder liner water (kg / s) refers to the mass of cylinder liner water flowing through the system per unit time. T represents the isobaric specific heat capacity of cylinder liner water at average temperature (kJ / (kg·K)), which represents the amount of heat absorbed (or released) by a unit mass of water when the temperature increases (or decreases) by 1 Kelvin under constant pressure; jw，in Indicates the inlet temperature of the cylinder liner water; T jw，out This indicates the outlet temperature of the cylinder liner water.

[0045] A represents the heat exchange area (m²), a fixed parameter of the equipment; ΔT lm The logarithmic mean temperature difference (K) is expressed by the following formula: Where ΔT lm It is an equivalent average value of the temperature difference between the hot and cold fluids across the entire heat exchange surface; T gas，out This indicates the inlet temperature of the "heat source" (actually the engine cylinder liner water) on the cylinder liner side; T gas，in This represents the outlet temperature of the "heat source" (actually the engine cylinder liner water) on the cylinder liner water side; however, in this model, the focus is on heat transfer from the engine to the cylinder liner water. More precisely, T gas，in This can be considered as the average temperature of the engine cylinder liner wall (heat source side) (which can be estimated using an empirical coefficient correlated with exhaust temperature or lubricating oil temperature); T gas，out It can be approximated as T jw，in (Assuming sufficient cooling); T jw，in Indicates the cylinder liner water inlet temperature (K); T jw，out Indicates the cylinder liner water outlet temperature (K); (T) gas，in -T jw，out (T) is the temperature difference between the hot fluid inlet and the cold fluid outlet (usually the large-end temperature difference); gas，out -Tjw，in (T) is the temperature difference between the hot fluid outlet and the cold fluid inlet (usually the smaller end temperature difference); gas，in -T jw，out )-(T gas，out -T jw，in () is the temperature difference between the two ends of the heat exchanger, which reflects the magnitude of the temperature difference change inside the heat exchanger; The logarithmic mean temperature difference (TT) is the natural logarithm of the ratio of the temperature differences at both ends. Taking the logarithm ensures that the calculated average temperature difference accurately reflects the logarithmic variation of the temperature difference along the heat transfer surface. It is suitable for steady-state heat transfer calculations of most indirect heat exchangers in counter-current or co-current arrangements. The logarithmic mean temperature difference formula is based on the integral solution of the differential equation of the heat transfer process under constant heat transfer coefficient and flow rate. It provides an accurate method to equate the integral of the temperature difference varying along the flow path to a constant temperature difference value ΔT. lm It is used to inversely calculate the overall heat transfer coefficient U from measured temperature and flow data, and the decay of U is precisely what assesses the health status of the heat exchanger (the degree of fouling) and generates the health index HI. jw The direct evidence.

[0046] Using the heat transfer coefficient U as the parameter to be identified, rearrange the formula: ; The current value of U current Relative to its cleanliness baseline value U clean The ratio of these ratios constitutes the HI (Health Index) of the cylinder liner water system. jw The core component, the dirt factor R f R can be calculated using the following formula: f =1 / U current -1 / U clean .

[0047] Key time-varying performance parameters (η) in the physical model ＿turbo＿eff The state (parameters) needs to be estimated and updated using real-time data. This invention employs an extended Kalman filter (EKF) because it can effectively handle nonlinear systems and provide optimal estimates of the state (parameters).

[0048] Turbocharger system η ＿turbo＿eff Estimates and updates: I. State-space model construction , will η ＿turbo＿eff It is considered as a slowly changing state variable χ (the same principle applies to the overall heat transfer coefficient U of the cylinder liner water heat exchanger).

[0049] 1) State equation: Assume its change is a noisy random walk process: χ k =χ k-1 +ω k-1,ω~N(0,Q); Where k is the discrete time step, k represents the current time, and k-1 represents the previous adjacent sampling time; χ is a state variable, representing a key, non-measurable quantity within the system that needs to be estimated or tracked. In this invention, η is specifically referred to as η. ＿turbo＿eff , χ k χ represents the true value (to be estimated) of the state variable χ at the current time k. k-1 This represents the true value of the state variable χ at the previous time k-1; ω represents process noise; Q is the covariance matrix, which reflects the possible rate of change of the parameters; N represents a normal distribution or a Gaussian distribution; (0, Q) represents the parameters of the normal distribution, the first number in parentheses represents the mean, and the second (matrix) represents the covariance matrix; The process noise ω ~ N(0, Q) follows a multidimensional normal distribution with a mean of 0 and a covariance matrix of Q.

[0050] In the state equation χ of the Extended Kalman Filter (EKF) k =χ k-1 +ω k-1 In ω~N(0,Q), this assumption is made in the statement: the turbocharger efficiency factor (state χ) changes between adjacent time steps, except for a deterministic slow evolution (by χ). k-1 In addition to the state η, there is a random, unbiased perturbation ω, the strength of which is characterized by Q. This Q is one of the key parameters in the EKF algorithm that needs to be pre-set or estimated online; it directly affects the filter's tradeoff between "trusting the model's prediction" and "trusting new observations" regarding state changes. By setting an appropriate Q, the state η can be effectively tracked. ＿turbo＿eff The true changes (i.e., performance degradation) are observed, while high-frequency noise in the measurement is smoothed out.

[0051] 2) Observation equation: Based on the physical model, our "observation" is the compressor power consumption P. comp The measured (or calculated) value z k The observation equation is: z k =h(χ k )+v k =(P tur，avail ) k ·χ k +v k v~N(0,R); Where v k This is the observation noise at the current moment; R is the covariance, which reflects the uncertainty of power measurement and model; (P) tur，avail ) k It is based on the currently measured T exh，in p exh，in p exh，out The calculated available power; h(χ) k ) is the observation function, which is a function that measures the state variable χ. k A function that maps to the expected observations; χ k It is the actual value of the state variable at time step k; (P) tur，avail ) k ·χ k It is the specific calculation formula of the observation function, that is, the theoretically predicted compressor power; v~N(0,R) indicates that the observed noise v follows a normal (Gaussian) distribution with a mean of 0 and a covariance of R; II. EKF Iteration Steps: For each sampling period (e.g., 1 second): 1) Prediction: Predicted state ; in This represents the predicted estimate of state χ at time k, based solely on all observation data up to time k-1 (i.e., prior information). (Subscript) It is read as "an estimate of time k given information at time k-1"; This represents the optimal estimate of the state χ at time k-1, which incorporates all observation data (i.e., posterior information) up to time k-1. This is the final output of the previous EKF iteration cycle (step). Since the state equation is assumed to be a random walk model Furthermore, the mean value of the noise ω is 0. Therefore, the optimal prediction of the state at the next moment is the optimal estimate of the current moment, which embodies the idea of ​​"inertial prediction".

[0052] Prediction error covariance: ; Represents the state prediction value The associated prediction error covariance matrix. It quantifies the magnitude of the uncertainty in the state estimate when based solely on model predictions (without incorporating new observations at time k). A larger P-value indicates higher prediction uncertainty; The posterior estimation error covariance matrix represents the uncertainty remaining after the previous estimation round. The process noise covariance matrix, as mentioned earlier, describes the intensity of the random disturbance ω that the state experiences during its evolution.

[0053] Uncertainty in forecasting It comes from two parts: 1) the uncertainty of the previous moment's estimate itself. 2) From the previous time step to the current time step, new uncertainties are introduced due to model imperfections and random perturbations. This formula shows that relying solely on model predictions can lead to a cumulative increase in uncertainty.

[0054] 2) Update: 2.1) Calculate the Kalman gain: ; Where K k Let represent the Kalman gain matrix at time k. It is a weight matrix that determines the extent to which the "surprise" of new observations should be used to correct the model's predictions during state updates; Represents the state prediction value The associated prediction error covariance matrix; H k For the observation equation z k =h(χ k )+v k The Jacobian matrix (first-order partial derivative matrix), due to z k =h(χ k )+v k =(P tur，avail ) k ·χ k +v k ,therefore This links the uncertainty of the state space with the uncertainty of the observation space; Representation matrix Transpose of; R represents the observation noise covariance matrix, which describes the intensity of noise v (measurement error, sensor noise) in the observed value z. A large R indicates that the observation data is unreliable and has high noise levels. This term represents the covariance of the prediction residuals (news) of the new observations. It consists of two parts: 1) the part where the uncertainty of the model prediction is mapped to the observation space. ;2) The uncertainty R of the observation itself: Formula meaning: Kalman gain K k The calculation follows an intuitive logic: if the model's predictions are very accurate ( If the noise level is low (R is large) or the observation noise is high (R is large), then K kIt will become smaller; this means the algorithm "trusts" the model's predictions more, and new observations have a smaller effect on correcting the final estimate. If the model's predictions are very accurate ( If the observation noise is large (R is small), then K k It will become larger; this means that the algorithm "trusts" the model's predictions more and uses them to significantly correct the model's predictions.

[0055] 2.2) Obtain new ; 2.3) Update state estimation: ; in At time k, the optimal estimate of state χ is obtained by incorporating all observations up to time k (i.e., posterior information). This is the final output of EKF at the current time. z k This represents the measured value at k, which is the compressor power calculated based on the sensor data. ; State prediction value Substituting these values ​​into the observation equation h(·), we obtain the predicted values ​​of the observed quantities. The new information, or observation residual, represents the difference between new observation data and model predictions, and is the direct source of information driving state updates.

[0056] The final optimal estimate is a weighted sum of the model predictions and corrections based on the new observations. The correction equals the "surprise from the new observations" (information) multiplied by a "confidence weight" (Kalman gain). This achieves an optimal fusion of model predictions and observed data.

[0057] 2.4) Update error covariance: ; in With posterior state estimates The associated updated error covariance matrix. It quantifies the magnitude of the remaining uncertainty in the state estimate after incorporating new observation information; I represents the identity matrix, whose dimensions are the same as those of the state vector; , , The meaning is the same as before.

[0058] This formula describes the decrease in estimation uncertainty after incorporating new observational data. Because the new information reduces the unknowns, the updated uncertainty... Typically less than the uncertainty of the forecast. .factor It can be viewed as a contraction factor caused by "information gain".

[0059] The input to EKF is: the efficiency estimate from the previous time step. Covariance And the measured data at the current moment (used to calculate z) k and H k The output of EKF is: the efficiency estimate at the current time after optimal correction. (i.e., η) ＿turbo＿eff,current ) and its uncertainty This estimate Used to calculate the health index HI turbo The data is then input into the time-series prediction model. The entire EKF iteration process enables real-time, online, and adaptive tracking of key equipment performance parameters, forming the core algorithmic foundation for building a high-fidelity digital twin and achieving predictive maintenance.

[0060] Estimation and update of the overall heat transfer coefficient U of the cylinder liner water heat exchanger: I. State-space model construction Consider U as a slowly changing state variable χ.

[0061] 1) State equation: Assume its change is a noisy random walk process: χ k =χ k-1 +ω k-1 ,ω~N(0,Q); Where k is the discrete time step, k represents the current time, and k-1 represents the previous adjacent sampling time; χ is a state variable, representing a key, non-measurable quantity within the system that needs to be estimated or tracked. In this invention, η is specifically referred to as η. ＿turbo＿eff , χ k χ represents the true value (to be estimated) of the state variable χ at the current time k. k-1 χ represents the true value of the state variable χ at the previous time k-1, where χ specifically refers to the overall heat transfer coefficient U. ω k-1 It is process noise, representing the random and minute changes in the heat transfer coefficient U from k-1 to k caused by unknown factors such as random accumulation of dirt and shedding of deposits; Q is the process noise covariance matrix, reflecting the possible rate of change of the parameters. It quantifies the intensity of random changes that the heat transfer coefficient U may undergo per unit time. A large Q value indicates large fluctuations and high uncertainty in the fouling growth process. N represents a normal distribution or a Gaussian distribution; (0, Q) represents the parameters of the normal distribution, the first number in parentheses represents the mean, and the second (matrix) represents the covariance matrix; The process noise ω ~ N(0, Q) follows a multidimensional normal distribution with a mean of 0 and a covariance matrix of Q.

[0062] In the state equation χ of the Extended Kalman Filter (EKF) k =χ k-1 +ω k-1 In ω~N(0,Q), this assumption is made in the statement: the overall heat transfer coefficient U (state χ) of the cylinder liner water heat exchanger changes between adjacent time steps, except for a deterministic slow evolution (by χ). k-1 In addition to the perturbation ω, there is a random, unbiased perturbation of strength characterized by Q. This Q is one of the key parameters in the EKF algorithm that needs to be preset or estimated online, and it directly affects the trade-off between "trusting model predictions" and "trusting new observation data" in response to state changes in the cylinder liner water heat exchanger. By setting an appropriate Q, the actual changes in state U (i.e., performance degradation) can be effectively tracked, while smoothing out high-frequency noise in the measurement.

[0063] 2) Observation equation: Based on the physical model, our "observation" is the actual heat transfer Q of the heat exchanger. measured The measured (or calculated) value z k The observation equation is: z k =h(χ k )+v k =A·△T lm，k ·χ k +v k v~N(0,R); Where z k It is an observed value, that is, based on real-time measurement data (cylinder liner water flow rate). Inlet and outlet water temperatures T jw，in T jw，out The calculated actual heat exchange: , where c p，w It is the specific heat capacity of cylinder liner water under constant pressure conditions; v k It is the observation noise at the current moment, representing the error in the heat exchange calculation, which originates from the measurement noise of temperature and flow sensors, as well as the simplification error of the heat transfer model (such as assuming a logarithmic mean temperature difference). A is the heat exchange area; △T lm，k It represents the logarithmic average temperature difference calculated at time k based on the measured inlet and outlet temperatures (for cylinder liner water heat exchangers, it is usually the temperature difference between the cylinder liner water and the cooling medium). This is a calculated quantity determined by real-time measurements.

[0064] R is the covariance, which reflects the uncertainty of the calculated heat exchange value; h(χ) k ) is the observation function, which is a function that measures the state variable χ. k The function mapped to the expected observations represents the state (heat transfer coefficient) at χ. k At that time, the theoretically predicted heat exchange rate; χ k It is the true value of the state variable at time step k, i.e., the overall heat transfer coefficient U; v~N(0,R) indicates that the observed noise v follows a normal (Gaussian) distribution with a mean of 0 and a covariance of R; II. EKF Iteration Steps: For each sampling period (e.g., 1 second): 1) Prediction: Predicted state ; in This represents the predicted estimate of the state χ (overall heat transfer coefficient U) at time k, based solely on all observation data up to time k-1 (i.e., prior information). (Subscript) It is read as "an estimate of time k given information at time k-1"; This represents the optimal estimate of the state χ (total heat transfer coefficient U) at time k-1, which incorporates all observation data up to time k-1 (i.e., posterior information). This is the final output of the previous EKF iteration cycle (step). Since the state equation is assumed to be a random walk model Furthermore, the mean value of the noise ω is 0. Therefore, the optimal prediction of the state at the next moment is the optimal estimate of the current moment, which embodies the idea of ​​"inertial prediction".

[0065] Prediction error covariance: ; Represents the state prediction value The associated prediction error covariance matrix. It quantifies the magnitude of the uncertainty in the state estimate when based solely on model predictions (without incorporating new observations at time k). A larger P-value indicates higher prediction uncertainty; The posterior estimation error covariance matrix represents the uncertainty remaining after the previous estimation round. The process noise covariance matrix, as mentioned earlier, describes the intensity of the random disturbance ω that the state experiences during its evolution.

[0066] Uncertainty in forecasting It comes from two parts: 1) the uncertainty of the previous moment's estimate itself. 2) From the previous time step to the current time step, new uncertainties are introduced due to model imperfections and random perturbations. This formula shows that relying solely on model predictions can lead to a cumulative increase in uncertainty.

[0067] 2) Update: 2.1) Calculate the Kalman gain: ; Where K k Let represent the Kalman gain matrix at time k. It is a weight matrix that determines the extent to which the "surprise" of new observations should be used to correct the model's predictions during state updates; Represents the state prediction value The associated prediction error covariance matrix; H k For the observation equation z k =h(χ k )+v k The Jacobian matrix (first-order partial derivative matrix), due to z k =h(χ k )+v k =A·△T lm，k ·χ k +v k ,therefore The uncertainty of the state space This is related to the uncertainty of the observation space (heat exchange); Representation matrix Transpose of; R represents the observation noise covariance matrix, which describes the intensity of noise v (measurement error, sensor noise) in the observed value z. A large R indicates that the observation data is unreliable and has high noise levels. This term represents the covariance of the prediction residuals (news) of the new observations. It consists of two parts: 1) the part where the uncertainty of the model prediction is mapped to the observation space. ;2) The uncertainty R of the observation itself: Formula meaning: Kalman gain K k The calculation follows an intuitive logic: if the model's predictions are very accurate ( If the noise level is low (R is large) or the observation noise is high (R is large), then K k It will become smaller; this means the algorithm "trusts" the model's predictions more, and new observations have a smaller effect on correcting the final estimate. If the model's predictions are very accurate ( If the observation noise is large (R is small), then K k It will become larger; this means that the algorithm "trusts" the model's predictions more and uses them to significantly correct the model's predictions.

[0068] 2.2) Obtain the new z k (The actual heat exchange rate Q is calculated based on the latest sensor data) measured ); 2.3) Update state estimation: ; in At time k, the optimal estimate of state χ is obtained by incorporating all observations up to time k (i.e., posterior information). This is the final output of EKF at the current time. It is based on the heat exchange predicted by the model, and the state prediction value is used. Substituting these values ​​into the observation equation h(·), we obtain the predicted values ​​of the observed quantities. The new information, or observation residual, represents the difference between new observation data and model predictions, and is the direct source of information driving state updates.

[0069] The final optimal estimate is a weighted sum of the model predictions and corrections based on the new observations. The correction equals the "surprise from the new observations" (new information) multiplied by a "confidence weight" (Kalman gain). This achieves the optimal fusion of model predictions and measured data, i.e., using the actual heat exchange z. k Heat exchange predicted by the model The difference (i.e., the innovation) is multiplied by the Kalman gain K. k This is used to correct the predicted value of the heat transfer coefficient χ. If the actual heat exchange is less than the predicted value, it indicates that the heat transfer performance may have deteriorated (U decreases), and the estimated value... It will be corrected downwards.

[0070] 2.4) Update error covariance: ; in With posterior state estimates The associated updated error covariance matrix. It quantifies the uncertainty in estimating the heat transfer coefficient U after incorporating new observation information. reduce; I represents the identity matrix, whose dimensions are the same as those of the state vector; , , The meaning is the same as before.

[0071] This formula describes the decrease in estimation uncertainty after incorporating new observational data. Because the new information reduces the unknowns, the updated uncertainty... Typically less than the uncertainty of the forecast. .factor It can be viewed as a contraction factor caused by "information gain".

[0072] The input to EKF is: the efficiency estimate from the previous time step. Covariance And the measured data at the current moment (used to calculate z) k and H k The output of EKF is: the efficiency estimate at the current time after optimal correction. (i.e., U) and its uncertainty This estimate Used to calculate the health index HI jw The data is then input into the time-series prediction model. The entire EKF iteration process enables real-time, online, and adaptive tracking of key equipment performance parameters, forming the core algorithmic foundation for building a high-fidelity digital twin and achieving predictive maintenance.

[0073] To obtain a comprehensive and easily understandable overall equipment health status, it is necessary to combine the health proxy indicators of multiple subsystems (in this embodiment, at least the cylinder liner water heat exchanger and turbocharger system) into a single overall health indicator. Where 1 represents a brand new, perfect state, and 0 represents a completely invalid state, the steps are as follows: 1) Calculation of sub-health indicators: HI turbo =η ＿turbo＿eff,current / η ＿turbo＿eff,baseline ; HI turbo For turbocharger health indicators; η ＿turbo＿eff,current This is the optimal estimate of the turbocharger's overall efficiency factor at the current moment. It is output in real time by an online parameter identification algorithm (such as an extended Kalman filter, EKF). This value incorporates the latest sensor data and reflects the true instantaneous performance of the equipment under the current operating conditions. η ＿turbo＿eff,baseline This is the nominal value of the overall efficiency factor of a turbocharger when it is in a healthy, new condition and operating under standard or typical conditions. This value is usually derived from factory test data, design values, or the average value obtained from long-term observation and statistics under stable conditions during the initial stage of system operation.

[0074] HI jw =U current / U baseline ; HI jw For the health indicators of cylinder liner water heat exchangers; U currentThis is an estimate of the overall heat transfer coefficient of the cylinder liner water heat exchanger at the current moment. Also derived from an online parameter identification algorithm, it reflects the heat exchanger's real-time heat transfer capacity affected by fouling, scaling, etc. U baseline This refers to the design value or initial calibration value of the overall heat transfer coefficient of the cylinder liner water heat exchanger under clean and foul-free conditions.

[0075] The above formula for calculating sub-health indicators transforms performance parameters (efficiency factor, heat transfer coefficient, etc.) with different physical units and dimensions into a dimensionless scalar value ranging around 1 by dividing the numerator (real-time value) by a baseline value. This makes performance degradation of different components and types comparable and compatible. The aforementioned HI turbo HI jw Sub-health indicators are used to synthesize the overall health indicator HI. overall Direct input (via weighted composite formula) This is also a key time series data input into time series prediction models (such as LSTM) to predict future degradation trajectories.

[0076] 2) Weighted synthesis: ; Where w i The weights are assigned based on the criticality of the components (∑w) i =1); i is the index of the sub-health indicator, representing the i-th monitored component or subsystem; n is the total number of sub-health indicators, that is, the number of key components or subsystems participating in the overall assessment. In this invention, n is at least 2, that is, the key components or subsystems participating in the overall assessment consist of the turbocharger system and the cylinder liner water heat exchanger. HI i It is the health indicator of the i-th subsystem or component; f i (·) is a nonlinear mapping function or normalization / standardization function for the i-th sub-health indicator, used to handle the threshold effect of certain HIs (such as vibration). It can be a sigmoid function (used to smoothly handle the threshold effect), an exponential function (used to amplify or suppress certain changes), a piecewise linear function (based on engineering experience, assigning different "health contributions" to different health intervals), or a simple linear function f. i (x) = one of x (used when the degradation of the component is considered linear and normalized). For example, in this embodiment, only HI is considered. turbo HI jw Given two health indicators, if we assume that the impact of the turbocharger efficiency decrease and the reduction in the cylinder liner water heat exchanger heat transfer coefficient on the overall system health is approximately linear, then the identity function f can be directly applied.i (HI) turbo )=HI turbo and f i (HI) jw )=HI jw .

[0077] 3) Standardization and smoothing For the synthesized HI overall Apply an exponentially weighted moving average (EWMA) filter to smooth short-term fluctuations and highlight long-term trends.

[0078] History overall Using sequences such as load rate sequences and ambient temperature sequences as inputs, a Long Short-Term Memory (LSTM) neural network is trained to predict HI levels over the next N hours (e.g., 24 hours or 168 hours). overall The trajectory of change. Input features include: HI data from the past M time steps. overall Engine power, ambient temperature, etc. Output: HI values ​​for the next N time steps. overall Predicted values. Training: Supervised training is performed using historical uptime data (excluding failures and downtime) with the goal of improving the predicted HI. overall Trajectory and actual observed HI overall Minimize trajectory error. Online prediction: Input the latest window data into the trained LSTM model to obtain the future HI. overall The predicted value HI pred (t) and its confidence interval. This prediction result is a key input for subsequent optimization controllers considering "long-term health costs". Another innovation of this application is the integration of LSTM with the Exponentially Smoothed State-Space Model (ETS). The ETS model is specifically designed to capture short-term trends (such as accelerated decline due to several consecutive days of high load operation) and cyclical fluctuations (such as regular changes associated with daily start-stop or seasonal maintenance) that may exist in health indicator data. The outputs of the two models are integrated in a learnable weighted fusion layer, the weights of which can be adjusted online according to the prediction error. This multi-scale fusion architecture is more robust than a single LSTM model and can more accurately decompose and predict the components of different time scales in changes in health indicators.

[0079] Step 3: Using a comprehensive function that includes the virtual health cost based on the prediction of future changes in health indicators and the short-term operational economic benefits as the objective function, and the equipment operation boundary that is dynamically related to health indicators as the constraint condition, solve for the optimal setpoint sequence of the generator and waste heat recovery system control strategy.

[0080] The health status-driven multi-objective real-time optimization controller is responsible for transforming health status information into real-time control commands. It takes a comprehensive function that includes virtual health costs based on predictions of future changes in health indicators and short-term operational economic benefits as its objective function. The optimization problem is solved once in each control cycle (e.g., 5 minutes).

[0081] The control variable u is defined as the set of all adjustable control commands that the health-state driven multi-objective real-time optimization controller of this invention needs to solve in each control cycle. It is a vector: u = [P gen V bypass f pump1 f pump2 ..., SOC tes T set，cond ,...] T .

[0082] Where P gen The active power setpoint of the generator instructs the electrical power (kW) output of the natural gas internal combustion engine (or gas turbine), controlling the primary energy input and main product output of the system. gen It directly determines power generation, fuel consumption, and the total amount of waste heat generated by the main unit. It is the most important lever for balancing electrical load demand and system operating status; V bypass The opening of the flue gas bypass valve controls the percentage of high-temperature flue gas discharged from the internal combustion engine that bypasses subsequent deep recovery equipment (such as flue gas heat exchangers and absorption chillers) and is directly discharged into the chimney. This percentage (%) is a key heat energy distributor. It determines whether the high-quality flue gas waste heat is used for deep recovery (heating / cooling) or is discarded. It is a core variable in reconciling the conflict between "power generation" and "heating / cooling".

[0083] f pump1 f pump2 ... represents the frequency converter frequency of each main water pump, controlling the flow rate (Hz) of the working fluid (water / solution) flowing through each heat exchanger (such as cylinder liner water heat exchanger, flue gas heat exchanger) or heating network. pump1 f pump2 ...to finely regulate the heat transfer process. By changing the flow rate, the heat exchange temperature difference can be optimized, the effects of heat exchanger fouling can be overcome, the heating / cooling power can be adjusted, and the power consumption of the water pump itself can be directly affected; SOC tes The target state of charge (SOC) for the thermal storage tank is defined as the desired thermal storage level to be achieved by the thermal storage tank (hot water tank) at the end of the control cycle, expressed as a percentage (%). tesThis enables energy transfer across time periods. By "peak shaving and valley filling," excess heat during off-peak hours is stored and released during peak hours. This allows generator loads to operate more smoothly, adapting to electricity price fluctuations and load changes, and protecting equipment from frequent load shocks. T set，cond For deep power generation devices such as organic Rankine cycles, set the condenser operating temperature (e.g., ORC) to the condenser setpoint. Optimize the efficiency of low-grade waste heat power generation cycles. By adjusting the condensing pressure, an optimal balance can be achieved between power generation efficiency and cooling costs.

[0084] u is the independent variable in the objective function J(u), and its value determines the system's position on the "economic-healthy" Pareto front. The result of optimizing the objective function J(u) is u. It is the optimal control command issued by the controller to the physical system, integrating dual objectives. Through u, the originally independent "operational optimization" (pursuing J...) short Maximum) and "Predictive Maintenance" (Assessment C) health The minimum cost is integrated into the same mathematical optimization problem. When solving the problem, the optimizer must simultaneously and in real time calculate the economic benefits and health costs of each candidate operation, thereby achieving true "lifecycle cost optimization" control.

[0085] The objective function (a composite function including a virtual health cost model based on predictions of future changes in health indicators and short-term operational economic benefits) is a hybrid function that balances short-term economic benefits with long-term equipment health costs: ; Among them, R elec R represents revenue from electricity generation. elec =P gen ·τ·Price elec P gen τ represents the generator active power setpoint (decision variable), in kilowatts (kW), indicating the power output of the generator as instructed by the controller; τ is the control cycle duration, in hours (h), representing the duration of the current optimization instruction (e.g., 5 minutes = 1 / 12 hour); Price elec The electricity price is a forecast, in yuan / kWh, and is derived from external market data or forecast data. R heat For heating / cooling revenue, R heat =Σ(Q delivered，i Price heat，i ). Among them Q delivered，iPrice represents the actual heat or cooling output delivered in the i-th heating or cooling network, measured in kilowatt-hours (kWh). It is a function of control variables in the deep recovery system, such as valve opening and pump speed. heat，i The price of heat or cold energy corresponding to the i-th heating / cold network, in yuan / kWh. C fuel Indicates fuel cost, .in This indicates the mass flow rate of natural gas, measured in kilograms per hour (kg / h). It is the generator output P. gen and the current unit power generation efficiency η gen (P) gen The function of HI) ∝P gen / η gen ); τ is the control cycle duration (h); Price fuel This is the price of fuel, in yuan / kg (¥ / kg); C fuel That is, the power generated P gen And the total cost of fuel consumed for the corresponding heat; C om，var This represents the variable maintenance costs associated with operation. These costs are directly related to the current operation, such as the power consumption costs of auxiliary equipment (pumps, fans) and the expected depreciation costs of consumables (such as spark plugs, filters) due to specific operating modes (such as frequent start-stop, high load). It is an empirical or statistical model, usually a function of the control variable u (such as pump frequency, number of start-stop). Short-term net income J short Overall meaning: The pre-tax gross profit (revenue minus direct variable costs) obtained from operating a combined heat and power system within a control period τ.

[0086] Where △HI j，pred (u) represents the predicted decline in the health index of a critical component j (such as a turbine or heat exchanger) over a future period (e.g., the next 24 hours) after strategy u is implemented (it is a negative value or zero). △HI j，pred (u) is calculated using a simplified virtual health cost model, which establishes control variables u (such as P). gen T exh The relationship between component stress / degradation rate.

[0087] For example, for a turbocharger, its degradation rate is related to the exhaust temperature T. exh and load P gen Related, the model is in the form of: Here, a and b are empirical indices, which best reflect the engineering experience and data-driven characteristics of the model. They quantify the nonlinear intensity of the influence of various driving factors on degradation. a>0 indicates that exhaust temperature accelerates aging, and typically a>1 indicates that the damage caused by high temperature increases nonlinearly (e.g., for every 10°C increase in temperature, the material creep life may be halved). b>0 indicates that load accelerates aging, and its magnitude reflects the coupling effect of mechanical and thermal loads. The initial values ​​of these indices can be set based on materials science knowledge (such as the Arrhenius equation) or component manufacturer life data, and are ultimately optimized by fitting historical operating data through a closed-loop learning mechanism; K turbo The proportionality coefficient is an empirical coefficient used to scale the combined effect of driving factors to the actual change in HI. Its value incorporates inherent factors such as material properties and design margins; T exh (u) The exhaust temperature reached by the turbine under the control variable u; P ren (u) Under control variable u, the generator's output power is an operating parameter directly determined by the generator power setpoint in control variable u. High load means higher mechanical stress, higher combustion pressure, and higher temperature; Δt is the length of the evaluation time window or the predicted time step; f turbo (T) exh (u), P ren (u), ...) is a multivariable function representing the instantaneous health degradation rate or health risk intensity of a turbocharger. Specifically, it refers to a function with multiple operating parameters as inputs and outputs as degradation rate drivers. The independent variables of the function are a series of key operating parameters determined by the control variable u, such as P. ren (u), T exh (u) and other possible variables: such as vibration amplitude, pressure, speed, start-stop state, etc. For turbochargers, the functional form is f turbo (T) exh P ren ) = (T exh -T ref ) a ·(P) ren / P rated ) b ;T ref This indicates a reference temperature threshold or low-stress reference temperature representing material degradation; P rated This indicates the generator's rated (nameplate) power; the formula uses temperature difference (T). exh -T ref (rather than absolute temperature T) exh This means that only the "overheating" portion exceeding the safety benchmark is included in health losses; the ratio P gen / P ratedThe load factor eliminates the influence of unit capacity differences, making the model more universal and intuitively reflecting the "working intensity" of the equipment relative to its design capacity. This is more in line with engineering reality and makes the model parameters more physically interpretable. The use of a product form instead of an additive form is based on an important engineering understanding: the combined effects of high temperature and high load are far more damaging than the sum of their individual effects (e.g., thermomechanical fatigue). This coupling effect is a typical aging mode for rotating machinery such as turbines. The function value f... turbo (·) is a dimensionless quantity or a quantity with a specific unit. It quantitatively describes the magnitude of the "health pressure" or "aging driving force" that the turbocharger experiences per unit time at the current operating point. The larger the value, the more severe the operating conditions and the faster the health deteriorates.

[0088] For example, the driving function for the health degradation rate of a cylinder liner heat exchanger is defined as: ; Among them, the change in health indicators predicted by the cylinder liner heat exchanger within the evaluation time window Δt after implementing the control variable u. for: ; Where f jw U(u) is the driving function for the health degradation rate of the cylinder liner heat exchanger, a dimensionless risk intensity index. A larger value indicates a faster rate of health degradation due to fouling or material stress at the current operating point. U(u) is the total heat transfer coefficient under the current operating conditions (unit: W / (m²·K)), calculated online based on real-time data: U(u) = Q(u) / [A·ΔT] lm [(u)], where Q(u) is the real-time heat exchange, A is the heat exchange area, and ΔT is the heat exchange area. lm (u) represents the real-time logarithmic mean temperature difference. U clean The overall heat transfer coefficient under clean conditions (unit: W / (m²·K)) is a fixed design parameter for the equipment. The heat transfer performance degradation factor is its reciprocal U(u) / U clean The health indicator HI of the heat exchanger jw This factor directly reflects the increase in heat transfer resistance caused by fouling and scaling. The worse the performance (the smaller U), the larger this factor, and the faster the degradation. m is the exponent (dimensionless) of the heat transfer degradation stress term, an empirical parameter greater than zero, reflecting the nonlinear impact of decreased heat transfer performance on maintenance needs (such as cleaning frequency) and operating energy consumption (increasing flow rate or temperature difference to achieve the same heat exchange). △p(u) is the actual pressure drop on the working fluid side of the heat exchanger (unit: Pa), obtained through measurement by a differential pressure sensor, and is a function of the flow rate. A function of fluid properties. △p cleanThe pressure drop (in Pa) is the pressure drop at the design flow rate under clean conditions, and the design parameters are for the equipment. The increased pressure drop, a factor contributing to flow resistance, is typically caused by channel blockage and fouling. This not only increases pump consumption but also exacerbates the risk of localized erosion corrosion. n is the exponent (dimensionless) of the flow resistance stress term, an empirical parameter greater than zero. T wall，max (u) is the estimated maximum metal temperature of the heat exchanger tube wall (unit: K), which can be estimated using a simplified one-dimensional heat transfer model based on the flue gas side and working fluid side temperatures, as well as the current overall heat transfer coefficient U(u). ref，jw This refers to the allowable temperature or critical temperature for low-temperature corrosion of the heat exchanger tube wall material (unit: K), determined based on the material properties. This is an overheating or low-temperature corrosion risk factor. When the wall temperature exceeds the allowable value (high-temperature side) or falls below the acid dew point (low-temperature side), this factor is positive, representing the material's overheating creep risk or low-temperature corrosion risk, respectively, and exhibits non-linear amplification (exponent p>0); where max() is the maximum value function, T wall，max (u) represents the estimated maximum metal temperature of the heat exchanger tube wall under control strategy u; T ref，jw The reference critical temperature (T) is a key engineering parameter set based on material properties and medium composition. wall，max (u)-T ref，jw This represents the deviation between the estimated wall temperature and the reference temperature, divided by the denominator T. ref，jw The normalization of temperature difference is indicated; the exponent p represents the nonlinear exponent of risk penalty, which determines the sensitivity and nonlinearity of health costs to the degree of overheating. When p=1, the penalty is linearly related to the overheating ratio; when p>1 (e.g., p=2, 3), the penalty is exponentially related to the overheating ratio. This means that a small overheating results in a limited increase in health costs, but once the overheating intensifies, the health costs will soar exponentially. This accurately simulates engineering reality: when materials approach their design limits, their lifespan deteriorates rapidly. K jw ΔHI is the heat exchanger health degradation proportionality coefficient (unit: 1 / h), an empirical coefficient related to fouling growth characteristics and material corrosion rate, used to convert the driving function into a rate of change of health indicators. Δt is the prediction time window for optimal control (unit: h). jw，pred (u) represents the predicted change in health indicators (dimensionless), which is negative and will be directly used to calculate the virtual health cost item; HI j，current This indicates the current real-time health indicators of component j; α jThis represents the maintenance cost coefficient for component j, in yuan. It converts a decrease in a health indicator (a technical indicator) into an equivalent economic cost. For example, a decrease of 0.1 in turbine HI might correspond to a major overhaul cost of 100,000 yuan, so α can be set as 1,000,000 yuan / HI-unit; β j It is a health status penalty factor. It is a coefficient greater than or equal to 1, and is HI. j，current A decreasing function. When HI j，current When it is already low (poor equipment condition), β j The value will increase. This allows the optimizer to adopt maintenance strategies for components that are already unhealthy. Its functions are twofold: 1) to ensure that the cost is always positive; 2) to impose a non-linear penalty on rapid relative degradation (i.e., the rate of decay per unit of current health), prompting the optimizer to select an operating point that can maintain a stable health state (slow degradation rate). C health (u, HI) pred The overall forecast estimates the negative impact of the current "aggressive" operating strategy on future equipment lifespan and maintenance bills. It guides the optimizer to proactively select operating points that are "milder" for the equipment, even if these points are less economical in the short term; λ represents the weighting coefficient (hyperparameter) of health costs, balancing short-term benefits. short With long-term health C health The trade-off between λ and λ is as follows: the larger λ is, the more "conservative" the system is, and the more it tends to protect the equipment, which may sacrifice some economic efficiency; the smaller λ is, the more "aggressive" the system is, and the more it pursues immediate economic benefits. Objective function J(u) = -J short +λJ health This is a scalarized form of a multi-objective optimization problem. By solving this problem, the optimization controller automatically seeks the immediate operating strategy that achieves the optimal total lifecycle cost (operating revenue - fuel cost - maintenance cost - future health depreciation cost) while satisfying all constraints (especially dynamic health constraints). This signifies a shift in control objectives from the traditional "highest instantaneous energy efficiency" or "lowest cost" to "maximizing asset value throughout its entire lifecycle." The health status-driven multi-objective real-time optimization controller uses the equipment operation boundary, which is dynamically associated with health indicators, as the constraint condition. The constraint is divided into static constraints and dynamic health constraints.

[0089] Static constraints include: 1) Generator power upper and lower limits: ;P gen For engine power, This is the lower limit of power. This is the upper limit of power; 2) Equipment physical limits: valve opening degree 0-100%, pump frequency range, etc.; 3) Power grid / heat network demand balance: P gen +P grid =P load Q supply =Q demand ±△Q storage ;where P grid P represents the exchange power (kW) with the public power grid; a positive value indicates purchasing electricity from the grid, and a negative value indicates selling electricity to the grid. load For local (or factory area) electricity load demand (kW); Q supply Q represents the actual heat (cooling) supplied by the system (kW). demand For heat (cooling) users' demand load (kW); △Q storage The power of the thermal storage tank (kW) is used for charging / discharging heat. A positive value indicates that heat is released from the thermal storage tank (increased supply), and a negative value indicates that heat is stored in the thermal storage tank (decreased net supply).

[0090] 4) Safe operation constraints: such as flue gas dew point temperature limits to prevent low-temperature corrosion.

[0091] Dynamic health constraints: The operating limits of a device are no longer fixed, but rather a function of its health status: ; Where g i (u) is a function of the operating or state variables that need to be constrained. For example, for a turbine g i (u) = T exh (Exhaust temperature); The design upper limit value of the operating variable or state variable that needs to be constrained when the equipment is new and in a healthy state (HI=1); Φ i (HI relevant ) is a dynamic scaling function with values ​​in the interval (0, 1). It is based on the relevant health indicator HI. relevant A scaling factor is calculated in real time. As HI decreases, Φ... i Reducing means lowering the operating limit. It then decreases.

[0092] The following is an example analysis of the above dynamic health constraint function: 1) Maximum exhaust temperature of turbine: ; Where T exh The actual / predicted exhaust temperature (K) of the turbine; The design of the new turbine allows for the highest exhaust temperature (K); θ turbo This is the sensitivity coefficient of turbine health to temperature constraints (0 < θ < 1). It determines the drastic decrease in the upper temperature limit when health indicators decline. HI turbo The current health indicator of the turbine subsystem (between 0 and 1); Turbine health (HI) turbo ≈1), allowing operation near the design maximum temperature. With health degradation (HI) turbo (Decrease), the maximum permissible temperature decreases linearly. For example, if HI turbo =0.8, θ turbo =0.3, then the maximum allowable temperature is 1-0.3 of the original design value. (1-0.8)=0.94 times. This actively reduces the operating stress of degraded equipment.

[0093] 2) Engine maximum load rate constraint: Among them, P gen The generator active power setpoint (kW); P rated The rated power of the generator (kW); θ load The sensitivity coefficient of overall health to load constraints; HI overall The overall health index of the equipment is a weighted average of the various sub-health indicators. When the overall health of the equipment declines, the maximum permissible power output decreases accordingly. This prevents the equipment from operating at full capacity when it is in a "sub-healthy" state, thus avoiding accelerated damage.

[0094] 3) Minimum temperature difference constraint for cylinder liner heat exchangers (to prevent further deterioration of heat transfer due to fouling): △T lm The actual logarithmic mean temperature difference (K) of the heat exchanger is the driving force for heat exchange. Minimum design temperature difference (K) required to ensure basic heat transfer in order to clean the heat exchanger; θ fouling The sensitivity coefficient of the dirt to the minimum temperature difference requirement (θ>0); HI jw The health index of the cylinder liner water heat exchanger reflects its degree of fouling (1 when clean). For cleaning heat exchangers (HI) jw=1), only the original minimum temperature difference requirement needs to be met. When heat exchanger fouling (HI) jw When the temperature difference is less than 1, its heat transfer capacity decreases. To ensure the necessary heat transfer, a larger minimum temperature difference is required (the right side of the inequality increases). This may force the system to increase the working fluid flow rate (increasing pump consumption) or adjust the temperature, which may sacrifice some efficiency but ensures the reliability of heat exchange; Static constraints define the absolute hard boundaries of system operation, while dynamic health constraints introduce adaptive soft boundaries. The latter is the key technical means for this invention to achieve a deep integration of "predictive maintenance" and "operation optimization": it uses real-time health status information to dynamically and proactively tighten operational constraints, guiding the optimization controller to automatically select a more "friendly" operating strategy for the equipment, thereby finding the optimal solution while ensuring the long-term reliability of the equipment, and realizing a mode shift from "post-fault repair" or "periodic over-maintenance" to "state-based adaptive operation and maintenance".

[0095] The health-state driven multi-objective real-time optimization controller uses sequential quadratic programming (SQP) and rolling time-domain solution to find the optimal setpoint sequence of control variables for the generator and waste heat recovery system.

[0096] Because the problem has a nonlinear objective and nonlinear constraints, the Sequential Quadratic Programming (SQP) algorithm is used to solve it. In each iteration, SQP constructs a quadratic programming (QP) subproblem to approximate the original problem, and updates the iteration points by solving the QP subproblem until convergence.

[0097] To handle forecast uncertainty and implement feedback control, a Model Predictive Control (MPC) framework is employed: The forecast time domain is, for example, the next hour, divided into four 15-minute intervals. The control time domain is typically equal to or shorter than the forecast time domain. In each control cycle, the optimizer solves for the optimal control sequence u within the next forecast time domain. (t), t=1, ...,N, but only the control command u for the first time period is implemented. (1) Control command issuance and execution is used to issue immediate commands from the optimal setpoint sequence to the underlying actuators. In the next cycle, prediction and optimization are re-performed based on new measurements (including the latest HI), and so on, rolling forward.

[0098] The MPC framework naturally incorporates forecasts of future load, electricity prices, and future health status (HI). pred .

[0099] Specific examples of deep recycling optimization strategies: Under the optimizer's decisions, the deep recycling system will exhibit dynamic intelligent behavior: Flue gas distribution under varying operating conditions: During periods of high electricity prices, even HI turboSlightly lower, the optimizer may still prioritize generation by selecting a slightly higher load and lower bypass, within the limits of dynamic constraints. This applies during periods of low electricity prices, or when HI... turbo At lower temperatures, the optimizer may choose to reduce the load and increase the bypass to direct more flue gas to the absorption chiller or thermal storage, thus protecting the turbine and preparing for subsequent peak heating / cooling.

[0100] Pump-assisted frequency conversion: based on real-time heat load and heat exchanger health status (HI). jw The pump speed of each loop is dynamically adjusted to minimize the total power consumption of all pumps while ensuring heat exchange requirements are met. When a heat exchanger experiences severe fouling (HI...),... jw At low temperatures, the optimizer may tend to increase its working fluid flow rate to compensate for the decrease in heat transfer coefficient, but this will increase pump consumption, therefore it is necessary to adjust the flow rate at J. short and J health We need to weigh the pros and cons.

[0101] Intelligent scheduling of thermal storage tanks: Thermal storage tanks (TES) serve as an important buffer for cross-period optimization. The optimizer decides when to store and release heat, taking into account not only electricity price differences and heat demand, but also equipment health. For example, if it is predicted that a high-load operation (which may exacerbate equipment wear and tear) will be required to meet heat demand in a future period, the optimizer may store heat in advance during low-load periods, thereby smoothing the operating curve and reducing equipment stress.

[0102] Step 4: Based on the actual operational feedback of the cogeneration system, perform online correction and adjustment of the parameters and weights of the objective function. The health-state-driven multi-objective real-time optimization controller is not a static optimizer, but an intelligent controller capable of continuously learning from operational experience. It can perform online correction and adjustment of the parameters and weights in the virtual health cost model based on the actual operational feedback of the cogeneration system.

[0103] Online calibration of virtual health cost models: The "virtual health cost model" used in the objective function (predicting ΔHI) j，pred (The simplified model of (u)) may initially be based on engineering experience or limited data. As the system runs, data can be accumulated to correct it, as follows: Data logging: Recording every control decision. k And the ΔHI actually observed within a period of time after execution (e.g., 24 hours). j，observed Change, ΔHI j，observed Indicating in strategy u k Under the influence of [the data], the health indicators of component j undergo real and actual changes over a period of time.

[0104] Model updates: periodically (e.g., weekly) update the accumulated {u k ΔHIj，observed The data pairs are used to update the parameters of the health impact assessment model (such as the indices a, b, and proportionality coefficient K in the aforementioned turbocharger long-term health model). j Linear regression or a simple neural network can be used for fitting.

[0105] This makes the system's understanding of "what kind of operation has what kind of impact on the equipment" increasingly accurate.

[0106] Adaptive adjustment of the weighting coefficient λ for health costs: The weighting coefficient λ controls the trade-off between short-term gains and long-term health. A fixed λ is unsuitable. We need to design a feedback-based adaptive rule: 1) Set health target range: for HI overall Define a healthy operating range, for example, [0.85, 0.95]. Ideally, the system should maintain a healthy operating range within this range.

[0107] 2) Monitor HI trends: Calculate the moving average and trend of HI (e.g., the slope of a linear regression over the past week). HI ).

[0108] 3) Adjust λ: If HI current <0.85 and slope HI If λ < 0 (poor health and continuous deterioration), then λ is significantly increased, forcing the optimizer to adopt a very conservative operating strategy and prioritize restoring health. If HI current >0.95 and slope HI If the value is >0 (health is very good), then λ can be appropriately reduced to allow the optimizer to more actively pursue short-term economic benefits; In other cases, λ remains relatively stable or is finely adjusted.

[0109] 4) Upper and lower limit protection: Set reasonable upper and lower limits for λ to avoid extreme decisions.

[0110] Through this mechanism, the system can dynamically adjust its decision preferences based on the actual performance of the device, achieving true personalized and adaptive optimization.

[0111] Continuous training of digital twin prediction models: Regularly (e.g., monthly) retrain the LSTM degradation prediction model with the latest operational data to ensure it can capture the latest aging characteristics and operating patterns of the equipment.

[0112] Step 5: Set the optimal setpoint sequence u The immediate instructions are issued to the underlying execution mechanism.

[0113] For example: Taking the specific operation of a 1MW natural gas internal combustion engine CHP system on a summer day as an example, the operation process of this system is explained in detail.

[0114] 05:00: The system is operating at low load. The optimizer receives forecast information: high electricity prices during the day and high cooling load demand; current HI turbo =0.88, HI jw =0.92. LSTM predicts that if the current medium load operation is maintained, HI will decrease gradually.

[0115] Decision: After the optimizer solves the problem, the instruction is to increase the generator output to 900kW, slightly reduce the flue gas bypass, and increase the share of flue gas going to the absorption chiller to provide cooling energy storage for the building in advance. Although this will slightly increase the turbine thermal stress (increase virtual health costs), the benefits from higher electricity prices and meeting cooling demand outweigh the gains, and the HI is still within a safe range.

[0116] 10:00: Electricity prices reach their peak, and cooling load demand surges. At this time, the cylinder liner water heat exchanger, due to slight fouling, HI... jw The value decreased slightly to 0.90. The dynamic constraints automatically and slightly increased the minimum temperature difference requirement for the heat exchanger.

[0117] Decision: The optimizer decides to maintain the generator output at 950kW (close to but not exceeding the dynamic upper limit under the health constraints) and start the thermal storage tank to release heat to supplement some of the air conditioning heat source, thereby allowing the absorption chiller to operate at full load and meet the cooling demand. At the same time, the cylinder liner water pump frequency is slightly increased to overcome the effects of fouling and meet the new temperature difference constraints.

[0118] 14:00: The weather forecast indicates thunderstorms this evening, which may cause power grid fluctuations. Meanwhile, the health prediction model shows that if peak operation continues, the HI (high-voltage) level will rise in the evening. turbo The predicted value will be close to the warning threshold of 0.85.

[0119] Decision: The optimizer proactively reduced the generator output to 800kW and increased flue gas bypass to store more heat in the thermal storage tank. Although some current power generation revenue was lost, potential grid risks were mitigated, and equipment stress was proactively reduced, bringing the predicted high-temperature (HI) trajectory back to a safe range. The system sent a notification to the operator: "To address grid fluctuation risks and maintain turbine health, output has been proactively reduced, switching to thermal storage mode." At night: Taking advantage of off-peak electricity prices and low load, the operation mode shifts to focus on equipment maintenance, maintaining low-stress operation, and may perform minor maintenance operations such as online cleaning cycles according to optimization instructions.

[0120] Throughout the process, the operator primarily acts as a supervisor. The system clearly displays the current health status, optimization goals, decision-making basis, and future predictions through a human-machine interface (HMI), and issues clear alarms and maintenance recommendations when significant intervention is required (such as a rapid drop in HI to the warning line).

[0121] See appendix Figure 2 A combined heat and power system integrating predictive maintenance and deep recovery is disclosed, comprising a first module, a second module, a third module, a fourth module and a fifth module; The first module acquires the operating process parameters, status monitoring parameters and external prediction data of the cogeneration system in real time. The second module constructs physical models of key components and uses real-time data to update model parameters online, generate real-time health indicators of the equipment, and predict its future trajectory. The third module uses a comprehensive function that includes a virtual health cost model based on the prediction of future changes in health indicators and short-term operating economic benefits as the objective function, and uses the equipment operating boundary that is dynamically related to health indicators as the constraint condition to solve the optimal setpoint sequence of control variables for generator and waste heat recovery system. The fourth module performs online correction and adjustment of the parameters and weights in the virtual health cost model based on the actual operation feedback of the cogeneration system. The fifth module sends the immediate instructions from the optimal setpoint sequence to the underlying execution mechanism.

[0122] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a combined heat and power method integrating predictive maintenance and deep recovery as described in any of the above methods.

[0123] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

Claims

1. A cogeneration method integrating predictive maintenance and deep recovery, characterized in that, Includes the following steps; Real-time acquisition of operating process parameters, status monitoring parameters and external prediction data of cogeneration systems; Construct physical models of key components and use real-time data to update model parameters online, generate real-time health indicators of the equipment, and predict its future trajectory. Using a comprehensive function that includes a virtual health cost model based on the prediction of future changes in health indicators and short-term operational economic benefits as the objective function, and the equipment operation boundary that is dynamically related to health indicators as the constraint, the optimal setpoint sequence of control variables for generator and waste heat recovery system is solved. The immediate instructions in the optimal setpoint sequence are sent to the underlying execution mechanism.

2. The cogeneration method integrating predictive maintenance and deep recovery as described in claim 1, characterized in that, After solving for the optimal setpoint sequence of control variables for the generator and waste heat recovery system, and before issuing the immediate instructions in the optimal setpoint sequence to the underlying actuators, the process also includes a step of online correction and adjustment of the parameters and weights in the virtual health cost model based on the actual operation feedback of the cogeneration system.

3. The cogeneration method integrating predictive maintenance and deep recovery as described in claim 2, characterized in that, The construction of physical models for key components includes building physical models for key components based on the first principles of thermodynamics and heat transfer; the online updating of model parameters using real-time data includes updating time-varying performance parameters in the model online through parameter estimation algorithms to generate real-time health indicators characterizing the health status of the equipment; the prediction of its future trajectory includes using a multi-scale time-series prediction model, which integrates a long short-term memory network and an exponential smoothing state-space model. The long short-term memory network is responsible for capturing the nonlinear long-term dependence between health indicators and high-dimensional operating conditions, while the exponential smoothing state-space model is responsible for modeling the local trends and seasonal fluctuations of health indicators. Finally, a weighted fusion layer outputs a comprehensive prediction of the health indicator degradation trajectory.

4. The cogeneration method integrating predictive maintenance and deep recovery as described in claim 3, characterized in that, The physical model includes a comprehensive efficiency factor physical model for turbochargers and a physical model for the overall heat transfer coefficient attenuation of cylinder liner water heat exchangers. The parameter estimation algorithm is the extended Kalman filter algorithm. The formula for the physical model of the turbocharger efficiency comprehensive factor is: ； Where, η ＿turbo＿eff This represents the overall efficiency factor of the turbocharger; Indicates intake mass flow rate; c P，air T represents the specific heat capacity of air at constant pressure. air，in Indicates the compressor inlet air temperature; T air，out This indicates the compressor outlet air temperature; Indicates exhaust mass flow rate; c P，exh T represents the isobaric specific heat capacity of exhaust gas at average temperature. exh，in T represents the turbine inlet exhaust temperature; exh，out，isen This represents the outlet temperature of the turbine after isentropic expansion. The formula for the physical model of the overall heat transfer coefficient decay in a cylinder liner water heat exchanger is: ； Among them U ＿fouling This indicates the overall heat transfer coefficient attenuation rate of the cylinder liner water heat exchanger. This indicates the mass flow rate of the cylinder liner water; T represents the isobaric specific heat capacity of cylinder liner water at average temperature; jw，out Indicates the outlet temperature of the cylinder liner water; T jw，in Indicates the inlet temperature of the cylinder liner water; A represents the heat exchange area; ΔT lm This represents the logarithmic mean temperature difference.

5. The cogeneration method integrating predictive maintenance and deep recovery as described in claim 4, characterized in that, Turbocharger Health Index (HI) turbo The calculation formula is: HI turbo =η ＿turbo＿eff,current / η ＿turbo＿eff,baseline ; Where η ＿turbo＿eff,current This is the optimal estimate of the turbocharger's overall efficiency factor, output in real time by the extended Kalman filter algorithm at the current moment. η ＿turbo＿eff,baseline The nominal value of the overall efficiency factor of a turbocharger when it is in a healthy, new condition and operating under standard or typical conditions; Cylinder liner water heat exchanger health index HI jw The calculation formula is: HI jw =U current / U baseline ; Where η ＿turbo＿eff,current η is the optimal estimate of the overall heat transfer coefficient of the cylinder liner water heat exchanger, output in real time by the extended Kalman filter algorithm at the current moment; ＿turbo＿eff,baseline This refers to the design value or initial calibration value of the overall heat transfer coefficient of the cylinder liner water heat exchanger under clean and foul-free conditions.

6. The cogeneration method integrating predictive maintenance and deep recovery as described in claim 5, characterized in that, The objective function is: ； Where R elec Represents electricity generation revenue; R heat For heating / cooling revenue; C fuel Indicates fuel cost; C om，var Indicates variable maintenance costs associated with operation; HI j，current This indicates the current real-time health indicator of component j; △HI j，pred (u) represents the predicted decrease in the health indicators of critical component j over a future period after strategy u is implemented; u represents the set of adjustable control commands, i.e., control variables; α j β represents the maintenance cost coefficient of component j; j λ represents the health status penalty factor; λ represents the weighting coefficient (hyperparameter) of health costs; j is the component index; Where the control variable u=[P] gen V bypass f pump1 f pump2 ..., SOC tes T set，cond ,...] T ; Where P gen V is the generator active power setpoint. bypass f is the opening degree of the flue gas bypass valve; pump1 f pump2 ... represents the frequency of the inverters for each major water pump; SOC tes The target state of charge of the thermal storage tank; T set，cond This is the setpoint for the condensation temperature.

7. A cogeneration method integrating predictive maintenance and deep recovery as described in claim 6, characterized in that, The constraints include static constraints and dynamic constraints. The static constraints include: 1) upper and lower limits of generator power; 2) physical limits of equipment; 3) grid / heat network demand balance; and 4) safe operation constraints. The dynamic constraint is a function of the sub-component's real-time health indicator HI: ; where g i (u) represents the constrained value of the control variable in the sub-component; The upper limit of the design value for this control variable when the equipment is new and in good health (HI=1); Φ i (HI relevant ) is a dynamic scaling function related to relevant health indicators.

8. The cogeneration method integrating predictive maintenance and deep recovery as described in claim 7, characterized in that, The optimal setpoint sequence of the control variables of the generator and waste heat recovery system is solved by sequential quadratic programming and rolling time domain. In each iteration, the sequential quadratic programming constructs a quadratic programming subproblem to approximate the original problem, and updates the iteration point by solving the quadratic programming subproblem until convergence. To handle forecast uncertainty and implement feedback control, a Model Predictive Control (MPC) framework is adopted: Forecast time domain: the next hour, divided into four 15-minute intervals; Control time domain: typically equal to or shorter than the forecast time domain; In each control cycle, the optimizer solves for the optimal control sequence u within the next forecast time domain. (t), t=1, ...,N, but only the control command u for the first time period is implemented. (1); In the next cycle, based on the new measurements, predictions and optimizations are performed again, and so on, rolling forward. The MPC framework naturally incorporates forecasts of future load, electricity prices, and future health status (HI). pred .

9. A cogeneration method integrating predictive maintenance and deep recovery as described in claim 8, characterized in that, Based on feedback from the actual operation of the cogeneration system, the parameters and weights in the virtual health cost model are corrected and adjusted online, including parameter adjustments: Data recording: Record the control variable u corresponding to each control decision, and the actual observed ΔHI within a certain period after execution. j，observed change; Model update: Periodically update the accumulated {u k ΔHI j，observed The data pair is used to update the parameters of the virtual health cost model, where k is the index of the control variable u; And the adaptive adjustment of the weighting coefficient λ: Set a health target range: Set a healthy operating range for health indicators; Monitor health indicator trends: Calculate the moving average and trend of health indicators; If the moving average of the health indicator is below the lower limit of the operating range and the trend is one of continuous deterioration, then increase λ. If the moving average of the health indicator is higher than the upper limit of the operating range and the trend is continuously improving, then decrease λ.

10. A combined heat and power system integrating predictive maintenance and deep recovery, comprising a first module, a second module, a third module, a fourth module and a fifth module; The first module acquires the operating process parameters, status monitoring parameters and external prediction data of the cogeneration system in real time. The second module constructs physical models of key components and uses real-time data to update model parameters online, generate real-time health indicators of the equipment, and predict its future trajectory. The third module uses a comprehensive function that includes a virtual health cost model based on the prediction of future changes in health indicators and short-term operating economic benefits as the objective function, and uses the equipment operating boundary that is dynamically related to health indicators as the constraint condition to solve the optimal setpoint sequence of control variables for generator and waste heat recovery system. The fourth module performs online correction and adjustment of the parameters and weights in the virtual health cost model based on the actual operation feedback of the cogeneration system. The fifth module sends the immediate instructions from the optimal setpoint sequence to the underlying execution mechanism.

Images