A steam turbine heat supply load optimization method based on adaptive variable working condition surface

Abstract

This invention relates to the field of thermal power engineering technology, and in particular to a method for optimizing the heating load of a steam turbine based on an adaptive variable operating condition surface. The method involves collecting main steam pressure and temperature, cooling water temperature, condenser vacuum, extraction steam flow rate, and exhaust steam parameters. Based on deviation and fuzzy C-means clustering, the system divides the turbine into three regions: design, sub-design, and deep variable operating condition. An RBF surface model is constructed, with Flueger constraints embedded in the sub-design region and Flueger, minimum flow rate, and exhaust enthalpy constraints embedded in the deep region. Membership degrees are truncated, normalized, and then weighted and fused to obtain the steam supply pressure setpoint. A simultaneous constraint coordination subproblem is constructed, and the correction value is solved in one step. The PID step size is adjusted according to the fused gradient magnitude. A dual-loop synchronous rolling correction and version locking are used to update the model and PID initial values. This invention improves the accuracy and safety of deep variable operating condition calculations and enhances thermal economy.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal power engineering, and specifically relates to a steam turbine heating load optimization method based on an adaptive off-design condition curved surface. Background Art

[0002] As an important form of efficient and cascade energy utilization, the combined heat and power generation technology enables a steam turbine to undertake power generation and heat supply tasks simultaneously, which significantly improves energy utilization efficiency. Extraction condensing steam turbines and back-pressure steam turbines are core equipment for realizing combined heat and power generation, and their operation efficiency highly depends on the reasonable matching between power generation output and heating load. In actual operation, the heat demand of heat users fluctuates significantly with seasons, day and night, and changes in external environmental conditions such as cooling water temperature and atmospheric pressure will cause the steam turbine to operate deviating from the designed working condition. Such off-design operation puts higher requirements on heating load optimization, and how to realize precise regulation of heating parameters under complex working conditions has become a key technical problem in the field of combined heat and power generation.

[0003] At present, the mainstream technical routes for steam turbine heating load optimization are divided into two categories. The first category is based on the Fleugel formula method, which establishes a quantitative relationship between flow rate and pressure by using the flow characteristics of steam in the flow passage part. This method has good accuracy only in the small-deviation linear region near the designed working condition. When the deviation amplitude of the working condition increases, its assumptions such as adiabatic flow and constant flow passage area are no longer valid, and the calculation accuracy drops sharply. The second category is based on the data-driven curved surface fitting method, which establishes the mapping relationship between working condition parameters and optimal steam supply pressure through machine learning methods such as neural networks and support vector machines. Although this type of method can cover a relatively wide range of working conditions, it has multiple defects as follows: the pure data-driven method lacks physical prior knowledge, and the output in the sparse working condition region may be physically infeasible; without embedding the first-principle constraints of thermodynamics, it may cause potential equipment safety hazards such as overheating of the last-stage blades; the fixed model parameters cannot track the drift of equipment characteristics caused by fouling, corrosion and the like in the flow passage part during long-term operation; the subsequent control link adopts a PID controller with fixed parameters, which cannot adapt to the drastic change of the curved surface gradient in different working condition regions, resulting in slow response in flat regions and easy overshoot and oscillation in steep regions.

[0004] The above defects seriously restrict the safe and economic operation of combined heat and power generation units under deep off-design conditions. The prior art still lacks a complete optimization method that can adaptively divide working condition partitions, organically embed thermodynamic constraints in curved surface fitting, realize the linkage adjustment of controller parameters and local characteristics of the curved surface, and have the capability of online updating of model parameters. In view of the above demands, the present invention proposes a steam turbine heating load optimization method based on an adaptive off-design condition curved surface. Contents of the Invention

[0005] The purpose of this invention is to provide a method and system for optimizing the heating load of a steam turbine based on an adaptive variable operating condition surface, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for optimizing the heating load of a steam turbine based on an adaptive variable operating condition surface, comprising the following steps: Step S1: Real-time acquisition of turbine operating condition input parameters, including main steam pressure, main steam temperature, cooling water inlet temperature, condenser vacuum, extraction steam flow rate, and exhaust steam parameters; Step S2: Based on the deviation between the current operating condition input parameters and the design operating condition, and combined with the cluster centers and cluster radii of each partition determined by fuzzy C-means clustering, the turbine operating area is dynamically divided into the effective domains of the design operating condition zone, the sub-design operating condition zone, and the deep variable operating condition zone. The effective domain of each partition is defined as within twice the cluster width from its cluster center. The deviation is not used as a hard partition boundary, but only to characterize the degree of deviation of the operating condition. Step S3: Construct data-driven surface models for the design operating condition zone, sub-design operating condition zone, and deep variable operating condition zone respectively. The surface model constructed for the sub-design operating condition zone embeds the Flueger flow conservation constraint, and the surface model constructed for the deep variable operating condition zone embeds the Flueger flow conservation constraint, the minimum flow constraint of the last stage blade, and the exhaust enthalpy constraint. Step S4: Calculate the membership degree of the current working condition to the design working condition area, sub-design working condition area and deep variable working condition area respectively. Perform truncation processing on the membership degree (set to zero if it is more than twice the width from the cluster center) and normalization. Weight and fuse the predicted steam supply pressure output by each partition surface model according to the normalized membership degree to obtain the comprehensive optimal steam supply pressure setting value. Step S5: Instead of sequentially checking the comprehensive optimal steam supply pressure setting value according to priority, a single-variable simultaneous coordination optimization subproblem is constructed with the objective of minimizing the sum of the squares of the Vlugel constraint deviation, the minimum flow constraint deviation of the last stage blade, and the exhaust enthalpy constraint deviation, and with the single correction magnitude as the constraint. The correction value that satisfies all constraints is solved at once, avoiding the problem of mutual destruction of constraints and incorrect correction direction caused by sequential correction. Step S6: Calculate the partial derivative vector of the weighted fused steam supply pressure setpoint with respect to the continuous operating condition input parameters as the surface gradient. This gradient is the weighted sum of the gradients of each partition model. Adjust the proportional coefficient of the PID controller dynamically according to the gradient amplitude, and execute the steam supply pressure control with the adjusted step size. Step S7: The parameters of the data-driven surface model are updated in a rolling manner with a first correction cycle, and the initial parameters of the PID controller are adaptively adjusted with a second correction cycle. The first correction cycle and the second correction cycle are synchronized and both are 5 minutes. When PID parameters are corrected, the model version number of the corresponding cycle is recorded, and the adjustment is made based on the performance index bound to the version to ensure the consistency between the benchmark of model update and PID parameter adjustment.

[0007] Preferably, the formula for calculating the deviation in step S2 is:

[0008] in, , , , , , These are the main steam pressure, main steam temperature, and cooling water inlet temperature under design conditions, respectively. The bias It is not used as a hard partition boundary, but only to characterize the degree of deviation from the operating condition; the actual effective scope of the partition is determined by the cluster center and cluster radius Ri obtained by fuzzy C-means clustering, and the effective scope of each partition is defined as the distance from the cluster center. The range.

[0009] Preferably, for the design operating condition region, a pure radial basis function neural network model is adopted, where the input layer neurons correspond to the operating condition input parameters, and the hidden layer adopts Gaussian radial basis functions; for the sub-design operating condition region, a model combining radial basis function neural network and Flüger constraint is adopted, and constraint embedding is achieved by adding a first penalty term to the loss function. The first penalty term is the absolute value of the difference between the Flüger formula calculated value and the network output value multiplied by a first penalty coefficient; for the deep variable operating condition region, a constraint optimization model is adopted, and based on the loss function of the sub-design operating condition region model, a second penalty term and a third penalty term are further added. The second penalty term is the squared penalty of the minimum flow constraint of the last stage blade, and the third penalty term is the squared penalty of the exhaust enthalpy constraint.

[0010] Preferably, in step S4, the membership degree is calculated using a Gaussian membership function:

[0011] in, The Euclidean distance from the current working condition input parameter vector to the i-th partition center vector is obtained by clustering historical working condition data using the fuzzy C-means clustering algorithm. Let be the cluster width parameter for the i-th partition; The Gaussian membership degree is calculated. Then, a truncation step is added:

[0012] Subsequently Normalization is performed: ,ensure

[0013] The weighted fusion calculation formula is as follows:

[0014] in, , , These are the predicted steam supply pressure values ​​output by the surface models for the design operating condition zone, sub-design operating condition zone, and deep variable operating condition zone, respectively.

[0015] Preferably, in step S5, the optimal steam supply pressure setting value is... Construct a constraint-coordinated optimization subproblem and complete the constraint correction in one go; The constraint deviation vector is:

[0016] in, The theoretical value is the value calculated using the Freuger formula. and The steam supply pressure is taken as follows under the current operating conditions. The corresponding exhaust steam flow rate and exhaust steam enthalpy; exist Solve the following single-variable constraint coordination problem in the vicinity:

[0017] in The maximum allowable pressure change for a single correction is set at 0.5 MPa to prevent over-adjustment. This embodiment uses the golden section method or parabolic interpolation method for rapid calculation to obtain the final correction value. The iteration termination condition is: ,in This represents the allowable deviation vector for each constraint.

[0018] Preferably, in step S6, the gradient direction of the surface model is defined as the weighted fusion combined steam supply pressure setpoint. Input parameters for continuous operation The partial derivative vector; given that Its gradient adopts the dominant simplification formula:

[0019] in, The values ​​were calculated separately for each partition's RBF model using the numerical difference method. gradient magnitude The calculation formula is:

[0020] in, For continuous variables in the operating condition input parameters, The total number of continuous variables, including main steam pressure, main steam temperature, cooling water inlet temperature, and condenser vacuum. Step size factor Based on gradient magnitude and preset gradient threshold The comparison results confirm: when hour, ;when hour, ,and ; The formula for updating the proportional gain of a PID controller is:

[0021] in, This is the proportional gain correction amount output by the fuzzy inference engine based on the current control deviation and the rate of change of the deviation. Adjust the weights for the proportional coefficient.

[0022] Preferably, the large step size factor The value is 0.8, which is the small step size factor. The value is set to 0.3; the integral and derivative coefficients of the PID controller are updated according to the following formulas:

[0023] in, and These are the adjustment weights for the integral coefficient and the derivative coefficient, respectively. and These are the correction values ​​for the integral and differential coefficients output by the fuzzy inference engine.

[0024] Preferably, in step S7, the first correction period is 5 minutes, and the most recently collected working condition data points are stored in a rolling training window with a capacity of 200 data points. When the window is full, the 50 data points with the farthest Euclidean distance from the current working condition center are removed, and the network parameters of each partition surface model are incrementally updated using the recursive least squares method. After each update, the system records the current model version number. ; The update formula for the recursive least squares method is:

[0025] in Let be the estimated vector of the output layer weights at time k. The regression vector is composed of the outputs of each radial basis function. This represents the actual observed value of the steam supply pressure at the current moment. It is a time-varying gain matrix; The second calibration cycle (PID parameter calibration loop) is uniformly adjusted to 5 minutes, strictly synchronized with the first calibration cycle; at each outer loop calibration time, the system statistically analyzes the control performance indicators of the previous complete cycle, including overshoot. Adjusting time and steady-state error It also records the version number of the REF model used during that period. When the overshoot is greater than 5% or the settling time is greater than 30 seconds, increase the initial value of the derivative parameter; when the steady-state error is greater than 0.05 MPa, increase the initial value of the integral parameter; the adjustment decision of the PID parameters is based solely on the... The performance metrics are version-bound, and each adjustment cannot exceed 20% of the current parameter value.

[0026] Preferably, in step S1, the data acquisition frequency is not less than 10Hz, and the acquired raw data is subjected to outlier removal and low-pass filtering smoothing processes in sequence. The main steam pressure is collected in the range of 0-18 MPa, the main steam temperature is collected in the range of 200℃-550℃, the cooling water inlet temperature is collected in the range of 5℃-40℃, the condenser vacuum degree is collected in the range corresponding to an absolute pressure of 0.005 MPa-0.1 MPa, and the extraction steam flow rate is measured by a standard orifice plate flow meter.

[0027] Preferably, when a change in the flow area of ​​the turbine's flow passage is detected due to scaling or corrosion, When changes occur, a switching logic with hysteresis is used: when When the value exceeds +5% for three consecutive sampling periods (sampling period of 1 minute), the constraint optimization model of the depth variable working condition zone is switched from the penalty function method to the safety domain constraint method. when When the percentage is below +3% for three consecutive sampling periods (or the absolute value is below 3%), it is permissible to switch back from the safety region constraint method to the penalty function method. Meanwhile, the flow area identification value Confidence assessment is required: If the linearity of the pressure-flow relationship under current operating conditions is... If the identification result has low confidence, then switching should be prohibited, the current optimization mode should be maintained, and the judgment should be re-evaluated after the working conditions stabilize. The safety domain constraint method includes: pre-establishing a safety domain boundary database for the safe operation of the last stage blades, which contains the minimum allowable exhaust flow rate and critical allowable exhaust enthalpy corresponding to different turbine speeds and different exhaust cylinder temperature combinations; formulating the training problem of the surface model as a constraint optimization problem with mean square error as the objective function and the safety domain boundary as an explicit inequality constraint, and solving it using a sequential quadratic programming algorithm.

[0028] Compared with the prior art, the beneficial effects of the present invention are: This invention divides the working space into a design working condition zone, a sub-design working condition zone, and a depth-variable working condition zone through a multi-working condition partitioning strategy. Differentiated surface fitting strategies are adopted for each zone, which effectively solves the problem of insufficient accuracy of pure data-driven methods in the depth-variable working condition zone, and the surface calculation error is significantly reduced. By introducing a truncated normalized membership degree fusion mechanism, the logical separation between hard partition boundaries and soft membership degrees is eliminated, ensuring that each partition model only participates in fusion within its own confidence radius, avoiding model jumps and constraint pollution problems at the boundaries. This invention embeds Flueger's law of flow conservation, minimum cooling flow rate constraint for the last-stage blades, and exhaust enthalpy constraint during the surface fitting process, ensuring that the surface output satisfies the first principles of thermodynamics and the conditions for safe equipment operation under any operating condition. Unlike traditional sequential verification and correction methods, this invention employs a simultaneous constraint coordination optimization strategy to solve for correction values ​​that satisfy all constraints in one go. This effectively avoids problems such as mutual constraint violation and correction directions that violate thermodynamic laws (e.g., incorrectly reducing steam supply pressure when flow is insufficient) that occur in sequential correction, significantly improving the convergence and safety of constraint processing. This invention calculates the gradient of the fused steam supply pressure setpoint (a weighted sum of the gradients of each partition model) and dynamically adjusts the PID proportional coefficient step size accordingly. In flat areas of the surface, a large step size is used to achieve rapid adjustment, while in steep areas, a small step size is used to achieve fine adjustment, effectively suppressing overshoot and significantly reducing the amount of adjustment overshoot. The gradient definition is strictly consistent with the final setpoint used for control, avoiding the problem of the controller response being out of sync with the model output. The dual closed-loop correction mechanism of this invention enables the surface model to track characteristic drift caused by equipment aging. The dual loops are strictly synchronized and a model version locking mechanism is introduced to ensure that the initial PID parameter adjustment is consistent with the current model version benchmark, avoiding the problem of mismatch between historical performance data and the current model, and maintaining the stability and traceability of long-term optimization effects. Under the premise of meeting safety constraints, this invention significantly reduces the heat consumption rate compared with the traditional fixed curve method; the multi-zone variable operating condition surface model can more accurately reflect the optimal operating point of the steam turbine under different operating conditions, making the heat load distribution more reasonable, thereby significantly improving the thermal economy of the cogeneration unit while ensuring equipment safety. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the steps according to the present invention; Figure 2 This is a schematic diagram illustrating the principle framework of surface model construction and membership degree weighted fusion according to the present invention; Figure 3 This is a schematic diagram of the process framework for thermodynamic constraint verification and gradient adaptive PID control according to the present invention. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” and “described” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the described features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0032] This invention provides a method for optimizing the heating load of a steam turbine based on an adaptive variable operating condition surface, comprising the following steps: Step S1: Real-time acquisition of turbine operating condition input parameters, including main steam pressure, main steam temperature, cooling water inlet temperature, condenser vacuum, extraction steam flow rate, and exhaust steam parameters. All parameters are acquired through a distributed data acquisition module deployed in the turbine control system. This module consists of multiple intelligent sensors located on the turbine body and key nodes of the thermal system, and maintains high-speed data communication with the central processing unit via an industrial fieldbus.

[0033] The main steam pressure is measured in the range of 0-18 MPa with a measurement accuracy of ±0.1% of full scale, using a high-precision pressure transmitter with temperature compensation. The main steam temperature is measured in the range of 200℃-550℃ with a measurement accuracy of ±0.5℃, using a sheathed thermocouple as the sensing element. The cooling water inlet temperature is measured in the range of 5℃-40℃ with a measurement accuracy of ±0.1℃, using a platinum resistance temperature sensor. The condenser vacuum level is measured in the range corresponding to an absolute pressure of 0.005 MPa-0.1 MPa. The extraction steam flow rate is measured using a standard orifice plate flow meter in conjunction with a differential pressure transmitter. Exhaust parameters include exhaust pressure, exhaust temperature, and exhaust humidity.

[0034] All data are collected at a frequency of no less than 10 times per second. Before entering the subsequent processing flow, the collected raw data undergoes outlier removal and low-pass filtering smoothing. Outlier removal uses a three-standard-deviation criterion based on statistical principles or a constraint criterion based on the physical feasible region to identify and mark abrupt data points; the filtering smoothing process uses a finite impulse response digital low-pass filter to suppress high-frequency noise components.

[0035] Step S2: Based on the deviation between the current operating condition input parameters and the turbine design operating condition, the entire operating range of the turbine is dynamically divided into three zones with different control strategies: the design operating condition zone, the sub-design operating condition zone, and the deep variable operating condition zone; in one embodiment, the formula for calculating the deviation is:

[0036] in, , , , , , These are the main steam pressure, main steam temperature, and cooling water inlet temperature under design conditions, respectively. The subscript "0" indicates the design condition parameter values ​​specified on the turbine nameplate or in the design documents.

[0037] In another embodiment, to improve the physical rationality of the partitioning criterion, the deviation is calculated using a weighted normalization formula:

[0038] Among them, For the j-th operating condition parameter, For the corresponding design operating conditions, Let j be the normalized baseline value of the j-th parameter. The weight coefficient of the j-th parameter and satisfying , This represents the number of operating parameters involved in the deviation calculation.

[0039] In the criteria for partition determination, deviation is... The calculation is still performed using the aforementioned formula, but it is not used as a hard partition boundary; it is only used to characterize the degree to which the operating condition deviates from the design point. The actual effective domain of each partition is determined by the cluster centers and cluster radii obtained from subsequent fuzzy C-means clustering. The effective scope of each partition is defined as follows: (i.e., within twice the width of the cluster center). In one embodiment, Take 10%, The 25% threshold is used only as a reference for initializing cluster centers and is not used as the final threshold for determining cluster affiliation.

[0040] Step S3: Construct a data-driven surface model for each partition and embed thermodynamic constraints. The data-driven surface model adopts a radial basis function neural network (RBF) model. An accurate digital simulation model of the turbine thermodynamic system is established based on an offline thermodynamic simulation platform, and a training sample set is generated by mesh sampling within each partition.

[0041] For the design operating condition region, a pure RBF model is adopted. Its input layer contains 5 neurons (corresponding to main steam pressure, main steam temperature, cooling water inlet temperature, condenser vacuum, and extraction steam flow rate, respectively), and the hidden layer uses Gaussian radial basis functions.

[0042] in, Let be the center vector of the Gaussian function of the i-th hidden layer neuron. This refers to the corresponding basis function width parameter. The basis function centers are determined by clustering historical operating data using the K-means clustering algorithm. The number of clusters is empirically set to 15 to 25, with the specific value optimized using cross-validation. The basis function width is calculated using the nearest neighbor distance method; that is, for each cluster center, the average distance to its nearest neighbor cluster center is calculated as the basis function width corresponding to that center.

[0043] For the sub-design load region, a hybrid surface model combining radial basis function neural networks and Frugell constraints is adopted. Its model structure is consistent with the design load region model; the difference lies in the objective function of the surface model for the sub-design load region:

[0044] in, Let i be the actual steam supply pressure value of the i-th training sample. This is the corresponding RBF network prediction value. This is the reference value of the steam supply pressure calculated using the Vlugel formula based on the current operating parameters. This is the first penalty coefficient (with a value of 10-50).

[0045] For the deep variable operating condition region, a constrained optimization surface model is adopted. This model further adds two penalty terms—the minimum flow constraint of the last-stage blades and the exhaust enthalpy constraint—to the objective function of the sub-design operating condition region. The extended objective function is expressed as:

[0046] in, This represents the actual exhaust steam flow rate. The minimum allowable exhaust steam flow rate, in t / h. For actual exhaust enthalpy, This is the critical exhaust enthalpy, expressed in kJ / kg. and The value range is 50-200.

[0047] Step S4: Calculate the membership degree of the current operating condition to the design operating condition zone, sub-design operating condition zone and deep variable operating condition zone respectively. Then, weight and fuse the predicted steam supply pressure values ​​output by the surface model of each zone according to the membership degree to obtain the comprehensive optimal steam supply pressure setpoint.

[0048] The membership function is a Gaussian membership function, and its mathematical expression is:

[0049] in, The Euclidean distance from the current operating condition input parameter vector to the center vector of the i-th partition is given. Let be the cluster width parameter for the i-th partition. This parameter is obtained by performing cluster analysis on historical operating data using an offline fuzzy C-means clustering algorithm. The cluster centers of each partition represent the distribution center points of typical operating conditions in that partition historically. The number of cluster centers is 15-25, and the width parameter... Take the average distance from each cluster center to its nearest neighbor cluster center.

[0050] The Gaussian membership degree is calculated. Then, a truncation step is added:

[0051] Subsequently Normalization is performed: ,ensure

[0052] The weighted fusion formula is:

[0053] in , , These are the predicted steam supply pressure values ​​output by the three partition surface models.

[0054] Step S5: Construct simultaneous constraints for the comprehensive optimal steam supply pressure setpoint. The constraint deviation vector is represented as follows:

[0055] in, and The steam supply pressure is taken as follows under the current operating conditions. The corresponding exhaust flow rate and exhaust enthalpy (obtained in real time through the thermodynamic calculation module).

[0056] exist Solve the following single-variable constraint coordination problem in the vicinity:

[0057] in, The maximum allowable pressure change for a single correction is set at 0.5 MPa to prevent over-adjustment. This embodiment uses the golden section method or parabolic interpolation method for rapid calculation to obtain the final correction value. .

[0058] The iteration termination condition is ( Each constraint allowable deviation vector corresponds to the original , , Maximum number of iterations Flow convergence tolerance enthalpy convergence tolerance .

[0059] Step S6: Calculate the gradient direction of the surface model at the current position, dynamically adjust the proportional coefficient of the PID controller according to the magnitude of the gradient direction, and execute steam supply pressure control with the adjusted step size.

[0060] The gradient direction of the surface model is defined as the weighted fusion combined steam supply pressure setpoint. Input parameters for continuous operation The vector of partial derivatives. Given... Its gradient should be a weighted sum of the gradients of the models in each partition. In practical engineering applications, to reduce computational complexity, the minor influence of membership degree variations with operating conditions is ignored, and a dominant simplification formula is adopted:

[0061] The calculation is performed separately by the RBF model of each partition using the numerical difference method, and the calculation formula is as follows:

[0062] in, The artificial disturbance amount for the j-th parameter is taken as 0.1%-1% of the parameter range in this embodiment.

[0063] The expression for the gradient magnitude is as follows:

[0064] in, These represent the four main continuously changing operating parameters: main steam pressure, main steam temperature, cooling water inlet temperature, and condenser vacuum.

[0065] In offline simulations or historical operating data, the gradient magnitude at each operating point is calculated, and the magnitude corresponding to the inflection point where the second derivative in the gradient magnitude distribution changes significantly is used as the gradient threshold. Alternatively, take the 70th percentile of the gradient magnitude at all operating points as the gradient threshold. In this embodiment, It is 0.8.

[0066] Step size factor Based on the gradient magnitude and the preset gradient threshold The comparison relationship between them is determined by segmentation. When When the current operating point is located in a relatively flat area of ​​the surface, the steam supply pressure setpoint is less sensitive to changes in input parameters. Therefore, a larger adjustment step size can be used to accelerate the response. (Step size factor setting) ;when When the current operating point is in a region of drastic surface change, the steam supply pressure setpoint is highly sensitive to changes in input parameters. To avoid over-adjustment and control oscillation, a smaller adjustment step size should be used to achieve fine adjustment. The step size factor is set to [value missing]. .

[0067] The PID controller is updated using an incremental adjustment strategy, and its update formula is as follows:

[0068] in, , and For the proportional, integral, and derivative parameters of the PID control at the current moment, , and For fuzzy inference engines based on control deviation and rate of change of deviation Real-time output of parameter correction values. The fuzzy inference engine adopts a dual-input, three-output structure, with input... and The universes of discourse are [-0.5, 0.5] MPa and [-0.1, 0.1] MPa / s, respectively, both divided into five fuzzy subsets: negative large, negative small, zero, positive small, and positive large. The membership function is a triangular function. The fuzzy rules are formulated according to the following principle: increase the membership function when the deviation is large. To improve response speed, the deviation decreases as it approaches zero. To prevent overshoot; increase when the rate of change of deviation is large. To enhance damping and suppress oscillations; to moderately increase damping when the deviation approaches zero. To eliminate steady-state error, when the deviation is large Zeros are used to avoid integral saturation. The centroid method is used for defuzzification. , and These are the weighting coefficients for the proportional, integral, and differential corrections, with values ​​of 0.6, 0.3, and 0.1, respectively.

[0069] Step S7: Construct two independent and strictly synchronized closed-loop correction mechanisms, which respectively update the parameters of the data-driven surface model in the first correction cycle and adaptively adjust the initial parameters of the PID controller in the second correction cycle.

[0070] The first inner loop is the surface model calibration loop, with a first calibration period of 5 minutes. The core data structure of the surface model calibration loop is a rolling training window with a capacity of 200 data points. Each data point contains a 5-dimensional working condition input vector at the current moment. , , , , The system collects the corresponding actual steam supply pressure observations and, at the end of each calibration cycle, appends newly acquired operating condition data points to the end of the training window. When the training window capacity reaches its limit, the system automatically calculates the Euclidean distance between all historical data points and the current operating condition center, discarding the 50 data points with the furthest distance to ensure that the training data always reflects the most recent operating condition distribution characteristics. Model parameters are updated incrementally using a recursive least squares method, with the following recursive formula:

[0071] in This is a parameter vector containing the center, width, and output weights of the RBF network. The time-varying gain matrix, This is the regression vector. After each update, the system records the current model version number. .

[0072] The second outer loop is the PID parameter calibration loop. The second calibration period is uniformly adjusted to 5 minutes, strictly synchronized with the first calibration period. The PID parameter calibration loop calculates the overshoot within the previous complete cycle (i.e., from the current moment back 5 minutes). Adjusting time and steady-state error Simultaneously record the RBF model version number used during that period. overshoot The calculation formula is:

[0073] in, This represents the maximum measured pressure during the step response process. Settling time. This refers to the period from the moment the step response occurs until the absolute value of the deviation enters and remains within the steady-state deviation range (±2%). Steady-state error The calculation formula is:

[0074] Based on the statistical results of the above performance indicators, the initial PID parameters are adaptively adjusted according to preset rules. If the overshoot exceeds 5% of the threshold or the settling time exceeds 30 seconds, it indicates that the derivative action of the current controller is insufficient, and the derivative parameter needs to be appropriately increased. The initial value is set to enhance system damping and suppress oscillations. If the steady-state error exceeds the threshold of 0.05 MPa, it indicates that the integral action of the current controller is insufficient, and the integral parameter needs to be appropriately increased. The initial values ​​are set to eliminate participation bias and improve control accuracy. The adjustment decisions for PID parameters are based solely on... Version-bound performance metrics are used to avoid mismatches between historical performance data and the current model due to model updates. Each adjustment should be limited to ±20%.

[0075] When a change in the flow area of ​​the turbine's flow passage is detected due to scaling or corrosion... When changes occur, a switching logic with hysteresis is used: when When the value exceeds +5% for three consecutive sampling periods (sampling period of 1 minute), the constraint optimization model of the depth variable working condition zone automatically switches from the penalty function method to the safety domain constraint method. when When the percentage is below +3% (or the absolute value is below 3%) for three consecutive sampling periods, it is permissible to switch back from the safety domain constraint method to the penalty function method.

[0076] Meanwhile, the flow area identification value Confidence assessment is required: If the linearity of the pressure-flow relationship under current operating conditions is... If the identification result has low confidence, then switching is prohibited, the current optimization mode is maintained, and the judgment is re-evaluated after the working conditions stabilize.

[0077] The safety domain constraint method includes: pre-establishing a safety domain boundary database for the safe operation of the last-stage blades, which contains the minimum allowable exhaust flow rate and critical allowable exhaust enthalpy corresponding to different turbine speeds and exhaust cylinder temperature combinations; formulating the training problem of the surface model as a constraint optimization problem with mean square error as the objective function and the safety domain boundary as an explicit inequality constraint, and solving it using a sequential quadratic programming algorithm. In one embodiment, the preset threshold +5% is only used for forward switching determination, and the back-switching is based on +3%.

[0078] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing the heating load of a steam turbine based on an adaptive variable operating condition surface, characterized in that, Includes the following steps: Step S1: Real-time acquisition of turbine operating condition input parameters, including main steam pressure, main steam temperature, cooling water inlet temperature, condenser vacuum, extraction steam flow rate, and exhaust steam parameters; Step S2: Based on the deviation between the current operating condition input parameters and the design operating condition, and combined with the cluster centers and cluster radii of each partition determined by fuzzy C-means clustering, the turbine operating area is dynamically divided into the effective domains of the design operating condition zone, the sub-design operating condition zone, and the deep variable operating condition zone. The effective domain of each partition is defined as within twice the cluster width from its cluster center. Step S3: Construct data-driven surface models for the design operating condition zone, sub-design operating condition zone, and deep variable operating condition zone respectively. The surface model constructed for the sub-design operating condition zone embeds the Flueger flow conservation constraint, and the surface model constructed for the deep variable operating condition zone embeds the Flueger flow conservation constraint, the minimum flow constraint of the last stage blade, and the exhaust enthalpy constraint. Step S4: Calculate the membership degree of the current working condition to the design working condition area, sub-design working condition area and deep variable working condition area respectively. Perform truncation processing on the membership degree (set to zero if it is more than twice the width from the cluster center) and normalization. Weight and fuse the predicted steam supply pressure output by each partition surface model according to the normalized membership degree to obtain the comprehensive optimal steam supply pressure setting value. Step S5: Instead of sequentially checking the comprehensive optimal steam supply pressure setting value according to priority, a single-variable simultaneous coordination optimization subproblem is constructed with the objective of minimizing the sum of the squares of the Vlugel constraint deviation, the minimum flow constraint deviation of the last stage blade, and the exhaust enthalpy constraint deviation, and with the single correction magnitude as the constraint. The correction value that satisfies all constraints is solved in one go. Step S6: Calculate the partial derivative vector of the weighted fused steam supply pressure setpoint with respect to the continuous operating condition input parameters as the surface gradient. This gradient is the weighted sum of the gradients of each partition model. Adjust the proportional coefficient of the PID controller dynamically according to the gradient amplitude, and execute the steam supply pressure control with the adjusted step size. Step S7: The parameters of the data-driven surface model are updated in a rolling manner with a first correction cycle, and the initial parameters of the PID controller are adaptively adjusted with a second correction cycle. The first correction cycle and the second correction cycle are synchronized and both are 5 minutes. When the PID parameters are corrected, the model version number of the corresponding cycle is recorded, and the adjustment is made based on the performance indicators bound to the version.

2. The method for optimizing turbine heating load based on adaptive variable operating condition surface according to claim 1, characterized in that, The formula for calculating the deviation in step S2 is as follows: ； in, , , , , , These are the main steam pressure, main steam temperature, and cooling water inlet temperature under design conditions, respectively. The deviation It is not used as a hard partition boundary, but only to characterize the degree of deviation from the operating condition; the actual effective scope of the partition is determined by the cluster center and cluster radius Ri obtained by fuzzy C-means clustering, and the effective scope of each partition is defined as the distance from the cluster center. The range.

3. The method for optimizing turbine heating load based on adaptive variable operating condition surface according to claim 1, characterized in that, For the design operating condition region, a pure radial basis function neural network model is adopted, with its input layer neurons corresponding to the operating condition input parameters and the hidden layer using Gaussian radial basis functions. For the sub-design operating condition region, a model combining radial basis function neural networks and Flüger constraints is adopted. Constraint embedding is achieved by adding a first penalty term to the loss function. The first penalty term is the absolute value of the difference between the Flüger formula calculated value and the network output value multiplied by a first penalty coefficient. For the deep variable operating condition region, a constraint optimization model is adopted. Based on the loss function of the sub-design operating condition region model, a second penalty term and a third penalty term are further added. The second penalty term is the squared penalty of the minimum flow constraint of the last stage blade, and the third penalty term is the squared penalty of the exhaust enthalpy constraint.

4. The method for optimizing turbine heating load based on adaptive variable operating condition surface according to claim 1, characterized in that, In step S4, the membership degree is calculated using a Gaussian membership function. ； in, The Euclidean distance from the current working condition input parameter vector to the i-th partition center vector is obtained by clustering historical working condition data using the fuzzy C-means clustering algorithm. Let be the cluster width parameter for the i-th partition; The Gaussian membership degree is calculated. Then, a truncation step is added: ； Subsequently Normalization is performed: ,ensure ; The weighted fusion calculation formula is as follows: ； in, , , These are the predicted steam supply pressure values ​​output by the surface models for the design operating condition zone, sub-design operating condition zone, and deep variable operating condition zone, respectively.

5. The method for optimizing turbine heating load based on adaptive variable operating condition surface according to claim 1, characterized in that, In step S5, the optimal steam supply pressure setting value is determined. Construct a simultaneous constraint coordination optimization subproblem and complete the constraint correction in one go; The constraint deviation vector is: ； in, The theoretical value is the value calculated using the Freuger formula. and The steam supply pressure is taken as follows under the current operating conditions. The corresponding exhaust steam flow rate and exhaust steam enthalpy; exist Solve the following single-variable constraint coordination problem in the vicinity: ； in This represents the maximum allowable pressure change for a single correction (taken as 0.5 MPa). The final correction value is obtained by quickly solving using the golden section method or parabolic interpolation method. The iteration termination condition is: ,in This represents the allowable deviation vector for each constraint.

6. The method for optimizing turbine heating load based on adaptive variable operating condition surface according to claim 1, characterized in that, In step S6, the gradient direction of the surface model is defined as the weighted fusion combined steam supply pressure setpoint. Input parameters for continuous operation The partial derivative vector; given that Its gradient adopts the dominant simplification formula: ； in, The calculations were performed separately for each partition's RBF model using the numerical difference method. gradient magnitude The calculation formula is: ； in, For continuous variables in the operating condition input parameters, The total number of continuous variables includes main steam pressure, main steam temperature, cooling water inlet temperature, and condenser vacuum; the condenser vacuum is collected in the range of absolute pressure 0.005MPa-0.1MPa. Step size factor Based on gradient magnitude and preset gradient threshold The comparison results confirm: when hour, ;when hour, ,and ; The formula for updating the proportional gain of a PID controller is: ； in, This is the proportional gain correction amount output by the fuzzy inference engine based on the current control deviation and the rate of change of the deviation. Adjust the weights for the proportional coefficient.

7. The method for optimizing turbine heating load based on adaptive variable operating condition surface according to claim 6, characterized in that, The large step size factor The value is 0.8, which is the small step size factor. The value is set to 0.3; the integral and derivative coefficients of the PID controller are updated according to the following formulas: ； in, and These are the adjustment weights for the integral coefficient and the derivative coefficient, respectively. and These are the correction values ​​for the integral and differential coefficients output by the fuzzy inference engine.

8. The method for optimizing turbine heating load based on adaptive variable operating condition surface according to claim 1, characterized in that, In step S7, the first correction period is 5 minutes. A rolling training window stores the most recently collected operating condition data points. The rolling training window has a capacity of 200 data points. When the window is full, the 50 data points with the furthest Euclidean distance from the current operating condition center are removed. The recursive least squares method is then used to incrementally update the network parameters of each partition surface model. After each update, the system records the current model version number. ; The update formula for the recursive least squares method is: ； in Let be the estimated vector of the output layer weights at time k. The regression vector is composed of the outputs of each radial basis function. This represents the actual observed value of the steam supply pressure at the current moment. It is a time-varying gain matrix; The second calibration cycle (PID parameter calibration loop) is uniformly adjusted to 5 minutes, strictly synchronized with the first calibration cycle; at each outer loop calibration time, the system statistically analyzes the control performance indicators of the previous complete cycle, including overshoot. Adjusting time and steady-state error It also records the version number of the RBF model used during that period. When the overshoot is greater than 5% or the settling time is greater than 30 seconds, increase the initial value of the derivative parameter; when the steady-state error is greater than 0.05 MPa, increase the initial value of the integral parameter; the adjustment decision of the PID parameters is based solely on the... The performance metrics are version-bound, and each adjustment cannot exceed 20% of the current parameter value.

9. The method for optimizing turbine heating load based on adaptive variable operating condition surface according to claim 1, characterized in that, In step S1, the data acquisition frequency is not less than 10Hz, and the acquired raw data is sequentially subjected to outlier removal and low-pass filtering smoothing. The main steam pressure is collected in the range of 0-18 MPa, the main steam temperature is collected in the range of 200℃-550℃, the cooling water inlet temperature is collected in the range of 5℃-40℃, the condenser vacuum degree is collected in the range corresponding to an absolute pressure of 0.005 MPa-0.1 MPa, and the extraction steam flow rate is measured by a standard orifice plate flow meter.

10. The method for optimizing turbine heating load based on adaptive variable operating condition surface according to claim 5, characterized in that, When a change in the flow area of ​​the turbine's flow passage is detected due to scaling or corrosion... When changes occur, a switching logic with hysteresis is used: when When the value exceeds +5% for three consecutive sampling periods (sampling period of 1 minute), the constraint optimization model of the depth variable working condition zone is switched from the penalty function method to the safety domain constraint method. when When the percentage is below +3% for three consecutive sampling periods (or the absolute value is below 3%), it is permissible to switch back from the safety region constraint method to the penalty function method. Meanwhile, the flow area identification value Confidence assessment is required: If the linearity of the pressure-flow relationship under current operating conditions is... If the identification result has low confidence, then switching should be prohibited, the current optimization mode should be maintained, and the judgment should be re-evaluated after the working conditions stabilize. The safety domain constraint method includes: pre-establishing a safety domain boundary database for the safe operation of the last stage blades, which contains the minimum allowable exhaust flow rate and critical allowable exhaust enthalpy corresponding to different combinations of turbine speed and exhaust cylinder temperature; The training problem of the surface model is formulated as a constrained optimization problem with mean square error as the objective function and the boundary of the security region as an explicit inequality constraint, and solved using a sequential quadratic programming algorithm.

Images