A method for monitoring risks of a thermal power generation pipeline system

Abstract

The present application provides a kind of risk monitoring method of thermal power pipeline system, it is related to pipeline risk monitoring technical field, the present application utilizes multidimensional working condition data to train deep neural network, constructs sediment change model;Subsequently, combined with shutdown inspection data, the task phase is divided, and the model error is determined by comparing the theoretical and actual increase;High error label is identified by analyzing the parameter distribution characteristics of error group;Finally, real-time working condition is matched with label, and the error tendency is determined by using Euclidean distance, and the model output is dynamically corrected, so as to obtain accurate current sediment volume.Through the establishment of error feedback and feature correction mechanism, the defects that the traditional model is difficult to capture complex coupling relationship are effectively made up, and the precision of thermal power pipeline sediment prediction is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of pipeline risk monitoring technology, specifically a risk monitoring method for thermal power generation pipeline systems. Background Technology

[0002] As the core carrier of energy conversion and transmission, the growth of deposits within thermal power generation pipeline systems directly affects the heat transfer efficiency and operational safety of the units. Therefore, accurate monitoring and risk warning of pipeline deposit volume are crucial aspects of power plant operation and maintenance. In actual production, the accumulation of pipeline deposits is influenced by a combination of factors, including steam temperature and pressure, feedwater quality, fuel composition, and load fluctuations, exhibiting highly nonlinear characteristics. Currently, combining sensor-collected thermal, working fluid, and operational parameters with data-driven methods such as deep neural networks to establish deposit change models, and using actual measurement data obtained from periodic shutdowns for statistical analysis of model prediction deviations, identification of high-error features, and logical correction, has become a technical approach for real-time evaluation and refined management of pipeline health status. This has significant engineering application implications for optimizing soot blowing frequency, extending pipeline maintenance cycles, and preventing pipeline failure risks.

[0003] An existing technology, namely the oil pipeline simulation analysis system disclosed in CN119323094A, constructs a three-dimensional visualization model to identify wax precipitation temperature and deposition risk in a timely manner, providing accurate data support for subsequent dynamic monitoring and pipeline cleaning decisions. Through the risk identification module, the system can analyze and calculate the wax precipitation temperature based on real-time acquired oil property data and promptly determine whether there is a deposition risk within the oil pipeline. Upon receiving a dynamic tracking command, the kinetic analysis module can acquire the deposition thickness at various locations within the pipeline in real time and analyze the impact of changes in oil temperature and flow rate on deposition distribution.

[0004] While the published documents have achieved the identification of the impact of certain fixed factors on sediment change factors, there are more influencing factors and more coupling relationships in thermal power pipelines. Therefore, the technical features cannot more accurately determine the sediment change relationship caused by the changes of several factors in the pipeline, resulting in a large error in the judgment results.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a risk monitoring method for thermal power generation pipeline systems to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A risk monitoring method for a thermal power generation pipeline system, comprising the following steps: Step 1: Obtain the operating condition data and corresponding shutdown inspection data for each historical pipeline task. The shutdown inspection data includes the deposition increase on the heated surface between two adjacent shutdown inspections. Using the deposition increase after each task as the sample label, input it into a deep neural network model for training to obtain a deposition change model. The output is the predicted deposition increase. Step 2: Obtain the actual task sequence and input its corresponding operating condition data into the deposition change model to obtain the predicted deposition increase after each task. Divide the actual task sequence into several task stages based on the shutdown inspection interval and obtain the actual stage deposition increase for each task stage. Simultaneously, calculate the predicted stage for each task stage based on the predicted deposition increase output by the deposition change model. Step 3: Obtain the model error for each task stage, and analyze the working data of each task based on the model error value to obtain the high error label that causes the large model error. The high error label is the parameter type in the working data. Step 4: Input the working data of each task in the current task stage into the sediment change model to obtain the predicted sediment increase. At the same time, match the specific data of the parameter type with the high error label to identify the high error results and their error tendency. Combine the model error and error tendency to correct the predicted sediment increase. Based on the corrected predicted sediment increase and the actual stage increase, form the current sediment volume. Output the sediment risk monitoring results based on the current sediment volume.

[0008] Furthermore, historical operating data of the pipeline is obtained. This operating data includes thermal parameters, working fluid parameters, fuel parameters, operating parameters, and current deposition volume. The thermal parameters include steam temperature, pressure, flow rate, and flue gas temperature. The working fluid parameters include feedwater pH, dissolved oxygen content, conductivity, and iron / copper ion concentration. The fuel parameters include coal calorific value and furnace feed rate. The operating parameters include unit load, soot blowing frequency, and pipe diameter. The current deposition volume is the volume of deposits in the pipeline before each task begins, using the operating data of each task as a sample. Based on the difference between the sediment volume at the start time and the sediment volume at the end time of the same task, the sediment increase corresponding to the task is determined, and the sediment increase is used as a label. The operational data of each task and the sediment volume at the start of the task are used as input samples for the model, and the sediment increase is used as a label. A pre-defined deep neural network is trained to obtain a sediment change model; wherein, the sediment change model is used to output the predicted sediment increase corresponding to the input sample.

[0009] Furthermore, the actual task sequence after pipeline replacement is obtained, which includes the operating data of each task after pipeline replacement. The actual task sequence is then input into the sediment change model according to the task, and the predicted sediment increase corresponding to each task is output.

[0010] Furthermore, the operating interval between two adjacent shutdown checks is defined as a task interval, and each task sequence within the task interval is defined as a task stage; The volume of sediment in the pipeline is acquired at the start of the shutdown inspection of the task phase and at the end of the shutdown inspection of the task phase. The difference between the sediment volume at the end of the shutdown inspection and the sediment volume at the beginning of the shutdown inspection is calculated to obtain the actual stage increase corresponding to the task stage.

[0011] Furthermore, the predicted increase in the pipeline within each stage of the task is calculated, based on the following logic: In each phase task, each task generates a deposition increase. The deposition increases of all tasks are summed to obtain the predicted phase increase.

[0012] Furthermore, for each stage task, the difference between the predicted stage increase and the actual stage increase is obtained and labeled as the model error. The absolute values ​​of the model errors are sorted from largest to smallest, and the top 20% of model errors are selected, with at least one selected.

[0013] Furthermore, high-error labels are obtained by analyzing the errors of the selected and unselected models, as follows: The selected model error corresponding to the stage task is divided into a positive error stage task group and a negative error stage task group. For multiple preset working condition parameter types, the working condition data corresponding to the positive error stage task group is mean-processed to obtain a positive mean vector, the working condition data corresponding to the negative error stage task group is mean-processed to obtain a negative mean vector, and the working condition data corresponding to the unselected model error is mean-processed to obtain a baseline mean vector. For each working condition parameter type, a first overlap is calculated based on the positive mean vector and the reference mean vector, and a second overlap is calculated based on the negative mean vector and the reference mean vector. When the first overlap is less than the preset overlap threshold, the operating condition parameter type is determined as the high error label corresponding to the positive error stage task group; when the second overlap is less than the preset overlap threshold, the operating condition parameter type is determined as the high error label corresponding to the negative error stage task group.

[0014] Furthermore, within the currently executing phase of the task, the operational data of each task is input into the sediment change model to obtain the predicted increase in sediment after the end of each task. The working data for each task is compared with the high-error labels to identify high-error results and their error tendencies. The identification logic is as follows: Identifying high-error results and their error tendencies includes: Within the currently ongoing phase of the mission, the operational data of each mission are input into the sediment change model to obtain the predicted sediment increase for each mission. For the working data of each task, for the high error label set of the task group in the positive error stage and the high error label set of the task group in the negative error stage, calculate the overlap between the working data and the benchmark mean vector in terms of the parameter type corresponding to each high error label. When the overlap of the working condition data across all parameter types corresponding to the high error labels is less than the preset overlap threshold, the task is determined to be a high error task; otherwise, the task is determined to be a non-high error task. For tasks identified as high-error tasks, calculate the first distance between the working condition data and the positive mean vector of the task group in the positive error phase, and the second distance between the working condition data and the negative mean vector of the task group in the negative error phase. When the first distance value is less than the second distance value, the error tendency is determined to be positive; when the second distance value is less than or equal to the first distance value, the error tendency is determined to be negative.

[0015] Furthermore, the predicted deposition increase for each task is corrected based on the high error results and error tendency. The correction logic is as follows: When the task's operating data matches the high error result, the mean of the model error corresponding to the error tendency is selected, and the predicted sediment increase output by the sediment change model corresponding to the high error result is corrected by adding the mean. When the task's operating data does not match the high error result, the value of the predicted sediment increase corresponding to the task's operating data is not changed. In each phase of the mission, the current sediment volume is obtained by summing up all corrected predicted sedimentary increases and all actual phase increases, and this is output as the risk monitoring result.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a deep neural network by integrating multi-dimensional data such as thermodynamic parameters, working fluid parameters, fuel parameters, and operational parameters. Utilizing the model's ability to fit complex nonlinear operating conditions, it initially characterizes the sediment variation patterns under the influence of multiple factors. Based on this, it further compares actual data obtained through shutdown checks with model predictions, classifying model errors into positive and negative groups. The overlapping area of ​​histograms is used to identify "high-error labels" that significantly contribute to the error, and Euclidean distance matching technology is combined to determine the error tendency of the current operating condition, thereby dynamically and in real-time correcting the model output. This invention's closed-loop mechanism of reverse feature identification and compensation based on actual error distribution effectively compensates for the shortcomings of traditional methods in accurately capturing the dynamic relationships of multiple factors, significantly reducing the bias in sediment prediction under complex coupled environments, and improving the accuracy and reliability of risk monitoring. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a schematic diagram of the depositional change model structure of the present invention; Figure 3 This is a schematic diagram illustrating the high-error label judgment method of the present invention; Figure 4 This is a schematic diagram illustrating the current sediment volume change in this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] Example: Please see Figure 1 The present invention provides a technical solution: A risk monitoring method for a thermal power generation pipeline system, comprising the following steps: Step 1: Obtain the operating condition data and corresponding shutdown inspection data for each historical task of the pipeline. The shutdown inspection data includes the amount of deposition increase on the heated surface between two adjacent shutdown inspections. The amount of deposition increase in the pipeline after each task is used as the sample label and input into the deep neural network model for training to obtain the deposition change model. The output is the predicted amount of deposition increase. Step 1 includes the following: Historical operating data of the pipeline is acquired. The operating data includes thermal parameters, working fluid parameters, fuel parameters, operating parameters, and current deposition volume. The thermal parameters include steam temperature, pressure, flow rate, and flue gas temperature. The working fluid parameters include feedwater pH, dissolved oxygen content, conductivity, and iron / copper ion concentration. The fuel parameters include coal calorific value and furnace feed rate. The operating parameters include unit load, soot blowing frequency, and pipe diameter. The current deposition volume is the volume of deposits in the pipeline before the start of each task, using the operating data of each task as a sample. Based on the difference between the sediment volume at the start time and the sediment volume at the end time of the same task, the sediment increase corresponding to the task is determined, and the sediment increase is used as a label. The working conditions data of each task and the sediment volume at the start of the task are used as input samples for the model, and the amount of sediment increase is used as a label. The preset deep neural network is trained to obtain a sediment change model. The sediment change model is used to output the predicted amount of sediment increase corresponding to the input sample.

[0021] By incorporating thermodynamic parameters such as steam temperature, pressure, flow rate, and flue gas temperature, working fluid parameters such as feedwater pH, dissolved oxygen content, conductivity, and iron / copper ion concentration, and operational and fuel parameters such as coal calorific value, furnace feed rate, unit load, soot blowing frequency, and pipe diameter into the data collection scope, a comprehensive characterization of the multi-dimensional influencing factors influencing pipeline sediment formation—including thermodynamics, chemistry, fluid dynamics, and operational factors—is achieved. This avoids the problem that a single type of parameter cannot capture the complex mechanisms of sediment formation. The volume of sediment already present in the pipeline before each task begins is included in the current sediment volume sample, providing initial boundary conditions for the sediment accumulation effect of the deep neural network. This allows the model to distinguish the intrinsic correlation between historical sediment states and newly added sediments in the current task. The resulting sample set unifies heterogeneous data from the unit's distributed control system, laboratory analysis, and maintenance logs into structured input, laying a data foundation for the subsequent deep neural network to learn the nonlinear mapping relationship from the high-dimensional operating space to the sediment increment space.

[0022] The actual sediment volume increase detected after each mission is labeled as a sample and used for deep neural network training. This allows the model to directly learn the quantitative correlation between the combination of operating parameters and the sediment growth rate from historical operational data, rather than relying on empirical formulas or simplified mass transfer models for rough estimation. The deep neural network automatically extracts the cross-coupling features between parameters through multi-layer nonlinear transformations, such as the synergistic effect of high-temperature steam parameters and iron ion concentration on the accelerated formation of oxide scale. This elevates sediment increment prediction from single-factor trend analysis to accurate prediction based on multi-factor fusion. The sediment change model obtained after training takes real-time operating data from each mission as input and outputs the corresponding predicted sediment increase. This transforms the changes in the internal state of the pipeline, which originally required shutdown for physical inspection, into values ​​that can be calculated in real time using operating parameters. This provides a continuous and quantifiable predictive benchmark for subsequent mission phase division and error analysis.

[0023] In a preferred embodiment, the number of input layer nodes in the sediment change model strictly corresponds to the dimensions of the acquired operating data, including five major categories of feature variables: thermodynamic parameters, working fluid parameters, fuel parameters, operating parameters, and current sediment volume. Thermodynamic parameters (such as steam temperature and pressure) determine the phase transition state of the working fluid; working fluid parameters (such as pH value and ion concentration) reflect the concentration of corrosion products; fuel parameters and operating parameters (unit load and soot blowing frequency) determine the heat load distribution and physical erosion intensity; and the introduction of "current sediment volume" is to simulate the positive feedback effect of sediment surface roughness on subsequent particle capture. These features are all Z-score normalized before input to eliminate the gradient vanishing problem caused by the order-of-magnitude difference between the feedwater pH value (single digit) and the unit load (hundred megawatts).

[0024] The hidden layers consist of four layers, with the number of neurons in each layer decreasing in a sequence of 512-256-128-64, enabling layer-by-layer abstraction and compression of features. Each fully connected layer is followed by a normalization layer to accelerate convergence and enhance the model's robustness to load fluctuations from different units. The ReLU activation function is chosen to capture the complex nonlinear critical relationship between operating parameters and deposition rate. Dropout technology is also incorporated into the hidden layers, with a dropout rate of 0.2, to prevent the model from overfitting to specific historical downtime data during training, ensuring good generalization and prediction capabilities when facing new operational task sequences.

[0025] The compiled historical task sample set was divided into training and validation sets in an 8:2 ratio. The Adam optimizer was used for weight updates, with an initial learning rate of 0.001, and a learning rate decay strategy was implemented, decreasing the learning rate by 0.1 times every 30 epochs to ensure a smooth slide to the global optimum in the later stages of training. Mean squared error was used as the loss function, forcing the model to prioritize reducing biases that could lead to severe pipeline blockage and prediction errors. The training process was set to 200 epochs with early stopping enabled; training automatically stopped when the validation set loss no longer decreased within 15 consecutive epochs.

[0026] The output layer is a single node, directly mapped to the predicted deposition increase (volume value) after this task. Through training the neural network model, the physicochemical processes inside the pipe, which were originally difficult to observe intuitively, are transformed into digital incremental outputs under a series of discrete tasks.

[0027] Figure 2 This is a schematic diagram of the deposition change model structure of the present invention. In this model, working condition data is used as a sample input to the input layer, processed by the hidden layer, and the deposition increase is output through the output layer.

[0028] Step 2: Obtain the actual task sequence and input the corresponding working condition data into the sediment change model to obtain the predicted sediment increase after each task. Divide the actual task sequence into several task stages through the interval of shutdown inspection and obtain the actual stage increase of each task stage. At the same time, calculate the predicted stage increase of each task stage based on the predicted sediment increase output by the sediment change model. Step 2 includes the following: Obtain the actual task sequence after pipeline replacement, which includes the operating data of each task after pipeline replacement. Input the actual task sequence into the sediment change model in sequence according to the task, and output the predicted sediment increase corresponding to each task.

[0029] The operating interval between two adjacent shutdown checks is defined as a task interval, and each task sequence within the task interval is defined as a task stage; The volume of sediment in the pipeline is acquired at the start of the shutdown inspection of the task phase and at the end of the shutdown inspection of the task phase. The difference between the sediment volume at the end of the shutdown inspection and the sediment volume at the beginning of the shutdown inspection is calculated to obtain the actual stage increase corresponding to the task stage.

[0030] The actual task sequence after pipeline replacement is sequentially input into the sediment change model trained in step 1. Utilizing the model's ability to output a predicted increase in sedimentation for each task, the discrete shutdown inspection results are extended into a continuous stream of predicted values ​​covering the entire task sequence. This allows maintenance personnel to obtain the contribution of each operation to the pipeline sedimentation state without performing shutdown inspections after each task. It solves the monitoring granularity problem that shutdown inspections can only obtain phased cumulative changes and cannot perceive the impact of individual tasks, providing precise temporal data support for subsequent error tracing using tasks as the smallest unit.

[0031] By conducting shutdown inspections at the end of each task phase, the actual increase in the volume of sediment inside the pipeline was obtained. All tasks performed within that inspection interval were calibrated as the corresponding phase tasks, establishing a one-to-one correspondence between phase divisions and physical detection results. The actual increase in sediment volume, as the objective true value of sediment growth within the maintenance window, directly reflects the true cumulative degree of physical changes on the pipeline inner wall. This provides an irreplaceable calibration benchmark for subsequently evaluating the predictive reliability of the model throughout the entire task phase, transforming the calculation of model error from a purely mathematical comparison into a quantitative verification with actual physical measurements.

[0032] Within the defined phased tasks, the predicted sediment increase output by the sediment change model for each task is successively accumulated. The single-task increments predicted by the model are summed and aggregated into a predicted phase increase, realizing the integration of the predicted values ​​over time. This transforms the model output from isolated single-task increments into a phased total change that matches the shutdown and inspection cycle, ensuring complete alignment between the predicted results and the actual maintenance measurements in both spatial scale and time window. The predicted phase increase constructed in this way provides a comparable calculation object for subsequent steps to directly perform difference calculations with the corresponding actual phase increase, laying the foundation for quantitative analysis of model errors.

[0033] Step 3: Obtain the model error for each task stage, and analyze the working data of each task based on the model error value to obtain the high error label that causes the large model error. The high error label is the parameter type in the working data. Step 3 includes the following: For each stage task, the difference between the predicted stage increase and the actual stage increase is obtained and labeled as the model error. The absolute values ​​of the model errors are sorted from largest to smallest, and the top 20% of model errors are selected, with at least one selected.

[0034] By subtracting the actual increase in the predicted stage from the increase in the actual stage for each task stage obtained in step 2, the overall prediction bias of the model throughout the entire stage is quantified into a model error value with positive and negative directions. This difference calculation transforms the vague qualitative judgment of "the model is sometimes too large and sometimes too small" into a clear numerical sequence, making the prediction reliability of each task stage a comparable and ranking metric. By sorting the absolute values ​​of the model errors from largest to smallest and selecting the top 20% of error terms, the focus is on a few abnormal task stages with the largest prediction biases. These selected stages collectively reflect the operating range where the operating conditions deviate significantly from the model training distribution, providing high-value analytical samples for subsequent reverse tracing of the types of operating parameters that caused model failure.

[0035] High-error labels are obtained by analyzing the errors of the selected and unselected models, as follows: The selected model error corresponding to the stage task is divided into a positive error stage task group and a negative error stage task group. For multiple preset working condition parameter types, the working condition data corresponding to the positive error stage task group is mean-processed to obtain a positive mean vector, the working condition data corresponding to the negative error stage task group is mean-processed to obtain a negative mean vector, and the working condition data corresponding to the unselected model error is mean-processed to obtain a baseline mean vector. For each working condition parameter type, a first overlap is calculated based on the positive mean vector and the reference mean vector, and a second overlap is calculated based on the negative mean vector and the reference mean vector. When the first overlap is less than the preset overlap threshold, the operating condition parameter type is determined as the high error label corresponding to the positive error stage task group; when the second overlap is less than the preset overlap threshold, the operating condition parameter type is determined as the high error label corresponding to the negative error stage task group.

[0036] Through a three-stage data analysis chain of "model error absolute value sorting - group mean vector construction - overlap comparison," the black-box prediction bias of sediment change models is transformed into interpretable working fluid parameter labels. First, the top 20% of stages are sorted by the absolute value of model errors. Essentially, this involves extreme value screening to separate "abnormal stages" where the operating conditions significantly deviate from the mainstream distribution of model training. This is because the prediction bias in these stages is far greater than what random fluctuations can explain, reflecting local failures in the model's mapping relationship under multi-parameter coupling. Next, the selected stages are divided into positive error stage task groups and negative error stage task groups according to the error direction, and their operating condition data mean vectors are calculated separately. Simultaneously, a baseline mean vector is constructed using the unselected normal stages. Distribution centers are established in the high-dimensional operating condition space for the two types of abnormal states: "significant overestimation" and "significant underestimation," with the center of the normal operating conditions serving as the control baseline. For each operating condition parameter type, the first overlap between the positive mean vector and the baseline mean vector, and the second overlap between the negative mean vector and the baseline mean vector are calculated. This overlap measures the degree of coincidence between the numerical distribution of the outlier group and the distribution under normal operating conditions for that parameter. The lower the value, the further the typical value of that parameter deviates from normal operating conditions in the outlier group; that is, the parameter is a characteristic driving factor causing model error. When the overlap is below a preset threshold, the parameter is marked as having a high error, effectively excluding parameters that are slightly offset by random disturbances but are essentially still within normal fluctuations, retaining only those key parameters with statistically significant differences. Maintenance personnel can thus understand that when the feedwater iron ion concentration enters a high-value range similar to stage three, the model is prone to producing a large positive error (overestimating the increase in deposition); when the feedwater copper ion concentration enters a high-value range similar to stage four, the model is prone to producing a large negative error (underestimating the increase in deposition). This directly maps the physicochemical mechanism of corrosion product migration and deposition to monitorable parameter boundaries, providing a transparent error attribution path for the hidden pipeline deterioration process. The high-error label set and the mean vectors of positive, negative and baseline types produced in this step provide a refined abnormal condition discrimination template for the subsequent step 401. They together define the judgment rule of "what parameter combination deviates from the normal to what extent, when the model will be unreliable". It is the key data infrastructure connecting offline error root cause analysis to online adaptive discrimination.

[0037] The threshold can be obtained by relying on the operating data of the unselected normal task phase in step 3, and by statistically analyzing the distribution of the parameter dimension of the baseline mean vector: calculate the standard deviation of the parameter under the baseline operating condition, and take the interval formed by the mean plus or minus a certain number of standard deviations (such as 2 times) as the typical envelope of normal fluctuation. Within this interval, calculate the overlap of the baseline sample itself as the reference lower limit, and take the lower limit value or a conservative value slightly lower than it as the preset overlap threshold. The unselected phase represents the set of operating conditions where the model prediction error is within the normal range. Its parameter fluctuation characteristics objectively characterize the inherent variability of the pipeline system under stable operating conditions. Under normal operating conditions, the value of any parameter is clustered near the mean, and its overlap with its own mean is naturally high. However, abnormal operating conditions that deviate significantly from the mean will inevitably lead to a sharp drop in overlap. There is a quantifiable jump between the two. Therefore, by setting a threshold using the distribution characteristics obtained from the self-statistics of the normal stage data, it can not only adapt to the natural dispersion of different pipeline system parameters, but also ensure that the extracted high error labels are statistically significant, providing a stable and interpretable numerical baseline for the subsequent step 401 to determine whether the task conditions fall into the unreliable region of the model.

[0038] The purpose of setting a predefined overlap threshold is to establish a clear decision boundary for the statistic of "overlap," quantifying the degree of overlap between the distribution of operating parameters in abnormal and normal phases into a quantifiable numerical standard, distinguishing between "normal fluctuations in parameter distribution" and "abnormal shifts that cause model errors." When the overlap between the distribution of positive or negative error groups on a certain parameter and the baseline normal distribution is too low, it means that the parameter frequently falls into a value range that is rarely seen under normal operating conditions during abnormal phases. This systematic shift exceeds the range that random fluctuations can explain, and therefore it is reasonable to identify the parameter as a characteristic driving factor that causes the prediction bias of the sediment change model. Conversely, if the overlap is higher than the threshold, even if there is a slight difference between the two distributions, it is more likely to be due to sampling noise or permissible variations during normal operation, and should not be marked as a high error label, thus avoiding incorrect attribution due to the accidental fluctuation of a single parameter.

[0039] Under normal operating conditions, the various operating parameters of a thermal power generating unit, such as feedwater pH, dissolved oxygen concentration, or iron / copper ion concentration, are not fixed values. Instead, they fluctuate within a relatively narrow and stable range determined by physical and chemical equilibrium. Influenced by numerous random and minute disturbances such as unit load, minor fuel variations, and laboratory errors, these parameter fluctuations typically follow or approximately follow a normal distribution. Based on the properties of the normal distribution, the probability that a parameter value falls within a range several times its mean standard deviation is definite and extremely high—meaning that even without any substantial anomalies, the parameter itself exhibits a predictable, normal fluctuation envelope characterized by statistical regularity.

[0040] When a significant deviation occurs in the sediment change model's prediction of a certain task phase, if the cause is indeed a substantial shift in the working fluid parameter, then the distribution of that parameter's value during the abnormal phase will no longer follow the steady-state fluctuation pattern. Instead, it will form a new distribution driven by abnormal physicochemical processes (such as accelerated corrosion or chemical dosing mismatch), with a mean significantly different from the normal state. As the mean of the abnormal phase drifts, the "overlap" between the abnormal distribution and the baseline distribution in probability density will decrease sharply. Therefore, the reduction in overlap is essentially a faithful statistical reflection of the parameter data generation mechanism switching from a "steady-state process controlled by random small perturbations" to a "deviation process driven by specific fault mechanisms."

[0041] In practical applications, the reason why data from the normal error phase of the unselected model can be directly used is that the threshold can be set by calculating its mean plus or minus a certain number of standard deviations. These normal phases represent a set of operating conditions where the prediction model is reliable and the system is in steady-state operation. Their parameter samples accurately depict the steady-state fluctuation range of the system. The standard deviation calculated in this way quantifies the inherent natural variability of the system itself, rather than an artificially defined absolute hard standard. Therefore, setting the threshold at a position several standard deviations (such as 2 or 3 standard deviations) around the mean is essentially defining a control interval in which "random fluctuations are likely to be included." When the mean of the parameter distribution of the abnormal group drifts significantly, causing its overlap with this control interval to be lower than the threshold value obtained through self-learning, we can be highly confident that this shift is statistically significant and caused by a substantial anomaly in engineering, rather than simple random noise. The threshold obtained in this way can adapt to the material, water quality and fuel characteristics of each specific pipeline, avoiding misjudgments that may be caused by applying general fixed limits, and providing a solid and interpretable engineering statistical baseline for the accurate identification of online high-error operating conditions.

[0042] As a preferred embodiment, some data on the analysis model error and its corresponding high error labels are shown in Table 1 below: Table 1 shows some model error data and parameter types. Figure 3 This is a schematic diagram of the high-error label judgment of the present invention, based on Table 1 and... Figure 3 It can be seen that stage three corresponds to the positive value group of the selected model error, stage four corresponds to the negative value group of the selected model error, and stage one and stage two correspond to the unselected model error. The high error label corresponding to the positive value group is the iron ion concentration in the feed water, and the high error label corresponding to the negative value group is the copper ion concentration in the feed water.

[0043] Step 4: Input the working condition data of each task in the current task stage into the sediment change model to obtain the predicted sediment increase. At the same time, match the specific data with high error labels according to the parameter type to identify high error results and their error tendencies. Combine the model error and error tendency to correct the predicted sediment increase. The current sediment volume is formed based on the corrected predicted sediment increase and the actual stage increase.

[0044] Step 4 includes the following: Step 401: Within the currently executing phase task, input the working condition data of each task into the sediment change model to obtain the predicted sediment increase after each task ends. During the current mission phase, operational data from each mission is continuously input into the sediment change model to obtain predicted sediment increase. This ensures that the model's predictions of sediment growth are updated synchronously with the actual operation of the unit, forming a dynamic tracking capability for the real-time evolution of sedimentation conditions on the pipeline inner wall. The invisible, slow accumulation process inside the pipeline is transformed into a digital incremental record updated with each mission completion through model output. This solves the problem of lagging status information only being available during shutdown inspections in traditional methods, providing a fundamental data stream for timely detection of sedimentation risks during operation.

[0045] The operating data for each task is compared with high-error labels to identify high-error results and their error tendencies, including: Within the currently ongoing phase of the mission, the operational data of each mission are input into the sediment change model to obtain the predicted sediment increase for each mission. For the working data of each task, for the high error label set of the task group in the positive error stage and the high error label set of the task group in the negative error stage, calculate the overlap between the working data and the benchmark mean vector in terms of the parameter type corresponding to each high error label. When the overlap of the working condition data across all parameter types corresponding to the high error labels is less than the preset overlap threshold, the task is determined to be a high error task; otherwise, the task is determined to be a non-high error task. For tasks identified as high-error tasks, calculate the first distance between the working condition data and the positive mean vector of the task group in the positive error phase, and the second distance between the working condition data and the negative mean vector of the task group in the negative error phase. When the first distance value is less than the second distance value, the error tendency is determined to be positive; when the second distance value is less than or equal to the first distance value, the error tendency is determined to be negative.

[0046] In real-time operation, the high-error label set and mean vector benchmark solidified in step 3 are used to perform cascaded inference of "overlap discrimination - distance double verification" on the working data of each task, thereby realizing the instant identification and direction determination of high-error working conditions online. First, for the working data of the current task, the overlap between the value of each parameter type corresponding to the high-error label and the benchmark mean vector is checked one by one. This overlap essentially measures whether the task, in the key parameter dimension that has been proven to cause model failure, is outside the distribution range of normal working conditions; only when the overlap of all high-error label parameters is lower than the threshold is the task determined to be a high-error task. The "full trigger" condition requires multiple sensitive parameters to be abnormal at the same time, avoiding false alarms caused by the random fluctuation of a single parameter, significantly improving the specificity of high-error identification, and accurately capturing those compound deviation states that may trigger model collaborative error mapping. Subsequently, for the labeled high-error tasks, the first and second distance values ​​between their operating condition vectors and the positive and negative mean vectors are calculated respectively. These distance metrics (such as Euclidean distance) reflect the similarity between the current operating point and historical positive and negative error clusters in a high-dimensional normalized space: the smaller the distance, the closer the operating condition is to the corresponding abnormal pattern in terms of overall parameter profile. When the first distance value is less than the second distance value, it is judged as a positive error tendency; otherwise, it is a negative error tendency. This comparison based on global vector distance can integrate all operating condition parameters rather than relying solely on a single high-error label, ensuring accurate inference of the deviation direction even when operating condition parameters interact. The system can automatically issue an early warning immediately after each task ends: the current task's operating condition has fallen into a region with low reliability of the sediment change model, and it clearly predicts whether the model output tends to be too high or too low. Operators can then decide whether to increase vigilance regarding sediment accumulation at this stage or arrange chemical adjustment measures in advance to suppress abnormal sedimentation trends, thereby shifting pipeline risk management from passively waiting for shutdown inspections to a dynamic perception level during operation. The high-error task markers and error tendencies generated in step 401 are directly used as the start signal and direction command for selective correction in step 402. Only when marked as high error are compensations made based on the corresponding historical error averages superimposed on the positive or negative error tendencies. This maintains the model's sensitive response to changes in operating conditions under normal conditions, and also uses statistical corrections to offset systematic mapping distortions in specific outlier domains. This ensures that the current sediment volume obtained by the final accumulation can integrate the three information of prediction, statistical correction, and subsequent actual measurement to form a robust and accurate dynamic estimate of the pipeline deposition state.

[0047] Step 402: Correct the predicted deposition increase for each task based on the high error results and error trends. The correction logic is as follows: When the task's operating data matches the high error result, the mean of the model error corresponding to the error tendency is selected, and the predicted sediment increase output by the sediment change model corresponding to the high error result is corrected by adding the mean. When the task's operating data does not match the high error result, the value of the predicted sediment increase corresponding to the task's operating data is not changed. For tasks with high error rates, the mean model error of the corresponding positive or negative value group is selected as the correction value based on its error tendency. This correction value is directly superimposed on the predicted sedimentary increase output by the sediment change model, achieving targeted compensation for the systematic bias of the model. By leveraging the complementary advantages of the multi-parameter cross-mapping capability of the sediment change model and the correction benchmark formed by historical error statistics, the model's sensitivity to changes in operating conditions is preserved while its systematic offset in specific operating condition regions is corrected using long-term accumulated error statistics. This makes the corrected single sedimentary increase value closer to the actual physical sedimentary amount.

[0048] In each phase of the task, the current sediment volume is obtained by summing up all corrected predicted sedimentary increases and all actual phase increases, and then output.

[0049] During the phased task progression, all corrected predicted sediment increase amounts are accumulated with the actual phase increase amounts measured during shutdown inspections to obtain the current total sediment volume, which is then output. This integrates three types of heterogeneous data—model predictions, error statistical corrections, and physical measurement benchmarks—into a unified state estimate. The accumulated corrected predicted sediment increase ensures that the dynamic tracking of the continuous accumulation of sediment on the pipeline wall between two shutdown inspections does not deviate due to local model errors. Meanwhile, the superimposed actual phase increase amounts periodically calibrate and reset the cumulative estimate using physical true values, eliminating the error accumulation effect caused by long-term pure reliance on model recursion. The final output of the current sediment volume provides operators with a reliable sediment state assessment that integrates information from the mechanistic model, data-driven corrections, and measured verifications, upgrading pipeline risk monitoring from periodic post-incident inspections to dynamic quantitative control during operation.

[0050] As a preferred embodiment, partial data on the predicted depositional increase and current sediment volume at a certain task stage are shown in Table 2: Table 2 Comparison of predicted sedimentary increase and current sediment volume Figure 4 The diagram shows the current sediment volume change. As shown in Table 2, the predicted sediment increase for tasks 2, 5, 7, and 9 has been revised. Before the revision, the single sediment volume showed drastic changes, which was obviously inconsistent with the actual application. The revised current sediment volume change is more accurate and in line with the actual pattern.

[0051] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0052] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0053] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0054] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A risk monitoring method for a thermal power generation pipeline system, characterized in that, The specific steps include: Step 1: Obtain the operating condition data and corresponding shutdown inspection data for each historical pipeline task. The shutdown inspection data includes the deposition increase on the heated surface between two adjacent shutdown inspections. Using the deposition increase after each task as the sample label, input it into a deep neural network model for training to obtain a deposition change model. The output is the predicted deposition increase. Step 2: Obtain the actual task sequence and input its corresponding operating condition data into the deposition change model to obtain the predicted deposition increase after each task. Divide the actual task sequence into several task stages based on the shutdown inspection interval and obtain the actual stage deposition increase for each task stage. Simultaneously, calculate the predicted stage for each task stage based on the predicted deposition increase output by the deposition change model. Step 3: Obtain the model error for each task stage, and analyze the working data of each task based on the model error value to obtain the high error label that causes the large model error. The high error label is the parameter type in the working data. Step 4: Input the working data of each task in the current task stage into the sediment change model to obtain the predicted sediment increase. At the same time, match the specific data of the parameter type with the high error label to identify the high error results and their error tendency. Combine the model error and error tendency to correct the predicted sediment increase. Based on the corrected predicted sediment increase and the actual stage increase, form the current sediment volume. Output the sediment risk monitoring results based on the current sediment volume.

2. The risk monitoring method for a thermal power generation pipeline system according to claim 1, characterized in that: Historical operating data of the pipeline is acquired. The operating data includes thermal parameters, working fluid parameters, fuel parameters, operating parameters, and current deposition volume. The thermal parameters include steam temperature, pressure, flow rate, and flue gas temperature. The working fluid parameters include feedwater pH, dissolved oxygen content, conductivity, and iron / copper ion concentration. The fuel parameters include coal calorific value and furnace feed rate. The operating parameters include unit load, soot blowing frequency, and pipe diameter. The current deposition volume is the volume of deposits in the pipeline before the start of each task, using the operating data of each task as a sample. Based on the difference between the sediment volume at the start time and the sediment volume at the end time of the same task, the sediment increase corresponding to the task is determined, and the sediment increase is used as a label. The working conditions data of each task and the sediment volume at the start of the task are used as input samples for the model, and the sediment increase is used as a label. The preset deep neural network is trained to obtain a sediment change model. The sediment change model is used to output the sediment increase corresponding to the input sample.

3. The risk monitoring method for a thermal power generation pipeline system according to claim 2, characterized in that: Obtain the actual task sequence after pipeline replacement, which includes the operating data of each task after pipeline replacement. Input the actual task sequence into the sediment change model in sequence according to the task, and output the predicted sediment increase corresponding to each task.

4. The risk monitoring method for a thermal power generation pipeline system according to claim 3, characterized in that: The operating interval between two adjacent shutdown checks is defined as a task interval, and each task sequence within the task interval is defined as a task stage; The volume of sediment in the pipeline is acquired at the start of the shutdown inspection of the task phase and at the end of the shutdown inspection of the task phase. The difference between the sediment volume at the end of the shutdown inspection and the sediment volume at the beginning of the shutdown inspection is calculated to obtain the actual stage increase corresponding to the task stage.

5. The risk monitoring method for a thermal power generation pipeline system according to claim 4, characterized in that: The calculation of the predicted increase in the pipeline within each stage of the task is as follows: In each phase task, each task generates a deposition increase. The deposition increases of all tasks are summed to obtain the predicted phase increase.

6. The risk monitoring method for a thermal power generation pipeline system according to claim 5, characterized in that: For each stage task, the difference between the predicted stage increase and the actual stage increase is obtained and labeled as the model error. The absolute values ​​of the model errors are sorted from largest to smallest, and the top 20% of model errors are selected, with at least one selected.

7. The risk monitoring method for a thermal power generation pipeline system according to claim 6, characterized in that: High-error labels are obtained by analyzing the errors of the selected and unselected models, as follows: The selected model error corresponding to the stage task is divided into a positive error stage task group and a negative error stage task group. For multiple preset working condition parameter types, the working condition data corresponding to the positive error stage task group is mean-processed to obtain a positive mean vector, the working condition data corresponding to the negative error stage task group is mean-processed to obtain a negative mean vector, and the working condition data corresponding to the unselected model error is mean-processed to obtain a baseline mean vector. For each working condition parameter type, a first overlap is calculated based on the positive mean vector and the reference mean vector, and a second overlap is calculated based on the negative mean vector and the reference mean vector. When the first overlap is less than the preset overlap threshold, the working condition parameter type is determined as the high error label corresponding to the positive error stage task group; When the second overlap is less than the preset overlap threshold, the working condition parameter type is determined as the high error label corresponding to the negative error stage task group.

8. The risk monitoring method for a thermal power generation pipeline system according to claim 7, characterized in that: Within the currently executing phase of the task, the operational data of each task is input into the sediment change model to obtain the amount of sediment increase after each task ends. The operational data for each task are compared with high-error labels to identify high-error results and their error tendencies, including: Within the currently ongoing phase of the mission, the operational data of each mission are input into the sediment change model to obtain the predicted sediment increase for each mission. For the working data of each task, for the high error label set of the task group in the positive error stage and the high error label set of the task group in the negative error stage, calculate the overlap between the working data and the benchmark mean vector in terms of the parameter type corresponding to each high error label. When the overlap of the working condition data across all parameter types corresponding to the high error labels is less than the preset overlap threshold, the task is determined to be a high error task; otherwise, the task is determined to be a non-high error task. For tasks identified as high-error tasks, calculate the first distance between the working condition data and the positive mean vector of the task group in the positive error phase, and the second distance between the working condition data and the negative mean vector of the task group in the negative error phase. When the first distance value is less than the second distance value, the error tendency is determined to be positive; when the second distance value is less than or equal to the first distance value, the error tendency is determined to be negative.

9. A risk monitoring method for a thermal power generation pipeline system according to claim 8, characterized in that: The predicted deposition increase for each task is corrected based on the high error results and error tendency. The correction logic is as follows: When the task's operating data matches the high error result, the mean of the model error corresponding to the error tendency is selected, and the predicted sediment increase output by the sediment change model corresponding to the high error result is corrected by adding the mean. When the task's operating data does not match the high error result, the value of the predicted sediment increase corresponding to the task's operating data is not changed. In each phase of the mission, the current sediment volume is obtained by summing up all corrected predicted sedimentary increases and all actual phase increases, and this is output as the risk monitoring result.

Images