Intelligent monitoring and optimization method and system for wind-solar-storage operation of centralized control center

Abstract

The present application relates to the technical field of new energy centralized control center intelligent monitoring and equipment health state evaluation, especially relates to a centralized control center wind, light and storage operation intelligent monitoring optimization method and system, comprising collecting operation physical components to construct four-dimensional discrete sequences to obtain an original data set; outputting a fine net operation time sequence through bidirectional long short-term memory network and physical mechanism logical comparison; using whale optimization algorithm to iterate variation mode decomposition parameters, separating weak high-frequency components according to energy distribution entropy value; mapping to high-dimensional Hilbert feature space, projecting to a degradation indicator subspace through local retention projection and asymmetric mapping mechanism, obtaining explicit sub-health state evolution characteristics; combining multi-scale permutation entropy and equipment degradation physical model comparison and verification, stripping early performance degradation characteristics. The present application solves the problem of high mode overlap caused by dynamic weather strong disturbance and equipment micro-loss, and realizes the accurate stripping of early performance degradation characteristics from random environmental noise.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and equipment health status assessment technology for new energy centralized control centers, and particularly to an intelligent monitoring and optimization method and system for wind, solar and energy storage operation in centralized control centers. Background Technology

[0002] The centralized control center, as a centralized monitoring and dispatch center, collects real-time operational data from wind farms, photovoltaic power stations, and energy storage power stations through a data acquisition and monitoring control system, enabling remote centralized monitoring and control. The power prediction system predicts future wind and solar power output based on meteorological data, while the energy management system formulates and executes optimal power generation and energy storage control strategies based on the prediction results, grid dispatch instructions, and real-time operating status to mitigate the fluctuations and intermittent nature of wind and solar power output. As a key flexible resource, the energy storage system can store and release electrical energy, effectively participate in grid peak shaving and frequency regulation, and coordinate with wind and solar power generation to achieve smooth output and plan tracking.

[0003] Existing intelligent monitoring technologies for wind, solar, and energy storage operations in centralized control centers suffer from the following technical challenges: In scenarios involving the extraction of weak signal features from massive operational data of wind, solar, and energy storage clusters and the application of refined health assessment technologies across the entire equipment lifecycle, the incremental power loss caused by the aerodynamic performance degradation of wind turbine blades or the increased internal resistance of energy storage cells is limited in its temporal distribution and spectral characteristics by highly random external environmental fluctuations. Sub-health is defined as minor damage or performance deviation in the physical structure of critical equipment components, which, although not yet triggering protection trip alarms, has already deviated from the critical stage of rated design. Taking photovoltaic array inverters as an example, when power semiconductor devices experience weak leakage current fluctuations due to thermal cycling fatigue, the characteristic signals generated by the actual degradation trend of the equipment are easily completely drowned out by broadband random environmental noise due to the strong dynamic weather disturbances caused by drastic changes in irradiance. Because the external power fluctuations induced by the environment and the early defect evolution inside the equipment exhibit a high degree of modal overlap on the power characteristic curve, conventional monitoring algorithms cannot effectively decouple potential faults from background clutter. Ultimately, this makes it difficult for the condition assessment system to accurately predict the evolution logic of new energy units from microscopic performance degradation to macroscopic fault triggering. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent monitoring and optimization method and system for wind, solar and energy storage operation in a centralized control center. This invention solves the technical problem that early performance degradation characteristics cannot be accurately separated from random environmental noise due to the high modal overlap between strong dynamic weather disturbances and incremental equipment losses on the power characteristic curve.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: In a first aspect, the intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center provided by this invention includes: Step 1: Using the data interface of the new energy centralized control system, collect the irradiance, wind speed, ambient temperature output by the meteorological station of the station, as well as the bus voltage, feeder current, and active power reported in real time by the unit terminal. Map the physical components to the preset three-dimensional Cartesian coordinate system, associate the time dimension observation values, spatial coordinates, and physical quantity intensity, establish a four-dimensional discrete sequence, and obtain the original dataset including the micro-environmental disturbance and macro-output characteristics of the station. Step 2: Input the original dataset into a bidirectional long short-term memory network to identify abrupt outliers in the data stream. Perform logical comparison on outliers that conform to the power output upper limit model and the energy conservation criterion. Perform linear interpolation repair on bad points that do not conform to the physical mechanism to obtain a clean running time series. Step 3: Load the refined operating time series, start the whale optimization algorithm to iteratively search for the penalty factor and mode number of variational mode decomposition, and decompose the refined operating time series into a series of intrinsic mode functions with center frequency distribution characteristics; based on the energy distribution entropy value of the intrinsic mode functions, identify and separate the meteorological trend component corresponding to the low frequency band, the regulation and control component corresponding to the mid frequency band, and the weak high frequency component representing the micro-evolution characteristics of the equipment from the intrinsic mode functions; Step 4: Map the weak high-frequency components to the high-dimensional Hilbert feature space, construct the neighborhood weight matrix using the local preservation projection algorithm, enhance the topological distortion features of the nonlinear manifold through the asymmetric mapping mechanism, project the weak signal components to the degradation indicator subspace, and obtain the explicit sub-health state evolution features. Step 5: Calculate the multi-scale permutation entropy value of the sub-health state evolution characteristics, convert the complexity variable into physical excitation force parameters and input them into the pre-constructed equipment degradation physical model, compare the theoretical degradation trajectory generated by simulation with the morphological residual of the sub-health state evolution characteristics, determine that the fluctuation is attributed to physical structural damage, and remove early performance degradation characteristics from the weak high-frequency components.

[0006] Furthermore, the intelligent monitoring and optimization method for wind, solar, and energy storage operation in the centralized control center described in this invention establishes a four-dimensional discrete sequence including: Real-time operating parameters of wind farms, photovoltaic power stations and energy storage power stations are asynchronously collected according to a preset sampling frequency; Spatial dimension alignment of the asynchronously collected real-time running parameters is performed using physical geographic location tags; Based on the time dimension observations, the real-time operating parameters that have been aligned in the spatial dimension are unified in terms of time reference, and a topological parameter matrix including time, spatial three-dimensional coordinates and physical quantity intensity is constructed.

[0007] Furthermore, the intelligent monitoring and optimization method for wind, solar, and energy storage operation in the centralized control center described in this invention identifies abrupt outliers in the data stream, including: The forward hidden layer is used to capture the causal logical trend of the original dataset over time; The collaborative backward hidden layer extracts the correlation feedback features between the current state and future time steps; By comparing the degree of deviation between the causal logical trend and the correlation feedback features, outliers in the original dataset that deviate from the preset physical evolution logic are located.

[0008] Furthermore, in the intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center described in this invention, step 3 includes: The initial parameter search space was determined by simulating whale encirclement behavior; Perform bubble net attack and random search within the initial parameter search space to update the candidate parameter combinations for variational mode decomposition; The candidate parameter combination that minimizes the frequency domain aliasing is selected as the optimal penalty factor and mode number.

[0009] Furthermore, in the intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center described in this invention, step 4 includes: The local neighborhood correlation degree of the weak high-frequency component is calculated in the Hilbert feature space to determine the weight distribution; The asymmetric mapping mechanism suppresses the weights of the linear drift components in the weight distribution and simultaneously amplifies the topological weights of the nonlinear twisted manifold. The weighted feature vectors are projected into the degradation indicator subspace to enhance the mapping difference between early loss features and background noise.

[0010] Furthermore, the intelligent monitoring and optimization method for wind, solar, and energy storage operation in the centralized control center described in this invention transforms complexity variables into physical excitation force parameters, including: The orderliness index of the evolutionary features of the sub-health state is extracted under multiple preset sampling time scales; The signal complexity variable is calculated based on the degree of abrupt change in the orderedness index. The signal complexity variable is used as an input to the aerodynamic model of the wind turbine generator or the internal resistance evolution formula of the energy storage battery for dynamic simulation.

[0011] Furthermore, in the intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center described in this invention, step 5 further includes: The extracted early performance degradation features are projected into a high-dimensional kernel space; Kernel principal component analysis algorithm is used to eliminate nonlinear correlations between projected features; The core feature vector reflecting the equipment's degradation trend is extracted by maximizing the feature variance.

[0012] Furthermore, in the intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center described in this invention, step 5 further includes: Retrieve reference feature fingerprints that match the current device type from the historical failure mode sample library; Calculate the correlation and matching degree between the core feature vector and the reference feature fingerprint; A multi-dimensional health assessment index system for the entire life cycle of equipment is constructed by combining real-time power deviation, temperature gradient, and internal resistance evolution rate.

[0013] Furthermore, in the intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center described in this invention, step 5 further includes: The data points of the real-time generated equipment life cycle health assessment index system are projected onto a preset fault evolution probability cloud map; Based on the probability distribution position of the data points in the fault evolution probability cloud map, the remaining life prediction value of the equipment evolving from the current sub-healthy state to the fault trigger threshold is calculated.

[0014] Secondly, the intelligent monitoring and optimization system for wind, solar, and energy storage operation in a centralized control center provided by this invention is applied to the intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center as described in any one of the claims, including: The data acquisition module is used to collect irradiance, wind speed, ambient temperature output by the meteorological station of the power station and bus voltage, feeder current and active power reported in real time by the end of the unit using the data interface of the new energy centralized control system. The physical components are mapped to a preset three-dimensional Cartesian coordinate system, and the time dimension observations, spatial coordinates and physical quantity intensity are associated to establish a four-dimensional discrete sequence, so as to obtain the original dataset including the micro-environmental disturbance and macro-output characteristics of the power station. The data cleaning module is used to input the original dataset into a bidirectional long short-term memory network, identify abrupt outliers in the data stream, perform logical comparisons on outliers that conform to the power output upper limit model and the energy conservation criterion, and perform linear interpolation repair on bad points that do not conform to the physical mechanism to obtain a clean running time series. The frequency domain extraction module is used to load the refined operating time series, start the whale optimization algorithm to iteratively search for the penalty factor and mode number of variational mode decomposition, and decompose the refined operating time series into a series of intrinsic mode functions with center frequency distribution characteristics; based on the energy distribution entropy value of the intrinsic mode functions, the module identifies and separates the meteorological trend component corresponding to the low frequency band, the regulation and control component corresponding to the mid frequency band, and the weak high frequency component representing the micro-evolution characteristics of the equipment from the intrinsic mode functions; The feature evolution mining module is used to map the weak high-frequency components to a high-dimensional Hilbert feature space, construct a neighborhood weight matrix using a local preservation projection algorithm, enhance the topological distortion features of the nonlinear manifold through an asymmetric mapping mechanism, project the weak signal components to the degradation indicator subspace, and obtain explicit sub-health state evolution features. The dynamic verification module is used to calculate the multi-scale permutation entropy value of the sub-health state evolution characteristics, convert the complexity variable into physical excitation force parameters and input them into the pre-constructed equipment degradation physical model, compare the theoretical degradation trajectory generated by simulation with the morphological residual of the sub-health state evolution characteristics, determine that the fluctuation is attributed to physical structural damage, and peel off the early performance degradation characteristics from the weak high-frequency components. The performance evaluation module projects the extracted early performance degradation features into a high-dimensional kernel space, uses kernel principal component analysis to eliminate the nonlinear correlation between the projected features, and extracts the core feature vectors that reflect the equipment's degradation trend. It also constructs a multi-dimensional equipment lifecycle health evaluation index system by combining historical failure mode sample library, real-time power deviation, temperature gradient, and internal resistance evolution rate, and projects the data points onto a preset fault evolution probability cloud map to calculate the predicted remaining lifespan of the equipment as it evolves from its current sub-healthy state to the fault trigger threshold.

[0015] Beneficial effects of this invention: The intelligent monitoring and optimization method and system for wind, solar, and energy storage operation in a centralized control center provided by this invention establishes a four-dimensional discrete sequence by mapping physical components to a preset three-dimensional Cartesian coordinate system and associating them with time-dimension observations. This unifies the spatial dimension and time reference of real-time operating parameters under asynchronous acquisition, eliminates temporal misalignment obstacles in the process of massive data fusion, and lays a spatiotemporally consistent underlying data support for cross-dimensional data fusion. The original dataset is input into a bidirectional long short-term memory network and compared with the power output upper limit model and energy conservation criterion. This eliminates false measurement noise while maintaining the continuity of the energy conservation criterion on the time axis, outputting a clean operating time series with real physical meaning. The clean operating time series is loaded and the whale optimization algorithm is iteratively searched for the optimal combination of decomposition parameters for variational mode decomposition. Based on the energy distribution entropy value of the intrinsic mode function, weak high-frequency components are separated from the intrinsic mode function, completing the initial decoupling of weak signals from meteorological trend components and regulation control components in the frequency domain. This overcomes the frequency domain aliasing defect caused by improper parameter selection in conventional algorithms. Weak high-frequency components are mapped to a high-dimensional Hilbert feature space. A neighborhood weight matrix is ​​constructed using a local preservation projection algorithm, and the topological distortion characteristics of the nonlinear manifold are enhanced through an asymmetric mapping mechanism. Weak signal components are projected onto the degradation indicator subspace to obtain explicit sub-health state evolution characteristics, which greatly amplifies the mapping difference between early loss characteristics and background environmental noise. This solves the problem of weak signal submersion caused by the high modal overlap of dynamic meteorological disturbances and equipment micro-incremental losses on the power characteristic curve from the manifold topology level. The multi-scale permutation entropy value of the sub-health state evolution characteristics is calculated, and the complexity variable is transformed into physical excitation force parameters and input into the pre-constructed equipment degradation physical model. By comparing the theoretical degradation trajectory generated by simulation with the morphological residual of the sub-health state evolution characteristics, false early warning signals induced by transient fluctuations in grid frequency are effectively eliminated at the physical mechanism level. Early performance degradation characteristics are accurately extracted from weak high-frequency components, and finally, accurate quantitative assessment and reliable early warning of the evolution logic of new energy equipment from micro-sub-health state to macro-fault triggering are realized. Attached Figure Description

[0016] 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.

[0017] Figure 1 This is a flowchart illustrating the intelligent monitoring and optimization method for wind, solar, and energy storage operation in the centralized control center according to the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention provided by various embodiments will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] Firstly, please refer to Figure 1 The intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center provided by this invention includes: Step 1: Using the data interface of the new energy centralized control system, collect the irradiance, wind speed, ambient temperature output by the meteorological station of the station, as well as the bus voltage, feeder current, and active power reported in real time by the unit terminal. Map the physical components to the preset three-dimensional Cartesian coordinate system, associate the time dimension observation values, spatial coordinates, and physical quantity intensity, establish a four-dimensional discrete sequence, and obtain the original dataset including the micro-environmental disturbance and macro-output characteristics of the station. Step 2: Input the original dataset into a bidirectional long short-term memory network to identify abrupt outliers in the data stream. Perform logical comparison on outliers that conform to the power output upper limit model and the energy conservation criterion. Perform linear interpolation repair on bad points that do not conform to the physical mechanism to obtain a clean running time series. Step 3: Load the refined operating time series, start the whale optimization algorithm to iteratively search for the penalty factor and mode number of variational mode decomposition, and decompose the refined operating time series into a series of intrinsic mode functions with center frequency distribution characteristics; based on the energy distribution entropy value of the intrinsic mode functions, identify and separate the meteorological trend component corresponding to the low frequency band, the regulation and control component corresponding to the mid frequency band, and the weak high frequency component representing the micro-evolution characteristics of the equipment from the intrinsic mode functions; Step 4: Map the weak high-frequency components to the high-dimensional Hilbert feature space, construct the neighborhood weight matrix using the local preservation projection algorithm, enhance the topological distortion features of the nonlinear manifold through the asymmetric mapping mechanism, project the weak signal components to the degradation indicator subspace, and obtain the explicit sub-health state evolution features. Step 5: Calculate the multi-scale permutation entropy value of the sub-health state evolution characteristics, convert the complexity variable into physical excitation force parameters and input them into the pre-constructed equipment degradation physical model, compare the theoretical degradation trajectory generated by simulation with the morphological residual of the sub-health state evolution characteristics, determine that the fluctuation is attributed to physical structural damage, and remove early performance degradation characteristics from the weak high-frequency components.

[0020] In the daily operation of a new energy centralized control center, devices from different manufacturers often have inconsistent sampling frequencies. Utilizing the data interface of the new energy centralized control system, irradiance, wind speed, and ambient temperature output from the station's meteorological station, as well as bus voltage, feeder current, and active power reported in real time by the generator units, are collected. To unify the spatial reference, the collected physical components are mapped to a preset three-dimensional Cartesian coordinate system, assigning specific spatial location labels to each component. A four-dimensional discrete sequence is established by associating time-dimensional observations, spatial coordinates, and physical quantity intensity. After alignment processing with the spatiotemporal reference, a raw dataset including the station's micro-environmental disturbances and macro-output characteristics is obtained. The logic between these steps is that only when the underlying data possesses unified spatiotemporal attributes can the subsequent neural network effectively uncover spatial correlations and temporal evolution patterns.

[0021] After obtaining the raw dataset, the challenge lies in handling communication jitter and sensor drift interference within the massive dataset. The raw dataset is input into a bidirectional long short-term memory network. A forward hidden layer captures the causal logic trend of the data evolution over time, while a backward hidden layer extracts the correlation feedback features between future moments and the current state. By comparing the deviations between the forward and backward features, abrupt outliers in the data stream are identified. Relying solely on data-driven algorithms is prone to misjudgment; therefore, a logical comparison of abrupt outliers is necessary, incorporating physical constraints. Outliers conforming to the power output upper limit model and energy conservation criterion are retained, while outliers that do not conform to physical mechanisms are repaired using linear interpolation. This interpolation repair process eliminates spurious measurement noise, resulting in a clean running time series, thus removing broadband background noise interference for subsequent frequency domain decomposition.

[0022] To address the challenge of highly overlapping modes between environmental fluctuations and micro-losses, a clean operating time series is loaded, and the whale optimization algorithm is used to iteratively search for the penalty factor and mode number of variational mode decomposition. Simulating whale encirclement behavior and bubble network attacks, global parameter optimization is performed, selecting the parameter combination that minimizes frequency domain aliasing. The clean operating time series is then decomposed into a series of intrinsic mode functions (IMFs) with center frequency distribution characteristics. To address the difficulty in intuitively defining IMFs, the degree of disorder in each frequency band component is quantified based on the energy distribution entropy value of the IMFs. From the IMFs, meteorological trend components corresponding to the low-frequency band, regulation and control components corresponding to the mid-frequency band, and weak high-frequency components representing the micro-evolutionary characteristics of the equipment's internal structure are identified and separated. This achieves preliminary decoupling at the frequency domain level, allowing fault precursors previously masked by environmental fluctuations to manifest independently.

[0023] The separated weak high-frequency components still exhibit strong non-stationary characteristics in the time domain, requiring nonlinear manifold mining. These weak high-frequency components are mapped to a high-dimensional Hilbert feature space, where local neighborhood correlations are calculated using a local preserving projection algorithm to construct a neighborhood weight matrix. To address the interweaving of linear drift trajectories caused by dynamic meteorological disturbances and nonlinear distorted manifolds resulting from incremental equipment losses, an asymmetric mapping mechanism is used to enhance the topological distortion features of the nonlinear manifold. The weights of linear drift components in the weight distribution are suppressed, while the topological weights of the nonlinear distorted manifold are simultaneously amplified. The weak signal components are projected onto the degradation indicator subspace, resulting in explicit sub-health state evolution characteristics. This explicit topological feature makes the weak signal completely stand out from background clutter, completing the leap from signal extraction to feature representation.

[0024] To prevent false early warnings caused by power grid transient fluctuations, closed-loop verification of the extracted features at the physical dimension is required. This involves calculating the multi-scale permutation entropy value of the sub-health state evolution features and extracting permutation orderliness indicators at different time scales. Complexity variables are generated based on the degree of abrupt change in the permutation orderliness indicators, and these complexity variables are transformed into physical excitation force parameters and input into a pre-constructed physical model of equipment degradation. Dynamic simulations are then performed using aerodynamic models of wind turbine generators or internal resistance evolution formulas for energy storage batteries, comparing the theoretical degradation trajectory generated by the simulation with the morphological residuals of the sub-health state evolution features. It is determined whether the morphological residuals are within a pre-set confidence interval; if they are, the fluctuations are attributed to physical structural damage, thus separating early performance degradation features from weak high-frequency components. This process, encompassing data fusion, deep cleaning, frequency domain decoupling, manifold enhancement, and mechanism verification, overcomes the technical bottleneck of early-stage faults in new energy equipment being overwhelmed by random environmental noise.

[0025] In actual new energy power plant operation environments, hardware interfaces from different suppliers often exhibit communication latency differences, resulting in inconsistent data reporting times. Real-time operating parameters of wind farms, photovoltaic power plants, and energy storage power plants are asynchronously collected according to a preset sampling frequency. To overcome information silos in the physical space dimension, physical geographic location tags are used to spatially align the asynchronously collected real-time operating parameters. This spatial alignment process assigns real physical space coordinates to isolated measurement point data. Based on these physical space coordinates, the time reference of the spatially aligned real-time operating parameters is unified according to the time dimension observations. This time reference unification eliminates the timing misalignment problem caused by asynchronous sampling, thereby constructing a topological parameter matrix including time, three-dimensional spatial coordinates, and physical quantity intensity.

[0026] After constructing the topology parameter matrix, complex communication jitter interference issues arise. A forward hidden layer is used to capture the causal logical trend of the original dataset over time. This causal logical trend reflects the smooth operating trajectory of physical devices under conditions free from external abrupt changes. Considering the lag effect in power system transient processes, a collaborative backward hidden layer is used to extract the correlation feedback features between future moments and the current state. These correlation feedback features supplement the backtracking information of state abrupt changes that cannot be perceived by unidirectional time series. After obtaining the causal logical trend and correlation feedback features, outliers in the original dataset that deviate from the preset physical evolution logic are located by comparing the degree of deviation between the two.

[0027] The parameters of conventional variational mode decomposition (MODED) are difficult to adaptively select, directly affecting the frequency domain decomposition performance. An initial parameter search space is determined by simulating whale encirclement behavior. This initial parameter search space defines the feasible solution range for the penalty factor and the number of modes. Within the defined initial parameter search space, bubble net attacks and random searches are performed to update the candidate parameter combinations for variational mode decomposition. The bubble net attack is used for in-depth exploration near local optima, while the random search is used to escape local extrema traps and conduct global exploration. After multiple iterations, multiple sets of candidate parameter combinations are generated, and the candidate parameter combination that minimizes frequency domain aliasing is selected as the optimal penalty factor and number of modes.

[0028] After acquiring weak high-frequency components, deep features often manifest as nonlinear topological structures in high-dimensional space. Local neighborhood correlations of these weak high-frequency components are calculated within the Hilbert feature space to determine the weight distribution. Local neighborhood correlations reflect the density of data point clustering on the manifold structure. Since meteorological disturbances typically exhibit linear drift, while equipment degradation manifests as nonlinear distortion, an asymmetric mapping mechanism is used to suppress the weights of linearly drifting components in the weight distribution, while simultaneously amplifying the topological weights of the nonlinearly distorted manifold. After weight redistribution, the weighted feature vectors are projected into the degradation indicator subspace to enhance the mapping difference between early loss features and background environmental noise.

[0029] To map abstract data features back to the physical mechanism level, an orderliness index of the evolution characteristics of sub-health states is extracted at multiple preset sampling time scales. The orderliness index characterizes the inherent randomness of the time series at different observation scales. A signal complexity variable is calculated based on the degree of abrupt change in the orderliness index. This signal complexity variable quantifies the severity of the deviation of the equipment's internal operating state from the normal operating baseline. To verify the physical rationality of the state deviation, the signal complexity variable is used as input to a wind turbine aerodynamic model or an energy storage battery internal resistance evolution formula for dynamic simulation.

[0030] After extracting early performance degradation features, excessively high feature dimensionality can lead to the curse of dimensionality in subsequent evaluation algorithms. The extracted early performance degradation features are projected into a high-dimensional kernel space. This high-dimensional kernel space mapping transforms originally linearly inseparable degradation features into linearly separable states. Kernel principal component analysis is then used to eliminate nonlinear correlations between the projected features. Eliminating nonlinear correlations removes redundant feature dimensionality information. During dimensionality reduction, the core feature vector reflecting the device's degradation trend is extracted by maximizing the feature variance.

[0031] After extracting the core feature vector, historical operating experience needs to be introduced as a health assessment benchmark. Reference feature fingerprints matching the current equipment type are retrieved from the historical failure mode sample library. These reference feature fingerprints include extensive state evolution data of similar equipment before typical failures occurred in previous operating cycles. After obtaining the reference benchmark, the correlation matching degree between the core feature vector and the reference feature fingerprint is calculated. The correlation matching measure quantifies the similarity level between the current equipment state and known failure modes. Based on the similarity level and combined with real-time power deviation, temperature gradient, and internal resistance evolution rate, a multi-dimensional equipment lifecycle health assessment index system is constructed.

[0032] After establishing a health assessment index system for the entire equipment lifecycle, it is necessary to transform these multi-dimensional indicators into intuitive lifespan predictions. The data points from the real-time generated health assessment index system are projected onto a pre-defined fault evolution probability cloud map. This cloud map displays the statistical probability distribution of equipment failure under different combinations of health indicators. Based on the probability distribution positions of the data points in the fault evolution probability cloud map, the predicted remaining lifespan of the equipment as it evolves from its current sub-healthy state to the fault trigger threshold is calculated.

[0033] In the data processing path for establishing a four-dimensional discrete sequence, the collected physical components are mapped to a preset three-dimensional Cartesian coordinate system, and the time dimension observations, spatial coordinates, and physical quantity intensity are correlated. The specific mapping calculation logic is expressed as follows: in, Indicates the first Each measuring point is at A four-dimensional discrete sequence constructed at each moment; Indicates the first The lateral spatial coordinates of each measuring point in a preset three-dimensional Cartesian coordinate system; Indicates the first The longitudinal spatial coordinates of each measuring point in a preset three-dimensional Cartesian coordinate system; Indicates the first The vertical spatial coordinates of each measuring point in a preset three-dimensional Cartesian coordinate system; Indicates the first Each measuring point is at The physical quantity intensity feature vectors collected at all times include irradiance, wind speed, ambient temperature, bus voltage, feeder current, and active power. This represents the current time-dimension observation. Through the coordinate association described above, independent data streams are transformed into a raw dataset with a unified spatiotemporal reference.

[0034] In the data processing path that yields the cleaned running time series, the specific calculation process for linear interpolation repair of bad points that do not conform to the physical mechanism, targeting identified abrupt outliers, is as follows: in, Indicates the point where a mutation occurs outlier. The refined running time series feature values ​​obtained after linear interpolation repair at each time point; The true observed value of a physical quantity at the effective sampling time before the occurrence of a mutation outlier. This represents the true observed value of a physical quantity at one effective sampling time after the occurrence of a mutation outlier. This indicates the time point at which a defect is determined to be inconsistent with physical mechanisms; This represents the time point at which an effective sampling moment occurs before the occurrence of a mutation outlier. This represents a valid sampling point after the occurrence of a mutation outlier. The missing sequence is filled in using an interpolation repair process, maintaining the continuity of the energy conservation principle along the time axis.

[0035] In the data processing path of disassembling and identifying intrinsic mode functions, the specific calculation logic for quantifying the degree of frequency band component disorder based on the energy distribution entropy value of the intrinsic mode functions is as follows: in, The first value represents the output of variational mode decomposition. The total energy of each intrinsic mode function; This represents the total length of discrete sampling points in the clean running time series; This represents the discrete time step in the clean running time series; This indicates that the variational mode decomposition occurs at time step [missing information]. The output of the first The amplitude values ​​of the intrinsic mode functions; Indicates the first The entropy value of the energy distribution of an intrinsic mode function is used to measure the non-stationarity of the signal distribution. This represents the total number of modes specified in the variational mode decomposition iterative search. This represents the sum of the energies of all intrinsic mode functions; This represents a logarithmic operation with the natural constant as the base. It achieves automatic decoupling and separation of weak high-frequency components and meteorological trend components by utilizing the numerical difference in energy distribution entropy values.

[0036] In the data processing path of manifesting the evolutionary characteristics of sub-health state, the specific calculation logic for constructing the neighborhood weight matrix using the local preserving projection algorithm is as follows: in, The neighborhood weight matrix constructed by the local preserving projection algorithm represents the first... The feature node and the first Topological connection weight values ​​between feature nodes; Represents an exponential function with the natural constant as its base; The first term represents the weak high-frequency component mapped to the high-dimensional Hilbert feature space. Local manifold data points; The first term represents the weak high-frequency component mapped to the high-dimensional Hilbert feature space. Local manifold data points; Indicates the first The local manifold data point and the first The square of the Euclidean distance between data points in a local manifold is used to measure the degree of geometric similarity. This represents the scale attenuation constant that controls the width of the local neighborhood. The weak signal components are projected onto the degradation indicator subspace by adjusting the connection weights through the aforementioned asymmetric mapping mechanism.

[0037] In the data processing path for extracting the core feature vectors reflecting the equipment's attenuation trend, the kernel principal component analysis algorithm is used to eliminate the nonlinear correlation between projected features, and the eigenvalue decomposition of the covariance matrix of the core feature vectors is calculated as follows: in, This represents the feature covariance matrix after mapping to a high-dimensional kernel space; This represents the total number of samples representing early performance degradation features projected onto the high-dimensional kernel space; Sample labels indicating early performance degradation characteristics; Indicates the first A vector of early performance degradation feature samples; This represents the vector of early performance degradation features. Nonlinear mapping functions that map to a high-dimensional kernel space; Indicates the first The transpose of the feature vectors of early performance degradation feature samples mapped to a high-dimensional kernel space; The first characteristic covariance matrix represents the... Each eigenvalue represents the feature variance energy along the projection direction of the principal component of the core feature vector. Indicates the relationship with the first The core feature vectors corresponding to each feature value are obtained. Based on the feature variance maximization extraction rule, the core feature vectors of the top features are obtained, thus completing the construction of a full life cycle health assessment index system.

[0038] Secondly, the intelligent monitoring and optimization system for wind, solar, and energy storage operation in a centralized control center provided by this invention is applied to the intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center as described in any one of the claims, including: The data acquisition module is used to collect irradiance, wind speed, ambient temperature output by the meteorological station of the power station and bus voltage, feeder current and active power reported in real time by the end of the unit using the data interface of the new energy centralized control system. The physical components are mapped to a preset three-dimensional Cartesian coordinate system, and the time dimension observations, spatial coordinates and physical quantity intensity are associated to establish a four-dimensional discrete sequence, so as to obtain the original dataset including the micro-environmental disturbance and macro-output characteristics of the power station. The data cleaning module is used to input the original dataset into a bidirectional long short-term memory network, identify abrupt outliers in the data stream, perform logical comparisons on outliers that conform to the power output upper limit model and the energy conservation criterion, and perform linear interpolation repair on bad points that do not conform to the physical mechanism to obtain a clean running time series. The frequency domain extraction module is used to load the refined operating time series, start the whale optimization algorithm to iteratively search for the penalty factor and mode number of variational mode decomposition, and decompose the refined operating time series into a series of intrinsic mode functions with center frequency distribution characteristics; based on the energy distribution entropy value of the intrinsic mode functions, the module identifies and separates the meteorological trend component corresponding to the low frequency band, the regulation and control component corresponding to the mid frequency band, and the weak high frequency component representing the micro-evolution characteristics of the equipment from the intrinsic mode functions; The feature evolution mining module is used to map the weak high-frequency components to a high-dimensional Hilbert feature space, construct a neighborhood weight matrix using a local preservation projection algorithm, enhance the topological distortion features of the nonlinear manifold through an asymmetric mapping mechanism, project the weak signal components to the degradation indicator subspace, and obtain explicit sub-health state evolution features. The dynamic verification module is used to calculate the multi-scale permutation entropy value of the sub-health state evolution characteristics, convert the complexity variable into physical excitation force parameters and input them into the pre-constructed equipment degradation physical model, compare the theoretical degradation trajectory generated by simulation with the morphological residual of the sub-health state evolution characteristics, determine that the fluctuation is attributed to physical structural damage, and peel off the early performance degradation characteristics from the weak high-frequency components. The performance evaluation module projects the extracted early performance degradation features into a high-dimensional kernel space, uses kernel principal component analysis to eliminate the nonlinear correlation between the projected features, and extracts the core feature vectors that reflect the equipment's degradation trend. It also constructs a multi-dimensional equipment lifecycle health evaluation index system by combining historical failure mode sample library, real-time power deviation, temperature gradient, and internal resistance evolution rate, and projects the data points onto a preset fault evolution probability cloud map to calculate the predicted remaining lifespan of the equipment as it evolves from its current sub-healthy state to the fault trigger threshold.

[0039] In the application scenario of intelligent monitoring of wind, solar, and energy storage operations in a centralized control center, the incremental loss in output power caused by the aerodynamic performance degradation of wind turbine blades or the increase in internal resistance of energy storage cells highly overlaps with the modes of random external environmental fluctuations in terms of time domain distribution and spectral characteristics. To decouple the complex intertwined signals, the data interface of the new energy centralized control system is used to collect irradiance, wind speed, ambient temperature output from the meteorological station at the wind farm, as well as bus voltage, feeder current, and active power reported in real time by the turbine terminals. Hardware interfaces often have communication delay differences, resulting in inconsistent data reporting time nodes at the lower levels. Real-time operating parameters of wind farms, photovoltaic power stations, and energy storage power stations are asynchronously collected according to a preset sampling frequency. The physical components are mapped to a preset three-dimensional Cartesian coordinate system, and the asynchronously collected real-time operating parameters are spatially aligned using physical geographic location labels. After the spatial benchmark is constructed, the time benchmark is unified for the spatially aligned real-time operating parameters based on the time dimension observations. By associating time-dimensional observations, spatial coordinates, and physical quantity intensity, a four-dimensional discrete sequence is established to obtain the original dataset that includes the micro-environmental disturbances and macro-output characteristics of the station.

[0040] Massive datasets inevitably contain communication jitter and sensor drift interference. The original dataset is input into a bidirectional long short-term memory network. A forward hidden layer captures the causal logic trend of the original dataset over time, extracting the smooth operating trajectory of physical devices under conditions free from external abrupt changes. Power system transient processes exhibit lag effects; a backward hidden layer is used to extract the correlation feedback features between future moments and the current state. By comparing the degree of deviation between the causal logic trend and the correlation feedback features, outliers in the original dataset that deviate from the preset physical evolution logic are located, identifying abrupt outliers in the data stream. Logical comparison of abrupt outliers is performed using physical constraint boundaries to verify that data fluctuations conform to the actual situation of the power output upper limit model and the energy conservation principle. Linear interpolation is performed to repair bad points that do not conform to the physical mechanism, eliminating spurious measurement noise and obtaining a clean operating time series.

[0041] The difficulty in adaptively selecting parameters for conventional algorithms affects the frequency domain decomposition effect. A clean running time series is loaded, and the whale optimization algorithm is initiated to iteratively search for the penalty factor and mode number of variational mode decomposition. The initial parameter search space is determined by simulating whale encirclement behavior, defining the feasible solution range for the penalty factor and mode number. Within the initial parameter search space, bubble net attacks and random searches are performed to update the candidate parameter combinations for variational mode decomposition. The candidate parameter combination that minimizes frequency domain aliasing is selected as the optimal penalty factor and mode number, decomposing the clean running time series into a series of intrinsic mode functions (IMFs) with center frequency distribution characteristics. Based on the energy distribution entropy values ​​of the IMFs, the disorder level of each frequency band component is quantified. From the IMFs, meteorological trend components corresponding to the low-frequency band, regulation and control components corresponding to the mid-frequency band, and weak high-frequency components representing the micro-evolutionary characteristics of the equipment are identified and separated.

[0042] The separated weak high-frequency components exhibit strong non-stationary characteristics in the time domain. These components are mapped to a high-dimensional Hilbert feature space. Local neighborhood correlations of the weak high-frequency components are calculated within the Hilbert feature space to determine the weight distribution, and a neighborhood weight matrix is ​​constructed using a local preserving projection algorithm. The linear drift trajectory formed by dynamic meteorological disturbances intertwines with the nonlinear distorted manifold generated by incremental equipment losses. An asymmetric mapping mechanism is used to suppress the weights of the linear drift components in the weight distribution, simultaneously amplifying the topological weights of the nonlinear distorted manifold and enhancing its topological distortion characteristics. The weighted eigenvectors are projected onto the degradation indicator subspace, amplifying the mapping difference between early loss characteristics and background environmental noise. By projecting weak signal components onto the degradation indicator subspace, explicit sub-health state evolution characteristics are obtained.

[0043] To prevent false early warnings caused by power grid transient fluctuations, a closed-loop verification based on physical dimensions is required, involving the calculation of multi-scale permutation entropy values ​​of sub-health state evolution characteristics. The permutation orderliness index of sub-health state evolution characteristics is extracted across multiple preset sampling time scales. The signal complexity variable is calculated based on the degree of abrupt change in the permutation orderliness index and then transformed into physical excitation force parameters. This signal complexity variable is then used as input to a wind turbine aerodynamic model or an energy storage battery internal resistance evolution formula for dynamic simulation, and input into a pre-constructed equipment degradation physical model. The simulated theoretical degradation trajectory is compared with the morphological residuals of the sub-health state evolution characteristics. If the morphological residuals are within a preset confidence interval, the fluctuations are attributed to physical structural damage, thus separating early performance degradation characteristics from weak high-frequency components.

[0044] To address the curse of dimensionality caused by excessively high feature dimensions, early performance degradation features are projected into a high-dimensional kernel space. Kernel principal component analysis (KPCA) is used to eliminate nonlinear correlations between projected features. Core feature vectors reflecting equipment degradation trends are extracted by maximizing feature variance. Historical operating experience is introduced as a benchmark, and reference feature fingerprints matching the current equipment type are retrieved from a historical failure mode sample library. The correlation matching degree between the core feature vector and the reference feature fingerprints is calculated. A multi-dimensional equipment lifecycle health assessment index system is constructed by combining real-time power deviation, temperature gradient, and internal resistance evolution rate. Data points from the real-time generated equipment lifecycle health assessment index system are projected onto a pre-defined fault evolution probability cloud map. Based on the probability distribution position of the data points in the fault evolution probability cloud map, the predicted remaining lifespan of the equipment from its current sub-healthy state to the fault trigger threshold is calculated.

[0045] Embodiment 1 of this invention: In the actual operation of large-scale wind farms, the slight power loss caused by minor icing on the blade surface or slight degradation of aerodynamic performance is often deeply embedded in the background noise of the power spectrum caused by gusts. Using the data interface of the new energy centralized control system, the wind speed, ambient temperature, and real-time bus voltage, feeder current, and active power reported by the wind turbine terminals are asynchronously collected at a preset sampling frequency of 100 milliseconds. The extracted physical components are mapped to a preset three-dimensional Cartesian coordinate system, and the asynchronously collected real-time operating parameters are spatially aligned using physical geographic location labels. After spatial alignment, the real-time operating parameters are unified based on the time dimension observations, establishing a topological parameter matrix including time, three-dimensional spatial coordinates, and physical quantity intensity. The topological parameter matrix is ​​transformed into the original dataset and input into a bidirectional long short-term memory network. The forward hidden layer captures the causal logic trend of the original dataset evolving over time, and the backward hidden layer extracts the correlation feedback features of future moments to the current state. By comparing the deviation between causal logical trends and correlation feedback characteristics, outliers in the original dataset that deviate from the preset physical evolution logic are located. Outliers conforming to the power output upper limit model and energy conservation criterion are logically compared, and bad points that do not conform to the physical mechanism are repaired using linear interpolation to obtain a clean operating time series. The clean operating time series is loaded, and the whale optimization algorithm is initiated to determine the initial parameter search space by simulating whale encirclement behavior. Within the initial parameter search space, bubble net attack and random search are performed to select the candidate parameter combination that minimizes frequency domain aliasing as the optimal penalty factor and mode number. Based on the optimal decomposition parameters, the clean operating time series is decomposed into a series of intrinsic mode functions (IMFs) with central frequency distribution characteristics. Based on the energy distribution entropy values ​​of the IMFs, weak high-frequency components representing the micro-evolutionary characteristics inside the wind turbine are separated from the IMFs. These weak high-frequency components are mapped to a high-dimensional Hilbert feature space, and the local neighborhood correlation degree of the weak high-frequency components is calculated within the feature space to determine the weight distribution. The weights of linear drift components in the weight distribution are suppressed by an asymmetric mapping mechanism, while the topological weights of the nonlinear distorted manifold are amplified simultaneously. The weighted feature vectors are projected onto the degradation indicator subspace to obtain explicit sub-health state evolution features. The orderliness index of the sub-health state evolution features is extracted at different preset sampling time scales, and the signal complexity variable is calculated based on the degree of mutation of the orderliness index. The signal complexity variable is then used as input to a wind turbine aerodynamic model for dynamic simulation. By comparing the theoretical degradation trajectory generated by the simulation with the morphological residuals of the sub-health state evolution features, the fluctuations are determined to be attributable to physical structural damage. Early performance degradation features are extracted from weak high-frequency components and projected onto a high-dimensional kernel space. Kernel principal component analysis is used to eliminate the nonlinear correlation between the projected features, extracting the core feature vectors reflecting the equipment degradation trend.The system retrieves reference feature fingerprints that match the current wind turbine generator from the historical failure mode sample library, and constructs a multi-dimensional equipment life cycle health assessment index system by combining real-time power deviation and temperature gradient.

[0046] Embodiment 2 of this invention: Photovoltaic array inverters face thermal cycling fatigue during long-term operation, leading to weak leakage current fluctuations in power semiconductor devices. Limited by dynamic weather disturbances caused by drastic changes in irradiance, the characteristic signals of the actual degradation trend are easily submerged by broadband random environmental noise. For photovoltaic power plant applications, the data interface of the new energy centralized control system is used to asynchronously collect irradiance, ambient temperature, and real-time bus voltage, feeder current, and active power reported by the photovoltaic inverter terminal from the meteorological station at a preset sampling frequency of fifty milliseconds. The corresponding physical components are mapped to a preset three-dimensional Cartesian coordinate system, and the time-dimensional observations, spatial coordinates, and physical quantity intensity are associated to establish a four-dimensional discrete sequence. The original dataset, including the micro-environmental disturbances and macro-output characteristics of the power plant, is input into a bidirectional long short-term memory network to identify abrupt outliers in the data stream. The identified abrupt outliers are logically compared with the photovoltaic power output upper limit model to filter out false measurement noise caused by communication jitter. A linear interpolation repair operation is performed to obtain a clean operating time series, which is then loaded into a variational mode decomposition engine. The whale optimization algorithm iteratively searches for the penalty factor and mode number of variational mode decomposition. Based on the energy distribution entropy value of the intrinsic mode function, it identifies and separates the meteorological trend component corresponding to the low-frequency band, the regulation and control component corresponding to the mid-frequency band, and the weak high-frequency component representing the micro-evolution characteristics of the inverter's internal structure from the intrinsic mode function. To deeply mine the information contained in the weak high-frequency component, it maps the weak high-frequency component to a high-dimensional Hilbert feature space. A neighborhood weight matrix is ​​constructed using a local preservation projection algorithm, and the topological distortion characteristics of the nonlinear manifold are enhanced through an asymmetric mapping mechanism. The weak signal component is projected onto the degradation indicator subspace to obtain explicit sub-health state evolution characteristics. The multi-scale permutation entropy value of the sub-health state evolution characteristics is calculated, and the complexity variable is transformed into physical excitation force parameters. The physical excitation force parameters are input into a pre-constructed photovoltaic inverter degradation physical model, and the morphological residuals of the simulated theoretical degradation trajectory and the sub-health state evolution characteristics are compared. When the morphological residuals are within a pre-set confidence interval, the fluctuation is determined to be attributable to damage to the physical structure of the semiconductor device. Early performance degradation features are extracted from weak high-frequency components and projected into a high-dimensional kernel space. Kernel principal component analysis (KPCA) is used to eliminate nonlinear correlations between features, and core feature vectors reflecting the inverter's degradation trend are extracted by maximizing feature variance. Reference feature fingerprints matching the current inverter type are retrieved from a historical failure mode sample library, and the correlation degree between the core feature vectors and the reference feature fingerprints is calculated.

[0047] Embodiment 3 of this invention: When an energy storage system participates in grid peak shaving and frequency regulation, the transient fluctuations in grid frequency and the incremental losses caused by the increased internal resistance of the energy storage cells exhibit high modal overlap on the output power characteristic curve. For energy storage power station application scenarios, the data interface of the new energy centralized control system is used to asynchronously collect the ambient temperature of the energy storage power station and the real-time reported bus voltage, feeder current, and active power from the energy storage converter terminal at a preset sampling frequency of ten milliseconds. The extracted physical components are mapped to a preset three-dimensional Cartesian coordinate system, and the real-time operating parameters of the battery cluster are spatially aligned using physical geographic location labels. A time reference is unified based on the time dimension observations, and a topology parameter matrix is ​​constructed. The topology parameter matrix representing the original dataset is input into a bidirectional long short-term memory network, and the forward hidden layer is used to capture the causal logical trend of the original dataset's evolution over time. The backward hidden layer, in conjunction with the network, extracts the correlation feedback features of future moments to the current state, locating outliers in the original dataset that deviate from the preset physical evolution logic. Outliers conforming to the energy conservation criterion are logically compared, and bad points that do not conform to the physical mechanism are repaired by linear interpolation, outputting a clean operating time series. The whale optimization algorithm is initiated iteratively to search for the penalty factor and mode number of variational mode decomposition, decomposing the clean operating time series into a series of intrinsic mode functions with center frequency distribution characteristics. Based on the energy distribution entropy value of the intrinsic mode functions, weak high-frequency components representing the micro-evolution characteristics inside the energy storage cell are separated from the intrinsic mode functions. These weak high-frequency components are mapped to a high-dimensional Hilbert feature space, and the local neighborhood correlation degree of the weak high-frequency components is calculated in the Hilbert feature space to determine the weight distribution. An asymmetric mapping mechanism is used to suppress the linear drift component weight representing the grid dispatch command response in the weight distribution, while simultaneously amplifying the nonlinear twisted manifold topology weight representing the internal resistance evolution. The feature vector is projected onto the degradation indicator subspace to obtain explicit sub-health state evolution characteristics, and the orderliness index of the sub-health state evolution characteristics is extracted. The signal complexity variable is calculated based on the degree of abrupt change in the orderliness index. This variable is then used as input to the internal resistance evolution formula of the energy storage battery for dynamic simulation. By comparing the theoretical degradation trajectory generated in the simulation with the measured residual, the fluctuation is determined to be caused by damage to the internal chemical structure of the cell. Early performance degradation features are extracted from weak high-frequency components and projected into a high-dimensional kernel space. Kernel principal component analysis is used to eliminate nonlinear correlations between features and extract core feature vectors. Matching reference feature fingerprints are retrieved from a historical failure mode sample library. A multi-dimensional equipment lifecycle health assessment index system is constructed by combining real-time power deviation and internal resistance evolution rate. The data points of the real-time generated equipment lifecycle health assessment index system are projected onto a preset fault evolution probability cloud map.

[0048] Four-dimensional discrete sequences refer to structured data matrices formed by mapping independent one-dimensional time-series signals collected by different sensors to a unified timestamp and incorporating spatial geographic location information. In the specific implementation logic, each meteorological station and terminal measurement element of a wind farm or energy storage power station is assigned preset three-dimensional Cartesian coordinates (horizontal, vertical, and longitudinal), and millisecond-level time dimension observations are used as alignment benchmarks. This transforms isolated physical quantity intensity feature vectors into data points with clear spatiotemporal labels. A multi-dimensional tensor containing spatial location information, temporal evolution information, and physical state information is established, enabling subsequent algorithms to fully perceive the correlation of environmental disturbances between spatially proximate units.

[0049] A clean operating time series refers to a smooth, high-quality data stream that fully conforms to the objective laws of energy conversion, after the original dataset has undergone dual filtering by neural network identification and physical rules to eliminate false anomalies caused by hardware defects such as sensor zero-point drift and communication latency jitter. At the underlying data processing path, a bidirectional long short-term memory network is used to capture the correlation between historical state evolution trends and future feedback, locating abrupt outliers. Combining the rated power output upper limit model of the power generation equipment with the system energy conservation criterion, logical consistency checks are performed on these abrupt outliers. Linear interpolation or mean repair is performed on outliers determined to violate physical mechanisms, filling in missing data while eliminating non-mechanistic high-frequency glitches, maintaining the strict continuity of energy flow along the time axis.

[0050] Intrinsic mode functions (IMFs) refer to the amplitude-frequency modulated (AM) signal components that satisfy specific center frequency band distribution characteristics obtained after iterative decomposition of a non-stationary power sequence using a variational mode decomposition engine. Addressing the severe mode aliasing issue in actual industrial field data, the whale optimization algorithm is employed to perform bubble net attacks and random searches within the global parameter search space to find the combination of penalty factors and mode numbers that minimizes frequency domain aliasing. Under optimal parameter constraints, the wideband runtime series is adaptively deconstructed into multiple narrowband eigenmodes with independent physical meaning. Based on the energy distribution entropy values ​​of each IMF, the degree of frequency band component disorder can be quantified, providing a clear frequency domain boundary criterion for subsequent accurate extraction of weak high-frequency components representing micro-device losses.

[0051] The degradation indicator subspace refers to a specific mathematical projection interval constructed using feature dimensionality reduction and manifold learning algorithms when processing high-dimensional non-stationary signals. Within this specific mathematical projection interval, signal components representing minor damage to the internal physical structure of the device are amplified, while strong interference components representing regular changes in the external meteorological environment are significantly suppressed. The underlying technical solution employs a local preservation projection algorithm to calculate the local neighborhood correlation of data points on the manifold structure, constructing a neighborhood weight matrix. An asymmetric mapping mechanism is used to forcibly adjust the projection direction of the feature basis vectors, suppressing the weights of linear drift components with periodic rotation and translation characteristics, and highlighting the topological structure with nonlinear distortion characteristics. This provides a dedicated mathematical expression space for separating weak signals from environmental background noise.

[0052] The explicit evolutionary characteristics of sub-health states refer to early performance degradation signals that were originally hidden beneath the background noise of strong dynamic weather disturbances and were difficult to detect by conventional threshold alarm systems. After manifold topological reconstruction, these signals acquire measurable and distinguishable feature vectors in terms of data form. When slight icing occurs on wind turbine blades or trace amounts of lithium plating appear inside energy storage battery cells, the resulting characteristic frequencies severely overlap with wind speed pulsations or grid transient frequency fluctuations. By projecting the weak high-frequency components onto the degradation indicator subspace, the hidden physical degradation behavior is restored to a clearly visible topological distortion on the characteristic manifold, achieving a substantial transformation from microscopic performance deviations in sub-health equipment to macroscopically observable data characteristics.

[0053] Physical excitation force parameters refer to the abstract complexity indices calculated from purely data-driven models, which are transformed into parameters that can be directly input into the digital simulation model of the equipment as dynamic or electrochemical excitation sources through dimensional transformation and mechanism mapping. In the specific implementation, a multi-scale permutation entropy algorithm is used to calculate the permutation orderliness index of signal sequences at different time scales. The higher the degree of abrupt change in the permutation orderliness index, the more severe the instability of the internal microstructure of the equipment. The entropy value variable characterizing the signal complexity is mapped to the mechanical stress parameters required for the aerodynamic model of the wind turbine generator, or to the thermal stress parameters required for the internal resistance evolution formula of the energy storage battery, through a preset transformation logic. The introduction of physical mechanism simulation comparison eliminates the risk of overfitting and false alarms that may arise from pure algorithm models.

[0054] A fault evolution probability cloud map is a probabilistic statistical distribution model constructed based on a large historical equipment failure mode sample library. It describes the evolution of a unit from a sub-healthy critical state to a complete loss of specified functions. The construction process employs kernel principal component analysis to eliminate nonlinear correlations between historical fault features and extracts the principal components of core feature vectors. Using the core feature vectors of different degradation stages as coordinate anchors, a continuous distribution cloud map reflecting the failure probability density is generated in a multi-dimensional space using a Gaussian distribution. Real-time extracted equipment health assessment index data points are projected onto the generated continuous cloud map distribution. Based on the probability density interval of the real-time data points, the remaining lifespan interval for macroscopic failures of key equipment components is deduced.

Claims

1. An optimization method for intelligent monitoring of wind, solar, and energy storage operation in a centralized control center, characterized in that: include: Step 1: Using the data interface of the new energy centralized control system, collect the irradiance, wind speed, ambient temperature output by the meteorological station of the station, as well as the bus voltage, feeder current, and active power reported in real time by the unit terminal. Map the physical components to the preset three-dimensional Cartesian coordinate system, associate the time dimension observation values, spatial coordinates, and physical quantity intensity, establish a four-dimensional discrete sequence, and obtain the original dataset including the micro-environmental disturbance and macro-output characteristics of the station. Step 2: Input the original dataset into a bidirectional long short-term memory network to identify abrupt outliers in the data stream. Perform logical comparison on outliers that conform to the power output upper limit model and the energy conservation criterion. Perform linear interpolation repair on bad points that do not conform to the physical mechanism to obtain a clean running time series. Step 3: Load the refined operating time series, start the whale optimization algorithm to iteratively search for the penalty factor and mode number of variational mode decomposition, and decompose the refined operating time series into a series of intrinsic mode functions with center frequency distribution characteristics; based on the energy distribution entropy value of the intrinsic mode functions, identify and separate the meteorological trend component corresponding to the low frequency band, the regulation and control component corresponding to the mid frequency band, and the weak high frequency component representing the micro-evolution characteristics of the equipment from the intrinsic mode functions; Step 4: Map the weak high-frequency components to the high-dimensional Hilbert feature space, construct the neighborhood weight matrix using the local preservation projection algorithm, enhance the topological distortion features of the nonlinear manifold through the asymmetric mapping mechanism, project the weak signal components to the degradation indicator subspace, and obtain the explicit sub-health state evolution features. Step 5: Calculate the multi-scale permutation entropy value of the sub-health state evolution characteristics, convert the complexity variable into physical excitation force parameters and input them into the pre-constructed equipment degradation physical model, compare the theoretical degradation trajectory generated by simulation with the morphological residual of the sub-health state evolution characteristics, determine that the fluctuation is attributed to physical structural damage, and remove early performance degradation characteristics from the weak high-frequency components.

2. The intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center according to claim 1, characterized in that, Establishing a four-dimensional discrete sequence includes: Real-time operating parameters of wind farms, photovoltaic power stations and energy storage power stations are asynchronously collected according to a preset sampling frequency; Spatial dimension alignment of the asynchronously collected real-time running parameters is performed using physical geographic location tags; Based on the time dimension observations, the real-time operating parameters that have been aligned in the spatial dimension are unified in terms of time reference, and a topological parameter matrix including time, spatial three-dimensional coordinates and physical quantity intensity is constructed.

3. The intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center according to claim 2, characterized in that, Identifying abrupt outliers in a data stream includes: The forward hidden layer is used to capture the causal logical trend of the original dataset over time; The collaborative backward hidden layer extracts the correlation feedback features between the current state and future time steps; By comparing the degree of deviation between the causal logical trend and the correlation feedback features, outliers in the original dataset that deviate from the preset physical evolution logic are located.

4. The intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center according to claim 3, characterized in that, Step 3 includes: The initial parameter search space was determined by simulating whale encirclement behavior; Perform bubble net attack and random search within the initial parameter search space to update the candidate parameter combinations for variational mode decomposition; The candidate parameter combination that minimizes the frequency domain aliasing is selected as the optimal penalty factor and mode number.

5. The intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center according to claim 4, characterized in that, Step 4 includes: The local neighborhood correlation degree of the weak high-frequency component is calculated in the Hilbert feature space to determine the weight distribution; The asymmetric mapping mechanism suppresses the weights of the linear drift components in the weight distribution and simultaneously amplifies the topological weights of the nonlinear twisted manifold. The weighted feature vectors are projected into the degradation indicator subspace to enhance the mapping difference between early loss features and background noise.

6. The intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center according to claim 5, characterized in that, Transforming complexity variables into physical excitation force parameters includes: The orderliness index of the evolutionary features of the sub-health state is extracted under multiple preset sampling time scales; The signal complexity variable is calculated based on the degree of abrupt change in the orderedness index. The signal complexity variable is used as an input to the aerodynamic model of the wind turbine generator or the internal resistance evolution formula of the energy storage battery for dynamic simulation.

7. The intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center according to claim 6, characterized in that, Step 5 also includes: The extracted early performance degradation features are projected into a high-dimensional kernel space; Kernel principal component analysis algorithm is used to eliminate nonlinear correlations between projected features; The core feature vector reflecting the equipment's degradation trend is extracted by maximizing the feature variance.

8. The intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center according to claim 7, characterized in that, Step 5 also includes: Retrieve reference feature fingerprints that match the current device type from the historical failure mode sample library; Calculate the correlation and matching degree between the core feature vector and the reference feature fingerprint; A multi-dimensional health assessment index system for the entire life cycle of equipment is constructed by combining real-time power deviation, temperature gradient, and internal resistance evolution rate.

9. The intelligent monitoring and optimization method for wind, solar, and energy storage operation in a centralized control center according to claim 8, characterized in that, Step 5 also includes: The data points of the real-time generated equipment life cycle health assessment index system are projected onto a preset fault evolution probability cloud map; Based on the probability distribution position of the data points in the fault evolution probability cloud map, the remaining life prediction value of the equipment evolving from the current sub-healthy state to the fault trigger threshold is calculated.

10. A centralized control center intelligent monitoring and optimization system for wind, solar, and energy storage operation, applied to the intelligent monitoring and optimization method for wind, solar, and energy storage operation as described in any one of claims 1 to 9, characterized in that... include: The data acquisition module is used to collect irradiance, wind speed, ambient temperature output by the meteorological station of the power station and bus voltage, feeder current and active power reported in real time by the end of the unit using the data interface of the new energy centralized control system. The physical components are mapped to a preset three-dimensional Cartesian coordinate system, and the time dimension observations, spatial coordinates and physical quantity intensity are associated to establish a four-dimensional discrete sequence, so as to obtain the original dataset including the micro-environmental disturbance and macro-output characteristics of the power station. The data cleaning module is used to input the original dataset into a bidirectional long short-term memory network, identify abrupt outliers in the data stream, perform logical comparisons on outliers that conform to the power output upper limit model and the energy conservation criterion, and perform linear interpolation repair on bad points that do not conform to the physical mechanism to obtain a clean running time series. The frequency domain extraction module is used to load the refined operating time series, start the whale optimization algorithm to iteratively search for the penalty factor and mode number of variational mode decomposition, and decompose the refined operating time series into a series of intrinsic mode functions with center frequency distribution characteristics; based on the energy distribution entropy value of the intrinsic mode functions, the module identifies and separates the meteorological trend component corresponding to the low frequency band, the regulation and control component corresponding to the mid frequency band, and the weak high frequency component representing the micro-evolution characteristics of the equipment from the intrinsic mode functions; The feature evolution mining module is used to map the weak high-frequency components to a high-dimensional Hilbert feature space, construct a neighborhood weight matrix using a local preservation projection algorithm, enhance the topological distortion features of the nonlinear manifold through an asymmetric mapping mechanism, project the weak signal components to the degradation indicator subspace, and obtain explicit sub-health state evolution features. The dynamic verification module is used to calculate the multi-scale permutation entropy value of the sub-health state evolution characteristics, convert the complexity variable into physical excitation force parameters and input them into the pre-constructed equipment degradation physical model, compare the theoretical degradation trajectory generated by simulation with the morphological residual of the sub-health state evolution characteristics, determine that the fluctuation is attributed to physical structural damage, and peel off the early performance degradation characteristics from the weak high-frequency components. The performance evaluation module is used to project the stripped early performance degradation features into a high-dimensional kernel space, use kernel principal component analysis algorithm to eliminate the nonlinear correlation between the projected features, and extract the core feature vector that reflects the equipment degradation trend. By combining historical failure mode sample library, real-time power deviation, temperature gradient and internal resistance evolution rate, a multi-dimensional equipment life cycle health assessment index system is constructed. The data points are projected onto a preset fault evolution probability cloud map to calculate the remaining life prediction value of the equipment from the current sub-healthy state to the fault trigger threshold.

Images