Method for optimizing intake filter system replacement period based on machine learning

Abstract

This application discloses a machine learning-based optimization method for the replacement cycle of an air intake filter system, relating to the field of machine learning technology. The method includes: acquiring operational data of the air intake filter system and inputting it into a pre-trained efficiency loss prediction model, outputting a predicted efficiency loss curve; determining a set of candidate replacement cycles based on the morphological characteristics of the predicted efficiency loss curve; for each candidate replacement cycle, calculating the power generation reduction loss and fuel consumption increase loss from the current moment to the end of the candidate replacement cycle, representing the cumulative indirect cost, and obtaining the direct cost; calculating the total cost corresponding to each candidate replacement cycle, and selecting the candidate replacement cycle with the minimum total cost as the optimal replacement cycle. This solves the technical problem that existing air intake filter systems struggle to dynamically adjust replacement cycles according to actual operating conditions and rely excessively on manual experience, leading to excessively high maintenance costs or decreased filtration efficiency.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, specifically to a machine learning-based method for optimizing the replacement cycle of an air intake filter system. Background Technology

[0002] With the widespread application of gas turbines in power generation, the intake filtration system, as a key component ensuring the safe and stable operation of gas turbines, directly affects the power generation efficiency and operating economy of the unit. The main function of the intake filtration system is to filter particulate matter, dust, and other impurities from the air, preventing them from entering the compressor and turbine components and causing blade wear, corrosion, and scale buildup. Over time, the filter screen gradually becomes clogged, leading to increased pressure differential and intake resistance. This, in turn, causes a decrease in the gas turbine's intake airflow and compression ratio, ultimately resulting in reduced output power and deteriorated thermal efficiency.

[0003] However, traditional air intake filter replacement strategies mainly rely on two methods: fixed-cycle replacement or differential pressure threshold alarm. Fixed-cycle replacement is usually based on empirical statistics or manufacturer recommendations, setting a fixed replacement interval. This method fails to fully consider the differences in actual operating environments. In areas with good air quality, it may cause premature filter replacement, increasing unnecessary maintenance costs. In high dust and high humidity environments, it may lead to delayed replacement, exacerbating unit efficiency loss and equipment wear.

[0004] The differential pressure threshold alarm method monitors the pressure difference before and after the filter. When the pressure difference exceeds a preset threshold, it triggers a replacement reminder. Although it can reflect the real-time degree of filter blockage, the threshold setting often lacks economic considerations and is difficult to balance the relationship between filter replacement costs and efficiency loss costs. Summary of the Invention

[0005] This application provides an optimization method for the replacement cycle of an air intake filter system based on machine learning, which solves the technical problem that the replacement cycle of existing air intake filter systems is difficult to dynamically adjust according to actual operating conditions and relies too much on human experience, resulting in excessive maintenance costs or decreased filtration efficiency.

[0006] The technical solution to the above-mentioned technical problems in this application is as follows: This application provides a machine learning-based method for optimizing the replacement cycle of an air intake filter system, the method comprising: The current filter differential pressure, rate of increase of differential pressure, acceleration of increase of differential pressure, filter degradation factor, and environmental parameters of the intake filtration system are obtained to form operating data; The operational data is input into a pre-trained efficiency loss prediction model, and the predicted efficiency loss curve of the intake filtration system efficiency loss over time is output within a future preset time window. Based on the morphological characteristics of the predicted efficiency loss curve, a set of candidate replacement cycles is dynamically determined; For each candidate replacement cycle, calculate the power generation reduction loss and fuel consumption increase loss from the current time to the end of the candidate replacement cycle to obtain the cumulative indirect cost, and obtain the filter purchase cost, replacement operation downtime loss, equipment life loss cost and replacement operation labor cost to obtain the direct cost; Based on the cumulative indirect costs and the direct costs, the total cost corresponding to each candidate replacement cycle is calculated, and the candidate replacement cycle with the minimum total cost is selected as the optimal replacement cycle.

[0007] This application provides one or more technical solutions, which have at least the following technical effects or advantages: This application provides a machine learning-based optimization method for the replacement cycle of an air intake filter system. First, it acquires operational data of the air intake filter system to construct a multi-dimensional state perception system. Second, it inputs the operational data into a pre-trained efficiency loss prediction model, utilizing the powerful modeling capabilities of long short-term memory networks for time-series data to output a predicted efficiency loss curve within a preset future time window, achieving a shift from static state assessment to dynamic trend prediction. Third, based on the morphological characteristics of the predicted efficiency loss curve, it dynamically determines a set of candidate replacement cycles through second-derivative inflection point identification and curvature adaptive sampling strategies, ensuring both comprehensiveness of the search and improved computational efficiency. Then, for each candidate replacement cycle, it comprehensively considers cumulative indirect costs such as reduced power generation and increased fuel consumption, as well as direct costs such as filter procurement costs, replacement downtime losses, equipment lifespan depreciation costs, and replacement operation labor costs, establishing a full life-cycle cost accounting model to quantify equipment lifespan depreciation costs. Finally, based on the principle of minimizing total cost, it selects the optimal replacement cycle, achieving a balance between economic benefits and equipment reliability.

[0008] Through the above technical solutions, the technical solutions of this application embodiment can adaptively adjust the replacement cycle according to the actual operating status and environmental conditions of the air intake filtration system, avoiding resource waste caused by premature replacement and performance degradation caused by late replacement, and improving the operation and maintenance management level of the gas turbine air intake filtration system. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1This is a flowchart illustrating the method for optimizing the replacement cycle of an air intake filter system based on machine learning, as provided in an embodiment of this application. Figure 2 This is a schematic diagram of the process for obtaining equipment life loss cost in the machine learning-based optimization method for the replacement cycle of an air intake filter system provided in this application embodiment. Detailed Implementation

[0011] This application provides a machine learning-based method for optimizing the replacement cycle of an air intake filter system. This method addresses the technical problem that existing air intake filter systems often fail to dynamically adjust their replacement cycles based on actual operating conditions, rely excessively on human experience, and thus result in high maintenance costs or decreased filtration efficiency.

[0012] like Figure 1 As shown in the embodiments of this application, a method for optimizing the replacement cycle of an air intake filter system based on machine learning is provided, including: S10: Acquire the current filter differential pressure, differential pressure rise rate, filter differential pressure rise acceleration, filter degradation factor, and environmental parameters of the intake filtration system to form operating data; In this embodiment, the specific method of collecting operating data is achieved through a sensor network deployed at each key node of the intake filtration system. The differential pressure sensor collects the pressure values ​​on both sides of the filter at a fixed frequency, and obtains the current filter differential pressure through differential calculation. The rate of increase of differential pressure reflects the development speed of filter blockage, and the acceleration of the increase of filter differential pressure is used to characterize the severity of the blockage trend.

[0013] The filter degradation factor comprehensively considers the filter's cumulative operating time, historical cleaning frequency, and material aging degree. It is obtained through joint evaluation of equipment maintenance records and material property database. Environmental parameters are acquired in real time through meteorological monitoring interface and flow meter, which together constitute a multi-dimensional operating data vector.

[0014] Specifically, step S10 in the method includes: The current filter pressure difference is collected in real time by differential pressure sensors installed before and after the filter screen of the intake air filtration system; Perform a sliding window linear regression on the current filter pressure difference to calculate the rate of increase of the pressure difference; Differential calculation of the pressure difference rise rate yields the filter pressure difference rise acceleration; Obtain the initial pressure difference of the filter screen in its initial clean state, calculate the ratio of the current filter screen pressure difference to the initial pressure difference, and obtain the filter screen degradation factor; Ambient temperature is collected in real time by a temperature sensor, ambient humidity is collected in real time by a humidity sensor, and particulate matter concentration data is collected in real time by a particulate matter concentration sensor. The ambient temperature, ambient humidity, and particulate matter concentration data are used as environmental parameters. The current filter pressure difference, the rate of increase of the pressure difference, the acceleration of the increase of the filter pressure difference, the filter degradation factor, and the environmental parameters are combined to form the operating data.

[0015] In this embodiment, firstly, differential pressure sensors are installed before and after the filter screen of the intake filtration system to collect the pressure values ​​on both sides of the filter screen in real time. The current filter screen differential pressure data is obtained through differential calculation. This differential pressure data serves as an indicator of the degree of filter screen clogging, and its value gradually increases as dust accumulates on the filter screen surface. To ensure the continuity and reliability of the data, the differential pressure sensor adopts a high-frequency sampling mode, with a sampling interval that can be set to 1 to 5 minutes, and the raw data is temporarily stored in the buffer area of ​​the edge computing node.

[0016] Secondly, regarding the rate of increase in differential pressure, this embodiment employs a sliding window linear regression method. Specifically, the time span of the sliding window is set to 2 to 4 hours, and the window contains all differential pressure data collected within that time period. The differential pressure sequence within the window is fitted with least squares linearly, and the slope of the fitted line is the rate of increase in differential pressure at that moment. This rate index can effectively filter out instantaneous fluctuation interference and smoothly present the development trend of filter blockage. As the window slides successively, a continuous time series curve of the rate of increase in differential pressure can be formed.

[0017] Furthermore, numerical differentiation is performed on the rate of increase of pressure difference to obtain the acceleration of the filter pressure difference increase. The dimension of this acceleration is the second derivative of the rate of change of pressure difference, which is used to capture the abrupt characteristics of the clogging trend. When the acceleration value is positive and continues to increase, it indicates that the filter clogging has entered the acceleration stage; when the acceleration tends to zero or negative, it suggests that the clogging process may tend to slow down or be disturbed by external factors.

[0018] Next, retrieve the baseline pressure difference value of the filter under the initial clean state recorded in the equipment management system. This value is usually calibrated after the filter is newly installed or deeply cleaned. Calculate the ratio between the current filter pressure difference measured in real time and this initial pressure difference to obtain a dimensionless filter degradation factor. This factor comprehensively reflects the physical clogging degree and structural performance degradation of the filter. The closer the value is to 1, the closer the filter state is to the initial clean level. A continuously increasing value indicates the gradual deterioration of the filter's service performance.

[0019] Furthermore, environmental parameters are collected through multi-source sensor fusion. A temperature sensor is deployed on the outside of the air inlet to monitor changes in ambient temperature in real time. Temperature fluctuations directly affect air density and viscosity, thereby altering the relationship between pressure difference and flow rate. A humidity sensor simultaneously collects the relative humidity of the environment. High humidity environments easily cause filter media fibers to expand and dust to absorb moisture and clump together, accelerating the clogging process. A particulate matter concentration sensor uses the laser scattering principle to measure the concentration levels of key particle sizes such as PM10 and PM2.5 in the intake air in real time. This parameter determines the dust load rate of the filter. These three types of environmental parameters, along with the real-time readings from the flow meter, are incorporated into the operating data vector to form a complete characterization of the filter's operating conditions.

[0020] Finally, the current filter pressure difference, pressure difference rise rate, filter pressure difference rise acceleration, filter degradation factor, and environmental parameters are structurally combined according to a preset data format to construct a multi-dimensional operating data vector. This vector serves as the input feature for the subsequent machine learning model, and its dimensional design and sampling synchronization mechanism ensure the alignment and integrity of the time-series data.

[0021] S20: Input the running data into the pre-trained efficiency loss prediction model and output the predicted efficiency loss curve of the intake filtration system efficiency loss over time within a future preset time window. In this embodiment, the efficiency loss prediction model adopts a deep temporal neural network architecture, specifically a Transformer variant with an encoder-decoder structure. This architecture effectively captures long-range dependencies and multi-scale temporal patterns in the operating data. The model input receives the constructed multi-dimensional operating data vector sequence. The encoder part consists of multiple layers of self-attention mechanisms stacked together, with each layer containing a multi-head attention module and a feedforward neural network. The decoder part adopts an autoregressive generation mechanism with causal masks, using the encoder output as a condition to progressively predict the efficiency loss sequence within a future preset time window. Based on the efficiency loss prediction model, a predicted efficiency loss curve of the intake filtration system efficiency loss changing with time within the future preset time window is obtained.

[0022] Specifically, step S20 in the method includes: Based on the statistical data of the filter's historical replacement cycle, the historical average replacement cycle is calculated, and the length of the historical average replacement cycle is used as the length of the future preset time window. The running data is input into a pre-trained efficiency loss prediction model. The efficiency loss prediction model takes the current time as the starting point and predicts the efficiency loss value at each time point in the future preset time window step by step with a preset time step, thereby generating a predicted efficiency loss sequence. By associating the time points in the predicted efficiency loss sequence with the corresponding efficiency loss values, a predicted efficiency loss curve is formed with time as the horizontal axis and efficiency loss value as the vertical axis.

[0023] In this embodiment, firstly, the filter replacement records of the past three to five years stored in the equipment maintenance management system are retrieved. After removing replacement events caused by abnormal factors such as abnormal failures and planned maintenance, the effective replacement cycle is calculated by arithmetic mean. If the sample size of historical data is insufficient or the operating conditions of the equipment have changed significantly, the industry benchmark value of the same type of equipment under similar operating conditions is used as a supplement. The length of this window is usually set to 30 to 90 days and can be dynamically adjusted according to actual application needs.

[0024] Secondly, the completed running data vector sequence is input into the pre-trained efficiency loss prediction model. The model takes the current time as the prediction starting point and performs time-by-time prediction according to the preset time step. The setting of the time step needs to take into account both prediction accuracy and computational efficiency. In this embodiment, 6 hours or 12 hours are used as the standard step size to generate a discrete future efficiency loss prediction value sequence within the window period.

[0025] During the prediction process, the model encoder first extracts deep features from the historical running data, and the decoder generates the efficiency loss prediction values ​​for each future time step by step based on the state vector output by the encoder and combined with the autoregressive mechanism. The output of each prediction step not only depends on the global information of the encoder, but also integrates the hidden state transmission of the previous prediction step to ensure the temporal coherence of the prediction sequence.

[0026] Furthermore, the generated predicted efficiency loss sequence is visualized and mapped. A two-dimensional coordinate system is established with the prediction time axis as the horizontal axis and the percentage of efficiency loss as the vertical axis. The discrete prediction points are connected in chronological order to form a smooth predicted efficiency loss curve. This curve intuitively presents the performance degradation trajectory of the air intake filtration system within the future window period. The slope of the curve reflects the rate of development of efficiency loss, while the curvature change indicates the stage transition of the degradation mode.

[0027] Specifically, the process of constructing an efficiency loss prediction model includes: Multiple sets of historical data samples were collected from the historical operation database. Each set of historical data samples included historical filter pressure difference, historical pressure difference rise rate, historical filter pressure difference rise acceleration, historical filter degradation factor, historical ambient temperature, historical ambient humidity, and historical particulate matter concentration, which constituted the training sample input set. Obtain the historical air intake filtration system output power corresponding to each group of historical data samples in the training sample input set. Calculate the historical efficiency loss value based on the difference between the historical air intake filtration system output power and the theoretical output power of the air intake filtration system, and construct the training sample output set. An initial efficiency loss prediction model is constructed using a long short-term memory network; Using the training sample input set as input features and the training sample output set as supervision labels, the initial efficiency loss prediction model is trained in a supervised manner until the verification convergence is obtained, thus obtaining the trained efficiency loss prediction model.

[0028] In this embodiment, firstly, multiple sets of historical data samples are systematically collected from the historical operation database to construct a complete training dataset. The time span of the data samples covers at least two complete filter replacement cycles to ensure that the model can learn the evolution law of the filter from its initial clean state to complete failure. Each set of historical data samples contains feature dimensions that are isomorphic to the real-time operation data, namely historical filter pressure difference, historical pressure difference rise rate, historical filter pressure difference rise acceleration, historical filter degradation factor, as well as historical ambient temperature, historical ambient humidity and historical particulate matter concentration, which together constitute the training sample input set.

[0029] Secondly, the training sample output set is constructed based on the actual operating power data of the historical intake filtration system. Historical output power records corresponding to the input feature timestamps are extracted from the unit control system logs. Simultaneously, the theoretical output power design value under this operating condition is retrieved. The theoretical output power is determined by the technical specifications provided by the equipment manufacturer or obtained through performance testing under ideal operating conditions. The difference between the historical output power and the theoretical output power is divided by the theoretical output power to obtain the normalized historical efficiency loss percentage, which serves as the supervision label for the training samples. This efficiency loss index comprehensively reflects the increased flow resistance, decreased ventilation volume, and resulting system energy efficiency degradation caused by filter clogging.

[0030] Furthermore, the architecture of the initial efficiency loss prediction model adopts a long short-term memory network. This network structure effectively alleviates the gradient vanishing problem of traditional recurrent neural networks through a gating mechanism, and is suitable for capturing long-term dependencies and multi-timescale dynamics in the filter degradation process.

[0031] Specifically, the network input layer dimension matches the feature dimension of the running data vector. The hidden layer consists of two stacked LSTM layers, each containing 128 memory units to balance the model's expressive power and generalization performance. The output layer is a single neuron that uses a linear activation function to directly output the efficiency loss prediction value. The Adam optimization algorithm is used, with an initial learning rate set to 0.001 and a learning rate decay strategy introduced. When the validation set loss does not improve for 10 consecutive training epochs, the learning rate is automatically reduced. The batch size is set to 32, and the sequence length is consistent with the historical observation window used in real-time prediction.

[0032] Furthermore, during supervised training, the training sample input set is divided into a training subset and a validation subset in chronological order, for example, with a ratio of 8:2. The time span of the validation subset is ensured to be later than that of the training subset to simulate real time-series prediction scenarios. The loss function uses mean squared error to measure the deviation between the prediction efficiency loss and the real historical efficiency loss. At the same time, an L2 regularization term is introduced to constrain the network weights to prevent overfitting. Training iterations continue until the validation set loss converges to a stable plateau or reaches the preset maximum number of training rounds, such as 100.

[0033] The converged model is then evaluated on an independent test set. Evaluation metrics include root mean square error, mean absolute percentage error, and the dynamic time warping distance between the predicted curve and the true curve. Only models that pass the tests can be deployed as pre-trained efficiency loss prediction models for real-time prediction inference.

[0034] S30: Based on the morphological characteristics of the predicted efficiency loss curve, dynamically determine a set of candidate replacement cycles; In this embodiment, multi-scale morphological analysis is performed on the predicted efficiency loss curve to extract key feature points characterizing the filter degradation stage transition, which serve as the basis for generating candidate replacement cycles. The morphological feature analysis includes various types such as curve slope abrupt change points, curvature extreme points, efficiency loss threshold crossing points, and the intersection points of the curve's stable and accelerating segments. Each feature point corresponds to different physical meanings and decision-making tendencies.

[0035] Specifically, step S30 in the method includes: Calculate the second derivative of the predicted efficiency loss curve at each time point; Identify the time point when the second derivative changes from a negative value to a positive value, take the time point as the inflection point, and take the time corresponding to the inflection point as the benchmark replacement cycle; With the benchmark replacement cycle as the center, a preset time range is extended to both sides to form a search interval for candidate replacement cycles; Calculate the curvature value of the predicted efficiency loss curve at each time point within the search interval; Calculate the average curvature value of all time points within the search interval, and use the average value as the curvature threshold. Based on the curvature threshold, the search interval is divided into an encrypted sampling region and a sparse sampling region; Within the encrypted sampling area, candidate replacement cycles are generated with a first sampling step size; within the sparse sampling area, candidate replacement cycles are generated with a second sampling step size. All candidate replacement cycles are merged to generate a set of candidate replacement cycles.

[0036] In this embodiment, firstly, the second derivative of the predicted efficiency loss curve at each time point is calculated to capture the sensitive position of the curve's concavity / convexity change. The second derivative is calculated using the central difference method. For the discrete prediction point sequence, forward or backward difference is used for extension at the boundary. The time point when the second derivative turns from negative to positive marks the curve's change from convex to concave, which corresponds to the critical point when the efficiency loss changes from slow growth to accelerated growth. This inflection point physically corresponds to the starting moment when the filter blockage enters the rapid deterioration stage.

[0037] Secondly, after identifying the inflection point, the corresponding time is set as the baseline replacement cycle. This baseline point represents the turning point when the filter performance degradation changes from slow to rapid. If the filter is replaced at this point, the performance decline trend can be stopped in time before the efficiency loss increases significantly. With this baseline replacement cycle as the center, a preset time range is extended to both sides of the time axis to form a search interval for candidate replacement cycles. The preset extension range is dynamically determined according to the filter type, operating conditions and historical maintenance data. It is used to cover the complete transition stage of filter performance from slight degradation to severe failure, ensuring that the search interval includes the preventive replacement window before the inflection point. For example, in this embodiment, a standard extension width of ±7 to 14 days is used to make the search interval cover a decision window of about 15 to 30 days.

[0038] Secondly, within the defined search interval, the curvature value of the predicted efficiency loss curve at each time point is calculated. Curvature, as a differential geometric quantity describing the degree of local bending of the curve, can effectively distinguish different patterns of efficiency loss change. The curvature calculation formula is based on a combination of the first and second derivatives. A larger curvature value indicates a more drastic change in efficiency loss near that time point, while a smaller curvature value indicates a more gradual change. By traversing the curvature distribution of all time points within the search interval, its arithmetic mean is calculated as the curvature threshold. This threshold binarizes the change characteristics within the interval.

[0039] Furthermore, based on the curvature threshold, the search interval is divided into a dense sampling region and a sparse sampling region. The efficiency loss in the sparse sampling region is relatively stable, allowing for a coarser sampling granularity to reduce subsequent computational burden. A first sampling step size is used in the dense sampling region, and a second sampling step size is used in the sparse sampling region. This organic combination of the two sampling strategies effectively controls the generation scale of candidate periods while ensuring decision accuracy.

[0040] Finally, all candidate replacement cycles generated in the encrypted sampling region and the sparse sampling region are merged in chronological order, duplicate time points are removed and sorted to form a set of candidate replacement cycles that fully covers the search interval. This set includes both dense candidate points near the inflection point to support fine-grained decision-making and sparse candidate points at the interval boundary to preserve the diversity of solutions.

[0041] The search interval is divided into an encrypted sampling region and a sparse sampling region, including: The search interval is divided into multiple consecutive sub-intervals, and the average curvature value of all time points in each sub-interval is obtained by filtering. When the average curvature value within the sub-interval is greater than the curvature threshold, the sub-interval is divided into an encrypted sampling region; When the average curvature value within the sub-interval is less than or equal to the curvature threshold, the sub-interval is divided into a sparse sampling region.

[0042] In this embodiment, the search interval is first divided into multiple consecutive sub-intervals, with the length of each sub-interval set to 3 to 5 days. This ensures that each sub-interval contains enough data points to calculate a stable average curvature, while avoiding excessively long sub-intervals that could mask local features. For each sub-interval, the curvature values ​​corresponding to all time points within it are extracted, and the arithmetic mean is calculated as the representative curvature value for that sub-interval. This average value effectively suppresses fluctuations at individual discrete points and reflects the overall drastic changes within the sub-interval.

[0043] Secondly, the average curvature of each sub-interval is compared with the global curvature threshold to complete the region classification. When the average curvature of a sub-interval is greater than the curvature threshold, it indicates that the efficiency loss curve is significantly curved during that period, and the filter performance is in a state of rapid evolution. It is classified into the dense sampling region. When the average curvature of a sub-interval is less than or equal to the curvature threshold, it indicates that the change is relatively gentle during that period, and the performance degradation trend is stable. It is classified into the sparse sampling region. For sub-intervals where the average curvature at the boundary is exactly equal to the threshold, they are uniformly classified into the sparse sampling region to avoid computational redundancy caused by excessive density.

[0044] Secondly, after the region is divided, adjacent sub-intervals of the same type are merged to form continuous encrypted sampling regions and sparse sampling regions. The boundaries of the merged regions may not be completely aligned with the boundaries of the original sub-intervals. At this time, fine-tuning is performed based on the boundaries of the original sub-intervals to ensure the integrity and continuity of the region division.

[0045] Further, within the encrypted sampling region, candidate replacement periods are generated with a first sampling step size, and within the sparse sampling region, candidate replacement periods are generated with a second sampling step size, including: For each time point within the encrypted sampling region, the ratio of the curvature value at that time point to the curvature threshold is calculated to obtain the first density coefficient; The first sampling step size is obtained by dividing the preset base step size by the first density coefficient; For each time point within the sparse sampling region, the ratio of the curvature value at that time point to the curvature threshold is calculated to obtain the second density coefficient; The second sampling step size is obtained by multiplying the preset base step size by the second density coefficient; Candidate replacement cycles are generated by sampling at equal intervals within the encrypted sampling region using the first sampling step size, and by sampling at equal intervals within the sparse sampling region using the second sampling step size.

[0046] In this embodiment, firstly, for each time point within the encrypted sampling area, the ratio of its curvature value to the curvature threshold is calculated. This ratio is defined as the first density coefficient, which reflects the deviation factor of the time point from the global average change. A larger density coefficient indicates a more significant curvature of the efficiency loss curve at that location, and a more critical evolution of the filter performance, thus requiring a higher sampling density to capture subtle changes. The preset base step size is typically set to 2 to 3 days as a reference benchmark for the standard sampling interval. Dividing the base step size by the first density coefficient yields the adaptively adjusted first sampling step size, which is inversely proportional to the curvature intensity; a larger curvature results in a shorter step size and denser sampling.

[0047] For example, when the curvature value at a certain time point is twice the threshold, the first density coefficient is 2. If the base step size is 2 days, the actual first sampling step size is shortened to 1 day. When the curvature value reaches 4 times the threshold, the first sampling step size is further reduced to 0.5 days, realizing ultra-fine sampling of key areas.

[0048] Secondly, for each time point within the sparse sampling region, the ratio of its curvature value to the curvature threshold is calculated. This ratio is defined as the second density coefficient. Since the curvature values ​​within the sparse sampling region are generally lower than or equal to the threshold, the second density coefficient usually falls within the range of 0 to 1. The smaller the value, the smoother the change. The second sampling step size is obtained by multiplying the preset base step size by the second density coefficient. This step size is directly proportional to the curvature intensity; the smaller the curvature, the longer the step size, and the sparser the sampling.

[0049] For example, when the curvature value at a certain time point is 0.5 times the threshold, the second density coefficient is 0.5. If the base step size is 3 days, the actual second sampling step size is shortened to 1.5 days. When the curvature value is only 0.2 times the threshold, the second sampling step size is extended to 0.6 days, effectively reducing redundant calculations in flat areas.

[0050] Furthermore, with an adaptively determined first sampling step size, equidistant sampling is performed within the encrypted sampling area to generate candidate replacement cycles. Sampling starts at the left boundary of the area and progresses step by step until the entire area is covered. If the remaining interval at the end is less than half a step size, the last sampling point is added to ensure complete coverage of the area.

[0051] Similarly, by using the second sampling step size to perform equally spaced sampling in the sparse sampling region, the candidate replacement cycles generated by the two sampling strategies come from different region types and have differentiated density distribution characteristics, together forming a multi-level candidate set covering the search interval.

[0052] S40: For each candidate replacement cycle, calculate the power generation reduction loss and fuel consumption increase loss from the current time to the end of the candidate replacement cycle to obtain the cumulative indirect cost, and obtain the filter purchase cost, replacement operation downtime loss, equipment life loss cost and replacement operation labor cost to obtain the direct cost; In this embodiment of the application, for each candidate replacement cycle generated above, the power generation reduction loss from the current time to the end of the candidate replacement cycle is calculated, that is, the power generation reduction loss at all time points within the selected time interval is accumulated, and the fuel consumption increase loss is the fuel consumption increase loss at all time points within the time interval. Then, the power generation reduction loss and the fuel consumption increase loss are added together to obtain the cumulative indirect cost.

[0053] Next, the direct cost is obtained by summing the filter purchase cost, downtime loss during replacement, equipment life loss cost, and replacement time cost. This direct cost remains basically constant in a single replacement operation and does not change significantly with the specific location of the candidate replacement cycle, but it will repeat at different frequencies in different replacement cycle decision schemes.

[0054] Specifically, step S40 in the method includes: For each candidate replacement cycle, extract the sequence of predicted efficiency loss values ​​from the predicted efficiency loss curve, within the time interval from the current moment to the end of the candidate replacement cycle. The theoretical output power of the intake filtration system under unclogging filter conditions is obtained. Each efficiency loss value in the efficiency loss prediction value sequence is multiplied by the theoretical output power to obtain the power loss value at each time point. The power loss value is multiplied by the corresponding power generation duration to obtain the power generation reduction value. The power generation reduction value is multiplied by the preset grid-connected electricity price to obtain the power generation reduction loss at the time point. The power generation reduction losses at all time points within the time interval are accumulated to obtain the power generation reduction loss. The theoretical fuel consumption of the intake filtration system under unclogging conditions is obtained. Each efficiency loss value in the efficiency loss prediction value sequence is multiplied by the theoretical fuel consumption to obtain the fuel consumption increase value at each time point. The fuel consumption increase value is multiplied by a preset fuel price to obtain the fuel consumption increase loss at that time point. The fuel consumption increase losses at all time points within the time interval are accumulated to obtain the fuel consumption increase loss. The cumulative indirect cost is obtained by adding the loss from the reduction in power generation to the loss from the increase in fuel consumption. The filter purchase price is used as the filter purchase cost; the downtime required for replacement is multiplied by the power generation revenue per unit time as the downtime loss for replacement; the estimated labor hours consumed for replacement are multiplied by the preset labor hour price as the labor hour cost for replacement; and the equipment life loss cost is obtained. The direct cost is obtained by summing up the filter purchase cost, the downtime loss during replacement, the equipment life loss cost, and the labor cost of replacement.

[0055] In this embodiment of the application, firstly, for each candidate replacement cycle, the time interval from the current moment to the end of the candidate replacement cycle is extracted from the predicted efficiency loss curve to form a sequence of predicted efficiency loss values ​​within the interval. This sequence reflects the degradation trajectory of filter performance over time.

[0056] Secondly, the theoretical output power of the intake filtration system under unclogging conditions is obtained. This theoretical value is determined based on unit design parameters, environmental operating conditions, and historical clean filter operation data, representing the baseline power generation capacity when the filter is in pristine condition. Each efficiency loss value in the efficiency loss prediction sequence is multiplied by the theoretical output power to obtain the power loss value at each time point due to increased intake resistance. This power loss value reflects the real-time impact of filter clogging on unit output; the more severe the clogging, the greater the power loss.

[0057] Next, the power loss value at each time point is multiplied by the corresponding power generation duration to obtain the power generation reduction value at that time point. The determination of the power generation duration needs to take into account the unit dispatch plan, maintenance arrangement and load forecast. For peak-shaving units, the operating differences between peak and valley periods should be distinguished. Then, the power generation reduction value is multiplied by the preset on-grid electricity price to obtain the power generation reduction loss at that time point. The on-grid electricity price can be the annual long-term contract average price, the spot market forecast average price or the weighted average price. The specific value is determined according to the company's power trading strategy. Finally, the power generation reduction losses at all time points within the time interval are accumulated to form the total power generation reduction loss within the candidate replacement cycle.

[0058] The theoretical fuel consumption of the intake filtration system under unclogging conditions is obtained synchronously. This value is based on the unit's thermal efficiency design value, fuel characteristic parameters, and heat balance calculation under standard operating conditions. Each efficiency loss value in the efficiency loss prediction value sequence is multiplied by the theoretical fuel consumption to obtain the amount of extra fuel consumed at each time point to maintain the same output, i.e., the fuel consumption increase value. This increase value reflects the fuel waste caused by the decrease in thermal efficiency due to filter clogging.

[0059] Then, the increase in fuel consumption at each time point is multiplied by the preset fuel price to obtain the fuel consumption increase loss at that time point. The fuel price should take into account the purchase contract price, transportation costs, taxes and price fluctuation terms. For units that use multiple fuels for co-firing, the weighted average fuel price is calculated according to the actual co-firing ratio. The fuel consumption increase loss at all time points within the time interval is accumulated to form the total fuel consumption increase loss within the candidate replacement cycle.

[0060] Furthermore, by adding the aforementioned losses from reduced power generation to the losses from increased fuel consumption, we obtain the cumulative indirect cost. This cost indicator quantifies the ongoing economic losses caused by performance degradation during the delayed replacement period of the filter, and its magnitude is positively correlated with the length of the candidate replacement cycle.

[0061] Next, the unit price of the filter is obtained as the filter procurement cost. This unit price comes from the spare parts master data or the latest purchase order price in the enterprise's ERP system. For imported filters, the impact of exchange rate fluctuations and tariffs needs to be considered. The downtime required for the replacement operation is obtained. This time consists of the filter disassembly and assembly time, system isolation and recovery time, and necessary safety inspection time as specified in the standard maintenance procedure. Combined with the unit capacity electricity price loss or grid assessment standards, the downtime loss for the replacement operation is calculated.

[0062] Specifically, the estimated labor hours for replacement work are obtained, including the direct working hours of maintenance personnel, technical preparation hours, and quality acceptance hours. These are multiplied by a preset hourly rate to obtain the replacement work's labor cost. The hourly rate is determined based on the company's internal salary standards or outsourcing service contracts, taking into account skill levels and work periods. Equipment lifespan depreciation costs are obtained based on historical failure data. The arithmetic mean of historical lifespan depreciation costs is calculated to arrive at the lifespan depreciation cost.

[0063] Finally, the costs of filter procurement, downtime losses during replacement, equipment lifespan depreciation, and replacement labor time are summarized to obtain the direct cost. This cost represents all fixed expenditures required for a single replacement operation and is not directly related to the timing of replacement. However, it occurs at different frequencies in different replacement cycle decision schemes and constitutes an important part of total cost optimization.

[0064] Furthermore, such as Figure 2 As shown, the cost of equipment lifespan depreciation includes: Obtain the predicted efficiency loss value from the predicted efficiency loss curve, input the predicted efficiency loss value and the corresponding environmental parameters and unit operating parameters into the pre-trained filter pressure difference inversion model, and output the predicted filter pressure difference value in the time interval from the current moment to the end of the candidate replacement cycle. Based on machine learning, a wear rate mapping model of the intake filtration system under filter clogging conditions is constructed, wherein the wear rate mapping model is used to describe the relationship between filter pressure difference and blade wear rate. The predicted filter pressure difference is input into the wear rate mapping model, and the predicted blade wear rate is output. The predicted blade wear rate is integrated over the time interval to obtain the blade wear increment. The blade wear increment is then multiplied by the blade replacement cost per unit to obtain the equipment life loss cost.

[0065] In this embodiment, firstly, the predicted efficiency loss value in the predicted efficiency loss curve is obtained, and it is input together with the corresponding environmental parameters and unit operating parameters into a pre-trained filter pressure difference inversion model. This inversion model is jointly trained based on computational fluid dynamics simulation data and on-site measured pressure difference samples, and can establish a nonlinear mapping relationship between filter efficiency loss and intake system pressure difference, and output a sequence of predicted filter pressure difference values ​​from the current moment to the end of the candidate replacement cycle. The unit operating parameters include at least the gas turbine load and intake flow rate.

[0066] Secondly, a wear rate mapping model of the intake filtration system under filter clogging conditions is constructed based on machine learning algorithms. This model is based on historical operation and maintenance data, and collects compressor blade morphology detection data, coating thickness measurement data and vibration monitoring data under different pressure difference conditions. The gradient boosting tree ensemble learning method is used to establish a quantitative relationship between filter pressure difference and blade wear rate. The input features of the wear rate mapping model include steady-state pressure difference, pressure difference fluctuation amplitude, pressure difference change rate and cumulative running time. The output is the volumetric wear rate of blade material or coating failure rate, reflecting the erosion intensity of filter clogging on compressor flow passage components.

[0067] For example, the wear rate mapping model is trained based on machine learning, and the steps are as follows: A gradient boosting decision tree algorithm was used to construct a wear rate prediction model. The maximum depth of the tree was set to 8 to 12 layers, the learning rate was controlled within the range of 0.05 to 0.1, and the subsampling ratio was set to 0.8 to prevent overfitting. The optimal number of iterations was determined by cross-validation. During the model training process, the root mean square error and the mean absolute percentage error were used as evaluation indicators to quantitatively evaluate the model's prediction accuracy. When the mean absolute percentage error on the validation set was less than 15% and the coefficient of determination was greater than 0.85, the model was deemed to meet the requirements for engineering applications.

[0068] Next, the filter pressure difference prediction sequence is input into the wear rate mapping model, and the blade wear rate prediction value at each moment is calculated point by point. This prediction value takes into account the influence of pressure difference level and its dynamic changes on the wear mechanism. High amplitude pulsating pressure difference will aggravate the fatigue damage and particle erosion of the blade, while continuous high pressure difference will lead to boundary layer separation and flow instability, accelerating coating peeling.

[0069] Furthermore, the predicted blade wear rate is numerically integrated over a time interval to obtain the blade wear increment within the candidate replacement cycle. The integration method adopts the trapezoidal rule or Simpson's rule to ensure the controllability of time discretization error. The physical meaning of the wear increment is the cumulative value of blade material loss or coating thickness reduction, which is directly related to the remaining service life of the blade.

[0070] Finally, the unit cost of blade replacement is obtained, which includes the cost of purchasing new blades, disassembly and assembly labor costs, dynamic balancing testing costs, and debugging costs. For blades using advanced coating technology or complex cooling structures, the cost options of repair or remanufacturing must also be considered. Multiplying the blade wear increment by the unit cost of blade replacement yields the equipment life loss cost, which quantifies the depreciation value of delayed filter replacement on the lifespan of the compressor's core components.

[0071] Furthermore, the process of constructing the filter pressure difference inversion model includes: Efficiency loss value, filter pressure difference value, and environmental parameters and unit operating parameters at the same time are obtained from historical operating data. The efficiency loss value, environmental parameters, and unit operating parameters are used as input features, and the filter pressure difference value is used as output label to form a pressure difference inversion training sample set. The environmental parameters include at least air density and air velocity, and the unit operating parameters include at least gas turbine load and intake air flow. Based on the filter resistance equation in fluid mechanics and the efficiency loss mechanism model in gas turbine thermodynamics, physical constraint terms are established. An initial filter pressure difference inversion model is constructed based on a deep neural network. The initial filter pressure difference inversion model uses efficiency loss value, environmental parameters and unit operating parameters as input layer nodes and filter pressure difference value as output layer node. During the training process, the predicted filter pressure difference output by the initial filter pressure difference inversion model is substituted into the physical constraint term to calculate the physical residual. The physical residual is used as the physical constraint loss term, and the filter pressure difference value in the pressure difference inversion training sample set is used as the supervision signal to calculate the data fitting loss term between the predicted filter pressure difference value and the filter pressure difference value. The total loss function is constructed by weighted summing of the physical constraint loss term and the data fitting loss term. The network parameters of the initial filter pressure difference inversion model and the unknown coefficients in the physical constraint terms are iteratively optimized until the total loss function is minimized, thus obtaining the pre-trained filter pressure difference inversion model.

[0072] In this embodiment, firstly, efficiency loss values, filter pressure differential values, and corresponding environmental and unit operating parameters are extracted from historical operating data. Environmental parameters include at least air density and air velocity. Air density is calculated from ambient temperature, atmospheric pressure, and humidity, reflecting the thermodynamic state of the intake working fluid. Air velocity is calculated based on the intake flow rate and filter flow area, characterizing the speed at which airflow passes through the filter. Unit operating parameters include at least gas turbine load and intake flow rate. Load reflects the unit's output level, while intake flow rate is coupled with compressor speed, adjustable guide vane opening, and environmental conditions. Efficiency loss values, environmental parameters, and unit operating parameters are used as input features, and filter pressure differential values ​​are used as output labels to construct a pressure differential inversion training sample set. Sample collection must cover different seasons, different load conditions, and different filter aging stages to ensure the representativeness and diversity of the sample distribution.

[0073] Secondly, according to the filter resistance equation in fluid mechanics, the filter pressure difference is proportional to the square of the air velocity and has a linear relationship with the air density. At the same time, affected by the degree of filter blockage, the efficiency loss mechanism model in gas turbine thermodynamics reflects the inherent law that the increase in intake pressure difference leads to the increase in compressor power consumption and the decrease in turbine work capacity. The above physical laws are transformed into mathematical constraints, and physical constraint terms are established. The pressure difference increment and efficiency loss should satisfy a monotonically increasing relationship.

[0074] For example, the filter resistance equation is expressed as ΔP=ξ·1 / 2ρv 2 Where ΔP is the filter pressure difference, ξ is the filter resistance coefficient, ρ is the air density, and v is the air velocity, the efficiency loss mechanism model is expressed as η=k·ΔP, where η is the efficiency loss value and k is the proportionality coefficient.

[0075] Furthermore, an initial filter pressure difference inversion model is constructed based on a deep neural network. The network structure adopts a combination of multi-layer fully connected layers and batch normalization layers. The input layer nodes correspond to efficiency loss values, environmental parameters, and unit operating parameters, while the output layer nodes are filter pressure difference values. The hidden layer introduces a nonlinear activation function to capture the complex coupling relationship between parameters, and the output layer adopts linear activation to ensure the continuous output range of the pressure difference value.

[0076] During training, the predicted filter pressure from the initial filter pressure inversion model is substituted into the physical constraint term to calculate the physical residual. The physical residual refers to the degree to which the predicted value violates physical laws, reflecting the deviation of the model's prediction from fluid dynamics equations and thermodynamic mechanisms. The smaller the residual, the more the prediction conforms to physical laws. For example, when efficiency loss increases while the predicted pressure decreases, or when the predicted pressure exceeds the theoretical boundary based on fluid dynamics equations, a positive residual penalty is generated. The physical residual is used as the physical constraint loss term, and the measured filter pressure values ​​in the pressure inversion training sample set are used as the supervision signal to calculate the mean square error or mean absolute error between the predicted and measured filter pressure values, forming the data fitting loss term.

[0077] Specifically, the calculation process for the physical residual is as follows: Assume that at a certain moment, the input characteristics are an efficiency loss value η = 0.015, i.e., 1.5%, and an air density ρ = 1.18 kg / m³. 3 The air velocity is v=12.5m / s, the gas turbine load is L=85% of the rated load, the intake air flow rate is Q=650kg / s, and the initial filter pressure difference inversion model predicts the output filter pressure difference value as 520Pa.

[0078] First, according to the filter resistance equation ΔP=ξ·1 / 2ρv 2 The current filter resistance coefficient ξ = 5.8, determined from historical data, varies with the degree of clogging. This coefficient is an empirical value based on the correlation with efficiency loss. The theoretical pressure difference baseline value is calculated as: 5.8 × 0.5 × 1.18 × 12.5 2 =535.9Pa. The physical constraint requires that the relative deviation between the predicted pressure difference and the theoretical pressure difference benchmark value be controlled within a reasonable range. The allowable deviation threshold is set at ±8%, that is, the effective range is [492.9Pa, 578.8Pa].

[0079] Secondly, based on the efficiency loss mechanism model η=k·ΔP, the proportionality coefficient k=2.8×10 is obtained from historical sample regression. -5 Pa -1 The efficiency loss value based on the predicted pressure difference inversion is 2.8 × 10⁻⁶. -5 ×520=0.01456, compared with the input efficiency loss value η=0.015, the efficiency loss residual is obtained as |0.015-0.01456| / 0.015=2.93%.

[0080] Next, the physical residuals are calculated comprehensively. The pressure deviation residual = |520-535.9| / 535.9 = 2.97%, which is within the allowable threshold and does not incur additional penalties. The efficiency consistency residual = 2.93%. With weighting coefficients of 0.6 and 0.4 respectively, the weighted physical residual = 0.6×max(0,2.97%-8%)+0.4×2.93% = 1.17%. Since the pressure deviation is within the allowable range, this residual is zero. Only the efficiency consistency residual contributes 1.17% of the physical constraint loss.

[0081] Then, the data fitting loss term is calculated. At this moment, the measured filter pressure difference is 548 Pa. The absolute error between the predicted value and the measured value is |520-548|=28 Pa, and the relative error is 5.11%. The data fitting loss in the form of mean square error is (520-548). 2 / 1000=0.784, the normalization coefficient is set to 1000 to balance the order of magnitude.

[0082] Meanwhile, the measured filter pressure difference values ​​in the pressure difference inversion training sample set are used as supervision signals. The mean square error or mean absolute error between the predicted and measured filter pressure difference values ​​is calculated as a data fitting loss term. The data fitting loss term ensures that the model can accurately reproduce historical observation data and avoids over-reliance on physical priors while ignoring actual operating characteristics.

[0083] Furthermore, the physical constraint loss term and the data fitting loss term are weighted and summed to construct the total loss function. The weight coefficients are dynamically adjusted based on the confidence level of the physical constraints and the data quality. Higher weights are assigned to physical laws with high confidence, while the influence weight of measured data with more noise is appropriately reduced. For example, if the weight of the physical constraint loss is set to 0.3 and the weight of the data fitting loss is set to 0.7, then the total loss = 0.3 × 1.17% + 0.7 × 0.784 ≈ 0.552.

[0084] The training process incorporates an early stopping mechanism and a learning rate decay strategy to prevent overfitting and accelerate convergence. When the total loss function no longer decreases on the validation set or reaches the preset number of iterations (e.g., 10 iterations), training is terminated, resulting in a pre-trained filter pressure difference inversion model.

[0085] For example, the initial filter pressure difference inversion model is constructed and trained based on a deep neural network, and the steps are as follows: First, a fully connected neural network with four hidden layers is constructed, with the number of neurons in each hidden layer being 128, 256, 128, and 64 respectively. Each fully connected layer is followed by a batch normalization layer to stabilize data distribution drift during training. The momentum coefficient of the batch normalization layer is set to 0.1. The hidden layer activation function is LeakyReLU, with a negative slope of 0.01 to alleviate the vanishing gradient problem while preserving the ability to transmit negative information. The output layer is a single neuron using the identity mapping activation function, directly outputting the predicted filter pressure difference. The network weights are initialized using a He normal distribution, the bias term is initialized to zero, and the initial learning rate is set to 0.001.

[0086] Furthermore, the filter resistance equation ΔP=ξ·1 / 2ρv 2 Combining this with the efficiency loss mechanism model η=k·ΔP, and eliminating the pressure difference variable ΔP, we obtain the theoretical relationship between efficiency loss and air density and air velocity: η=k·ξ·1 / 2ρv 2 The physical residual is defined as the deviation between the estimated efficiency loss of the model's implicit output and the above theoretical relationship, i.e., Lp=|ηp-k·ξ·1 / 2ρv 2 | where ηp is obtained by forward propagation of the input features through the network, and k and ξ are learnable physical parameters that are optimized synchronously with the network weights. For example, k can be 0.8 and ξ can be initialized to 1.2.

[0087] During training, the Adam optimizer is used to update parameters. The training batch size is set to 32 or 64 based on the total number of samples. After each training cycle, the total loss function value and the mean physical residual are evaluated on the validation set. If the validation loss does not decrease for 5 consecutive cycles, an early stopping mechanism is triggered, and the current optimal model weights are saved. Otherwise, the learning rate decays exponentially with a decay coefficient of 0.95, and iterative optimization continues until the early stopping condition is met or the maximum number of iterations of 100 is reached. Finally, the pre-trained filter pressure difference inversion model is output.

[0088] S50: Based on the cumulative indirect costs and the direct costs, calculate the total cost corresponding to each candidate replacement cycle, and select the candidate replacement cycle with the minimum total cost as the optimal replacement cycle.

[0089] In this embodiment of the application, based on the cumulative indirect costs and direct costs, the cumulative indirect costs and direct costs corresponding to the period are summed to obtain the total cost under the candidate replacement period. The two are combined to form a complete life cycle cost assessment framework.

[0090] Furthermore, by comparing and selecting the candidate replacement cycle with the lowest total cost as the optimal replacement cycle, this total cost function comprehensively reflects the economic trade-offs in choosing the timing of filter replacement. That is, although extending the replacement cycle can reduce the amortization frequency of direct costs, it will lead to a continuous increase in cumulative indirect costs; although shortening the replacement cycle can suppress performance degradation losses, it will increase the frequency of direct cost expenditures.

[0091] In summary, compared with existing technologies, this application achieves accurate prediction of the trend of efficiency loss of the intake filtration system within a future preset time window, breaking through the limitation of the traditional fixed-cycle replacement mode being unable to adapt to dynamic operating conditions.

[0092] In summary, the embodiments of this application have at least the following technical effects: This application provides a machine learning-based optimization method for the replacement cycle of an air intake filter system. First, it acquires operational data of the air intake filter system to construct a multi-dimensional state perception system. Second, it inputs the operational data into a pre-trained efficiency loss prediction model, utilizing the powerful modeling capabilities of long short-term memory networks for time-series data to output a predicted efficiency loss curve within a preset future time window, achieving a shift from static state assessment to dynamic trend prediction. Third, based on the morphological characteristics of the predicted efficiency loss curve, it dynamically determines a set of candidate replacement cycles through second-derivative inflection point identification and curvature adaptive sampling strategies, ensuring both comprehensiveness of the search and improved computational efficiency. Then, for each candidate replacement cycle, it comprehensively considers cumulative indirect costs such as reduced power generation and increased fuel consumption, as well as direct costs such as filter procurement costs, replacement downtime losses, equipment lifespan depreciation costs, and replacement operation labor costs, establishing a full life-cycle cost accounting model to quantify equipment lifespan depreciation costs. Finally, based on the principle of minimizing total cost, it selects the optimal replacement cycle, achieving a balance between economic benefits and equipment reliability.

[0093] Through the above technical solutions, the technical solutions of this application embodiment can adaptively adjust the replacement cycle according to the actual operating status and environmental conditions of the air intake filtration system, avoiding resource waste caused by premature replacement and performance degradation caused by late replacement, and improving the operation and maintenance management level of the gas turbine air intake filtration system.

[0094] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0095] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for optimization of an intake filter system replacement period based on machine learning, characterized in that, The method includes: The current filter differential pressure, rate of increase of differential pressure, acceleration of increase of differential pressure, filter degradation factor, and environmental parameters of the intake filtration system are obtained to form operating data; The operational data is input into a pre-trained efficiency loss prediction model, and the predicted efficiency loss curve of the intake filtration system efficiency loss over time is output within a future preset time window. Based on the morphological characteristics of the predicted efficiency loss curve, a set of candidate replacement cycles is dynamically determined; For each candidate replacement cycle, calculate the power generation reduction loss and fuel consumption increase loss from the current time to the end of the candidate replacement cycle to obtain the cumulative indirect cost, and obtain the filter purchase cost, replacement operation downtime loss, equipment life loss cost and replacement operation labor cost to obtain the direct cost; Based on the cumulative indirect costs and the direct costs, the total cost corresponding to each candidate replacement cycle is calculated, and the candidate replacement cycle with the minimum total cost is selected as the optimal replacement cycle.

2. The method of claim 1, wherein, The current filter differential pressure, differential pressure rise rate, degradation factor, and environmental parameters of the intake filtration system are acquired to form operational data, including: The current filter pressure difference is collected in real time by differential pressure sensors installed before and after the filter screen of the intake air filtration system; Perform a sliding window linear regression on the current filter pressure difference to calculate the rate of increase of the pressure difference; Differential calculation of the pressure difference rise rate yields the filter pressure difference rise acceleration; Obtain the initial pressure difference of the filter screen in its initial clean state, calculate the ratio of the current filter screen pressure difference to the initial pressure difference, and obtain the filter screen degradation factor; Ambient temperature is collected in real time by a temperature sensor, ambient humidity is collected in real time by a humidity sensor, and particulate matter concentration data is collected in real time by a particulate matter concentration sensor. The ambient temperature, ambient humidity, and particulate matter concentration data are used as environmental parameters. The current filter pressure difference, the rate of increase of the pressure difference, the acceleration of the increase of the filter pressure difference, the filter degradation factor, and the environmental parameters are combined to form the operating data.

3. The method of claim 1, wherein: The operational data is input into a pre-trained efficiency loss prediction model, which outputs a predicted efficiency loss curve showing the change in the efficiency loss of the intake filtration system over time within a preset future time window, including: Based on the statistical data of the filter's historical replacement cycle, the historical average replacement cycle is calculated, and the length of the historical average replacement cycle is used as the length of the future preset time window. The running data is input into a pre-trained efficiency loss prediction model. The efficiency loss prediction model takes the current time as the starting point and predicts the efficiency loss value at each time point in the future preset time window step by step with a preset time step, thereby generating a predicted efficiency loss sequence. By associating the time points in the predicted efficiency loss sequence with the corresponding efficiency loss values, a predicted efficiency loss curve is formed with time as the horizontal axis and efficiency loss value as the vertical axis.

4. The method of claim 3, wherein, The process of constructing an efficiency loss prediction model includes: Multiple sets of historical data samples were collected from the historical operation database. Each set of historical data samples included historical filter pressure difference, historical pressure difference rise rate, historical filter pressure difference rise acceleration, historical filter degradation factor, historical ambient temperature, historical ambient humidity, and historical particulate matter concentration, which constituted the training sample input set. Obtain the historical air intake filtration system output power corresponding to each group of historical data samples in the training sample input set. Calculate the historical efficiency loss value based on the difference between the historical air intake filtration system output power and the theoretical output power of the air intake filtration system, and construct the training sample output set. An initial efficiency loss prediction model is constructed using a long short-term memory network; Using the training sample input set as input features and the training sample output set as supervision labels, the initial efficiency loss prediction model is trained in a supervised manner until the verification convergence is obtained, thus obtaining the trained efficiency loss prediction model.

5. The method of claim 1, wherein, Based on the morphological characteristics of the predicted efficiency loss curve, a set of candidate replacement cycles is dynamically determined, including: Calculate the second derivative of the predicted efficiency loss curve at each time point; Identify the time point when the second derivative changes from a negative value to a positive value, take the time point as the inflection point, and take the time corresponding to the inflection point as the benchmark replacement cycle; With the benchmark replacement cycle as the center, a preset time range is extended to both sides to form a search interval for candidate replacement cycles; Calculate the curvature value of the predicted efficiency loss curve at each time point within the search interval; Calculate the average curvature value of all time points within the search interval, and use the average value as the curvature threshold. Based on the curvature threshold, the search interval is divided into an encrypted sampling region and a sparse sampling region; Within the encrypted sampling area, candidate replacement cycles are generated with a first sampling step size; within the sparse sampling area, candidate replacement cycles are generated with a second sampling step size. All candidate replacement cycles are merged to generate a set of candidate replacement cycles.

6. The method of claim 5, wherein: Based on the curvature threshold, the search interval is divided into a dense sampling region and a sparse sampling region, including: The search interval is divided into multiple consecutive sub-intervals, and the average curvature value of all time points in each sub-interval is obtained by filtering. When the average curvature value within the sub-interval is greater than the curvature threshold, the sub-interval is divided into an encrypted sampling region; When the average curvature value within the sub-interval is less than or equal to the curvature threshold, the sub-interval is divided into a sparse sampling region.

7. The method of claim 5, wherein: Within the encrypted sampling region, candidate replacement periods are generated with a first sampling step size; within the sparse sampling region, candidate replacement periods are generated with a second sampling step size, including: For each time point within the encrypted sampling region, the ratio of the curvature value at that time point to the curvature threshold is calculated to obtain the first density coefficient; The first sampling step size is obtained by dividing the preset base step size by the first density coefficient; For each time point within the sparse sampling region, the ratio of the curvature value at that time point to the curvature threshold is calculated to obtain the second density coefficient; The second sampling step size is obtained by multiplying the preset base step size by the second density coefficient; Candidate replacement cycles are generated by sampling at equal intervals within the encrypted sampling region using the first sampling step size, and by sampling at equal intervals within the sparse sampling region using the second sampling step size.

8. The method of claim 1, wherein: For each candidate replacement cycle, calculate the power generation reduction loss and fuel consumption increase loss from the current time to the end of the candidate replacement cycle to obtain the cumulative indirect costs. Also, obtain the filter purchase cost, replacement operation downtime loss, equipment lifespan depreciation cost, and replacement operation labor cost to obtain the direct costs, including: For each candidate replacement cycle, extract the sequence of predicted efficiency loss values ​​from the predicted efficiency loss curve, within the time interval from the current moment to the end of the candidate replacement cycle. The theoretical output power of the intake filtration system under unclogging filter conditions is obtained. Each efficiency loss value in the efficiency loss prediction value sequence is multiplied by the theoretical output power to obtain the power loss value at each time point. The power loss value is multiplied by the corresponding power generation duration to obtain the power generation reduction value. The power generation reduction value is multiplied by the preset grid-connected electricity price to obtain the power generation reduction loss at the time point. The power generation reduction losses at all time points within the time interval are accumulated to obtain the power generation reduction loss. The theoretical fuel consumption of the intake filtration system under unclogging conditions is obtained. Each efficiency loss value in the efficiency loss prediction value sequence is multiplied by the theoretical fuel consumption to obtain the fuel consumption increase value at each time point. The fuel consumption increase value is multiplied by a preset fuel price to obtain the fuel consumption increase loss at that time point. The fuel consumption increase losses at all time points within the time interval are accumulated to obtain the fuel consumption increase loss. The cumulative indirect cost is obtained by adding the loss from the reduction in power generation to the loss from the increase in fuel consumption. The filter purchase price is used as the filter purchase cost; the downtime required for replacement is multiplied by the power generation revenue per unit time as the downtime loss for replacement; the estimated labor hours consumed for replacement are multiplied by the preset labor hour price as the labor hour cost for replacement; and the equipment lifespan loss cost is obtained. The direct cost is obtained by summing up the filter purchase cost, the downtime loss during replacement, the equipment life loss cost, and the labor cost of replacement.

9. The method of claim 8, wherein: Obtain the cost of equipment lifespan depreciation, including: Obtain the predicted efficiency loss value from the predicted efficiency loss curve, input the predicted efficiency loss value and the corresponding environmental parameters and unit operating parameters into the pre-trained filter pressure difference inversion model, and output the predicted filter pressure difference value in the time interval from the current moment to the end of the candidate replacement cycle. Based on machine learning, a wear rate mapping model of the intake filtration system under filter clogging conditions is constructed, wherein the wear rate mapping model is used to describe the relationship between filter pressure difference and blade wear rate. The predicted filter pressure difference is input into the wear rate mapping model, and the predicted blade wear rate is output. The predicted blade wear rate is integrated over the time interval to obtain the blade wear increment. The blade wear increment is then multiplied by the blade replacement cost per unit to obtain the equipment life loss cost.

10. The method of claim 9, wherein: The process of constructing the filter pressure difference inversion model includes: Efficiency loss value, filter pressure difference value, and environmental parameters and unit operating parameters at the same time are obtained from historical operating data. The efficiency loss value, environmental parameters, and unit operating parameters are used as input features, and the filter pressure difference value is used as output label to form a pressure difference inversion training sample set. The environmental parameters include at least air density and air velocity, and the unit operating parameters include at least gas turbine load and intake air flow. Based on the filter resistance equation in fluid mechanics and the efficiency loss mechanism model in gas turbine thermodynamics, physical constraint terms are established. An initial filter pressure difference inversion model is constructed based on a deep neural network. The initial filter pressure difference inversion model uses efficiency loss value, environmental parameters and unit operating parameters as input layer nodes and filter pressure difference value as output layer node. During the training process, the predicted filter pressure difference output by the initial filter pressure difference inversion model is substituted into the physical constraint term to calculate the physical residual. The physical residual is used as the physical constraint loss term, and the filter pressure difference value in the pressure difference inversion training sample set is used as the supervision signal to calculate the data fitting loss term between the predicted filter pressure difference value and the filter pressure difference value. The total loss function is constructed by weighted summing of the physical constraint loss term and the data fitting loss term. The network parameters of the initial filter pressure difference inversion model and the unknown coefficients in the physical constraint terms are iteratively optimized until the total loss function is minimized, thus obtaining the pre-trained filter pressure difference inversion model.

Images