Vector processing method, device and optical storage system for time series data

Abstract

Embodiments of the present application provide a vector processing method, device and optical storage system for time series data, wherein the method comprises: obtaining original time series data corresponding to the optical storage system; automatically cleaning the original time series data to obtain preprocessed data; performing feature extraction on the preprocessed data to determine feature data related to the working condition of the optical storage system; determining fusion features, and the fusion weight is configured to enhance the distinguishability of the normal working condition and the abnormal working condition of the optical storage system in the vector space; obtaining a preconfigured target vector dimension; performing dimension reduction processing on the fusion features according to the target vector dimension, and normalizing the reduced features to obtain standardized vectors with consistent lengths; through the scheme of the present application, the distinguishability of the vectorization processing for different working conditions is improved to adapt to different scene precision requirements.

Description

Technical Field

[0001] This invention relates to the field of energy storage technology, and more specifically, to a vector processing method, apparatus, and photovoltaic-energy storage system for time-series data. Background Technology

[0002] With the widespread deployment of industrial equipment such as portable energy storage and home photovoltaic systems in energy management, intelligent analysis of massive time-series data has become key to improving equipment operation and maintenance efficiency and fault diagnosis capabilities.

[0003] However, time series data processing methods in related technologies typically only use simple statistical features or general text embedding models for vectorization. Their built-in feature extraction schemes are relatively fixed and cannot fully capture the multi-level semantic information in time series data (such as time domain fluctuation patterns, frequency domain periodic features, and deep latent patterns). As a result, the generated vectors have insufficient discriminative power when distinguishing between normal and abnormal operating conditions and cannot adapt to the accuracy requirements of different scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a vector processing method, apparatus, and optical storage system for time-series data, which can solve the problem that the vectors generated for massive time-series data have insufficient distinguishability when differentiating between normal and abnormal operating conditions, and cannot adapt to the accuracy requirements of different scenarios.

[0005] In view of this, an embodiment of the first aspect of the present invention provides a vector processing method for time-series data.

[0006] A second aspect of the present invention provides a vector processing apparatus for time-series data.

[0007] An embodiment of the third aspect of the present invention provides a photoelectric storage system.

[0008] To achieve the above objectives, an embodiment of the first aspect of the present invention provides a vector processing method for time-series data in a photovoltaic-storage system. The photovoltaic-storage system includes at least one of energy storage equipment and photovoltaic equipment. The vector processing method for time-series data includes: acquiring raw time-series data corresponding to the photovoltaic-storage system, the raw time-series data including a time sequence of at least one signal channel of battery temperature, state of charge, charge / discharge power, and grid state; automatically cleaning the raw time-series data to obtain preprocessed data; extracting time-domain features, frequency-domain features, and time-series embedding features from the preprocessed data to determine feature data related to the operating conditions of the photovoltaic-storage system, the feature data including time-domain statistical features, frequency-domain features, and time-series embedding features; weighting and concatenating the time-domain statistical features, frequency-domain features, and time-series embedding features according to preconfigured fusion weights to determine fused features, the fusion weights being configured to enhance the distinguishability between normal and abnormal operating conditions of the photovoltaic-storage system in the vector space; acquiring a preconfigured target vector dimension; performing dimensionality reduction processing on the fused features according to the target vector dimension, and normalizing the dimensionality-reduced features to obtain a standardized vector of uniform length.

[0009] The temporal embedding features are obtained by processing the pre-processed data using a pre-trained temporal neural network embedding model.

[0010] The vector processing method for time-series data provided by this invention is specifically applied to photovoltaic-storage systems. These systems include at least one of energy storage devices and photovoltaic devices. The purpose of this vector processing method is to transform the large amount of multi-channel time-series data generated during the operation of the photovoltaic-storage system into standardized vectors, facilitating subsequent similarity retrieval, fault diagnosis, and intelligent analysis. The entire method begins with data acquisition, and proceeds through cleaning, feature extraction, fusion, dimensionality reduction, and normalization, ultimately outputting standardized vectors suitable for machine learning.

[0011] Understandably, by fully capturing the temporal fluctuation patterns, frequency domain periodic features, and deep latent patterns in time-series data through three-layer feature fusion, the problem of insufficient vector discrimination is solved, and the discrimination of vectorization processing for different working conditions is improved to adapt to the accuracy requirements of different scenarios.

[0012] In some technical solutions, optionally, before acquiring the original time-series data corresponding to the photovoltaic-storage system, the vector processing method for the time-series data further includes: acquiring common operating point data, which is the time-series data of different types of photovoltaic-storage systems under the same operating conditions; jointly fine-tuning the time-series neural network embedding model of the first type of photovoltaic-storage system and the time-series neural network embedding model of the second type of photovoltaic-storage system based on the common operating point data to minimize the vector distribution difference of the time-series neural network embedding models corresponding to different types of photovoltaic-storage systems on the common operating point data; and extracting features from the original time-series data of the first type of photovoltaic-storage system and the second type of photovoltaic-storage system based on the fine-tuned time-series neural network embedding model to ensure that the standardized vectors of different types of photovoltaic-storage systems are in the same vector space.

[0013] In this scheme, by acquiring common operating point data of different types of photovoltaic storage systems under the same operating conditions, the temporal neural network embedding models of the first and second models are jointly fine-tuned to minimize the difference in vector distribution between the two on the common operating point data, so that the standardized vectors generated by different models of systems after fine-tuning are in the same vector space.

[0014] Understandably, by jointly fine-tuning and aligning the vector spaces of different types of optical storage systems, the standardized vectors generated by the new system can be directly retrieved from the historical case library of the old system. The similarity measurement error is controlled within 5%, which solves the problem of incomparable vectors caused by hardware differences between different equipment models. This enables the reuse of historical cases across models and reduces the knowledge transfer cost of fault diagnosis for new equipment models.

[0015] In some technical solutions, the original time-series data may optionally be automatically cleaned to obtain preprocessed data, including: removing glitch values ​​from the original time-series data that exceed the physical feasibility range of the optical storage system to obtain first preprocessed data; performing linear interpolation on the first preprocessed data to obtain second preprocessed data; and resampling the second preprocessed data from different signal channels to a uniform time interval to determine the preprocessed data.

[0016] In this scheme, the first preprocessed data is obtained by removing glitch values ​​that are beyond the physical feasibility range of the optical storage system. The first preprocessed data is then linearly interpolated to obtain the second preprocessed data. Finally, the second preprocessed data from different signal channels are resampled to a uniform time interval to determine the final preprocessed data.

[0017] Understandably, removing glitch values ​​eliminates abnormal data points caused by sensor noise and communication errors, linear interpolation restores the data gaps caused by glitch removal, and resampling solves the time axis misalignment problem caused by inconsistent sampling frequencies of different sensors. This provides clean, continuous, and time-aligned preprocessed data for subsequent feature extraction, improving the accuracy and stability of vectorization processing.

[0018] In some technical solutions, optionally, time-domain feature extraction, frequency-domain feature extraction, and time-series embedding feature extraction are performed on the preprocessed data to determine feature data related to the operating conditions of the photovoltaic-storage system. This includes: performing time-domain statistical analysis on the preprocessed data to determine time-domain statistical features, which at least include mean, variance, standard deviation, skewness, kurtosis, peak value, valley value, rate of change, cumulative value, autocorrelation coefficient, and for the photovoltaic-storage system, the mean, maximum, rate of change of temperature difference between cells, and slope of state of charge change; performing frequency-domain analysis on the preprocessed data to determine frequency-domain features, which at least include main frequency components, spectral energy distribution, main frequency proportion, and power frequency interference signals used to judge sampling anomalies in the photovoltaic-storage system; obtaining a pre-trained time-series neural network embedding model, and inputting the preprocessed data into the time-series neural network embedding model to determine the time-series embedding features; wherein the time-series neural network embedding model is replaceably integrated via a standard model interface.

[0019] In this scheme, time-domain statistical features, frequency-domain features, and time-series embedding features are determined by preprocessing the data. The time-domain statistical features capture the central tendency, dispersion, distribution pattern, and extreme value behavior of the signal. The frequency-domain features reveal periodic components and power frequency interference. The time-series embedding features learn deep latent patterns.

[0020] Understandably, the three-layer feature comprehensively represents time-series data from different dimensions, solving the problem of insufficient semantic representation capability of single feature extraction schemes.

[0021] In some technical solutions, optionally, before inputting the preprocessed data into the temporal neural network embedding model, the vector processing method for the temporal data further includes: obtaining triplet samples, which include anchor samples, positive samples, and negative samples, where positive samples and anchor samples belong to the same type of working condition, and negative samples and anchor samples belong to different types of working conditions; determining a contrastive loss function based on the triplet samples, and updating the parameters of the temporal neural network embedding model through backpropagation, so as to narrow the distance between vectors of the same type of working condition and widen the distance between vectors of different types of working conditions.

[0022] In this scheme, a triplet sample consisting of anchor point samples, positive samples of the same working condition, and negative samples of different working conditions is obtained. The contrast loss function is determined based on the triplet, and the model parameters are updated through backpropagation to narrow the distance between vectors of the same working condition and widen the distance between vectors of different working conditions.

[0023] Understandably, by explicitly optimizing the distance relationships in the vector space through contrastive learning, normal operating condition vectors are clustered together and abnormal operating condition vectors are far apart, which fundamentally improves the vectors' ability to distinguish between normal and abnormal operating conditions. This solves the problem that the general embedding model is insensitive to differences in operating conditions and provides a training-level guarantee for the high discriminativeness of subsequent standardized vectors.

[0024] In some technical solutions, optionally, the pre-configured target vector dimension is obtained, including: determining the target vector dimension according to the application scenario of the optical storage system, which includes real-time stream processing scenario, offline diagnostic scenario and intermediate scenario; wherein, the target vector dimension corresponding to the real-time stream processing scenario is smaller than the target vector dimension corresponding to the intermediate scenario, and the target vector dimension corresponding to the intermediate scenario is smaller than the target vector dimension corresponding to the offline diagnostic scenario.

[0025] In this scheme, the target vector dimension is determined according to the application scenario of the optical storage system. The application scenarios include real-time stream processing scenario, offline diagnostic scenario and intermediate scenario. The target vector dimension corresponding to the real-time stream processing scenario is smaller than the target vector dimension corresponding to the intermediate scenario, and the target vector dimension corresponding to the intermediate scenario is smaller than the target vector dimension corresponding to the offline diagnostic scenario.

[0026] Understandably, by configuring different target vector dimensions according to the application scenario, smaller dimensions are used in real-time stream processing scenarios to reduce computation and retrieval latency and meet millisecond-level response requirements, while larger dimensions are used in offline diagnostic scenarios to retain more information and improve retrieval accuracy. Intermediate scenarios use a compromise dimension to balance speed and accuracy, thereby solving the problem that fixed dimensions cannot adapt to the needs of different scenarios and improving the scenario adaptability and resource utilization efficiency of the vectorized processing system.

[0027] In some technical solutions, optionally, the fusion features are dimensionality reduced according to the target vector dimension, and the dimensionality-reduced features are normalized to obtain standardized vectors of consistent length. This includes: dimensionality reduction of the fusion features according to the target vector dimension to determine the dimensionality-reduced feature vectors; normalization of the dimensionality-reduced feature vectors to obtain candidate vectors of consistent length; quality verification of the candidate vectors, including dimensionality integrity checks and normalization checks; and if the quality verification passes, determining the standardized vectors based on the candidate vectors.

[0028] In this scheme, the fusion features are reduced in dimensionality according to the target vector dimension to obtain a dimensionality-reduced feature vector. The dimensionality-reduced feature vector is then normalized to obtain candidate vectors of consistent length. The candidate vectors are then checked for dimensionality integrity and normalization. Once the checks are passed, the candidate vectors are determined as standardized vectors.

[0029] Understandably, dimensionality reduction compresses high-dimensional fusion features to the target dimension, reducing storage and retrieval computational overhead and mitigating the curse of dimensionality. Normalization eliminates the interference of vector magnitude differences on similarity metrics, ensuring that cosine similarity calculations only reflect directional differences. Dimensionality integrity checks and normalization verification ensure the standardization of output vectors; invalid vectors are intercepted and reprocessed, thus guaranteeing the input quality for downstream retrieval and diagnostic tasks.

[0030] In some technical solutions, optionally, after obtaining standardized vectors of uniform length, the vector processing method for time-series data further includes: acquiring historical sample data of the photovoltaic storage system, including labeled samples of similar and dissimilar operating conditions; determining the aggregation degree of the vector set of similar operating conditions and the separation degree of the vector set of dissimilar operating conditions based on the historical sample data; determining the anomaly discrimination degree based on the aggregation degree and separation degree, wherein the anomaly discrimination degree is the value obtained by dividing the difference between the center distance of dissimilar operating condition vectors and the radius of similar operating condition vectors by the radius of similar operating condition vectors; if the anomaly discrimination degree is not lower than a preset anomaly discrimination threshold, then outputting standardized vectors, wherein the anomaly discrimination threshold is not lower than 95%;

[0031] Otherwise, the original time series data with an anomaly discrimination score lower than the anomaly discrimination threshold will be re-sent into the automatic cleaning process, and the feature extraction parameters or fusion weights will be adjusted. The time domain feature extraction to normalization step will be performed again until the anomaly discrimination score of the standardized vector is not lower than the anomaly discrimination threshold, or the preset maximum number of retries is reached.

[0032] In this scheme, historical sample data of the photovoltaic-storage system is acquired. Anomaly discrimination is determined based on the aggregation degree of similar operating condition vector sets and the separation degree of dissimilar operating condition vector sets. Anomaly discrimination is calculated as the difference between the center distance of dissimilar operating condition vectors and the radius of similar operating condition vectors, divided by the radius of similar operating condition vectors. The anomaly discrimination threshold is no less than 95%. If the discrimination is not lower than the threshold, a standardized vector is output; otherwise, the corresponding original data is re-processed, the feature extraction parameters or fusion weights are adjusted, and all steps are repeated until the threshold is met or the maximum number of retries is reached.

[0033] Understandably, using anomaly differentiation metrics to assess the quality of standardized vectors ensures that the net margin between normal and abnormal conditions in the vector space reaches over 95% of the scatter radius of the same type, guaranteeing the upper limit of subsequent retrieval recall and diagnostic accuracy from the source. For batches that fail to meet the standards, iterative optimization is achieved through automatic parameter adjustment and reprocessing, preventing inferior vectors from polluting downstream systems. Simultaneously, the maximum retry mechanism prevents infinite loops, realizing closed-loop control of vector quality.

[0034] A second aspect of the present invention provides a vector processing apparatus for time-series data, comprising: a data acquisition module for acquiring raw time-series data corresponding to a photovoltaic-storage system, the raw time-series data including a time sequence of at least one signal channel among battery temperature, state of charge, charge / discharge power, and grid state; a data cleaning module for automatically cleaning the raw time-series data to obtain preprocessed data; and a feature extraction module for performing time-domain feature extraction, frequency-domain feature extraction, and time-series embedding feature extraction on the preprocessed data to determine feature data related to the operating conditions of the photovoltaic-storage system, the feature data including time-domain statistical features, frequency-domain features, and other features. The system comprises three modules: a feature fusion module, a target determination module, and a vector processing module. The first module performs dimensionality reduction on the fused features according to the target vector dimension and normalizes the reduced features to obtain standardized vectors of uniform length. The second module consists of a time-domain statistical feature, a frequency domain feature, and a time-domain embedded feature. The first module uses a pre-trained time-domain neural network embedding model to process the pre-processed data. The third module uses a time-domain embedded feature to obtain the fused features based on the target vector dimension.

[0035] An embodiment of the third aspect of this application provides a photovoltaic-storage system, which includes at least one of an energy storage device and a photovoltaic device, and further includes the vector processing device for time-series data described in the second aspect.

[0036] Additional aspects and advantages of the technical solutions of the present invention will become apparent in the following description or may be learned by practice of the invention. Attached Figure Description

[0037] Figure 1 One of the flowcharts of a vector processing method for time-series data according to this application is shown;

[0038] Figure 2 A second flowchart illustrating the vector processing method for time-series data according to this application is shown.

[0039] Figure 3 The third flowchart illustrates the vector processing method for time-series data according to this application;

[0040] Figure 4 The fourth flowchart illustrates the vector processing method for time-series data according to this application;

[0041] Figure 5 The fifth flowchart illustrates the vector processing method for time-series data according to this application;

[0042] Figure 6 A flowchart of the vector processing method for time-series data according to this application is shown in diagram six.

[0043] Figure 7 A schematic block diagram of a vector processing apparatus for time-series data according to this application is shown;

[0044] Figure 8 A schematic block diagram of the optical energy storage system according to this application is shown;

[0045] Figure 9 The diagram shows a three-way feature extraction architecture according to this application;

[0046] Figure 10 A schematic diagram of the time-series data standard vectorization process according to this application is shown.

[0047] Among them, 900: Vector processing device for time series data; 902: Data acquisition module; 904: Data cleaning module; 906: Feature extraction module; 908: Feature fusion module; 910: Target determination module; 912: Vector processing module;

[0048] 1000: Photovoltaic and energy storage system; 1002: Energy storage equipment; 1004: Photovoltaic equipment. Detailed Implementation

[0049] To better understand the above-described objectives, features, and advantages of the embodiments of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0050] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, embodiments of the invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below.

[0051] The vector processing method, apparatus, and optical storage system for time-series data provided in this application will be described in detail below with reference to specific embodiments and application scenarios.

[0052] This embodiment provides a vector processing method for time-series data, used in a photovoltaic-storage system, which includes at least one of energy storage devices and photovoltaic devices, such as... Figure 1 As shown, vector processing methods for time series data include:

[0053] Step S100: Obtain the raw time-series data corresponding to the photovoltaic-storage system. The raw time-series data includes the time sequence of at least one signal channel among battery temperature, state of charge, charge / discharge power, and grid state.

[0054] Step S102: Automatically clean the raw time series data to obtain preprocessed data;

[0055] Step S104: Perform time-domain feature extraction, frequency-domain feature extraction, and time-series embedding feature extraction on the preprocessed data to determine the feature data related to the operating conditions of the photovoltaic-storage system;

[0056] Step S106: The time-domain statistical features, frequency-domain features, and time-series embedding features are weighted and concatenated according to the pre-configured fusion weights to determine the fusion features. The fusion weights are configured to enhance the distinguishability between normal and abnormal operating conditions of the photovoltaic-storage system in the vector space.

[0057] Step S108: Obtain the pre-configured target vector dimension;

[0058] Step S110: Reduce the dimensionality of the fused features according to the target vector dimension, and normalize the reduced features to obtain a standardized vector with consistent length.

[0059] Among them, the temporal embedding features are obtained by processing the pre-processed data using a pre-trained temporal neural network embedding model;

[0060] The feature data includes time-domain statistical features, frequency-domain features, and time-series embedding features.

[0061] The vector processing method for time-series data provided by this invention is specifically applied to photovoltaic-storage systems. These systems include at least one of energy storage devices and photovoltaic devices. The purpose of this vector processing method is to transform the large amount of multi-channel time-series data generated during the operation of the photovoltaic-storage system into standardized vectors, facilitating subsequent similarity retrieval, fault diagnosis, and intelligent analysis. The entire method begins with data acquisition, and proceeds through cleaning, feature extraction, fusion, dimensionality reduction, and normalization, ultimately outputting standardized vectors suitable for machine learning.

[0062] Understandably, by fully capturing the temporal fluctuation patterns, frequency domain periodic features, and deep latent patterns in time-series data through three-layer feature fusion, the problem of insufficient vector discrimination is solved, and the discrimination of vectorization processing for different working conditions is improved to adapt to the accuracy requirements of different scenarios.

[0063] Specifically, a photovoltaic-storage system includes photovoltaic equipment and energy storage equipment. The energy storage equipment includes portable energy storage equipment, balcony photovoltaic energy storage equipment, and home energy storage equipment.

[0064] Portable energy storage devices are mobile energy storage power supplies that integrate a large-capacity battery pack, inverter output, DC fast charging, solar input, and battery management system. Battery capacities typically range from 500Wh to 3000Wh. They support multiple charging methods, including AC power, solar power, and car chargers, and can power devices such as mobile phones, laptops, projectors, car refrigerators, and even induction cookers. They are widely applicable to outdoor camping, road trips, home emergencies, and film production.

[0065] Balcony photovoltaic energy storage systems are designed specifically for users in small spaces such as urban apartments and balconies. They integrate photovoltaic input, energy storage, inverter technology, and intelligent management into a single system, supporting plug-and-play operation and requiring no professional electrician for installation. Basic capacity typically ranges from 1kWh to 4kWh, and can be expanded to over 10kWh as needed. Utilizing lithium iron phosphate battery cells, they allow for remote monitoring and intelligent scheduling via a mobile app, effectively reducing household electricity bills.

[0066] Home energy storage devices are complete household energy management systems, typically consisting of an energy storage inverter, battery packs, and an Energy Management System (EMS), such as a Battery Management System (BMS). Battery capacity generally ranges from 5kWh to 30kWh, supporting three core functions: self-consumption of photovoltaic power generation, peak-valley arbitrage, and off-grid backup power. They possess rapid grid-connected / off-grid switching capabilities, allowing for millisecond-level switching to backup power mode during power outages to ensure the continued operation of critical loads such as refrigerators and air conditioners.

[0067] Acquire raw time-series data corresponding to the photovoltaic-storage system. This raw time-series data comes from various sensors and monitoring units deployed in the photovoltaic-storage system.

[0068] The operating status of a photovoltaic-storage system is described by multiple key physical quantities. Therefore, the raw time-series data includes the time sequence of at least one of the following signal channels: battery temperature, state of charge, charge / discharge power, and grid status.

[0069] For example, battery temperature reflects the thermodynamic state of the energy storage battery and is a core indicator for assessing the risk of thermal runaway. State of charge (SOC) indicates the current percentage of remaining battery capacity and directly affects charging and discharging strategies. Charging and discharging power reflects the instantaneous magnitude and direction of energy flow and is a direct representation of the system's operating mode. Grid status records parameters such as connection status to the external power grid, voltage, and frequency, which is particularly important for grid-connected photovoltaic-energy storage systems.

[0070] These signal channels were chosen because they directly determine the operating conditions and safety boundaries of the photovoltaic energy storage system, and there are significant differences in the performance of these channels between normal and abnormal operating conditions.

[0071] For example, in the early stages of thermal runaway, the battery temperature will rise abnormally rapidly, while the temperature rises gradually during normal charging. The raw time-series data is continuously generated at a fixed sampling frequency, such as once every five seconds or once per minute, forming a multi-channel time series.

[0072] These data may be affected by sensor noise, communication interruptions or electromagnetic interference during the acquisition process, so they need to be cleaned.

[0073] The purpose of automatic cleaning is to remove noise, outliers, and missing values ​​from the original data, ensuring the accuracy and reliability of subsequent feature extraction.

[0074] The cleaning process consists of three steps. The first step is to remove glitch values ​​from the original time-series data that are outside the physical feasible range of the optical storage system.

[0075] Each physical quantity in a photovoltaic-storage system has a reasonable range of values. For example, battery temperature is typically neither below -40 degrees Celsius nor above 150 degrees Celsius, state of charge is limited to between 0% and 100%, charging and discharging power is limited by the rated power of the equipment, and grid voltage also has a specified fluctuation range. Any data point outside these ranges is considered a spike value and is directly discarded.

[0076] The second step is to fill in the missing data points caused by communication interruption or sensor failure using linear interpolation.

[0077] Linear interpolation uses the nearest valid data points before and after the missing point to calculate the estimated value of the missing point according to the time interval ratio, thereby restoring the continuity of the data.

[0078] The third step is to resample the timing data from different signal channels to a uniform time interval.

[0079] Because different sensors may have different sampling frequencies—for example, a temperature sensor samples every five seconds, while a power sensor samples every second—direct analysis can lead to time axis misalignment. Resampling adjusts all channels to the same time base through interpolation or aggregation, for example, unifying them to one data point every five seconds.

[0080] After automatic cleaning, well-organized, continuous, and time-aligned preprocessed data is obtained.

[0081] Preprocessed data undergoes time-domain feature extraction, frequency-domain feature extraction, and time-series embedding feature extraction to determine feature data relevant to the operating conditions of the photovoltaic-storage system. The feature data includes time-domain statistical features, frequency-domain features, and time-series embedding features. These three types of features characterize the essential attributes of time-series data from different perspectives, complementing each other to form a complete feature representation.

[0082] Time-domain statistical characteristics are obtained by performing time-domain statistical analysis on the preprocessed data. Time-domain statistical analysis involves calculating a series of basic statistics for the numerical sequence of each signal channel within a fixed time window.

[0083] These statistics include mean, variance, standard deviation, skewness, kurtosis, peak value, trough value, rate of change, cumulative value, and autocorrelation coefficient. The mean reflects the central tendency of the signal; variance and standard deviation measure the dispersion of the signal; skewness describes the asymmetry of the distribution; kurtosis indicates the sharpness of the distribution; peak value and trough value capture extreme events; rate of change reflects the dynamic speed of the signal's change; cumulative value reflects the total amount over a period of time; and the autocorrelation coefficient measures the correlation between the signal and its lagged version, revealing time dependence.

[0084] To address the unique characteristics of photovoltaic-storage systems, additional time-domain statistical features include the average temperature difference between cells, the maximum temperature difference, the rate of change of temperature difference, and the slope of the state of charge (SCC) change. The average temperature difference between cells calculates the average difference in temperature among multiple cells; the maximum temperature difference records the largest temperature difference; and the rate of change of temperature difference tracks the rate of change of temperature over time. These features are crucial for identifying abnormalities in battery thermal management. The slope of the SCC change reflects whether the charge / discharge rate is abnormal; for example, excessively rapid discharge may indicate an internal short circuit. These time-domain features have clear physical meaning, can be directly understood and verified by engineers, and provide an interpretable supplement to subsequent deep learning features.

[0085] Frequency domain characteristics are obtained by performing frequency domain analysis on the preprocessed data. Frequency domain analysis uses Fast Fourier Transform to transform the time-domain signal from the time domain to the frequency domain, thereby revealing the hidden periodic components in the signal.

[0086] The Fast Fourier Transform (FFT) decomposes a finite-length time-domain sequence into a superposition of sine waves of different frequencies. Each frequency corresponds to a complex coefficient, the amplitude of which represents the intensity of the frequency component. Frequency domain characteristics include the dominant frequency components, spectral energy distribution, and the proportion of the dominant frequency.

[0087] The dominant frequency components refer to the frequency components with the largest amplitudes, and their frequency and amplitude values ​​are extracted. Spectral energy distribution describes the energy proportion of each frequency interval, reflecting the frequency distribution characteristics of the signal energy. The dominant frequency proportion is the proportion of energy at the frequency with the largest amplitude to the total spectral energy, used to measure the strength of the signal's periodicity. Furthermore, frequency domain characteristics are specifically used to detect the presence of power frequency interference signals. Power frequency interference signals are typically 50 Hz or its harmonics. In photovoltaic-storage systems, the presence of power frequency interference often indicates an abnormality in the sampling circuit of the battery management system or poor grounding. Frequency domain characteristics can capture periodic abnormal patterns that are difficult to visually detect in the time domain, such as inverter harmonic oscillations or intermittent fluctuations in battery internal resistance. These patterns may appear as small fluctuations on the time domain curve, but as significant spectral peaks in the frequency domain.

[0088] Temporal embedding features are obtained by processing pre-processed data using a pre-trained temporal neural network embedding model.

[0089] Temporal neural network embedding models are deep learning models specifically designed for time series data. These models are pre-trained on historical data from large-scale photovoltaic storage systems, learning hierarchical semantic representations ranging from micro-level sampling point fluctuations to macro-level trend patterns. After inputting the preprocessed data into the model, it extracts local patterns, short-term dependencies, and long-term trends through forward propagation of a multi-layer neural network, ultimately outputting a fixed-dimensional vector, which is the temporal embedding feature.

[0090] Temporal embedding features contain deep features automatically learned by the model, which can express complex patterns that are difficult to capture by time-domain and frequency-domain features, such as subtle transition patterns under different operating conditions.

[0091] The time-domain statistical features, frequency-domain features, and time-series embedding features are weighted and concatenated according to the pre-configured fusion weights to determine the fusion features.

[0092] The fusion weights are a pre-configured set of values, each corresponding to the contribution of the three features.

[0093] The configuration of fusion weights aims to enhance the distinction between normal and abnormal operating conditions of the photovoltaic-storage system in the vector space.

[0094] In practical implementation, automatic machine learning methods can be used to optimize within a pre-defined weight search space, such as grid search or Bayesian optimization. The optimization objective is to find the optimal weight combination that aggregates similar samples and separates dissimilar samples, using the discriminant metric between normal and abnormal samples. Weighted concatenation involves multiplying the three feature vectors by their respective weights and then concatenating them end-to-end to form a higher-dimensional initial fusion feature.

[0095] The initial fusion features retain the physical interpretability of the time-domain features while incorporating the periodic information of the frequency-domain features and the high-order semantics of the deep learning features, thus comprehensively characterizing the operating status of the photovoltaic storage system.

[0096] The dimension of the target vector is determined based on the application scenario of the photovoltaic storage system.

[0097] For example, real-time stream processing scenarios require rapid response, such as millisecond-level retrieval latency in battery management systems. Therefore, the target vector dimension is set to a low value, such as 64 or 128 dimensions, to reduce computational and storage overhead. Offline diagnostic scenarios prioritize accuracy; for example, when performing deep analysis on historical data, it is necessary to retain as much information as possible to improve retrieval recall. Therefore, the target vector dimension is set to a higher value, such as 512 or 768 dimensions. Intermediate scenarios use a 256-dimensional target vector, balancing speed and accuracy. The configuration of the target vector dimension is achieved by adjusting the parameters of the dimensionality reduction algorithm without changing the model architecture or feature extraction process.

[0098] The fused features are reduced in dimensionality according to the target vector dimension, and then normalized to obtain standardized vectors of uniform length. These reduced feature vectors are called dimensionality-reduced feature vectors. Subsequently, Euclidean norm (L2) normalization is performed on the dimensionality-reduced feature vectors. This involves summing the squares of each element of each vector, taking the square root, to obtain the Euclidean length of the vector, and then dividing each element by the length, making the length of all vectors equal to one.

[0099] The purpose of normalization is to eliminate the influence of differences in vector magnitude on similarity calculation, so that subsequent cosine similarity calculation only reflects similarity in direction.

[0100] After dimensionality reduction and normalization, the candidate vectors still need to be quality checked.

[0101] Quality verification includes dimensionality integrity checks and normalization checks. Dimensionality integrity checks confirm that the dimensions of the candidate vectors are consistent with the dimensions of the target vectors.

[0102] Once the validation is successful, the candidate vector is determined as a standardized vector, which can be used for downstream retrieval, clustering, or classification tasks.

[0103] For example, for a portable energy storage device, its battery management system reports a set of data every five seconds, including the temperatures of four battery cells and the ambient temperature. After automatic cleaning, the time-domain statistical features within the past 30-minute window are extracted, including 15 features such as the mean, variance, maximum, minimum, and mean temperature difference between cells for each cell; the frequency-domain features are obtained through Fast Fourier Transform (FFT) to obtain 10 features such as the amplitude of the first five main frequency components and the spectral energy concentration; the time-series embedding features are output as a 128-dimensional vector by the Time-Series Neural Network Embedding Model (TS2Vec). The fusion weights are set to 0.3 for the time domain, 0.2 for the frequency domain, and 0.5 for the embedding. After weighted concatenation, the dimensions are reduced to 128 dimensions through principal component analysis and normalized to obtain the final standardized vector.

[0104] The standardized vectors were then used to retrieve similar historical operating conditions, enabling rapid fault warning.

[0105] Through the above method, this invention transforms the raw time-series data of the photovoltaic storage system into high-quality, highly discriminative standardized vectors, providing a reliable data foundation for subsequent intelligent analysis and decision-making.

[0106] In some embodiments, optionally, standardized vectors generated by the same optical storage system at different time periods are serialized and compared using time alignment and dynamic time warping methods to identify equipment performance degradation trends or intermittent failure modes.

[0107] In some embodiments, optionally, after obtaining the standardized vector, the vector is also indexed and stored according to the device identifier and timestamp of the optical storage system, and an inverted index based on vector similarity is established to accelerate subsequent real-time retrieval and fault matching.

[0108] In some embodiments, the standardized vector is optionally associated with the operating condition label of the photovoltaic-storage system. The operating condition label includes at least one of normal charging, normal discharging, standby, over-temperature warning, insulation fault, and thermal runaway precursor, so that the subsequent supervised learning model can be directly trained using the labeled data.

[0109] In some embodiments, the target vector dimension is optionally dynamically determined based on the current computing resource load of the optical storage system. When the utilization rate of the central processing unit or graphics processor exceeds a preset threshold, the dimension is automatically reduced, and when the resources are idle, the dimension is automatically increased, thereby achieving a real-time balance between computing efficiency and representation accuracy.

[0110] In some embodiments, optionally, such as Figure 2 As shown, before obtaining the original time-series data corresponding to the optical storage system in step S100, the vector processing method for the time-series data further includes:

[0111] Step S10: Obtain common operating point data, which is time-series data of different types of photovoltaic storage systems under the same operating conditions;

[0112] Step S12: Based on the common operating point data, jointly fine-tune the temporal neural network embedding model of the first type of photovoltaic storage system and the temporal neural network embedding model of the second type of photovoltaic storage system to minimize the vector distribution difference of the temporal neural network embedding models corresponding to different types of photovoltaic storage systems on the common operating point data;

[0113] Step S14: Based on the fine-tuned temporal neural network embedding model, feature extraction is performed on the original time-series data of the first type of photovoltaic storage system and the second type of photovoltaic storage system, so that the standardized vectors of different types of photovoltaic storage systems are in the same vector space.

[0114] In this embodiment, by acquiring common operating point data of different types of photovoltaic storage systems under the same operating conditions, the temporal neural network embedding models of the first and second models are jointly fine-tuned to minimize the difference in vector distribution between the two on the common operating point data, so that the standardized vectors generated by the different models of systems after fine-tuning are in the same vector space.

[0115] Understandably, by jointly fine-tuning and aligning the vector spaces of different types of optical storage systems, the standardized vectors generated by the new system can be directly retrieved from the historical case library of the old system. The similarity measurement error is controlled within 5%, which solves the problem of incomparable vectors caused by hardware differences between different equipment models. This enables the reuse of historical cases across models and reduces the knowledge transfer cost of fault diagnosis for new equipment models.

[0116] Specifically, common operating point data is obtained, which refers to time-series data collected from different models of optical storage systems under the same operating conditions.

[0117] Same operating status means that the key parameters affecting the operating conditions of the photovoltaic and energy storage system remain consistent, such as the same state of charge range, the same ambient temperature range, the same charging and discharging power level, or the same grid connection status.

[0118] Due to differences in hardware design, sensor layout, and control strategies, different models of photovoltaic energy storage systems may exhibit systematic deviations in their sensor output time-series data, even under the same physical conditions. By collecting data from common operating points, a common reference benchmark can be provided for different system models.

[0119] The common operating condition data is obtained by placing the first and second model photovoltaic-energy storage systems under the same operating conditions in a laboratory or field environment, and synchronously recording the time sequence of signal channels such as battery temperature, state of charge, charge / discharge power, and grid status. This data is stored separately by model and labeled with corresponding operating condition tags.

[0120] Based on common operating point data, the time-series neural network embedding models of the first type of photovoltaic storage system and the second type of photovoltaic storage system are jointly fine-tuned. At the same time, the parameters of the two models are adjusted so that their output vector distributions on the common operating point data are as consistent as possible.

[0121] In practice, common operating point data are input into the embedding models of the first and second models respectively, resulting in two sets of vectors. A distribution difference metric function, such as the maximum mean difference, is defined to calculate the distribution difference between the two sets of vectors in the feature space.

[0122] The optimization objective of joint fine-tuning is to minimize this distribution difference metric function. Simultaneously, the original contrastive learning loss function can be preserved to ensure the vector's ability to distinguish between different operating conditions. The parameters of both models are updated simultaneously using the backpropagation algorithm, iterating multiple times until the distribution difference converges to below a preset threshold.

[0123] During joint fine-tuning, the parameters of both models can be updated in their entirety, or only the top layer or a specific layer can be updated, in order to balance the alignment effect and the preservation of the original representational ability.

[0124] Based on the fine-tuned temporal neural network embedding model, feature extraction was performed on the original time-series data of the first type of photovoltaic storage system and the second type of photovoltaic storage system, respectively.

[0125] At this point, the two models have undergone joint fine-tuning, and their output vector distributions at common operating points tend to be consistent. For data at non-common operating points, since the model parameters have been adjusted, their output vectors will also converge towards the common space to some extent.

[0126] Therefore, regardless of whether it's the first or second type of optical-storage system, the generated temporal embedded features reside in the same vector space. Subsequent steps such as temporal feature extraction, frequency domain feature extraction, feature fusion, dimensionality reduction, and normalization are all based on the aligned embedded features, and the resulting standardized vectors naturally reside in the same vector space.

[0127] Through the above-described cross-model vector space alignment steps, the standardized vectors of different models of optical storage systems can be directly compared and retrieved in the same vector space.

[0128] For example, when a new model of photovoltaic and energy storage system experiences abnormal operating conditions, the generated standardized vector can be used to search the historical case library of older models to find similar abnormal patterns, thereby quickly locating the cause of the failure. The similarity measurement error is controlled within 5%, ensuring the reliability of cross-model retrieval, avoiding the huge cost of building a separate case library for each model, and realizing the reuse of historical knowledge across models.

[0129] In some embodiments, optionally, for three or more models of optical storage systems, a pairwise joint fine-tuning method is adopted, in which two models are selected for alignment each time, and then the aligned model is used as a reference to align with the next model, until the vector space of all models is unified to the same reference space.

[0130] In some embodiments, the common operating point data may optionally include a combination of different ambient temperature ranges and different state of charge ranges, so that the alignment process covers the main operating area of ​​the photovoltaic-storage system and avoids large differences in vector distribution under other operating conditions due to alignment only under a single operating condition.

[0131] In some embodiments, the aligned vector space may optionally include a model identifier field to distinguish the vector source when searching across models. When vectors of different models appear in the search results, the similarity score is corrected according to the model identifier to compensate for residual systematic bias.

[0132] In some embodiments, optionally, when a new model of optical storage system is added, only a small amount of data on the new model and the existing model at common operating points needs to be collected to fine-tune the embedding model of the new model, while the embedding model of the existing model remains unchanged, thereby supporting incremental vector space expansion.

[0133] In some embodiments, optionally, such as Figure 3 As shown, step S102: Automatically clean the original time series data to obtain preprocessed data, including:

[0134] Step S1020: Remove glitch values ​​from the original time-series data that exceed the physical feasibility range of the optical storage system to obtain the first preprocessed data;

[0135] Step S1022: Perform linear interpolation on the first preprocessed data to obtain the second preprocessed data;

[0136] Step S1024: Resample the second preprocessed data from different signal channels to a uniform time interval to determine the preprocessed data.

[0137] In this embodiment, the first preprocessed data is obtained by removing glitch values ​​that are beyond the physical feasible range of the optical storage system. The first preprocessed data is then linearly interpolated to obtain the second preprocessed data. The second preprocessed data from different signal channels are then resampled to a uniform time interval to finally determine the preprocessed data.

[0138] Understandably, removing glitch values ​​eliminates abnormal data points caused by sensor noise and communication errors, linear interpolation restores the data gaps caused by glitch removal, and resampling solves the time axis misalignment problem caused by inconsistent sampling frequencies of different sensors. This provides clean, continuous, and time-aligned preprocessed data for subsequent feature extraction, improving the accuracy and stability of vectorization processing.

[0139] The purpose of the cleaning step is to eliminate invalid data in the raw data caused by sensor noise, communication interruption, electromagnetic interference or equipment malfunction, and to ensure that the data on which subsequent feature extraction depends is accurate, continuous and time-aligned.

[0140] In a photovoltaic energy storage system, each signal channel has a reasonable range of values ​​for its corresponding physical quantity, which is determined by the physical characteristics of the equipment and its safe operating boundaries.

[0141] For example, the minimum battery temperature is determined by the ambient low-temperature limit and the lower limit of the battery material's tolerance, and is usually not lower than minus forty degrees Celsius; the maximum value is determined by the thermal runaway protection threshold, and is usually not higher than one hundred and fifty degrees Celsius. The physically feasible range of the state of charge is 0% to 100%, and data below 0% or above 100% are considered measurement errors.

[0142] The feasible range of charging and discharging power is determined by the rated power of the inverter and the battery. Data points exceeding the rated power may be due to sensor saturation or communication errors.

[0143] The voltage and frequency of the power grid also have a specified fluctuation range, for example, the voltage can be within plus or minus 10 percent of the rated value.

[0144] These physical feasible ranges can be obtained from equipment manuals, industry standards, or statistical distributions of historical operating data, and configured during system initialization.

[0145] The automatic cleaning program scans the raw time series data channel by channel, identifies any values ​​that exceed the corresponding physical feasible range as glitch values ​​and removes them, and marks the data position corresponding to the removed time point as a void.

[0146] After glitch removal, a null defect appeared in the data sequence. Linear interpolation uses the two most recent valid data points before and after the null defect to calculate the estimated value of the null defect according to the time interval.

[0147] Specifically, let the value of the previous valid data point before the empty defect be V1, the corresponding time be T1, the value of the next valid data point be V2, the corresponding time be T2, and the time of the empty defect be T. Then the interpolation result of the empty defect is V1 plus the difference between V2 and V1 multiplied by the difference between T and T1 divided by the difference between T2 and T1.

[0148] Linear interpolation assumes that the data changes linearly between two valid points. This assumption is reasonable for time-series data from optical storage systems with high sampling frequencies because the changes in physical quantities between adjacent sampling points are usually small. After interpolation, the data sequence is restored to a continuous sequence without gaps, i.e., the second preprocessed data.

[0149] The second preprocessed data from different signal channels is resampled to a uniform time interval to determine the preprocessed data. Different sensors in a photovoltaic energy storage system may have different sampling frequencies; for example, a temperature sensor samples every five seconds, a power sensor samples every second, and the state of charge is updated every ten seconds. This asynchronous data cannot be directly analyzed using multi-channel joint analysis.

[0150] Resampling aligns the data from all channels onto the time grid by selecting a uniform target time interval.

[0151] The principle for selecting the target time interval is to retain the effective information of all channels without generating too much redundant data.

[0152] Typically, the greatest common divisor or least common multiple of the original sampling intervals of all channels is chosen, or it is determined based on the time resolution requirements of the downstream application.

[0153] For example, if the temperature is one point every five seconds and the power is one point every second, then one second can be chosen as the uniform interval. In this case, the temperature channel needs to be supplemented to one point per second by interpolation or forward padding.

[0154] After resampling, all signal channels are fully aligned on the time axis, and each time point contains the synchronization values ​​of all channels, thus obtaining well-organized preprocessed data, which provides standardized input for subsequent feature extraction.

[0155] In some embodiments, the upper and lower limits of each channel in the physically feasible range are optionally read from the equipment specification parameters of the optical storage system and cross-validated with the statistical distribution of historical operating data of the corresponding model. The statistical distribution is obtained by calculating the mean and standard deviation of historical valid data. The intersection of the range obtained by adding or subtracting three times the standard deviation from the mean and the nominal range in the specification is taken as the final physically feasible range to be compatible with the theoretical safety boundary and the actual operating deviation.

[0156] In some embodiments, optionally, in addition to relying on the physically feasible range, the removal of glitch values ​​also introduces local anomaly detection based on a sliding window. Within a preset window length, a standard score is calculated for each data point. The standard score is the difference between the point value and the mean of the values ​​within the window divided by the standard deviation of the values ​​within the window. When the absolute value of the standard score is greater than a preset threshold, the corresponding point is determined as a local glitch and removed. The sliding window length and the standard score threshold are determined by looking up a table based on the sampling frequency and noise level of the optical storage system.

[0157] In some embodiments, optionally, for a signal channel containing multiple cell temperatures, an inter-channel consistency check is inserted between glitch removal and linear interpolation. The deviation between each cell temperature and the average of all cell temperatures is calculated at each time point. When the deviation of a certain channel exceeds a preset temperature difference threshold, the channel time point value is replaced with the average of the remaining normal channels and recorded in the correction count. The temperature difference threshold is set according to the maximum temperature difference allowed by the thermal equalization design within the battery pack, thereby eliminating the contamination of subsequent inter-cell temperature difference characteristic calculations by a single failed temperature sensor.

[0158] In some embodiments, the physical feasible range may optionally support an online update mode. The online update mode continuously updates the mean and standard deviation and refreshes the upper and lower bounds using historical valid data within a rolling time window as samples. The refresh amplitude is limited by the learning rate parameter to avoid the boundary being skewed by a single segment of abnormal data. The learning rate parameter decays inversely with the cumulative running time of the optical storage system, so that the boundary adapts to the actual working conditions more quickly in the early stage of equipment commissioning and tends to stabilize after long-term operation.

[0159] In some embodiments, optionally, such as Figure 4 As shown, step S104: Time-domain feature extraction, frequency-domain feature extraction, and time-series embedding feature extraction are performed on the preprocessed data to determine feature data related to the operating conditions of the photovoltaic-storage system, including:

[0160] Step S1042: Perform time-domain statistical analysis on the preprocessed data to determine the time-domain statistical characteristics. The time-domain statistical characteristics include at least the mean, variance, standard deviation, skewness, kurtosis, peak value, valley value, rate of change, cumulative value, autocorrelation coefficient, as well as the mean, maximum, rate of change of temperature difference between cells, and slope of state of charge change for the photovoltaic-storage system.

[0161] Step S1044: Perform frequency domain analysis on the preprocessed data to determine the frequency domain characteristics. The frequency domain characteristics include at least the main frequency components, spectral energy distribution, main frequency ratio, and power frequency interference signals used to determine sampling anomalies in the photovoltaic-storage system.

[0162] Step S1046: Obtain the pre-trained temporal neural network embedding model, and input the preprocessed data into the temporal neural network embedding model to determine the temporal embedding features;

[0163] The temporal neural network embedding model can be replaced by a standard model interface.

[0164] In this embodiment, time-domain statistical features, frequency-domain features, and time-series embedding features are determined by preprocessing the data. The time-domain statistical features capture the central tendency, dispersion, distribution pattern, and extreme value behavior of the signal. The frequency-domain features reveal periodic components and power frequency interference. The time-series embedding features learn deep latent patterns.

[0165] Understandably, the three-layer feature comprehensively represents time-series data from different dimensions, solving the problem of insufficient semantic representation capability of single feature extraction schemes.

[0166] The features such as the temperature difference between cells and the slope of the state of charge change introduced by the photovoltaic-storage system directly serve the early warning of thermal runaway and the detection of abnormal charging and discharging. The power frequency interference signal is used to diagnose the health status of the sampling circuit of the battery management system, so that the feature extraction is closely related to the actual fault mode of the photovoltaic-storage system.

[0167] Temporal feature extraction is achieved through temporal statistical analysis of the preprocessed data. Temporal statistical analysis involves calculating a series of values ​​describing the statistical characteristics of the numerical sequence for each signal channel within a fixed time window. The mean is the arithmetic average of all sampled values, reflecting the central trend of the signal.

[0168] Variance is the average of the squares of the differences between each sample value and the mean, and standard deviation is the square root of variance. Together, they measure the degree of dispersion of a signal.

[0169] Skewness is the cubic standard moment, which describes the symmetry of the data distribution relative to the normal distribution. Positive values ​​indicate right skewness, and negative values ​​indicate left skewness.

[0170] Kurtosis is a fourth-order standard moment that describes the sharpness of a data distribution, while kurtosis indicates the presence of extreme values ​​in the data.

[0171] Peak values ​​are the maximum values ​​within a window, and valley values ​​are the minimum values, used to capture extreme events.

[0172] The rate of change is the difference between adjacent sampling points divided by the time interval, reflecting the instantaneous speed of change of the signal.

[0173] The cumulative value is the sum of all sampled values ​​within the window, used to measure the total amount.

[0174] The autocorrelation coefficient is the correlation coefficient between a signal and its lagged version, revealing time dependence.

[0175] In response to the unique characteristics of photovoltaic-storage systems, the time-domain statistical features also include the average temperature difference between cells, the maximum temperature difference, the rate of change of temperature difference, and the slope of the change in state of charge.

[0176] The average temperature difference between cells is obtained by calculating the arithmetic mean of the pairwise temperature differences of all cells. The maximum temperature difference is the maximum value of the temperature differences of all cells. The rate of change of temperature difference is the rate of change of temperature difference over time. The slope of change of state of charge is the rate of change of state of charge over time.

[0177] These features were included because the temperature difference between cells and its rate of change directly reflect the uniformity of battery thermal management and the risk of thermal runaway, while the slope of the state of charge change indicates whether the charge and discharge rates are abnormal. For example, excessively fast discharge may indicate an internal short circuit.

[0178] All time-domain statistical features are calculated window by window within a sliding time window. The window length is preset according to the dynamic response time and sampling frequency of the photovoltaic storage system. For example, for temperature data sampled every five seconds, the window length can be set to 360 sampling points corresponding to 30 minutes.

[0179] Frequency domain feature extraction is achieved by performing frequency domain analysis on the preprocessed data.

[0180] Frequency domain analysis uses Fast Fourier Transform to convert the time-domain signal from the time domain to the frequency domain, thus obtaining the signal's spectrum.

[0181] The number of points in a Fast Fourier Transform is determined by the required frequency resolution and signal length. More points result in higher frequency resolution, but also greater computational complexity. Typically, the number of points is an integer power of two and not less than the signal length.

[0182] Frequency domain characteristics include major frequency components, spectral energy distribution, dominant frequency proportion, and power frequency interference signals. Major frequency components refer to the top few frequency components with the largest amplitude in the spectrum. Extracting their frequency and amplitude values ​​reveals the most prominent periodic components of the signal.

[0183] Spectral energy distribution divides the entire frequency range into several intervals and calculates the proportion of the sum of the squares of the spectral amplitudes in each interval to the total energy, which is used to describe the distribution characteristics of energy on the frequency axis.

[0184] The dominant frequency ratio is the proportion of energy corresponding to the maximum amplitude frequency to the total energy of the entire spectrum, used to measure the strength of the signal's periodicity.

[0185] Power frequency interference signals refer to spectral components with a frequency of 50 Hz or 60 Hz. When the amplitude exceeds a preset threshold, it indicates the presence of power frequency interference. The interference usually originates from abnormal sampling circuits or poor grounding in the battery management system. Therefore, it is used as a characteristic to judge the health status of the sampling system.

[0186] Frequency domain characteristics can capture periodic anomalous patterns that are difficult to detect in the time domain. For example, inverter harmonic oscillations are manifested as abnormally increased amplitudes at specific frequencies, and intermittent fluctuations in battery internal resistance are manifested as changes in the spectral energy distribution.

[0187] Temporal embedding feature extraction is achieved through a pre-trained temporal neural network embedding model.

[0188] Temporal neural network embedding models are deep learning models specifically designed for time series data.

[0189] These models were pre-trained on historical data from large-scale photovoltaic and energy storage systems, learning hierarchical semantic representations from micro-sampling point fluctuations to macro-trend patterns.

[0190] When acquiring a temporal neural network embedding model, pre-trained weights are loaded from a model repository. The model is integrated through a standard model interface, which defines a unified input and output format, allowing seamless replacement of models with different structures. Preprocessed data is organized according to the model's requirements, for example, multi-channel temporal data is organized into a three-dimensional tensor with batch, time, and channel dimensions, and then input into the model.

[0191] The model extracts local patterns, short-term dependencies, and long-term trends through forward propagation of a multi-layer neural network, and finally outputs a fixed-dimensional vector, which is the temporal embedded feature.

[0192] Temporal embedding features contain deep features automatically learned by the model, which can express complex patterns that are difficult to capture by time and frequency domain features, such as subtle transition patterns under different operating conditions or early signs of complex faults.

[0193] Because the model is integrated with a standard model interface, when a more advanced temporal embedding model emerges, only the model file needs to be replaced and interface compatibility ensured; no other parts of the feature extraction process need to be modified.

[0194] In some embodiments, optionally, when there are multiple rated power or multiple cell topology versions of the same photovoltaic energy storage system, version identifier embedding or operating condition normalization parameters are added to the front end of the embedding model. The operating condition normalization parameters are the ratio of the average state of charge in the current window to the rated capacity, and the deviation between the ambient temperature and the nominal temperature, in order to compress the stretch of vector space caused by individual differences within the model, so as to make subsequent cross-version retrieval more stable.

[0195] In some embodiments, the time-domain statistical analysis is optionally performed window-by-window within an analysis window that slides along the time axis. The analysis window length and the sliding step size are jointly determined by the dominant dynamic response time scale and the sampling frequency of the photovoltaic-storage system. The dominant dynamic response time scale is taken from the calibration results of the battery thermal model or the step response of the power control loop. The window length is taken as an integer multiple of the number of sampling points corresponding to the time scale, and the sliding step size is no greater than half of the window length to avoid missing short anomalous segments.

[0196] In some embodiments, optionally, a non-uniformity index of temperature spatial distribution is introduced for the temperature difference correlation between cells in the photovoltaic energy storage system. The non-uniformity index is obtained by the ratio of the variance or range of all cell temperatures to the average cell temperature. The ratio is used as a dimensionless quantity to reduce the influence of different ambient temperatures on the absolute value of the benchmark, so that the same characteristic value is still comparable under different seasons and different installed power.

[0197] In some embodiments, optionally, such as Figure 5 As shown, before inputting the preprocessed data into the temporal neural network embedding model in step S1046, the vector processing method for temporal data further includes:

[0198] Step S10442: Obtain triplet samples. Triplet samples include anchor samples, positive samples, and negative samples. Positive samples and anchor samples belong to the same type of working condition, while negative samples and anchor samples belong to different types of working conditions.

[0199] Step S10444: Determine the contrastive loss function based on the triplet samples, and update the parameters of the temporal neural network embedding model through backpropagation to narrow the distance between vectors of the same working condition and widen the distance between vectors of different working conditions.

[0200] In this embodiment, a triplet sample consisting of anchor point samples, positive samples of the same working condition, and negative samples of different working conditions is obtained. The contrast loss function is determined based on the triplet, and the model parameters are updated through backpropagation to narrow the distance between vectors of the same working condition and widen the distance between vectors of different working conditions.

[0201] Understandably, by explicitly optimizing the distance relationships in the vector space through contrastive learning, normal operating condition vectors are clustered together and abnormal operating condition vectors are far apart, which fundamentally improves the vectors' ability to distinguish between normal and abnormal operating conditions. This solves the problem that the general embedding model is insensitive to differences in operating conditions and provides a training-level guarantee for the high discriminativeness of subsequent standardized vectors.

[0202] The purpose of performing model training based on contrastive learning is to explicitly optimize the distance relationship in the vector space so that vectors of the same working conditions are closer to each other and vectors of different working conditions are farther apart, thereby improving the ability of subsequent standardized vectors to distinguish between normal and abnormal working conditions.

[0203] The triplet sample consists of three samples: an anchor sample, a positive sample, and a negative sample.

[0204] An anchor sample is a preprocessed data segment selected as a reference benchmark, while a positive sample is another preprocessed data segment belonging to the same operating condition as the anchor sample, such as both coming from normal charging conditions or both coming from standby conditions.

[0205] Negative samples are preprocessed data segments that belong to different operating conditions than anchor samples. For example, when the anchor sample is a normal charging condition, the negative sample can be a data segment of a thermal runaway precursor condition or an over-temperature warning condition.

[0206] The triplet sample is constructed by sampling from the historical operating data of the photovoltaic storage system based on the operating condition labels.

[0207] Operating condition labels are pre-labeled by maintenance personnel or automated diagnostic systems, with each data segment corresponding to a specific operating condition category.

[0208] For each anchor point sample, a sample is randomly selected from its similar working condition data as a positive sample, and a sample is randomly selected from its dissimilar working condition data as a negative sample.

[0209] The number and sampling ratio of triplet samples are adjusted according to the data abundance of various working conditions. For abnormal working condition categories with few samples, an oversampling strategy is adopted to increase their frequency of occurrence in triplets in order to avoid the model ignoring a few categories.

[0210] The contrastive loss function is determined based on the triplet samples. The contrastive loss function is a metric for learning loss, and its design goal is to impose constraints on the distance relationships in the triplet.

[0211] Specifically, the anchor sample, positive sample, and negative sample are input into the temporal neural network embedding model to obtain three vectors, namely the anchor vector, the positive sample vector, and the negative sample vector.

[0212] Calculate the Euclidean or cosine distance between the anchor vector and the positive sample vector, denoted as the class distance. Calculate the distance between the anchor vector and the negative sample vector, denoted as the out-of-class distance. The contrastive loss function requires that the difference between the out-of-class distance and the class distance be greater than a preset margin parameter.

[0213] When the difference between the distance of different classes and the distance of the same class is less than the margin, the loss function outputs a positive value, indicating that the model needs to be adjusted; when the difference is greater than or equal to the margin, the loss function outputs zero, indicating that the current distance relationship meets the requirements.

[0214] The value of the margin parameter is determined by searching the retrieval precision grid on the validation set. It is usually selected between 0.5 and 1.0. The larger the margin, the stricter the requirement for separation, but it may lead to training difficulties. The smaller the margin, the looser the constraint, but the discrimination may be insufficient.

[0215] Backpropagation is the standard training mechanism for deep learning models. It involves calculating the gradient of the loss function with respect to each trainable parameter in the model, and then updating the parameter values ​​in the opposite direction of the gradient with a certain step size. The step size is controlled by the learning rate, which is initially set to one-thousandth and gradually decays to zero during training using a cosine annealing strategy.

[0216] Each update uses a small batch of triplet samples. The batch size is determined by both the GPU memory capacity and the training speed, typically thirty-two or sixty-four. Through multiple training iterations, the model parameters are gradually adjusted, causing the distance between vectors of the same working condition to continuously decrease, and the distance between vectors of different working conditions to continuously increase.

[0217] Training can be terminated when the discrimination metric on the validation set stops improving for several consecutive rounds, or when the preset maximum number of training rounds is reached.

[0218] Through comparative learning and training, the temporal neural network embedding model learns to cluster according to working condition categories in the vector space, giving the subsequent standardized vectors an inherent discriminative ability.

[0219] In some embodiments, the operating condition labels in the triplet samples are optionally derived from the joint annotation of the operation and maintenance records and automatic protection events of the photovoltaic and energy storage system. The automatic protection events include at least over-temperature protection, over-current protection, insulation impedance alarm and off-grid tripping. When the operation and maintenance records and automatic protection events give the same category to the same data segment, they are marked as a reliable set of the same type. When they are inconsistent, they are classified into a set to be confirmed and their participation weight is reduced or temporarily not used during training. The weight is obtained from the consistency confidence mapping table. The consistency confidence is normalized to between zero and one according to the number of times the two types of sources are triggered in the same time period.

[0220] In some embodiments, optionally, when there are multiple firmware versions or control parameter versions in the optical storage system, the positive samples in the triplet are preferentially taken from the data segments of the same firmware version as the anchor sample. When there are insufficient positive samples of the same version, they are supplemented with aligned cross-version data. The upper limit of the supplementation ratio is controlled by the historical statistical value of the cross-version alignment error. The alignment error is taken as the average cosine distance between the output vectors of the two versions at the same common operating point. The smaller the average cosine distance, the higher the allowable supplementation ratio. The upper limit of the ratio is taken as the monotonically inverse mapping from the alignment error to the ratio and truncated between zero and 0.5 to prevent version drift from masking true anomalies.

[0221] In some embodiments, optionally, obtaining the pre-configured target vector dimension includes:

[0222] The target vector dimension is determined based on the application scenarios of the photovoltaic storage system. The application scenarios include real-time stream processing scenarios, offline diagnostic scenarios, and intermediate scenarios.

[0223] Among them, the target vector dimension corresponding to the real-time stream processing scenario is smaller than the target vector dimension corresponding to the intermediate scenario, and the target vector dimension corresponding to the intermediate scenario is smaller than the target vector dimension corresponding to the offline diagnostic scenario.

[0224] In this embodiment, the target vector dimension is determined according to the application scenario of the optical storage system. The application scenarios include real-time stream processing scenario, offline diagnostic scenario and intermediate scenario. The target vector dimension corresponding to the real-time stream processing scenario is smaller than the target vector dimension corresponding to the intermediate scenario, and the target vector dimension corresponding to the intermediate scenario is smaller than the target vector dimension corresponding to the offline diagnostic scenario.

[0225] Understandably, by configuring different target vector dimensions according to the application scenario, smaller dimensions are used in real-time stream processing scenarios to reduce computation and retrieval latency and meet millisecond-level response requirements, while larger dimensions are used in offline diagnostic scenarios to retain more information and improve retrieval accuracy. Intermediate scenarios use a compromise dimension to balance speed and accuracy, thereby solving the problem that fixed dimensions cannot adapt to the needs of different scenarios and improving the scenario adaptability and resource utilization efficiency of the vectorized processing system.

[0226] In the vector processing method for time series data, the pre-configured target vector dimension is determined based on the application scenario of the optical storage system. The purpose is to enable the dimension of the standardized vector to adapt to the different requirements of the representation accuracy in different scenarios, thereby achieving a balance between resource-constrained real-time scenarios and offline scenarios that pursue accuracy.

[0227] First, it is necessary to clarify that there are a variety of typical application scenarios for photovoltaic energy storage systems in actual operation, mainly including real-time stream processing scenarios, offline diagnostic scenarios, and intermediate scenarios between the two.

[0228] Real-time stream processing scenarios occur when battery management systems or energy management controllers need to monitor device status in real time and issue rapid alarms. For example, during the operation of portable energy storage devices, data generated every five seconds or every minute needs to be immediately converted into vectors and compared with historical anomaly vectors. If the retrieval delay is too high, the fault warning window will be missed.

[0229] Offline diagnostic scenarios occur when maintenance personnel perform batch retrospective analysis of historical operational data, such as fully vectorizing the data of residential photovoltaic inverters from the past six months to build a case knowledge base. In this case, computation time is not the bottleneck, but it is necessary to preserve the integrity of the original information as much as possible to improve the retrieval recall rate.

[0230] The intermediate scenario refers to applications that have certain requirements for response speed but are not so demanding as to be at the millisecond level, while also having high expectations for accuracy, such as daily routine health assessments.

[0231] Based on the characteristics of the above scenarios, the configuration of the target vector dimension follows a basic principle: the target vector dimension is the smallest for real-time stream processing scenarios, the target vector dimension is in the middle for intermediate scenarios, and the target vector dimension is the largest for offline diagnostic scenarios.

[0232] This principle stems from the fact that vector dimension directly affects computational complexity and information retention. Low-dimensional vectors discard more detailed information during dimensionality reduction, but the computational cost of vector inner product operations and nearest neighbor searches is reduced, which is beneficial for meeting latency constraints in real-time scenarios. High-dimensional vectors retain more subtle differences in the original features, making similarity metrics more able to distinguish small differences between operating conditions, but the computational cost also increases significantly, making them more suitable for offline scenarios. An intermediate value is chosen to strike a balance between speed and accuracy.

[0233] For example, in real-time stream processing scenarios, 64-dimensional or 128-dimensional vectors are typically chosen. 64-dimensional vectors are one of the smallest dimensions that many embedded vector retrieval libraries can efficiently handle, while 128-dimensional vectors can provide slightly better discrimination without significantly increasing latency.

[0234] For offline diagnostic scenarios, 512 or 768 dimensions are typically chosen. 512 dimensions are the default output dimension for many pre-trained temporal embedding models, while 768 dimensions can accommodate more detailed information after fusion.

[0235] For intermediate scenarios, 256 dimensions is a common compromise. These values ​​are determined based on factors including the computing power of the hardware platform, the indexing efficiency of the vector database, and the discrimination curves of typical operating conditions.

[0236] After feature fusion is generated, the dimensionality reduction step typically employs principal component analysis, which retains the first few principal components to achieve the specified target dimension.

[0237] The target dimension value serves as an input parameter for principal component analysis, controlling the number of principal components retained.

[0238] When the target dimension changes, only the input parameters need to be modified; there is no need to retrain the model or adjust the feature extraction process. Therefore, the configuration of the target vector dimension is lightweight and flexible, and can be dynamically switched according to the current scenario during system operation.

[0239] Through the above methods, the standardized vector of the optical storage system can flexibly adjust its dimensions according to actual application needs, ensuring both rapid response in real-time scenarios and accurate analysis in offline scenarios, thereby improving the scenario adaptability and resource utilization efficiency of the entire vectorized processing system.

[0240] In some embodiments, the application scenario identification is optionally accomplished by the task type identifier and scheduling priority of the optical storage system. The task type identifier includes at least real-time alarm, daily health assessment, and offline knowledge base construction. The scheduling priority is given by the current load level of the task queue of the operating system or controller. When the same device undertakes multiple types of tasks at the same time, the target vector dimension takes effect according to the scenario with the highest requirement among the currently active tasks, and is hot-updated according to the configuration of the new scenario after the task switch is completed. The hot update only modifies the dimension reduction output dimension and the retrieval index handle, and does not trigger the retraining of the embedded model.

[0241] In some embodiments, optionally, such as Figure 6 As shown, step S110: Dimensionality reduction of the fused features is performed according to the target vector dimension, and the dimensionality-reduced features are normalized to obtain standardized vectors of uniform length, including:

[0242] Step S1102: Perform dimensionality reduction processing on the fused features according to the target vector dimension to determine the dimensionality-reduced feature vector;

[0243] Step S1104: Normalize the dimensionality-reduced feature vectors to obtain candidate vectors of uniform length;

[0244] Step S1106: Perform quality verification on the candidate vectors. Quality verification includes dimensionality integrity check and normalization verification.

[0245] Step S1108: If the quality verification passes, determine the standardized vector based on the candidate vector.

[0246] In this embodiment, the fusion features are reduced in dimensionality according to the target vector dimension to obtain a dimensionality-reduced feature vector. The dimensionality-reduced feature vector is normalized to obtain candidate vectors of consistent length. The candidate vectors are then checked for dimensionality integrity and normalization. After the checks pass, the candidate vectors are determined as standardized vectors.

[0247] Understandably, dimensionality reduction compresses high-dimensional fusion features to the target dimension, reducing storage and retrieval computational overhead and mitigating the curse of dimensionality. Normalization eliminates the interference of vector magnitude differences on similarity metrics, ensuring that cosine similarity calculations only reflect directional differences. Dimensionality integrity checks and normalization verification ensure the standardization of output vectors; invalid vectors are intercepted and reprocessed, thus guaranteeing the input quality for downstream retrieval and diagnostic tasks.

[0248] In the vector processing method for time series data, the fusion features are reduced in dimensionality according to the target vector dimension and the reduced features are normalized to obtain standardized vectors with consistent length. While retaining the main information, the fusion features are compressed to a specified dimension, and the influence of vector magnitude differences on subsequent similarity calculations is eliminated by normalization. At the same time, the standardization of the output vector is ensured by quality verification.

[0249] Specifically, the fusion features are dimensionality-reduced according to the target vector dimension to determine the dimensionality-reduced feature vector. The fusion feature is a high-dimensional vector obtained by weighting and concatenating time-domain statistical features, frequency-domain features, and time-series embedded features according to pre-configured fusion weights. Its original dimension is usually much higher than the target vector dimension.

[0250] The necessity of dimensionality reduction lies in the fact that high-dimensional vectors bring high storage costs and retrieval computation overhead, and there is a lot of redundant information in high-dimensional space. When using high-dimensional vectors directly for similarity retrieval, the resolution of the distance metric will decrease due to the curse of dimensionality.

[0251] Dimensionality reduction is achieved using principal component analysis (PCA). PCA projects the original high-dimensional data onto several orthogonal directions with the largest variance through linear transformation. The number of projection directions is controlled by the dimension of the target vector.

[0252] In practice, the covariance matrix of the fused features over all samples is first calculated. Then, the covariance matrix is ​​decomposed into eigenvalues, and the eigenvalues ​​are sorted from largest to smallest. The eigenvectors corresponding to the k largest eigenvalues ​​are selected as the projection matrix, where k is equal to the dimension of the target vector.

[0253] Multiply each fused feature vector by the projection matrix to obtain a k-dimensional reduced feature vector.

[0254] The acquisition of the projection matrix relies on a batch of representative samples, which are collected from the historical operating data of the photovoltaic storage system, covering common operating conditions to ensure that the projection direction is universal.

[0255] Once the projection matrix is ​​determined, it is fixed. All subsequent fusion features use the same projection matrix for dimensionality reduction to ensure that the geometric relationships of the vector space are consistent after dimensionality reduction.

[0256] Normalizing the dimensionality-reduced feature vectors yields candidate vectors of uniform length. The purpose of normalization is to eliminate the interference of vector magnitude differences on similarity measurement.

[0257] In vector retrieval, the commonly used similarity metric is cosine similarity, which is defined as the cosine of the angle between two vectors. The cosine value depends only on the direction of the vectors and is independent of their magnitude.

[0258] If vectors are not normalized, vectors with larger magnitudes will dominate in certain distance metrics such as Euclidean distance, causing search results to favor vectors with larger magnitudes, which contradicts the original intention of judging the similarity of working conditions.

[0259] The normalization method is L2 normalization, which calculates the Euclidean length of each dimensionality-reduced eigenvector. The Euclidean length is the square root of the sum of the squares of the vector's elements. Then, each element in the vector is divided by the Euclidean length, so that the magnitude of all vectors becomes one.

[0260] After normalization, the length of all candidate vectors is unit length. At this point, the cosine similarity is equivalent to the vector dot product, which is more computationally efficient.

[0261] Dimension integrity checks are used to confirm that the actual dimensions of the candidate vector are consistent with the dimensions of the target vector, preventing dimensional deviations caused by calculation errors or data transmission errors during the dimensionality reduction process.

[0262] Normalization verification is used to confirm whether the error between the L2 norm of the candidate vector and one is within the allowable range. Due to the precision limitations of floating-point operations, the magnitude of the normalized vector may deviate slightly from one. Therefore, an error tolerance threshold is set, usually 10 to the power of -6. Normalization is considered successful when the absolute difference between the magnitude and one is less than the error tolerance threshold. When both verifications pass, the candidate vector is considered successful.

[0263] If the quality verification passes, the standardized vector is determined based on the candidate vector. The standardized vector is the candidate vector that has passed the verification and can be directly output for use by downstream tasks.

[0264] If the quality check fails, the candidate vector will not be adopted. The system will trigger a reprocessing process, which will resend the corresponding original time series data into the cleaning step, adjust the parameters, and then execute the subsequent steps again until the check passes or the maximum number of retries is reached.

[0265] In some embodiments, the dimensionality reduction process can optionally be implemented using a principal component analysis projection matrix. The projection matrix is ​​obtained offline by a batch of representative fusion feature samples covering common operating conditions of the photovoltaic storage system. The calibration process first calculates the sample covariance matrix and performs eigenvalue decomposition, and then arranges the eigenvectors corresponding to the first few largest eigenvalues ​​according to the target vector dimension to form the projection matrix. The projection matrix and the corresponding mean vector are stored together. During online processing, only a lightweight operation of subtracting the mean and multiplying the projection matrix is ​​performed to ensure that the vectors generated in different batches are in the same linear subspace.

[0266] In some embodiments, optionally, after obtaining standardized vectors of uniform length, the vector processing method for time-series data further includes: acquiring historical sample data of the photovoltaic storage system, the historical sample data including labeled samples of similar operating conditions and samples of dissimilar operating conditions; determining the aggregation degree of the vector set of similar operating conditions and the separation degree of the vector set of dissimilar operating conditions based on the historical sample data; determining the anomaly discrimination degree based on the aggregation degree and separation degree, the anomaly discrimination degree being the value obtained by dividing the difference between the center distance of dissimilar operating condition vectors and the radius of similar operating condition vectors by the radius of similar operating condition vectors; if the anomaly discrimination degree is not lower than a preset anomaly discrimination threshold, then outputting a standardized vector, the anomaly discrimination threshold being not lower than 95%;

[0267] Otherwise, the original time series data with an anomaly discrimination score lower than the anomaly discrimination threshold will be re-sent into the automatic cleaning process, and the feature extraction parameters or fusion weights will be adjusted. The time domain feature extraction to normalization step will be performed again until the anomaly discrimination score of the standardized vector is not lower than the anomaly discrimination threshold, or the preset maximum number of retries is reached.

[0268] In this embodiment, historical sample data of the photovoltaic storage system is acquired. The anomaly discrimination degree is determined based on the aggregation degree of similar operating condition vector sets and the separation degree of dissimilar operating condition vector sets. The anomaly discrimination degree is the value obtained by dividing the difference between the center distance of dissimilar operating condition vectors and the radius of similar operating condition vectors by the radius of similar operating condition vectors. The anomaly discrimination threshold is not lower than 95%. If the discrimination degree is not lower than the threshold, a standardized vector is output; otherwise, the corresponding original data is re-sent for cleaning, and the feature extraction parameters or fusion weights are adjusted before all steps are executed again until the standard is met or the maximum number of retries is reached.

[0269] Understandably, using anomaly differentiation metrics to assess the quality of standardized vectors ensures that the net margin between normal and abnormal conditions in the vector space reaches over 95% of the scatter radius of the same type, guaranteeing the upper limit of subsequent retrieval recall and diagnostic accuracy from the source. For batches that fail to meet the standards, iterative optimization is achieved through automatic parameter adjustment and reprocessing, preventing inferior vectors from polluting downstream systems. Simultaneously, the maximum retry mechanism prevents infinite loops, realizing closed-loop control of vector quality.

[0270] After obtaining standardized vectors of consistent length, anomaly discrimination evaluation and quality control steps are performed. The purpose is to quantitatively evaluate the generated standardized vectors and determine whether they have sufficient working condition discrimination capability. Only vectors that pass the evaluation can be output and used, while vectors that fail the evaluation trigger a reprocessing process for iterative optimization.

[0271] Historical sample data is a set of labeled time-series data selected from the past operation records of the photovoltaic storage system. This data has undergone the same cleaning, feature extraction, fusion, dimensionality reduction, and normalization steps as the current processing flow, and has been transformed into standardized vectors, with each vector accompanied by an operating condition label.

[0272] Operating condition labels are created by maintenance personnel or automated diagnostic systems based on the actual equipment status, indicating the specific operating condition category corresponding to the data.

[0273] Samples under the same operating conditions refer to those samples with the same label that belong to normal operating conditions, such as normal charging conditions or normal discharging conditions.

[0274] Anomaly condition samples refer to those labeled as abnormal operating conditions, such as over-temperature warning conditions, thermal runaway precursor conditions, insulation fault conditions, or over-discharge conditions.

[0275] The scale of historical sample data needs to be large enough to ensure statistical significance. Typically, each operating condition category contains at least several hundred samples. The number of samples is determined based on the frequency of occurrence of the operating condition in historical operation. For rare and abnormal operating conditions, the data can be expanded through data augmentation or oversampling.

[0276] Based on historical sample data, the aggregation degree of similar working condition vector sets and the separation degree of dissimilar working condition vector sets are determined. The aggregation degree is used to measure the compactness of similar working condition vectors in the vector space.

[0277] Calculate the geometric center of the set of vectors representing similar working conditions. The geometric center is the mean vector of all vectors representing similar working conditions. Then, calculate the Euclidean distance from each vector representing similar working conditions to the geometric center, and take the maximum value of all distances as the radius of the vector representing similar working conditions.

[0278] A smaller radius indicates that similar vectors are more clustered and have a higher degree of clustering.

[0279] Separation is used to measure the distance between heterogeneous operating condition vectors and homogeneous operating condition vectors.

[0280] Calculate the geometric center of the set of vectors representing different operating conditions, and take the mean vector of all vectors representing different operating conditions. Then calculate the Euclidean distance between the geometric centers of vectors representing the same operating conditions and the geometric centers of vectors representing different operating conditions. This Euclidean distance is called the distance between the centers of vectors representing different operating conditions.

[0281] The greater the center distance, the more separated the two types of vectors are in space, and the higher the degree of separation.

[0282] Anomaly discrimination is a dimensionless comprehensive index used to quantify the degree to which normal operating condition vectors and abnormal operating condition vectors are distinguishable in vector space.

[0283] The formula for calculating the anomaly discrimination index is the difference between the center distance of the vectors of different working conditions and the radius of the vectors of the same working conditions, divided by the radius of the vectors of the same working conditions.

[0284] The numerator of the formula represents the net interval between the two types of vectors, and the denominator normalizes the dispersion of the same type of working conditions, so that the discrimination is not affected by the vector magnitude or the absolute distance scale.

[0285] A higher anomaly discrimination value indicates better spatial separation between normal and abnormal data. The anomaly discrimination threshold is preset to be no less than 95%.

[0286] The anomaly discrimination threshold is set based on the trade-off between false alarm rate and false negative rate in practical applications. 95% means that the net interval between normal operating condition vectors and abnormal operating condition vectors is at least 95% of the same type of dispersion radius. Under this condition, the retrieval recall and diagnostic accuracy can meet the requirements of industrial applications.

[0287] The threshold can be determined by iterating through different values ​​on the validation set and observing the change curve of retrieval accuracy. The critical value that makes the F1 score reach its peak is selected as the threshold.

[0288] If the anomaly discrimination score is not lower than the preset anomaly discrimination threshold, a standardized vector is output. This means that the standardized vectors of the current batch have achieved the expected discrimination ability in spatial distribution and can be used by downstream retrieval or diagnostic systems.

[0289] Otherwise, the original time series data with an anomaly discrimination score lower than the anomaly discrimination threshold will be re-sent into the automatic cleaning process, and the feature extraction parameters or fusion weights will be adjusted. The time domain feature extraction to normalization step will be performed again until the anomaly discrimination score of the standardized vector is not lower than the anomaly discrimination threshold, or the preset maximum number of retries is reached.

[0290] The specific execution method of reprocessing is to locate those original time-series data segments that cause insufficient distinguishability, retrieve these data segments from storage, and re-enter the automatic cleaning step.

[0291] During reprocessing, the parameters that can be adjusted can be the window length or the number of frequency domain transform points in feature extraction, or the allocation ratio of the three features in the fusion weight.

[0292] The direction of adjustment is determined by the analysis of the reasons for insufficient discrimination. For example, if similar vectors are too scattered, the weight of temporal features should be increased appropriately to strengthen common constraints; if dissimilar vectors are too close to similar vectors, the weight of temporal embedding features should be increased appropriately to enhance deep discrimination.

[0293] After each reprocessing, the anomaly discrimination score is recalculated. If the score still does not meet the target, the parameters are adjusted and the process is repeated until the discrimination score meets the target or the preset maximum number of retries is reached. The maximum number of retries is usually set to three or five to prevent infinite loops from consuming computing resources.

[0294] Original data segments that fail to meet the maximum number of retries are marked as non-vectorizable and are then handled manually.

[0295] Through the above-mentioned anomaly discrimination evaluation and reprocessing mechanism, this invention ensures that each batch of output standardized vectors has high discrimination, thereby providing reliable quality assurance for subsequent similarity retrieval and fault diagnosis.

[0296] In one specific embodiment, optionally, the present invention proposes a standardized vectorization and anomaly differentiation method and system for time-series data. The present invention does not merely perform a simple mapping from data to vectors, but rather constructs a vectorized transmission pipeline with multiple configurable schemes, full-process automation, and proactive optimization of anomaly differentiation.

[0297] The system consists of the following core modules:

[0298] Multi-scheme temporal feature extraction module: Supports three major feature extraction schemes simultaneously, which can be used independently or in combination.

[0299] Time-domain statistical feature scheme: Automatically calculates more than 15 statistical features such as mean, variance, standard deviation, skewness, kurtosis, peak value, valley value, rate of change, cumulative value, and autocorrelation;

[0300] Frequency domain feature extraction scheme: Based on Fast Fourier Transform (FFT), extract frequency domain features such as main frequency components, spectral energy distribution, and main frequency proportion;

[0301] Temporal embedding model solution: Supports pluggable integration of dedicated temporal neural network models such as Time Series to Vector (TS2Vec), Patch Time Series Transformer (PatchTST), and Time-series Network (TimeNet) deep learning model for multi-period temporal representation, and automatically learns the latent representation of temporal sequence through deep learning.

[0302] Vectorized Pipeline Automation Engine: Defines a standardized vectorized processing pipeline: raw data, data cleaning, feature extraction (configurable scheme), feature fusion, vectorized encoding, and normalization.

[0303] The entire process is automated, with the raw time series input upstream and a standardized vector output downstream.

[0304] Feature fusion and dimensionality configurable module: This module performs weighted fusion of temporal features, frequency domain features, and deep learning embedded features to generate the final vector. The vector dimension can be configured as needed according to the downstream scenario (e.g., low dimension for real-time scenarios, high dimension for offline diagnosis).

[0305] Anomaly Discrimination Active Optimization Module: Introduces a contrastive learning or metric learning mechanism during vectorization to explicitly narrow the distance between vectors of the same working condition and widen the distance between vectors of different working conditions when training the sequential embedding model, so that normal working condition vectors and abnormal working condition vectors have high discrimination in the vector space.

[0306] Vector quality verification module: Performs quality assessment on the generated vectors, including dimensional integrity check, normalization verification, and anomaly differentiation metric assessment (calculating the aggregation degree of similar categories and the separation degree of dissimilar categories). Batches that fail to meet the standards are automatically reprocessed.

[0307] This invention does not rely on general text embedding, but instead designs a three-layer feature representation system for time-series data:

[0308] The first layer consists of time-domain statistical characteristics: capturing the central tendency (mean), dispersion (variance), distribution pattern (skewness, kurtosis), extreme value behavior (peak, trough), and time dependence (autocorrelation coefficient) of the waveform. These characteristics are physically interpretable and can be directly understood by engineers.

[0309] The second layer, frequency domain features: The time-domain waveform is transformed into the frequency domain using FFT to extract the dominant frequency and spectral energy distribution. This is crucial for identifying periodic anomalies (such as inverter harmonic oscillations and intermittent internal resistance fluctuations in batteries), information that pure time-domain features cannot capture.

[0310] The third layer involves deep learning embeddings: training or fine-tuning specialized temporal embedding models such as TS2Vec. These models are pre-trained on large amounts of temporal data and can automatically learn hierarchical temporal semantics—from pixel-level sampling point fluctuations to mid-level trend patterns and high-level condition types. Unlike general text models, they understand temporal concepts such as "acceleration," "period," and "mutation."

[0311] The vector generated by weighted fusion of the three layers of features retains both statistical interpretability and the high-order semantic representation capabilities of deep learning. For example, the cosine similarity between the temperature curve in the early stage of thermal runaway and the normal charging temperature curve in the fused vector space is only 0.2 (high discriminative power), rather than the 0.7 of the general model (easily confused).

[0312] The three-way feature extraction architecture diagram is as follows: Figure 9As shown, the original time-series input serves as the initial data source for system processing. It is first distributed in parallel to three independent feature extraction branches: time-domain statistical feature extraction, frequency-domain FFT feature extraction, and time-series embedding model extraction, in order to capture the multi-dimensional characteristics of the optical storage system.

[0313] In the time-domain statistical feature extraction branch, the system first calculates the basic statistics, then extracts the distribution morphology features, extreme values ​​and location features in sequence, and finally extracts the time-series structure features. After four layers of progressive processing in this branch, a 17-dimensional time-domain feature vector is finally output.

[0314] In the frequency domain FFT feature extraction branch, the system first obtains basic spectral features, then further extracts spectral statistical features and spectral distribution features, and finally mines sub-band energy features. This branch transforms the signal to the frequency domain through Fast Fourier Transform, ultimately outputting an N / 2+M+10 dimensional frequency domain vector, where N represents the original signal length and M represents the number of manually defined frequency band divisions.

[0315] In the temporal embedding model extraction branch, the system introduces various deep temporal model architectures. First, local temporal patterns are extracted using PatchTST or TimeNet. Second, long-range dependencies are captured using Informer or Reformer. Finally, a temporal contrastive learning encoder enhances the discriminative power of the features. The final output is an embedding vector with 32 to 256 dimensions, with the specific dimensions dynamically adjusted according to the complexity of the selected model.

[0316] Finally, the 17-dimensional time-domain feature vector, the N / 2+M+10-dimensional frequency-domain vector, and the 32-256-dimensional embedding vector output by the three branches are all merged into the feature fusion layer.

[0317] The feature fusion layer concatenates or weights the aforementioned heterogeneous features to output a high-dimensional fusion feature vector. This fusion feature vector comprehensively integrates the signal's time-domain statistical regularity, frequency-domain energy distribution, and deep semantic features, providing a high-quality data foundation for subsequent normalization and anomaly discrimination evaluation.

[0318] This invention incorporates an anomaly discrimination optimization mechanism during the vectorization process:

[0319] Contrastive learning loss function: When fine-tuning embedding models such as TS2Vec, construct triplet samples (anchor samples, positive samples - similar working conditions, negative samples - different working conditions), and explicitly optimize the vector space distribution through contrastive loss.

[0320] Domain-Adaptive Enhancement: For portable energy storage and home photovoltaic scenarios, the model is fine-tuned using historical fault data from each domain to make the model more sensitive to abnormal patterns of specific devices.

[0321] Effectiveness quantification index: Define anomaly discrimination index = (distance between the centers of vectors of different working conditions - radius of vectors of the same working conditions) / radius of vectors of the same working conditions.

[0322] This invention ensures that the anomaly differentiation rate is not less than 95%.

[0323] This invention supports flexible configuration of vector dimensions:

[0324] Real-time stream processing scenario (Flink): Configure as a low-dimensional vector (64 or 128 dimensions) to prioritize retrieval speed, with single retrieval latency controlled within 10ms.

[0325] Offline diagnostic scenarios (Spark): Configure as a high-dimensional vector (512 or 768 dimensions) to retain more information and improve retrieval accuracy.

[0326] Intermediate scenario: Configurable to 256 dimensions, balancing speed and accuracy.

[0327] Dimension configuration is achieved by adjusting feature fusion weights and Principal Component Analysis (PCA) dimensionality reduction parameters, independent of any specific model architecture. Upper-layer business logic can specify the target dimension via the Application Programming Interface (API) parameters.

[0328] A schematic diagram of the time series data standardization vectorization process is shown below. Figure 10 As shown, it includes:

[0329] Step S200: Raw timing input;

[0330] Step S202: Data cleaning;

[0331] Step S204: Extract time-domain statistical features;

[0332] Step S206: Extract frequency domain FFT features;

[0333] Step S208: Input the temporal embedding model;

[0334] Step S210: Feature fusion;

[0335] Step S212: Dimension Configuration;

[0336] Step S214: Normalization;

[0337] Step S216: Quality verification;

[0338] Step S218: Standardize the vector output.

[0339] Specifically, the vectorized pipeline of this invention is fully automated:

[0340] 1. Input: Raw time series data (arbitrary sampling rate, arbitrary number of channels);

[0341] 2. Automatic cleaning: deburring, missing value interpolation, and resampling to a uniform frequency;

[0342] 3. Feature extraction: Time-domain, frequency-domain, and embedding model-based feature extraction are performed in parallel according to the configuration.

[0343] 4. Feature fusion: Weighted splicing according to preset weights;

[0344] 5. Normalization: L2 normalization ensures consistent vector length;

[0345] 6. Quality verification: If it passes, output it; if it fails, reprocess it.

[0346] When adapting to new device types, only the embedded model needs to be replaced or fine-tuned (by plugging and unplugging through the standard model interface), and the time-domain and frequency-domain feature extraction does not need to be changed.

[0347] The entire pipeline requires no manual redesign.

[0348] In one specific embodiment, optionally, in Embodiment 1: the BMS of the portable energy storage device reports a set of temperature data (including the temperatures of 4 battery cells and 1 ambient temperature) every 5 seconds. Each reported temperature snapshot and the sliding window data from the past 30 minutes need to be converted into a vector in real time for use by the subsequent Retrieval-Augmented Generation (RAG) engine.

[0349] Step 1: Data Cleaning

[0350] Flink receives temperature data in real time and performs automatic cleaning: removing glitches that are outside the physically feasible range (-40℃ to 150℃); performing linear interpolation for single-point missing values; and resampling the temperature of the four cells to a uniform 5-second interval.

[0351] Step 2: Feature Extraction (Three-way Parallel):

[0352] Temporal feature extraction (15 dimensions):

[0353] Average, variance, maximum, minimum, and peak-to-valley temperature of each cell; average, maximum, and rate of change of temperature difference between cells; difference between ambient temperature and average cell temperature; slope of temperature change over the past 30 minutes.

[0354] Frequency domain feature extraction (10 dimensions):

[0355] Perform FFT on the temperature difference sequence to extract the amplitude and frequency values ​​of the first 5 main frequency components; spectral energy concentration (energy percentage of the first 3 main frequencies); and check for the presence of 50Hz power frequency interference signals (used to determine BMS sampling anomalies).

[0356] Temporal Embedding Model (128-dimensional, TS2Vec): Input four temperature sequences into the TS2Vec model and output a 128-dimensional latent representation vector.

[0357] Step 3: Feature fusion and dimensionality reduction:

[0358] Since this is a real-time stream processing scenario, the target dimension is configured to be 128. The fusion weights are: time domain 0.3, frequency domain 0.2, and embedding 0.5. After weighted concatenation, the dimensionality is reduced to 128 dimensions using PCA, followed by L2 normalization.

[0359] Step 4: Quality Verification

[0360] Dimension verification: Check if the output vector is 128-dimensional;

[0361] Normalization check: Check if the L2 norm is 1 (error < 1e-6);

[0362] Anomaly discrimination assessment: Take 100 known normal samples and 100 abnormal samples of this model, and calculate the aggregation degree of the same type and the separation degree of different types.

[0363] In this embodiment, the anomaly discrimination rate is 96.2%, which exceeds the 95% threshold, and the verification is passed.

[0364] Step 5: Vector Output:

[0365] Standardized vectors are written to the message queue (Kafka) output topic, and then stored and consumed by the downstream RAG engine and high-performance vector database (Milvus).

[0366] Example 2: In the offline batch vectorization of historical data of home photovoltaic inverters, it is necessary to perform offline batch vectorization of the historical operation data (shadow data stored in MongoDB) of all home photovoltaic inverters in the past 6 months to build a historical case knowledge base.

[0367] The differences between Example 2 and Example 1 include:

[0368] Computing engine: Spark offline batch processing, capable of processing terabytes of historical data;

[0369] Feature scheme: Increase the detail of frequency domain features (increase the number of FFT points from 128 to 1024), use the 512-dimensional version of the embedding model (PatchTST), and aim for a 512-dimensional output.

[0370] Quality verification: After the offline task is completed, a quality assessment is performed on all vectors, and a report is generated. This embodiment processed a total of 230 million records, with a vectorization success rate of 99.97% and an anomaly discrimination evaluation value of 97.1%.

[0371] In one specific embodiment, optionally, when a company launches a new model of energy storage device, it needs to align its vector space with the vector spaces of existing models to facilitate cross-model retrieval of historical cases. This invention achieves this through the following means:

[0372] Data was collected at common operating points for both models, such as the same State of Charge (SOC) range and the same ambient temperature.

[0373] The two embedding models are jointly fine-tuned at these common points to minimize the distribution difference between the two vector spaces at these common points;

[0374] After fine-tuning, the anomaly vectors of the new energy storage device can be used to search for historical cases of existing devices across space, with a similarity measurement error of no more than 5%.

[0375] like Figure 7 As shown in the illustration, this application embodiment also provides a vector processing device 900 for time-series data, including: a data acquisition module 902, used to acquire raw time-series data corresponding to a photovoltaic-storage system, the raw time-series data including a time sequence of at least one signal channel among battery temperature, state of charge, charge / discharge power, and grid state; a data cleaning module 904, used to automatically clean the raw time-series data to obtain preprocessed data; and a feature extraction module 906, used to perform time-domain feature extraction, frequency-domain feature extraction, and time-series embedding feature extraction on the preprocessed data to determine feature data related to the operating conditions of the photovoltaic-storage system, the feature data including time-domain statistical features, frequency-domain features, and other features. The system comprises: a feature fusion module 908, which weights and concatenates time-domain statistical features, frequency-domain features, and time-series embedded features according to pre-configured fusion weights to determine the fused features, wherein the fusion weights are configured to enhance the distinguishability between normal and abnormal operating conditions of the photovoltaic storage system in the vector space; a target determination module 910, which obtains the pre-configured target vector dimension; and a vector processing module 912, which performs dimensionality reduction processing on the fused features according to the target vector dimension and normalizes the dimensionality-reduced features to obtain standardized vectors of consistent length; wherein the time-series embedded features are obtained by processing the pre-processed data using a pre-trained time-series neural network embedding model.

[0376] like Figure 8 As shown in the figure, this application embodiment also provides a photovoltaic energy storage system 1000, which includes at least one of an energy storage device 1002 and a photovoltaic device 1004, and further includes a vector processing device 900 for time-series data.

[0377] The photovoltaic-storage system 1000 includes photovoltaic equipment 1004 and energy storage equipment 1002. The energy storage equipment 1002 includes portable energy storage equipment, balcony photovoltaic energy storage equipment, and home energy storage equipment, etc.

[0378] Portable energy storage devices are mobile energy storage power supplies that integrate a large-capacity battery pack, inverter output, DC fast charging, solar input, and battery management system. Battery capacities typically range from 500Wh to 3000Wh. They support multiple charging methods, including AC power, solar power, and car chargers, and can power devices such as mobile phones, laptops, projectors, car refrigerators, and even induction cookers. They are widely applicable to outdoor camping, road trips, home emergencies, and film production.

[0379] Balcony photovoltaic energy storage systems are designed specifically for users in small spaces such as urban apartments and balconies. They integrate photovoltaic input, energy storage, inverter technology, and intelligent management into a single system, supporting plug-and-play operation and requiring no professional electrician for installation. Basic capacity typically ranges from 1kWh to 4kWh, and can be expanded to over 10kWh as needed. Utilizing lithium iron phosphate battery cells, they allow for remote monitoring and intelligent scheduling via a mobile app, effectively reducing household electricity bills.

[0380] Home energy storage devices are complete household energy management systems, typically consisting of energy storage inverters, battery packs, and energy management systems (BMS, EMS). Battery capacity generally ranges from 5kWh to 30kWh, supporting three core functions: self-consumption of photovoltaic power generation, peak-valley arbitrage, and off-grid backup power. They possess rapid grid-connected / off-grid switching capabilities, allowing for millisecond-level switching to backup power mode during power outages, ensuring the continuous operation of critical loads such as refrigerators and air conditioners.

[0381] In this invention, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance; the term "multiple" refers to two or more unless otherwise explicitly defined. The terms "install," "connect," "link," and "fix" should be interpreted broadly. For example, "connect" can be a fixed connection, a detachable connection, or an integral connection; "link" can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0382] In the description of this invention, it should be understood that the terms "upper," "lower," "left," "right," "front," "rear," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or unit referred to must have a specific orientation or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0383] In the description of this specification, the terms "one embodiment," "some embodiments," "specific embodiment," etc., refer to specific features, structures, materials, or characteristics described in connection with an embodiment or example that are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0384] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. 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.

Claims

1. A vector processing method for time-series data, characterized in that, For a photovoltaic-storage system, the photovoltaic-storage system including at least one of energy storage devices and photovoltaic devices, the vector processing method for the time-series data includes: Obtain the raw time-series data corresponding to the photovoltaic-storage system, wherein the raw time-series data includes the time sequence of at least one signal channel among battery temperature, state of charge, charge / discharge power, and grid state; The original time-series data is automatically cleaned to obtain preprocessed data; The preprocessed data is subjected to time-domain feature extraction, frequency-domain feature extraction and time-series embedding feature extraction to determine feature data related to the operating conditions of the optical storage system. The feature data includes time-domain statistical features, frequency-domain features and time-series embedding features. The time-domain statistical features, frequency-domain features, and time-series embedding features are weighted and concatenated according to the pre-configured fusion weights to determine the fusion features. The fusion weights are configured to enhance the distinguishability between normal and abnormal operating conditions of the photovoltaic-storage system in the vector space. Obtain the pre-configured target vector dimension; The fused features are reduced in dimensionality according to the target vector dimension, and the reduced features are normalized to obtain a standardized vector of uniform length. The temporal embedding features are obtained by processing the preprocessed data using a pre-trained temporal neural network embedding model.

2. The vector processing method for time-series data according to claim 1, characterized in that, Before acquiring the raw time-series data corresponding to the optical storage system, the vector processing method for the time-series data further includes: Acquire common operating condition point data, which is time-series data of different models of photovoltaic storage systems under the same operating conditions; Based on the common operating point data, the temporal neural network embedding model of the first type of photovoltaic storage system and the temporal neural network embedding model of the second type of photovoltaic storage system are jointly fine-tuned to minimize the vector distribution difference of the temporal neural network embedding models corresponding to different types of photovoltaic storage systems on the common operating point data. Based on the fine-tuned temporal neural network embedding model, feature extraction is performed on the original time-series data of the first type of photovoltaic storage system and the second type of photovoltaic storage system, so that the standardized vectors of different types of photovoltaic storage systems are in the same vector space.

3. The vector processing method for time-series data according to claim 1, characterized in that, The automatic cleaning of the original time-series data to obtain preprocessed data includes: Remove glitch values ​​from the original time-series data that are outside the physical feasible range of the optical storage system to obtain the first preprocessed data; Linear interpolation is performed on the first preprocessed data to obtain the second preprocessed data; The second preprocessed data from different signal channels are resampled to a uniform time interval to determine the preprocessed data.

4. The vector processing method for time-series data according to claim 1, characterized in that, The preprocessed data is subjected to time-domain feature extraction, frequency-domain feature extraction, and time-series embedding feature extraction to determine feature data related to the operating conditions of the photovoltaic storage system, including: The preprocessed data is subjected to time-domain statistical analysis to determine time-domain statistical characteristics. The time-domain statistical characteristics include at least the mean, variance, standard deviation, skewness, kurtosis, peak value, valley value, rate of change, cumulative value, autocorrelation coefficient, as well as the mean, maximum, rate of change of temperature difference between cells, and slope of state of charge change for the photovoltaic-storage system. The preprocessed data is subjected to frequency domain analysis to determine frequency domain characteristics. The frequency domain characteristics include at least the main frequency components, spectral energy distribution, main frequency ratio, and power frequency interference signals used to determine sampling anomalies in the photovoltaic energy storage system. Obtain the pre-trained temporal neural network embedding model, and input the preprocessed data into the temporal neural network embedding model to determine the temporal embedding features; The temporal neural network embedding model is replaceably integrated via a standard model interface.

5. The vector processing method for time-series data according to claim 4, characterized in that, Before inputting the preprocessed data into the temporal neural network embedding model, the vector processing method for temporal data further includes: Obtain triplet samples, wherein the triplet samples include anchor samples, positive samples and negative samples, the positive samples and the anchor samples belong to the same type of working condition, and the negative samples and the anchor samples belong to different types of working conditions; The contrast loss function is determined based on the triplet samples, and the parameters of the temporal neural network embedding model are updated through backpropagation to narrow the distance between vectors of the same type of working conditions and widen the distance between vectors of different types of working conditions.

6. The vector processing method for time-series data according to claim 1, characterized in that, The process of obtaining the pre-configured target vector dimension includes: The dimension of the target vector is determined based on the application scenario of the optical storage system, which includes real-time stream processing scenario, offline diagnostic scenario, and intermediate scenario. Wherein, the target vector dimension corresponding to the real-time stream processing scenario is smaller than the target vector dimension corresponding to the intermediate scenario, and the target vector dimension corresponding to the intermediate scenario is smaller than the target vector dimension corresponding to the offline diagnostic scenario.

7. The vector processing method for time-series data according to claim 1, characterized in that, The step of reducing the dimensionality of the fused features according to the target vector dimension, and then normalizing the reduced features to obtain standardized vectors of uniform length, includes: The fusion features are reduced in dimensionality according to the target vector dimension to determine the dimensionality-reduced feature vector. The reduced feature vectors are normalized to obtain candidate vectors of uniform length; The candidate vectors are subjected to quality verification, which includes dimensionality integrity check and normalization check. If the quality verification passes, the standardized vector is determined based on the candidate vector.

8. The vector processing method for time-series data according to any one of claims 1 to 7, characterized in that, After obtaining standardized vectors of uniform length, the vector processing method for time series data further includes: Obtain historical sample data of the photovoltaic storage system, including labeled samples of similar operating conditions and samples of different operating conditions; The aggregation degree of similar working condition vector sets and the separation degree of dissimilar working condition vector sets are determined based on the historical sample data. The anomaly discrimination degree is determined based on the aggregation degree and the separation degree. The anomaly discrimination degree is the value obtained by dividing the difference between the center distance of the heterogeneous working condition vector and the radius of the same working condition vector by the radius of the same working condition vector. If the anomaly discrimination score is not lower than a preset anomaly discrimination threshold, then the standardized vector is output, where the anomaly discrimination threshold is not lower than 95%. Otherwise, the original time-series data with an anomaly discrimination score lower than the anomaly discrimination threshold will be re-sent into the automatic cleaning process, and the feature extraction parameters or the fusion weights will be adjusted. The step of extracting time-domain features to normalization will be performed again until the anomaly discrimination score of the normalized vector is not lower than the anomaly discrimination threshold, or the preset maximum number of retries is reached.

9. A vector processing device for time-series data, characterized in that, include: The data acquisition module is used to acquire the raw time-series data corresponding to the photovoltaic-storage system. The raw time-series data includes the time sequence of at least one signal channel among battery temperature, state of charge, charge and discharge power, and grid state. The data cleaning module is used to automatically clean the raw time-series data to obtain preprocessed data; The feature extraction module is used to perform time-domain feature extraction, frequency-domain feature extraction, and time-series embedding feature extraction on the preprocessed data to determine feature data related to the operating conditions of the photovoltaic storage system. The feature data includes time-domain statistical features, frequency-domain features, and time-series embedding features. The feature fusion module is used to perform weighted concatenation of the time-domain statistical features, the frequency-domain features and the time-series embedded features according to the pre-configured fusion weights to determine the fused features. The fusion weights are configured to enhance the distinguishability between normal and abnormal operating conditions of the photovoltaic storage system in the vector space. The target determination module is used to obtain the pre-configured target vector dimensions; The vector processing module is used to perform dimensionality reduction processing on the fused features according to the target vector dimension, and to normalize the dimensionality-reduced features to obtain a standardized vector with consistent length. The temporal embedding features are obtained by processing the preprocessed data using a pre-trained temporal neural network embedding model.

10. A photovoltaic energy storage system, characterized in that, The photovoltaic-storage system includes at least one of energy storage equipment and photovoltaic equipment, and the photovoltaic-storage system further includes a vector processing device for time-series data as described in claim 9.

Images