A multi-parameter dual-channel time sequence reconstruction-based energy storage power station operation state early warning method and system

Abstract

The application discloses a kind of based on multi-parameter dual-channel time series reconstruction energy storage power station operating state early warning method and system.The method of the application includes: obtaining the multi-parameter operating time series data of energy storage power station, extracting key early warning index as input feature from it, constructing original sequence channel and first-order difference sequence channel in sliding time window, constructing dual-channel time series reconstruction early warning model, and training model and calculating reconstruction anomaly score by the weighted sum of original sequence reconstruction error, difference sequence reconstruction error and time series recursive prediction error;Again, the final early warning main detection score is obtained by smoothing the reconstruction anomaly score, the early warning threshold is determined, and normal, first-level early warning, second-level early warning, third-level early warning and fourth-level early warning are output.The application realizes the stability of effective identification and hierarchical early warning of operating state anomaly by constructing original sequence channel and first-order difference sequence channel, and jointly modeling the multi-parameter operating state of energy storage power station.

Description

Technical Field

[0001] This invention belongs to the field of energy storage system status assessment and data-driven analysis technology, specifically a method and system for early warning of the operating status of energy storage power stations based on multi-parameter dual-channel time sequence reconstruction. Background Technology

[0002] During the operation of an energy storage system, the battery management system can continuously collect multi-dimensional operating data, including information such as current, voltage, temperature, and state of charge. This operating data reflects the dynamic changes in the internal electrochemical reaction process of the battery and the external operating conditions. Therefore, analyzing the operating data to identify changes in the system's operating status is of great significance for ensuring the safe operation of the energy storage system.

[0003] During the operation of an energy storage system, when the system's operating state becomes abnormal, the operating variables typically exhibit abnormal characteristics at multiple levels. Specifically: (1) at the statistical level, the mean level, fluctuation amplitude, and directional drift of the operating variables may deviate from the statistical distribution characteristics under normal operating conditions; (2) at the temporal dynamic level, the time evolution pattern of the operating variables may deviate from the dynamic behavior pattern under normal operating conditions, manifesting as abnormal transient fluctuations or increased short-term prediction deviations. These two types of changes often occur simultaneously and require comprehensive analysis from both the amplitude state and the rate of change state to accurately characterize the process of changes in the system's operating state.

[0004] In existing technologies, common anomaly identification methods mainly include methods based on statistical features and methods based on time series modeling. Methods based on statistical features typically calculate statistical indicators such as the mean, standard deviation, or rate of change of variables, and use thresholds to determine whether variables are abnormal. While these methods are computationally simple, they have the following shortcomings: First, traditional statistical methods usually use a globally fixed reference distribution, failing to fully consider the impact of differences in operating conditions on the normal fluctuation range, leading to normal fluctuations under different operating conditions being misjudged as abnormal. Second, a single statistical indicator only describes the operating status from one dimension, making it difficult to comprehensively reflect multiple abnormal manifestations such as amplitude deviation, increased volatility, and trend changes. Time series modeling methods typically employ neural network models such as autoencoders to model time series data and identify abnormal states through reconstruction errors. While these methods can identify dynamic changes in time series, they suffer from the following shortcomings: First, traditional single-channel autoencoders only reconstruct the amplitude of the original sequence, failing to explicitly model the rate of change between adjacent time points, thus limiting their ability to capture dynamic evolutionary features. Second, traditional methods typically use equal weighting for different performance indicators, failing to allocate differentiated weights based on the degree of change of each indicator within the current window, resulting in insufficient ability to distinguish key abnormal features. Third, traditional methods usually use only a single reconstruction loss as the training objective, lacking explicit modeling of short-term state evolution trends.

[0005] Furthermore, existing methods have two shortcomings in terms of early warning output. Firstly, in processing early warning scores, most methods directly determine early warnings based on anomaly scores within a single time window, without performing time-series smoothing of the scores. This makes the early warning results susceptible to sampling noise and local random fluctuations, leading to frequent jumps. Secondly, in setting early warning thresholds, most methods use globally fixed thresholds or empirical thresholds, failing to adaptively determine thresholds based on the distribution of healthy samples under the same operating conditions. This makes it difficult to adapt to differences in data distribution under different operating conditions. Additionally, existing methods often lack explanations of the source of anomalies when outputting anomaly judgments, making it difficult to explain whether the early warning results are mainly caused by deviations in the level of operating indicators, increased volatility, or trend drift. This affects the judgment of operation and maintenance personnel regarding the causes of changes in the energy storage power station's status. Summary of the Invention

[0006] The technical problem this invention aims to solve is that existing early warning methods for the operating status of energy storage power stations typically rely on single statistical indicators, empirical thresholds, or single-channel time series models for anomaly detection. These methods struggle to simultaneously characterize the amplitude changes, short-term fluctuations, and dynamic evolution characteristics of energy storage system operating variables, making them susceptible to changes in operating conditions, local disturbances, and random fluctuations, resulting in insufficient stability of the early warning results. Furthermore, existing methods often lack explanations of the sources of anomalies when outputting anomaly judgments, making it difficult to determine whether the early warning results are primarily caused by deviations in operating indicator levels, increased volatility, or trend drift. This, in turn, affects the judgment of operation and maintenance personnel regarding the causes of changes in the energy storage power station's status.

[0007] To address the aforementioned technical issues, this invention proposes a method and system for early warning of the operating status of energy storage power stations based on multi-parameter dual-channel time-series reconstruction. By constructing original sequence channels and first-order differential sequence channels, it jointly models the multi-parameter operating status of energy storage power stations, enabling effective identification and hierarchical early warning of abnormal operating status. Furthermore, this invention combines a three-dimensional attribution method under statistical health reference to explain the source of anomalies.

[0008] To achieve the above-mentioned objectives, the present invention adopts the following technical solution.

[0009] In a first aspect, the present invention provides a method for early warning of the operating status of an energy storage power station based on multi-parameter dual-channel time-series reconstruction, comprising the following steps: S1. Obtain multi-parameter operating time series data of energy storage power station, determine key early warning indicators, construct multivariate time series of key early warning indicators based on sliding time window, and perform first-order difference on multivariate time series to obtain differential time series. S2, input the multivariate time series and the difference time series into the original sequence channel and the first-order difference sequence channel respectively, and perform joint modeling through the feature attention module, the dual-channel long short-term memory network autoencoder, the dual-channel decoder and the time series recursive prediction branch to obtain the dual-channel time series reconstruction early warning model; S3, the dual-channel time-series reconstruction early warning model is trained by using the weighted sum of the original sequence reconstruction error, the differential sequence reconstruction error, and the time-series recursive prediction error as the joint training loss; S4. The reconstructed anomaly score is calculated window by window using the trained dual-channel temporal reconstruction early warning model, and the reconstructed anomaly score is processed by exponential weighted moving average to obtain the final early warning main detection score. S5 determines the warning threshold based on the quantile of the final warning main detection score under the same working conditions, and outputs normal, first-level warning, second-level warning, third-level warning or fourth-level warning accordingly.

[0010] Furthermore, in step S1, the key early warning indicator is a pre-determined variable representing the battery's operating state, which is at least one of the following: state of charge, voltage, current, temperature, and their statistical characteristics.

[0011] Furthermore, in step S2, the specific construction steps of the dual-channel time-series reconstruction early warning model include: S21, Construct a dual-channel input structure, wherein the original sequence channel is used to receive multivariate time series, and the first-order difference sequence channel is used to receive difference time series; S22, Construct a feature attention module, generate feature attention weights by pooling, concatenating and softmax mapping multivariate time series and difference time series, and use feature attention weights to weight each feature of the two input channels; S23, construct a dual-channel long short-term memory network autoencoder, and encode the weighted original sequence channel and the first-order difference sequence channel through two long short-term memory networks respectively to obtain the original sequence encoding features and the difference sequence encoding features; concatenate the original sequence encoding features and the difference sequence encoding features and input them into a fully connected mapping layer to obtain a low-dimensional latent feature vector used to characterize the current time window running state; S24, construct a dual-channel decoder and a temporal recursive prediction branch, and input the low-dimensional latent feature vector into the dual-channel decoder and the temporal recursive prediction branch respectively; wherein, the dual-channel decoder is used to output the original sequence reconstruction result and the difference sequence reconstruction result, and the temporal recursive prediction branch is used to output the temporal recursive prediction value after the window.

[0012] Furthermore, in step S4, the weighted composition method of the reconstructed anomaly score is the same as that of the joint training loss in step S3, both using a weighted sum of the original sequence reconstruction error, the differential sequence reconstruction error, and the time-series recursive prediction error.

[0013] Furthermore, the reconstructed anomaly score is processed by an exponentially weighted moving average to obtain the final early warning main detection score S. alarm (t), where S alarm (t)= α S rec (t)+(1- α )S alarm (t-1), S rec (t) represents the reconstruction anomaly score, S alarm (t-1) represents the warning main detection score at time t-1. α Represents the smoothing coefficient, and within the same set of operating conditions, it is based on the final warning main detection score S. alarm The distribution of (t) is selected from preset quantiles, which are used as the first-level warning threshold, the second-level warning threshold, the third-level warning threshold and the fourth-level warning threshold.

[0014] Furthermore, the aforementioned energy storage power station operation status early warning method also includes: step S6, for the time window of the output early warning level, matching a set of healthy neighbors in the historical healthy window that has not triggered an early warning under the same operating conditions, using the statistical distribution of the set of healthy neighbors as the health reference baseline, and calculating the attribution information of each key early warning indicator in the current window in three dimensions: horizontal deviation, abnormal fluctuation, and trend drift.

[0015] Further, in step S6, based on the healthy neighbor set, the attribution information for each dimension is calculated as follows: the horizontal deviation attribution information is the Z-score of the mean of each indicator in the current window relative to the distribution of the mean of the corresponding indicator in the healthy neighbor set; the volatility anomaly attribution information is the larger of the standard deviation and range of each indicator in the current window relative to the Z-score of the distribution of the corresponding indicator in the healthy neighbor set; the trend drift attribution information is the larger of the absolute value of the slope and the mean of the point-by-point difference amplitude of each indicator in the current window relative to the Z-score of the distribution of the corresponding indicator in the healthy neighbor set.

[0016] Furthermore, in step S6, the output attribution information is output synchronously with the early warning time window, reconstruction anomaly score, early warning main detection score, early warning threshold, early warning level, and reconstruction error component. The reconstruction error component includes the original sequence reconstruction error, the differential sequence reconstruction error, and the time-series recursive prediction error, which are used to indicate the main contribution indicators and contribution dimensions of the current early warning.

[0017] Secondly, the present invention provides an early warning system for the operation status of an energy storage power station based on multi-parameter dual-channel timing reconstruction, used to realize the above-mentioned early warning method for the operation status of an energy storage power station throughout its entire life cycle, comprising: The data input module is used to acquire multi-parameter operating time series data of the energy storage power station; The time window construction module is used to construct multivariate time series of key early warning indicators based on sliding time windows, and to construct the difference time series obtained by first-order difference of the multivariate time series, and to construct the original sequence channel and the first-order difference sequence channel respectively. The indicator input module is used to receive pre-determined key early warning indicators; The dual-channel time series reconstruction detection module is used to build a dual-channel time series reconstruction early warning model. It performs joint encoding and reconstruction of the multivariate time series and differential time series of key early warning indicators. It obtains the reconstruction anomaly score through feature attention mechanism and three-objective joint training. The three objectives are the original sequence reconstruction error, the differential sequence reconstruction error and the time series recursive prediction error. The smoothing and threshold determination module is used to smooth the reconstructed anomaly score, determine the first-level warning threshold, second-level warning threshold, third-level warning threshold and fourth-level warning threshold based on the quantile of the final warning main detection score, and output the warning level.

[0018] Furthermore, the aforementioned energy storage power station operation status early warning system also includes: The statistical attribution module is used to calculate the degree of deviation of statistical features of a time window from a healthy reference in terms of level, volatility and trend by matching the healthy neighbor set, and output statistical attribution information. The results output module is used to output the early warning time window, reconstruction anomaly score, early warning main detection score, early warning threshold, early warning level, reconstruction error components, and statistical three-dimensional attribution information. The reconstruction error components include the original sequence reconstruction error, the differential sequence reconstruction error, and the time-series recursive prediction error.

[0019] Compared with existing technologies, the present invention has the following advantages: It constructs a dual-channel time-series reconstruction early warning model, simultaneously introducing multivariate time series and differential time series, enabling more stable identification of operational anomalies by utilizing both amplitude and rate-of-change information of energy storage power station operating variables; the joint modeling of reconstruction and prediction tasks enhances the model's ability to perceive abnormal time-series patterns; smoothing the early warning master detection score into a final early warning master detection score improves the continuity of early warning output; setting thresholds based on healthy samples or samples under similar operating conditions enhances the adaptability of early warning judgments to actual operating conditions; and attributing the results through three dimensions—horizontal deviation, volatility anomalies, and trend drift—improves the interpretability and engineering application value of the early warning results. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of an early warning method for the operating status of an energy storage power station according to the present invention; Figure 2 This is a schematic diagram of the dual-channel timing reconstruction early warning model in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the three-dimensional attribution of statistical health reference in Embodiment 2 of the present invention; Figure 4 This is a comparison chart of the standard deviation changes of SOC on normal operation days and warning days in Embodiment 2 of the present invention; Figure 5 This is a comparison chart of the rate of change of voltage standard deviation between normal operation days and warning days in Embodiment 2 of the present invention; Figure 6 This is a comparison chart of the voltage variation coefficient changes between normal operation days and warning days in Embodiment 2 of the present invention; Figure 7 This is a comparison chart of the voltage standard deviation changes between normal operation days and warning days in Embodiment 2 of the present invention; Figure 8 This is a comparison chart of the standard deviation of current on normal operation days and warning days in Embodiment 2 of the present invention; Figure 9 This is a comparison chart of the temperature variation coefficient changes between normal operation days and warning days in Embodiment 2 of the present invention; Figure 10 This is a schematic diagram of the early warning system for the operation status of the energy storage power station of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described below with reference to embodiments. It should be understood that the following embodiments are only for illustrating the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or modifications made by those skilled in the art without departing from the principle of the present invention should all fall within the scope of protection of the present invention.

[0023] Example 1 This embodiment presents a method for early warning of the full lifecycle operation status of an energy storage power station based on multi-parameter dual-channel time-series reconstruction. Figure 1 As shown, the steps are as follows: S1. Obtain multi-parameter operating time series data of energy storage power station, determine key early warning indicators, construct multivariate time series (i.e., original series) of key early warning indicators based on sliding time window, and obtain the differential time series by first-order difference of multivariate time series.

[0024] The operational time series data comes from multi-dimensional time series data collected by the battery management system or energy storage power station monitoring system. The time series data includes one or more of the battery's current, voltage, temperature, state of charge and their derived statistical characteristics, and an operational time series dataset is constructed according to a unified time benchmark.

[0025] The collected operational time-series data undergoes preprocessing, including time alignment, outlier identification, missing value completion, duplicate timestamp handling, and standardization. This ensures a consistent time benchmark and comparable data scale across different operational variables, resulting in a standardized operational dataset. Key early warning indicators are extracted from this standardized dataset. These indicators can be pre-determined based on the energy storage power station's operating mechanism, historical operational data, and engineering experience. They are at least one of the following: current, voltage, temperature, state of charge, and statistical characteristics calculated from current, voltage, temperature, and state of charge.

[0026] A sample of key early warning indicators is constructed based on a sliding time window. Each sliding time window corresponds to a multivariate operating state segment, used to describe the operating state of the energy storage power station within a local time range. This serves as the input variable for the subsequent dual-channel time-series reconfiguration early warning model. The key early warning indicators characterize the changes in the operating state of the energy storage power station, and the final early warning determination is based on the reconfiguration anomaly score and threshold judgment results output by the early warning model.

[0027] S2. The multivariate time series and the difference time series are input into the original sequence channel and the first-order difference sequence channel, respectively. The dual-channel time series reconstruction early warning model is obtained by joint modeling through the feature attention module, the dual-channel long short-term memory network autoencoder, the dual-channel decoder and the time series recursive prediction branch.

[0028] Based on window-level samples, a dual-channel time-series reconstruction early warning model is constructed. The dual channels include an original sequence channel and a first-order difference sequence channel. The original sequence channel describes the amplitude level, overall distribution, and steady-state offset characteristics of the operating variables within the current time window; the first-order difference sequence channel describes the rate of change, short-term fluctuations, and dynamic evolution characteristics of the operating variables between adjacent sampling times. By simultaneously introducing multivariate time series and difference time series, the operation process of the energy storage power station can be characterized from both amplitude and rate of change perspectives, avoiding the problem of insufficient representation of abnormal states by a single input form.

[0029] The dual-channel time-series reconstruction early warning model includes a feature attention module, a dual-channel long short-term memory network autoencoder, a dual-channel decoder, and a time-series recursive prediction branch. Specifically, the feature attention module assigns feature weights based on the changes in different operating data within the current time window; the dual-channel long short-term memory network autoencoder extracts time-series features from multivariate time series and difference time series respectively; the dual-channel decoder reconstructs the input sequence; and the time-series recursive prediction branch learns the short-term evolution patterns of the energy storage power station's operating state.

[0030] The specific construction steps of the dual-channel time-series reconstruction early warning model are as follows: S21, Construct a dual-channel input structure, wherein the original sequence channel is used to receive multivariate time series, and the first-order difference sequence channel is used to receive the difference time series obtained after performing first-order difference on the multivariate time series; S22, Construct a feature attention module, generate feature attention weights by pooling, concatenating and softmax mapping multivariate time series and difference time series, and use feature attention weights to weight each feature of the two input channels; S23, construct a dual-channel long short-term memory network autoencoder, and encode the weighted original sequence channel and the first-order difference sequence channel through two long short-term memory networks respectively to obtain the original sequence encoding features and the difference sequence encoding features; concatenate the original sequence encoding features and the difference sequence encoding features and input them into a fully connected mapping layer to obtain a low-dimensional latent feature vector used to characterize the current time window running state; S24, construct a dual-channel decoder and a temporal recursive prediction branch, and input the low-dimensional latent feature vector into the dual-channel decoder and the temporal recursive prediction branch respectively; wherein, the dual-channel decoder is used to output the original sequence reconstruction result and the difference sequence reconstruction result, and the temporal recursive prediction branch is used to output the temporal recursive prediction value after the window.

[0031] S3. The dual-channel time-series reconstruction early warning model is trained by using the weighted sum of the original sequence reconstruction error, the differential sequence reconstruction error, and the time-series recursive prediction error as the joint training loss.

[0032] During the model training phase, a weighted sum of the original sequence reconstruction error, the differential sequence reconstruction error, and the time-series recursive prediction error is used as the joint training loss, enabling the model to learn the multidimensional temporal variation patterns under normal operating conditions. In this way, the model can not only characterize the overall state of the operating variables but also their rate of change and short-term evolution trends, thereby improving its ability to identify anomalous time-series patterns.

[0033] S4. The reconstructed anomaly score is calculated window by window using the trained dual-channel time-series reconstruction early warning model, and the reconstruction anomaly score is processed by exponential weighted moving average to obtain the final early warning main detection score.

[0034] During the early warning detection phase, the time window to be detected is input into the trained dual-channel time series reconstruction early warning model, and the weighted sum of the original sequence reconstruction error, the differential sequence reconstruction error, and the time series recursive prediction error is taken as the reconstruction anomaly score.

[0035] The reconstructed anomaly score is processed by an exponentially weighted moving average to obtain the final early warning main detection score S. alarm (t), where S alarm (t)= α S rec (t)+(1- α )S alarm (t-1), S rec (t) represents the reconstruction anomaly score, S alarm (t-1) represents the warning main detection score at time t-1. α Denotes the smoothing coefficient, and S is taken respectively within the same set of working conditions. alarm The preset quantiles of (t) are used as the first-level, second-level, third-level, and fourth-level warning thresholds. 0 < α ≤1, used to control the weight of the current reconstruction anomaly score in the smoothing result.

[0036] The reconstruction anomaly score satisfies:

[0037] in, Indicates the reconstruction anomaly score. This represents the reconstruction error of the original sequence. For the difference sequence reconstruction error, This is the time-series recursive prediction error. , and All are non-negative weights and + + =1.

[0038] When the energy storage power station is in normal operation, the input sequence can be reconstructed by the model well, and the prediction results are close to the actual operating state. Therefore, the warning main detection score is low. When the operating state of the energy storage power station deviates from the normal mode, the model reconstruction ability and prediction consistency decrease, and the corresponding warning main detection score increases.

[0039] Considering the short-term disturbances, sampling noise, and local random fluctuations that occur during the actual operation of energy storage power stations, directly relying on the reconstruction anomaly score of a single time window for early warning can easily lead to frequent jumps in warning results. This invention performs time smoothing processing on the window-level early warning main detection score to obtain the final early warning judgment score, thereby enhancing the continuity and stability of the early warning results.

[0040] S5 determines the warning threshold based on the quantile of the final warning main detection score under the same working conditions, and outputs normal, first-level warning, second-level warning, third-level warning or fourth-level warning accordingly.

[0041] After obtaining the final warning judgment score, a warning threshold is set based on the score distribution in healthy samples or samples under the same operating conditions. The operating status of the energy storage power station is then classified into normal, Level 1 warning threshold, Level 2 warning threshold, Level 3 warning threshold, and Level 4 warning threshold according to these thresholds. By determining the threshold based on healthy samples or samples under the same operating conditions, the warning judgment can adapt to the differences in data distribution under different operating conditions, reducing false alarms caused by operating condition switching or occasional fluctuations.

[0042] S6: For the time window of the output warning level, match the healthy neighbor set in the historical healthy window that has not triggered the warning under the same working conditions. Use the statistical distribution of the healthy neighbor set as the health reference baseline, and calculate the attribution information of each key warning indicator in the current window in three dimensions: horizontal deviation, abnormal fluctuation and trend drift.

[0043] To improve the interpretability of early warning results, this invention constructs a three-dimensional attribution method based on statistical health reference while outputting the early warning level. The three dimensions include horizontal deviation, volatility anomaly, and trend drift. Horizontal deviation describes the degree of deviation of the overall operating indicator value from the healthy state; volatility anomaly describes the short-term volatility of the operating indicator within a local time window; and trend drift describes whether the operating indicator exhibits continuous change or an abnormal rate of change. Through the above attribution information, the reconstructed anomaly score output by the early warning model can be interpreted as a specific operating indicator and its change type, providing a basis for the analysis of the operating status of energy storage power stations.

[0044] Based on the healthy neighbor set, the attribution information for each dimension is calculated as follows: the horizontal deviation attribution information is the Z-score of the mean of each indicator in the current window relative to the distribution of the mean of the corresponding indicator in the healthy neighbor set; the volatility anomaly attribution information is the larger of the standard deviation and range of each indicator in the current window relative to the Z-score of the distribution of the corresponding indicator in the healthy neighbor set; the trend drift attribution information is the larger of the absolute value of the slope and the mean of the point-by-point difference amplitude of each indicator in the current window relative to the Z-score of the distribution of the corresponding indicator in the healthy neighbor set.

[0045] The attribution information is not a necessary triggering condition for determining the warning level, but is used to explain the source of the warning and assist in operational analysis; the warning results include the time window, the final warning main detection score, the warning level, the reconstruction error components, and the main attribution dimensions.

[0046] The final output includes the warning time window, the final warning main detection score, the warning level, threshold information, reconstruction error components, and statistical three-dimensional attribution information. The reconstruction error components include the original sequence reconstruction error, the difference sequence reconstruction error, and the time-series recursive prediction error. The statistical three-dimensional attribution information is output synchronously with the warning time window, reconstruction anomaly score, warning main detection score, warning threshold, warning level, and reconstruction error components. It is used to indicate the main contributing indicators and dimensions of the current warning, that is, to explain that the current warning mainly originates from horizontal deviation, fluctuation anomaly, or trend drift. Thus, the warning results can not only determine whether the operating status is abnormal, but also reflect the main manifestations of the abnormal status.

[0047] Example 2 This embodiment presents a method for early warning of the full life cycle operation status of an energy storage power station based on multi-parameter dual-channel time-series reconstruction. The steps are as follows: First, multi-dimensional time-series operational data are collected from the battery management system or monitoring system of the energy storage power station. This operational data includes at least current, voltage, temperature, and state of charge. The raw data undergoes timestamp alignment, duplicate timestamp removal, missing value completion, outlier handling, and standardization to ensure that different operational variables have a unified time reference and comparable data scale.

[0048] In this embodiment, to avoid interference from data during the shutdown or idle phases of the energy storage system on model training, the effective operating intervals are first identified based on the operating current. When the absolute value of the current is greater than a preset current threshold, the corresponding time is determined to be an effective operating state. Time points that continuously meet this condition are merged to form continuous operating segments. When the length of a continuous operating segment is less than 60 minutes, the segment is discarded. This method preserves data that reflects the actual charging and discharging operation process of the energy storage power station.

[0049] After obtaining valid runtime segments, a sliding time window is used to construct window-level samples. The sliding window length is 60 minutes, the sliding step size is 30 minutes, and adjacent windows overlap by 30 minutes, i.e., the overlap rate is 50%. If the original sampling interval is Δ... t Each window contains 60 min / Δ t There are sampling points. For the t-th time window, its multivariate time series... Represented as:

[0050] in,L Indicates the number of sampling points within the window. m This indicates the number of key early warning indicators. Indicates the first in the window The value of the j-th key early warning indicator at each sampling point. Represents a real number.

[0051] Key early warning indicators can be pre-selected based on the operating mechanism, historical operating data, and engineering experience of the energy storage power station. The following indicators are selected as key early warning indicators: soc_std_60 (characterizing state of charge fluctuations), voltage_std_abs_rate (characterizing abnormal voltage rate of change), voltage_cv (characterizing voltage consistency), voltage_std (characterizing voltage dispersion), current_std_60 (characterizing current fluctuations), and temperature_cv (characterizing temperature dispersion). These key early warning indicators serve as input variables for the dual-channel time-series reconstruction model. The final early warning determination is based on the anomaly score and threshold judgment results output by the model. These key early warning indicators are only used as input conditions for the early warning model and are not considered as the main detection component of the early warning model of this invention.

[0052] In order to simultaneously characterize the magnitude and rate of change of the running variables, in multivariate time series... X Construct a first-order difference time series based on t:

[0053] Among them, multivariate time series X t is used to describe the amplitude level, overall distribution, and steady-state offset characteristics of the running variable within a 60-minute window; the difference time series Δ X t is used to describe the rate of change, short-term fluctuations, and dynamic evolution characteristics of the running variable between adjacent sampling times. This dual-channel input avoids the problem of insufficient characterization of dynamic changes when using only a single original sequence.

[0054] like Figure 2 As shown, after obtaining the window-level multivariate time series and the difference time series, a feature attention module is constructed. This feature attention module is used to assign weights based on the changes in different performance indicators within the current window. Specifically, pooling is performed on the original sequence channels and the first-order difference sequence channels respectively, and the pooling results are concatenated to obtain the context vector.

[0055] Calculate feature attention weights based on context vectors:

[0056] Where W1, W2, b1, and b2 are model parameters. Represents a non-linear activation function. This represents the attention weight of the j-th key warning indicator within the current window. This weight is then used to perform feature-wise weighting on the two input channels:

[0057] In the formula, This indicates that the model represents the first time within the t-th time window. The original sequence values ​​of the j-th sampling point and the j-th key early warning indicator; This represents the channel values ​​of the original sequence after being weighted by feature attention weights; This indicates the time window containing the first (t) time unit. The first-order difference sequence values ​​of the j-th sampling point and the j-th key early warning indicator; This represents the channel values ​​of the first-order difference sequence after being weighted by feature attention weights.

[0058] Subsequently, a dual-channel long short-term memory network autoencoder was constructed. (Original sequence channels) and Input the data into two independent LSTM encoders to obtain the corresponding hidden states. and The two hidden states are concatenated, and The window-level latent feature vector is obtained through bottleneck mapping:

[0059] in, Low-dimensional time-series features used to characterize the operating status of energy storage power stations under the current window. This represents the weight moments of the fully connected mapping layer. This represents the bias term of the fully connected mapping layer. In one specific implementation, the LSTM hidden representation dimension is 32, the batch size is 16, and the maximum number of training epochs is 120.

[0060] During the decoding stage, the latent feature vectors The input is a dual-channel decoder, from which the reconstructed original sequence results are obtained. Sum of difference sequence reconstruction results At the same time, Input the time-series recursive prediction branch to obtain the predicted value at the next sampling time after the window.

[0061] During the model training phase, the original sequence reconstruction error, the difference sequence reconstruction error, and the time-series recursive prediction error are calculated respectively: In one specific implementation, the model training samples are selected from historical stable operation windows or healthy sample pools. The training samples are used to characterize typical time-series patterns under normal operating conditions of energy storage power stations.

[0062]

[0063]

[0064]

[0065] in, This indicates that the model represents the first time within the t-th time window. The original values ​​of the j-th sampling point and the j-th key early warning indicator The reconstructed value; This represents the first-order difference between adjacent sampling points of the j-th key early warning indicator within the t-th time window; The model represents the first-order difference value. The reconstructed value; This represents the actual value of the j-th key early warning indicator at the next sampling time after the t-th time window; This represents the model's predicted value for the j-th key early warning indicator at the next sampling time; is the loss weight of the j-th key early warning indicator in the joint training loss, used to weight the error terms of different indicators; The same value can be used (equal weighting), or the values ​​can be set differently according to the indicators based on engineering experience and historical operational data statistics. The joint training loss of the model is:

[0066] in, , and All are non-negative weights and + + =1. In one specific implementation, = = =1 / 3, by minimizing the above joint loss, the model can learn the amplitude pattern, rate of change pattern, and short-term evolution trend under normal operating conditions of the energy storage power station. After the model training is completed, the time window to be evaluated is input into the dual-channel time series reconstruction model, and the window-level reconstruction anomaly score is calculated in the same way:

[0067] When the energy storage power station is in normal operation, the input sequence can be reconstructed by the model well, and the time-series recursive prediction value is close to the actual value, so Srec(t) is small. When the operation deviates from the normal mode, the model reconstruction error and prediction error increase, and the corresponding Srec(t) increases. Therefore, Srec(t) can be used to characterize the degree to which the current window deviates from the normal time-series mode.

[0068] Considering the sampling noise, short-term disturbances, and local random fluctuations that occur during the actual operation of energy storage power stations, this embodiment uses an exponentially weighted moving average to process the window-level reconstruction anomaly score, resulting in the early warning main detection score:

[0069] in, The smoothing coefficient can be determined based on a preset smoothing window span, for example... =2 / (span+1). In one specific implementation, span is 5, then... =1 / 3. Through the above smoothing process, the impact of abnormal fluctuations in a single window on the early warning results can be reduced, and the continuity of early warning output can be improved.

[0070] After obtaining the final warning main detection score, the warning threshold is determined based on the score distribution in healthy samples or samples under the same operating conditions. Let the set of operating conditions be Ωm, then the four warning thresholds can be expressed as follows:

[0071] in, , , and Let represent the corresponding quantile functions. Based on the relationship between Saltm(t) and the threshold, determine the warning level for the t-th window: when Saltm(t) < 0. When, it is judged as a normal state; when ≤Salarm(t)< When, it is determined to be a Level 1 warning; when ≤Salarm(t)< When, it is determined to be a Level II warning; when ≤Salarm(t)< When the alarm is at its highest level, it is classified as a Level 3 warning; when Salarm(t) ≥ At that time, it was determined to be a Level IV warning.

[0072] To improve the interpretability of early warning results, this embodiment constructs a three-dimensional attribution method based on statistical health references while outputting the early warning level. For the first... t The first window j Calculate the average of several key early warning indicators. Standard deviation Range linear trend slope and average difference magnitude A healthy neighbor set N(t) is obtained by matching healthy windows under the same operating conditions. In one specific implementation, an operating condition matching space is first constructed based on contextual features such as SOC, current, and temperature. Standardized Euclidean distance is used as the contextual distance metric, and windows with a contextual distance less than the 0.60 quantile are selected as the healthy sample pool. For each window to be evaluated, the 25 most similar healthy windows are selected from the healthy sample pool as the healthy neighbor set N(t). Let the mean and standard deviation of the corresponding statistics in the healthy neighbor set be denoted as follows: and .

[0073] The deviation of the horizontal characteristic is defined as:

[0074] Volatility anomalies are defined as follows:

[0075] Trend drift is defined as:

[0076] in, To prevent small positive numbers with a denominator of zero. level,j Z(t) is used to characterize the degree of deviation of the overall level of the indicator in the current window from the healthy state. vol,j (t) is used to characterize the degree of short-term fluctuation anomalies, Z trend,j (t) is used to characterize the degree of anomalousness in trend changes or rates of change. After obtaining the three-dimensional attribution values ​​for each indicator, the window-level attribution strength is calculated:

[0077] Among them, A level (t), A vol (t) and A trend (t) represents the combined attribution strength of the current window across the three dimensions of horizontal deviation, volatility anomaly, and trend drift, respectively. Figure 3 The calculation relationship of the three-dimensional attribution under the statistical health reference is shown below. Horizontal deviation characterizes the degree of deviation of the current time window mean from the healthy neighbor set; volatility anomaly characterizes the degree of anomaly of the standard deviation and range relative to the healthy neighbor set; and trend drift characterizes the degree of anomaly of the slope and mean difference magnitude relative to the healthy neighbor set. When outputting the early warning result, the dimension with the larger value among the three can be used as the primary attribution dimension to explain which type of operational status change mainly caused the current early warning.

[0078] The final output includes the warning time window and the reconstructed anomaly score S. rec (t), Early Warning Main Detection Score alarm(t), warning threshold, warning level, reconstruction error components, and statistical three-dimensional attribution information, such as Figures 4 to 9 As shown, the comparisons of SOC standard deviation, voltage standard deviation rate of change, voltage coefficient of variation, voltage standard deviation, current standard deviation, and temperature coefficient of variation between normal operation days and warning days illustrate the increased volatility, greater dispersion, or trend changes of relevant key indicators on warning days compared to normal operation days, thus helping to explain the source of the warning results. The above three-dimensional statistical attribution information is not used as the final warning trigger condition, but rather to explain the source of the warning, providing a basis for status analysis for energy storage power station operation and maintenance personnel.

[0079] Example 3 This embodiment provides a full lifecycle operational status early warning system for energy storage power stations based on multi-parameter dual-channel timing reconstruction, used to implement the aforementioned operational status early warning method for energy storage power stations, such as... Figure 10 As shown, this system can be deployed in energy storage power station monitoring platforms, cloud analytics platforms, or edge computing devices to perform real-time or offline analysis of energy storage power station operating data, thereby generating operational status early warning results.

[0080] The system is built based on a data flow driven approach and includes a data input module, a preprocessing module, a time window construction module, an indicator input module, a dual-channel time series reconstruction detection module, a smoothing and threshold determination module, a statistical attribution module, and a result output module. The modules communicate with each other through a data interface to realize the early warning process of the energy storage power station's operating status.

[0081] The data input module is used to receive multi-parameter operating time series data collected by the battery management system or monitoring system of the energy storage power station. The data includes at least variables such as current, voltage, temperature and state of charge.

[0082] The preprocessing module is used to perform time alignment, outlier handling, duplicate timestamp handling, missing value completion, and standardization on the input data to obtain a running dataset with a unified time base and a unified data scale.

[0083] The time window construction module is used to construct multivariate time series of key early warning indicators based on sliding time windows, and to construct differential time series obtained by first-order difference of multivariate time series, and to construct original sequence channels and first-order difference sequence channels respectively.

[0084] The indicator input module is used to receive pre-determined key early warning indicators, which serve as input variables for the dual-channel time-series reconstruction model to characterize the changes in the operating status of the energy storage power station.

[0085] The dual-channel temporal reconstruction detection module is used to construct a dual-channel temporal reconstruction early warning model. It includes a feature attention module, a raw sequence encoder, a differential sequence encoder, a dual-channel decoder, and a temporal recursive prediction branch. It utilizes the joint encoding and reconstruction of multivariate time series and differential time series of key early warning indicators. The reconstruction anomaly score is obtained through feature attention mechanism and joint training with three objectives: raw sequence reconstruction error, differential sequence reconstruction error, and temporal recursive prediction error.

[0086] The smoothing and threshold determination module is used to smooth the reconstructed anomaly score, determine the first-level warning threshold, second-level warning threshold, third-level warning threshold and fourth-level warning threshold based on the quantile of the final warning main detection score, and output the warning level.

[0087] The statistical attribution module is used to calculate the degree of deviation of statistical features of a time window from a healthy reference in terms of level, volatility, and trend by matching the healthy neighbor set, and output statistical attribution information.

[0088] The results output module is used to output the early warning time window, reconstruction anomaly score, early warning main detection score, early warning threshold, early warning level, reconstruction error components, and statistical three-dimensional attribution information. The reconstruction error components include the original sequence reconstruction error, the differential sequence reconstruction error, and the time-series recursive prediction error.

[0089] In some implementations, the system can periodically receive new operational data and automatically update window-level samples, reconstruct anomaly scores and early warning master detection scores, thereby achieving dynamic early warning of the energy storage power station's operational status. Early warning results can be displayed through a visual interface, including the early warning master detection score curve, early warning threshold, changes in early warning level, and key attribution dimensions.

[0090] The system described in this embodiment uses a dual-channel time-series reconstruction early warning model as its core and employs statistical three-dimensional attribution for explanation and output. Through this system structure, the operating status of energy storage power stations can be continuously monitored without relying on fault labels, and early warning levels and anomaly source descriptions can be output when the operating status deviates from the normal mode, thereby improving the stability and interpretability of early warnings for the operating status of energy storage power stations.

[0091] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. For those skilled in the art, various modifications, substitutions, or variations can be made to the specific embodiments of the present invention without departing from the spirit and substance of the present invention, and all such modifications, substitutions, or variations should be included within the scope of protection of the present invention.

Claims

1. A method for early warning of the operating status of an energy storage power station based on multi-parameter dual-channel time-series reconstruction, characterized in that, Including the following steps: S1. Obtain multi-parameter operating time series data of energy storage power station, determine key early warning indicators, construct multivariate time series of key early warning indicators based on sliding time window, and perform first-order difference on multivariate time series to obtain differential time series. S2, input the multivariate time series and the difference time series into the original sequence channel and the first-order difference sequence channel respectively, and perform joint modeling through the feature attention module, the dual-channel long short-term memory network autoencoder, the dual-channel decoder and the time series recursive prediction branch to obtain the dual-channel time series reconstruction early warning model; S3, the dual-channel time-series reconstruction early warning model is trained using the weighted sum of the original sequence reconstruction error, the differential sequence reconstruction error, and the time-series recursive prediction error as the joint training loss; S4. The reconstructed anomaly score is calculated window by window using the trained dual-channel temporal reconstruction early warning model, and the reconstructed anomaly score is processed by exponential weighted moving average to obtain the final early warning main detection score. S5 determines the warning threshold based on the quantile of the final warning main detection score under the same working conditions, and outputs normal, first-level warning, second-level warning, third-level warning or fourth-level warning accordingly.

2. The method according to claim 1, characterized in that, In step S1, the key early warning indicator is a pre-determined variable representing the battery's operating state, which is at least one of the following: state of charge, voltage, current, temperature, and their statistical characteristics.

3. The method according to claim 1, characterized in that, In step S2, the specific construction steps of the dual-channel time-series reconstruction early warning model include: S21, Construct a dual-channel input structure, wherein the original sequence channel is used to receive multivariate time series, and the first-order difference sequence channel is used to receive difference time series; S22, Construct a feature attention module, generate feature attention weights by pooling, concatenating and softmax mapping multivariate time series and difference time series, and use feature attention weights to weight each feature of the two input channels; S23, construct a dual-channel long short-term memory network autoencoder, and encode the weighted original sequence channel and the first-order difference sequence channel through two long short-term memory networks respectively to obtain the original sequence encoding features and the difference sequence encoding features; concatenate the original sequence encoding features and the difference sequence encoding features and input them into a fully connected mapping layer to obtain a low-dimensional latent feature vector used to characterize the current time window running state; S24, construct a dual-channel decoder and a temporal recursive prediction branch, and input the low-dimensional latent feature vector into the dual-channel decoder and the temporal recursive prediction branch respectively; wherein, the dual-channel decoder is used to output the original sequence reconstruction result and the difference sequence reconstruction result, and the temporal recursive prediction branch is used to output the temporal recursive prediction value after the window.

4. The method according to claim 1, characterized in that, In step S4, the weighted composition of the reconstructed anomaly score is the same as that of the joint training loss in step S3, both using a weighted sum of the original sequence reconstruction error, the differential sequence reconstruction error, and the time-series recursive prediction error.

5. The method according to claim 1, characterized in that, The reconstructed anomaly score is processed by an exponentially weighted moving average to obtain the final early warning main detection score S. alarm (t), where S alarm (t)= α S rec (t)+(1- α )S alarm (t-1), S rec (t) represents the reconstruction anomaly score, S alarm (t-1) represents the warning main detection score at time t-1. α Represents the smoothing coefficient, and within the same set of operating conditions, it is based on the final early warning main detection score S. alarm The distribution of (t) is selected from preset quantiles, which are used as the first-level warning threshold, the second-level warning threshold, the third-level warning threshold and the fourth-level warning threshold.

6. The method according to claim 1, characterized in that, It also includes: Step S6, for the time window of the output warning level, matching the healthy neighbor set in the historical operation window that did not trigger the warning under the same working conditions, using the statistical distribution of the healthy neighbor set as the health reference baseline, and calculating the attribution information of each key warning indicator in the current window in the three dimensions of horizontal deviation, abnormal fluctuation and trend drift.

7. The method according to claim 6, characterized in that, In step S6, based on the healthy neighbor set, the attribution information for each dimension is calculated as follows: the horizontal deviation attribution information is the Z-score of the mean of each indicator in the current window relative to the distribution of the mean of the corresponding indicator in the healthy neighbor set; the volatility anomaly attribution information is the larger of the standard deviation and range of each indicator in the current window relative to the Z-score of the distribution of the corresponding indicator in the healthy neighbor set. The trend drift attribution information is the larger of the absolute value of the slope and the mean of the point-by-point difference magnitude of each indicator in the current window relative to the Z-score of the corresponding indicator distribution in the healthy neighbor set.

8. The method according to claim 6, characterized in that, In step S6, the output attribution information is output synchronously with the early warning time window, reconstruction anomaly score, early warning main detection score, early warning threshold, early warning level and reconstruction error component. The reconstruction error component includes the original sequence reconstruction error, the differential sequence reconstruction error and the time series recursive prediction error, which are used to indicate the main contribution indicators and contribution dimensions of the current early warning.

9. A pre-warning system for the operation status of an energy storage power station based on multi-parameter dual-channel timing reconstruction, used to implement the method described in any one of claims 1-8, characterized in that, include: The data input module is used to acquire multi-parameter operating time series data of the energy storage power station; The time window construction module is used to construct multivariate time series of key early warning indicators based on sliding time windows, and to construct the difference time series obtained by first-order difference of the multivariate time series, and to construct the original sequence channel and the first-order difference sequence channel respectively. The indicator input module is used to receive pre-determined key early warning indicators; The dual-channel time series reconstruction detection module is used to build a dual-channel time series reconstruction early warning model. It performs joint encoding and reconstruction of the multivariate time series and differential time series of key early warning indicators. The dual-channel time series reconstruction early warning model is trained through a feature attention mechanism and three objectives to obtain a reconstruction anomaly score. The three objectives are the original sequence reconstruction error, the differential sequence reconstruction error, and the time series recursive prediction error. The smoothing and threshold determination module is used to smooth the reconstructed anomaly score, determine the first-level warning threshold, second-level warning threshold, third-level warning threshold and fourth-level warning threshold based on the quantile of the final warning main detection score, and output the warning level.

10. The energy storage power station operation status early warning system according to claim 9, characterized in that, Also includes: The statistical attribution module is used to calculate the degree of deviation of statistical features of a time window from a healthy reference in terms of level, volatility and trend by matching the healthy neighbor set, and output statistical attribution information. The results output module is used to output the early warning time window, reconstruction anomaly score, early warning main detection score, early warning threshold, early warning level, reconstruction error components, and statistical three-dimensional attribution information. The reconstruction error components include the original sequence reconstruction error, the differential sequence reconstruction error, and the time-series recursive prediction error.

Images