Methods for Low-Load, Full-Condition Coupling Process of Ultra-High Pressure Coal-Fired Thermal Power Generator Units
By constructing a third-order inertial link transfer function matrix model with pure delay, and combining chaotic quantum sparrow search and hierarchical Bayesian parameter identification, the modeling problem of the boiler-machine coupling process in the low-load section of ultra-high pressure coal-fired heating and power generation units was solved, achieving efficient modeling and stable operation under all operating conditions.
Patent Information
- Application Number
- CN202611218492.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies make it difficult to model the boiler-generator coupling process under all operating conditions in the low-load section of ultra-high pressure coal-fired heating and power generating units, which limits the improvement of the unit's load change rate and affects peak-shaving capacity and heating quality.
A transfer function matrix model is constructed using a third-order inertial element with pure delay. The pure delay element is simplified by combining the first-order Pad approximation. The chaotic quantum sparrow search algorithm is used for offline parameter identification. A hierarchical Bayesian parameter online identification mechanism is constructed by combining operator evolution theory and polynomial matrix theory. An automatic matching mechanism for model structure and adaptive parameter update mechanism for multiple load intervals is designed.
It achieves high-precision modeling within the 20%-100% full load range, improving the model's adaptability and real-time performance, and supporting the improvement of unit load change rate and operational stability.
Smart Images

Figure CN123113501A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of coal-fired power generation unit operation control technology, specifically a method for the low-load full-condition turbine-boiler coupling process of ultra-high pressure coal-fired heating power generation units. Background Technology
[0002] During the construction of the new power system, the proportion of new energy power generation such as wind power and photovoltaic power continues to increase, and coal-fired power generating units need to undertake deep peak-shaving tasks to ensure the safe and stable operation of the power grid. Ultra-high voltage coal-fired cogeneration units simultaneously undertake the dual functions of power supply and heating, and their operating conditions cover a wide range, requiring stable operation within the 20%-100% rated load range. The boiler-turbine coupling process is the core dynamic process of coal-fired unit operation. The combustion and steam-water heat exchange processes on the boiler side are coupled with the energy conversion and steam extraction heating processes on the turbine side, exhibiting characteristics of large inertia, large delay, and strong coupling of multiple variables. Especially at low loads, the boiler combustion stability decreases, the steam-water flow characteristics change, and the nonlinear and time-varying characteristics of the boiler-turbine coupling are significantly enhanced.
[0003] Current modeling methods for boiler-turbine coupling processes are mainly divided into two categories: mechanistic modeling and data-driven identification. Mechanistic modeling is based on the equations of mass conservation, energy conservation, and momentum conservation, and constructs a set of partial differential equations for the boiler steam-water system and combustion system. It can describe the process characteristics from the physical essence, but the model structure is complex, has many parameters, and requires a huge amount of computation, making it difficult to apply directly to real-time control. Moreover, factors such as changes in the physical properties of the working fluid and combustion instability under low load can lead to a significant decrease in the accuracy of the mechanistic model, making it unable to adapt to the requirements of full-condition operation.
[0004] Data-driven identification methods rely on historical operational data to build process models, requiring no in-depth understanding of internal physical mechanisms and thus having lower engineering implementation difficulty. Traditional least squares identification and gradient correction identification methods are only applicable to steady-state or slowly changing operating conditions, and their identification accuracy is insufficient in rapidly changing loads and low-load strongly nonlinear scenarios. Conventional intelligent optimization algorithms such as genetic algorithms and particle swarm optimization can handle nonlinear identification problems, but they suffer from slow convergence speeds and a tendency to get trapped in local optima. Offline identification is time-consuming, and online identification cannot meet control requirements in terms of real-time performance.
[0005] Furthermore, most existing modeling methods are only designed for conventional load segments of over 40%, lacking adaptability to deep peak-shaving load segments of 20%-40%, and lacking a unified modeling and parameter adaptation mechanism for all operating conditions. The lack of model accuracy in low-load segments directly limits the improvement of unit load change rate, affecting the unit's peak-shaving capacity and heating quality. Summary of the Invention
[0006] The purpose of this invention is to provide a method for the low-load, full-condition turbine-boiler coupling process of an ultra-high pressure coal-fired thermal power generator set, so as to solve the problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for the coupling process of a boiler and turbine under low load conditions in an ultra-high pressure coal-fired thermal power generation unit, comprising the following steps:
[0008] Step 1: Collect historical operating data of the ultra-high pressure coal-fired heating and power generation unit during the 20%-100% full load range. After preprocessing the data, analyze the dynamic characteristics of the boiler steam-water power process, screen the key input and output variables that affect the stability and economy of the unit under low load operation, and construct the transfer function matrix model structure of the boiler-machine coupling process.
[0009] Step 2: Using the mean square error function between the actual values of multiple output variables and the model output values as the optimization objective, the chaotic quantum sparrow search algorithm is used to identify the structural parameters of the transfer function matrix model offline based on the preprocessed historical data, and the parameters obtained offline are used as the initial values for online identification under each typical working condition.
[0010] Step 3: Based on the model structure and initial parameter values obtained offline, an online parameter identification mechanism based on hierarchical Bayesian estimation is constructed using operator evolution theory and multinomial matrix theory to quickly identify the model parameters online under dynamic conditions.
[0011] Step 4: Design an automatic matching and parameter adaptive update program for the model structure under different steady-state operating conditions to achieve adaptive adjustment of the operating conditions of the boiler-turbine coupling process model under full load.
[0012] Step 5: For the furnace-machine coupling process with different input and output dimensions, computer numerical simulation and simulator experiments are carried out in sequence to verify the effectiveness of the method.
[0013] Furthermore, in step one, each element of the transfer function matrix model is a third-order inertial element with a pure delay, and its Laplace domain transfer function expression is:
[0014]
[0015] In the formula, The channel transfer function from the i-th input variable to the j-th output variable; This is the static gain of the channel; , , These are the three time constants of the third-order inertial element; This is the pure delay time of the channel; For the Laplace operator.
[0016] Furthermore, in step one, the pure delay element in the transfer function is equivalently transformed using the first-order Pad approximation method. The expression for the pure delay element after transformation is:
[0017]
[0018] After transformation, the transfer function unit is equivalent to a fourth-order inertial element with zeros, which simplifies subsequent parameter identification and calculation.
[0019] Furthermore, in step two, the expression for the objective function for optimizing the mean square error is:
[0020]
[0021] In the formula, This represents the mean square error value. This represents the total number of sampling points for historical data. This represents the total number of output variables in the furnace-machine coupling process. This represents the actual running value of the m-th output variable at the k-th sampling time. This represents the model prediction value of the m-th output variable at the k-th sampling time.
[0022] Furthermore, the data preprocessing in step one specifically includes: using a low-order polynomial sliding fitting method for zero initial value processing, using a soft thresholding method for denoising the running data, and using a low-order difference method for removing coarse values; wherein the low-order polynomial sliding fitting uses a third-order polynomial, and selects 5 sampling points before and after the current sampling point to form a sliding window, and fits to obtain the correction value of the current point.
[0023] Furthermore, the offline identification of the chaotic quantum sparrow search algorithm in step two specifically includes:
[0024] S21: The position vectors of the sparrow population are initialized using Tent chaotic mapping, and each position vector corresponds to a set of parameters to be identified in the transfer function matrix;
[0025] S22: Using the mean squared error function as the fitness function, calculate the fitness value of each individual sparrow and classify them into discoverers, followers and watchers according to their fitness.
[0026] S23: The position of the discoverer is updated iteratively using a quantum behavior update method, the position of the follower is updated using a dimension-by-dimensional following strategy, and the position of the vigilant is updated using a random walk strategy.
[0027] S24: Introduce a chaotic perturbation every 10 iterations to perform chaotic mutation on the current optimal solution. If the fitness is better after mutation, replace the original optimal solution.
[0028] S25: When the maximum number of iterations is reached or the change in fitness is less than the set threshold, the iteration is terminated and the optimal parameter set is output as the offline identification result.
[0029] Furthermore, in step three, a hierarchical Bayesian parameter identification mechanism is constructed based on the inverse form of a polynomial matrix, and the expression for the posterior estimation of the parameters is:
[0030]
[0031] In the formula, is the posterior estimation vector of the parameter to be identified; The information matrix is composed of input and output data; The output variable is the observation vector; is the regularization coefficient, derived from the hierarchical Bayesian prior distribution; Let be the identity matrix; the above equation can be solved quickly using a polynomial matrix inversion algorithm.
[0032] Furthermore, in step three, the application of operator evolution theory is specifically as follows: for dynamic variable load conditions, differential operators are used to perform time-domain transformation on the transfer function matrix, transforming the continuous-time model into a discrete-time operator form, so that the model parameters can continuously evolve with the changes in operating conditions, adapting to the parameter identification requirements of dynamic processes.
[0033] Furthermore, the automatic model structure matching and adaptive parameter update in step four specifically includes: pre-dividing the load into five load intervals: 20%-30%, 30%-40%, 40%-60%, 60%-80%, and 80%-100%, with each interval corresponding to a set of baseline model parameters; real-time acquisition of unit load values; and triggering model parameter switching when the load crosses the interval boundary and the duration exceeds the set time. The switching process uses a smooth transition algorithm to achieve disturbance-free parameter switching.
[0034] Furthermore, the different input and output dimensions in step five include three typical boiler-machine coupling scenarios: two-input single-output, two-input two-output, and three-input three-output. The input variables include boiler fuel quantity, feedwater flow rate, and air supply flow rate, while the output variables include main steam pressure, main steam temperature, and unit active power.
[0035] Compared with the prior art, the beneficial effects of the present invention are:
[0036] This invention uses a third-order inertial element with pure delay to construct a multivariable transfer function matrix, and combines the first-order Pad approximation to simplify the pure delay element. This not only accurately characterizes the large inertia, large delay, and multi-capacity dynamic characteristics of the boiler-generator coupling process of ultra-high pressure coal-fired heating units, but also reduces the computational complexity of parameter identification, and can adapt to the modeling requirements of 20%-100% full load range.
[0037] This invention employs a chaotic quantum sparrow search algorithm for offline parameter identification. It enhances population diversity through chaotic mapping initialization, strengthens global optimization capabilities by combining quantum behavior updates, and introduces periodic chaotic perturbations to escape local optima. This effectively improves the offline identification accuracy and convergence efficiency of multivariate strongly coupled models and reduces the offline modeling time.
[0038] This invention combines operator evolution theory with multinomial matrix theory to construct a hierarchical Bayesian parameter online identification mechanism. It achieves rapid updating of model parameters under dynamic conditions through algebraic solution, which significantly improves the online identification speed compared with traditional iterative identification methods. It can adapt to rapidly changing load conditions and ensure the model tracking accuracy of dynamic processes.
[0039] This invention designs an automatic model structure matching and parameter-free switching mechanism for multiple load ranges, covering the entire operating range from 20% deep peak load to 100% rated load. By using hysteresis logic and smooth transition algorithm, it avoids frequent model switching and parameter jumps, effectively improving the adaptability of the full load range model and supporting the improvement of unit load change rate and operational stability. Attached Figure Description
[0040] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0042] Please see Figure 1 This invention provides a method for the boiler-turbine coupling process of an ultra-high pressure coal-fired thermal power generation unit under low load and full operating conditions, comprising the following steps:
[0043] Step 1: Run data acquisition and preprocessing
[0044] 1.1 Data Acquisition for Full-Condition Operation
[0045] Historical operating data of ultra-high pressure coal-fired thermal power generating units were collected within the 20%-100% full load range. The data sampling period was consistent with the inherent sampling period of the unit's distributed control system to ensure that the data time resolution met the requirements for dynamic characteristic identification. The range of variables collected covered core operating parameters on the boiler side, turbine side, and heating side, specifically including boiler fuel total command, feedwater pump outlet flow rate, total air supply flow rate, primary air header pressure, secondary air box pressure, furnace outlet flue gas temperature, water-cooled wall outlet temperature, steam drum water level, main steam pressure, main steam temperature, reheat steam temperature, turbine regulating stage pressure, high-pressure cylinder exhaust pressure, unit active power, heating extraction steam pressure, and heating extraction steam flow rate.
[0046] The data collection period covers a complete operating cycle of at least 12 months, including operating samples under different seasons and heating loads; the operating conditions cover typical steady-state load points, dynamic processes of load increase, dynamic processes of load decrease, heating steam extraction switching processes, and conventional disturbance conditions. Through multi-condition, long-term data collection, the dataset ensures that it can comprehensively reflect the dynamic characteristics of the boiler-turbine coupling across the entire load range, avoiding insufficient model generalization ability caused by single operating conditions.
[0047] 1.2 Multi-level data preprocessing
[0048] To address common issues in raw operational data such as missing zero values, random measurement noise, and gross error interference, a three-stage preprocessing process is implemented in the order of "zero value repair, noise suppression, and anomaly removal" to ensure the quality of the modeling data. The specific process is as follows:
[0049] (1) Zero initial value and missing value repair: The low-order polynomial sliding fitting method is used to repair zero values, breakpoints and missing points in the data. A third-order polynomial is selected as the fitting basis function. Taking the sampling point with missing points as the center, five valid sampling points before and after it are selected to form a sliding window of length 11. Based on the valid data in the window, the coefficients of the third-order polynomial are obtained by least squares fitting. The time corresponding to the missing point is substituted into the polynomial equation to calculate the repair value of the point. For segments with no more than five consecutive missing points, this method can achieve high-precision repair while preserving the dynamic trend of the data and avoid the interference of data breakage on subsequent identification results.
[0050] (2) Random Noise Suppression: The wavelet soft thresholding method is used to suppress high-frequency random measurement noise in the running data. First, the Daubechies wavelet basis is selected to perform a 5-level wavelet decomposition on the original data sequence to obtain the approximation coefficients and detail coefficients at each scale; then, the threshold of each level of detail coefficients is calculated using a general threshold criterion. Detail coefficients with absolute values less than the threshold are set to zero, and detail coefficients with absolute values greater than the threshold are shrunk to the corresponding threshold size towards zero; finally, the processed coefficients are reconstructed using wavelet to obtain the denoised running data sequence. Soft thresholding can avoid signal oscillation and pseudo-Gibbs phenomenon caused by hard thresholding, and retain the dynamic response characteristics of the process while removing noise.
[0051] (3) Gross Error Removal: A low-order difference method combined with the 3σ criterion is used to identify and remove gross errors in the data. First-order difference calculations are performed on the data sequence to obtain the difference sequence between adjacent sampling points. The mean and standard deviation of the difference sequence are statistically analyzed, and an anomaly detection threshold is set as the mean plus three times the standard deviation. If the difference value corresponding to a sampling point exceeds this threshold, the point is determined to be a gross error point. For sampling points determined to be anomalous, the linear interpolation result of the two adjacent valid sampling points is used to replace them. This method can quickly identify abnormal data caused by measurement faults, signal interference, equipment impact, etc., ensuring the reliability of the modeling dataset.
[0052] Step 2: Construction of the furnace-machine coupling process model structure
[0053] 2.1 Key Variable Screening
[0054] Based on preprocessed historical operating data, we conducted an in-depth analysis of the dynamic characteristics of the steam-water power process and combustion system inside the boiler, sorted out the coupling relationship between fuel combustion, heat transfer, steam-water phase change, steam work, and steam extraction for heating, and clarified the inherent action path of the boiler-machine coupling.
[0055] The mutual information method is used to quantify the degree of nonlinear correlation between input and output variables; a larger mutual information value indicates a stronger coupling correlation between variables. By calculating the mutual information values of all candidate input-output variable pairs, a correlation threshold is set, and variables with mutual information values higher than the threshold are selected as key variables in the furnace-machine coupling model.
[0056] Ultimately, the core input variables for typical operating conditions were determined to be boiler fuel quantity, feedwater flow rate, and air supply flow rate, while the core output variables were main steam pressure, main steam temperature, and unit active power. For operating conditions where heating steam extraction is activated, the heating steam extraction flow rate was used as an additional input variable, and the heating steam extraction pressure was used as an additional output variable, thus constructing an extended coupled model for heating operating conditions.
[0057] By filtering out redundant and weakly coupled variables, the model structure can be effectively simplified, the computational dimension and workload of subsequent parameter identification can be reduced, and the model can accurately reflect the core dynamic characteristics of the furnace-machine coupling.
[0058] 2.2 Multivariable Transfer Function Matrix Model
[0059] The multivariable machine-furnace coupling process is described using a transfer function matrix. The dimension of the transfer function matrix is determined by the number of input and output variables selected. For a three-input, three-output system, the transfer function matrix is a 3×3 matrix. Each element in the matrix corresponds to a channel transfer function from one input variable to one output variable, representing the dynamic relationship between the input and the output. The transfer function of each channel adopts the form of a third-order inertial element with a pure delay element, and its Laplace domain expression is shown in equation (1):
[0060] (1)
[0061] In the formula, The channel transfer function from the i-th input variable to the j-th output variable; This is the static gain of the channel; , , These are the three time constants of the third-order inertial element; This is the pure delay time of the channel; For the Laplace operator.
[0062] The boiler-machine coupling process of a coal-fired unit involves multiple energy storage links, such as furnace heat storage, water-cooled wall heat storage, steam drum heat storage, and superheater heat storage. The third-order inertial link can fit the cascade response characteristics of multiple capacity links well. Combined with the pure delay link to characterize the hysteresis effect brought about by fuel delivery, heat transfer, and working fluid flow, it can accurately fit the dynamic response of the process under the premise of an appropriate number of parameters, taking into account both model accuracy and identification efficiency.
[0063] 2.3 First-order Pad approximation for pure delay elements
[0064] The pure delay element is an exponential non-minimum phase element, and its parameters cannot be directly solved using conventional linear identification methods, which would increase the complexity of the identification algorithm. To simplify the subsequent parameter identification calculation, the first-order Pade approximation method is used to perform a rational polynomial equivalent transformation on the pure delay element. The first-order Pade approximation uses Taylor expansion to transform the exponential function into a rational fractional form, and its expression is shown in equation (2):
[0065] (2)
[0066] Substituting equation (2) into equation (1) and rearranging it, we can obtain the equivalent transfer function form, whose numerator is a first-order polynomial and whose denominator is a fourth-order polynomial, which is equivalent to a fourth-order inertial element containing a zero.
[0067] The first-order Pad approximation has sufficient engineering accuracy in thermal processes with a moderate ratio of delay time to time constant. It can transform pure delay elements into combinations of inertial elements without significantly increasing the number of parameters, adapting to the solution framework of traditional system identification algorithms and greatly reducing the difficulty of subsequent identification.
[0068] Step 3: Offline parameter identification based on the chaotic quantum sparrow search algorithm
[0069] 3.1 Identifying and optimizing the objective function
[0070] The mean square error between the actual running values of multiple output variables and the model output values is used as the performance optimization objective of model identification, i.e., the fitness function of the optimization algorithm. The expression of the objective function is shown in equation (3):
[0071] (3)
[0072] In the formula, This represents the mean square error value. This represents the total number of sampling points for historical data. This represents the total number of output variables in the furnace-machine coupling process. This represents the actual running value of the m-th output variable at the k-th sampling time. The model prediction value of the m-th output variable at the k-th sampling time is obtained by numerical simulation calculation from the transfer function matrix model under the current parameter group.
[0073] The optimization objective of parameter identification is to find an optimal set of model parameters that minimizes the value of the objective function J, i.e., minimizes the overall deviation between the model output sequence and the actual running sequence. This objective function comprehensively considers the fitting accuracy of all output channels, ensuring the overall identification effect of the multivariate coupled model.
[0074] 3.2 Chaotic Quantum Sparrow Search Algorithm Identification Process
[0075] The chaotic quantum sparrow search algorithm is used to identify all parameters of the transfer function matrix offline. The specific implementation steps are as follows:
[0076] (1) Population Chaotic Initialization: The gain, time constant, and pure delay time of all channels in the transfer function matrix are sequentially encoded into a one-dimensional position vector, and the position of each sparrow corresponds to a set of parameters to be identified. The initial population position is generated using the Tent chaotic mapping. The Tent mapping has the characteristics of uniform traversal and strong randomness, which can make the initial individuals evenly distributed within the parameter feasible region, improve population diversity, and avoid the algorithm from being concentrated in a local area in the initial stage. Set algorithm hyperparameters such as population size, maximum number of iterations, discoverer ratio, and watcher ratio.
[0077] (2) Fitness calculation and ranking: Substitute the parameters corresponding to each sparrow individual into the transfer function matrix model, and obtain the model response sequence of each output variable through numerical simulation. Calculate the fitness value of each individual according to equation (3). Sort the individuals in the population according to their fitness values from smallest to largest. The individuals ranked at the top are classified as discoverers, responsible for global search; the individuals ranked at the bottom are classified as followers, responsible for following the optimal solution; at the same time, a small number of individuals are randomly selected as vigilants, responsible for local reconnaissance and early warning.
[0078] (3) Iterative position update:
[0079] Discoverer Update: A quantum behavior update strategy is employed, using the current globally optimal position as an attractor and updating the discoverer's position based on a quantum potential well model. Quantum behavior update breaks through the traditional particle swarm optimization velocity update framework, enabling the discoverer to search over a wider range and effectively improving the algorithm's global optimization capability.
[0080] Follower update: Followers adopt a dimension-by-dimensional following strategy. Each dimension parameter independently follows the position of the discoverer with the best fitness in the current dimension, while introducing adaptive random perturbation. For followers with lower fitness ranking, the position is reinitialized and the search is moved to other regions within the feasible domain to avoid overcrowding of the population.
[0081] Watcher Update: Watchers randomly walk around the optimal position of the population to sense changes in fitness in local areas; when the optimal fitness does not improve significantly for several consecutive generations, the population escape mechanism is triggered, causing some individuals to jump to more distant areas to avoid the algorithm getting stuck in local optima.
[0082] (4) Periodic Chaotic Perturbation: A chaotic perturbation operation is introduced every 10 iterations. A chaotic perturbation vector with the same dimension as the position vector is generated based on the Tent mapping, and it is superimposed on the current global optimum to obtain a new mutated solution. The fitness of the new solution is calculated. If it is better than the original optimum, it is replaced; otherwise, the original optimum is retained. Chaotic perturbation can break the local equilibrium when the algorithm tends to converge, help the algorithm escape local optima, and further improve the optimization accuracy.
[0083] (5) Iteration termination and result output: When the algorithm reaches the preset maximum number of iterations, or when the relative change of the optimal fitness over multiple generations is less than the set convergence threshold, the iteration process is terminated and the current global optimal position vector, i.e., the offline identification result of the transfer function matrix parameters, is output.
[0084] For each typical load interval, offline identification is performed using historical data of the corresponding interval to obtain the benchmark model parameter set for each load interval, forming a full-condition benchmark model library, which serves as the initial benchmark value for subsequent online identification.
[0085] Step 4: Data-driven online model parameter identification
[0086] 4.1 Time-domain transformation of operator evolution theory
[0087] To address the time-varying nature of model parameters under dynamic load conditions, operator evolution theory is used to discretize the continuous-time transfer function model, constructing a discrete-domain model of the turbine-furnace coupling process, thus providing a mathematical basis for online identification.
[0088] In practice, the bilinear transformation method is used to transform the Laplace operator *s* in the continuous domain into a shift operator *z* in the discrete domain, resulting in a discrete transfer function matrix model. The bilinear transformation avoids frequency aliasing and accurately maps the dynamic characteristics of the continuous system. Based on this, a parameter evolution operator is introduced to describe the continuous evolution of model parameters with changes in operating conditions such as load and heating conditions. This allows the model parameters to be smoothly adjusted to adapt to changes in operating conditions, meeting the parameter identification requirements of dynamic load variations.
[0089] By using operator evolution transformation, the continuous-time parameter identification problem is transformed into a discrete-time algebraic problem, and the dynamic model in the form of differential equations is transformed into a system of linear equations, which facilitates rapid online identification.
[0090] 4.2 Hierarchical Bayesian Parameter Identification Based on Polynomial Matrices
[0091] Based on multinomial matrix theory and combined with a hierarchical Bayesian estimation framework, a fast online identification mechanism for model parameters is constructed, revealing the dynamic parameter identification mechanism driven by real-time running data. The expression for the posterior estimation of parameters is shown in equation (4):
[0092] (4)
[0093] In the formula, is the posterior estimation vector of the parameter to be identified; The information matrix is composed of input and output data; The output variable is the observation vector; is the regularization coefficient, derived from the hierarchical Bayesian prior distribution; It is an identity matrix.
[0094] The hierarchical Bayesian framework consists of two layers: the first layer is the parameter layer, which assumes that the parameters to be identified follow a Gaussian prior distribution, the prior mean is taken as the baseline parameter for offline identification, and the prior variance is set according to the fluctuation range of the working conditions; the second layer is the hyperparameter layer, which adopts a non-informative prior for the variance parameter of the prior distribution. By recursively deriving the Bayesian formula, the posterior distribution of the parameters is finally obtained, and its posterior mean is the estimation result of equation (4).
[0095] The polynomial matrix inversion algorithm is used to quickly solve equation (4). By utilizing the structural characteristics of the transfer function matrix and the polynomial operation rules, the calculation process of matrix inversion is simplified. Compared with conventional numerical matrix inversion, the calculation speed is greatly improved, which meets the real-time requirements of online identification.
[0096] 4.3 Implementation of Online Recognition for Scrolling Windows
[0097] A sliding time window mechanism is used to achieve rolling online identification of parameters. A fixed-length time window is set, the latest running data is collected in real time and added to the end of the window, and the earliest data at the beginning of the window is removed to keep the window length constant. After each sampling period, the latest parameter estimate is calculated based on the updated window data according to equation (4), so as to realize the real-time rolling update of model parameters.
[0098] The length of the sliding window can be adjusted according to the unit's operating characteristics and control requirements: a longer window length results in stronger anti-interference capabilities and higher stability of the identification results, but a slower response speed to changes in operating conditions; a shorter window length results in faster tracking of changes in operating conditions, but is more susceptible to measurement noise interference. By reasonably setting the window length, both identification accuracy and response speed can be balanced, adapting to the operating conditions requiring different load rate variations.
[0099] Step 5: Automatic matching and adaptive parameter updating of the full-condition model structure
[0100] 5.1 Load Zone Division and Benchmark Model Library
[0101] To achieve accurate adaptation across the 20%-100% full load range, the unit load is pre-divided into five typical ranges: 20%-30% rated load, 30%-40% rated load, 40%-60% rated load, 60%-80% rated load, and 80%-100% rated load.
[0102] For each load range, typical steady-state and dynamic historical data within that range are selected, and offline identification is performed using the aforementioned chaotic quantum sparrow search algorithm to obtain the baseline model parameter set for that range. The baseline parameter sets for all ranges together constitute the full-condition baseline model library. The baseline model for the low-load range focuses on parameter changes caused by unstable boiler combustion, low working fluid flow, and altered heat transfer characteristics; the baseline model for the high-load range focuses on parameter saturation and enhanced coupling strength, ensuring that the baseline model for each range accurately corresponds to the boiler-turbine coupling characteristics of that load segment.
[0103] 5.2 Automatic Working Condition Matching Logic
[0104] The system collects real-time active power data from the generating units, calculates the percentage of the current actual load relative to the rated load, and determines the current load interval. To avoid frequent model switching caused by small load fluctuations near interval boundaries, a hysteresis-based decision logic is employed: during load increases, interval switching is triggered only when the load exceeds the upper limit of the previous interval and remains stable for a set duration; during load decreases, interval switching is triggered only when the load falls below the lower limit of the current interval and remains stable for a set duration. The hysteresis width and stabilization duration can be set according to the unit's load fluctuation characteristics, effectively improving the stability of model operation.
[0105] When a change in load range is detected, the baseline parameter set for the new range is retrieved from the baseline model library and used as the new initial value for the online identification module. The parameter adaptive update is then initiated, enabling the model to quickly adapt to the characteristics of the new load range.
[0106] 5.3 Parameter Disturbance-Free Switching Mechanism
[0107] When model parameter switching is triggered, a smooth transition algorithm is used to achieve a bumpless parameter switching, avoiding sudden parameter changes that could cause abrupt changes in model output and affect the stable operation of the control system. A transition duration is set, during which the model's parameters gradually change from the current value in the original interval to the reference parameter value in the new interval according to an exponential law. The time constant of the transition curve is set according to the unit's inertial characteristics to ensure that the model output is continuous and smooth without abrupt changes during the switching process.
[0108] By automatically matching operating conditions and seamlessly switching parameters, continuous adaptation of the model across the 20%-100% full load range is achieved, ensuring the modeling accuracy of each load range while avoiding disturbances caused by operating condition switching.
[0109] Step Six: Multi-Scenario Algorithm Verification
[0110] For the boiler-turbine coupling process with different input and output dimensions, the effectiveness of the algorithm was verified sequentially, covering three typical coupling scenarios: two-input single-output, two-input two-output, and three-input three-output. The performance of the algorithm under different coupling complexities was verified step by step. The input variables include three core regulating variables: fuel quantity, feedwater flow rate, and air supply flow rate. The output variables include three core controlled variables: main steam pressure, main steam temperature, and unit active power.
[0111] The verification was conducted using two methods: computer numerical simulation and simulator testing, as detailed below:
[0112] (1) Computer numerical simulation: A standard turbine-furnace coupled process model was built in numerical simulation software as the true value model of the controlled object. Measurement noise and load disturbances of different intensities were added to the object model, and the method of this invention and the traditional identification method were used for modeling respectively. The steady-state fitting accuracy, dynamic tracking capability, algorithm convergence speed, anti-interference capability and other indicators of the model were compared and analyzed. The identification effect under the whole working condition was verified by setting simulation conditions with different load levels.
[0113] (2) Simulator test: A full-range simulator based on ultra-high pressure coal-fired heating and power generation unit was used. This simulator can accurately reproduce the dynamic characteristics of the real unit under all operating conditions. The identification algorithm of this invention was embedded into the simulation system in the form of a module. Steady-state operating conditions and dynamic operating conditions with rising and falling loads at different load segments were selected for testing. The modeling accuracy of the 20%-40% low load segment, the parameter tracking capability of the load change process, and the smoothness of the operating condition switching were verified. The engineering applicability of the algorithm under the characteristics of the real unit was verified.
[0114] Through comprehensive verification across multiple scenarios and methods, we ensure that the algorithm maintains stable and reliable performance under varying coupling complexities and load conditions, thereby guaranteeing the effectiveness of practical engineering applications.
[0115] Example:
[0116] This embodiment uses a 300MW ultra-high pressure coal-fired thermal power generating unit as the application object to implement the low-load full-condition boiler-turbine coupling process modeling method of the present invention. The specific implementation process is as follows:
[0117] First, historical operating data of the unit over the past 12 months was collected, with a sampling period of 1 second. The data covered the full load range of 20%-100%, including pure condensing operation and heating operation with different extraction steam flow rates. The collected raw data were then subjected to third-order polynomial moving zero-value repair, 5-layer wavelet soft thresholding denoising, and first-order difference 3σ coarse value removal to obtain a high-quality modeling dataset.
[0118] Then, variable correlation analysis was performed using the mutual information method to select and determine the input variables as boiler fuel quantity, feedwater flow rate, and air supply flow rate, and the output variables as main steam pressure, main steam temperature, and unit active power, thus constructing a three-input, three-output transfer function matrix model. Each channel adopts a third-order inertial element structure with pure delay, and the pure delay element is equivalently transformed using the first-order Pad approximation.
[0119] Next, five typical load intervals were defined, and corresponding historical data were selected for each interval. The chaotic quantum sparrow search algorithm was used for offline parameter identification. The algorithm population size was set to 50, the maximum number of iterations to 200, with 70% being discoverers and 10% being vigilant. A chaotic perturbation was introduced every 10 iterations. After iterative optimization, the baseline model parameter set for each load interval was obtained, forming a full-condition baseline model library.
[0120] Subsequently, an online identification module was built, employing a sliding time window mechanism with a window length of 100 sampling points. Based on operator evolution theory, the continuous model was discretized, and combined with a multinomial matrix hierarchical Bayesian algorithm, real-time rolling updates of model parameters were achieved. Automatic matching logic for five load intervals was designed, setting the hysteresis width to 2% of rated load, the stabilization switching time to 30 seconds, and employing exponential smooth transition for parameter switching to achieve adaptive adjustment of the full-condition model.
[0121] Finally, the model was validated on a full-range simulator for the unit. Five steady-state load points (25%, 35%, 50%, 75%, and 100%) and four dynamic processes (20%-40% load increase, 40%-20% load decrease, 40%-100% load increase, and 100%-40% load decrease) were selected for testing. The test results show that the model established by this method can accurately track the dynamic response of the unit under all operating conditions, maintaining high modeling accuracy across all load segments; the online parameter identification speed can adapt to rapid load change conditions, and the model tracking error is small during dynamic processes; the operating condition switching process is smooth and undisturbed, effectively supporting the improvement of the unit's load change rate and operational stability.
[0122] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for low-load, full-condition turbine-boiler coupling process of an ultra-high pressure coal-fired thermal power generator unit, characterized in that: Includes the following steps: Step 1: Collect historical operating data of the ultra-high pressure coal-fired heating and power generation unit during the 20%-100% full load range. After preprocessing the data, analyze the dynamic characteristics of the boiler steam-water power process, screen the key input and output variables that affect the stability and economy of the unit under low load operation, and construct the transfer function matrix model structure of the boiler-machine coupling process. Step 2: Using the mean square error function between the actual values of multiple output variables and the model output values as the optimization objective, the chaotic quantum sparrow search algorithm is used to identify the structural parameters of the transfer function matrix model offline based on the preprocessed historical data, and the parameters obtained offline are used as the initial values for online identification under each typical working condition. Step 3: Based on the model structure and initial parameter values obtained offline, an online parameter identification mechanism based on hierarchical Bayesian estimation is constructed using operator evolution theory and multinomial matrix theory to quickly identify the model parameters online under dynamic conditions. Step 4: Design an automatic matching and parameter adaptive update program for the model structure under different steady-state operating conditions to achieve adaptive adjustment of the operating conditions of the boiler-turbine coupling process model under full load. Step 5: For the furnace-machine coupling process with different input and output dimensions, computer numerical simulation and simulator experiments are carried out in sequence to verify the effectiveness of the method.
2. The method for low-load, full-condition turbine-boiler coupling process of an ultra-high pressure coal-fired thermal power generator unit according to claim 1, characterized in that: In step one, each element of the transfer function matrix model is a third-order inertial element with a pure delay, and its Laplace domain transfer function expression is: In the formula, The channel transfer function from the i-th input variable to the j-th output variable; This is the static gain of the channel; , , These are the three time constants of the third-order inertial element; This is the pure delay time of the channel; For the Laplace operator.
3. The method for low-load full-condition turbine-boiler coupling process of ultra-high pressure coal-fired thermal power generation unit according to claim 2, characterized in that: In step one, the pure delay element in the transfer function is equivalently transformed using the first-order Pad approximation method. The expression for the pure delay element after transformation is: After transformation, the transfer function unit is equivalent to a fourth-order inertial element with zeros, which simplifies subsequent parameter identification and calculation.
4. The method for low-load full-condition turbine-boiler coupling process of ultra-high pressure coal-fired thermal power generation unit according to claim 1, characterized in that: In step two, the expression for the objective function for optimizing the mean square error is: In the formula, This represents the mean square error value. This represents the total number of sampling points for historical data. This represents the total number of output variables in the furnace-machine coupling process. This represents the actual running value of the m-th output variable at the k-th sampling time. This represents the model prediction value of the m-th output variable at the k-th sampling time.
5. The method for low-load, full-condition turbine-boiler coupling process of an ultra-high pressure coal-fired thermal power generator unit according to claim 1, characterized in that: The data preprocessing in step one specifically includes: using a low-order polynomial sliding fit method for zero initial value processing, using a soft thresholding method for denoising the running data, and using a low-order difference method for removing coarse values; wherein the low-order polynomial sliding fit uses a third-order polynomial, and selects 5 sampling points before and after the current sampling point to form a sliding window, and fits to obtain the correction value of the current point.
6. The method for low-load full-condition turbine-boiler coupling process of ultra-high pressure coal-fired thermal power generation unit according to claim 1, characterized in that: The offline identification using the chaotic quantum sparrow search algorithm in step two specifically includes: S21: The position vectors of the sparrow population are initialized using Tent chaotic mapping, and each position vector corresponds to a set of parameters to be identified in the transfer function matrix; S22: Using the mean squared error function as the fitness function, calculate the fitness value of each individual sparrow and classify them into discoverers, followers and watchers according to their fitness. S23: The position of the discoverer is updated iteratively using a quantum behavior update method, the position of the follower is updated using a dimension-by-dimensional following strategy, and the position of the vigilant is updated using a random walk strategy. S24: Introduce a chaotic perturbation every 10 iterations to perform chaotic mutation on the current optimal solution. If the fitness is better after mutation, replace the original optimal solution. S25: When the maximum number of iterations is reached or the change in fitness is less than the set threshold, the iteration is terminated and the optimal parameter set is output as the offline identification result.
7. The method for low-load full-condition turbine-boiler coupling process of ultra-high pressure coal-fired thermal power generation unit according to claim 1, characterized in that: In step three, a hierarchical Bayesian parameter identification mechanism is constructed based on the inverse form of a polynomial matrix. The expression for the posterior estimation of the parameters is as follows: In the formula, is the posterior estimation vector of the parameter to be identified; The information matrix is composed of input and output data; The output variable is the observation vector; is the regularization coefficient, derived from the hierarchical Bayesian prior distribution; Let be the identity matrix; the above equation can be solved quickly using a polynomial matrix inversion algorithm.
8. The method for low-load full-condition turbine-boiler coupling process of ultra-high pressure coal-fired thermal power generation unit according to claim 1, characterized in that: In step three, the application of operator evolution theory is as follows: for dynamic variable load conditions, differential operators are used to perform time-domain transformation on the transfer function matrix, transforming the continuous-time model into a discrete-time operator form, so that the model parameters can continuously evolve with the change of operating conditions, adapting to the parameter identification requirements of dynamic processes.
9. The method for low-load full-condition turbine-boiler coupling process of ultra-high pressure coal-fired thermal power generation unit according to claim 1, characterized in that: The automatic model structure matching and adaptive parameter update in step four specifically include: pre-dividing the load into five load intervals: 20%-30%, 30%-40%, 40%-60%, 60%-80%, and 80%-100%, with each interval corresponding to a set of baseline model parameters; real-time acquisition of unit load values; and triggering model parameter switching when the load crosses the interval boundary and the duration exceeds the set time. The switching process uses a smooth transition algorithm to achieve disturbance-free parameter switching.
10. The method for low-load full-condition turbine-boiler coupling process of ultra-high pressure coal-fired thermal power generation unit according to claim 1, characterized in that: The different input and output dimensions in step five include three typical boiler-machine coupling scenarios: two-input single-output, two-input two-output, and three-input three-output. The input variables include boiler fuel quantity, feedwater flow rate, and air supply flow rate, while the output variables include main steam pressure, main steam temperature, and unit active power.
