A power plant frequency modulation control method based on a hollow causal time sequence convolution encoder

Abstract

The application discloses a thermal power plant frequency modulation control method based on a hollow causal time sequence convolution encoder, and the method comprises the following steps: synchronously collecting system frequency, active power, load power, new energy output and key equipment measurement to form a multi-dimensional sliding time window interference operation sequence; inputting the sequence into the hollow causal time sequence convolution encoder to extract short-time mutation patterns and sustained swing patterns, and gathering interference pattern feature vectors; performing cosine similarity matching on a template library containing interference type and interference intensity reference templates, calculating the discrimination reliability according to the difference between the maximum and the second maximum similarity scores; when the discrimination reliability exceeds a preset threshold, updating the corresponding template and adjusting the coupling mapping weight of the encoder to obtain a corrected feature vector, and then performing matching again and outputting the final dominant interference type, intensity and reliability, which are used as the interference direction input for frequency modulation control. The application solves the problems that the dominant interference type and intensity are difficult to be determined in real time under the multivariable linkage in the frequency modulation process, and the accuracy is easy to be lost with the change of working conditions.

Description

Technical Field

[0001] This invention relates to the field of frequency regulation operation disturbance identification technology in thermal power plants, specifically to a frequency regulation control method for thermal power plants based on a hollow causal temporal convolutional encoder. Background Technology

[0002] Thermal power units undertake primary / secondary frequency regulation and AGC (Automatic Generation Control) tracking tasks, requiring continuous adjustment of control parameters based on system frequency deviation and unit power response during operation. Grid load fluctuations, renewable energy output fluctuations, and disturbances in key components of the unit speed control system and the boiler-turbine-generator link will cause interconnected changes in frequency, active power, load, and multiple equipment measurements, which may manifest as short-term abrupt changes or continuous oscillations. Because the sources and intensities of disturbances are superimposed and similar in form under different operating conditions, frequency regulation control without a clear indication of the "current dominant disturbance" is prone to insufficient parameter reference and mismatch between control actions and disturbance mechanisms. Therefore, a technical support is typically needed in the field that can utilize multi-source measurement sequences to determine the type and intensity of the dominant disturbance within real-time boundaries, providing usable disturbance-oriented input for frequency regulation control.

[0003] In existing technologies, common approaches for frequency regulation disturbance identification and classification include: threshold / logic rule-based discrimination based on frequency and power deviations; extraction of oscillation features based on frequency domain or time-frequency analysis to identify disturbance patterns; parameter estimation based on simplified unit or grid models to infer the disturbance source; and feature engineering using historical data followed by classifiers such as support vector machines or random forests, or sequence models such as recurrent neural networks, to output the category. Other approaches solidify feature vectors of typical events into templates and perform online matching through similarity calculations to determine the corresponding disturbance type or level.

[0004] The above solutions still have some limitations in engineering applications and are not well-suited to frequency regulation scenarios: First, many methods model variables separately or focus on a single measurement, making it difficult to express the linkage between frequency, power, and key equipment measurements in the same link; second, some sequence modeling or feature extraction relies on backtracking of entire historical periods or non-causal processing, resulting in unclear boundaries for online real-time judgment; third, templates or models are mostly fixed offline, lacking a controlled update mechanism when facing changes in unit operating conditions and equipment status, leading to the risk of erroneous writing due to the drift or update of judgment criteria over time. Summary of the Invention

[0005] To address this issue, this invention provides a frequency control method for thermal power plants based on a hollow causal temporal convolutional encoder, which solves the problem that the dominant interference type and intensity are difficult to determine in real time under multi-variable linkage during frequency regulation in thermal power plants, and that the method is prone to inaccuracy as operating conditions change.

[0006] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions: According to a first aspect of the present invention, an embodiment of the present invention provides a frequency regulation control method for thermal power plants based on a dilated causal temporal convolutional encoder, the method comprising: Collect system frequency, active power, load power, new energy output and key equipment measurements during frequency regulation operation of thermal power plants, and form a time-arranged sequence of interference operation. The interference sequence is input into a dilated causal temporal convolutional encoder. Short-term mutation patterns and continuous oscillation patterns are extracted by multiple causal convolutional layers that advance according to a preset dilation rate. An interference pattern feature vector representing the temporal pattern of the interference is output. An interference reference template library is established, which includes interference type reference templates and interference intensity reference templates. The interference morphology feature vector is compared with the reference templates in the interference reference template library to obtain the most similar reference template. The confidence level is calculated based on the difference between the maximum similarity score and the second largest similarity score obtained from the similarity calculation. The discrimination confidence is compared with a preset threshold to determine whether the discrimination confidence exceeds the preset threshold. If it does, the interference reference template library is updated and the dilated causal temporal convolutional encoder is tuned. The interference running sequence is then re-inputted into the tuned dilated causal temporal convolutional encoder to obtain the corrected interference morphological feature vector. The corrected interference morphological feature vector is then compared with the reference template in the interference reference template library for similarity calculation and confidence comparison. The process is iterated until the confidence is lower than the corresponding preset threshold, and a final dominant interference determination result containing interference type, interference intensity and confidence is generated. Based on the final dominant interference determination result, determine whether to activate the frequency regulation control of the thermal power plant and the corresponding control parameters to realize the frequency regulation control of the thermal power plant.

[0007] Furthermore, data on system frequency, active power, load power, renewable energy output, and key equipment measurements during frequency regulation operation of thermal power plants are collected and compiled into a time-ordered sequence of disturbance operation data, specifically including: During the frequency regulation operation of thermal power plants, the system frequency, active power, load power, new energy output and key equipment measurements are collected synchronously at a preset sampling period, and the measurement data at each sampling time are assigned the same time identifier. Based on the same time marker, the measurement data are concatenated into a multidimensional feature vector at a single discrete time k according to a preset feature order. ; Set the sequence length corresponding to the preset time window. and at the current discrete time At the end, the multidimensional feature vectors of multiple consecutive discrete moments are arranged chronologically to form a disturbance operation sequence. ,in Represented as ; When receiving the multidimensional feature vector at a new time step, the preset sequence length of the interference running sequence is kept unchanged by sliding update to meet the real-time decision boundary of the dilated causal temporal convolutional encoder.

[0008] Furthermore, the interference sequence is input into a dilated causal temporal convolutional encoder. Short-term abrupt changes and persistent oscillations are extracted through multiple causal convolutional layers progressively increasing according to a preset dilation rate. The resulting interference morphology feature vector, representing the temporal morphology of the interference, is output, specifically including: The dilated causal temporal convolutional encoder includes a multivariate response coupling mapping layer, a dilated causal convolutional layer group, and a temporal convergence layer; wherein the dilated causal convolutional layer group includes at least three causal convolutional layers connected in series and progressively increasing according to a preset dilation rate. Interference running sequence The input is fed into a multivariate response coupling mapping layer, which is used to process the disturbance running sequence. The multi-dimensional channel transformation is performed, and the neuron response at each time step is calculated using multiple one-dimensional causal convolution kernels to obtain a first encoding sequence with a preset number of channels. ,in , Indicates the first The channel response vector at each time point, with dimension . , t∈{1,2,3,…,L}; The first encoded sequence The input is fed into a dilated causal convolutional layer group, which utilizes multiple cascaded causal convolutional layers to first perform short-time mutation morphology extraction to obtain intermediate coding sequences. Then process the intermediate encoded sequence Performing persistent oscillation pattern extraction yields a second coding sequence containing both short-term mutation patterns and persistent oscillation patterns. ; The second encoded sequence The input is fed to a temporal convergence layer, which is used to adjust the sequence length in the time dimension. The interference morphological feature vector is obtained by adaptive weighted convergence of the channel responses at each time point. , for dimensional vector, the interference morphological feature vector It also carries the linkage relationship formed by the multivariate response coupling mapping and the temporal morphological information formed by the group of dilated causal convolutional layers.

[0009] Furthermore, an interference reference template library is established, containing interference type reference templates and interference intensity reference templates. The interference morphological feature vector is compared with the reference templates in the interference reference template library to obtain the most similar reference template. The confidence level is then calculated based on the difference between the highest and second-highest similarity scores obtained from the similarity calculation. Specifically, this includes: Establish an interference reference template library containing interference type reference templates and interference intensity reference templates. The dimensions of the interference type reference templates and interference intensity reference templates are both the same as those of the interference morphology feature vector. The dimensions are the same; Set the reference template index range to to , This indicates the total number of reference template indexes, in each reference template index. Below, each corresponds to a reference template vector for the interference type. and interference intensity reference template vector And the interference type refers to the template vector. Bind interference type identifier Interference intensity reference template vector Binding interference intensity identifier ; Calculate using cosine similarity respectively With each reference template index i , The similarity scores are calculated, and the two similarity scores are fused according to the preset fusion weight to obtain the final similarity score under the corresponding reference template index i. The final similarity scores under all reference template indices are sorted to obtain the most similar reference template index corresponding to the maximum similarity score and the second most similar reference template index corresponding to the second largest similarity score. The discrimination confidence is obtained by calculating the difference between the maximum similarity score and the second largest similarity score.

[0010] Furthermore, cosine similarity is used to calculate... With each reference template index i , The similarity scores are calculated, and the two similarity scores are fused according to a preset fusion weight to obtain the final similarity score under the corresponding reference template index i. The final similarity scores under all reference template indices are sorted to obtain the most similar reference template index corresponding to the maximum similarity score and the second most similar reference template index corresponding to the second largest similarity score. The discrimination confidence is obtained by calculating the difference between the maximum similarity score and the second largest similarity score, specifically including:

[0011] in, Represents the morphological feature vector of interference and is dimensional vector; express The Dimensional components; Indicates that the reference template index is The type of interference is referenced in the template vector; express The Dimensional components; Indicates that the reference template index is The interference intensity is referenced to the template vector; express The Dimensional components; Indicates the vector dimension; Indicates the vector dimension index; Indicates the total number of reference template indexes; express and The cosine similarity score; express and The cosine similarity score; This represents the weighting of the similarity score fusion and is a preset constant; Indicates that the reference template index is Similarity score; Indicates the most similar reference template index; Indicates a secondary similarity reference template index; This represents the maximum similarity score; This represents the second-highest similarity score; This indicates the degree of credibility in the judgment.

[0012] Further, the discrimination confidence level is compared with a preset threshold to determine whether the discrimination confidence level exceeds the preset threshold. If it does, the interference reference template library is updated and the dilated causal temporal convolutional encoder is tuned. The interference running sequence is then re-inputted into the tuned dilated causal temporal convolutional encoder to obtain the corrected interference morphological feature vector, specifically including: The update of the interference reference template library includes: Based on the most similar reference template index Obtain the corresponding interference type reference template vector from the interference reference template library. With interference intensity reference template vector and obtain The bound interference type identifier and The bound interference strength identifier ; The weights are updated based on the preset template. respectively with , Perform weighted fusion to generate an updated interference type reference template. and the updated interference strength reference template and according to And the updated interference type reference template Updated interference strength reference template The difference vector is calculated. ; The updated interference type is referenced to the template vector. Compared with the updated interference strength reference template vector Backfill to the interference reference template library index is At the corresponding position, complete the update of the corresponding reference template in the interference reference template library.

[0013] Furthermore, the weights are updated based on a preset template. respectively with , Perform weighted fusion to generate an updated interference type reference template. and the updated interference strength reference template and according to And the updated interference type reference template Updated interference strength reference template The difference vector is calculated. Specifically, it includes:

[0014] in, Represents the morphological feature vector of interference and is dimensional vector; Indicates the most similar reference template index; Indicates that the index is The interference type refers to the template vector and is dimensional vector; Indicates that the index is The interference intensity is referenced to the template vector and is dimensional vector; This indicates that the preset template updates the weights and that the weights are real numbers. The updated interference type reference template vector; This represents the updated interference strength reference template vector; This indicates the preset fusion weights and is a real number; Represents the difference vector and is Dimensional vector.

[0015] Furthermore, the tuning of the dilated causal temporal convolutional encoder includes: The difference vector is used as an error signal to perform controlled adjustments on the encoding parameters of the dilated causal temporal convolutional encoder; the controlled adjustment is limited to: adjusting the mapping weights of the multivariate response coupling mapping layer in the dilated causal temporal convolutional encoder. Adjustments were made while maintaining the progressive structure of the void ratio in the dilated causal convolutional layer group; controlled adjustments were then performed to obtain... The adjustment amount of each weight term is determined, and the optimized dilated causal temporal convolutional encoder is obtained.

[0016] Furthermore, the corrected interference morphological feature vector is then compared with the reference templates in the interference reference template library for similarity calculation and confidence level comparison. This process is iterated until the confidence level falls below a corresponding preset threshold, and a final dominant interference determination result containing interference type, interference intensity, and confidence level is generated, specifically including: When the iteration continues until the confidence level falls below a preset threshold, the index of the most similar reference template corresponding to the maximum similarity score calculated in the last iteration is obtained. If multiple similarity scores are tied for the highest value, the minimum index priority rule will be used to determine the score. ; Index based on the most similar reference template Retrieve the most similar interference type reference template and the most similar interference intensity template from the interference reference template library, and obtain the interference type identifier corresponding to the most similar interference type reference template and the most similar interference intensity template. With interference intensity label ; Based on the obtained interference type identifier Interference intensity label And the final round of iteration determines the credibility of the discrimination, generating the final dominant interference judgment result.

[0017] Furthermore, based on the final dominant interference determination result, it is determined whether to activate the frequency regulation control of the thermal power plant and the corresponding control parameters to realize the frequency regulation control of the thermal power plant, specifically including: The confidence level is read from the final dominant interference determination result. It is then determined whether the confidence level exceeds the control activation threshold. If it does, the interference type identifier is read from the final dominant interference determination result. and interference intensity marking And identified by interference type and interference intensity marking The corresponding frequency modulation control parameters are retrieved from a pre-established frequency modulation control parameter table as an index. The frequency modulation control parameters include primary frequency modulation gain, secondary frequency modulation gain, frequency modulation dead zone, power ramp-up limit, and command filtering time constant.

[0018] Compared with existing technologies, this invention provides a frequency regulation control method for thermal power plants based on a hollow causal temporal convolutional encoder, comprising: collecting system frequency, active power, load power, new energy output, and key equipment measurements during frequency regulation operation of the thermal power plant, and forming a time-arranged disturbance operation sequence; inputting the disturbance operation sequence into a hollow causal temporal convolutional encoder, extracting short-term abrupt change patterns and continuous oscillation patterns through multiple causal convolutional layers progressively according to a preset void ratio, and outputting a disturbance pattern feature vector characterizing the disturbance temporal pattern; establishing a disturbance reference template library containing disturbance type reference templates and disturbance intensity reference templates, performing similarity calculations between the disturbance pattern feature vector and the reference templates in the disturbance reference template library to obtain the most similar reference template, and calculating the maximum similarity score based on the similarity calculation. The difference in the second-largest similarity score is used to calculate the confidence level. This confidence level is then compared to a preset threshold to determine if it exceeds the threshold. If it does, an update to the interference reference template library and tuning of the dilated causal temporal convolutional encoder are triggered. The interference running sequence is then re-inputted into the tuned dilated causal temporal convolutional encoder to obtain a corrected interference morphological feature vector. This corrected interference morphological feature vector is then compared with a reference template in the interference reference template library for similarity calculation and confidence level comparison. This process is iterated until the confidence level falls below the corresponding preset threshold, generating a final dominant interference determination result containing the interference type, interference intensity, and confidence level. Based on the final dominant interference determination result, it is determined whether to activate the frequency regulation control of the thermal power plant and the corresponding control parameters to achieve frequency regulation control of the thermal power plant. This invention, in the frequency regulation operation of a thermal power plant, achieves reliable determination of the dominant interference type and intensity within the real-time boundary relying solely on historical measurements, based on multi-source measurement sequences. Under the condition of confidence level, the reference basis is updated in a controlled manner to provide executable interference direction input to the frequency regulation control. Attached Figure Description

[0019] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic diagram illustrating the workflow of a frequency modulation control method for thermal power plants based on a hollow causal temporal convolutional encoder, provided for an embodiment of the present invention; Figure 2 A strictly causal receptive field coverage map with progressive void rate in a frequency modulation control method for thermal power plants based on a void causal temporal convolutional encoder provided in an embodiment of the present invention; Figure 3This invention provides a method for frequency regulation control of thermal power plants based on a hollow causal temporal convolutional encoder, which includes "difference (confidence) gating" and triggering updates for similarity discrimination. Figure 4 This is a schematic diagram of the closed loop formed by "hollow causal time-series convolutional encoder + controlled adjustable W" in a frequency modulation control method for thermal power plants based on a hollow causal time-series convolutional encoder provided in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0021] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0022] The first embodiment of this invention provides a frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder, which is described below in conjunction with... Figure 1 Please provide a detailed explanation.

[0023] like Figure 1 As shown, in step S100, the system frequency, active power, load power, new energy output and key equipment measurements during the frequency regulation operation of the thermal power plant are collected and a time-arranged interference operation sequence is formed.

[0024] In this embodiment, a preset sampling period is set during the frequency regulation operation of a thermal power plant. and with This serves as a unified sampling trigger cycle for system frequency, active power, load power, renewable energy output, and key equipment measurements. The sampling trigger is driven by the same timing source, and the sampling time is identified using discrete time markers. Characterize and identify discrete time points. Write the various measurement data obtained from this sampling to ensure that the system frequency, active power, load power, new energy output and key equipment measurements have a consistent time identifier at the data layer.

[0025] Identification at discrete time The system frequency is denoted as follows: Active power is recorded as The load power is denoted as The output of new energy sources will be recorded as The key measurements of the equipment consist of fourteen preset measurements, each denoted as follows: to The fourteen measured quantities consist of key state quantities of frequency regulation-related equipment, specifically selected sequentially from governor setpoint, governor output, main steam pressure, main steam temperature, reheat steam pressure, boiler feedwater flow rate, fuel flow rate, turbine main steam valve opening, generator terminal voltage, reactive power, speed deviation, primary frequency regulation output, automatic generation control command, and unit output limiting status. This results in 18 dimensions, as detailed in Table 1 below. Any... Identification at discrete time If missing, use the identifier from the previous discrete time step. of The substitution method keeps the input dimensions fixed and preserves the meaning of each feature dimension.

[0026] Table 1 Key Measurement Items for Systems and Equipment

[0027] According to the preset feature order , , , and to By sequentially piecing them together, a single discrete time marker is formed. eigenvectors ,in Represented as:

[0028] The preset feature order is fixed during system initialization and remains consistent across all discrete time points, ensuring that the 18-dimensional features are consistent in terms of measurement source, arrangement order, and dimension definition. This enables the multivariate response coupling mapping layer of the dilated causal temporal convolutional encoder to perform stable channel transformations on fixed channel meanings.

[0029] Set the sequence length corresponding to the preset time window. And identified by the current discrete time. At the end, the feature vectors identified by multiple consecutive discrete time points are arranged chronologically to form an interference operation sequence. ,in Represented as When the new discrete time identifier Arrival and formation of new feature vectors At that time, the interference sequence will be updated to... This ensures that the interference sequence always maintains the preset sequence length. Furthermore, the temporal order remains unchanged. When the interference sequence is input into the dilated causal temporal convolutional encoder, the temporal order is consistent with the historical dependency boundary of the causal convolution, ensuring that the encoded output of the dilated causal temporal convolutional encoder at any discrete moment depends only on the historical measurement input and satisfies the real-time decision boundary.

[0030] like Figure 1 As shown, in step S200, the interference running sequence is input to the dilated causal temporal convolutional encoder, and short-term mutation morphology and continuous oscillation morphology are extracted by multiple causal convolutional layers that are progressively increased according to a preset dilation rate, and the interference morphology feature vector representing the temporal morphology of the interference is output.

[0031] In this embodiment, the interfering running sequence is used as the unified input to the dilated causal temporal convolutional encoder. The interfering running sequence is denoted as... ,in Indicates a discrete time marker. And the sequence length is The feature vector at each time step The system is 18-dimensional, with features consisting of system frequency, active power, load power, renewable energy output, and key equipment measurements. This is maintained during input to the dilated causal temporal convolutional encoder. The temporal order remains unchanged, so that the encoding calculation at any given time position uses only... to From historical data to current measurements, the real-time decision boundary for satisfying causal convolution is determined from the input side.

[0032] The first layer of the dilated causal temporal convolutional encoder is set as a multivariate response coupling map layer. The multivariate response coupling map layer receives... The system has eighteen input channels, and performs channel transformation on the input channels using a preset number of one-dimensional causal convolution kernels; the preset number is denoted as... , The number of channels corresponds to the first encoded sequence. The multivariate response coupling mapping layer uses causal convolution to calculate the neuron response at each time step in the time dimension, and pads the required historical length of the convolution kernel with left padding to use only measurements prior to the current time step. This allows system frequency, active power, load power, renewable energy output, and key equipment measurements to form a joint response within the same convolution kernel. The first encoded sequence output by the multivariate response coupling mapping layer is denoted as... ,in Each Indicates the first The channel response vector at each time point, with dimension . , t∈{1,2,3,…,L}.

[0033] The first encoded sequence Input a dilated causal convolutional layer array. The dilated causal convolutional layer array consists of at least three one-dimensional dilated causal convolutional layers in a cascaded structure. Each dilated causal convolutional layer introduces a dilated interval in the time dimension to expand the temporal perception range, while maintaining causal constraints in each layer, ensuring that the convolutional output at any given time depends only on historical measurements prior to that time. The dilation rate is set progressively, and the dilation rate is denoted as [missing information]. , , and satisfy Using void ratio Hollow causal convolution pairs Perform short-term mutation morphology extraction to obtain the intermediate coding sequence, denoted as . Using void ratio and The dilated causal convolution is sequentially applied to Perform continuous oscillation pattern extraction and output the second encoded sequence as follows: ,in Each Indicates the first The channel response vector at each time point. (Through...) The local mutation time series morphology obtained by covering adjacent time ranges and through , The oscillating temporal pattern obtained by expanding the time perception range is expressed within the same causal coding link, and the convolution output at any time position maintains the constraint relationship that depends only on historical measurements.

[0034] Figure 2 It demonstrates the strict causal receptive field coverage with progressive void rate, and the output point uses only historical sampling points. The "future disabled" on the right reflects strict causality. The three-layer void rate progression is as follows: dense sampling of nearest neighbors captures short-term mutations, and sparse sampling of mesoscale / long span captures continuous oscillations and long-period associations. The receptive field expands within the same link. Figure 2 This embodiment demonstrates how to prevent future information leakage and support real-time online discrimination; multi-scale morphology (mutation + oscillation) is expressed in the same vector, which is more robust than "single-scale / direct classification".

[0035] The dilated causal temporal convolutional encoder ends with a temporal convergence layer. The temporal convergence layer receives the second encoded sequence. and in the time dimension, the sequence length Adaptive weighted aggregation is performed on the channel responses at each time point. The adaptive weighted aggregation first calculates a scalar score for each time point, then normalizes the scores for all time points to obtain aggregation weights, and finally sums the channel responses at each time point according to the aggregation weights to obtain the interference morphology feature vector. The calculation is as follows:

[0036] in, Represents the morphological feature vector of interference; Indicates the length of the interfering sequence; and Indicated at the sequence length The time location index within; Indicates the second encoded sequence At time position The channel response vector; This represents the convergence weight vector of the temporal convergence layer, with dimension 1. ; Indicates the time position offset coefficient; Indicates exponentiation; This represents the calculation of the inner product score of the channel response vectors; Indicates the number of channels in the second encoded sequence and corresponds to The vector dimension. and The encoding parameters of the dilated causal temporal convolutional encoder are determined during the model training phase and remain fixed during the online inference phase of frequency modulation. This allows the temporal convergence layer to simultaneously utilize channel response content scores and temporal position biases within the same input window to complete convergence, thereby maintaining... It also carries the linkage relationship formed by the multivariate response coupling mapping and the temporal morphological information formed by the group of dilated causal convolutional layers.

[0037] The functionalities of each module in the dilated causal temporal convolutional encoder are summarized in Table 2 below: Table 2 Module Function Description Table

[0038] like Figure 1 As shown, in step S300, an interference reference template library is established, which includes interference type reference templates and interference intensity reference templates. The interference morphology feature vector is compared with the reference templates in the interference reference template library to obtain the most similar reference template. The judgment confidence is calculated based on the difference between the maximum similarity score and the second largest similarity score obtained from the similarity calculation.

[0039] In this embodiment, the interference morphological feature vector output by the dilated causal temporal convolutional encoder is used as the reference template matching input, and the interference morphological feature vector is denoted as... ,in Indicates a discrete time marker. for dimensional vector, Represents the vector dimension. When building the interference reference template library, the reference template index range is set to... to ,in This indicates the total number of reference templates in the interference reference template library; for each reference template index... In the interference reference template library, interference type reference templates and interference intensity reference templates are set respectively. The interference type reference template adopts... 3D template vector representation and denoted as And bind the interference type identifier. Interference intensity reference template adopted 3D template vector representation and denoted as And bind the interference strength identifier. By making , and Maintaining the same dimension, we achieve same-dimensional matching between the interference morphological feature vector and the interference reference template library.

[0040] The information contained in the interference reference template library is summarized in Table 3 below: Table 3 Template Information Summary Table

[0041] The initial template vector of the reference template is composed of multiple interference morphological feature vectors obtained by passing labeled interference running sequences through a dilated causal temporal convolutional encoder. This is for sequences with the same interference type identifier. The set of interference morphological feature vectors is used to generate the mean value for each dimension component. For identifiers with the same interference intensity The set of interference morphological feature vectors is used to generate the mean value for each dimension component. ,make and Each component of the is a real value that can be directly used in numerical calculations.

[0042] Will Input similarity calculation layer and read all from interference reference template library With all A matching set is formed. The similarity calculation layer uses cosine similarity to calculate the similarity separately. and Similarity score and and The similarity scores are calculated and fused with preset fusion weights to form a similarity score vector, maintaining a one-to-one correspondence between the similarity score vector and the reference template index. The similarity score vectors are then sorted by similarity to determine the most similar reference template index corresponding to the highest similarity score and the second most similar reference template index corresponding to the second highest similarity score. Based on the most similar reference template index, the corresponding interference type and interference intensity identifiers are read to generate dominant interference candidate results. The difference between the highest and second highest similarity scores is used to generate a discrimination confidence score, which is then compared with a template advantage threshold to obtain a template advantage association result. This result is used to determine whether to trigger the update of the interference reference template library and the adjustment of the encoding parameters of the dilated causal temporal convolutional encoder to enter the same processing link. The above similarity calculation, similarity sorting, discrimination confidence score calculation, and template advantage association calculation adopt the following unified calculation:

[0043] in, Represents the morphological feature vector of interference; express The Dimensional components; Indicates that the reference template index is The interference type is referenced in the template vector of the template. express The Dimensional components; Indicates that the reference template index is The interference intensity is referenced to the template vector of the template. express The Dimensional components; Indicates the vector dimension; Indicates the vector dimension index; Indicates the total number of reference templates; express and The cosine similarity score; express and The cosine similarity score; This represents the weighting of the similarity score fusion and is a preset constant; Indicates that the reference template index is Similarity score; Indicates the most similar reference template index; Indicates a secondary similarity reference template index; This represents the maximum similarity score; This represents the second-highest similarity score; Indicates the degree of credibility in the judgment; Indicates the template dominance threshold; Indicates the template's advantage in associating results; This operation selects the index corresponding to the maximum similarity score within a specified set of indices. This indicates an indicator function that takes the value when the condition is true. When the condition is not met, the value is taken as... .

[0044] Index based on the most similar reference template The index in the localization interference reference template library is Refer to the interference type template and read the interference type identifier. And locate the interference reference template library index as The interference strength reference template was used to read the interference strength identifier. ,Will and Combine to generate dominant interference candidate results and determine their credibility. Correlate with dominant interference candidate results to provide template advantage correlation results Call.

[0045] like Figure 1 As shown, in step S400, the discrimination confidence is compared with a preset threshold to determine whether the discrimination confidence exceeds the preset threshold. If it exceeds the preset threshold, the interference reference template library is updated and the dilated causal temporal convolutional encoder is tuned. The interference running sequence is then re-inputted into the tuned dilated causal temporal convolutional encoder to obtain the corrected interference morphological feature vector.

[0046] In this embodiment, the credibility will be determined. Template dominance threshold Comparison yields template advantage triggering results. .in, This is the difference between the maximum similarity score and the second maximum similarity score in the similarity score vector; The template dominance threshold is a constant value. To trigger or not trigger. As a unified threshold for updating the interference reference template library and adjusting the coding parameters of the dilated causal temporal convolutional encoder, the update of the interference reference template library and the adjustment of coding parameters are brought into the same processing link and subject to the same threshold constraint.

[0047] Template advantage trigger result Upon triggering, the corresponding reference template is located in the interference reference template library based on the dominant interference candidate result. The dominant interference candidate result includes an interference type identifier and an interference intensity identifier, denoted as [insert identifier here]. and ,in The most similar reference template index is used. Locking Interference Type Reference Template Vector With interference intensity reference template vector and lock Bound interference type identifier and Bound interference strength identifier This ensures that the template update process remains consistent in terms of the binding relationship between the reference template index and the identifier.

[0048] interference morphological feature vector The corresponding reference template is written as the update input. (Interference morphological feature vector) The temporal convergence layer is used to converge the second encoded sequence in the time dimension. 3D vector. Weights are updated using a pre-defined template. For template vector and Perform weighted fusion to generate an updated corresponding reference template, and then use the updated corresponding reference template in conjunction with... Construct the difference vector As the target constraint for adjusting encoding parameters, the following calculation is used:

[0049] in, Represents the morphological feature vector of interference and is dimensional vector; Indicates the vector dimension; Indicates the most similar reference template index; Indicates that the index is The interference type refers to the template vector and is dimensional vector; Indicates that the index is The interference intensity is referenced to the template vector and is dimensional vector; This indicates that the preset template updates the weights and that the weights are real numbers. The updated interference type reference template vector; This represents the updated interference strength reference template vector; This indicates the preset fusion weights and is a real number; Represents the difference vector and is Dimensional vector.

[0050] The updated interference type is referenced to the template vector. Compared with the updated interference strength reference template vector Backfill to the interference reference template library index is The position of the reference template library restricts the update objects to the corresponding reference template. The difference vector is then used to... As a target constraint for adjusting the coding parameters, the controlled adjustment process of the dilated causal temporal convolutional encoder is entered, so that the adjustment of the coding parameters is performed around the difference convergence direction in the reference template space.

[0051] Figure 3 This embodiment demonstrates the "difference (confidence) gating" and triggering update principle for similarity discrimination. The similarity between z and the reference template library is calculated one by one to obtain a score vector s. The difference Δ between the maximum and second-largest similarity scores is taken as the discrimination confidence, and compared with the advantage threshold τ: if Δ is sufficiently large, a dominant type / intensity candidate is output; if Δ is insufficient, the judgment is unstable, triggering template updates or entering the review / re-encoding adjustment chain. Based on the above, we can see the advantages of this invention: the output not only provides the result but also provides an actionable confidence basis; the Δ-τ gating suppresses erroneous updates and noise drift, achieving "updateable but not haphazard updates," which is more reliable and auditable than direct classification labels.

[0052] Based on the difference vector Controlled adjustments are performed on the encoding parameters of the dilated causal temporal convolutional encoder. These controlled adjustments are limited to modifying the mapping weights of the multivariate response coupling mapping layer while maintaining the progressive structure of the dilation rate in the dilated causal convolutional layer group. The mapping weights of the multivariate response coupling mapping layer are denoted as... The number of channels in the first encoded sequence is denoted as... The length of the one-dimensional causal convolution kernel of the multivariable response coupling mapping layer is denoted as... The learning rate is denoted as .Will The organization is of size The weight tensor is such that each output channel corresponds to an 18-dimensional input channel. Channel mapping weights for each historical position. Based on the difference vector. When constructing the error signal and performing reverse propagation, the error signal is reverse-distributed from the timing convergence layer to the second coded sequence. The response of each time-position channel is then back-propagated to the first encoded sequence through a group of dilated causal convolutional layers while maintaining the causal dependency direction. The error signal of the first encoded sequence and the 18-dimensional input of the interference sequence are then processed in the time dimension according to the convolution kernel length. After alignment, multiply each term and sum them in the time dimension to obtain the result. The adjustment amount for each weight term. The learning rate is used. Each weight term is adjusted once according to its corresponding adjustment amount, achieving controlled adjustment of the multivariate response coupling mapping relationship; the porosity of the dilated causal convolutional layer group. , , Remain unchanged, where , , For progressive void ratio parameters and satisfying .

[0053] Use complete mapping weights A controlled-adjustment dilated causal temporal convolutional encoder operates on the same disturbance sequence. Perform dilated causal temporal convolutional coding again and then pass it through a temporal convergence layer to obtain the corrected interference morphological feature vector, which is denoted as . .in, It represents the sequence of disturbances and is composed of 18-dimensional feature vectors from multiple consecutive moments arranged chronologically. for The dimension of the vector is kept the same as that of the reference template vectors in the interference reference template library. Template advantage triggers the result. To prevent triggering, the interference reference template library is kept unchanged, and the encoding parameters of the dilated causal temporal convolutional encoder are not adjusted. The interference morphological feature vector is then... Determined as the morphological feature vector of the correction interference .

[0054] Figure 4 This demonstrates a closed loop formed by a "diffuse causal temporal convolutional encoder + controlled adjustable weight W". The input to the dilated causal temporal convolutional encoder is a T×18 interference sequence; ① The coupling mapping layer uses causal 1D convolution to transform the 18-dimensional channels into the first encoding sequence, and the weight W is adjustable to write frequency / power / new energy / measurement linkage; ② The dilated causal convolutional layer group progressively extracts short mutations and continuous oscillations; ③ Temporal convergence yields an M-dimensional morphological vector z; feedback only acts on W, and other structures remain stable. Figure 3 The advantages of this invention can be seen: it unifies "multivariate coupling + multi-scale temporal morphology" into z; the controlled adjustment of parameters reduces the risk of drift, and it is more controllable and interpretable than end-to-end classification.

[0055] like Figure 1 As shown, in step S500, the corrected interference morphological feature vector is then compared with the reference template in the interference reference template library for similarity calculation and confidence comparison. The process is iterated until the confidence is lower than the corresponding preset threshold, and a final dominant interference determination result containing interference type, interference intensity and confidence is generated.

[0056] In this embodiment, the corrected interference morphological feature vector is input into the similarity calculation layer, and the corrected interference morphological feature vector is denoted as... ,in Indicates a discrete time marker. And for 1D vector; each dimension component For the temporal convergence layer, the second encoded sequence is... The real values ​​obtained by converging the responses of each channel in the time dimension make It carries the channel convergence response obtained by encoding the linkage relationship between system frequency, active power, load power, new energy output and key equipment measurements and their timing morphology.

[0057] The similarity calculation layer reads all reference templates from the interfering reference template library as the matching set. The total number of reference templates is denoted as... Refer to the template index as ,in Each reference template index Binding interference type reference template vector Interference intensity reference template vector Interference type identifier Interference intensity label ,in And for dimensional vector, And for dimensional vector, and All of these are real values ​​that can be directly used in numerical calculations. The interference type reference template matrix is ​​formed by stacking the templates in the order of their indexes. ,in The Behavior ; All The interference intensity reference template matrix is ​​formed by stacking templates in index order. ,in The Behavior This allows the matching set to enter the similarity calculation layer in matrix form and maintain its correlation with the index. A one-to-one correspondence.

[0058] Correcting the morphological eigenvector of the interference Calculate the similarity score for each reference template within the matching set. For each reference template index... First calculate and cosine similarity score : In dimensional index Calculate the product item by item And sum to obtain the inner product value; on the same dimension index respectively and Perform summation of squares and take the square root to obtain the modulus; divide the inner product by the product of the two moduli to obtain... Then, using the same component product summation and modulus calculation process, calculate... and cosine similarity score To ensure that the linkage pattern and duration characteristics of interference sources are jointly evaluated within the same ranking scale, a preset fusion weight is used. The fusion yields a reference template index. Similarity score ,in A preset constant whose value is between zero and one. Depend on and The similarity scores are summed to obtain the vector. The similarity scores of all reference template indices are then arranged in index order to form a similarity score vector. ,in and keep The Components and Reference Template Index One-to-one correspondence.

[0059] Similarity score vector A similarity ranking is performed to determine the final reference template corresponding to the highest similarity score. The similarity ranking is based on the index set. This is achieved by comparing each item in the database to obtain the index corresponding to the maximum similarity score. The index corresponding to the maximum similarity score is denoted as... When multiple similarity scores are tied for the highest value, the minimum index priority rule is used to determine the score. This makes the final reference template index determination process deterministic. The index will be... The reference template is determined as the final reference template, and the interference type identifier bound to the final reference template is read. With interference intensity label Identify the type of interference. With interference intensity label The final dominant interference determination result is generated by combining the results.

[0060] Compared to the structure that directly feeds the interference sequence into the classifier to output the category label, the improvements in the model structure of this embodiment are reflected in two aspects: 1) The dilated causal temporal convolutional encoder writes the linkage relationship between system frequency, active power, load power, new energy output and key equipment measurements into the same encoding link through a multivariate response coupling mapping layer, and then writes short-term mutation patterns and continuous oscillation patterns into the same interference pattern feature vector through a dilated causal convolutional layer group with progressively increasing dilation rate; 2) The updatable reference template discrimination architecture solidifies the judgment criteria of the dominant interference type and intensity into reference templates through the interference reference template library, and uses the similarity score vector and discrimination credibility to form an executable judgment basis, so that the template advantage threshold drives the update of the interference reference template library and the adjustment of the encoding parameters of the dilated causal temporal convolutional encoder into the same link.

[0061] like Figure 1As shown, in step S600, based on the final dominant interference determination result, it is determined whether to enable the frequency regulation control of the thermal power plant and the corresponding control parameters to realize the frequency regulation control of the thermal power plant.

[0062] In this embodiment, the interference type identifier and interference intensity identifier are read from the final dominant interference determination result, and the judgment confidence level is read as a confidence field. The interference type identifier corresponding to the final dominant interference determination result is denoted as... The interference intensity identifier corresponding to the final dominant interference determination result is denoted as ,in Indicates the final reference template index. and All data are stored using an enumerable numerical encoding format. The confidence level is denoted as... ,in Represents the morphological feature vector of the correction interference The corresponding time marker, This is a real number generated from the difference in similarity scores.

[0063] The interference type identifier, interference intensity identifier, and judgment confidence level are combined in a preset field order to generate the dominant interference judgment output, and a time identifier is associated with the dominant interference judgment output. The dominant interference determination output is denoted as... Using a vectorized field organization method makes It consists of four fields, in a fixed order: time identifier, interference type identifier, interference strength identifier, and trusted field. It is generated using an explicit writing method. :Will The Write to each field ,Will The Write to each field ,Will The Write to each field ,Will The Write to each field .Will and By binding the identifier field at the same time, With correction of interference morphological eigenvectors Maintain consistency across the timeline.

[0064] The output of the dominant interference determination The dominant disturbance direction for frequency regulation control in thermal power plants is input. The frequency regulation control system of thermal power plants is pre-configured with a frequency regulation control parameter table, which uses a binary key. The frequency modulation control parameter vector is indexed and denoted as . . It adopts a vector field organization method and consists of a fixed field order, with the fixed field order being the primary frequency modulation gain field. Secondary frequency modulation gain field FM dead zone field Power ramp-up limiting field Instruction filtering time constant field ,in , , , , All are real values ​​that can be directly used for control calculations. Read The Each field obtains an interference type identifier. Read The Each field obtains the interference strength identifier. Using keys The frequency modulation control parameter vector is retrieved from the frequency modulation control parameter table. .

[0065] Frequency regulation control parameters are selected for activation based on the reliability of the judgment. Activation threshold for frequency regulation control settings in thermal power plants. ,in This is a preset real number. Read. The Each field is used to determine the credibility. ,Will and Compare generated parameters enable flags ,when When Set to enabled when When Set to disabled. When When enabled, The parameters are written into the frequency regulation control register of the thermal power plant and participate in the frequency regulation control calculation; when To maintain the default frequency regulation control parameter vector in the parameter register area of ​​the thermal power plant when not enabled. ,in For a preset fixed vector and the field order is the same as Consistent. The fields Write the time register of the frequency regulation control in the thermal power plant so that the frequency regulation control calculation adopts the same... Aligned timing markers execute parameter calls and parameter enabling.

[0066] The interference detection output and frequency modulation control parameters are shown in Table 4 below: Table 4 Interference Judgment Output and Frequency Modulation Control Parameter Table

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder, characterized in that, The method includes: Collect system frequency, active power, load power, new energy output and key equipment measurements during frequency regulation operation of thermal power plants, and form a time-arranged sequence of interference operation. The interference sequence is input into a dilated causal temporal convolutional encoder. Short-term mutation patterns and continuous oscillation patterns are extracted by multiple causal convolutional layers that advance according to a preset dilation rate. An interference pattern feature vector representing the temporal pattern of the interference is output. An interference reference template library is established, which includes interference type reference templates and interference intensity reference templates. The interference morphology feature vector is compared with the reference templates in the interference reference template library to obtain the most similar reference template. The confidence level is calculated based on the difference between the maximum similarity score and the second largest similarity score obtained from the similarity calculation. The discrimination confidence is compared with a preset threshold to determine whether the discrimination confidence exceeds the preset threshold. If it does, the interference reference template library is updated and the dilated causal temporal convolutional encoder is tuned. The interference running sequence is then re-inputted into the tuned dilated causal temporal convolutional encoder to obtain the corrected interference morphological feature vector. The corrected interference morphological feature vector is then compared with the reference template in the interference reference template library for similarity calculation and confidence comparison. The process is iterated until the confidence is lower than the corresponding preset threshold, and a final dominant interference determination result containing interference type, interference intensity and confidence is generated. Based on the final dominant interference determination result, determine whether to activate the frequency regulation control of the thermal power plant and the corresponding control parameters to realize the frequency regulation control of the thermal power plant.

2. The frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder according to claim 1, characterized in that, The system frequency, active power, load power, renewable energy output, and key equipment measurements during frequency regulation operation of thermal power plants are collected and arranged into a time-ordered sequence of disturbance operation data, specifically including: During the frequency regulation operation of thermal power plants, the system frequency, active power, load power, new energy output and key equipment measurements are collected synchronously at a preset sampling period, and the measurement data at each sampling time are assigned the same time identifier. Based on the same time marker, the measurement data are concatenated into a multidimensional feature vector at a single discrete time k according to a preset feature order. ; Set the sequence length corresponding to the preset time window. and at the current discrete time At the end, the multidimensional feature vectors of multiple consecutive discrete moments are arranged chronologically to form a disturbance operation sequence. ; When receiving the multidimensional feature vector at a new time step, the preset sequence length of the interference running sequence is kept unchanged by sliding update to meet the real-time decision boundary of the dilated causal temporal convolutional encoder.

3. The frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder according to claim 2, characterized in that, The interference sequence is input into a dilated causal temporal convolutional encoder. Short-term abrupt changes and persistent oscillations are extracted through multiple causal convolutional layers progressively increasing according to a preset dilation rate. The resulting output is an interference morphology feature vector representing the temporal morphology of the interference, specifically including: The dilated causal temporal convolutional encoder includes a multivariate response coupling mapping layer, a dilated causal convolutional layer group, and a temporal convergence layer; wherein the dilated causal convolutional layer group includes at least three causal convolutional layers connected in series and progressively increasing according to a preset dilation rate. Interference running sequence The input is fed into a multivariate response coupling mapping layer, which is used to process the disturbance running sequence. The multi-dimensional channel transformation is performed, and the neuron response at each time step is calculated using multiple one-dimensional causal convolution kernels to obtain a first encoding sequence with a preset number of channels. ,in , Indicates the first The channel response vector at each time point, with dimension . , t∈{1,2,3,…,L}; The first encoded sequence The input is fed into a dilated causal convolutional layer group, which utilizes multiple cascaded causal convolutional layers to first perform short-time mutation morphology extraction to obtain intermediate coding sequences. Then process the intermediate encoded sequence Performing persistent oscillation pattern extraction yields a second coding sequence containing both short-term mutation patterns and persistent oscillation patterns. ; The second encoded sequence The input is fed to a temporal convergence layer, which is used to adjust the sequence length in the time dimension. The interference morphological feature vector is obtained by adaptive weighted convergence of the channel responses at each time point. , for dimensional vector, the interference morphological feature vector It also carries the linkage relationship formed by the multivariate response coupling mapping and the temporal morphological information formed by the group of dilated causal convolutional layers.

4. The frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder according to claim 3, characterized in that, An interference reference template library is established, containing interference type reference templates and interference intensity reference templates. The interference morphological feature vector is compared with the reference templates in the interference reference template library to obtain the most similar reference template. The confidence level is then calculated based on the difference between the highest and second-highest similarity scores obtained from the similarity calculation. Specifically, this includes: Establish an interference reference template library containing interference type reference templates and interference intensity reference templates. The dimensions of the interference type reference templates and interference intensity reference templates are both the same as those of the interference morphology feature vector. The dimensions are the same; Set the reference template index range to to , This indicates the total number of reference template indexes, in each reference template index. Below, each corresponds to a reference template vector for the interference type. and interference intensity reference template vector And the interference type refers to the template vector. Bind interference type identifier Interference intensity reference template vector Binding interference intensity identifier ; Calculate using cosine similarity respectively With each reference template index i , The similarity scores are calculated, and the two similarity scores are fused according to the preset fusion weight to obtain the final similarity score under the corresponding reference template index i. The final similarity scores under all reference template indices are sorted to obtain the most similar reference template index corresponding to the maximum similarity score and the second most similar reference template index corresponding to the second largest similarity score. The discrimination confidence is obtained by calculating the difference between the maximum similarity score and the second largest similarity score.

5. A frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder according to claim 4, characterized in that, Calculate using cosine similarity respectively With each reference template index i , The similarity scores are calculated, and the two similarity scores are fused according to a preset fusion weight to obtain the final similarity score under the corresponding reference template index i. The final similarity scores under all reference template indices are sorted to obtain the most similar reference template index corresponding to the maximum similarity score and the second most similar reference template index corresponding to the second largest similarity score. The discrimination confidence is obtained by calculating the difference between the maximum similarity score and the second largest similarity score, specifically including: in, Represents the morphological feature vector of interference and is dimensional vector; express The Dimensional components; Indicates that the reference template index is The type of interference is referenced in the template vector; express The Dimensional components; Indicates that the reference template index is The interference intensity is referenced to the template vector; express The Dimensional components; Indicates the vector dimension; Indicates the vector dimension index; Indicates the total number of reference template indexes; express and The cosine similarity score; express and The cosine similarity score; This represents the weighting of the similarity score fusion and is a preset constant; Indicates that the reference template index is Similarity score; Indicates the most similar reference template index; Indicates a secondary similarity reference template index; This represents the maximum similarity score; This represents the second-highest similarity score; This indicates the degree of credibility in the judgment.

6. The frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder according to claim 4, characterized in that, The discrimination confidence level is compared with a preset threshold to determine whether the discrimination confidence level exceeds the preset threshold. If it does, the interference reference template library is updated and the dilated causal temporal convolutional encoder is tuned. The interference running sequence is then re-inputted into the tuned dilated causal temporal convolutional encoder to obtain the corrected interference morphological feature vector, specifically including: The update of the interference reference template library includes: Based on the most similar reference template index Obtain the corresponding interference type reference template vector from the interference reference template library. With interference intensity reference template vector and obtain The bound interference type identifier and The bound interference strength identifier ; The weights are updated based on the preset template. respectively with , Perform weighted fusion to generate an updated interference type reference template. and the updated interference strength reference template and according to And the updated interference type reference template Updated interference strength reference template The difference vector is calculated. ; The updated interference type is referenced to the template vector. Compared with the updated interference strength reference template vector Backfill to the interference reference template library index is At the corresponding position, complete the update of the corresponding reference template in the interference reference template library.

7. A frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder according to claim 6, characterized in that, The weights are updated based on the preset template. respectively with , Perform weighted fusion to generate an updated interference type reference template. and the updated interference strength reference template and according to And the updated interference type reference template Updated interference strength reference template The difference vector is calculated. Specifically, it includes: in, Represents the morphological feature vector of interference and is dimensional vector; Indicates the most similar reference template index; Indicates that the index is The interference type refers to the template vector and is dimensional vector; Indicates that the index is The interference intensity is referenced to the template vector and is dimensional vector; This indicates that the preset template updates the weights and that the weights are real numbers. The updated interference type reference template vector; This represents the updated interference strength reference template vector; This indicates the preset fusion weights and is a real number; Represents the difference vector and is Dimensional vector.

8. A frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder according to claim 6, characterized in that, The tuning of the dilated causal temporal convolutional encoder includes: The difference vector is used as an error signal to perform controlled adjustments on the encoding parameters of the dilated causal temporal convolutional encoder; the controlled adjustment is limited to: adjusting the mapping weights of the multivariate response coupling mapping layer in the dilated causal temporal convolutional encoder. Adjustments were made while maintaining the progressive structure of the void ratio in the dilated causal convolutional layer group; controlled adjustments were then performed to obtain... The adjustment amount of each weight term is determined, and the optimized dilated causal temporal convolutional encoder is obtained.

9. A frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder according to claim 1, characterized in that, The corrected interference morphological feature vector is then compared with the reference templates in the interference reference template library for similarity calculation and confidence level assessment. This process is iterated until the confidence level falls below a corresponding preset threshold, and a final dominant interference determination result containing interference type, interference intensity, and confidence level is generated. Specifically, this includes: When the iteration continues until the confidence level falls below a preset threshold, the index of the most similar reference template corresponding to the maximum similarity score calculated in the last iteration is obtained. If multiple similarity scores are tied for the highest value, the minimum index priority rule will be used to determine the score. ; Index based on the most similar reference template Retrieve the most similar interference type reference template and the most similar interference intensity template from the interference reference template library, and obtain the interference type identifier corresponding to the most similar interference type reference template and the most similar interference intensity template. With interference intensity label ; Based on the obtained interference type identifier Interference intensity label And the final round of iteration determines the credibility of the discrimination, generating the final dominant interference judgment result.

10. A frequency modulation control method for thermal power plants based on a dilated causal temporal convolutional encoder according to claim 9, characterized in that, Based on the final dominant interference determination result, determine whether to activate the frequency regulation control of the thermal power plant and the corresponding control parameters to realize the frequency regulation control of the thermal power plant, specifically including: The confidence level is read from the final dominant interference determination result. It is then determined whether the confidence level exceeds the control activation threshold. If it does, the interference type identifier is read from the final dominant interference determination result. and interference intensity marking And identified by interference type and interference intensity marking The corresponding frequency modulation control parameters are retrieved from a pre-established frequency modulation control parameter table as an index. The frequency modulation control parameters include primary frequency modulation gain, secondary frequency modulation gain, frequency modulation dead zone, power ramp-up limit, and command filtering time constant.

Images