Operation safety monitoring method and system of wind and light energy storage platform

Abstract

The invention discloses an operation safety monitoring method of a wind and light energy storage platform. According to the method, operation parameters are collected and standardized in real time, and power and environment features are extracted to construct a multi-dimensional feature vector; a key feature subset is determined through dimension reduction, and a normal operation mode is established through clustering analysis; calculating the health deviation degree between the real-time state track and the clustering center, and triggering an alarm sequence; constructing a parameter interaction directed graph to reversely search and position an abnormal source; a monitoring model is updated in combination with historical abnormal logs, and the real-time abnormal risk probability is calculated through recursive filtering. According to the invention, intelligent safety monitoring and dynamic risk quantification of the wind and light energy storage system are realized, and the operation safety and the accuracy of abnormal early warning are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of operational safety monitoring and big data analysis technology, specifically to a method and system for operational safety monitoring of wind and solar energy storage platforms based on multi-dimensional feature fusion and reverse tracing. Background Technology

[0002] Currently, in the new energy field, wind-solar-storage platforms, as complex systems integrating power generation, energy storage, and grid connection, directly affect the stability and reliability of the energy system through their operational safety. With the continuous expansion of renewable energy installed capacity, the platform's operating environment is becoming increasingly complex, and equipment operating parameters exhibit high-dimensionality, strong coupling, and dynamic changes. This places higher demands on the accuracy and real-time performance of safety monitoring, making safety monitoring a crucial technical foundation for ensuring the stable operation of wind-solar-storage platforms.

[0003] In existing technologies, Supervisory Control and Data Acquisition (SCADA) systems are typically used to monitor various operating parameters of wind, solar, and energy storage platforms in real time, and to issue alarms for abnormal states based on fixed thresholds, single-parameter limit judgments, or simple statistical models. However, due to the massive and high-dimensional nature of the parameters generated by the platform's operation, and the complex correlations and dynamic fluctuations between different parameters (such as wind power output, photovoltaic power, and energy storage charging and discharging status), existing technologies struggle to effectively extract key features reflecting the system's health status from this complex operational data. This results in an inability to accurately judge the system's state and capture subtle anomalies when faced with external disturbances or hidden abnormal patterns between parameters.

[0004] Existing technologies have the problem of making it difficult to accurately and timely identify and locate potential risks in the complex operating environment of wind and solar energy storage platforms. Summary of the Invention

[0005] This invention proposes a method and system for monitoring the operational safety of wind and solar energy storage platforms, in order to solve the problem that existing technologies are unable to accurately identify and locate potential risks in the complex operating environment of wind and solar energy storage platforms.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for monitoring the operational safety of a wind-solar energy storage platform, comprising:

[0007] Real-time operating parameters of the wind-solar-storage energy platform are collected, and noise and outliers are removed through standardization to obtain a preprocessed dataset. Power operation features and environmental impact features are extracted from the preprocessed dataset, and a multidimensional feature matrix is ​​constructed through multidimensional spatial mapping. The multidimensional feature matrix is ​​subjected to dimensionality reduction processing, and the key feature subset affecting the system operation is determined according to the projection weight of each feature dimension; Clustering algorithms are used to cluster the preset historical operating data within the key feature subset space to determine multiple cluster centers representing normal operating modes; The real-time running status trajectory of the key feature subset is obtained, the health status deviation between the real-time running status trajectory and the cluster center is calculated, and the alarm trigger sequence is determined based on the continuity of the health status deviation exceeding the preset deviation threshold. A directed graph of parameter interaction is constructed based on the alarm trigger sequence. An anomaly source identifier is obtained by performing a reverse search on the directed graph of parameter interaction using a graph traversal algorithm. Retrieve historical anomaly log sequences that match the anomaly source identifier, and extract the attenuation coefficient and residual feedback weight, which characterize the anomaly evolution. State variables are constructed based on the key feature subset, and improved Kalman filtering is performed using the parameter decay coefficient and the residual feedback weight to obtain posterior state estimates. Risk probability mapping is then performed on the posterior state estimates to obtain the abnormal risk probability.

[0008] Secondly, the present invention provides an operational safety monitoring system for a wind-solar energy storage platform, comprising: The data preprocessing module is used to collect real-time operating parameters of the wind and solar energy storage platform, and remove noise and outliers through standardization to obtain a preprocessed dataset. The feature construction module is used to extract power operation features and environmental impact features from the preprocessed dataset, and to construct a multi-dimensional feature matrix through multi-dimensional spatial mapping. The feature dimensionality reduction module is used to perform dimensionality reduction processing on the multidimensional feature matrix and determine the key feature subset that affects the operation of the system based on the projection weight of each feature dimension. The pattern clustering module is used to cluster the preset historical operating data in the key feature subset space using a clustering algorithm to determine multiple cluster centers that represent normal operating modes. The status monitoring module is used to acquire the real-time running status trajectory of the key feature subset, calculate the health status deviation between the real-time running status trajectory and the cluster center, and determine the alarm trigger sequence based on the continuity of the health status deviation exceeding the preset deviation threshold. The reverse tracing module is used to construct a directed graph of parameter interactions based on the alarm trigger sequence, and then use a graph traversal algorithm to construct the directed graph of parameter interactions. The reverse search shown in the diagram yields the identifier of the anomaly source. The evolution characteristic extraction module is used to retrieve historical abnormal log sequences that match the abnormal source identifier, and extract the parameter decay coefficient and residual feedback weight that characterize the abnormal evolution characteristics. The risk quantification module is used to construct state variables based on the key feature subset and perform improved Kalman filtering using the parameter decay coefficient and the residual feedback weight to obtain posterior state estimates. The posterior state estimates are then mapped to risk probabilities to obtain abnormal risk probabilities.

[0009] Compared with the prior art, the present invention has the following advantages: (1) This invention eliminates the impact of environmental noise and dimensional differences on monitoring results by performing wavelet denoising and standardization on multi-source heterogeneous data from wind-solar-storage platforms. By calculating the load weights in the principal component space for feature dimensionality reduction, core features such as power fluctuations and state of charge (SOC) change rates can be accurately identified from massive, strongly coupled power operation and environmental parameters. This not only reduces the computational load of the system but also significantly improves the system's sensitivity to hidden anomaly patterns.

[0010] (2) By using the density peak clustering algorithm to analyze historical data, this invention constructs multiple cluster centers and their distribution boundaries representing normal operation modes. Compared with traditional fixed threshold monitoring, this dynamic measurement method based on "health status deviation" can better adapt to changes in normal operating conditions of wind and solar energy storage platforms caused by weather fluctuations or load switching, effectively reducing false alarm rates and achieving accurate capture of subtle health deviations in the system.

[0011] (3) This invention constructs a directed graph of parameter interaction to mathematically represent the complex physical topology and parameter causal logic, and uses a reverse search algorithm to automatically trace the fault chain when an alarm occurs. This mechanism solves the problem of "knowing the anomaly but not the source" in the prior art. Through quantitative analysis of the cumulative impact intensity, it can quickly locate the initial source node that causes system fluctuations, providing intuitive and efficient decision support for operation and maintenance personnel.

[0012] (4) This invention upgrades the static monitoring model to a dynamic recursive filtering model by extracting the parameter attenuation coefficient and residual weight from historical logs. Combined with the threshold adaptive adjustment technology of residual distribution, the system can calculate the quantitative probability of abnormal risks in real time. This method fully considers the temporal evolution characteristics of abnormal development, making safety early warning more forward-looking and accurate, and ensuring the long-term stability of the wind and solar energy storage platform in a dynamic and ever-changing operating environment. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating an embodiment of the active recording interference method based on composite ultrasound provided by the present invention. Figure 2 This is a schematic diagram of an embodiment of the active recording interference system based on composite ultrasound provided by the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0015] Reference Figure 1 The first embodiment of the present invention provides a method for monitoring the operational safety of a wind-solar energy storage platform, including steps S11 to S18: S11: Collect real-time operating parameters of the wind and solar energy storage platform, remove noise and outliers through standardization processing, and obtain a preprocessed dataset; S12, extract power operation features and environmental impact features from the preprocessed dataset, and construct a multidimensional feature matrix through multidimensional spatial mapping; S13, perform dimensionality reduction processing on the multidimensional feature matrix, and determine the key feature subset affecting the system operation based on the projection weight of each feature dimension; S14, A clustering algorithm is used to cluster the preset historical operating data in the key feature subset space to determine multiple cluster centers representing normal operating modes; S15, obtain the real-time running status trajectory of the key feature subset, calculate the health status deviation between the real-time running status trajectory and the cluster center, and determine the alarm trigger sequence based on the continuity of the health status deviation exceeding the preset deviation threshold. S16, construct a directed graph of parameter interaction based on the alarm trigger sequence, and perform a reverse search in the directed graph of parameter interaction using a graph traversal algorithm to obtain the abnormal source identifier; S17, retrieve the historical anomaly log sequence that matches the anomaly source identifier, and extract the parameter decay coefficient and residual feedback weight that characterize the anomaly evolution characteristics from it. S18. Construct state variables based on the key feature subset and perform improved Kalman filtering using the parameter attenuation coefficient and the residual feedback weight to obtain posterior state estimates. Map the posterior state estimates to risk probabilities to obtain abnormal risk probabilities.

[0016] In step S11, real-time operating parameters of the wind-solar-storage energy platform are collected, and noise and outliers are removed through standardization processing to obtain a preprocessed dataset, including: Acquire multi-source heterogeneous characteristic data from wind, solar and energy storage measurement points, and construct a time-continuous real-time operation sequence by linear interpolation filling; Wavelet transform is used to filter out high-frequency noise components in the real-time running sequence, and a smoothed and denoised signal that retains trend characteristics is extracted. Abnormal outliers in the smoothed and denoised signal are removed using statistical discriminant analysis to obtain a corrected data sequence. The corrected data sequence is normalized and mapped to obtain unified dimensional data that eliminates the influence of dimensionality; The unified dimensional data is timestamped and reorganized in a multidimensional format to obtain the preprocessed dataset.

[0017] Specifically, the wind-solar-storage energy platform comprises wind power generation units, photovoltaic power generation units, and energy storage units. Each unit's measuring sensors collect operating parameters at different sampling frequencies and data formats. The measuring points for the wind power generation unit include wind speed sensors, pitch angle encoders, and generator temperature probes; for the photovoltaic power generation unit, they include irradiance sensors, module temperature sensors, and inverter current sensors; and for the energy storage unit, they include battery pack voltage sensors, charge / discharge current sensors, and battery temperature sensors. These multi-source, heterogeneous characteristic data suffer from timestamp misalignment due to inconsistent sampling frequencies, necessitating linear interpolation to construct a time-continuous real-time operating sequence.

[0018] Linear interpolation fills the data by selecting 1Hz as the uniform target frequency. For low-frequency measurement points, interpolation is calculated linearly between adjacent sampling points. For example, if the pitch angle is 15° at time t1 and 19° at time t2 (t1+2 seconds), then the interpolation value at the intermediate time t1.5 is 15° + (19° - 15°) × (1 / 2) = 17°. For high-frequency measurement points, the arithmetic mean of all sampling points is calculated within a one-second window as the representative value. After processing, a continuous time sequence with a uniform sampling frequency is formed, typically with a length of 3600 data points (corresponding to 1 hour).

[0019] To filter out high-frequency noise introduced by sensor electromagnetic interference and mechanical vibration in the sequence, wavelet transform is employed. The db4 wavelet from the Daubechies wavelet family is used as the mother wavelet function, with a decomposition level of 5. The wavelet transform decomposes the signal into low-frequency approximation coefficients and high-frequency detail coefficients. By analyzing the energy distribution of the detail coefficients at each level, it is identified that high-frequency noise is mainly concentrated in the first and second level detail coefficients (the sum of the energy of the first and second levels exceeds 70% of the total energy, thus being determined as high-frequency dominant noise). The first and second level detail coefficients are set to zero, while the third to fifth level detail coefficients and the fifth level approximation coefficient are retained. An inverse transform is then performed on the zeroed wavelet coefficients to reconstruct the signal, resulting in a smoothed and denoised signal.

[0020] To further improve data quality, it is necessary to remove outliers caused by sensor failures, communication packet loss, or extreme operating conditions from the smoothed and denoised signal. The specific method involves setting a sliding window for the signal, with a window length of 300 data points (corresponding to 5 minutes), sliding in steps of 60 data points (1 minute). Within each window, the mean and standard deviation of the data, as well as the first and third quartiles, are calculated. The interquartile range is then calculated based on the first and third quartiles. Subsequently, two normal data intervals are defined according to the statistical 3σ principle and the interquartile range outlier determination range. When a data point exceeds either normal data interval, it is identified as an outlier and corrected using linear interpolation of the preceding and following data points. If three or more consecutive data points are identified as outliers, that time period is marked as a missing data segment and removed. After processing, a corrected data sequence is obtained, with the outlier rate reduced to below 0.5%.

[0021] Because the operating parameters of wind-solar-storage energy platforms encompass various physical quantities such as power (kW), temperature (°C), voltage (V), and current (A), and the numerical ranges of these parameters differ significantly, normalization mapping is required to eliminate the influence of dimensions and obtain data with uniform dimensions. This scheme employs a minimum-maximum normalization method to map each parameter to the [0, 1] interval, with the normalized parameters being the minimum and maximum values ​​of historical operating data (usually selected from the past 30 days).

[0022] Finally, the data with unified dimensions is timestamped and reorganized into a multidimensional format to form a structured preprocessed dataset. Timestamp alignment uses Coordinated Universal Time (UTC) as the unified benchmark, performing time zone conversion and network clock synchronization correction on the local timestamps of each measurement point, achieving millisecond-level accuracy. Multidimensional format reorganization integrates the one-dimensional time series of each measurement point into a multidimensional matrix. Assuming the platform has M measurement points, each containing data from N time points, the reorganized data matrix has dimensions M×N. The i-th row of the matrix represents the time series of the i-th measurement point, and the j-th column represents a snapshot of the parameters of all measurement points at the j-th time. Simultaneously, timestamp vectors and measurement point identifier vectors are constructed as indices to complete the construction of the preprocessed dataset.

[0023] In step S12, power operation features and environmental impact features are extracted from the preprocessed dataset, and a multidimensional feature matrix is ​​constructed through multidimensional spatial mapping, including: Power operation data is extracted from the preprocessed dataset, and power fluctuation amplitude, voltage deviation sequence and current imbalance are calculated as power operation characteristics. Based on the environmental monitoring dimensions in the preprocessed dataset, the frequency of sudden wind speed changes and the gradient of light intensity are calculated as environmental impact characteristics. By combining the timestamps of the power operation characteristics and environmental impact characteristics, the corresponding energy storage state of charge change rate, temperature compensation offset and frequency response delay are extracted and determined as system response characteristics. The power operation characteristics, environmental impact characteristics, and system response characteristics are synchronized and aligned, and then mapped to a high-dimensional feature space through matrix recombination to construct a multi-dimensional feature matrix.

[0024] Specifically, power operation data, including time series of parameters such as wind power generation, photovoltaic power generation, energy storage charging and discharging power, three-phase voltage, and three-phase current, are extracted from the preprocessed dataset. The power fluctuation amplitude is calculated using a sliding window standard deviation method, setting a sliding window of 120 data points (2 minutes) and calculating the standard deviation of power data within the window with a step size of 30 data points (30 seconds). A larger standard deviation indicates more severe power fluctuations. The voltage deviation sequence is obtained by calculating the relative deviation between the actual voltage and the rated voltage. Current imbalance is used to measure the symmetry of the three-phase current, expressed as the ratio of negative-sequence current to positive-sequence current. These calculation results constitute power operation characteristics, reflecting the stable operating state of the system. Environmental impact characteristics are extracted based on the environmental monitoring dimension in the preprocessed dataset. The wind speed change frequency is obtained by counting the number of times the wind speed change rate exceeds a threshold per unit time. The wind speed change rate threshold is set to 2 m / s² (i.e., the wind speed changes by more than 2 m / s within 1 second), and the number of times this condition is met within a 1-hour time window is used as the change frequency. This threshold was determined based on statistical analysis of historical operating data. When the wind speed change rate exceeds 2 m / s², the standard deviation of wind power generation fluctuations increases by more than 1.5 times compared to normal operating conditions, and the probability of abnormal system conditions rises from 5% under normal conditions to over 25%, indicating that this change rate has a significant impact on system stability. Higher frequency of sudden wind speed changes indicates more unstable wind conditions and a more significant impact on wind power generation. The solar intensity gradient characterizes the severity of solar illumination changes and is obtained by calculating the difference in irradiance between adjacent moments. The average and maximum values ​​of the solar intensity gradient are calculated over one hour. The average gradient reflects the overall trend of solar illumination changes, while the maximum gradient reflects the intensity of the impact of sudden factors such as cloud cover.

[0025] By combining timestamps of power operation characteristics and environmental impact characteristics, response characteristic parameters of the energy storage system are extracted. The state of charge (SOC) change rate represents the battery's charging and discharging speed, obtained by calculating the difference in SOC between adjacent moments and dividing by the time interval, with units of % / min. A larger absolute value of the SOC change rate indicates a faster energy storage system response. Temperature compensation offset reflects the impact of ambient temperature on battery performance. It is calculated based on the deviation between the battery temperature and the standard operating temperature (typically 25℃), specifically by subtracting the standard operating temperature from the battery temperature and multiplying by a predetermined temperature coefficient (typically 0.005 / ℃ for lithium batteries). Frequency response delay measures the time difference between receiving a dispatch command and the actual power output of the energy storage system. It is calculated by comparing the moment of the dispatch command with the moment when the power reaches 95% of the target value; the delay time is typically in the range of 0.5-3 seconds. These parameters characterize the dynamic response capability of the energy storage system under power operation and environmental changes.

[0026] All extracted features are synchronized and aligned according to their timestamps. Since the calculation methods and time granularities of different features vary, they need to be unified to the same time reference. This scheme uses 1 minute as the unified time granularity. For features with a calculation granularity less than 1 minute (e.g., the original sampling of the voltage deviation sequence is 1 second), the average value within each 1-minute window is taken as the representative value for that minute. For features with a calculation granularity greater than 1 minute (e.g., the sliding window for power fluctuation amplitude is 2 minutes), the nearest neighbor interpolation method is used to map them to a 1-minute granularity. After timestamp alignment, a complete set of feature values ​​corresponds to each 1-minute interval.

[0027] The aligned features are mapped to a high-dimensional feature space through matrix recombination. Assuming K features are extracted (including power fluctuation amplitude, voltage deviation, current imbalance, wind speed abrupt change frequency, light intensity gradient, SOC change rate, temperature compensation offset, frequency response delay, etc.), with a time span of T minutes (typically T=60), a K×T dimensional feature matrix is ​​constructed. The i-th row of the matrix represents the time series of the i-th feature at time T, and the j-th column represents the state vector of all K features at time j.

[0028] In step S13, the multidimensional feature matrix is ​​subjected to dimensionality reduction processing, and the key feature subset affecting system operation is determined according to the projection weights of each feature dimension, including: Calculate the covariance matrix of the multidimensional feature matrix, and perform eigenvalue decomposition on the covariance matrix to obtain an eigenvector matrix arranged in descending order of eigenvalues; The loading weights of each feature dimension in the principal component space are calculated based on the eigenvector matrix. The feature dimensions are sorted from largest to smallest according to the load weight, and the features whose cumulative contribution meets the preset ratio or whose load value exceeds the preset weight threshold are selected as the key feature subset. Calculate the covariance matrix of the multidimensional feature matrix, perform eigenvalue decomposition on the covariance matrix, and solve for its eigenvalues ​​and eigenvectors. Arrange the obtained eigenvalues ​​in descending order, and adjust the column order of the eigenvector matrix accordingly, so that the first column corresponds to the largest eigenvalue, the second column to the second largest eigenvalue, and so on. Calculate the loading weights of each feature dimension in the principal component space based on the eigenvector matrix. For a feature, calculate the sum of squared loading weights on the first few principal components (usually selecting the number of principal components with a cumulative variance contribution rate of 85%-95%). The larger the overall loading weight, the stronger the projection intensity of the feature along the main principal component directions, and the stronger its ability to characterize the system's operating state.

[0029] K features are sorted from largest to smallest based on their comprehensive load weight, and a subset of key features is selected according to preset rules. The preset rules include two methods: Method 1 sets a cumulative contribution threshold, starting with the feature with the largest load weight and accumulating until the sum of accumulated load weights reaches a preset threshold (e.g., 80%), at which point the selected features constitute the key feature subset. Method 2 sets a load value threshold and directly selects all features whose comprehensive load weight exceeds this threshold (e.g., 0.15). This scheme prioritizes Method 1. If the number of features selected by Method 1 is too small (e.g., less than 3), Method 2 is used to supplement it, ensuring that the key feature subset contains at least 3-5 features. Based on the operational mechanism analysis of wind-solar-storage energy systems, the key feature subset typically includes features such as power fluctuation amplitude (reflecting power generation fluctuations), energy storage state of charge change rate (reflecting energy storage response speed), light intensity gradient (reflecting environmental disturbances), and frequency response delay (reflecting system response capability).

[0030] In step S14, a clustering algorithm is used to cluster the preset historical operating data within the key feature subset space to determine multiple cluster centers representing normal operating modes, including: Historical operational data is projected onto a feature space composed of the aforementioned key feature subsets to generate an operational data sample set, and an Euclidean distance matrix between the samples is constructed. Based on the Euclidean distance matrix, calculate the local density value of each sample point and its relative distance to higher density sample points; Based on the local density value and the relative distance, sample points with density peak attributes are identified as initial cluster centers. Based on the initial cluster centers, the remaining samples are divided into corresponding operating mode clusters using distance clustering; Calculate the statistical center and distribution radius of each of the aforementioned operating mode clusters to determine multiple cluster centers.

[0031] Specifically, historical operational data is projected onto a feature space composed of a subset of key features. The historical operational data is selected from normal operating periods over the past 3-6 months, excluding abnormal periods such as equipment failures and maintenance downtime. For the subset of key features determined in step S13 (e.g., containing 5 features), the Z-score standardized values ​​of these 5 features at each time point are extracted from the historical data. The 5 feature values ​​at each time point constitute a sample point. Assuming the historical data spans M time points, an operational data sample set containing M sample points is generated. Each sample point can be represented as a 5-dimensional vector (x1, x2, x3, x4, x5), corresponding to the feature values ​​of each dimension in the key feature subset.

[0032] Construct an Euclidean distance matrix between samples based on the running data sample set. Calculate the pairwise Euclidean distance between all sample points, constructing an M×M dimensional symmetric distance matrix D with diagonal elements set to 0. Calculate the local density value of each sample point based on the distance matrix using a truncated kernel method, as shown in the formula: ,in Let i be the Euclidean distance between sample point i and sample point j. χ(x) is the cutoff distance, and χ(x) is the step function (χ=1 when x<0, otherwise χ=0). Cutoff distance The method is to sort all distance values ​​in the distance matrix D from smallest to largest, and select the distance values ​​located at the 1%-2% quantile as... This makes the average of each sample point The neighborhood contains 1%-2% of the total number of sample points. Local density value. This represents the number of samples around sample point i within the cutoff distance range.

[0033] Calculate the relative distance of each sample point to higher-density sample points. For sample points whose density value is not the highest, the relative distance is defined as the minimum Euclidean distance between that point and all sample points with higher density values; for sample points with the highest density value, the relative distance is defined as the maximum Euclidean distance between that point and all other sample points. The relative distance reflects the degree of isolation between a sample point and a higher-density region. Sample points with cluster center properties typically have both high local density (located in a dense region) and a large relative distance (far from other dense regions).

[0034] Initial cluster centers are identified by multiplying the local density value by the relative distance. The decision value for each sample point is calculated by multiplying the local density value by the relative distance. All sample points are sorted from largest to smallest decision value, and the mean μ and standard deviation σ of the decision value sequence are calculated. Sample points whose decision values ​​satisfy γ > μ + 2σ are selected as initial cluster centers. The number of initial cluster centers corresponds to the number of operating mode clusters, typically between 3 and 8, corresponding to different typical operating states of the wind-solar-storage energy system (such as high wind speed and high solar radiation, low wind speed and high solar radiation, and energy storage-discharge-dominated modes).

[0035] The remaining samples are partitioned based on the initial cluster centers using distance-based clustering rules. For each non-cluster center sample point, the Euclidean distance to each initial cluster center is calculated, and the sample point is assigned to the operating pattern cluster corresponding to the nearest cluster center. After traversing all remaining sample points to complete the partitioning, each operating pattern cluster contains one cluster center and several member sample points. If the cluster containing a certain cluster center has too few samples (e.g., less than 3% of the total number of samples), the cluster center is considered a noise point and is removed, and its member samples are reassigned to the next nearest cluster center cluster.

[0036] Calculate the statistical center and distribution radius of each operating mode cluster to determine the cluster center and feature distribution boundary. The statistical center is obtained by calculating the mean value of all sample points within the cluster across each feature dimension. The statistical center more accurately represents the overall characteristics of the cluster than the initial cluster center. The distribution radius describes the dispersion of samples within the cluster and is calculated as the 90th quantile of the distances from all sample points within the cluster to the statistical center; that is, 90% of the sample points are less than the distribution radius from the statistical center. The feature distribution boundary is defined as a hypersphere with the statistical center as its center and the distribution radius as its radius. Operating states falling within this boundary are considered to belong to the normal operating mode.

[0037] In step S15, the real-time running status trajectory of the key feature subset is obtained, the health status deviation between the real-time running status trajectory and the cluster center is calculated, and an alarm trigger sequence is determined based on the continuity of the health status deviation exceeding a preset deviation threshold, including: Real-time acquisition of the operational data corresponding to the key feature subset, and construction of a real-time operational state trajectory representing the multidimensional dynamic evolution trend of the system using a time sliding window; Calculate the minimum Euclidean distance between the real-time running status trajectory and each of the cluster centers, and use it as the health status deviation degree to generate a real-time deviation distance sequence; If the values ​​in the real-time deviation distance sequence continuously exceed the preset deviation threshold, the corresponding sample points are determined as abnormal state points and aggregated to form a potential health state deviation set. Based on the deviation magnitude and duration of each abnormal state point in the potential health status deviation set, the corresponding risk level is matched and arranged in chronological order to generate an alarm trigger sequence.

[0038] Specifically, the runtime data corresponding to the key feature subset is acquired in real time. Following the feature extraction methods determined in steps S12 and S13, the real-time values ​​of each key feature are calculated from the preprocessed dataset at the current moment. Assuming the key feature subset contains 5 features, a 5-dimensional feature vector is obtained at each moment. A real-time runtime trajectory is constructed using a time-sliding window. The sliding window length is set to L time points (typically L = 10-30, corresponding to 10-30 minutes), and the window slides forward in steps of 1 time point. At each window position, feature vectors from the L time points within the window are extracted and concatenated in chronological order to form a 5×L-dimensional trajectory matrix. Each row of this matrix represents the evolution trend of a feature over L time points, and each column represents a 5-dimensional state snapshot at a certain moment, capable of capturing the directional and continuous characteristics of parameter changes.

[0039] The minimum Euclidean distance between the real-time operating trajectory and each cluster center is calculated as the health status deviation. For the 5-dimensional feature vector at the current time t, the Euclidean distance between it and the C cluster centers (statistical centers) determined in step S14 is calculated. The minimum value among the C distance values ​​is selected as the health status deviation at that time. The smaller the deviation, the closer the system is to a certain normal mode; the larger the deviation, the further the system deviates from all normal modes, and there may be an anomaly or failure risk. The health status deviation is continuously calculated at each time point to generate a real-time deviation distance sequence.

[0040] A preset deviation threshold is set to determine abnormal states. This threshold is determined based on the distribution radius of each cluster center, typically 1.2-1.5 times the maximum or average value of all cluster center distribution radii. The real-time deviation distance sequence is analyzed point-by-point. If the deviation at a given moment exceeds the preset deviation threshold, that moment is marked as a suspected abnormal point. A continuity condition is further checked. If the deviation at N consecutive time points (N is usually set to 3-5) exceeds the threshold, these N time points are identified as abnormal state points, forming an abnormal state interval. Setting a continuity condition prevents occasional short-term fluctuations from being mistakenly identified as abnormal. All abnormal state points that meet the continuity condition are aggregated to form a potential health state deviation set. Each element in the set contains the start and end times of the abnormal interval, as well as the deviation values ​​at each time point within that interval.

[0041] Risk levels are matched based on the deviation amplitude and duration of each abnormal state point in the potential health status deviation set. Deviation amplitude is quantified by the ratio of the average deviation within the abnormal interval to a preset deviation threshold; a larger ratio indicates a more severe deviation. Duration refers to the length of the abnormal interval; a longer duration indicates a stronger persistence of the abnormality. Three risk levels are defined: low risk, medium risk, and high risk. A risk level determination matrix is ​​established: low risk is defined as deviation amplitude ≤ 1.5 and duration ≤ 10 minutes; medium risk is defined as 1.5 < deviation amplitude ≤ 2.5 or 10 < duration ≤ 20 minutes; and high risk is defined as deviation amplitude > 2.5 or duration > 20 minutes. The risk level classification is based on the degradation mechanism of system health status: deviation amplitude reflects the severity of the system's deviation from normal operating mode; an amplitude exceeding 1.5 times the threshold indicates that system parameters have significantly deviated from the normal range and may enter a performance degradation stage; an amplitude exceeding 2.5 times the threshold indicates a severely abnormal system operating state, with a high risk of equipment failure or process runaway. The duration reflects the stability of the abnormal state. Short-term fluctuations (≤10 minutes) may originate from operating condition adjustments or environmental disturbances, and the system has self-recovery capabilities. Medium-duration deviations (10-20 minutes) indicate weakened system adjustment capabilities, requiring manual intervention. Long-duration deviations (>20 minutes) indicate that the system can no longer recover to normal operation on its own, and immediate corrective measures must be taken to avoid equipment damage or production accidents. This judgment matrix incorporates statistical analysis of historical failure cases to ensure that the risk level matches the actual failure development pattern. Each abnormal interval in the potential health state deviation set is matched with its corresponding risk level according to the judgment matrix.

[0042] Anomalies are arranged chronologically, along with their risk levels, to generate an alarm trigger sequence. This sequence is a time-ordered list, where each entry includes the start and end times, risk level, deviation magnitude, and duration of the anomaly interval. The sequence is ordered from earliest to latest start time, allowing operations personnel to analyze and handle anomalies in chronological order. For high-risk alarms, the system triggers real-time notifications; for medium- and low-risk alarms, the system logs them for periodic review by operations personnel.

[0043] In step S16, a directed graph of parameter interactions is constructed based on the alarm trigger sequence. A reverse search is performed on this graph using a graph traversal algorithm to obtain the anomaly source identifier, including: The alarm trigger sequence is analyzed, the abnormal parameter dimensions that trigger the alarm and their corresponding temporal correlations are extracted, and the abnormal parameter dimensions are mapped to a set of candidate nodes; Using the candidate node set as the core and each parameter in the key feature subset as a node, a parameter interaction directed graph is constructed based on the power grid physical topology and the causal influence logic between parameters. Extract the node transition trajectories from historical operating states, construct the state transition probability matrix between adjacent nodes, and use the eigenvectors of the state transition probability matrix to determine the propagation weights of each edge, and update the parameter interaction directed graph. Using the first abnormal parameter node in the alarm triggering sequence as the search starting point, perform a reverse breadth-first search in the updated directed graph of parameter interactions to calculate the cumulative impact intensity of each backtracking path; The source node pointed to by the backtracking path with the highest cumulative impact intensity is identified as the anomaly source identifier. Specifically, the alarm trigger sequence is analyzed to extract the anomaly parameter dimensions and temporal correlation of the triggering alarm. Each anomaly interval in the alarm trigger sequence corresponds to a deviation of a set of key features. By comparing the deviation of each feature value within the interval with its corresponding cluster center, the parameter dimension with the most significant deviation is identified as the anomaly parameter. The significance of the deviation is quantified by calculating the normalized deviation of each parameter, i.e., subtracting the mean of the parameter dimension in each cluster center from the real-time value of each parameter, and then dividing by the standard deviation of the parameter dimension in the cluster center. The parameter with the largest deviation significance value among all parameters is selected as the dominant anomaly parameter at that moment. Within the anomaly interval, the frequency of each parameter becoming the dominant anomaly parameter is counted, and the 1-2 parameters with the highest frequency are the anomaly parameter dimensions of that anomaly interval. These identified anomaly parameter dimensions are mapped to a set of candidate nodes, with each candidate node corresponding to an anomaly parameter that triggers the alarm. The temporal correlation is determined by analyzing the time sequence of the deviations of each parameter within the abnormal interval. If the deviation of parameter A occurs earlier than the deviation of parameter B, and the time difference is within the preset correlation window (e.g., 3-5 minutes), then parameter A and parameter B are considered to have a temporal correlation and are recorded as a temporal correlation pair.

[0044] Using a candidate node set as the core, a directed graph of parameter interactions is constructed for each node using parameters from a key feature subset. The directed edges between nodes are established based on the physical topology of the wind-solar-storage energy system and the causal influence logic between parameters. The physical topology reflects the direction of energy flow and control signal transmission; for example, light intensity affects photovoltaic power generation, photovoltaic power generation affects energy storage charging and discharging power, and energy storage charging and discharging affects grid voltage. The rule for establishing directed edges is that if a change in parameter A directly causes a change in parameter B, then a directed edge is established from node A to node B, and the direction of the edge represents the direction of influence propagation. During graph construction, edges related to the candidate node set are prioritized, i.e., the inbound edges (edges pointing to candidate nodes) and outbound edges (edges originating from candidate nodes) are constructed to ensure efficient tracing of anomaly sources during subsequent reverse search. The initially constructed directed graph is a topological graph, and no weight values ​​are assigned to the edges.

[0045] Extract node transition trajectories from historical operating states and construct a state transition probability matrix between adjacent nodes. Node transition trajectories represent the propagation path of parameter anomalies between different nodes during system operation; extract anomaly propagation sequences from historical anomaly event records. Statistically analyze the transition trajectories of all historical anomaly events and calculate the conditional probability of transitioning from node i to node j. Construct the state transition probability matrix.

[0046] The propagation weights of each edge are determined using the eigenvectors of the state transition probability matrix. The principal eigenvector of the probability matrix P (the eigenvector whose eigenvalue is closest to 1) is calculated; each component of this eigenvector reflects the steady-state importance of each node in long-term state transitions. The propagation weight of the edge from node i to node j in the directed graph is obtained by multiplying the transition probability of that edge by the steady-state importance of the target node. This weight comprehensively considers both the transition probability and the importance of the target node. By traversing all directed edges in the directed graph and assigning a propagation weight to each edge, the parameter-interactive directed graph is updated.

[0047] The first anomalous parameter node in the alarm trigger sequence is used as the starting point for the search. A reverse breadth-first search is performed in the updated directed graph of parameter interactions. The reverse search refers to searching in the opposite direction of the directed edges, tracing back from the anomalous node to its possible source nodes. The breadth-first search unfolds layer by layer according to the distance from the starting point, recording all paths back to each node during the search. For each backtracking path, its cumulative influence strength is calculated, specifically the product of the weights of all edges along the path.

[0048] Compare the cumulative impact strength of all backtracking paths and select the path with the highest cumulative impact strength. The source node of this path is the anomaly source identifier. The anomaly source identifier represents the parameters corresponding to the most likely root cause of the current anomalous state. If there are multiple paths with the highest cumulative impact strength, the relative difference between the cumulative impact strength of each path and the maximum value is used for discrimination. When the relative difference is less than 10%, the impact strength of multiple paths is considered to be equivalent. In this case, the shortest path is selected, as a shorter path indicates a more direct causal relationship. If the cumulative impact strength and length of multiple paths are the same, the path with the highest corresponding component in the principal feature vector of the source node is selected. This component value reflects the steady-state importance of the node in long-term state transitions.

[0049] In step S17, historical anomaly log sequences matching the anomaly source identifier are retrieved, and parameter decay coefficients and residual feedback weights characterizing the anomaly evolution are extracted from them, including: Based on the anomaly source identifier, relevant historical operation records are retrieved from the historical anomaly log database and divided into multiple historical time window slices according to a preset duration. Extract the real-time feature sequence related to the anomaly source identifier at the current moment, calculate the similarity distance between the real-time feature sequence and each historical time window slice, and select the historical matching sub-sequence with the highest similarity; Curve fitting is performed on the historical matching subsequences to extract the attenuation coefficient, a parameter characterizing the abnormal recovery rate; Calculate the residual error sequence between the historical matching subsequence and the fitted curve, and determine the residual feedback weight based on the variance distribution of the residual error sequence.

[0050] Specifically, relevant historical operational records are retrieved from the historical anomaly log database based on the anomaly source identifier. This database stores records of all anomaly events that occurred within the past 6-12 months. Each record includes information such as the time of the anomaly occurrence, the anomaly parameter identifier, the duration of the anomaly, the parameter change trajectory, and the final recovery status. Based on the anomaly source identifier, all historical events in the database originating from that parameter are retrieved, typically yielding 20-50 relevant records. For each historical record, complete operational data is extracted from 30 minutes before the anomaly's onset to 60 minutes after its end. This time span includes the normal state before the anomaly, the anomaly's development process, and the recovery process after the anomaly was resolved.

[0051] Historical operation records are divided into multiple historical time window slices according to a preset duration. The preset duration is determined based on the time scale of anomaly evolution, typically set to 15 minutes for wind, solar, and energy storage systems. Each time window slice contains 15 data points (equivalent to a 1-minute aggregate of 900 original sampling points at a sampling frequency of 1Hz). For a 90-minute historical operation record, it is divided into 6 time window slices, corresponding to the pre-anomaly stage, anomaly occurrence stage, anomaly development stage, anomaly peak stage, anomaly decay stage, and recovery stabilization stage. Time series data of all parameters in the key feature subset are extracted from each time window slice, forming a K×15 dimensional slice matrix (K being the number of key features). Each row of the slice matrix represents the evolution trajectory of a parameter within that time window. All historical event time window slices are aggregated to form a historical time window slice set, typically ranging from 120 to 300 slices.

[0052] Extract the real-time feature sequence related to the anomaly source identifier at the current moment. Starting from the start time of the anomaly interval identified in step S15, extract real-time running data of the same duration (15 minutes) as the historical slices, and construct a K×15 dimensional real-time feature sequence matrix. Calculate the similarity distance between the real-time feature sequence and each slice in the historical time window slice set, using the Dynamic Time Warping (DTW) distance metric. For the real-time sequence matrix X and the historical slice matrix Y, first perform row-by-row DTW distance calculations on both to obtain the DTW distances for each of the K parameters. Then, take a weighted average of the K distances to obtain the comprehensive similarity distance. The weights are determined based on the influence weight of each parameter at the anomaly source node (the anomaly source parameter has the highest weight, set to 0.4-0.5, and the weights of other parameters are evenly distributed). Traverse the historical slice set to calculate all similarity distances, and select the top 3 historical slices with the smallest distance values ​​as candidates. Further comparison of the recovery characteristics of the complete historical events corresponding to these three slices was conducted. The time required for the anomaly source parameter to return from its peak to the normal range (distance from the cluster center less than the distribution radius) was calculated as the recovery time indicator. At the same time, whether the parameter exceeded the normal range again within 30 minutes after recovery was counted as a stability indicator. The slice corresponding to the historical event with the shortest recovery time and no secondary anomalies after recovery was selected as the historical matching subsequence.

[0053] Curve fitting was performed on historical matching subsequences to extract the decay coefficient, a parameter characterizing the anomaly recovery rate. The time series of the anomaly source parameter during the recovery phase was extracted, showing a trend of gradually decaying from the peak back to normal values. An exponential decay model was used for fitting, with the fitting function being... Let y(t) be the parameter value at time t, A be the initial deviation amplitude, λ be the attenuation coefficient, and B be the steady-state value (corresponding to the parameter value under normal operating conditions). The optimal values ​​of parameters A, λ, and B are determined by least squares optimization. The Levenberg-Marquardt algorithm or trust region algorithm can be used for nonlinear parameter estimation, and the fitting quality is evaluated using indicators such as R² value (usually required to be >0.85) and residual analysis. The attenuation coefficient λ reflects the rate of anomaly recovery; a larger λ value indicates a faster return to normal parameters. It is usually necessary to set reasonable initial parameter values ​​and constraints (e.g., λ>0, A>0) to ensure that the optimization converges to a physically reasonable solution, and to retain the fitting confidence interval to quantify the uncertainty of parameter estimation.

[0054] Calculate the residual error sequence between the historical matching subsequence and the fitted curve. The residual error sequence is defined as the sequence of differences between the actual observed values ​​and the fitted values. The residual sequence reflects the random fluctuations and systematic biases present in the actual anomaly evolution process. Calculate the variance σ² of the residual error sequence. A larger variance indicates higher prediction uncertainty of the fitted model, requiring higher feedback weights to be assigned to the residuals to correct the prediction results. Residual feedback weights Defined as , where σ²0 is a normalization constant (usually set to the median of the residual variances of all historical matching sequences). This formula makes the feedback weight closer to 1 when the residual variance is larger, and the weight closer to 0 when the variance is smaller.

[0055] In step S18, state variables are constructed based on the key feature subset, and an improved Kalman filter is performed using the parameter attenuation coefficient and the residual feedback weight to obtain a posterior state estimate. The posterior state estimate is then mapped to a risk probability to obtain the anomaly risk probability, including: Construct a state-space model with the aforementioned subset of key features as state variables, predict the state based on the posterior state estimation result of the previous time step, and obtain the prior state prediction for the current time step. The transition matrix of the prior state prediction is corrected using the parameter attenuation coefficient to generate a dynamic prediction state that considers the abnormal evolution trend. Calculate the state residual between the dynamically predicted state and the real-time observed features, adjust the measurement noise covariance according to the residual feedback weight, and perform confidence weighting on the state residual to obtain the reliable state residual. The prediction error covariance matrix is ​​updated based on the adjusted measurement noise covariance, and the filter gain matrix at the current time is calculated. The credible state residual is processed using the filter gain matrix to correct the dynamic prediction state, thereby obtaining the posterior state estimate of the monitoring model. The probability distribution characteristics of the reliable state residuals are statistically analyzed in real time, and the anomaly judgment threshold boundary is adaptively adjusted according to the degree of deviation of the residual distribution within the preset confidence interval. The statistical distance between the posterior state estimate and the cluster center is calculated, and probability mapping is performed in combination with the adaptively adjusted threshold boundary to solve for the probability of abnormal risk.

[0056] Specifically, a state-space model is constructed with a subset of key features as state variables. Assuming the subset contains K parameters, the state vector is defined as x(t) = [x1(t), x2(t), ..., xK(t)]^T, representing the state values ​​of the K parameters at time t. The state-space model consists of two parts: the state transition equation and the observation equation, established according to the standard Kalman filter framework. The observation matrix H(t) is typically the identity matrix I, representing the directly observed state variables. The state transition matrix F(t) is obtained through system identification of historical normal operation data. The system identification process employs the least squares method or subspace identification method, utilizing the state observation data {x(t1), x(t2), ..., x(t)} within a continuous time window. n The values ​​of each element in the transition matrix F are estimated by solving the optimization problem min‖X(t)-F×X(t-1)‖² (where X is the state matrix). For wind-solar-storage energy systems, the sparse structure of the F matrix can be preset by combining physical mechanism knowledge, that is, determining which elements should be zero according to the grid topology and energy flow law, thereby improving the accuracy and interpretability of identification. Based on the posterior state estimation result of the previous time step (t-1), the state prediction is performed according to the prediction steps of standard Kalman filtering to obtain the prior state prediction x(t|t-1) and the prior error covariance matrix P(t|t-1) at the current time step.

[0057] Standard Kalman filtering uses a fixed transfer matrix F(t), assuming the system's dynamic characteristics remain constant. However, the evolution of system parameters changes under anomalous conditions. This technical solution uses the parameter attenuation coefficient λ extracted in step S17 to dynamically correct the transfer matrix. The correction method involves multiplying the elements in the transfer matrix related to the anomaly source parameters by an attenuation factor e^(-λΔt), where Δt is the time interval. The corrected transfer matrix F'(t) is used to recalculate the prior state prediction, resulting in a dynamic predicted state x'(t|t-1) considering the anomaly evolution trend. The state residual between the dynamic predicted state and the real-time observed feature z(t) is calculated. The i-th component of the residual vector reflects the prediction error of the i-th parameter. A larger absolute value of the residual indicates a larger prediction bias, potentially indicating unmodeled disturbances or an aggravated anomaly. The residual feedback weights are obtained based on the residuals obtained in step S17. Adjust the measurement noise covariance matrix using the following formula: A higher residual feedback weight indicates higher measurement uncertainty. The state residuals are then weighted by confidence, and the weighted residuals are... When the residual feedback weight is low (the residual fluctuation of the historical matching sequence is small), the influence weight of the residual is reduced, making the posterior estimation more dependent on the dynamic prediction state; when the residual feedback weight is high, the larger influence of the residual is retained, making the posterior estimation more corrective of prediction bias. Based on this, Substitute the covariance matrix to update the covariance matrix, and calculate the filter gain matrix K(t) using the Kalman gain formula.

[0058] The state residuals are processed using the filter gain matrix to correct the dynamically predicted state and obtain the posterior state estimate. The posterior error covariance matrix is ​​updated to The posterior state estimate integrates historical state evolution patterns, abnormal decay trends, and current observation information, making it more accurate and reliable than simple observations or predictions.

[0059] The probability distribution characteristics of state residuals are statistically analyzed in real time, and the anomaly detection threshold boundary is adaptively adjusted based on the degree of residual distribution shift. State residual sequences within the most recent time window (e.g., 30 minutes) are collected, and the residual mean μy and standard deviation σy are calculated. The residual mean shift is also calculated. Ratio of residual standard deviation ,in The standard deviation of the residuals during historical normal operation is used as the baseline value. The anomaly detection threshold is adjusted based on the degree of residual distribution deviation, using the formula Dth'(t) = Dth × (1 + 0.3 × δμ + 0.2 × δσ), where Dth is the initial deviation threshold set in step S15. The coefficients 0.3 and 0.2 in the formula are reference values ​​and can be determined through optimization using historical data based on the actual system characteristics. The optimization goal is to maintain a high recall rate (approximately 92%) while controlling the false alarm rate at a reasonable level (e.g., below 8%). Recall rate is defined as the proportion of actual anomalies that are successfully detected. The mean deviation δμ reflects systematic bias, and its coefficient is usually slightly higher than the coefficient of the standard deviation ratio δσ to reflect the dominant role of systematic bias in threshold adjustment. This adaptive adjustment appropriately relaxes the threshold to reduce false alarms when systematic bias or increased fluctuations occur in the system, and maintains a stricter threshold to improve anomaly detection sensitivity when the system is running smoothly.

[0060] The statistical distance between the posterior state estimate and the center of the normal operation mode is calculated, and the probability of anomaly risk is solved by probability mapping using an adaptive threshold. The minimum Mahalanobis distance Dm(t) between the posterior state estimate and each cluster center determined in step S14 is calculated. Dm(t) and the adaptive threshold Dth'(t) are then probabilistically mapped using a logistic function, and the probability of anomaly risk is calculated using the following formula: , where α is the steepness parameter (usually taken as 2 - 4). When Dm(t) < Dth'(t), the probability P(t) < 0.5 and the system is in a safe state. When Dm(t) > Dth'(t), the probability P(t) > 0.5 and the system is in an abnormal risk state. When Dm(t) >> Dth'(t), the probability P(t) approaches 1 and the system is in a high-risk abnormal state. The abnormal risk probability provides a continuous quantitative risk assessment index for operation and maintenance decisions. Key information such as the abnormal risk probability, identified abnormal source parameters, and minimum Mahalanobis distance Dm(t) is integrated to generate an abnormal diagnosis report, which is presented to operation and maintenance personnel in a visual manner through the monitoring interface in real time to assist them in quickly judging the system state and taking corresponding maintenance measures. At the same time, the diagnosis results are stored in the historical database for subsequent abnormal pattern analysis and model optimization.

[0061] Although the present invention has been described in detail with general descriptions and specific embodiments above, modifications or improvements can be made based on the present invention, which are obvious to those skilled in the art. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection required by the present invention.

[0062] Referring to Figure 2 , the second embodiment of the present invention provides an operation safety monitoring system for a wind-solar energy storage platform, including: A data preprocessing module, configured to collect real-time operation parameters of the wind-solar energy storage platform, remove noise and outliers through standardization processing, and obtain a preprocessed data set; A feature construction module, configured to extract power operation features and environmental impact features from the preprocessed data set, and construct a multi-dimensional feature matrix through multi-dimensional space mapping; A feature dimensionality reduction module, configured to perform dimensionality reduction processing on the multi-dimensional feature matrix, and determine a key feature subset affecting system operation according to the projection weights of each feature dimension; A mode clustering module, configured to use a clustering algorithm to cluster preset historical operation data in the space of the key feature subset, and determine multiple clustering centers representing normal operation modes; A state monitoring module, configured to obtain the real-time operation state trajectory of the key feature subset, calculate the health state deviation degree between the real-time operation state trajectory and the clustering center, and determine an alarm trigger sequence according to the continuity of the health state deviation degree exceeding a preset deviation threshold; A reverse traceability module, configured to construct a parameter interaction directed graph according to the alarm trigger sequence, perform reverse search in the parameter interaction directed graph through a graph traversal algorithm, and obtain an abnormal source identifier; The evolution characteristic extraction module is used to retrieve historical abnormal log sequences that match the abnormal source identifier, and extract the parameter decay coefficient and residual feedback weight that characterize the abnormal evolution characteristics. The risk quantification module is used to construct state variables based on the key feature subset and perform improved Kalman filtering using the parameter decay coefficient and the residual feedback weight to obtain posterior state estimates. The posterior state estimates are then mapped to risk probabilities to obtain abnormal risk probabilities.

[0063] It should be noted that the wind and solar energy storage platform operation safety monitoring system provided in this embodiment of the invention is used to execute all the process steps of the wind and solar energy storage platform operation safety monitoring method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0064] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a safety monitoring program for a wind-solar energy storage platform. When the processor executes the computer program, it implements various steps in the above-described embodiments of the safety monitoring method for the wind-solar energy storage platform, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the data preprocessing module.

[0065] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0066] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for monitoring the operational safety of a wind-solar energy storage platform, characterized in that, include: Real-time operating parameters of the wind-solar-storage energy platform are collected, and noise and outliers are removed through standardization to obtain a preprocessed dataset. Power operation features and environmental impact features are extracted from the preprocessed dataset, and a multidimensional feature matrix is ​​constructed through multidimensional spatial mapping. The multidimensional feature matrix is ​​subjected to dimensionality reduction processing, and the key feature subset affecting the system operation is determined according to the projection weight of each feature dimension; Clustering algorithms are used to cluster the preset historical operating data within the key feature subset space to determine multiple cluster centers representing normal operating modes; The real-time running status trajectory of the key feature subset is obtained, the health status deviation between the real-time running status trajectory and the cluster center is calculated, and the alarm trigger sequence is determined based on the continuity of the health status deviation exceeding the preset deviation threshold. A directed graph of parameter interaction is constructed based on the alarm trigger sequence. An anomaly source identifier is obtained by performing a reverse search on the directed graph of parameter interaction using a graph traversal algorithm. Retrieve historical anomaly log sequences that match the anomaly source identifier, and extract the attenuation coefficient and residual feedback weight, which characterize the anomaly evolution. State variables are constructed based on the key feature subset, and improved Kalman filtering is performed using the parameter decay coefficient and the residual feedback weight to obtain posterior state estimates. Risk probability mapping is then performed on the posterior state estimates to obtain the abnormal risk probability.

2. The method for monitoring the operational safety of a wind-solar-storage energy platform according to claim 1, wherein the process of collecting real-time operational parameters of the wind-solar-storage energy platform and removing noise and outliers through standardization to obtain a preprocessed dataset includes: The real-time operating parameters of wind, solar and energy storage monitoring points are obtained, and a time-continuous real-time operating sequence is constructed by linear interpolation. Wavelet transform is used to filter out high-frequency noise components in the real-time running sequence, and a smoothed and denoised signal that retains trend characteristics is extracted. Abnormal outliers in the smoothed and denoised signal are removed using statistical discriminant analysis to obtain a corrected data sequence. The corrected data sequence is normalized and mapped to obtain unified dimensional data that eliminates the influence of dimensionality; The unified dimensional data is timestamped and reorganized in a multidimensional format to obtain a preprocessed dataset.

3. The method for monitoring the operational safety of a wind-solar-storage energy platform according to claim 1, wherein extracting power operation features and environmental impact features from the preprocessed dataset and constructing a multidimensional feature matrix through multidimensional spatial mapping includes: Power operation data is extracted from the preprocessed dataset, and power fluctuation amplitude, voltage deviation sequence and current imbalance are calculated as power operation characteristics. Based on the environmental monitoring dimensions in the preprocessed dataset, the frequency of sudden wind speed changes and the gradient of light intensity are calculated as environmental impact characteristics. By combining the timestamps of the power operation characteristics and the environmental impact characteristics, the corresponding energy storage state of charge change rate, temperature compensation offset and frequency response delay are extracted and determined as system response characteristics. The power operation characteristics, environmental impact characteristics, and system response characteristics are synchronized and aligned, and then mapped to a high-dimensional feature space through matrix recombination to construct a multi-dimensional feature matrix.

4. The method for monitoring the operational safety of a wind-solar energy storage platform according to claim 1, wherein the step of dimensionality reduction of the multidimensional feature matrix and determining the key feature subset affecting system operation based on the projection weights of each feature dimension includes: Calculate the covariance matrix of the multidimensional feature matrix, and perform eigenvalue decomposition on the covariance matrix to obtain an eigenvector matrix arranged in descending order of eigenvalues; The loading weights of each feature dimension in the principal component space are calculated based on the eigenvector matrix. The feature dimensions are sorted from largest to smallest according to the load weight, and the features whose cumulative contribution meets the preset ratio or whose load weight exceeds the preset weight threshold are selected as the key feature subset.

5. The method for monitoring the operational safety of a wind-solar-storage energy platform according to claim 1, wherein the step of using a clustering algorithm to cluster preset historical operational data within the key feature subset space to determine multiple cluster centers representing normal operation modes includes: Historical operational data is projected onto a feature space composed of the aforementioned key feature subsets to generate an operational data sample set, and an Euclidean distance matrix between the samples is constructed. Based on the Euclidean distance matrix, calculate the local density value of each sample point and its relative distance to higher density sample points; Based on the local density value and the relative distance, sample points with density peak attributes are identified as initial cluster centers. Based on the initial cluster centers, the remaining samples are divided into corresponding operating mode clusters using distance clustering; Calculate the statistical center and distribution radius of each of the aforementioned operating mode clusters to determine multiple cluster centers.

6. The method for monitoring the operational safety of a wind-solar-storage energy platform according to claim 1, wherein acquiring the real-time operational status trajectory of the key feature subset, calculating the health status deviation between the real-time operational status trajectory and the cluster center, and determining an alarm trigger sequence based on the continuity of the health status deviation exceeding a preset deviation threshold, comprises: Real-time acquisition of the operational data corresponding to the key feature subset, and construction of a real-time operational state trajectory representing the multidimensional dynamic evolution trend of the system using a time sliding window; Calculate the minimum Euclidean distance between the real-time running status trajectory and each of the cluster centers, and use it as the health status deviation degree to generate a real-time deviation distance sequence; If the values ​​in the real-time deviation distance sequence continuously exceed the preset deviation threshold, the corresponding sample points are determined as abnormal state points and aggregated to form a potential health state deviation set. Based on the degree of health status deviation and duration of each abnormal state point in the potential health status deviation set, the corresponding risk level is matched and arranged in chronological order to generate an alarm trigger sequence.

7. The method for monitoring the operational safety of a wind-solar-storage energy platform according to claim 1, wherein constructing a directed parameter interaction graph based on the alarm trigger sequence, and performing a reverse search in the directed parameter interaction graph using a graph traversal algorithm to obtain the anomaly source identifier, includes: The alarm trigger sequence is analyzed, the abnormal parameter dimensions that trigger the alarm and their corresponding temporal correlations are extracted, and the abnormal parameter dimensions are mapped to a set of candidate nodes; Using the candidate node set as the core and each parameter in the key feature subset as a node, a parameter interaction directed graph is constructed based on the power grid physical topology and the causal influence logic between parameters. Extract the node transition trajectories from historical operating states, construct the state transition probability matrix between adjacent nodes, and use the eigenvectors of the state transition probability matrix to determine the propagation weights of each edge, and update the parameter interaction directed graph. Using the first abnormal parameter node in the alarm triggering sequence as the search starting point, perform a reverse breadth-first search in the updated directed graph of parameter interactions to calculate the cumulative impact intensity of each backtracking path; The source node pointed to by the backtracking path with the highest cumulative impact intensity is identified as the abnormal source identifier.

8. The method for monitoring the operational safety of a wind-solar-storage energy platform according to claim 1, wherein retrieving historical anomaly log sequences matching the anomaly source identifier and extracting parameter attenuation coefficients and residual feedback weights characterizing the anomaly evolution features from these sequences includes: Based on the anomaly source identifier, relevant historical operation records are retrieved from the historical anomaly log database and divided into multiple historical time window slices according to a preset duration. Extract the real-time feature sequence related to the anomaly source identifier at the current moment, calculate the similarity between the real-time feature sequence and each historical time window slice, and select the historical matching sub-sequence with the highest similarity; Curve fitting is performed on the historical matching subsequences to extract the attenuation coefficient, a parameter characterizing the abnormal recovery rate; Calculate the residual error sequence between the historical matching subsequence and the fitted curve, and determine the residual feedback weight based on the variance distribution of the residual error sequence.

9. The method for monitoring the operational safety of a wind-solar energy storage platform according to claim 1, wherein the step of constructing state variables based on the key feature subset and performing improved Kalman filtering using the parameter attenuation coefficient and the residual feedback weight to obtain a posterior state estimate, and mapping the posterior state estimate to a risk probability to obtain an anomaly risk probability, includes: Construct a state-space model with the aforementioned subset of key features as state variables, predict the state based on the posterior state estimation result of the previous time step, and obtain the prior state prediction for the current time step. The transition matrix of the prior state prediction is corrected using the parameter attenuation coefficient to generate a dynamic prediction state that considers the abnormal evolution trend. Calculate the state residual between the dynamically predicted state and the real-time observed features, adjust the measurement noise covariance according to the residual feedback weight, and perform confidence weighting on the state residual to obtain the reliable state residual. The prediction error covariance matrix is ​​updated based on the adjusted measurement noise covariance, and the filter gain matrix at the current time is calculated. The credible state residual is processed using the filter gain matrix to correct the dynamic prediction state, thereby obtaining the posterior state estimate of the monitoring model. The probability distribution characteristics of the reliable state residuals are statistically analyzed in real time, and the anomaly judgment threshold boundary is adaptively adjusted according to the degree of deviation of the residual distribution within the preset confidence interval. The statistical distance between the posterior state estimate and the cluster center is calculated, and probability mapping is performed in conjunction with the anomaly detection threshold boundary to solve for the anomaly risk probability.

10. A safety monitoring system for the operation of a wind-solar energy storage platform, characterized in that, include: The data preprocessing module is used to collect real-time operating parameters of the wind and solar energy storage platform, and remove noise and outliers through standardization to obtain a preprocessed dataset. The feature construction module is used to extract power operation features and environmental impact features from the preprocessed dataset, and to construct a multi-dimensional feature matrix through multi-dimensional spatial mapping. The feature dimensionality reduction module is used to perform dimensionality reduction processing on the multidimensional feature matrix and determine the key feature subset that affects the operation of the system based on the projection weight of each feature dimension. The pattern clustering module is used to cluster the preset historical operating data in the key feature subset space using a clustering algorithm to determine multiple cluster centers that represent normal operating modes. The status monitoring module is used to acquire the real-time running status trajectory of the key feature subset, calculate the health status deviation between the real-time running status trajectory and the cluster center, and determine the alarm trigger sequence based on the continuity of the health status deviation exceeding the preset deviation threshold. The reverse tracing module is used to construct a directed graph of parameter interaction based on the alarm trigger sequence, and perform a reverse search in the directed graph of parameter interaction using a graph traversal algorithm to obtain the anomaly source identifier; The evolution characteristic extraction module is used to retrieve historical abnormal log sequences that match the abnormal source identifier, and extract the parameter decay coefficient and residual feedback weight that characterize the abnormal evolution characteristics. The risk quantification module is used to construct state variables based on the key feature subset and perform improved Kalman filtering using the parameter decay coefficient and the residual feedback weight to obtain posterior state estimates. The posterior state estimates are then mapped to risk probabilities to obtain abnormal risk probabilities.

Images