An abnormality diagnosis method and device for a waste heat boiler of a combined cycle unit and a storage medium

Abstract

This invention discloses a method, device, and storage medium for anomaly diagnosis of waste heat boilers in combined cycle power units, relating to the field of power equipment condition monitoring technology. The method includes: acquiring historical operating data of the waste heat boiler; calculating heat parameters based on time-series partitioning and fusing them with historical data to obtain an operating dataset; training a parameter prediction model using the dataset; inputting the operating data to be diagnosed to obtain predicted values ​​of key state parameters and acquiring time-series deviation data between these predicted and actual values; inputting the deviation data into an anomaly detection model to identify anomalies; combining a seasonal trend decomposition algorithm to clarify the anomaly type; converting the anomaly point data into labels and inputting them into a random forest model; combining feature attribution value analysis to quantify the marginal contribution of each operating feature parameter contained in the anomaly point data, locating the root cause parameters and their direction of action; and formulating a feedforward optimization and control strategy based on the root cause parameters and trend components to adjust the parameters of the controllable execution system to eliminate early faults.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to a method, device and storage medium for diagnosing abnormalities in a combined cycle unit waste heat boiler. Background Technology

[0002] With the increasing proportion of renewable energy, gas-fired combined cycle power units have become the main force for deep peak shaving in power grids due to their significant advantages such as flexible start-up and shutdown and rapid response. As a core hub connecting the gas turbine and steam turbine, the operating status of the waste heat boiler directly affects the safety and economic benefits of the entire power plant. However, due to the highly nonlinear, highly coupled, and significant thermal inertia and time lag characteristics of the internal steam-water heat exchange process of the waste heat boiler, the equipment is prone to hidden performance degradation or deviation from design benchmarks under frequent load changes during peak shaving.

[0003] In practice, power plants primarily rely on fixed threshold alarms set by the Distributed Control System (DCS) and regular manual inspections for anomaly monitoring of waste heat boilers. However, this traditional approach has significant limitations. On the one hand, fixed threshold alarms suffer from severe lag, often triggering only when equipment has already experienced a significant malfunction or parameters have severely exceeded limits, failing to provide early warnings of early deterioration phenomena such as minor leaks in heat exchange tubes, sensor drift, or ash buildup on the heat exchange surface. On the other hand, manual inspections are labor-intensive and highly subjective, easily leading to missed or incorrect diagnoses when faced with massive amounts of multi-dimensional operational data.

[0004] To overcome the drawbacks of manual and traditional monitoring, existing technologies have introduced automated data-driven models. However, these tools have significant technical shortcomings. First, most existing models only make overall predictions based on macroscopic indicators such as total unit power, failing to delve into high-dimensional time-series data and effectively separating normal load fluctuations from actual equipment degradation, resulting in poor anti-interference capabilities and a high false alarm rate. Second, existing models generally have "black box" characteristics, often only outputting binary classification results, unable to quantify the contribution of each parameter, and even more difficult to trace and locate the specific root cause parameters leading to anomalies. Furthermore, existing diagnostic systems typically lack closed-loop control mechanisms linked to lower-level control systems, preventing diagnostic results from being directly translated into practical operational guidance for production.

[0005] In summary, existing waste heat boiler monitoring methods, due to their over-reliance on manual labor, limited model functionality, and low levels of intelligence and interpretability, are unable to accurately, proactively, and comprehensively address early anomalies. This not only increases the risk of unplanned unit shutdowns but also results in significant energy waste, ultimately impacting the overall operational efficiency of the power plant. Therefore, there is an urgent need for an anomaly diagnosis solution that can intelligently decouple temporal characteristics, possesses high interpretability, and enables closed-loop optimization. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method, device, and storage medium for diagnosing abnormalities in waste heat boilers of combined cycle units.

[0007] In a first aspect, embodiments of the present invention propose a method for diagnosing abnormalities in a waste heat boiler of a combined cycle unit, the method comprising: Historical operating data of the waste heat boiler of the combined cycle unit is obtained, noise reduction and time-series segmentation are performed on the historical operating data, heat parameters are calculated based on the segmented data, and the historical operating data and heat parameters are fused to obtain the operating dataset. A parameter prediction model is trained using the operational dataset. The operational data to be diagnosed is input into the parameter prediction model to obtain the predicted values ​​of the key state parameters of the waste heat boiler. The time-series deviation data between the predicted values ​​and actual values ​​of each key state parameter is also obtained. The time series deviation data is input into the anomaly detection model to identify outlier data points. The time series deviation data is then split into trend components, seasonal components, and residual components using a seasonal trend decomposition algorithm to clarify the anomaly type. Outlier data is converted into Boolean anomaly labels and input into a random forest model for parameter feature interaction analysis. The feature attribution value method is combined to quantify the marginal contribution of each running feature parameter contained in the outlier data to the prediction results. Based on the marginal contribution, the root cause parameters that cause the anomalies and the direction of the effect of the root cause parameters on the abnormal state are located. Based on the identified root cause parameters and trend components, a feedforward optimization and control strategy for multi-objective parameters is formulated, and early failures are eliminated by adjusting the controllable execution system parameters of the waste heat boiler.

[0008] In one optional implementation, a parameter prediction model is trained using an operational dataset. The operational data to be diagnosed is then input into the parameter prediction model to obtain predicted values ​​of the key state parameters of the waste heat boiler, including: The parameter prediction model includes three parallel recurrent neural network sub-models: a flue gas heat prediction sub-model, a turbine power prediction sub-model, and a steam temperature prediction sub-model. Environmental parameters, compressor inlet guide vane opening, and overall unit load command are extracted from the operational data to be diagnosed as input features of the flue gas heat prediction sub-model. Long-term dependencies are extracted through the control mechanisms of forget gate, input gate, and output gate in the sub-model, and the predicted value of flue gas heat at the inlet of the waste heat boiler is output. The waste heat boiler inlet flue gas parameters and condenser pressure are extracted from the operating data to be diagnosed as input features of the turbine power prediction sub-model, and the turbine dynamic power prediction value is output. The inlet flue gas parameters of the waste heat boiler and the pressure and flow rate of each stage of feedwater are extracted from the operating data to be diagnosed as input features of the steam temperature prediction sub-model, and the predicted value of the outlet steam temperature of the waste heat boiler is output. The predicted values ​​of flue gas heat at the inlet of the waste heat boiler, the predicted values ​​of dynamic power of the steam turbine, and the predicted values ​​of steam temperature at the outlet of the waste heat boiler together constitute the key state parameters.

[0009] In one optional implementation, time-series deviation data is input into an anomaly detection model to identify outlier data, including: An unsupervised temporal reconstruction network containing an encoder and a decoder is constructed to convert temporal deviation data into a deviation sequence matrix to be diagnosed. By utilizing the attention mechanism in the unsupervised temporal reconstruction network, the similarity weights between the current deviation sequence matrix and the historical normal operation sequence key matrix are calculated, and the similarity weights are transformed into the contribution distribution of the historical sequence to the current diagnostic sequence. The time series deviation data are reconstructed based on the contribution distribution, and the mean absolute error between the original deviation data and the reconstructed data is calculated as the reconstruction error loss value. Within the set time window length, when the reconstruction error loss value is greater than the set dynamic adaptive threshold, it is determined that there is a deviation in the operating mode at the current time step, and abnormal point data is identified and output.

[0010] In one alternative implementation, a seasonal trend decomposition algorithm is used to split the time-series deviation data into trend components, seasonal components, and residual components to identify anomaly types, including: The local weighted regression scatter smoothing method is used to perform additive decomposition on the time series deviation data in order to extract the trend component, seasonal component and residual component from the time series deviation data respectively; Determine whether outlier data is caused by a continuous deviation of the trend component from the baseline; If the anomaly is determined to be caused by a continuous deviation of the trend component, then the interference of the seasonal component and the residual component is filtered out, and the anomaly type is locked as a performance degradation anomaly.

[0011] In one alternative implementation, outlier data is converted into Boolean outlier labels and input into a random forest model for parametric feature interaction analysis, including: The healthy operating condition time steps output by the unsupervised time series reconstruction network are marked as zero, and the time steps of data identified as outliers are marked as one, thus constructing a Boolean outlier label dataset containing binary categorical variables. Cross-validation algorithm is used to optimize the hyperparameter combination of the random forest model; Calculate the decrease in Gini impurity of the input feature to be evaluated before and after splitting the data at each decision tree node of the random forest model, and obtain the information gain of the feature; Information gain is used to evaluate the correlation weight of steam and feedwater parameters with the occurrence of abnormal states, and a multi-parameter feature importance ranking matrix is ​​output.

[0012] In one optional implementation, the marginal contribution of each operational feature parameter contained in the outlier data to the prediction result is quantified using the feature attribution value method. Based on the marginal contribution, the root cause parameters leading to the anomaly and the direction of their influence on the anomaly state are located, including: Traverse subsets of feature parameters in the random forest model and calculate the marginal change in the model's anomaly probability prediction output value before and after introducing specific state parameters. Calculate the weighted average of the marginal changes corresponding to all possible permutations and combinations of feature subsets to obtain the marginal contribution. Extract the parameters whose absolute value of marginal contribution ranks among the top by a preset limit as the root parameters; Calculate the standardized feature quantization value of the root cause parameter, and determine whether the root cause parameter's effect on the abnormal state is to exacerbate or suppress the abnormality based on the mapping relationship between the standardized feature quantization value and the positive or negative sign of the marginal contribution.

[0013] In one alternative implementation, a multi-objective parameter feedforward optimization control strategy is formulated based on the identified root cause parameters and trend components. Early failures are eliminated by adjusting the controllable execution system parameters of the waste heat boiler, including: Based on the set of root cause parameters that exacerbate the anomalies, they are mapped to the corresponding physical actuators of the waste heat boiler. If the root cause parameter is determined to be high-pressure water flow or medium-pressure water pressure, and the value of the corresponding parameter is lower than the health threshold, then the first feedforward optimization instruction is generated to increase the speed of the water pump inverter or adjust the opening of the water supply regulating valve. If the root cause parameter is the reheat desuperheating water flow rate parameter, then a second feedforward optimization instruction is generated to execute the desuperheating water valve action based on temperature deviation feedforward compensation. The first and second feedforward optimization instructions are sent to the distributed control system for execution.

[0014] In one alternative implementation, after formulating a feedforward optimization control strategy for multi-objective parameters, the method further includes: Collect and update system feedback data after implementing the feedforward optimization and control strategy; Replace the trend component in the system feedback update data with the baseline prediction data of the parametric prediction model under healthy conditions, and calculate the expected difference of performance parameters before and after the replacement. Calculate the average increase in steam temperature and the average increase in steam cycle output of the waste heat boiler under steady-state conditions based on the expected difference in performance parameters. By comparing the expected increase with the actual monitoring feedback, the step size and gain parameters of the feedforward optimization control strategy are dynamically fine-tuned.

[0015] Secondly, embodiments of the present invention provide an abnormal diagnosis device for a combined cycle unit waste heat boiler, the device comprising: The data preprocessing module is used to acquire historical operating data of the waste heat boiler of the combined cycle unit, perform noise reduction and time series division on the historical operating data, calculate heat parameters based on the divided data, and fuse the historical operating data and heat parameters to obtain the operating dataset. The parameter prediction module is used to train and build a parameter prediction model using the running dataset. The running data to be diagnosed is input into the parameter prediction model to obtain the predicted values ​​of the key state parameters of the waste heat boiler, and to obtain the time-series deviation data between the predicted values ​​and actual values ​​of each key state parameter. The reconstruction detection module is used to input time series deviation data into the anomaly detection model, identify outlier data, and combine the seasonal trend decomposition algorithm to split the time series deviation data into trend component, seasonal component and residual component to clarify the anomaly type. The attribution analysis module is used to convert outlier data into Boolean outlier labels and input them into the random forest model for parameter feature interaction analysis. It combines feature attribution value method to quantify the marginal contribution of each running feature parameter contained in the outlier data to the prediction results, and locates the root cause parameter that caused the anomaly and the direction of the effect of the root cause parameter on the abnormal state based on the marginal contribution. The feedforward optimization module is used to formulate a multi-objective parameter feedforward optimization control strategy based on the identified root cause parameters and trend components, and to eliminate early failures by adjusting the controllable execution system parameters of the waste heat boiler.

[0016] Beneficial effects: This invention provides a method for diagnosing anomalies in waste heat boilers of combined cycle power units. It integrates calculated thermodynamic parameters with historical time-series data to construct a high-quality operational dataset that considers both physical mechanisms and data-driven approaches. This dataset is then used to train a parameter prediction model for high-precision inference on the data to be diagnosed, obtaining time-series deviation data. By combining the anomaly detection model with seasonality and trend decomposition algorithms to clarify anomaly types, it achieves automatic decoupling of the deep logic within multi-dimensional operational data, effectively overcoming the limitations of existing technologies that rely on fixed threshold alarms leading to delayed warnings, and the difficulty of existing models in distinguishing between normal operating fluctuations and actual equipment degradation. By accurately identifying anomalies using the anomaly detection model, and based on this, constructing a random forest model combined with feature attribution value analysis, it intelligently and quantitatively analyzes the marginal contribution of each operational feature parameter in the anomaly point data to the prediction results, greatly breaking the "black box" characteristic of traditional deep learning models. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention; Figure 2 This is a flowchart of the steps in a method for diagnosing abnormalities in a combined cycle unit waste heat boiler, provided by an embodiment of the present invention. Figure 3 This is a schematic diagram of the functional modules of a combined cycle unit waste heat boiler abnormality diagnosis device provided in an embodiment of the present invention. Detailed Implementation

[0018] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the drawings is only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the invention, but does not constitute a limitation on the invention.

[0019] In the description of this invention, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Furthermore, terms such as "module," "model," and "network" used in this specification generally refer to computer-related entities as hardware, software, or a combination of both. The invention will be further described below with reference to the accompanying drawings.

[0020] Reference Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0021] like Figure 1As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0022] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0023] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and electronic programs.

[0024] exist Figure 1 In the electronic device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the combined cycle unit waste heat boiler abnormal diagnosis device stored in the memory 1005 through the processor 1001 and executes the combined cycle unit waste heat boiler abnormal diagnosis method provided in the embodiment of the present invention.

[0025] This invention proposes a method for diagnosing anomalies in waste heat boilers of combined cycle power units. This method is primarily executed by computer programs or instructions deployed on edge computing servers or cloud data centers. (See attached...) Figure 2 As shown, this method includes a complete data closed-loop process, specifically including the following steps S101 to S105: S101: Obtain historical operating data of the waste heat boiler of the combined cycle unit, perform noise reduction and time-series division on the historical operating data, calculate heat parameters based on the divided data, and fuse the historical operating data and heat parameters to obtain the operating dataset. In specific industrial applications, waste heat boilers operate in extremely harsh environments with numerous measurement points. In this step, using the Supervisory Information System (SIS) and the historical database of the DCS, and employing the Object Linking and Embedded Process Control (OLE) protocol, multi-dimensional time-series sensor data from the combined cycle unit's waste heat boiler are continuously extracted at sampling periods of 1 to 10 seconds. This data covers hundreds of measurement points, including but not limited to: ambient temperature, ambient humidity, atmospheric pressure, compressor inlet guide vane (IGV) opening, gas turbine exhaust temperature, exhaust flow rate, high / medium / low pressure feedwater pressure and flow rate, and main steam temperature.

[0026] Because industrial sensors are inevitably affected by electromagnetic interference and measurement thermal inertia, the raw data collected contains a large number of high-frequency spikes and missing values. Therefore, rigorous noise reduction and cleaning are necessary. First, an outlier removal algorithm based on the Wright criterion (3-Sigma rule) is used to treat abrupt data points deviating more than three standard deviations from the mean as non-physical true values ​​caused by sensor communication failures and remove them. For the missing values ​​resulting from the removal, cubic spline interpolation is used to fill them. Subsequently, in order to eliminate high-frequency random fluctuations during normal load changes, a minute-level moving average filter is used to smooth the data, with the moving window length set to 5 to 15 minutes, thereby obtaining time-series data that reflects the true thermodynamic trend of the equipment.

[0027] Furthermore, to enable subsequent machine learning models to directly learn energy conservation and heat transfer mechanisms, this embodiment does not rely solely on the original temperature and pressure values, but introduces thermodynamic standard conversion calculations. Based on the industry standard published by the International Association for the Properties of Water and Steam (IAPWS-IF97) and the property database provided by the National Institute of Standards and Technology (NIST REFPROP), the specific enthalpy and specific entropy of the working fluid at each node are dynamically calculated using temperature and pressure data, thereby calculating the total heat carried by the flue gas at the waste heat boiler inlet and the heat parameters absorbed by steam at each stage.

[0028] After completing the above cleaning and feature engineering (i.e., calculating heat parameters based on the partitioned data), the preprocessed high-dimensional dataset is fused with the heat parameters according to the time series sequence, and further partitioned into training sets to obtain the final high-quality running dataset, providing reliable data support for the subsequent construction of prediction models.

[0029] S102, a parameter prediction model is trained and constructed using the running dataset. The running data to be diagnosed is input into the parameter prediction model to obtain the predicted values ​​of the key state parameters of the waste heat boiler, and the time-series deviation data between the predicted values ​​and actual values ​​of each key state parameter is obtained. Traditional prediction models often employ multiple linear regression or shallow artificial neural networks, which struggle to capture the dynamic time lags in the thermal process of waste heat boilers. This embodiment employs deep learning technology, constructing an architecture based on Long Short-Term Memory (LSTM) networks within the parameter prediction model. To achieve decoupling of multiple variables and accurate prediction, this step specifically includes the following sub-steps S1021 to S1024: S1021, the parameter prediction model includes three parallel recurrent neural network sub-models, namely the flue gas heat prediction sub-model, the turbine power prediction sub-model, and the steam temperature prediction sub-model. Waste heat boilers are complex nonlinear multi-input multi-output systems. Training all parameters in a single unified network can easily lead to gradient vanishing and conflicting objectives in multi-task learning. Therefore, this step instantiates three independent, parallel LSTM sub-networks in the software architecture, which are trained with specialized feature maps for thermal input (flue gas heat), energy output (turbine power), and core state variables (steam temperature), respectively. This significantly reduces the hypothesis space complexity of each sub-model and improves prediction accuracy.

[0030] S1022, extract environmental parameters, compressor inlet guide vane opening degree and unit overall load command from the operating data to be diagnosed as input features of the flue gas heat prediction sub-model, extract long-term dependencies through the control mechanism of forget gate, input gate and output gate in the sub-model, and output the predicted value of flue gas heat at the inlet of the waste heat boiler. Since the exhaust state of a gas turbine is mainly constrained by the atmospheric environment and the boundary conditions of the gas turbine operation, atmospheric ambient temperature, relative humidity, and ambient pressure (i.e., the environmental parameters) are extracted and combined with the compressor inlet guide vane opening and the overall unit load command issued by the dispatch center to form a multi-dimensional input feature vector matrix.

[0031] The input matrix is ​​sequentially fed into the LSTM cells of the flue gas heat prediction sub-model at each time step. Inside each cell, a weight vector between 0 and 1 is first calculated using the sigmoid activation function in the "forget gate," determining which lagging thermal inertia information from historical cell states needs to be discarded. Next, the sigmoid and hyperbolic tangent (tanh) functions in the "input gate" extract newly added valid information from the current time step's environment and load commands, updating the cell state accordingly. Finally, the "output gate" evaluates the current cell state and hidden layer state, generating the predicted flue gas heat value for the waste heat boiler inlet for the next time step. This mechanism perfectly matches the thermal response delay characteristics of gas turbine exhaust systems, which can last for several minutes.

[0032] S1023, extract the waste heat boiler inlet flue gas parameters and condenser pressure from the operating data to be diagnosed as input features of the turbine power prediction sub-model, and output the turbine dynamic power prediction value. When predicting the work capacity of a combined cycle unit, the energy input at the inlet side of the waste heat boiler and the cold-end boundary conditions are decisive factors. Therefore, the inlet flue gas flow rate, temperature, and composition (i.e., the inlet flue gas parameters of the waste heat boiler) involved in step S1022, along with the condenser pressure indicating the cold-end back pressure of the thermal cycle, are input into the turbine power prediction sub-model. This sub-model also undergoes multi-step temporal convolution and mapping using LSTM layers to finally fit the real-time dynamic power prediction value of the turbine under dynamic load conditions.

[0033] S1024, extract the inlet flue gas parameters of the waste heat boiler and the pressure and flow rate of each stage of feedwater from the operating data to be diagnosed as... output the predicted value of the outlet steam temperature of the waste heat boiler; the predicted value of the inlet flue gas heat of the waste heat boiler, the predicted value of the turbine dynamic power, and the predicted value of the outlet steam temperature of the waste heat boiler together constitute the key state parameters.

[0034] Steam temperature is the most direct indicator for evaluating the heat exchange efficiency of the heating surface of a waste heat boiler. Using the inlet flue gas parameters of the waste heat boiler, as well as the feedwater pressure and flow rate delivered by the high-pressure, medium-pressure, and low-pressure feedwater pumps, as forward inputs, the steam temperature prediction sub-model outputs predicted values ​​for the outlet steam temperature, including the high-pressure main steam temperature and the reheat steam temperature.

[0035] After all sub-models have completed forward inference, the system performs element-wise subtraction on the predicted value matrix output by these three sub-models and the actual measurement value matrix collected by the DCS. The resulting time series residual is the aforementioned time series deviation data. Under ideal healthy conditions, this time series deviation data should appear as a white noise sequence with a mean close to zero.

[0036] S103, input the time series deviation data into the anomaly detection model, identify outlier data, and combine the seasonal trend decomposition algorithm to split the time series deviation data into trend component, seasonal component and residual component to clarify the anomaly type; The timing deviation data contains a large number of normal fluctuations caused by unit start-up and shutdown, and grid frequency regulation. To accurately extract the abnormal features characterizing equipment degradation from these complex fluctuations, this step specifically includes two major stages: anomaly identification and feature decomposition, namely sub-steps S1031 to S1038: S1031, Construct an unsupervised temporal reconstruction network containing an encoder and a decoder to convert the temporal deviation data into a deviation sequence matrix to be diagnosed; By slicing the time-series aberration data using a sliding time window, a one-dimensional time series is mapped to a high-dimensional matrix of aberration sequences to be diagnosed. Subsequently, an unsupervised time-series reconstruction network based on deep learning is constructed. This network employs an encoder-decoder architecture, the core idea of ​​which is to compress the healthy sequence matrix into low-dimensional latent variables, and then reconstruct the high-dimensional sequence from the latent variables. For unseen aberration sequences, the network will be unable to accurately reconstruct them.

[0037] S1032, using the attention mechanism in the unsupervised temporal reconstruction network, calculate the similarity weight between the current deviation sequence matrix and the historical normal operation sequence key matrix, and transform the similarity weight into the contribution distribution of the historical sequence to the current diagnostic sequence; Internally, a Scaled Dot-Product Attention mechanism is introduced. The deviation sequence matrix of the current diagnosis is mapped to a query matrix through a linear transformation, and the latent space representation of historical data obtained by the model under healthy operating conditions is mapped to a key matrix and a value matrix. The similarity weights are obtained by calculating the inner product of the query matrix and the key matrix, dividing by the square root of the dimensionality scaling factor, and then normalizing using the Softmax function. This weight matrix intuitively reflects the distribution of the contribution of historical normal operating conditions to the data at the moment of the current diagnosis, enabling the model to automatically focus on the most important feature bands in the time series data and ignore irrelevant measurement noise.

[0038] S1033, reconstruct the time-series deviation data based on the contribution distribution, and calculate the average absolute error between the original deviation data and the reconstructed data as the reconstruction error loss value; The decoder uses the context vector and contribution distribution output from the attention layer to perform inverse projection reconstruction of the original signal, obtaining the reconstructed time series sequence. Subsequently, the difference between the original undiagnosed bias data and the model-reconstructed data is calculated step-by-step. This embodiment uses Mean Absolute Error (MAE) to quantify this reconstruction error loss value, calculated as follows:

[0039] In the formula, This represents the reconstruction error loss value within the time window; This indicates the total number of data sampling points within the set time window length; Indicates the first The original timing deviation data values ​​for each time step; This represents the output of the unsupervised temporal reconstruction network. Reconstructed data values ​​at each time step.

[0040] S1034, within the set time window length, when the reconstruction error loss value is greater than the set dynamic adaptive threshold, it is determined that there is a deviation in the running mode at the current time step, and the abnormal point data is identified and output. Traditional fixed thresholds are prone to false alarms. This method employs a dynamic adaptive threshold setting mechanism based on Extreme Value Theory (EVT). By fitting the probability density distribution of reconstruction errors from the validation set's normal data, a threshold curve that varies with the confidence interval is dynamically generated. Once the current time step... If the data exceeds the upper limit of the dynamic adaptive threshold, the system will immediately trigger an alarm, determine that the current time series is in an abnormal mode, and mark this data segment as an anomaly.

[0041] S1035, The time-series deviation data are additively decomposed using the local weighted regression scatter smoothing method; After identifying the anomaly, it remains unclear whether the anomaly is caused by accidental fluctuations in the sensor or by continuous degradation such as scaling inside the equipment. At this point, the STL algorithm is introduced. The STL algorithm is a highly robust time-series decomposition method based on Loess-based locally weighted regression scatter smoothing.

[0042] S1036, to extract the trend component, seasonal component and residual component from the time series deviation data respectively; Through a nested Loess smoothing process involving inner and outer loops, this model transforms one-dimensional time-series skew data... According to the additive model, it can be decomposed into three independent eigenvectors:

[0043] In the formula, Indicates time step The original time series deviation data below; The extracted trend component smoothly reflects the macroscopic evolution of the deviation data over a longer period. The seasonal component is used to characterize the regular fluctuations in data caused by the daily periodic increases and decreases in load due to grid dispatch requirements. The residual component contains high-frequency random noise or transient electromagnetic interference that cannot be explained by trends and seasonality.

[0044] S1037, determine whether the outlier data is caused by the continuous deviation of the trend component from the baseline; The system calculates the separated trend component. The first derivative (slope) and cumulative integral area within the abnormal time window. If the slope remains continuously monotonic and the integral area exceeds a certain tolerance band of the healthy baseline, then the anomaly is not a sudden disturbance.

[0045] S1038, if it is determined that the problem is caused by the continuous deviation of the trend component, then the interference of the seasonal component and the residual component is filtered out, and the anomaly type is locked as performance degradation anomaly.

[0046] Through this decoupling of features with clear physical meaning, the algorithm intelligently filters... (Normal peak-shaving fluctuations) and (Sensor noise), only for The system traces long-term deviations. This allows the system to confidently classify anomalies as "early performance degradation anomalies caused by controllable parameters" (such as a long-term decrease in heat transfer coefficient due to ash accumulation on heat exchanger tube walls), rather than pseudo-anomalies caused by power grid fluctuations.

[0047] S104. Convert the outlier data into Boolean outlier labels and input them into the random forest model for parameter feature interaction analysis. Combine the feature attribution value method to quantify the marginal contribution of each running feature parameter contained in the outlier data to the prediction results. Based on the marginal contribution, locate the root parameters that cause the anomalies and the direction of the effect of the root parameters on the abnormal state. To overcome the "black box" problem of uninterpretability in deep learning neural networks, this step utilizes ensemble tree models and game theory principles from machine learning to trace the source of abnormal states. Specifically, it includes the following sub-steps S1041 to S1048: S1041, mark the health condition time step output by the unsupervised temporal reconstruction network as zero, mark the time step that is determined to be the abnormal point data as one, and construct the Boolean type abnormal label dataset containing binary classification variables. The operating condition segments previously defined by unsupervised detection are digitally encoded. A safe and stable operating state is assigned a Boolean value of 0, while time-series features exhibiting performance degradation are assigned a Boolean value of 1. This label column is then combined with the corresponding multi-dimensional operating input features to create a generalized state table for supervised learning.

[0048] S1042, The cross-validation algorithm is used to optimize the hyperparameter combination of the random forest model; A random forest algorithm based on decision tree ensembles is introduced. To prevent the model from overfitting on massive industrial data, K-fold (e.g., 5-fold) grid search cross-validation is employed. During training, the system automatically traverses and tests hundreds of parameter combinations, ultimately selecting the optimal hyperparameters that minimize the generalization error. These parameters mainly include: the total number of decision trees (n_estimators), the maximum growth depth of each tree (max_depth), and the minimum number of samples that a leaf node must contain (min_samples_leaf).

[0049] S1043, calculate the decrease in Gini impurity of the input feature to be evaluated before and after splitting the data at each decision tree node of the random forest model, and obtain the information gain of the feature; After the RF model is trained, for any feature parameter (e.g., a certain level of water supply pressure), the algorithm backtracks to see how the confounding degree of the target category (abnormal label 1 and normal label 0) is improved when that feature is used as a splitting condition in all nodes within the decision tree. Specifically, this involves calculating the difference between the Gini Impurity before the node split and the weighted sum of the Gini Impurities of the left and right child nodes after the split.

[0050] S1044, use the information gain to evaluate the correlation weight of steam parameters and feedwater parameters with the occurrence of abnormal states, and output a multi-parameter feature importance ranking matrix.

[0051] By averaging the Gini impurity reduction (i.e., information gain) of a particular feature across all decision trees, the global correlation weight of that feature can be obtained. After normalizing the correlation weights of all steam and feedwater parameters, they are arranged from largest to smallest to generate a feature importance ranking matrix. This step enables preliminary screening of hundreds or thousands of boiler measuring points, eliminating a large number of redundant parameters unrelated to deterioration anomalies.

[0052] S1045, Traverse the subset combinations of feature parameters in the random forest model and calculate the marginal change in the model's anomaly probability prediction output value before and after introducing specific state parameters. While RF outputs an importance matrix, it cannot specify whether a particular feature "initiates" or "suppresses" anomalies. Therefore, a Shapley Additive Explanations (SHAP) attribution value method based on cooperative game theory is introduced. This algorithm treats each parameter as a "player" in the game and the anomaly probability predicted by the final model as the "total prize." By exhaustively enumerating all subset combinations of feature parameters, the algorithm calculates the impact of adding a specific parameter to a subset. When, the model outputs an increment of probability.

[0053] S1046, Calculate the weighted average of the marginal changes corresponding to all possible permutations and combinations of feature subsets to obtain the marginal contribution. For specific operating parameters Its marginal contribution of SHAP to the anomaly prediction results at the current time step The specific calculations are as follows:

[0054] In the formula, To represent factorial, This represents the complete set of all state parameter features involved in training the random forest model; Represents the feature set Specific parameters have been removed. A specific subset of parameters thereafter; and These represent the number of parameter dimensions contained in the entire set and the subset, respectively. Indicates using only a subset of parameters The features in the model are the output predicted values ​​of the probability of anomalies occurring. Indicates in subset Add parameters to the base The output value predicted by the model; the difference inside the parentheses is the marginal change mentioned in the previous step; the preceding fractional term is the combination weight factor assigned to this subset based on permutation and combination theory.

[0055] S1047, extract the parameters whose absolute value of marginal contribution ranks among the top in a preset order as the root source parameters; Calculate the SHAP value of all parameters at a given anomalous moment. . The larger the absolute value, the stronger the effect of the parameter on the current model's judgment of system anomalies. The system sorts the absolute values ​​in descending order and usually extracts the top 3 to 5 characteristic parameters, which are then identified as the root cause parameters of the current waste heat boiler deterioration (such as the outlet temperature of a certain stage desuperheater, the opening degree of a certain high-pressure feedwater regulating valve, etc.).

[0056] S1048, calculate the standardized feature quantization value of the root cause parameter, and determine whether the direction of the effect of the root cause parameter on the abnormal state is to exacerbate the abnormality or suppress the abnormality based on the mapping relationship between the standardized feature quantization value and the positive and negative signs of the marginal contribution.

[0057] The current measured value is subtracted from the historical average value of the identified root cause parameter under healthy conditions, and then divided by the historical standard deviation to obtain a level of standard normalization (quantization value). If a parameter is currently abnormally high (positive feature quantization value) and its calculated SHAP marginal contribution is also positive (increasing the anomaly probability output), it means that the high parameter "exacerbates" the anomaly, and this is the parameter we need to suppress. Conversely, if the SHAP value is negative, it means that it is trying to bring the system back to normal, which is the direction of "suppressing anomalies". This step transforms cold probability numbers into a mechanical direction with great engineering guidance.

[0058] S105, based on the identified root cause parameters and trend components, formulates a feedforward optimization control strategy for multi-objective parameters, and eliminates early failures by adjusting the controllable execution system parameters of the waste heat boiler.

[0059] Traditional big data diagnostics often stop at issuing diagnostic reports. This further breaks down the physical barriers between the information layer and the control execution layer, forming a closed-loop architecture with self-healing capabilities. Specifically, it covers sub-steps S1051 to S1058: S1051, based on the set of root cause parameters that exacerbate the anomaly, map them to the corresponding physical actuator of the waste heat boiler; Retrieve the pipeline and instrumentation topology database of the unit, map and bind the "exacerbation anomaly" parameters traced in the previous stage to the actual operable actuators on site (such as variable frequency feedwater pump motors, pneumatic regulating valves, inlet flue gas dampers, etc.).

[0060] S1052, if it is determined that the root cause parameter is high pressure water flow or medium pressure water pressure, and the value of the corresponding parameter is lower than the health threshold, then a first feedforward optimization instruction is generated to increase the speed of the water pump frequency converter or adjust the opening of the water supply regulating valve. In the specific control logic generation stage, if the system attribution indicates that the source of the oscillation in the high-pressure steam drum water level and the resulting abnormal main steam temperature is insufficient pressure in the high-pressure feedwater header, the system will automatically calculate the required pressure gain compensation. Subsequently, it will generate a first feedforward optimization instruction directly targeting the high-pressure variable frequency feedwater pump controller (which has proportional-integral-derivative PID functionality). This instruction adds a feedforward signal on top of the original cascade water level control in the DCS, increasing the inverter speed in advance to compensate for the flow loss caused by the pipeline fluid resistance.

[0061] S1053, if the root cause parameter is the reheat desuperheating water flow rate parameter, then generate a second feedforward optimization instruction and execute the desuperheating water valve action based on temperature deviation feedforward compensation. If the reheater outlet temperature exceeds the limit due to a shift in the heat distribution on the boiler flue gas side, and SHAP analysis indicates that the sluggish response of the reheat desuperheating water is the core cause of the exacerbation of the anomaly, the system uses the slope of the "trend component" extracted in step S1036 as a derivative feedforward to predict the peak temperature to be reached several minutes later and generates a second feedforward optimization command to slightly open the desuperheating water regulating valve in advance. This effectively overcomes the measurement time constant lag problem caused by the outer sheath of the thermocouple and prevents thermal stress fatigue of the tube wall caused by drastic cooling.

[0062] S1054, the first feedforward optimization command and the second feedforward optimization command are sent to the distributed control system for execution.

[0063] By utilizing bidirectional industrial communication standards such as the OPC UA unified architecture protocol, the generated optimized bias commands are written as external setpoints into the plant-level DCS network, and the field actuators drive the water pumps or valves to respond to the action.

[0064] S1055, Collect system feedback update data after executing the feedforward optimization control strategy; After the system issues a command, it does not stop there, but continues to monitor the effect. The latest time-series measurement data of the sensors after the controlled object responds is collected within a set observation window (e.g., 15 to 30 minutes after the optimized action), which is the system feedback update data.

[0065] S1056, replace the trend component in the system feedback update data with the prediction baseline data of the parameter prediction model in a healthy state, and calculate the expected difference of performance parameters before and after the replacement; The latest feedback data is fed back into the STL decomposition module to extract the trend term resulting from the executed operation. Then, the baseline data representing ideal healthy operating conditions, generated by the original LSTM prediction model, is used to compare this new trend term. This substitution comparison mechanism removes the interference of external operating condition changes, yielding the expected difference in performance parameters purely resulting from the optimization strategy.

[0066] S1057, Calculate the average increase in steam temperature and the average increase in power output of the waste heat boiler under steady-state conditions based on the expected difference in the performance parameters. This transforms microscopic deviations into macroscopic economic and technical indicators. For example, calculations within a specific peak-shaving steady-state range reveal that, after optimization, the average decrease in high-pressure main steam temperature and reheat steam temperature compared to the uninterrupted trend increased by approximately 2.30°C to 4.68°C. This, in turn, leads to an increase in the enthalpy drop of the steam entering the turbine to perform work, resulting in an average increase in the output power of the steam cycle system at the megawatt (MW) level. This directly quantifies the economic benefits of eliminating early boiler deviations.

[0067] S1058, by comparing the expected increase with the actual monitoring feedback, dynamically fine-tunes the step size and gain parameters of the feedforward optimization control strategy.

[0068] If the actual temperature rise calculated by monitoring has not reached the ideal recovery range predicted by the LSTM model, or if over-adjustment occurs (e.g., excessive valve opening leading to excessive cooling), it indicates that the proportional gain of the feedforward command is too small or too large. In this case, the algorithm inputs the closed-loop feedback error into the adaptive controller to fine-tune the action step size coefficient and gain proportional coefficient of the subsequently sent first and second commands, thereby ensuring that the waste heat boiler control loop can achieve smooth and stable self-healing operation throughout the entire peak-shaving life cycle.

[0069] Through the complete software and hardware execution process described above, this invention creatively integrates deep time-series feature mining and model interpretability algorithms, successfully and accurately pinpointing the root causes of equipment degradation in the vast ocean of power plant data, and eliminating early hidden dangers through automated feedforward intervention, thus ensuring the efficient and safe operation of generator sets under complex and ever-changing power grid demands.

[0070] This invention also provides an abnormal diagnosis device for waste heat boilers in combined cycle units, referring to... Figure 3 The diagram shows a functional unit diagram of a combined cycle unit waste heat boiler anomaly diagnosis device 200 according to the present invention. The device may include the following modules: The data preprocessing module 201 is used to acquire historical operating data of the waste heat boiler of the combined cycle unit, perform noise reduction and time series division on the historical operating data, calculate heat parameters based on the divided data, and fuse the historical operating data and heat parameters to obtain the operating dataset. The parameter prediction module 202 is used to train and build a parameter prediction model using the running dataset, input the running data to be diagnosed into the parameter prediction model, obtain the predicted values ​​of the key state parameters of the waste heat boiler, and obtain the time series deviation data between the predicted values ​​and actual values ​​of each key state parameter. The reconstruction detection module 203 is used to input time series deviation data into the anomaly detection model, identify outlier data, and combine the seasonal trend decomposition algorithm to split the time series deviation data into trend component, seasonal component and residual component to clarify the anomaly type. The attribution analysis module 204 is used to convert outlier data into Boolean outlier labels and input them into the random forest model for parameter feature interaction analysis. It combines feature attribution value method to quantify the marginal contribution of each running feature parameter contained in the outlier data to the prediction results, and locates the root cause parameter that causes the anomaly and the direction of the effect of the root cause parameter on the abnormal state based on the marginal contribution. The feedforward optimization module 205 is used to formulate a multi-objective parameter feedforward optimization control strategy based on the located root cause parameters and trend components, and to eliminate early faults by adjusting the controllable execution system parameters of the waste heat boiler.

[0071] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the combined cycle unit waste heat boiler abnormality diagnosis method of the present invention.

[0072] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory can include Random Access Memory (RAM), or non-volatile memory, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0073] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0074] Furthermore, to achieve the above objectives, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the combined cycle unit waste heat boiler anomaly diagnosis method of the present invention.

[0075] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0076] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce a product including instruction means, the instruction means being implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0079] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for diagnosing abnormalities in a waste heat boiler of a combined cycle power unit, characterized in that, The method includes: Historical operating data of the waste heat boiler of the combined cycle unit is obtained, noise reduction and time-series segmentation are performed on the historical operating data, heat parameters are calculated based on the segmented data, and the historical operating data and the heat parameters are fused to obtain the operating dataset. A parameter prediction model is trained using the aforementioned operational dataset. The operational data to be diagnosed is input into the parameter prediction model to obtain the predicted values ​​of the key state parameters of the waste heat boiler. The time-series deviation data between the predicted values ​​and actual values ​​of each key state parameter is also obtained. The time-series deviation data is input into the anomaly detection model to identify outlier data points. The time-series deviation data is then split into trend components, seasonal components, and residual components using a seasonal trend decomposition algorithm to clarify the anomaly type. The outlier data is converted into Boolean outlier labels and input into a random forest model for parameter feature interaction analysis. The marginal contribution of each running feature parameter contained in the outlier data to the prediction result is quantified by combining feature attribution value method. Based on the marginal contribution, the root cause parameter that causes the anomaly and the direction of the effect of the root cause parameter on the abnormal state are located. Based on the identified root cause parameters and trend components, a feedforward optimization and control strategy for multi-objective parameters is formulated, and early failures are eliminated by adjusting the controllable execution system parameters of the waste heat boiler.

2. The method for diagnosing abnormalities in a combined cycle unit waste heat boiler according to claim 1, characterized in that, A parameter prediction model is trained using the aforementioned operational dataset. The operational data to be diagnosed is then input into the parameter prediction model to obtain predicted values ​​for the key state parameters of the waste heat boiler, including: The parameter prediction model includes three parallel recurrent neural network sub-models: a flue gas heat prediction sub-model, a turbine power prediction sub-model, and a steam temperature prediction sub-model. Environmental parameters, compressor inlet guide vane opening, and overall unit load command are extracted from the operational data to be diagnosed as input features of the flue gas heat prediction sub-model. Long-term dependencies are extracted through the control mechanisms of forget gate, input gate, and output gate in the sub-model, and the predicted value of flue gas heat at the inlet of the waste heat boiler is output. The waste heat boiler inlet flue gas parameters and condenser pressure are extracted from the operating data to be diagnosed as input features of the turbine power prediction sub-model, and the turbine dynamic power prediction value is output. The inlet flue gas parameters of the waste heat boiler and the pressure and flow rate of each stage of feedwater are extracted from the operating data to be diagnosed as input features of the steam temperature prediction sub-model, and the predicted value of the outlet steam temperature of the waste heat boiler is output. The predicted values ​​of flue gas heat at the inlet of the waste heat boiler, the predicted dynamic power of the steam turbine, and the predicted steam temperature at the outlet of the waste heat boiler together constitute the key state parameters.

3. The method for diagnosing abnormalities in a combined cycle unit waste heat boiler according to claim 2, characterized in that, The time-series deviation data is input into the anomaly detection model to identify outlier data, including: An unsupervised temporal reconstruction network containing an encoder and a decoder is constructed to convert the temporal deviation data into a deviation sequence matrix to be diagnosed. Using the attention mechanism in the unsupervised temporal reconstruction network, the similarity weight between the current deviation sequence matrix and the historical normal operation sequence key matrix is ​​calculated, and the similarity weight is transformed into the contribution distribution of the historical sequence to the current diagnostic sequence. The time-series deviation data is reconstructed based on the contribution distribution, and the average absolute error between the original deviation data and the reconstructed data is calculated as the reconstruction error loss value. Within a set time window, when the reconstruction error loss value is greater than the set dynamic adaptive threshold, it is determined that there is a deviation in the operating mode at the current time step, and the abnormal point data is identified and output.

4. The method for diagnosing abnormalities in a combined cycle unit waste heat boiler according to claim 3, characterized in that, The time-series deviation data is split into trend components, seasonal components, and residual components using a seasonal trend decomposition algorithm to identify anomaly types, including: The time-series deviation data is additively decomposed using a local weighted regression scatter smoothing method to extract the trend component, seasonal component, and residual component from the time-series deviation data, respectively. Determine whether the outlier data is caused by the continuous deviation of the trend component from the baseline; If the anomaly is determined to be caused by the continuous deviation of the trend component, then the interference of the seasonal component and the residual component is filtered out, and the anomaly type is locked as a performance degradation anomaly.

5. The method for diagnosing abnormalities in a combined cycle unit waste heat boiler according to claim 4, characterized in that, The outlier data is converted into Boolean outlier labels and input into a random forest model for parametric feature interaction analysis, including: The healthy operating condition time step output by the unsupervised temporal reconstruction network is marked as zero, and the time step that is determined to be the abnormal point data is marked as one, thus constructing the Boolean type abnormal label dataset containing binary classification variables. The hyperparameter combination of the random forest model is optimized using a cross-validation algorithm. The information gain of the feature is obtained by calculating the decrease in Gini impurity of the input feature to be evaluated before and after splitting the data at each decision tree node of the random forest model. The information gain is used to evaluate the correlation weight of steam parameters and feedwater parameters with the occurrence of abnormal states, and a multi-parameter feature importance ranking matrix is ​​output.

6. The method for diagnosing abnormalities in a combined cycle unit waste heat boiler according to claim 5, characterized in that, The marginal contribution of each operational feature parameter contained in the outlier data to the prediction result is quantified using the feature attribution value method. Based on the marginal contribution, the root cause parameters leading to the anomaly and the direction of their influence on the anomaly state are located, including: Traverse the subset combinations of feature parameters in the random forest model and calculate the marginal change in the model's anomaly probability prediction output value before and after introducing specific state parameters. The marginal contribution is obtained by calculating the weighted average of the marginal changes corresponding to all possible permutations and combinations of feature subsets. The parameters whose absolute value of marginal contribution ranks among the top in a preset range are extracted as the root source parameters; Calculate the standardized feature quantization value of the root cause parameter, and determine whether the direction of the effect of the root cause parameter on the abnormal state is to exacerbate the abnormality or suppress the abnormality based on the mapping relationship between the standardized feature quantization value and the positive and negative signs of the marginal contribution.

7. The method for diagnosing abnormalities in a combined cycle unit waste heat boiler according to claim 6, characterized in that, Based on the identified root cause parameters and trend components, a multi-objective parameter feedforward optimization and control strategy is formulated, and early faults are eliminated by adjusting the controllable execution system parameters of the waste heat boiler, including: The set of root cause parameters that exacerbate the anomalies are mapped to the corresponding physical actuators of the waste heat boiler. If the root cause parameter is determined to be high-pressure water flow or medium-pressure water pressure, and the value of the corresponding parameter is lower than the health threshold, a first feedforward optimization instruction is generated to increase the speed of the water pump inverter or adjust the opening of the water supply regulating valve. If the root cause parameter is the reheat desuperheating water flow rate parameter, then a second feedforward optimization instruction is generated to execute the desuperheating water valve action based on temperature deviation feedforward compensation. The first feedforward optimization command and the second feedforward optimization command are sent to the distributed control system for execution.

8. The method for diagnosing abnormalities in a combined cycle unit waste heat boiler according to claim 7, characterized in that, After formulating a feedforward optimization control strategy for multi-objective parameters, the method further includes: Collect system feedback update data after executing the feedforward optimization control strategy; Replace the trend component in the system feedback update data with the prediction baseline data of the parameter prediction model in a healthy state, and calculate the expected difference of the performance parameters before and after the replacement. Calculate the average increase in steam temperature and the average increase in steam cycle output power of the waste heat boiler under steady-state conditions based on the expected difference in the performance parameters. By comparing the expected increase with the actual monitoring feedback, the step size and gain parameters of the feedforward optimization control strategy are dynamically fine-tuned.

9. A diagnostic device for abnormalities in a waste heat boiler of a combined cycle power unit, characterized in that, The device includes: The data preprocessing module is used to acquire historical operating data of the waste heat boiler of the combined cycle unit, perform noise reduction and time-series segmentation on the historical operating data, calculate heat parameters based on the segmented data, and fuse the historical operating data with the heat parameters to obtain the operating dataset. The parameter prediction module is used to train and construct a parameter prediction model using the running dataset, input the running data to be diagnosed into the parameter prediction model, obtain the predicted values ​​of the key state parameters of the waste heat boiler, and obtain the time-series deviation data between the predicted values ​​and actual values ​​of each key state parameter. The reconstruction detection module is used to input the time series deviation data into the anomaly detection model, identify outlier data, and combine the seasonal trend decomposition algorithm to split the time series deviation data into trend component, seasonal component and residual component to clarify the anomaly type. The attribution analysis module is used to convert the outlier data into Boolean outlier labels and input them into the random forest model for parameter feature interaction analysis. It combines the feature attribution value method to quantify the marginal contribution of each running feature parameter contained in the outlier data to the prediction result, and locates the root cause parameter that causes the anomaly and the direction of the effect of the root cause parameter on the anomaly state based on the marginal contribution. The feedforward optimization module is used to formulate a multi-objective parameter feedforward optimization control strategy based on the located root cause parameters and the trend component, and to eliminate early faults by adjusting the controllable execution system parameters of the waste heat boiler.

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 abnormal diagnosis of waste heat boiler in combined cycle units as described in any one of claims 1 to 8.

Images