Precise quantification and response margin dynamic prediction method for primary frequency modulation contribution electric quantity

Abstract

The invention discloses a primary frequency modulation contribution electric quantity accurate quantification and response margin dynamic prediction method comprising the following steps: collecting power grid real-time operation data, environment association data and digital twin mirror image data, and carrying out data preprocessing; based on the preprocessed data, calculating frequency modulation contribution electric quantity data by adopting a multi-target self-adaptive weighted integral algorithm model; based on the frequency modulation contribution electric quantity data, constructing a bidirectional LSTM prediction model fusing transfer learning and an attention mechanism, and outputting response margin probability distribution and a confidence interval; and the predicted response margin probability distribution and confidence interval realize prediction result encryption transmission and distributed optimization through a block chain-edge computing collaborative architecture, and a real-time feedback channel is established to dynamically adjust model parameters. Through the multi-source heterogeneous data fusion and digital twin mirroring technology, the holographic sensing of the power grid state is realized, the data noise is reduced, and the abnormal value repair accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system frequency modulation, in particular to a method for precise quantification of primary frequency modulation contribution and dynamic prediction of response margin. BACKGROUND

[0002] With the increase of new energy grid-connected proportion and the complexity of power grid operation conditions, the traditional frequency modulation contribution quantification method has three technical bottlenecks: first, the data acquisition dimension is single, relying only on real-time operation data of the power grid, without integrating environmental factors and equipment health status, resulting in serious measurement noise interference; second, the quantification model is static and fixed, using a fixed weight integral algorithm, which cannot dynamically adapt to the frequency fluctuation characteristics of the power grid and the response characteristic differences of different frequency modulation resources; third, the prediction accuracy of the response margin is insufficient, and the traditional time series model cannot capture the time-frequency domain correlation of multi-source characteristics, and lacks cross-scene migration ability. In the prior art, such as patent CN202310212345.6, the basic integral algorithm is used without considering dynamic optimization of the weight; the prediction model of patent CN202211567890.X does not introduce attention mechanism and spatiotemporal feature fusion, resulting in a prediction error rate of more than 15%, which cannot meet the demand of new power system for fine management of frequency modulation resources. SUMMARY

[0003] The purpose of the present application is to provide a method for precise quantification of primary frequency modulation contribution and dynamic prediction of response margin to solve the above problems.

[0004] To achieve the above purpose, the present application adopts the following technical solutions: In a first aspect, the present application provides a method for precise quantification of primary frequency modulation contribution and dynamic prediction of response margin, comprising: Collecting real-time operation data of the power grid, environment-related data and digital twin mirror image data, and performing data preprocessing through Kalman filter-wavelet threshold joint denoising and isolated forest-generative adversarial network fusion algorithm; Based on the preprocessed data, a multi-objective adaptive weighted integral algorithm model is used to calculate the frequency modulation contribution data; Based on the frequency modulation contribution data, a bidirectional LSTM prediction model is constructed by integrating transfer learning and attention mechanism, the time-frequency domain correlation features are extracted through the multi-head self-attention layer, the time series prediction is performed through the bidirectional LSTM layer, and the response margin probability distribution and confidence interval are outputted; The predicted response margin probability distribution and confidence interval are transmitted and distributed through the blockchain-edge computing collaborative architecture to realize prediction result encryption transmission and distributed optimization, and a real-time feedback channel is established to dynamically adjust the model parameters.

[0005] Furthermore, the collected real-time power grid operation data, environmental correlation data, and digital twin mirror data are preprocessed using a Kalman filter-wavelet thresholding joint denoising algorithm and an isolated forest-generative adversarial network fusion algorithm, including: A multi-source heterogeneous data acquisition system is constructed to collect real-time power grid operation data, environmental correlation data, and digital twin mirror data. The real-time operation data includes frequency, voltage, current, and output data of each generator unit. The environmental correlation data includes meteorological data, load forecast data, and equipment health data. The digital twin mirror data is used to map the power grid topology and component parameter changes in real time through a 3D modeling engine. The collected data undergoes spatiotemporal alignment preprocessing, and measurement noise is eliminated using a Kalman filter-wavelet threshold joint denoising algorithm. An outlier detection and repair are performed using an isolated forest-generative adversarial network fusion algorithm.

[0006] Furthermore, the digital twin mirror data acquired by the multi-source heterogeneous data acquisition system is mapped in real time through a 3D modeling engine to reflect changes in the power grid topology and component parameters, with a sampling frequency of not less than 50Hz. The specific implementation of the Kalman filter-wavelet threshold joint denoising algorithm is as follows: first, state estimation is performed by Kalman filtering, then the residual signal is decomposed into three levels using the db4 wavelet basis, and the signal is reconstructed after processing the high-frequency coefficients by the heuristic threshold function. In the isolated forest-generative adversarial network fusion algorithm, the isolated forest is used to initially detect outliers, the generator of the generative adversarial network is used to repair the outlier data, and the discriminator simultaneously evaluates the authenticity of the original data and the repaired data.

[0007] Furthermore, the calculation of frequency regulation contribution power data based on the preprocessed data using a multi-objective adaptive weighted integral algorithm model includes: The multi-objective adaptive weighted integral algorithm model is Hi=∫(t0-t1)W(t,θ,λ)×[Pt-P0-Si×Ri×(t-t0)]dt, where Hi is the primary frequency regulation weighted integral power of unit i, t0 is the time when the system frequency exceeds the primary frequency regulation dead zone of unit i, t1 is the time when the system frequency enters the primary frequency regulation dead zone of unit i, Pt is the actual power generation of unit i at time t, P0 is the actual power generation of unit i at time t0, Si is the unit state parameter, 0 indicates that AGC or planned command does not adjust the load, 1 indicates that load adjustment is in progress, Ri is the average ramp rate of unit i; W(t,θ,λ) is the weighted coefficient vector based on multi-objective deep reinforcement learning dynamic optimization, θ is the grid state feature vector, λ is the dual-objective optimization coefficient of economy and response speed, and its element Wn=1 / (σn√2π)×exp[-(ln(n)-μ(θ,λ))]. 2 / (2σ(θ,λ) 2], wherein n is the sequence number of the sampling time, μ(θ, λ) and σ(θ, λ) are the position parameters and scale parameters updated in real time by the MO-DQN algorithm, and satisfy ∑Wn=1.

[0008] Further, the double-target optimization coefficient λ of the multi-target adaptive weighted integration algorithm model is determined by the analytic hierarchy process, wherein the economic weight accounts for 40%-60%, and the response speed weight accounts for 40%-60%; the normalization processing of the weighted coefficient vector W(t, θ, λ) is realized by the min-max standardization, and satisfies ∑Wn=1 and Wn∈[0, 1]; the experience replay pool capacity of the MO-DQN algorithm is set to 10,000-50,000 samples, the exploration rate decay is performed by using the ε-greedy strategy, the initial exploration rate is 0.9, and the decay factor is 0.995.

[0009] Further, based on the frequency modulation contribution power data, a bidirectional LSTM prediction model fusing transfer learning and attention mechanism is constructed, time-frequency domain correlation features are extracted through a multi-head self-attention layer, time series prediction is performed through a bidirectional LSTM layer, and response margin probability distribution and confidence interval are output, including: Based on the quantized frequency modulation contribution power data, a bidirectional LSTM prediction model fusing transfer learning and attention mechanism is constructed, the model input includes historical contribution power sequence, real-time running state features, environment correlation features and cross-scene transfer features, time-frequency domain correlation features are extracted through a multi-head self-attention layer, time series prediction is performed through a bidirectional LSTM layer, a meta-learner is introduced to dynamically adjust model hyperparameters, and response margin probability distribution and confidence interval of each frequency modulation resource in the next 15 minutes are output; The cross-scene transfer features include wind power cluster output characteristics, photovoltaic consumption rate and energy storage SOC state, and are mapped through a transfer learning network; the number of heads of the multi-head self-attention layer is set to 4-8, the feature dimension of each attention head is 16-32 dimensions, and the number of hidden units of the bidirectional LSTM layer is 64-128; the meta-learner adopts the MAML algorithm framework.

[0010] Further, the predicted response margin probability distribution and confidence interval are transmitted and distributed through a blockchain-edge computing collaborative architecture to realize prediction result encryption transmission and distributed optimization, a real-time feedback channel is established to dynamically adjust model parameters, including: The predicted response margin probability distribution and confidence interval are transmitted to the power grid dispatching system through the blockchain-edge computing collaborative architecture, a lightweight prediction model is deployed on the edge node to realize distributed real-time optimization, a real-time feedback channel is established, and the weighted coefficient vector W(t, θ, λ) and the prediction model parameters are dynamically adjusted according to the dispatching instructions and the edge node calculation results; the blockchain-edge computing collaborative architecture adopts a consortium chain structure, and the consensus mechanism is a practical Byzantine fault tolerance algorithm.

[0011] In a second aspect, the present application provides a system for accurate quantification of frequency modulation contribution and dynamic prediction of response margin, comprising: A data acquisition module is configured to acquire real-time operation data of a power grid, environment-related data and digital twin mirror image data, and perform data preprocessing through Kalman filtering-wavelet threshold joint denoising and isolation forest-generative adversarial network fusion algorithm. A calculation module is configured to calculate frequency modulation contribution data based on the preprocessed data using a multi-objective adaptive weighted integral algorithm model. A prediction output module is configured to construct a bidirectional LSTM prediction model integrating transfer learning and attention mechanism based on the frequency modulation contribution data, extract time-frequency domain correlation features through a multi-head self-attention layer, perform time series prediction through a bidirectional LSTM layer, and output response margin probability distribution and confidence interval. A feedback adjustment module is configured to realize encrypted transmission and distributed optimization of prediction results through a blockchain-edge computing collaborative architecture based on the predicted response margin probability distribution and confidence interval, and establish a real-time feedback channel to dynamically adjust model parameters.

[0012] In a third aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for accurate quantification of frequency modulation contribution and dynamic prediction of response margin.

[0013] In a fourth aspect, the present application provides a computer-readable storage medium storing a computer program, wherein the computer program is executable by a processor to implement the steps of the method for accurate quantification of frequency modulation contribution and dynamic prediction of response margin.

[0014] Compared with the prior art, the present application has the following technical effects: The present application realizes holographic perception of power grid state through multi-source heterogeneous data fusion and digital twin mirror image technology, reduces data noise and improves accuracy of outlier repair; the multi-objective adaptive weighting mechanism of the present application improves the calculation accuracy of frequency modulation contribution, especially reduces the error rate in the scene of severe frequency fluctuation; the bidirectional LSTM model integrating transfer learning and attention mechanism of the present application reduces the prediction error of response margin and improves the coverage rate of confidence interval; the blockchain-edge computing collaborative architecture of the present application realizes real-time encrypted transmission of prediction results, shortens the delay of dynamic adjustment of model parameters, and meets the real-time demand of power grid dispatching. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0016] The application is further illustrated below in conjunction with the drawings: Embodiment 1, please refer to Figure 1 The application provides a precise quantification and dynamic prediction method of one-time frequency modulation contribution power, comprising: Collecting real-time operation data of power grid, environment-related data and digital twin mirror image data, and performing data preprocessing through Kalman filter-wavelet threshold joint denoising and isolated forest-generative adversarial network fusion algorithm; Based on the preprocessed data, a multi-objective adaptive weighted integral algorithm model is used to calculate the frequency modulation contribution power data; Based on the frequency modulation contribution power data, a bidirectional LSTM prediction model integrating transfer learning and attention mechanism is constructed, the time-frequency domain correlation features are extracted through the multi-head self-attention layer, the time series prediction is performed through the bidirectional LSTM layer, and the response margin probability distribution and confidence interval are outputted; The predicted response margin probability distribution and confidence interval realize encrypted transmission and distributed optimization of prediction results through the blockchain-edge computing collaborative architecture, and a real-time feedback channel is established to dynamically adjust the model parameters.

[0017] Specifically: Step 1: Construct a multi-source heterogeneous data acquisition system to collect real-time operation data of power grid, environment-related data and digital twin mirror image data, the real-time operation data including frequency, voltage, current and output data of each generator unit, the environment-related data including meteorological data, load prediction data and equipment health data, and the digital twin mirror image data being mapped to the topological structure and element parameter changes of the power grid in real time through a three-dimensional modeling engine; the collected data is preprocessed through time-space alignment, the measurement noise is eliminated through Kalman filter-wavelet threshold joint denoising algorithm, and the isolated forest-generative adversarial network fusion algorithm is used for abnormal value detection and repair; Step 2: Based on the preprocessed data from Step 1, a multi-objective adaptive weighted integral algorithm model is used to quantify the frequency regulation contribution of each frequency regulation resource at different times. The multi-objective adaptive weighted integral algorithm model is Hi=∫(t0-t1)W(t,θ,λ)×[Pt-P0-Si×Ri×(t-t0)]dt, where Hi is the primary frequency regulation weighted integral power of unit i, t0 is the time when the system frequency exceeds the primary frequency regulation dead zone of unit i, t1 is the time when the system frequency enters the primary frequency regulation dead zone of unit i, Pt is the actual power generation of unit i at time t, P0 is the actual power generation of unit i at time t0, and S... i represents the unit status parameter, where 0 indicates that AGC or planned instructions do not adjust the load, and 1 indicates that load adjustment is in progress. Ri represents the average ramp rate of unit i. W(t,θ,λ) is a weighted coefficient vector based on multi-objective deep reinforcement learning dynamic optimization, where θ is the power grid status feature vector, and λ is the dual-objective optimization coefficient of economy and response speed. Its element Wn=1 / (σn√2π)×exp[-(ln(n)-μ(θ,λ))² / (2σ(θ,λ)²)], where n is the sampling time sequence number, and μ(θ,λ) and σ(θ,λ) are the location and scale parameters updated in real time by the MO-DQN algorithm, and satisfy ∑Wn=1. Step 3: Based on the frequency modulation contribution power data obtained in Step 2, construct a bidirectional LSTM prediction model that integrates transfer learning and attention mechanisms. The model input includes historical contribution power sequence, real-time operating status features, environmental correlation features and cross-scene transfer features. The time-frequency domain correlation features are extracted through a multi-head self-attention layer, and time-series prediction is performed through a bidirectional LSTM layer. A meta-learner is introduced to dynamically adjust the model hyperparameters and output the response margin probability distribution and confidence interval of each frequency modulation resource in the next 15 minutes. Step 4: The response margin probability distribution and confidence interval predicted in Step 3 are encrypted and transmitted to the power grid dispatching system through a blockchain-edge computing collaborative architecture. A lightweight prediction model is deployed at the edge node to achieve distributed real-time optimization, and a real-time feedback channel is established. The weighted coefficient vector W(t,θ,λ) in Step 2 and the prediction model parameters in Step 3 are dynamically adjusted according to the dispatching instructions and the calculation results of the edge node.

[0018] This invention employs a multi-objective adaptive weighted integral algorithm model to quantify the electricity contribution of frequency regulation; utilizes a bidirectional LSTM model integrating transfer learning and attention mechanisms to predict the probability distribution and confidence interval of response margin; and achieves distributed optimization through a blockchain-edge computing collaborative architecture. This invention solves the problems of single data dimension, static and fixed quantification models, and insufficient prediction accuracy in existing technologies, improving the refinement level and cross-scenario adaptability of frequency regulation resource management. It can be widely applied to complex power grid frequency regulation scenarios with a high proportion of renewable energy grid connection.

[0019] Embodiment 2 provides a precise quantification of frequency modulation contribution power and a dynamic prediction method of response margin, comprising: Step 1: Construct a multi-source heterogeneous data acquisition system to collect real-time operation data, environment-related data and digital twin mirror image data of the power grid, wherein the real-time operation data includes frequency, voltage, current and output data of each generator set, the environment-related data includes meteorological data, load prediction data and equipment health data, and the digital twin mirror image data is mapped by a three-dimensional modeling engine in real time to reflect the topological structure and element parameter changes of the power grid; the collected data is preprocessed by time-space alignment, the measurement noise is eliminated by Kalman filtering-wavelet threshold joint denoising algorithm, and the abnormal value detection and repair are performed by isolated forest-generative adversarial network fusion algorithm; Step 2: Based on the data preprocessed in step 1, the frequency modulation contribution power of each frequency modulation resource at different time is quantitatively calculated by using a multi-objective adaptive weighted integral algorithm model, wherein the multi-objective adaptive weighted integral algorithm model is Hi=∫(t0-t1)W(t,θ,λ)×[Pt-P0-Si×Ri×(t-t0)]dt, wherein Hi is the weighted integral power of the unit i, t0 is the time when the system frequency exceeds the dead zone of the primary frequency modulation action of the unit i, t1 is the time when the system frequency enters the dead zone of the primary frequency modulation action of the unit i, Pt is the actual power of the unit i at time t, P0 is the actual power of the unit i at time t0, Si is the state parameter of the unit, 0 indicates that the AGC or the planned instruction does not adjust the load, 1 indicates that the load is being adjusted, Ri is the average climbing rate of the unit i; W(t,θ,λ) is a weighted coefficient vector based on the dynamic optimization of multi-objective deep reinforcement learning, θ is a power grid state feature vector, λ is an economic-response speed double target optimization coefficient, its element Wn=1 / (σn√2π)×exp[-(ln(n)-μ(θ,λ))² / (2σ(θ,λ)²)], wherein n is the sampling time sequence number, μ(θ,λ) and σ(θ,λ) are the position parameters and scale parameters updated in real time by the MO-DQN algorithm, and satisfy ∑Wn=1; Step 3: Based on the frequency modulation contribution power data quantified in step 2, a bidirectional LSTM prediction model integrating transfer learning and attention mechanism is constructed, wherein the model input includes historical contribution power sequence, real-time operation state feature, environment-related feature and cross-scene transfer feature, the time-frequency domain correlation feature is extracted through the multi-head self-attention layer, the time series prediction is performed through the bidirectional LSTM layer, the meta-learner is introduced to dynamically adjust the model hyperparameters, and the response margin probability distribution and confidence interval of each frequency modulation resource in the next 15 minutes are output. Step 4: The response margin probability distribution and confidence interval predicted in step 3 are encrypted and transmitted to the power grid dispatching system through the blockchain-edge computing collaborative architecture, a lightweight prediction model is deployed in the edge node to realize distributed real-time optimization, a real-time feedback channel is established, and the weighted coefficient vector W(t, theta, lambda) in step 2 and the prediction model parameters in step 3 are dynamically adjusted according to the dispatching instructions and the edge node calculation results.

[0020] In the embodiment of the application, taking a provincial power grid frequency modulation system as an application scenario, the specific implementation manner is as follows: Step 1: deploying a multi-source heterogeneous data acquisition system based on 5G+ edge computing, including 200 PMU devices (sampling rate 50 Hz), 50 weather stations (collecting temperature / wind speed / illumination intensity), 100 generator SCADA interfaces and digital twin platforms, processing measurement noise through a Kalman filter-wavelet threshold joint denoising algorithm, wherein the Kalman filter process noise covariance Q=diag([0.01, 0.01, 0.01]), and the observation noise covariance R=0.1; the wavelet decomposition uses db4 wavelet basis for 3-level decomposition, and the high-frequency coefficients are processed by a soft threshold function.

[0021] The working principle of the application is: through the multi-source heterogeneous data acquisition system, the state of the power grid is holographically perceived, and after preprocessing, the multi-objective adaptive weighted integral model is input, the MO-DQN algorithm dynamically optimizes the weighted coefficient vector W(t, theta, lambda) according to the state of the power grid theta and the double-objective coefficient lambda, and the frequency modulation contribution electric quantity is quantized; the bidirectional LSTM model fused with transfer learning uses historical data and cross-scene features to predict the response margin, realizes distributed optimization through the blockchain-edge computing architecture, feeds back the quantization model and the prediction model parameters, and forms a "collection-quantization-prediction-optimization" closed-loop control.

[0022] In another embodiment of the application, a system for precise quantization of frequency modulation contribution electric quantity and dynamic prediction of response margin is provided, which can be used to realize the above-mentioned method for precise quantization of primary frequency modulation contribution electric quantity and dynamic prediction of response margin, and specifically, the system comprises: A data acquisition module is configured to acquire real-time operation data of a power grid, environment-related data and digital twin mirror image data, and perform data preprocessing through a Kalman filter-wavelet threshold joint denoising algorithm and an isolated forest-generative adversarial network fusion algorithm; A calculation module is configured to calculate frequency modulation contribution electric quantity data based on the preprocessed data by using a multi-objective adaptive weighted integral algorithm model; A prediction output module is configured to construct a bidirectional LSTM prediction model fused with transfer learning and an attention mechanism based on the frequency modulation contribution electric quantity data, extract time-frequency domain correlation features through a multi-head self-attention layer, perform time series prediction through a bidirectional LSTM layer, and output a response margin probability distribution and a confidence interval; The feedback adjustment module is used for predicting the response margin probability distribution and the confidence interval to realize encrypted transmission and distributed optimization of the prediction result through the blockchain-edge computing collaborative architecture, and establish a real-time feedback channel to dynamically adjust the model parameters.

[0023] The division of the modules in the embodiments of the present application is illustrative, and is merely a logical functional division.

[0024] In another embodiment of the present application, a computer device is provided, which comprises a processor and a memory, the memory is used to store a computer program, the computer program comprises program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions in the computer storage medium to implement a corresponding method process or a corresponding function; the processor in the embodiments of the present application can be used for the operation of the accurate quantization of the frequency contribution electric quantity and the dynamic prediction method of the response margin.

[0025] In still another embodiment of the present application, the present application also provides a storage medium, specifically a computer readable storage medium (Memory), which is a memory device in a computer device, used for storing programs and data. It can be understood that the computer readable storage medium here can include the built-in storage medium in the computer device, and of course can also include the extended storage medium supported by the computer device. The computer readable storage medium provides a storage space, which stores the operating system of the terminal. Moreover, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium here can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory. One or more instructions stored in the computer readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the above-mentioned embodiment of the method for accurate quantification of frequency contribution electric quantity and dynamic prediction of response margin.

[0026] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0027] The present application is described with reference to the flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device implemented in the flowcharts and / or block diagrams. Figure 1 The function of one or more flows and / or blocks Figure 1 The device for implementing the function specified in one or more flows and / or blocks.

[0028] These computer program instructions can also be stored in a computer readable memory capable of guiding the computer or other programmable data processing devices to work in a specific way, so that the instructions stored in the computer readable memory produce a manufactured product including instruction devices, which implement the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0029] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operation steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, so that the instructions executed on the computer or other programmable data processing devices provide processes for implementing the flow Figure 1 one or more processes and / or blocks Figure 1 the steps of the function specified in the one or more blocks.

[0030] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit it, although the present application has been described in detail with reference to the above examples, those skilled in the art should understand: the specific embodiments of the present application can still be modified or replaced by the equivalent, without departing from the spirit and scope of the present application, any modification or equivalent replacement, which should be covered within the protection scope of the claims of the present application.

Claims

1. A method for accurate quantification of primary frequency regulation contribution and dynamic prediction of response margin, characterized in that, include: Real-time power grid operation data, environmental correlation data, and digital twin mirror data are collected, and data preprocessing is performed using a Kalman filter-wavelet thresholding joint denoising and an isolated forest-generative adversarial network fusion algorithm. Based on the preprocessed data, a multi-objective adaptive weighted integral algorithm model is used to calculate the frequency regulation contribution power data; Based on frequency modulation contribution power data, a bidirectional LSTM prediction model integrating transfer learning and attention mechanism is constructed. The time-frequency domain correlation features are extracted through a multi-head self-attention layer, and time-series prediction is performed through a bidirectional LSTM layer to output the response margin probability distribution and confidence interval. The predicted response margin probability distribution and confidence interval are transmitted and optimized in a distributed manner through a blockchain-edge computing collaborative architecture, and a real-time feedback channel is established to dynamically adjust the model parameters.

2. The method for accurate quantification of primary frequency regulation contribution and dynamic prediction of response margin according to claim 1, characterized in that, The collected real-time power grid operation data, environmental correlation data, and digital twin mirror data are preprocessed using a Kalman filter-wavelet thresholding joint denoising algorithm and an isolated forest-generative adversarial network fusion algorithm, including: A multi-source heterogeneous data acquisition system is constructed to collect real-time power grid operation data, environmental correlation data, and digital twin mirror data. The real-time operation data includes frequency, voltage, current, and output data of each generator unit. The environmental correlation data includes meteorological data, load forecast data, and equipment health data. The digital twin mirror data is used to map the power grid topology and component parameter changes in real time through a 3D modeling engine. The collected data undergoes spatiotemporal alignment preprocessing, and measurement noise is eliminated using a Kalman filter-wavelet threshold joint denoising algorithm. An outlier detection and repair are performed using an isolated forest-generative adversarial network fusion algorithm.

3. The method for accurate quantification of primary frequency regulation contribution and dynamic prediction of response margin according to claim 2, characterized in that, The digital twin mirror data acquired by the multi-source heterogeneous data acquisition system is mapped in real time through a 3D modeling engine to reflect changes in the power grid topology and component parameters, with a sampling frequency of no less than 50Hz. The specific implementation of the Kalman filter-wavelet threshold joint denoising algorithm is as follows: first, state estimation is performed by Kalman filtering, then the residual signal is decomposed into three levels using the db4 wavelet basis, and the signal is reconstructed after processing the high-frequency coefficients by the heuristic threshold function. In the isolated forest-generative adversarial network fusion algorithm, the isolated forest is used to initially detect outliers, the generator of the generative adversarial network is used to repair outlier data, and the discriminator simultaneously evaluates the authenticity of the original data and the repaired data.

4. The method for accurate quantification of primary frequency regulation contribution and dynamic prediction of response margin according to claim 1, characterized in that, The calculation of frequency regulation contribution power data based on the preprocessed data using a multi-objective adaptive weighted integral algorithm model includes: The multi-objective adaptive weighted integral algorithm model is Hi=∫(t0-t1)W(t,θ,λ)×[Pt-P0-Si×Ri×(t-t0)]dt, where Hi is the primary frequency regulation weighted integral power of unit i, t0 is the time when the system frequency exceeds the primary frequency regulation dead zone of unit i, t1 is the time when the system frequency enters the primary frequency regulation dead zone of unit i, Pt is the actual power generation of unit i at time t, P0 is the actual power generation of unit i at time t0, Si is the unit state parameter, 0 indicates that AGC or planned command does not adjust the load, 1 indicates that load adjustment is in progress, Ri is the average ramp rate of unit i; W(t,θ,λ) is the weighted coefficient vector based on multi-objective deep reinforcement learning dynamic optimization, θ is the grid state feature vector, λ is the dual-objective optimization coefficient of economy and response speed, and its element Wn=1 / (σn√2π)×exp[-(ln(n)-μ(θ,λ))]. 2 / (2σ(θ,λ) 2 )], where n is the sampling time sequence number, μ(θ,λ) and σ(θ,λ) are the position parameters and scale parameters updated in real time by the MO-DQN algorithm, and satisfy ∑Wn=1.

5. The method for accurate quantification of primary frequency regulation contribution and dynamic prediction of response margin according to claim 4, characterized in that, The bi-objective optimization coefficient λ of the multi-objective adaptive weighted integral algorithm model is determined by the analytic hierarchy process, where the economic weight accounts for 40%-60% and the response speed weight accounts for 40%-60%; the normalization of the weighted coefficient vector W(t,θ,λ) is achieved by min-max standardization, satisfying ∑Wn=1 and Wn∈[0,1]; the experience replay pool capacity of the MO-DQN algorithm is set to 10000-50000 samples, and the exploration rate decay is performed using the ε-greedy strategy, with an initial exploration rate of 0.9 and a decay factor of 0.

995.

6. The method for accurate quantification of primary frequency regulation contribution and dynamic prediction of response margin according to claim 1, characterized in that, The bidirectional LSTM prediction model, based on frequency modulation contribution power data, is constructed by integrating transfer learning and attention mechanisms. It extracts time-frequency domain correlation features through a multi-head self-attention layer, performs time-series prediction via a bidirectional LSTM layer, and outputs the response margin probability distribution and confidence interval, including: Based on the frequency modulation contribution power data obtained by quantization, a bidirectional LSTM prediction model integrating transfer learning and attention mechanism is constructed. The model input includes historical contribution power sequence, real-time operating status features, environmental correlation features and cross-scene transfer features. Time-frequency domain correlation features are extracted through a multi-head self-attention layer, and time-series prediction is performed through a bidirectional LSTM layer. A meta-learner is introduced to dynamically adjust the model hyperparameters and output the response margin probability distribution and confidence interval of each frequency modulation resource in the next 15 minutes. The cross-scenario transfer features include wind power cluster output characteristics, photovoltaic absorption rate, and energy storage SOC status, which are mapped through a transfer learning network; the number of heads in the multi-head self-attention layer is set to 4-8, the feature dimension of each attention head is 16-32, and the number of hidden units in the bidirectional LSTM layer is 64-128; the meta-learner adopts the MAML algorithm framework.

7. The method for accurate quantification of primary frequency regulation contribution and dynamic prediction of response margin according to claim 1, characterized in that, The predicted response margin probability distribution and confidence interval are encrypted and distributed through a blockchain-edge computing collaborative architecture to achieve the prediction results. A real-time feedback channel is established to dynamically adjust the model parameters, including: The predicted response margin probability distribution and confidence interval are encrypted and transmitted to the power grid dispatching system through a blockchain-edge computing collaborative architecture. A lightweight prediction model is deployed at the edge nodes to achieve distributed real-time optimization and establish a real-time feedback channel. The weighted coefficient vector W(t,θ,λ) and prediction model parameters are dynamically adjusted according to the dispatching instructions and the calculation results of the edge nodes. The blockchain-edge computing collaborative architecture adopts a consortium blockchain structure and the consensus mechanism is a practical Byzantine fault-tolerant algorithm.

8. A system for precise quantification of frequency modulation contribution power and dynamic prediction of response margin, characterized in that, include: The data acquisition module is used to collect real-time power grid operation data, environmental correlation data, and digital twin mirror data. Data preprocessing is performed using a Kalman filter-wavelet thresholding joint denoising algorithm and an isolated forest-generative adversarial network fusion algorithm. The calculation module is used to calculate the frequency regulation contribution power data based on the preprocessed data using a multi-objective adaptive weighted integral algorithm model. The prediction output module is used to construct a bidirectional LSTM prediction model that integrates transfer learning and attention mechanisms based on frequency modulation contribution power data. It extracts time-frequency domain correlation features through a multi-head self-attention layer, performs time-series prediction through a bidirectional LSTM layer, and outputs the response margin probability distribution and confidence interval. The feedback adjustment module is used to predict the probability distribution of the response margin and the confidence interval. Through the blockchain-edge computing collaborative architecture, the prediction results are encrypted and distributed for transmission and optimization. A real-time feedback channel is established to dynamically adjust the model parameters.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for accurate quantification of primary frequency modulation contribution power and dynamic prediction of response margin as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for accurately quantifying the primary frequency modulation contribution power and dynamically predicting the response margin as described in any one of claims 1 to 7.

Images