Coal consumption prediction and operation optimization method, device and equipment for thermal power generating unit and medium

Abstract

The application relates to a thermal power unit coal consumption prediction and operation optimization method, device, equipment and medium, and relates to the technical field of artificial intelligence, and comprises the following steps: performing Gaussian kernel density estimation based on a heat supply proportion feature, determining minimum points between bimodal structures of a probability density function as self-adaptive thresholds of heating and non-heating seasons, combining fuzzy membership to divide heating season and non-heating season working conditions, and constructing a heating season and non-heating XGBoost coal consumption prediction model; performing weighted fusion on output results of the coal consumption prediction model based on fuzzy membership to obtain a power generation coal consumption prediction value; taking minimization of the power generation coal consumption prediction value as an objective function, adopting a particle swarm algorithm based on a thermal safety boundary, performing reverse optimization under the constraint of meeting a current load instruction, and outputting recommended values of operation parameters that make the power generation coal consumption of a current real-time working condition lowest. The application realizes accurate coal consumption prediction and operation optimization of a unit under full working conditions by constructing a seasonal prediction model and combining an improved particle swarm algorithm.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a method, apparatus, equipment, and medium for predicting coal consumption and optimizing the operation of thermal power units. Background Technology

[0002] Currently, to optimize the operation and control strategies of complex industrial equipment such as thermal power units and reduce coal consumption for power generation, existing technologies often employ an optimization method based on three-dimensional operating space discrete grid matching. The basic implementation process of this method is as follows: First, a standard sample library is established by cleaning historical operating data. Then, three key parameters—heat-to-power ratio, unit load, and ambient temperature—are selected to construct a three-dimensional feature space, which is then discretized into a grid. During actual unit operation, the current operating parameters are acquired in real time, and their corresponding grid subclass is located. The historical state point with the lowest coal consumption within that grid is traversed and retrieved. Finally, the operating parameters corresponding to this "historically optimal state" are used as target reference values ​​to guide the real-time adjustment of the unit.

[0003] However, the existing optimization methods based on three-dimensional discrete mesh matching in operating space have many defects and drawbacks in practical applications. First, the coupling relationship of the thermal system of the unit differs fundamentally between the heating season and the non-heating season. Existing technologies often ignore this physical characteristic, simply mixing historical data from the whole year for training and mesh generation, or introducing only a fixed "heat-to-power ratio" as the basis for mesh generation. This rigid modeling approach makes it impossible for the model to accurately represent the actual operating boundaries of the unit in different seasons. Especially in the transition season or the deep heating period, the optimization results often shift towards the annual average state, producing a serious "averaging" error, resulting in a significant decrease in prediction accuracy.

[0004] Secondly, existing 3D mesh matching methods are essentially offline optimization methods based on "historical experience lookup tables," and the target reference values ​​they provide are limited by the upper limit of historical data (i.e., they can only find the optimal values ​​that "have happened in the past," and cannot break through historical limitations). Most existing technologies stop at matching operating conditions or calculating the optimal point offline, lacking the ability to perform online closed-loop optimization based on the current real-time load command, using intelligent algorithms to perform reverse solving within a multi-dimensional controllable parameter space, and dynamically find the "parameters corresponding to the current lowest coal consumption." Therefore, existing technologies struggle to achieve real-time, dynamic, and groundbreaking global optimization of unit operation strategies.

[0005] In summary, there is an urgent need for a unit operation optimization method that can overcome the above-mentioned defects, possess high precision, strong seasonal adaptability, and online real-time closed-loop optimization capabilities. Summary of the Invention

[0006] In view of this, this application provides a method, device, equipment and medium for predicting coal consumption and optimizing operation of thermal power units. The main purpose is to solve the technical problems of existing technologies ignoring the fundamental differences in the coupling relationship of the thermal system of the unit between the heating season and the non-heating season, which leads to the model being unable to accurately represent the actual operating boundary of the unit in different seasons, resulting in a significant decrease in prediction accuracy, and the difficulty of existing technologies in achieving real-time, dynamic and groundbreaking global optimization of unit operation strategies.

[0007] Firstly, this application provides a method for predicting coal consumption and optimizing the operation of thermal power units, including: Acquire DCS high-frequency data and plant-level low-frequency data of thermal power units, as well as current real-time operating condition data. Perform steady-state screening and multi-source time alignment processing on the DCS high-frequency data and the plant-level low-frequency data to construct a high-dimensional feature matrix containing aligned DCS high-frequency and low-frequency features. Gaussian kernel density estimation is performed based on the heating ratio characteristics in the high-dimensional feature matrix to identify the bimodal structure of the probability density function. Adaptive thresholds for heating season and non-heating season are determined based on the minimum points between the bimodal structures. The operating conditions of the thermal power unit are dynamically divided by combining the fuzzy membership function. The operating conditions include heating season conditions, non-heating season conditions, and transition conditions. For the heating season operating conditions and the non-heating season operating conditions, respectively, an XGBoost coal consumption prediction model for the heating season and an XGBoost coal consumption prediction model for the non-heating season are constructed, which integrate thermodynamic cross-features and customized loss functions. Calculate the fuzzy membership degree corresponding to the heating ratio of the current real-time operating data, input the current real-time operating data into the trained heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model respectively, and perform weighted fusion of the output results of the heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model based on the fuzzy membership degree to obtain the predicted value of power generation coal consumption under the current operating conditions; Using the minimization of the predicted coal consumption for power generation as the objective function, an improved particle swarm optimization algorithm based on thermal safety boundaries is employed to perform reverse optimization under the constraints of the current load command, and output recommended operating parameter values ​​that minimize the coal consumption for power generation under the current real-time operating conditions.

[0008] Secondly, this application provides a device for predicting coal consumption and optimizing the operation of thermal power units, including: The acquisition module is used to acquire DCS high-frequency data and plant-level low-frequency data of thermal power units, as well as current real-time operating condition data. It performs steady-state screening and multi-source time alignment processing on the DCS high-frequency data and the plant-level low-frequency data to construct a high-dimensional feature matrix containing aligned DCS high-frequency and low-frequency features. The determination module is used to perform Gaussian kernel density estimation based on the heating ratio characteristics in the high-dimensional feature matrix, identify the bimodal structure of the probability density function, determine the adaptive threshold for heating season and non-heating season based on the minimum point between the bimodal structures, and dynamically classify the operating conditions of the thermal power unit in combination with the fuzzy membership function. The operating conditions include heating season conditions, non-heating season conditions and transition conditions. The module is used to construct XGBoost coal consumption prediction models for heating season and non-heating season, respectively, which integrate thermodynamic cross-features and customized loss functions for the heating season and non-heating season operating conditions. The calculation module is used to calculate the fuzzy membership degree corresponding to the heating ratio of the current real-time operating data, input the current real-time operating data into the trained heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model respectively, and perform weighted fusion of the output results of the heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model based on the fuzzy membership degree to obtain the predicted value of power generation coal consumption under the current operating conditions; The output module is used to minimize the predicted coal consumption for power generation as the objective function, and employs an improved particle swarm optimization algorithm based on thermal safety boundaries to perform reverse optimization under the constraints of the current load command, and outputs recommended operating parameters that minimize the coal consumption for power generation under the current real-time operating conditions.

[0009] Thirdly, this application provides an electronic device, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the computer program to implement the coal consumption prediction and operation optimization method for thermal power units described in the first aspect.

[0010] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the coal consumption prediction and operation optimization method for thermal power units described in the first aspect.

[0011] By utilizing the above technical solutions, this application provides a method, apparatus, equipment, and medium for predicting coal consumption and optimizing the operation of thermal power units. Compared with existing technologies, this application can acquire DCS high-frequency data and plant-level low-frequency data of thermal power units, as well as current real-time operating condition data. It performs steady-state screening and multi-source time alignment processing on the DCS high-frequency data and plant-level low-frequency data to construct a high-dimensional feature matrix containing aligned DCS high-frequency and low-frequency features. Based on the heating ratio feature in the high-dimensional feature matrix, it performs Gaussian kernel density estimation to identify the bimodal structure of the probability density function. Based on the minimum points between the bimodal structures, it determines adaptive thresholds for the heating season and non-heating season, and dynamically classifies the operating conditions of the thermal power unit using fuzzy membership functions. The operating conditions include heating season conditions, non-heating season conditions, and transitional conditions. For the heating season conditions and non-heating season conditions, it constructs... A heating season XGBoost coal consumption prediction model and a non-heating season XGBoost coal consumption prediction model were constructed, integrating thermodynamic cross-features and a customized loss function. Fuzzy membership degrees corresponding to the heating ratio under current real-time operating conditions were calculated. The current real-time operating conditions data were then input into the trained heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model, respectively. The outputs of the heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model were weighted and fused based on the fuzzy membership degrees to obtain the predicted coal consumption for power generation under the current operating conditions. With minimizing the predicted coal consumption for power generation as the objective function, an improved particle swarm optimization algorithm based on thermal safety boundaries was used to perform reverse optimization under the constraint of the current load command, outputting recommended operating parameter values ​​that minimize the coal consumption for power generation under the current real-time operating conditions.

[0012] By adopting the above technical solution, this application does not rely on manual experience to set fixed thresholds. Instead, it directly extracts the heating ratio features from a high-dimensional feature matrix that integrates high-frequency DCS and low-frequency plant-level data for probability density analysis. This accurately captures the natural aggregation state of data during the heating season and non-heating season, avoiding the distortion caused by averaging model parameters. The fuzzy membership mechanism enables the unit to smoothly transition prediction results during seasonal transitions, significantly improving the accuracy of coal consumption prediction under all operating conditions (especially complex transitional seasons).

[0013] This application, based on an online reverse optimization mechanism, no longer relies on historical data lookups. Instead, it uses minimizing the predicted coal consumption for power generation as the objective function and employs an improved particle swarm optimization algorithm based on thermal safety boundaries. Under the premise of satisfying current load commands and thermal safety constraints, it performs reverse optimization (i.e., deriving the optimal operating parameters from the target coal consumption) within a multi-dimensional controllable parameter space, directly outputting recommended operating parameter values ​​that minimize coal consumption under current real-time operating conditions. This application can explore better combinations of operating parameters not found in historical data, achieving breakthrough global optimization rather than simply replicating historical bests. This application possesses online closed-loop optimization capabilities, dynamically adjusting according to real-time load commands to ensure the unit always operates at its optimal energy efficiency point under current conditions, significantly improving the unit's economy and flexibility.

[0014] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a method for predicting coal consumption and optimizing operation of thermal power units, provided in an embodiment of this application; Figure 2 This is a schematic diagram of a thermal power unit coal consumption prediction and operation optimization device provided in an embodiment of this application. Detailed Implementation

[0018] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0019] The following description, with reference to the accompanying drawings, describes a method, apparatus, equipment, and medium for predicting coal consumption and optimizing operation of thermal power units according to embodiments of this application.

[0020] This application provides a method, apparatus, equipment, and medium for predicting coal consumption and optimizing the operation of thermal power units. Its main purpose is to address the technical problems of low accuracy in coal consumption prediction, significant interference from unsteady-state data, and lack of real-time optimization guidance for operating parameters in combined heat and power (CHP) units under different seasonal operating conditions. By constructing a seasonal prediction model and combining it with an improved particle swarm optimization algorithm, accurate coal consumption prediction and operation optimization under all operating conditions of the unit are achieved.

[0021] like Figure 1 As shown, embodiments of this application provide a method for predicting coal consumption and optimizing the operation of thermal power units, including: Step 101: Obtain DCS high-frequency data and plant-level low-frequency data of the thermal power unit, as well as the current real-time operating condition data. Perform steady-state screening and multi-source time alignment processing on the DCS high-frequency data and plant-level low-frequency data to construct a high-dimensional feature matrix containing the aligned DCS high-frequency and low-frequency features.

[0022] In thermal power plants, DCS system data (such as load, temperature, and pressure) is typically sampled at the second or minute level, while plant-level management data (such as heating ratio and plant power consumption rate) are monthly statistical data. There is a serious frequency mismatch between the two in the time dimension. If causal alignment is not performed and the low-frequency monthly statistical values ​​are directly assigned to all high-frequency DCS records within that month, a "ghost feature" phenomenon will occur.

[0023] To avoid the "ghost feature" phenomenon (i.e., the DCS operating status at a certain instant is associated with a monthly statistical value that does not correspond to the actual operating conditions at that instant, causing the XGBoost model to learn a false mapping relationship; for example, the plant power consumption rate in a certain month is 6.5%, but the actual rate is about 8% at the beginning of the month when the load is low and about 5.5% at the end of the month when the load is full. If no load weighting correction is made, the model will overestimate the impact of the plant power consumption rate under full load conditions and underestimate it under low load conditions, resulting in a systematic bias in the predicted value at the beginning and end of the month), this application can perform the following processing: First, the coefficient of variation of key parameters such as main steam pressure and unit load can be calculated using a sliding window. Non-steady-state (variable load) process data with coefficients of variation exceeding the preset steady-state threshold are eliminated, and only steady-state operating time data (i.e., steady-state data) in the DCS high-frequency data are retained. For monthly cumulative statistics (such as heating ratio) in plant-level low-frequency data, a load-weighted time allocation method can be adopted. The monthly cumulative statistics are allocated to each steady-state time slice where the steady-state data is located according to the proportion of the unit load corresponding to each steady-state time slice in the month to the total load of the month (that is, the monthly cumulative statistics are allocated to each steady-state time slice in the month by using time integration inverse operation), so as to complete the upward alignment in the time dimension.

[0024] Specifically, the monthly statistical value for the m-th month can be set as follows: There are N steady-state time slices in this month, and the unit load corresponding to the i-th time slice is... Then the feature value assigned to this time slice is:

[0025] In the formula, For the eigenvalues ​​assigned to the i-th steady-state time slice, This represents the monthly statistics for the m-th month. Let i be the unit load corresponding to the i-th steady-state time slice. The sum of the unit loads corresponding to all steady-state time slices within the month (i.e., the total load for the month) is calculated by weighting the loads at each time point according to their proportion to the total load for the month.

[0026] For monthly ratio statistics (such as plant power consumption rate) in plant-level low-frequency data, since this ratio is strongly correlated with the unit load level (the plant power consumption rate is higher when the load is low), the same monthly value cannot be directly assigned to all DCS records within that month.

[0027] This application establishes a time-based correction model for the load correction coefficient function of plant power consumption rate and load, thereby correcting monthly ratio statistics (such as monthly plant power consumption rate) into time-based characteristic values ​​that match the load level at each moment, and assigning them to the corresponding steady-state data to complete the alignment of load levels; wherein, the formula of the time-based correction model is as follows:

[0028] In the formula, For time-by-time correction model, Let m be the monthly plant power consumption rate. Let t be the rated load of the unit at time t. This is the rated load of the unit. This is a load correction coefficient function fitted based on historical data; Aligned DCS high-frequency data (such as load, main steam temperature and pressure, feedwater flow, etc.) and plant-level low-frequency data (such as heating ratio, plant power consumption rate, etc.) can be fused based on time indexing to construct a high-dimensional feature matrix.

[0029] Step 102: Based on the heating ratio feature in the high-dimensional feature matrix, Gaussian kernel density estimation is performed to identify the bimodal structure of the probability density function. Based on the minimum point between the bimodal structures, adaptive thresholds for heating season and non-heating season are determined. The operating conditions of thermal power units are dynamically divided by combining the fuzzy membership function. The operating conditions include heating season conditions, non-heating season conditions, and transition conditions.

[0030] In this embodiment of the disclosure, unlike the traditional method of dividing seasons based on calendar months (which is prone to failure due to abnormal weather), this application uses "heating ratio," which more accurately reflects the actual thermal characteristics of the unit, as the core criterion for data-driven adaptive division. An adaptive threshold is set. When the proportion of heating supply is greater than When the heating supply ratio is less than 1%, it is considered a heating season operating condition (high back pressure, large steam extraction); when the heating supply ratio is less than 1%. When this occurs, it is determined to be a non-heating season operating condition (pure condensing or low steam condition). This method can more accurately reflect the actual thermodynamic characteristics of the unit and avoid calendar division failures caused by climate anomalies. Specifically, it may include: Plot a two-dimensional scatter distribution of the full steady-state data with the heating ratio as the horizontal axis and power generation coal consumption as the vertical axis. Gaussian kernel density estimation is performed on the heating ratio dimension to identify the bimodal structure of the probability density function. The first peak of the probability density function corresponds to the heating ratio of non-heating season conditions (pure condensation or low steam conditions) concentrated in the low value range, while the second peak of the probability density function corresponds to the heating ratio of heating season conditions (high back pressure, large steam extraction) concentrated in the high value range. Extract the minimum point (i.e., the "natural breakpoint") of the probability density function between the first and second peaks, and determine the heating ratio corresponding to the minimum point as the adaptive threshold for the heating season and the non-heating season. Among them, the adaptive threshold is not preset by humans, but is automatically determined by the distribution characteristics of the actual operating data of the unit, so it can adapt to the differences in heating characteristics of different units and different years. In actual operation, the heating ratio is within the adaptive threshold. When there are fluctuations in the surrounding area (such as changes in heating load due to the transition between spring and autumn, or the temperature difference between day and night), directly switching between the heating season model and the non-heating season model based on a hard threshold will cause discontinuous jumps in the predicted values ​​at the switching point, affecting the stability of the operational guidance. To address this, this solution introduces a smooth transition mechanism based on fuzzy membership functions.

[0031] Set transition band width parameter And based on the transition band width parameter and adaptive threshold Constructing a transition interval ; When the heating ratio r is greater than or equal to the transition range maximum value When determining the fuzzy membership function The function value is 1, which determines that the operating condition of the thermal power unit is the heating season condition; When the heating proportion r is less than or equal to the transition range minimum value When determining the function value of the fuzzy membership function. The value is 0, indicating that the operating condition of the thermal power unit is the non-heating season condition; When the heating ratio r is within the transition range When determining the function value of the fuzzy membership function. The objective function value This refers to linear transition, which determines the operating condition of the thermal power unit as the transitional operating condition. This fuzzy fusion mechanism ensures that, near threshold boundaries such as the spring and autumn transition seasons, there will be no jumps in predicted values ​​or oscillations in the operating guidance signal due to hard switching.

[0032] Step 103: For both heating season and non-heating season operating conditions, construct XGBoost coal consumption prediction models for heating season and non-heating season respectively, which integrate thermodynamic cross-features and customized loss functions.

[0033] In this embodiment of the present disclosure, before inputting the original DCS parameters into the model, high-order cross features of the input original DCS parameters can be pre-constructed based on thermodynamic principles, so that the model can directly learn the state representation with physical meaning and avoid relying solely on multiple splits of the tree model to capture interaction effects; the high-order cross features may include: main steam temperature deviation, comprehensive index of flue gas heat loss and heating intensity index. Among them, the main steam temperature deviation is In the formula, Main steam temperature deviation To obtain the design main steam temperature from the design sliding pressure curve under the current load P, The current measured value of the main steam temperature is used. The deviation of the main steam temperature characterizes the degree of thermal efficiency loss caused by the main steam temperature deviating from the design operating conditions. The greater the deviation, the greater the thermal efficiency loss usually means. The comprehensive index of flue gas heat loss is In the formula, As a comprehensive index of flue gas heat loss, The exhaust gas temperature, Oxygen content in flue gas This is an empirical coefficient that comprehensively characterizes the heat loss from flue gas exhaust. Since flue gas heat loss is not only related to flue gas temperature, but also closely related to the excess air coefficient (reflected by oxygen content), XGBoost requires multiple tree splits to capture the interaction effect of flue gas temperature and oxygen content when they are input alone. However, by directly constructing the interaction term, the model can directly utilize this comprehensive index in a single split.

[0034] Heating intensity index is In the formula, As a heating intensity indicator, For the proportion of heating supply, This represents the current actual load of the unit. The product of the heating ratio and the load rate is used as the rated load of the unit to comprehensively reflect the degree to which the unit simultaneously undertakes the "dual burden" of power generation and heating. The standard XGBoost objective function is the mean squared error (MSE), which is a purely data-driven black-box fitting method and may produce prediction results that violate fundamental thermodynamic laws. This approach adds a physical constraint penalty term based on the first and second laws of thermodynamics to the standard MSE loss function, constructing a customized loss function. That is, a customized loss function is constructed based on the standard mean square error term of predicted coal consumption and actual coal consumption, the thermodynamic monotonicity penalty term, and the operating boundary penalty term to overcome the results that may violate thermodynamic laws due to pure data-driven black-box fitting. The customized loss function is as follows:

[0035] In the formula, To customize the loss function, The standard mean square error term between predicted coal consumption and actual coal consumption. , and These are the hyperparameters determined through cross-validation. This is a boundary penalty term applied when the predicted coal consumption exceeds the theoretical lower and lower limits based on the unit's heat balance calculations. As a thermodynamic monotonicity penalty term, after sorting the training samples by load, the predicted coal consumption trend between adjacent load levels is detected. When the model predicts that "under the condition that other parameters remain unchanged, i.e. when the flue gas temperature increases..." However, the predicted coal consumption value actually decreased. When a violation of the second law of thermodynamics occurs, a penalty proportional to the degree of violation is imposed, expressed as:

[0036] In the formula, This is a thermodynamic monotonicity penalty term. This represents the change in the predicted coal consumption value. This represents the change in flue gas temperature.

[0037] The training data for models used in heating seasons and those used in non-heating seasons are different, and their hyperparameter configurations should also be differentiated. Specifically, based on high-order cross features and a customized loss function, an XGBoost model is trained on a dataset of heating season operating conditions to obtain an XGBoost coal consumption prediction model for the heating season. The heating season operating conditions are affected by heating steam extraction and feature interactions are complex. Therefore, the XGBoost coal consumption prediction model for the heating season is configured with a large first tree depth threshold to capture the high-order interaction effects between multiple variables in the heating season. Based on high-order cross features and customized loss functions, XGBoost models are trained on datasets for non-heating season operating conditions to obtain non-heating season XGBoost coal consumption prediction models. Since the non-heating season is a pure condensing operating condition with relatively simple feature relationships, the non-heating season XGBoost coal consumption prediction model is configured with a second tree depth threshold that is less than the first tree depth threshold to avoid overfitting and improve generalization ability. Both the heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model employ grid search combined with 5-fold cross-validation for hyperparameter optimization. The model can automatically learn the nonlinear influence weights of various features (such as ambient temperature and heating steam extraction rate) on coal consumption under different seasons.

[0038] Step 104: Calculate the fuzzy membership degree corresponding to the heating ratio of the current real-time operating data. Input the current real-time operating data into the trained heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model respectively. Based on the fuzzy membership degree, perform weighted fusion of the output results of the heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model to obtain the predicted value of power generation coal consumption under the current operating conditions.

[0039] Specifically, after obtaining the current real-time operating data, the fuzzy membership degree corresponding to its actual heating ratio r can be calculated. The current features are input into the XGBoost coal consumption prediction model for the heating season. XGBoost Coal Consumption Prediction Model for Non-Heating Seasons Finally, the predicted coal consumption value for power generation under the current operating conditions is calculated using the predicted coal consumption formula, which is shown below:

[0040] In the formula, For the fuzzy membership function values ​​during the heating season, This represents the output value of the XGBoost coal consumption prediction model for the heating season. This represents the output value of the XGBoost coal consumption prediction model for the non-heating season. The values ​​represent the fuzzy membership function values ​​for the non-heating season.

[0041] This fuzzy fusion mechanism ensures that the predicted output transitions smoothly and continuously between the two models near the threshold boundary, eliminating the problem of predicted value jumps caused by hard switching. It has important practical value in guiding operations during the spring and autumn transition seasons.

[0042] Step 105: Taking the minimization of the predicted coal consumption for power generation as the objective function, an improved particle swarm optimization algorithm based on the thermal safety boundary is adopted to perform reverse optimization under the constraint of the current load command, and output the recommended operating parameter values ​​that minimize the coal consumption for power generation under the current real-time operating conditions.

[0043] In this embodiment of the disclosure, this step aims to utilize the trained model as the fitness function for inverse solution within a multidimensional controllable parameter space, rather than relying on historical experience for table lookup. Specifically, it may include: Minimize the predicted coal consumption for power generation ( Using the objective function as the current load command (unit load = current commanded load) and the safe operating range of each parameter as constraints, the particle swarm (decision variable x is an adjustable operating parameter such as main steam pressure, oxygen content, etc.) is initialized and iteratively optimized. During the iterative optimization process, an adaptive inertial weight adjustment strategy and a directional constraint mutation strategy based on the thermal safety boundary are introduced to output recommended operating parameter values ​​that satisfy the constraints and minimize the predicted value of power generation coal consumption.

[0044] The adaptive inertia weight adjustment strategy can be implemented using the following formula:

[0045] In the formula, The adaptive inertia weight at time t, This represents the maximum value of the inertia weight. This represents the minimum value of the inertia weight. This represents the maximum number of iterations for the particle swarm optimization algorithm. For adjustment coefficients, This represents the optimal fitness value for the current population. This represents the average fitness value of the current population. This represents the current iteration number.

[0046] Among them, inertia weight Instead of using a fixed value or a simple linear decreasing strategy, it adjusts the value based on the current search status and the unit's operating conditions.

[0047] When population fitness is dispersed ( When the population converges (large), it indicates that the search space still has significant exploration value, and the inertia weight remains large to enhance the global search capability; when the population converges (large), it indicates that the search space still has significant exploration value, and the inertia weight remains large to enhance the global search capability. When the inertia weight approaches 0, it decreases rapidly to enhance the ability to perform fine-grained local searches. The directional constraint mutation strategy involves not using general random perturbation mutation when a particle gets trapped in a local optimum, but instead performing directional mutation based on the thermal safety operation rules of the thermal power unit. Specifically, this may include: When the main steam pressure parameter of a certain particle deviates from the design value of the sliding pressure curve under the current load by more than the preset allowable deviation (i.e. ,in, This represents the main steam pressure parameter value corresponding to the particle. This represents the design value of the main steam pressure corresponding to the unit's design sliding pressure curve under the current load. When the preset allowable deviation is reached, the sliding pressure operation safety protection variation (i.e., directional variation) is triggered, pulling the main steam pressure parameter back towards the design value of the sliding pressure curve; When the parameter combination corresponding to a certain particle causes the temperature difference between the upper and lower cylinders of the steam turbine When the preset safety threshold is exceeded, the turbine thermal stress protection variation (i.e., directional variation) is triggered, and the main steam temperature and reheat steam temperature are simultaneously finely adjusted towards the design value to reduce the thermal stress of the cylinder block.

[0048] The aforementioned directional mutation rules encode the safe operating experience of power plant operators into the constraint logic of the algorithm, ensuring that the parameter combinations generated during the PSO search process are always within the thermally safe feasible domain, thus avoiding the possibility that a general PSO might give parameter recommendations that are "lowest in coal consumption but unsafe in operation".

[0049] Furthermore, to overcome the "conceptual drift" in the actual coal consumption characteristics of thermal power units over long-term operation due to factors such as equipment aging (e.g., slagging on heating surfaces, changes in turbine flow area), coal type changes (fluctuations in calorific value), and changes in equipment condition after maintenance, the model trained on historical data gradually deviates from the current actual equipment characteristics. Therefore, this application establishes an online self-learning model and dynamic baseline curve correction mechanism to achieve closed-loop adaptive operation of the system in the time dimension. Specifically, the deviation between the predicted values ​​of the system's sustainable monitoring model and the actual values ​​collected by the DCS is analyzed. The average absolute percentage error over the most recent W steady-state time slices is calculated using the sliding window statistical method. Preset drift alarm threshold When the mean absolute percentage error The drift alarm threshold is exceeded continuously. The number of times reached the preset number of times When this happens, the model update process is triggered. In addition, when the online calorific value analyzer of the coal entering the furnace detects that the change in the calorific value of the coal exceeds the preset deviation, or when the unit restarts after completing A / B / C level maintenance, a model update is also forcibly triggered.

[0050] After triggering the model update process, the system automatically collects steady-state DCS data within the most recent preset time period, merges it with the original training set, and uses XGBoost's incremental training function to fine-tune the existing model parameters instead of retraining from scratch. Incremental training uses a small learning rate to prevent new data from excessively overwriting historical knowledge. After fine-tuning, the performance of the updated model is tested on the validation set. Replacement is only performed if the new model's accuracy is not lower than the old model; otherwise, the old model is retained and an alarm log is recorded.

[0051] Each time the model update process is triggered, the system can automatically re-execute the characteristic curve generation process described above, i.e., fix the current environmental parameters, scan the entire load segment, and generate an updated load-coal consumption baseline characteristic curve. The system also retains the historical version of the baseline curve and calculates the deviation index between the old and new curves (such as the mean and maximum value of the coal consumption difference for each load segment). This deviation can serve as a quantitative assessment indicator of the degree of equipment degradation. If the curve continues to rise (coal consumption increases under the same load), it indicates that the equipment is experiencing efficiency degradation, which can trigger maintenance and repair decision recommendations.

[0052] Based on the above Figure 1 The specific implementation of the method shown in this embodiment provides a device for predicting coal consumption and optimizing the operation of thermal power units, such as... Figure 2 As shown, the device includes: an acquisition module 31, a determination module 32, a construction module 33, a calculation module 34, and an output module 35; The acquisition module 31 is used to acquire DCS high-frequency data and plant-level low-frequency data of thermal power units as well as current real-time operating condition data, and to perform steady-state screening and multi-source time alignment processing on the DCS high-frequency data and the plant-level low-frequency data to construct a high-dimensional feature matrix containing aligned DCS high-frequency features and low-frequency features. The determination module 32 is used to perform Gaussian kernel density estimation based on the heating ratio characteristics in the high-dimensional feature matrix, identify the bimodal structure of the probability density function, determine the adaptive threshold for heating season and non-heating season based on the minimum point between the bimodal structures, and dynamically classify the operating conditions of the thermal power unit in combination with the fuzzy membership function. The operating conditions include heating season conditions, non-heating season conditions and transition conditions. The construction module 33 is used to construct, for the heating season working condition and the non-heating season working condition respectively, the heating season XGBoost coal consumption prediction model integrating thermodynamic cross-features and customized loss function, and the non-heating season XGBoost coal consumption prediction model. The calculation module 34 is used to calculate the fuzzy membership degree corresponding to the heating ratio of the current real-time operating condition data, input the current real-time operating condition data into the trained heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model respectively, and perform weighted fusion on the output results of the heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model based on the fuzzy membership degree to obtain the predicted value of power generation coal consumption under the current operating condition. Output module 35 is used to minimize the predicted coal consumption for power generation as the objective function, and adopts an improved particle swarm optimization algorithm based on thermal safety boundary to perform reverse optimization under the constraint of the current load command, and outputs the recommended operating parameter values ​​that minimize the coal consumption for power generation under the current real-time operating conditions.

[0053] In specific application scenarios, the acquisition module 31 can be used to calculate the coefficient of variation of key parameters of main steam pressure and unit load using a sliding window, and remove non-steady-state data whose coefficient of variation exceeds a preset steady-state threshold to retain steady-state data in the DCS high-frequency data; for the monthly cumulative statistics in the plant-level low-frequency data, a load-weighted time allocation method is adopted to allocate the monthly cumulative statistics to each steady-state time slice in which the steady-state data is located according to the proportion of the unit load corresponding to each steady-state time slice in the month to the total load of the month, so as to complete the upward alignment in the time dimension; for the monthly ratio statistics in the plant-level low-frequency data, an hourly correction model based on the load correction coefficient function is established to correct the monthly ratio statistics to hourly feature values ​​that match the load level at each time moment, and assign them to the corresponding steady-state data to complete the alignment of the load level; the aligned DCS high-frequency data and the plant-level low-frequency data are fused based on the time index to construct the high-dimensional feature matrix.

[0054] In specific application scenarios, module 32 can be used to perform Gaussian kernel density estimation on the heating ratio dimension, identify the bimodal structure of the probability density function, wherein the first peak of the probability density function corresponds to the concentrated interval of the heating ratio in the non-heating season, and the second peak of the probability density function corresponds to the concentrated interval of the heating ratio in the heating season; extract the minimum point of the probability density function between the first peak and the second peak, and determine the heating ratio value corresponding to the minimum point as the adaptive threshold between the heating season and the non-heating season; set the transition band width parameter, and based on the transition band width parameter... The transition zone is constructed using the transition width parameter and the adaptive threshold. When the heating ratio is greater than or equal to the maximum value of the transition zone, the function value of the fuzzy membership function is determined to be 1, and the operating condition of the thermal power unit is determined to be the heating season condition. When the heating ratio is less than or equal to the minimum value of the transition zone, the function value of the fuzzy membership function is determined to be 0, and the operating condition of the thermal power unit is determined to be the non-heating season condition. When the heating ratio is within the transition zone, the function value of the fuzzy membership function is determined to be the target function value, and the operating condition of the thermal power unit is determined to be the transition condition.

[0055] In specific application scenarios, module 33 can be used to construct high-order cross-features of the input DCS raw parameters based on thermodynamic principles. These high-order cross-features include: main steam temperature deviation, comprehensive flue gas heat loss index, and heating intensity index. A customized loss function is constructed based on the standard mean square error term of predicted coal consumption and actual coal consumption, the thermodynamic monotonicity penalty term, and the operating boundary penalty term. Based on the high-order cross-features and the customized loss function, an XGBoost model is trained on the dataset of the heating season operating conditions to obtain the heating season XGBoost coal consumption prediction model. The model is configured with a first tree depth threshold to capture higher-order interaction effects among multiple variables during the heating season. Based on the higher-order cross features and the customized loss function, an XGBoost model is trained on the dataset of the non-heating season operating conditions to obtain the non-heating season XGBoost coal consumption prediction model. The non-heating season XGBoost coal consumption prediction model is configured with a second tree depth threshold that is less than the first tree depth threshold to avoid overfitting under the pure condensing conditions in the non-heating season. Both the heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model use grid search combined with 5-fold cross-validation for hyperparameter optimization.

[0056] In specific application scenarios, the output module 35 can be used to initialize the particle swarm for iterative optimization with the objective function of minimizing the predicted value of power generation coal consumption and the constraints of the current load command and the safe operating range of each parameter. During the iterative optimization process, an adaptive inertial weight adjustment strategy and a directional constraint mutation strategy based on the thermal safety boundary are introduced to output the recommended values ​​of operating parameters that satisfy the constraints and minimize the predicted value of power generation coal consumption.

[0057] In specific application scenarios, the output module 35 can be used to trigger a sliding pressure operation safety protection variation when the main steam pressure parameter of the particle deviates from the design value of the sliding pressure curve under the current load by more than a preset allowable deviation, thereby pulling the main steam pressure parameter back towards the design value of the sliding pressure curve; when the parameter combination corresponding to the particle causes the temperature difference between the upper and lower cylinders of the turbine to exceed a preset safety threshold, it can trigger a turbine thermal stress protection variation to simultaneously fine-tune the main steam temperature and reheat steam temperature towards the design value, thereby reducing the thermal stress of the cylinder.

[0058] like Figure 2 As shown, the device also includes: a generation module 36; The generation module 36 is used to calculate the average absolute percentage error between the predicted coal consumption of power generation and the actual DCS collected value in each steady-state time slice using the sliding window statistical method; when the average absolute percentage error exceeds the preset drift alarm threshold for a preset number of times, the coal consumption prediction model update process is triggered; and / or, when the change in the calorific value of the coal fed into the furnace is detected to exceed the preset deviation, or when the unit completes maintenance and restarts, the coal consumption prediction model update process is forcibly triggered; after triggering the model update process, steady-state data within a preset time period is collected, and the XGBoost incremental training function is used to fine-tune the model parameters of the coal consumption prediction model; after each trigger of the model update process, the entire load segment is rescanned to generate the updated load-coal consumption baseline characteristic curve, and historical versions are retained for deviation analysis.

[0059] It should be noted that other corresponding descriptions of the functional units involved in the coal consumption prediction and operation optimization device for thermal power units provided in this embodiment can be found in [reference]. Figure 1 The corresponding descriptions in [the document] will not be repeated here.

[0060] Based on the above, Figure 1 Accordingly, this embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Figure 1 The method shown.

[0061] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0062] Based on the above, Figure 1 The method shown, and Figure 2 To achieve the above objectives, the present application also provides an electronic device, comprising a storage medium and a processor; the storage medium for storing a computer program; and the processor for executing the computer program to implement the above-described virtual device embodiments. Figure 1 The method shown.

[0063] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0064] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0065] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical equipment, supporting the operation of the coal consumption prediction and operation optimization program for thermal power units, as well as other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software within the coal consumption prediction and operation optimization physical equipment for thermal power units.

[0066] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms, or it can be implemented by hardware. By applying the solution of this embodiment, compared with the existing technology, this application can obtain DCS high-frequency data and plant-level low-frequency data of thermal power units and current real-time operating condition data, perform steady-state screening and multi-source time alignment processing on DCS high-frequency data and plant-level low-frequency data, and construct a high-dimensional feature matrix containing aligned DCS high-frequency and low-frequency features; perform Gaussian kernel density estimation based on the heating ratio feature in the high-dimensional feature matrix, identify the bimodal structure of the probability density function, determine the adaptive threshold for heating season and non-heating season based on the minimum point between the bimodal structures, and dynamically divide the operating conditions of thermal power units by combining fuzzy membership functions. The operating conditions include heating season conditions, non-heating season conditions, and transition conditions; for heating season conditions and non-heating season conditions, respectively construct integrated thermodynamic cross-features and customized loss The system employs two XGBoost coal consumption prediction models: one for the heating season and one for the non-heating season. It calculates the fuzzy membership degree corresponding to the heating ratio based on the current real-time operating data. The current real-time operating data is then input into the trained XGBoost coal consumption prediction models for both the heating and non-heating seasons. The outputs of these models are weighted and fused based on the fuzzy membership degree to obtain the predicted coal consumption for power generation under the current operating conditions. Using the minimization of the predicted coal consumption for power generation as the objective function, an improved particle swarm optimization algorithm based on thermal safety boundaries is used to perform reverse optimization under the constraints of the current load command, outputting recommended operating parameters that minimize the coal consumption for power generation under the current real-time operating conditions.

[0067] By adopting the above technical solution, this application does not rely on manual experience to set fixed thresholds. Instead, it directly extracts the heating ratio features from a high-dimensional feature matrix that integrates high-frequency DCS and low-frequency plant-level data for probability density analysis. This accurately captures the natural aggregation state of data during the heating season and non-heating season, avoiding the distortion caused by averaging model parameters. The fuzzy membership mechanism enables the unit to smoothly transition prediction results during seasonal transitions, significantly improving the accuracy of coal consumption prediction under all operating conditions (especially complex transitional seasons).

[0068] This application, based on an online reverse optimization mechanism, no longer relies on historical data lookups. Instead, it uses minimizing the predicted coal consumption for power generation as the objective function and employs an improved particle swarm optimization algorithm based on thermal safety boundaries. Under the premise of satisfying current load commands and thermal safety constraints, it performs reverse optimization (i.e., deriving the optimal operating parameters from the target coal consumption) within a multi-dimensional controllable parameter space, directly outputting recommended operating parameter values ​​that minimize coal consumption under current real-time operating conditions. This application can explore better combinations of operating parameters not found in historical data, achieving breakthrough global optimization rather than simply replicating historical bests. This application possesses online closed-loop optimization capabilities, dynamically adjusting according to real-time load commands to ensure the unit always operates at its optimal energy efficiency point under current conditions, significantly improving the unit's economy and flexibility.

[0069] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0070] The above are merely specific embodiments of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to these embodiments, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for predicting coal consumption and optimizing the operation of thermal power units, characterized in that, The method includes: Acquire DCS high-frequency data and plant-level low-frequency data of thermal power units, as well as current real-time operating condition data. Perform steady-state screening and multi-source time alignment processing on the DCS high-frequency data and the plant-level low-frequency data to construct a high-dimensional feature matrix containing aligned DCS high-frequency and low-frequency features. Gaussian kernel density estimation is performed based on the heating ratio characteristics in the high-dimensional feature matrix to identify the bimodal structure of the probability density function. Adaptive thresholds for heating season and non-heating season are determined based on the minimum points between the bimodal structures. The operating conditions of the thermal power unit are dynamically divided by combining the fuzzy membership function. The operating conditions include heating season conditions, non-heating season conditions, and transition conditions. For the heating season operating conditions and the non-heating season operating conditions, respectively, an XGBoost coal consumption prediction model for the heating season and an XGBoost coal consumption prediction model for the non-heating season are constructed, which integrate thermodynamic cross-features and customized loss functions. Calculate the fuzzy membership degree corresponding to the heating ratio of the current real-time operating data, input the current real-time operating data into the trained heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model respectively, and perform weighted fusion of the output results of the heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model based on the fuzzy membership degree to obtain the predicted value of power generation coal consumption under the current operating conditions; Using the minimization of the predicted coal consumption for power generation as the objective function, an improved particle swarm optimization algorithm based on thermal safety boundaries is employed to perform reverse optimization under the constraints of the current load command, and output recommended operating parameter values ​​that minimize the coal consumption for power generation under the current real-time operating conditions.

2. The method according to claim 1, characterized in that, The steady-state screening and multi-source time alignment processing of the DCS high-frequency data and the plant-level low-frequency data to construct a high-dimensional feature matrix containing the aligned DCS high-frequency and low-frequency features specifically includes: The coefficient of variation of key parameters of main steam pressure and unit load is calculated using a sliding window. Non-steady-state data with coefficients of variation exceeding a preset steady-state threshold are removed in order to retain steady-state data in the DCS high-frequency data. For the monthly cumulative statistics in the plant-level low-frequency data, a load-weighted time allocation method is adopted to allocate the monthly cumulative statistics to each steady-state time slice in which the steady-state data is located according to the proportion of the unit load corresponding to each steady-state time slice in the month to the total load of the month, so as to complete the upward alignment in the time dimension. For the monthly ratio statistics in the plant-level low-frequency data, an hourly correction model based on the load correction coefficient function is established to correct the monthly ratio statistics into hourly characteristic values ​​that match the load level at each time point, and assign them to the corresponding steady-state data to complete the alignment of the load level. The aligned DCS high-frequency data and the plant-level low-frequency data are fused based on a time index to construct the high-dimensional feature matrix.

3. The method according to claim 1, characterized in that, The process involves Gaussian kernel density estimation based on the heating ratio characteristics in the high-dimensional feature matrix, identifying the bimodal structure of the probability density function, determining adaptive thresholds for the heating season and non-heating season based on the minimum points between the bimodal structures, and dynamically classifying the operating conditions of the thermal power units using fuzzy membership functions. Specifically, this includes: Gaussian kernel density estimation is performed on the heating ratio dimension to identify the bimodal structure of the probability density function, wherein the first peak of the probability density function corresponds to the concentrated interval of the heating ratio in the non-heating season, and the second peak of the probability density function corresponds to the concentrated interval of the heating ratio in the heating season. Extract the minimum point of the probability density function between the first peak and the second peak, and determine the heating ratio corresponding to the minimum point as the adaptive threshold for the heating season and the non-heating season; Set the transition band width parameter, and construct a transition interval based on the transition band width parameter and the adaptive threshold; When the heating ratio is greater than or equal to the maximum value of the transition interval, the function value of the fuzzy membership function is determined to be 1, and the operating condition of the thermal power unit is determined to be the heating season operating condition. When the heating ratio is less than or equal to the minimum value of the transition interval, the function value of the fuzzy membership function is determined to be 0, and the operating condition of the thermal power unit is determined to be the non-heating season condition. When the heating ratio is within the transition range, the function value of the fuzzy membership function is determined to be the target function value, and the operating condition of the thermal power unit is determined to be the transition condition.

4. The method according to claim 3, characterized in that, The aforementioned XGBoost coal consumption prediction models for the heating season and the non-heating season are constructed respectively, integrating thermodynamic cross-features and customized loss functions. Specifically, these include: The high-order cross-features of the input DCS raw parameters are constructed based on thermodynamic principles. The high-order cross-features include: main steam temperature deviation, comprehensive index of flue gas heat loss, and heating intensity index. The customized loss function is constructed based on the standard mean square error term of predicted coal consumption and actual coal consumption, the thermodynamic monotonicity penalty term, and the operating boundary penalty term; Based on the higher-order cross features and the customized loss function, an XGBoost model is trained on the dataset of the heating season operating conditions to obtain the heating season XGBoost coal consumption prediction model. The heating season XGBoost coal consumption prediction model is configured with a first tree depth threshold to capture the higher-order interaction effects among multiple variables in the heating season. Based on the higher-order cross features and the customized loss function, an XGBoost model is trained on the dataset of the non-heating season operating conditions to obtain the non-heating season XGBoost coal consumption prediction model. The non-heating season XGBoost coal consumption prediction model is configured with a second tree depth threshold that is less than the first tree depth threshold to avoid overfitting under the non-heating season pure condensing operating conditions. The heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model both use grid search combined with 5-fold cross-validation for hyperparameter optimization.

5. The method according to claim 1, characterized in that, The objective function is to minimize the predicted coal consumption for power generation. An improved particle swarm optimization algorithm based on thermal safety boundaries is used to perform reverse optimization under the constraints of the current load command, outputting recommended operating parameters that minimize the coal consumption for power generation under the current real-time operating conditions. Specifically, this includes: With minimizing the predicted coal consumption for power generation as the objective function and the current load command and the safe operating range of each parameter as constraints, the particle swarm optimization is initialized for iterative optimization. During the iterative optimization process, an adaptive inertial weight adjustment strategy and a directional constraint mutation strategy based on the thermal safety boundary are introduced to output recommended operating parameter values ​​that satisfy the constraints and minimize the predicted value of power generation coal consumption.

6. The method according to claim 5, characterized in that, The targeted constraint mutation strategy specifically includes: When the main steam pressure parameter of the particle deviates from the design value of the sliding pressure curve under the current load by more than the preset allowable deviation, the sliding pressure operation safety protection variation is triggered, and the main steam pressure parameter is pulled back towards the design value of the sliding pressure curve. When the parameter combination corresponding to the particle causes the temperature difference between the upper and lower cylinders of the turbine to exceed the preset safety threshold, the turbine thermal stress protection variation is triggered, and the main steam temperature and reheat steam temperature are simultaneously finely adjusted towards the design value to reduce the thermal stress of the cylinder.

7. The method according to claim 1, characterized in that, The method further includes: The sliding window statistical method was used to calculate the average absolute percentage error between the predicted value of power generation coal consumption and the actual DCS collected value in each steady-state time slice. When the average absolute percentage error exceeds the preset drift alarm threshold for a preset number of consecutive times, the coal consumption prediction model update process is triggered; and / or, when the change in the calorific value of the coal fed into the furnace is detected to exceed the preset deviation, or when the unit completes maintenance and restarts, the coal consumption prediction model update process is forcibly triggered. After triggering the model update process, steady-state data within a preset time period is collected, and the XGBoost incremental training function is used to fine-tune the model parameters of the coal consumption prediction model. Each time the model update process is triggered, the entire load segment is rescanned to generate an updated load-coal consumption baseline characteristic curve, and historical versions are retained for deviation analysis.

8. A device for predicting coal consumption and optimizing operation of thermal power units, characterized in that, include: The acquisition module is used to acquire DCS high-frequency data and plant-level low-frequency data of thermal power units, as well as current real-time operating condition data. It performs steady-state screening and multi-source time alignment processing on the DCS high-frequency data and the plant-level low-frequency data to construct a high-dimensional feature matrix containing aligned DCS high-frequency and low-frequency features. The determination module is used to perform Gaussian kernel density estimation based on the heating ratio characteristics in the high-dimensional feature matrix, identify the bimodal structure of the probability density function, determine the adaptive threshold for heating season and non-heating season based on the minimum point between the bimodal structures, and dynamically classify the operating conditions of the thermal power unit in combination with the fuzzy membership function. The operating conditions include heating season conditions, non-heating season conditions and transition conditions. The module is used to construct XGBoost coal consumption prediction models for heating season and non-heating season, respectively, which integrate thermodynamic cross-features and customized loss functions for the heating season and non-heating season operating conditions. The calculation module is used to calculate the fuzzy membership degree corresponding to the heating ratio of the current real-time operating data, input the current real-time operating data into the trained heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model respectively, and perform weighted fusion of the output results of the heating season XGBoost coal consumption prediction model and the non-heating season XGBoost coal consumption prediction model based on the fuzzy membership degree to obtain the predicted value of power generation coal consumption under the current operating conditions; The output module is used to minimize the predicted coal consumption for power generation as the objective function, and employs an improved particle swarm optimization algorithm based on thermal safety boundaries to perform reverse optimization under the constraints of the current load command, and outputs recommended operating parameters that minimize the coal consumption for power generation under the current real-time operating conditions.

9. An electronic device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for predicting coal consumption and optimizing operation of thermal power units as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for predicting coal consumption and optimizing operation of thermal power units as described in any one of claims 1 to 7.

Images