Run length domain multi-feature constrained power signal analysis system

Abstract

The present application relates to the technical field of new energy grid-connected electric energy metering and complex dynamic signal run domain analysis, and provides a run domain multi-feature constraint electric energy signal analysis system, which comprises an acquisition module for obtaining original data flow, a pretreatment module for applying Heine-Borel theorem to divide signal sub-intervals and combining wavelet packet denoising to generate clean signals, a modal decomposition module for separating quasi-steady-state items and dynamic items by using an improved empirical mode decomposition algorithm, a feature extraction module for generating amplitude parameter sequences through non-uniform down-sampling and fuzzy logic processing, a collaborative quantization module for extracting run feature parameters by using a hidden Markov model, a fusion modeling module for realizing multi-constraint feature classification by using a support vector machine, and an error compensation module for performing dynamic error correction by using a reinforcement learning strategy. The modules are cooperatively optimized through a multi-layer closed-loop feedback network, and the metering error problem caused by rapid fluctuation and multi-run mode change of dynamic electric energy signals in the new energy grid-connected scene is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of new energy grid-connected power metering and run-domain analysis of complex dynamic signals, and particularly to a run-domain multi-feature constrained power signal analysis system. Background Technology

[0002] Existing renewable energy grid-connected power metering technologies suffer from the following technical challenges: In application scenarios such as wind power, photovoltaic power, and high-speed rail traction, complex dynamic power signals exhibit rapid random fluctuations and multi-run mode changes due to the intermittency of renewable energy generation and the sudden changes in load. For example, the current amplitude of wind power substations generates short-term run impacts under wind speed changes, the light fluctuations of photovoltaic substations trigger long-term cyclic modes, and the load start-stop of high-speed rail traction stations causes wide-range changes in run length distribution. These characteristics make it difficult to quantify the implicit run length, modulation depth, and impact intensity of the signal using existing uniform sampling methods, resulting in a significant increase in the dynamic error of power metering devices and affecting metering fairness and grid stability. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a run-domain multi-feature constrained power signal analysis system, which solves the technical problem of increased dynamic error in power metering devices caused by the rapid random fluctuations, multi-run mode changes, and difficulty in quantifying implicit features of complex dynamic power signals.

[0004] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows: The run-domain multi-feature constrained power signal analysis system provided by this invention includes: The acquisition module is used to perform signal acquisition, dynamically adjust the sampling frequency using a recursive least squares algorithm, capture voltage and current signals, and generate a raw data stream with timestamps. The preprocessing module is used to receive the raw data stream, perform collaborative preprocessing, apply the Heine-Borel finite coverage theorem to divide the signal time axis into sub-intervals, and apply wavelet packet denoising to filter out high-frequency noise within the sub-intervals to generate a clean signal; it is also used to send the noise statistical features extracted during the preprocessing process back to the acquisition module. The mode decomposition module is used to receive the clean signal, perform bidirectional mode decomposition, and use an improved empirical mode decomposition algorithm to separate the quasi-steady-state term and dynamic term as bimodal components; it is also used to calculate the amplitude domain energy difference, loss degree and orthogonality index during the decomposition process, and send the index back to the preprocessing module; The feature extraction module is used to receive the dual-modal components, perform integrated feature extraction, calculate the amplitude 1-norm in the preprocessed sub-interval, and generate a downsampled amplitude parameter sequence after non-uniform downsampling and fuzzy logic defuzzification. The collaborative quantization module is used to receive the amplitude parameter sequence, perform run-domain collaborative quantization, generate a run-length symbol sequence through N-ary interval discretization, and extract the run-length, modulation depth and impulse intensity feature parameters by applying a hidden Markov model and encapsulating them into feature vectors. The fusion modeling module is used to receive the feature vector, perform multi-constraint fusion modeling, apply internal constraints in the run-length domain and external constraints in the amplitude domain, perform dynamic threshold classification of the feature parameters through a support vector machine, and output constrained feature parameters; it is also used to send the model parameters generated by modeling back to the collaborative quantization module. The error compensation module is used to receive the constrained feature parameters, perform dynamic error collaborative compensation, drive the adjustment of the power metering algorithm parameters through reinforcement learning strategy, and output the corrected power value; it is also used to send the residual data generated during the compensation process back to the acquisition module.

[0005] Furthermore, in the run-domain multi-feature constrained power signal analysis system of the present invention, the acquisition module is used to: analyze the spectral characteristics of the captured voltage and current signals using a recursive least squares algorithm, dynamically calculate and implement the adjustment amount of the sampling frequency; when signal fluctuations are detected to be aggravated, increase the sampling rate according to the adjustment amount; evaluate the stability of the signal using a sliding window mechanism, add timestamps and quality identifiers to the acquired signal data, and generate a raw data stream with timestamps.

[0006] Furthermore, the run-domain multi-feature constrained power signal analysis system of the present invention is characterized in that the preprocessing module is used to: apply the Heine-Borel finite coverage theorem to divide the signal time axis of the original data stream into a finite number of sub-intervals; within each sub-interval, apply wavelet packet denoising and eliminate high-frequency noise through a soft threshold function; wherein the denoising threshold is dynamically calibrated according to the signal-to-noise ratio of the sub-interval; output a clean signal, and extract the statistical features of the high-frequency noise as noise statistical features for transmission.

[0007] Furthermore, in the run-domain multi-feature constrained power signal analysis system of the present invention, the mode decomposition module is used to: process the clean signal using an improved empirical mode decomposition algorithm; predict the signal trend component through a long short-term memory network before decomposition, and initialize the decomposition boundary conditions with the trend component; calculate the amplitude domain energy difference, loss degree, and orthogonality index during the decomposition iteration process; input the amplitude domain energy difference, loss degree, and orthogonality index into a proportional-integral controller to dynamically adjust the number of decomposition iterations; and output the separated quasi-steady-state term and dynamic term as dual-mode components.

[0008] Furthermore, in the run-domain multi-feature constrained power signal analysis system of the present invention, the feature extraction module is used to: calculate the 1-norm of the amplitude sequence of the dual-modal components within a sub-interval defined by collaborative preprocessing; generate an amplitude parameter sequence by applying 1-norm non-uniform downsampling; dynamically search for downsampling frequency combinations to update the downsampling process using a particle swarm optimization algorithm; and perform membership mapping and defuzzification processing on the amplitude values ​​output by the updated downsampling process through a fuzzy logic system to generate a downsampled amplitude parameter sequence.

[0009] Furthermore, in the run-domain multi-feature constrained power signal analysis system of the present invention, the collaborative quantization module is used to: set an N-ary quantization interval, discretize the continuous amplitude parameter sequence into a run-symbol sequence; apply a hidden Markov model to analyze the run-symbol sequence, train model parameters through the Baum-Welch algorithm, and extract run length, modulation depth, and impulse intensity feature parameters; perform consistency verification between the extracted run length, modulation depth, and impulse intensity feature parameters and the independently calculated run length probability density distribution and autocorrelation function analysis results; when the verification deviation exceeds a threshold, trigger recalibration of the quantization interval boundary; and encapsulate the verified feature parameters into a feature vector for output.

[0010] Furthermore, in the run-domain multi-feature constraint power signal analysis system of the present invention, the fusion modeling module is used to: define the run length range as an internal constraint condition and define the amplitude variance range as an external constraint condition; use a support vector machine to perform dynamic threshold classification on the feature parameters in the feature vector; introduce a federated learning framework to aggregate feature data from multiple scenarios and update the support vector machine model parameters through a distributed gradient descent algorithm; output the feature parameters that satisfy the internal and external constraint conditions as constrained feature parameters, and send the updated model parameters.

[0011] Furthermore, in the run-domain multi-feature constrained power signal analysis system of the present invention, the error compensation module is used to: select gain correction or phase offset adjustment actions to drive the power metering algorithm parameter adjustment by using a reinforcement learning strategy, taking the constrained feature parameters and historical error data as input; update the action value function using a Q-learning algorithm; verify the metering results using a long m-sequence dynamic power reference test signal, and adjust the reinforcement learning strategy according to the verification results; output the corrected power value, and send the residual data generated by the compensation.

[0012] Furthermore, in the run-domain multi-feature constrained power signal analysis system of the present invention, the preprocessing module is further used to send the extracted noise statistical features back to the acquisition module; the mode decomposition module is further used to send the calculated amplitude domain energy difference, loss degree, and orthogonality index back to the preprocessing module; the feature extraction module is further used to send the optimal downsampling frequency parameters obtained by the particle swarm optimization algorithm back to the mode decomposition module; the cooperative quantization module is further used to send the state transition probability matrix obtained by the hidden Markov model analysis back to the feature extraction module; the fusion modeling module is further used to send the updated model parameters back to the cooperative quantization module; and the error compensation module is further used to send the residual data generated by the compensation back to the acquisition module.

[0013] Furthermore, in the run-domain multi-feature constrained power signal analysis system of the present invention, the preprocessing module, mode decomposition module, feature extraction module, collaborative quantization module, fusion modeling module, and error compensation module, while outputting the clean signal, dual-mode components, amplitude parameter sequence, feature vector, constrained feature parameters, and corrected power value, respectively send the generated noise statistical characteristics, amplitude domain energy difference, loss degree and orthogonality index, optimal downsampling frequency parameter, state transition probability matrix, updated model parameters, and residual data back to the acquisition module, preprocessing module, mode decomposition module, feature extraction module, collaborative quantization module, and acquisition module, forming a multi-layer closed-loop feedback.

[0014] The beneficial effects of this invention are: This invention effectively reduces spectral leakage by dynamically adjusting the sampling frequency using a recursive least squares algorithm; it significantly improves high-frequency noise suppression by employing the Heine-Borel finite coverage theorem to divide signal sub-intervals and combining it with wavelet packet denoising; it accurately separates quasi-steady-state and dynamic terms through an improved empirical mode decomposition algorithm and a long short-term memory network; it optimizes feature extraction accuracy through non-uniform downsampling and fuzzy logic processing; it enhances the reliability of feature quantization by constructing a multi-constraint fusion model using a hidden Markov model and a support vector machine; it achieves adaptive adjustment of metering parameters through an error compensation mechanism driven by a reinforcement learning strategy; and it enables the system's modules to form a collaborative optimization whole, ultimately significantly reducing dynamic power signal metering errors and improving grid operation stability in new energy grid-connected scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0016] Figure 1This is a system architecture diagram of a run-domain multi-feature constrained power signal analysis system. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0018] To better understand the purpose of this invention, the invention will now be described in further detail.

[0019] Please see Figure 1 The run-domain multi-feature constrained power signal analysis system provided by the present invention includes: The acquisition module is used to perform signal acquisition, dynamically adjust the sampling frequency using a recursive least squares algorithm, capture voltage and current signals, and generate a raw data stream with timestamps. The preprocessing module is used to receive the raw data stream, perform collaborative preprocessing, apply the Heine-Borel finite coverage theorem to divide the signal time axis into sub-intervals, and apply wavelet packet denoising to filter out high-frequency noise within the sub-intervals to generate a clean signal; it is also used to send the noise statistical features extracted during the preprocessing process back to the acquisition module. The mode decomposition module is used to receive the clean signal, perform bidirectional mode decomposition, and use an improved empirical mode decomposition algorithm to separate the quasi-steady-state term and dynamic term as bimodal components; it is also used to calculate the amplitude domain energy difference, loss degree and orthogonality index during the decomposition process, and send the index back to the preprocessing module; The feature extraction module is used to receive the dual-modal components, perform integrated feature extraction, calculate the amplitude 1-norm in the preprocessed sub-interval, and generate a downsampled amplitude parameter sequence after non-uniform downsampling and fuzzy logic defuzzification. The collaborative quantization module is used to receive the amplitude parameter sequence, perform run-domain collaborative quantization, generate a run-length symbol sequence through N-ary interval discretization, and extract the run-length, modulation depth and impulse intensity feature parameters by applying a hidden Markov model and encapsulating them into feature vectors. The fusion modeling module is used to receive the feature vector, perform multi-constraint fusion modeling, apply internal constraints in the run-length domain and external constraints in the amplitude domain, perform dynamic threshold classification of the feature parameters through a support vector machine, and output constrained feature parameters; it is also used to send the model parameters generated by modeling back to the collaborative quantization module. The error compensation module is used to receive the constrained feature parameters, perform dynamic error collaborative compensation, drive the adjustment of the power metering algorithm parameters through reinforcement learning strategy, and output the corrected power value; it is also used to send the residual data generated during the compensation process back to the acquisition module.

[0020] The run-domain multi-feature constrained power signal analysis system achieves accurate measurement of dynamic power signals through the collaborative work of multiple modules. The system begins with the acquisition module, which uses a recursive least squares algorithm to perform real-time spectral analysis on voltage and current signals, dynamically calculating the optimal sampling frequency. When increased signal fluctuations are detected, the acquisition module automatically increases the sampling rate to avoid spectral aliasing. Simultaneously, it evaluates signal stability through a sliding window mechanism, adding timestamps and quality identifiers to the raw data to form a standardized raw data stream.

[0021] After the raw data stream is transmitted to the preprocessing module, the Heine-Borel finite coverage theorem is first applied to divide the signal time axis into several sub-intervals, with the length of each sub-interval synchronized with the signal's fundamental period. Within each sub-interval, wavelet packet denoising is used for multi-scale decomposition, and a soft thresholding function is employed to eliminate high-frequency noise and transient interference. The denoising threshold is dynamically calibrated based on the real-time signal-to-noise ratio of each sub-interval to ensure the adaptability of the denoising effect. The preprocessed clean signal includes signal-to-noise ratio and distortion indices, and the extracted noise statistical features are fed back to the acquisition module to optimize the sampling strategy.

[0022] After the clean signal enters the mode decomposition module, the signal trend component is first predicted through a long short-term memory network, which serves as the boundary condition for improving the empirical mode decomposition algorithm. During the decomposition iteration process, the amplitude domain energy difference, loss degree, and orthogonality index are calculated in real time. Each index is dynamically adjusted by the controller to control the number of decomposition iterations, effectively suppressing mode aliasing. The quasi-steady-state term and dynamic term generated by the decomposition constitute a bimodal component. The calculated index parameters are fed back to the preprocessing module for optimizing sub-interval partitioning and denoising threshold setting.

[0023] After receiving the dual-modal components, the feature extraction module calculates the 1-norm of the amplitude sequence within the pre-processed sub-intervals and generates an initial amplitude parameter sequence using a 1-norm non-uniform downsampling method. The particle swarm optimization algorithm dynamically searches for the optimal combination of downsampling frequencies to minimize the reconstruction error. The optimized parameters are fed back to the mode decomposition module to improve boundary handling. The amplitude parameters undergo membership mapping and defuzzification processing via a fuzzy logic system, ultimately forming the downsampled amplitude parameter sequence.

[0024] The collaborative quantization module discretizes the amplitude parameter sequence into N-ary intervals, generating a run-length symbol sequence. The Hidden Markov Model (HMM) is trained using the Baum-Welch algorithm to extract run length, modulation depth, and impulse intensity feature parameters from the run-length symbol sequence. The extracted feature parameters are then compared with independently calculated run length probability density distributions and autocorrelation function analysis results for consistency verification. If the deviation exceeds the limit, a quantization interval boundary recalibration mechanism is triggered. The verified feature parameters are encapsulated as feature vectors, and the state transition probability matrix is ​​fed back to the feature extraction module to optimize the downsampling strategy.

[0025] The fusion modeling module first defines the run length range as an internal constraint and the amplitude variance range as an external constraint. The support vector machine dynamically selects a kernel function based on the separability of the feature space and performs dynamic threshold classification on the feature parameters. The federated learning framework aggregates feature data from multiple scenarios such as wind power and photovoltaics, and updates the global model parameters using a distributed gradient descent algorithm. The constrained feature parameters are output to the error compensation module, while the updated model parameters are fed back to the collaborative quantization module for optimizing feature selection.

[0026] The error compensation module employs a reinforcement learning strategy, using constrained feature parameters and historical error data as input, and updates the action value function through a Q-learning algorithm. The agent selects gain correction or phase shift adjustment actions to drive the optimization of the energy metering algorithm parameters. A long m-sequence dynamic energy reference test signal is used periodically to verify the validity of the metering results; the verification results are used to adjust the reinforcement learning strategy. The compensated corrected energy value serves as the final output of the system, while the compensation residual is fed back to the acquisition module, forming a closed-loop control system.

[0027] The acquisition module dynamically adjusts the sampling frequency using a recursive least squares algorithm. This algorithm continuously analyzes the spectral characteristics of voltage and current signals, updating filter weights by minimizing the trace of the prediction error covariance matrix. When increased signal fluctuations are detected, the module increases the sampling rate in real time based on the spectral analysis results; the adjustment amount is calculated recursively. A sliding window mechanism extracts signal segments of fixed time length, calculating the variance and peak value within each segment as stationarity evaluation metrics. The acquisition module appends a millisecond-level timestamp to each data point and generates an identifier including a signal quality index, ultimately forming a raw data stream with time-domain markings.

[0028] The preprocessing module applies the Heine-Borel finite coverage theorem to process the raw data stream. This theorem guarantees the existence of finite open coverage within a finite time interval, and the module adaptively determines the length of the coverage sub-interval based on the signal fundamental period. Within each sub-interval, wavelet packet denoising employs a 3-level decomposition using the Db4 wavelet basis function, and a soft thresholding function is used to process the detail coefficients. The denoising threshold is dynamically calculated based on the signal-to-noise ratio (SNR) of the sub-interval, which is derived from the ratio of the energy of the approximate coefficients after decomposition to the energy of the detail coefficients. The clean signal output after preprocessing includes the denoised time-domain waveform, and the extracted noise statistical features include noise power spectral density and amplitude distribution parameters.

[0029] The mode decomposition module employs an improved empirical mode decomposition algorithm to process clean signals. A long short-term memory (LSTM) network predicts the signal trend components through forget gates, input gates, and output gates. The network input is a signal sequence of the past 128 sampling points. The improved ESM algorithm initializes the envelope calculation using the predicted trend components and calculates the amplitude domain energy difference, loss degree, and orthogonality index of the intrinsic mode functions after each sieving iteration. A proportional-integral (PI) controller adjusts the number of sieving iterations based on the index deviation, with the proportional coefficient and integral time constant preset according to the signal characteristics. After decomposition, the output is a bimodal component consisting of a quasi-steady-state term and a dynamic term, while simultaneously feeding back index data including instantaneous frequency and modal energy distribution information.

[0030] The feature extraction module processes the dual-modal components within the pre-defined sub-intervals. First, the module calculates the 1-norm of the amplitude sequence, i.e., the sum of the absolute values ​​of the sequence, as the fundamental feature quantity. 1-norm non-uniform downsampling determines the sampling point location based on the gradient change of the feature quantity. A particle swarm optimization algorithm searches for the optimal downsampling frequency combination with a population size of 20 and 50 iterations. The fuzzy logic system uses a triangular membership function to map the amplitude values ​​to three fuzzy sets: "low," "medium," and "high." The centroid method is used to defuzzify the values ​​and obtain clear numerical values. The final generated downsampling amplitude parameter sequence retains the key feature points of the original signal while providing optimal sampling parameters as feedback to the upstream module.

[0031] When processing amplitude parameter sequences, the collaborative quantization module first establishes an N-ary quantization interval to discretize the continuous signal. The boundary values ​​of the quantization interval are dynamically set according to the statistical distribution characteristics of the signal amplitude, converting the continuous amplitude parameter sequence into a run-length symbol sequence composed of discrete symbols. The Hidden Markov Model uses a forward-backward algorithm to calculate the probability of the observation sequence, and iteratively updates the state transition matrix and the observation probability matrix using the Baum-Welch algorithm. During model training, features such as run-length (characterizing the length of time the signal remains within a specific amplitude range), modulation depth (reflecting the drasticness of signal amplitude changes), and impulse intensity (describing the energy magnitude of the signal's abrupt change components) are extracted.

[0032] The fusion modeling module defines a run length range as an internal constraint, determined based on historical data statistical characteristics; it also defines an amplitude variance range as an external constraint, with its threshold set based on the signal's steady-state characteristics. The support vector machine uses a radial basis function kernel to map features to a high-dimensional space and utilizes the margin maximization principle to find the optimal classification hyperplane. The federated learning framework integrates feature data from multiple renewable energy power plants through a secure aggregation algorithm. Each local model is trained using stochastic gradient descent, and the server updates the global model parameters using a weighted average. Feature parameters that satisfy the constraints are marked as valid features, and the model update information includes feature weights and classification threshold parameters.

[0033] The error compensation module constructs a reinforcement learning environment. The agent's observation state space includes constrained feature parameters, historical error sequences, and the operating status of the metering equipment. The Q-learning algorithm employs an ε-greedy strategy to explore the action space, with the learning rate decaying over time to ensure convergence. The long m-sequence dynamic energy reference test signal possesses ideal autocorrelation characteristics; the error index is calculated by comparing the actual metering results with the standard values ​​of the reference signal. The policy network parameters are updated using gradient descent based on the validation results, and the compensation actions include adjustments to the gain correction coefficient and the phase shift angle.

[0034] In the system feedback network, the preprocessing module sends noise statistical characteristics (including noise power spectrum and time-domain characteristics) to the acquisition module; the mode decomposition module returns the amplitude-domain energy difference, loss degree, and orthogonality index to the preprocessing module; the feature extraction module feeds back the optimal downsampling frequency parameters to the mode decomposition module; the collaborative quantization module sends the state transition probability matrix to the feature extraction module; the fusion modeling module returns the updated model parameters to the collaborative quantization module; and the error compensation module sends the compensation residual data to the acquisition module. All feedback data constitute a closed-loop optimization circuit.

[0035] The multi-layer closed-loop feedback mechanism achieves cross-module collaboration through six bidirectional channels. When the preprocessing module outputs a clean signal, it synchronously transmits noise features to the acquisition module to optimize the sampling strategy. When the mode decomposition module outputs dual-mode components, it feeds back decomposition indices to the preprocessing module to adjust signal segmentation parameters. When the feature extraction module generates the amplitude parameter sequence, it transmits downsampling parameters to the mode decomposition module to improve boundary handling. When the collaborative quantization module outputs feature vectors, it feeds back state transition probabilities to the feature extraction module to optimize feature selection. When the fusion modeling module outputs constrained feature parameters, it transmits model parameters to the collaborative quantization module to adjust the quantization strategy. When the error compensation module outputs the corrected energy value, it transmits residual data to the acquisition module to achieve system-level calibration.

[0036] When the run-domain multi-feature constrained power signal analysis system is implemented in a wind power grid-connected substation, the acquisition module captures voltage and current signals through a multi-channel synchronous acquisition unit. The recursive least squares algorithm dynamically adjusts the sampling frequency based on signal spectrum changes caused by wind speed fluctuations. When sudden wind speed changes lead to an expansion of signal frequency components, the algorithm immediately increases the sampling rate to the maximum allowable value. A sliding window mechanism extracts signal segments in minute-level intervals, assessing stationarity by calculating the variance and crest factor within each segment. A millisecond-level timestamp and signal-to-noise ratio (SNR) quality identifier are appended to each data point, forming a standardized raw data stream.

[0037] After receiving the raw data stream, the preprocessing module applies the Heine-Borel finite coverage theorem to divide the continuous time axis into sub-intervals synchronized with the fundamental wave period of the wind turbine. In photovoltaic power plant applications, to address signal fluctuations caused by changes in light intensity, wavelet packet denoising is used within each sub-interval for multi-scale decomposition, and a soft thresholding function is used to eliminate high-frequency noise generated by the inverter. The denoising threshold is dynamically calibrated based on the real-time signal-to-noise ratio of the sub-intervals. When a signal abrupt change caused by cloud obstruction is detected, the threshold sensitivity is automatically reduced to retain effective signal components. The clean signal output after preprocessing includes a distortion index, and the extracted noise statistical features include harmonic distribution parameters. These parameters are fed back to the acquisition module to optimize the sampling strategy.

[0038] When processing clean signals, the mode decomposition module first predicts the signal trend components using a long short-term memory network. In scenarios involving sudden load changes at high-speed rail traction stations, the network analyzes historical data to predict signal distortion caused by the start-up and shutdown of traction motors. The improved empirical mode decomposition algorithm uses the prediction results to initialize envelope calculations. During the decomposition iteration process, it calculates the amplitude domain energy difference and orthogonality index in real time. When mode aliasing is detected, the proportional-integral controller automatically increases the number of screening iterations. The quasi-steady-state term generated by the decomposition includes the fundamental component, while the dynamic term covers harmonic and interharmonic components. Simultaneously, the calculated index parameters are sent back to the preprocessing module to optimize the sub-interval partitioning scheme.

[0039] The feature extraction module calculates the 1-norm of the amplitude sequence within the preprocessed sub-interval and extracts key feature points through non-uniform downsampling. The particle swarm optimization algorithm dynamically searches for the optimal downsampling frequency with the goal of minimizing reconstruction error. During wind farm scheduling, the algorithm adaptively adjusts the search strategy based on power fluctuation characteristics. The fuzzy logic system maps amplitude values ​​to multiple membership functions and uses the centroid method to defuzzify measurement uncertainties. The generated downsampling amplitude parameter sequence retains key signal features while eliminating the influence of random noise.

[0040] The collaborative quantization module establishes an N-ary quantization interval to discretize the amplitude parameter sequence. The Hidden Markov Model (HMM) analyzes the state transition probability of the run-length symbol sequence using a forward-backward algorithm. In the analysis of photovoltaic power plant output power fluctuations, the run-length feature extracted by the model reflects the persistence of illumination changes, the modulation depth feature characterizes the power fluctuation amplitude, and the impact intensity feature captures abrupt changes caused by cloud movement. The extracted feature parameters are consistent with the run-length statistical distribution. When the consistency deviation exceeds a threshold, quantization interval recalibration is triggered to ensure the reliability of feature extraction.

[0041] The fusion modeling module defines the run length range as an internal constraint and the amplitude variance range as an external constraint. Support vector machines use radial basis function kernels to classify feature parameters. In the collaborative analysis of cross-regional new energy power stations, the federated learning framework aggregates data from stations in different geographical locations and updates global model parameters using a distributed gradient descent algorithm. Constrained feature parameters are marked as valid features, and the model update information includes feature weights and classification thresholds. These parameters are fed back to the collaborative quantization module to guide feature selection.

[0042] The error compensation module constructs a reinforcement learning environment, where the agent's observation state space includes constrained feature parameters and historical error sequences. The Q-learning algorithm employs an exploratory strategy to select gain correction or phase shift actions. A long m-sequence dynamic energy reference test signal provides a standard reference value, and the compensation strategy is adjusted by comparing the deviation between the actual metering results and the standard value. In actual wind farm operation, the system predictively adjusts metering parameters based on wind speed trends, and the compensation residual data is fed back to the acquisition module in real time, forming a complete closed-loop control from signal acquisition to error correction.

[0043] The system achieves inter-module collaborative optimization through six feedback channels. Noise statistical characteristics output by the preprocessing module guide the acquisition module to adjust its sampling strategy; decomposition indices from the mode decomposition module optimize the signal segmentation parameters of the preprocessing module; downsampling parameters from the feature extraction module improve boundary handling in mode decomposition; state transition probabilities from the collaborative quantization module optimize the feature extraction strategy; model parameters from the fusion modeling module guide feature quantization; and residual data from the error compensation module calibrates the entire system operation. This mesh feedback structure enables the system to maintain adaptive optimization capabilities in complex scenarios such as wind power, photovoltaics, and high-speed rail traction, significantly improving the accuracy of dynamic energy metering.

[0044] In a run-domain multi-feature constrained power signal analysis system, the run-domain is defined as a feature space with the signal run length as the core analysis dimension. The run length represents the duration for which the signal amplitude value remains within a specific quantization interval. This definition is applicable to dynamic power signal analysis in scenarios such as wind power, photovoltaics, and high-speed rail traction. The amplitude domain energy difference refers to the difference in amplitude energy between adjacent intrinsic mode functions during mode decomposition. It is obtained by calculating the difference between the sums of squares of each mode signal and is used to measure the degree of mode separation. The loss ratio represents the proportion of signal energy lost during mode decomposition and is calculated using the energy ratio of the original signal to the reconstructed signal. The orthogonality index evaluates the degree of orthogonality between intrinsic mode functions and quantifies the independence between modes through inner product operations. The specific collaborative mechanism of the system's multi-layer closed-loop feedback network is as follows: The acquisition module receives noise statistical features from the preprocessing module, including noise power spectral density and amplitude distribution parameters, and dynamically adjusts the sampling frequency parameters of the recursive least squares algorithm; the preprocessing module receives amplitude domain energy difference, loss degree, and orthogonality indexes from the mode decomposition module, and optimizes the sub-interval partitioning strategy of the Heine-Borel finite coverage theorem and the wavelet packet denoising threshold; the mode decomposition module receives the optimal downsampling frequency parameters from the feature extraction module, and improves the boundary condition initialization process of the trend component of the predicted signal in the long short-term memory network; the feature extraction module receives the state transition probability matrix from the collaborative quantization module, and adjusts the downsampling frequency search space of the particle swarm optimization algorithm; the collaborative quantization module receives updated model parameters from the fusion modeling module, including the classification threshold and feature weights of the support vector machine, and recalibrates the quantization interval boundary of the hidden Markov model; the error compensation module receives the updated signal quality identifier and historical residual data from the acquisition module, and optimizes the action selection mechanism of the Q-learning algorithm of the reinforcement learning strategy. The aforementioned feedback data stream ensures that the system maintains adaptive optimization in dynamic signal processing through continuous iteration, and the parameter adjustment of each module is based on real-time calculated indicators to achieve closed-loop control.

[0045] In the collaborative quantization module of the run-domain multi-feature constrained power signal analysis system, N, representing the number of quantization levels in the N-ary quantization interval, is a positive integer. The value of N is dynamically determined based on the dynamic range of signal amplitude and historical data statistical characteristics in new energy scenarios. For example, in wind power or photovoltaic signal analysis, N is typically 8 or 16 to balance feature resolution and computational complexity. The boundary values ​​of the quantization interval are adaptively adjusted by analyzing the probability distribution of the amplitude parameter sequence, using equal probability or equal interval methods to set the boundaries, ensuring that the run-length symbol sequence can effectively capture the dynamic characteristics of the signal. In the error compensation module, the long m-sequence is a pseudo-random sequence generated by a linear feedback shift register, where m represents the order of the shift register, determining the sequence length. The long m-sequence has ideal autocorrelation characteristics and is used to generate dynamic power reference test signals. The value of m is selected according to the test accuracy requirements; for example, when m=10, the sequence length is 1023 bits. The sequence generating polynomial uses a primitive polynomial, such as x^10 + x^3 + 1, to ensure the pseudo-randomness and periodicity of the sequence. Data transmission between system modules is achieved through standard communication protocols, such as TCP / IP or shared memory mechanisms, to ensure real-time transmission of feedback data. When the preprocessing module sends noise statistical features to the acquisition module, the data format includes timestamps, noise power spectral density, and amplitude distribution parameters, and is encapsulated in JSON or binary format to ensure data integrity and parsing efficiency.

[0046] In a run-domain multi-feature constrained power signal analysis system, the run length feature is calculated by counting the number of sampling points where the signal amplitude value appears consecutively within an N-ary quantization interval, specifically representing the maximum continuous time length for which the signal amplitude value remains within the same quantization interval. The modulation depth feature is obtained by calculating the difference between the maximum and minimum signal amplitude values ​​within the run period and normalizing it to the average amplitude value, reflecting the relative depth of amplitude changes. The impulse intensity feature is quantified by the square of the amplitude abrupt change component of the integrated signal over a time window, capturing the absolute value of the abrupt energy. The amplitude domain energy difference is calculated as the difference between the sum of squares of the amplitudes of adjacent intrinsic mode functions during mode decomposition, used to measure the degree of mode separation. The loss is calculated as the positive value of the ratio of the original signal energy to the reconstructed signal energy, characterizing the proportion of energy loss during the decomposition process. The orthogonality index is obtained by calculating and summing the inner products of each pair of intrinsic mode functions, evaluating the independence between modes. The non-uniform downsampling algorithm adaptively selects the sampling point position based on the rate of change of the signal's first-norm gradient, ensuring the preservation of key feature points. The particle swarm optimization algorithm aims to minimize the reconstruction error, dynamically searching for combinations of downsampling frequencies with a population size of 20 and 50 iterations. The fuzzy logic system uses a triangular membership function for membership mapping and defuzzifies the data using the centroid method to generate clear values. The Hidden Markov Model (HMM) training uses the Baum-Welch algorithm to iteratively update the state transition matrix and observation probability matrix, maximizing the likelihood probability of the observed sequence. Support Vector Machine (SVM) classification uses a radial basis function kernel to map features to a high-dimensional space, finding the optimal classification hyperplane through the margin maximization principle; the kernel function parameters are dynamically determined through cross-validation. In the Q-learning algorithm, the action value function update uses the Bellman equation, with an initial learning rate of 0.1 that decays over time, a discount factor of 0.9, and an ε-greedy strategy to balance exploration and exploitation. The long m-sequence dynamic energy reference test signal is generated by a linear feedback shift register. The register order m is selected based on the test accuracy requirements; for example, when m=10, the sequence length is 1023 bits. The generator polynomial is the primitive polynomial x^10 + x^3 + 1, ensuring the pseudo-randomness and periodicity of the sequence. Data transfer between system modules is achieved through TCP / IP protocol or shared memory mechanism. Feedback data is encapsulated in JSON format, including timestamp, parameter values ​​and quality indicators, to ensure real-time transmission and parsing efficiency.

Claims

1. A run-domain multi-feature constrained electrical energy signal analysis system, characterized in that, include: The acquisition module is used to perform signal acquisition, dynamically adjust the sampling frequency using a recursive least squares algorithm, capture voltage and current signals, and generate a raw data stream with timestamps. The preprocessing module is used to receive the raw data stream, perform collaborative preprocessing, apply the Heine-Borel finite coverage theorem to divide the signal time axis into sub-intervals, and apply wavelet packet denoising to filter out high-frequency noise within the sub-intervals to generate a clean signal. It is also used to send the noise statistical features extracted during the preprocessing process back to the acquisition module; The mode decomposition module is used to receive the clean signal, perform bidirectional mode decomposition, and use an improved empirical mode decomposition algorithm to separate the quasi-steady-state term and dynamic term as bimodal components. It is also used to calculate the amplitude domain energy difference, loss degree and orthogonality index during the decomposition process, and send the index back to the preprocessing module; The feature extraction module is used to receive the dual-modal components, perform integrated feature extraction, calculate the amplitude 1-norm in the preprocessed sub-interval, and generate a downsampled amplitude parameter sequence after non-uniform downsampling and fuzzy logic defuzzification. The collaborative quantization module is used to receive the amplitude parameter sequence, perform run-domain collaborative quantization, generate a run-length symbol sequence through N-ary interval discretization, and extract the run-length, modulation depth and impulse intensity feature parameters by applying a hidden Markov model and encapsulating them into feature vectors. The fusion modeling module is used to receive the feature vector, perform multi-constraint fusion modeling, apply internal constraints in the run-length domain and external constraints in the amplitude domain, perform dynamic threshold classification of the feature parameters through a support vector machine, and output constrained feature parameters; it is also used to send the model parameters generated by modeling back to the collaborative quantization module. The error compensation module is used to receive the constrained feature parameters, perform dynamic error collaborative compensation, drive the adjustment of the power metering algorithm parameters through reinforcement learning strategy, and output the corrected power value; it is also used to send the residual data generated during the compensation process back to the acquisition module.

2. The run-domain multi-feature constrained power signal analysis system according to claim 1, characterized in that, The acquisition module is used to: analyze the spectral characteristics of the captured voltage and current signals using a recursive least squares algorithm, dynamically calculate and implement the adjustment amount of the sampling frequency; when the signal fluctuation is detected to be aggravated, increase the sampling rate according to the adjustment amount; evaluate the stability of the signal using a sliding window mechanism, add timestamps and quality identifiers to the acquired signal data, and generate a raw data stream with timestamps.

3. The run-domain multi-feature constrained electrical energy signal analysis system according to claim 1, characterized in that, The preprocessing module is used to: apply the Heine-Borel finite coverage theorem to divide the signal time axis of the original data stream into a finite number of sub-intervals; within each sub-interval, apply wavelet packet denoising and eliminate high-frequency noise through a soft threshold function; wherein the denoising threshold is dynamically calibrated based on the signal-to-noise ratio of the sub-interval; output a clean signal and extract the statistical features of the high-frequency noise as noise statistical features for transmission.

4. The run-domain multi-feature constrained electrical energy signal analysis system according to claim 1, characterized in that, The mode decomposition module is used to process the clean signal using an improved empirical mode decomposition algorithm; Before decomposition, the trend component of the signal is predicted by a long short-term memory network, and the decomposition boundary conditions are initialized with the trend component. During the decomposition iteration process, the amplitude domain energy difference, loss degree, and orthogonality index are calculated; the amplitude domain energy difference, loss degree, and orthogonality index are input into the proportional-integral controller to dynamically adjust the number of decomposition iterations; The quasi-steady-state term and dynamic term obtained from the output separation are used as dual-mode components.

5. The run-domain multi-feature constrained electrical energy signal analysis system according to claim 1, characterized in that, The feature extraction module is used to: calculate the 1-norm of the dual-modal component amplitude sequence within the sub-interval defined by collaborative preprocessing; generate an amplitude parameter sequence by applying 1-norm non-uniform downsampling; and dynamically search for downsampling frequency combinations to update the downsampling process using a particle swarm optimization algorithm. The amplitude value output by the updated downsampling process is then processed by a fuzzy logic system for membership mapping and defuzzification to generate a downsampling amplitude parameter sequence.

6. The run-domain multi-feature constrained electrical energy signal analysis system according to claim 1, characterized in that, The collaborative quantization module is used to: define an N-ary quantization interval; discretize the continuous amplitude parameter sequence into a run-length symbol sequence; analyze the run-length symbol sequence using a Hidden Markov Model (HMM), train model parameters using the Baum-Welch algorithm, and extract run-length, modulation depth, and impact intensity feature parameters; verify the consistency of the extracted run-length, modulation depth, and impact intensity feature parameters with the independently calculated run-length probability density distribution and autocorrelation function analysis results; trigger recalibration of the quantization interval boundary when the verification deviation exceeds a threshold; and encapsulate the verified feature parameters into a feature vector for output.

7. The run-domain multi-feature constrained electrical energy signal analysis system according to claim 1, characterized in that, The fusion modeling module is used to: define the run length range as an internal constraint and the amplitude variance range as an external constraint; use a support vector machine to perform dynamic threshold classification on the feature parameters in the feature vector; introduce a federated learning framework to aggregate feature data from multiple scenarios and update the support vector machine model parameters through a distributed gradient descent algorithm; output the feature parameters that satisfy the internal and external constraints as constrained feature parameters, and send the updated model parameters.

8. The run-domain multi-feature constrained power signal analysis system according to claim 1, characterized in that, The error compensation module is used to: select gain correction or phase offset adjustment actions to drive the adjustment of the power metering algorithm parameters by using the constrained feature parameters and historical error data as input through a reinforcement learning strategy; update the action value function by applying the Q learning algorithm; verify the metering results using a long m-sequence dynamic power reference test signal, and adjust the reinforcement learning strategy according to the verification results. Output the corrected energy value and send the residual data generated by the compensation.

9. The run-domain multi-feature constrained electrical energy signal analysis system according to claim 1, characterized in that, The preprocessing module is further configured to send the extracted noise statistical features back to the acquisition module; the mode decomposition module is further configured to send the calculated amplitude domain energy difference, loss degree, and orthogonality index back to the preprocessing module; the feature extraction module is further configured to send the optimal downsampling frequency parameters obtained by the particle swarm optimization algorithm back to the mode decomposition module; the cooperative quantization module is further configured to send the state transition probability matrix obtained by the hidden Markov model analysis back to the feature extraction module; the fusion modeling module is further configured to send the updated model parameters back to the cooperative quantization module; and the error compensation module is further configured to send the residual data generated by the compensation back to the acquisition module.

10. The run-domain multi-feature constrained electrical energy signal analysis system according to claim 1, characterized in that, The preprocessing module, mode decomposition module, feature extraction module, collaborative quantization module, fusion modeling module, and error compensation module, while outputting the clean signal, dual-modal components, amplitude parameter sequence, feature vector, constrained feature parameters, and corrected energy value, also send the generated noise statistical characteristics, amplitude domain energy difference, loss degree and orthogonality index, optimal downsampling frequency parameter, state transition probability matrix, updated model parameters, and residual data back to the acquisition module, preprocessing module, mode decomposition module, feature extraction module, collaborative quantization module, and acquisition module, respectively, forming a multi-layer closed-loop feedback.

Images