Multi-dimensional state monitoring and early warning method for networked energy storage integrated system

Abstract

The application provides a multi-dimensional state monitoring and early warning method for a network-constructed energy storage integrated system, belongs to the field of energy storage system safety monitoring, and is used for solving the problems of single monitoring dimension, lack of physical consistency and difficulty in adapting to dynamic working conditions in the related art. The application maps real-time observation data to an initial probability distribution of system hidden states through an encoding neural network; the hidden states are continuously evolved in continuous time based on a neural random differential equation embedded with electrochemical and thermodynamic physical constraints, and an adaptive coupling term is dynamically generated from the hidden states to represent the time-varying coupling relationship between monitoring units; the intrinsic health indicators and comprehensive risk scores with uncertainty measures are decoded from the evolved hidden state distribution through a decoding neural network; and hierarchical early warning decisions are made based on the multi-dimensional monitoring indicators and their uncertainty measures. The application realizes unified modeling of electrochemical-thermal-electrical multi-physical fields, and improves the accuracy, foresight and adaptive ability of early warning.

Description

Technical Field

[0001] This application relates to the technical field of safety monitoring of energy storage systems, and in particular to a multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system. Background Technology

[0002] Grid-based energy storage systems, due to their ability to actively support grid frequency and voltage, have become key equipment for high-proportion renewable energy grid integration. The health status of their internal cells is closely coupled with the system's grid support performance, placing extremely high demands on the real-time performance, accuracy, and multi-dimensional integration of monitoring.

[0003] In existing technologies, battery management systems typically employ threshold-based methods to monitor single physical quantities such as voltage, current, and temperature, or utilize equivalent circuit models to estimate state of charge and state of health. In recent years, data-driven methods such as neural networks have also been used for fault prediction, but these are mostly limited to data fitting and lack mechanistic modeling of the internal electrochemical processes and thermal behavior of the battery. Furthermore, monitoring of grid-type converters primarily focuses on electrical response, which is independent of battery state monitoring and fails to form an integrated and coordinated system.

[0004] However, the aforementioned existing technologies have significant drawbacks: threshold-based monitoring cannot provide early warnings of faults; the equivalent circuit model suffers severe mismatch under high-rate charge and discharge conditions; purely data-driven methods lack physical consistency and are difficult to extrapolate to unknown operating conditions; and the separate monitoring of battery status and grid support performance disrupts the system's coupling relationship, making it impossible to accurately assess high-risk operating modes such as those with faulty batteries. Therefore, there is an urgent need for an early warning method that can integrate multi-physics mechanisms, quantify uncertainties, and achieve coordinated monitoring at the cell and system levels. Summary of the Invention

[0005] This application provides a multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system. It can achieve coordinated monitoring of cell-level health status and system-level support capabilities in a unified modeling manner, solving the problems of existing technologies having a single monitoring dimension and being unable to adapt to dynamic operating conditions.

[0006] Firstly, this application provides a multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system. The method includes the following steps: acquiring real-time observation data of each monitoring unit in a grid-type energy storage system, wherein the monitoring unit includes cells or modules, and the observation data includes at least voltage, current, and temperature; inputting the real-time observation data into an encoding neural network to obtain an initial probability distribution of the system's hidden state, wherein the system's hidden state is used to characterize the internal state of each monitoring unit and the coupling relationship between monitoring units; performing continuous-time evolution of the system's hidden state based on a physically constrained neural stochastic differential equation to obtain the system's hidden state distribution at the current or future time; wherein the drift term of the physically constrained neural stochastic differential equation embeds electrochemical physical constraints and thermodynamic physical constraints, and the drift term includes an adaptive coupling term dynamically generated by the system's hidden state, wherein the adaptive coupling term is used to characterize the coupling relationship between monitoring units evolving over time; inputting the evolved system hidden state distribution into a decoding neural network to obtain multi-dimensional state monitoring indicators and their uncertainty measures for each monitoring unit, wherein the multi-dimensional state monitoring indicators include at least intrinsic health indicators and a comprehensive risk score; and performing graded early warning decisions based on the multi-dimensional state monitoring indicators and their uncertainty measures.

[0007] By adopting the above technical solution, multimodal observation data is encoded into a hidden state distribution. Continuous-time evolution is performed using neural stochastic differential equations with embedded physical constraints. This can simultaneously characterize the electrochemical-thermal coupling behavior inside the battery and the dynamic coupling relationship between cells. Multi-dimensional monitoring indicators with confidence are decoded from the same hidden state, realizing integrated monitoring of grid-type energy storage systems from microscopic mechanisms to macroscopic states, and improving the accuracy and foresight of early warning.

[0008] Further, the drift term is represented as the sum of an electrochemical physical constraint term, a thermodynamic physical constraint term, and an adaptive coupling term; wherein, the electrochemical physical constraint term is used to calculate the rate of state change based on the electrochemical parameters decoded from the hidden state, the thermodynamic physical constraint term is used to calculate the rate of state change based on the temperature field decoded from the hidden state, and the adaptive coupling term is used to calculate the mutual influence between monitoring units based on the adjacency matrix dynamically generated from the hidden state.

[0009] By adopting the above technical solution, the drift function is decomposed into three components with clear physical meaning, so that the evolution process is simultaneously driven by electrochemistry, thermodynamics and coupling relationship, which enhances the interpretability and physical consistency of the model.

[0010] Furthermore, the adaptive coupling term is implemented through a graph neural network. The adjacency matrix of the graph neural network is dynamically generated by the system's hidden state at the current moment, and the adjacency matrix is ​​a time-varying matrix that is updated as the system's hidden state evolves.

[0011] By adopting the above technical solution, adaptive learning of the coupled structure is realized, which can capture the "risk community" that emerges during the aging process, making the modeling of risk propagation more realistic.

[0012] Furthermore, the physically constrained neural stochastic differential equation also includes a diffusion term, which is represented by the product of the diffusion function and the Wiener process differential. The diffusion function is parameterized by the neural network and is used to characterize the effects of unmodeled dynamics and external random perturbations.

[0013] By adopting the above technical solution and introducing a random diffusion term, we can quantify model uncertainty and external disturbances, making the evolution process more consistent with the stochastic characteristics of the actual system and providing confidence information for early warning.

[0014] Furthermore, the initial probability distribution is a Gaussian distribution, and the encoding neural network outputs the mean and variance of the Gaussian distribution.

[0015] By adopting the above technical solution, the initial distribution of the hidden state output by the encoding network provides a probabilistic starting point for subsequent Bayesian evolution, enabling the entire model to have end-to-end uncertainty transmission capability.

[0016] Furthermore, the decoding neural network includes an intrinsic health index decoder, a support capability index decoder, and a comprehensive risk score decoder; wherein, the intrinsic health index decoder is used to decode the intrinsic health index and its uncertainty measure of each monitoring unit from the system latent state distribution, the support capability index decoder is used to decode the support capability index of the grid-type converter from the system latent state distribution, the support capability index including at least virtual inertia response speed and voltage support strength, and the comprehensive risk score decoder is used to decode the comprehensive risk score and its uncertainty measure of each monitoring unit from the system latent state distribution.

[0017] By adopting the above technical solution, multi-dimensional indicators are decoded from the unified hidden state, realizing the coordinated monitoring of battery intrinsic health and grid support capability, reflecting the dual requirements of grid-based energy storage.

[0018] Furthermore, the graded early warning decision includes: determining the early warning level through a fuzzy rule base based on the mean of the comprehensive risk score, the standard deviation of the comprehensive risk score, the mean of the intrinsic health index, the standard deviation of the intrinsic health index, and the temperature change rate; the early warning level includes a warning level, a severe level, and an emergency level; and outputting an early warning signal and root cause information.

[0019] By adopting the above technical solutions and using uncertainty measurement in early warning decision-making, it is possible to distinguish between high-confidence risks and data noise, reduce false alarms, provide root cause information, and enhance the interpretability and practicality of early warnings.

[0020] Furthermore, it also includes a model training step: based on variational inference, maximizing the lower bound of the logarithmic evidence of the observed data sequence, and performing end-to-end joint training on the encoding neural network, the learnable parameters in the physical constraint neural stochastic differential equation, and the decoding neural network, wherein the observed data sequence includes voltage, current, and temperature observations at multiple time points.

[0021] By adopting the above technical solution and using variational inference for end-to-end training, the encoding and evolutionary decoding processes can be optimized simultaneously, enabling the entire model to approximate the real physical system under data-driven conditions and improving the model's generalization ability.

[0022] Furthermore, the method is executed by an edge computing node, which acquires the real-time observation data collected by a sensor array deployed within the energy storage power station.

[0023] By adopting the above technical solution and deploying the method on edge computing nodes, local real-time inference is achieved, reducing the dependence on cloud communication and ensuring the real-time performance and reliability of early warning.

[0024] Furthermore, it also includes a cloud-side update step: using the compressed hidden state or model gradient uploaded by the edge computing nodes of multiple energy storage power stations, updating the global model parameters through distributed training, and then distributing the updated model to the edge computing nodes.

[0025] By adopting the above technical solutions and introducing a cloud-based distributed update mechanism, it is possible to aggregate the experience of multiple power plants and continuously optimize the global model while protecting data privacy, enabling the system to have cross-domain adaptive evolution capabilities.

[0026] In summary, this application has at least the following beneficial effects:

[0027] A multi-dimensional state monitoring and early warning method for grid-type integrated energy storage systems is provided, which realizes unified modeling of electrochemical-thermal-electrical multi-physics field coupling, and improves the accuracy and foresight of monitoring;

[0028] By introducing physical constraints and uncertainty quantification, the interpretability and robustness of the model are enhanced.

[0029] Through structural adaptive learning and distributed updates, the system possesses the ability to continuously evolve and adapt to the differentiated needs of different power plants and operating conditions.

[0030] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0031] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0032] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0033] Figure 2 A flowchart of a multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system is shown in an embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0036] This application provides a multi-dimensional state monitoring and early warning method for grid-type integrated energy storage systems. It achieves unified modeling of electrochemical-thermal-electrical multi-physics fields through physically constrained neural stochastic differential equations, which solves the problems of single monitoring dimensions and lack of physical consistency in existing technologies. This improves the accuracy, foresight and adaptability of early warning, and provides a reliable guarantee for the high-safety operation of grid-type energy storage systems.

[0037] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0038] Reference Figure 1 The operating environment includes a three-layer collaborative hardware architecture consisting of end-side devices, edge-side devices, and cloud-side devices, providing physical support for multi-dimensional status monitoring and early warning methods for grid-type integrated energy storage systems.

[0039] The end-side equipment is deployed at the battery clusters and converter side of the energy storage power station site, including cell-level sensor arrays and grid connection point monitoring units. The cell-level sensor array includes voltage sensors, current sensors, multi-point temperature sensors, pressure sensors, and high-frequency acoustic wave sensors. The voltage sensors employ high-precision Hall effect sensors with a measurement range of 0 to 5 volts and an accuracy better than ±0.05%, used to collect the terminal voltage of each cell. The current sensors employ closed-loop Hall current sensors with a measurement range of -500 A to 500 A and an accuracy better than ±0.1%, used to collect the current flowing through the cell or module. The multi-point temperature sensors use thin-film platinum resistance elements, with three measuring points arranged on the surface of each cell, a measurement range of -40 degrees Celsius to 150 degrees Celsius, and an accuracy better than ±0.5 degrees Celsius, used to collect the surface temperature distribution of the cell. The pressure sensors employ thin-film pressure sensors with a range of 0 to 10 MPa and an accuracy better than ±1%, used to collect the expansion force on the cell terminals or casing. The high-frequency acoustic signature sensor employs a piezoelectric ultrasonic sensor with a center frequency of 100 kHz and a sampling rate of 1 MHz. It is used to acquire the ultrasonic flight time and attenuation coefficient within the battery cell, reflecting the electrolyte wetting state and lithium plating degree. The grid connection monitoring unit includes an electrical quantity acquisition unit and a synchronization phasor measurement unit. The electrical quantity acquisition unit uses a three-phase power converter with a measurement range of 0 to 1000 kW and an accuracy better than ±0.2%, used to acquire the active and reactive power of the grid-connected converter. The synchronization phasor measurement unit uses a high-precision global positioning system for synchronization, with voltage measurement accuracy better than ±0.1% and phase angle measurement accuracy better than ±0.01 degrees, used to acquire the grid connection frequency and voltage phasor. All sensors are connected via a gigabit fiber optic ring network built using an industrial Ethernet switch, employing a precise time protocol to achieve microsecond-level synchronization accuracy, ensuring multimodal data alignment under a unified timestamp.

[0040] The edge-side device is an intelligent early warning gateway deployed inside the energy storage power station. It is equipped with an NVIDIA Jetson AGX Orin series embedded GPU computing card, boasting 275 TOPS of integer computing power, 64 GB of memory, and a 2 TB solid-state drive for local data caching. The intelligent early warning gateway connects to all edge-side sensors via a 10 Gigabit fiber optic switch, acquiring real-time observation data at 10-millisecond intervals and performing the following processing locally: first, outlier removal and linear interpolation imputation of the raw data; then, constructing an observation vector for each battery cell. The intelligent early warning gateway internally deploys a pre-trained physically constrained neural stochastic differential equation model, which includes an encoding neural network, a drift function network, a diffusion function network, and a decoding neural network. Based on the decoded mean risk score, standard deviation of the risk score, mean health score, standard deviation of the health score, and temperature change rate, the intelligent early warning gateway determines the early warning level using a fuzzy rule base. Early warning signals are displayed through the power station monitoring system's human-machine interface and simultaneously uploaded to the station control system via Modbus TCP protocol.

[0041] The cloud-side equipment consists of a central server cluster equipped with eight NVIDIA A100 GPU servers, each with eight GPUs containing 40 GB of video memory, and a 100 TB distributed storage system. The cloud-side equipment connects to all intelligent early warning gateways via a 5G private network, receiving compressed hidden state sequences uploaded by each gateway weekly. The cloud-side equipment performs end-to-end joint training based on a variational inference framework. After training, the updated model parameters are packaged into a model file and distributed to each intelligent early warning gateway via the 5G private network. The gateways then load and replace the original model, completing the model update and iteration.

[0042] Throughout the operating environment, end-side devices collect real-time data such as voltage, current, temperature, pressure, ultrasonic characteristics, power, and frequency at 10-millisecond intervals, and transmit this data synchronously to edge-side devices via a gigabit fiber optic ring network. The edge-side devices complete the entire process of encoding, evolution, decoding, and early warning within 10 milliseconds, outputting the health status, risk score, and warning level of each battery cell. Cloud-side devices aggregate the compressed hidden states of each gateway weekly for a global model update, with each update taking approximately 2 hours. The updated model takes effect the following week. End-side devices, edge-side devices, and cloud-side devices are connected via industrial Ethernet and a 5G private network, jointly supporting the acquisition of real-time observation data, the continuous temporal evolution of the system's hidden states, the decoding of multi-dimensional monitoring indicators, and the complete process of hierarchical early warning decision-making.

[0043] Based on the above operating environment, this application also discloses a multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system. Figure 2 A flowchart illustrating a multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system according to an embodiment of this application is shown. (Refer to...) Figure 2 This method is executed by an edge computing node, which acquires real-time observation data collected by a sensor array deployed within an energy storage power station. The method specifically includes the following steps:

[0044] S1: Obtain real-time observation data of each monitoring unit in the grid-type energy storage system. The monitoring unit includes cells or modules, and the observation data includes at least voltage, current, and temperature.

[0045] In this step, the edge computing node acquires real-time observation data from a sensor array deployed within the energy storage power station via an industrial Ethernet or fiber optic ring network at 10-millisecond intervals. The sensor array includes a cell-level sensor array and a grid connection point monitoring unit. The cell-level sensor array contains voltage sensors, current sensors, multi-point temperature sensors, pressure sensors, and high-frequency acoustic wave sensors. The grid connection point monitoring unit includes an electrical quantity acquisition unit and a synchronous phasor measurement unit. For the cell-level sensor array, each monitoring unit i at time... The observed values ​​form a multidimensional vector, represented as:

[0046]

[0047] in, The terminal voltage value collected by the voltage sensor is expressed in volts (V). The current value collected by the current sensor is in amperes (A). The surface temperature values ​​collected by a multi-point temperature sensor are in degrees Celsius (°C). The expansion force value collected by the pressure sensor is expressed in megapascals (MPa). The ultrasonic feature value extracted by the high-frequency acoustic signature sensor is obtained through envelope detection and peak hold of the original ultrasonic signal and is dimensionless. All cell-level sensors are synchronized using a precise time protocol to ensure accurate data acquisition. It has microsecond-level consistency and the interval between adjacent sampling times. Milliseconds. The grid-connected monitoring unit synchronously collects electrical quantity data from the grid-connected converter, including active power. reactive power and grid connection frequency These data are at the same timestamp as the cell-level data. The data is then aligned to form a complete multimodal observation dataset. The observation data includes at least voltage, current, and temperature, as well as expansion force, ultrasonic characteristics, and active power, reactive power, and frequency response data of the grid-type converter.

[0048] S2: Input the real-time observation data into the encoding neural network to obtain the initial probability distribution of the system hidden state. The system hidden state is used to characterize the internal state of each monitoring unit and the coupling relationship between monitoring units.

[0049] In this step, the edge computing nodes first normalize the raw observation data collected in step S1. For each monitoring unit i at time... 5-dimensional observation vector Z-score normalization is performed by calculating the mean and standard deviation over the entire data window to obtain the normalized input vector. .all The input matrix is ​​composed of the normalized observation vectors of each monitoring unit. ,in This represents the total number of monitoring units. The edge computing nodes will normalize the observation matrix. Input encoding neural network. The encoding neural network adopts the encoder structure of variational autoencoder, and its goal is to map high-dimensional observation data to a probability distribution on a low-dimensional latent space. The encoding neural network consists of three fully connected layers. The first layer has an input dimension of 5 and an output dimension of 128, and uses ReLU as the activation function; the second layer has an input dimension of 128 and an output dimension of 128, and uses ReLU as the activation function; the third layer is the output layer, with an input dimension of 128 and an output dimension of 64. The first 32 dimensions correspond to the logarithm of the latent variable mean, and the last 32 dimensions correspond to the logarithm and variance of the latent variable. The forward propagation process of the encoding neural network can be represented as:

[0050]

[0051]

[0052]

[0053] in, , , This is the weight matrix. , , These are bias vectors; the parameters are all learnable parameters of the encoding neural network, obtained through subsequent model training steps. Output vector. The logarithm of the mean of the latent variables. Let logarithm and variance be the latent variables. The actual mean and variance of the latent variables are obtained through exponential transformation:

[0054]

[0055]

[0056] in, For the first The mean matrix of the hidden states of the system at time t. For the first The variance matrix of the hidden state of the system at time t. From this, the initial probability distribution of the hidden state of the system is obtained:

[0057]

[0058] in, For the first The hidden state matrix of the system at time t, each row of which Corresponding to the The hidden state vector of each monitoring unit is 32-dimensional. The initial probability distribution is Gaussian, and the encoding neural network outputs the mean and variance of the Gaussian distribution. This hidden state vector not only encodes the electrochemical internal state of the monitoring unit, but also implicitly encodes the coupling relationship between monitoring units through the topological structure of the hidden space. The introduction of the initial probability distribution gives the system's hidden state a probabilistic property, and its mean... The variance represents the most likely value of the hidden state. It characterizes the degree of uncertainty caused by observation noise and model uncertainty.

[0059] S3: Based on the physical constraint neural stochastic differential equation, the hidden state of the system is continuously evolved over time to obtain the distribution of the hidden state of the system at the current time or in the future time.

[0060] The drift term of the physical constraint neural stochastic differential equation incorporates electrochemical physical constraints and thermodynamic physical constraints, and the drift term includes an adaptive coupling term dynamically generated by the hidden state of the system. The adaptive coupling term is used to characterize the coupling relationship between monitoring units that evolves over time.

[0061] In this step, edge computing nodes perform continuous-time evolution of the system's hidden states based on physically constrained neural stochastic differential equations. This is achieved using the method obtained in step S2. Initial distribution of hidden states at time 1 Starting from the given point, any future time can be obtained by solving the following stochastic differential equation. System hidden state distribution:

[0062]

[0063] in, for The hidden state matrix of the system at time t. The total number of monitoring units. For hidden state dimensions; Let be the drift function, parameterized by the neural network, and let the parameter set be denoted as . It is used to characterize the deterministic evolution trend of the hidden state of a system; The diffusion function is also parameterized by the neural network, and the parameter set is denoted as . It is used to characterize the random uncertainty in the evolutionary process; For the standard Wiener process increment, each dimension is independent. Drift function. Specifically, it can be represented as the sum of three items:

[0064]

[0065] The drift term is represented as the sum of the electrochemical physical constraint term, the thermodynamic physical constraint term, and the adaptive coupling term. Electrochemical physical constraint term This study reveals the kinetics of electrochemical reactions within lithium-ion batteries. It first utilizes a lightweight electrochemical decoder to analyze the reactions from their latent states. The current electrochemical parameters are decoded: lithium-ion concentration in the negative electrode solid phase. Concentration of lithium ions in the positive electrode solid phase and overpotential Based on a simplified single-particle model, the rate of change of electrochemical state is calculated using the discretized form of the Butler-Folmer equation and Fick's second law, specifically expressed as:

[0066]

[0067] in, The linear transformation matrix from the electrochemical parameter space to the hidden state space is obtained through training; the partial derivative terms are calculated using the electrochemical mechanism formula based on the current current density and temperature. Thermodynamic physical constraint terms are also included. This demonstrates the thermal balance between heat generation and dissipation in a battery. This process is first implemented using a thermodynamic decoder from its latent state. Decoding the internal temperature field of the battery cell and heat production rate Based on the lumped heat capacity model, the rate of temperature change is determined by the energy conservation equation, with thermodynamic and physical constraint terms. This is the mapping from the temperature change rate obtained after discretizing the heat conduction equation to the hidden state change rate. The electrochemical physical constraint term is used to calculate the state change rate based on the electrochemical parameters decoded from the hidden state, and the thermodynamic physical constraint term is used to calculate the state change rate based on the temperature field decoded from the hidden state.

[0068] Adaptive Coupling Terms This is used to characterize the interactions between monitoring units, such as heat conduction and electrical stress sharing. It is implemented using a graph neural network, with the adjacency matrix at its core. Dynamically generated from the system's hidden state at the current moment:

[0069]

[0070] in, It is a two-layer fully connected network, and the input is the hidden state matrix. The output is The attention score matrix is ​​used as an example, and the softmax function is used to normalize each row to obtain the adjacency matrix. . elements in Indicates at time Monitoring Unit For monitoring units The influence weight, and satisfying , The adjacency matrix is ​​a time-varying matrix, updated as the system's hidden state evolves. The adaptive coupling term is used to calculate the mutual influence between monitoring units based on the dynamically generated adjacency matrix from the hidden state. This adaptive coupling term is implemented using a graph neural network, whose adjacency matrix is ​​dynamically generated by the system's hidden state at the current moment, and is a time-varying matrix updated as the system's hidden state evolves. Based on the dynamically generated adjacency matrix, the adaptive coupling term is calculated as follows:

[0071]

[0072] in, The weight matrix is ​​a learnable weight matrix, obtained through training. Diffusion function. Implemented by a two-layer fully connected network, the input is the hidden state at the current time. and time The embedding vector is output as a diagonal covariance matrix. The diagonal elements, that is:

[0073]

[0074] The physically constrained neural stochastic differential equations also include a diffusion term, which is represented by the product of a diffusion function and the Wiener process differential. The diffusion function is parameterized by the neural network to characterize the effects of unmodeled dynamics and external random perturbations. Edge computing nodes use a neural stochastic differential equation solver to numerically integrate the above equations. From the initial time... Start with step size Integrate forward from millisecond to the current time. This yields the hidden state distribution at each time step. The Euler-Maruyama method is used in the solution process, namely:

[0075]

[0076] in, This is a standard normally distributed random vector used to implement random sampling of the diffusion term. Through multiple parallel samplings (e.g., 32 Monte Carlo samplings), we can obtain... The sample set of the hidden states of the system at time t is used to estimate its probability distribution. This distribution contains the posterior uncertainty of all information from the initial observation to the current time. Through the parallel sampling described above, the current time can be obtained. The system's hidden state sample set is used to calculate the mean matrix of the hidden states. Sum of standard deviation matrix As a hidden state distribution The parameters are used in subsequent steps.

[0077] S4: Input the evolved system hidden state distribution into the decoding neural network to obtain the multi-dimensional state monitoring indicators and their uncertainty measures for each monitoring unit. The multi-dimensional state monitoring indicators include at least intrinsic health indicators and comprehensive risk scores.

[0078] In this step, the edge computing node takes the system latent state distribution evolved in step S3 as input and maps it to physically meaningful multi-dimensional monitoring indicators and their uncertainty measures through a decoding neural network. The decoding neural network includes an intrinsic health indicator decoder, a support capability indicator decoder, and a comprehensive risk score decoder. Since step S3 outputs the probability distribution of the latent state, this step employs a Monte Carlo sampling method to obtain the distribution characteristics of the monitoring indicators: starting from the current time... System hidden state distribution Extraction Sample ,in Take 32. Each sample Obtained through reparameterization techniques, i.e. ,in , This represents element-wise multiplication. This set of samples captures the posterior uncertainty of the system state.

[0079] The intrinsic health indicator decoder receives hidden state samples from each monitoring unit. As input, the output is the intrinsic health index estimate corresponding to this sample. The decoder is a three-layer fully connected network with the following structure: input layer dimension... The hidden layer has a dimension of 64 and uses the ReLU activation function; the output layer has a dimension of 1 and no activation function. To integrate health information from both electrochemical aging and thermal state dimensions, the intrinsic health index is defined as:

[0080]

[0081] in, From hidden state samples The negative electrode solid-phase diffusion coefficient is decoded using a lightweight electrochemical decoding head; This is the baseline value for the diffusion coefficient at the beginning of the battery's lifespan, preset according to the battery's factory data. ; The lifetime end diffusion coefficient threshold is set to . ; The highest internal temperature of the battery cell, decoded from the hidden state sample, is obtained using a thermodynamic decoding head. The ambient temperature is collected in real time by the power plant's environmental sensors, with a typical value of 25°C. The critical temperature for thermal runaway is set to 80°C based on the characteristics of the battery materials. and For the weighting coefficients, satisfying In this embodiment, we take , Through the analysis of The health indicators of each sample were statistically analyzed to obtain the mean of the intrinsic health indicators for each monitoring unit. and standard deviation :

[0082]

[0083]

[0084] in, Characterization monitoring unit The most likely value for the health status is 0, and the value ranges from 1 to 0. The closer it is to 1, the healthier the person is. This characterizes the uncertainty of the estimate.

[0085] The support capability index decoder is used to extract support performance indices related to grid-type converters from the system's hidden states. The input to this decoder is either the portion of the hidden state corresponding to the PCS node, or a PCS-level feature vector obtained by aggregating all cell hidden states through global pooling. (For example, average pooling is used). The decoder is a two-layer fully connected network with a hidden layer dimension of 32 and an output layer dimension of 2, corresponding to the virtual inertia response speed, respectively. and voltage support strength :

[0086]

[0087] The support capability indicators include at least virtual inertia response speed and voltage support strength. This represents the virtual inertia response speed, measured in seconds. A smaller value indicates a faster inertia-supported response. Voltage support strength, dimensionless, characterizes reactive current injection gain.

[0088] A comprehensive risk scoring decoder is used to assess the immediate risk of failure for each monitoring unit. This decoder relies not only on the hidden state of the current node but also incorporates information from surrounding nodes through an adaptive neighborhood aggregation mechanism. For each node... Its risk scoring sample Calculated in the following way:

[0089]

[0090] in, For nodes Hidden state samples; For nodes exist The neighborhood at time step S3 is the adjacency matrix dynamically generated in step S3. Decide; For attention weights, you can directly take... elements in ; This indicates vector concatenation; This is a three-layer fully connected network with a hidden layer dimension of 64 and an output layer dimension of 1. The activation function is sigmoid, and the risk score is compressed to... Interval. The risk scores of all samples are statistically analyzed to obtain the mean of the comprehensive risk score. and standard deviation :

[0091]

[0092]

[0093] in, For nodes The instantaneous risk probability, the closer to 1, the higher the risk; This addresses the uncertainty in risk estimation. Through parallel computation of the three decoders mentioned above, this step decouples three monitoring indicators from the unified hidden state distribution: intrinsic health indicators, support capability indicators, and comprehensive risk scores, each accompanied by an uncertainty measure.

[0094] S5: Based on the multi-dimensional state monitoring indicators and their uncertainty measures, execute hierarchical early warning decisions.

[0095] In this step, the edge computing node performs tiered early warning decisions based on the multi-dimensional monitoring indicators and their uncertainty measures obtained in step S4. The tiered early warning decisions include: determining the early warning level using a fuzzy rule base based on the mean of the comprehensive risk score, the standard deviation of the comprehensive risk score, the mean of the intrinsic health indicators, the standard deviation of the intrinsic health indicators, and the temperature change rate; the early warning levels include alert level, severe level, and emergency level; and outputting early warning signals and root cause information.

[0096] Temperature change rate The temperature data collected in step S1 is obtained through differential calculation:

[0097]

[0098] in, and Monitoring units The surface temperature at the current moment and the previous sampling moment, The sampling interval is milliseconds. The fuzzy rule base is the core of this step, employing a Mamdani-type fuzzy inference system, consisting of three stages: fuzzification, rule inference, and defuzzification. The fuzzification stage converts input variables into membership degrees of fuzzy sets. For each input variable, three fuzzy subsets are defined: Low, Medium, and High. The membership function uses a Gaussian function, with the general form:

[0099]

[0100] The membership function parameters for each input variable are pre-set based on historical data and expert experience. The rule inference phase employs 27 fuzzy rules, in the form of "IF (condition) THEN (conclusion)", where the conclusion is a fuzzy set of warning levels. Warning levels are defined as three fuzzy subsets: Alert, Warning, and Critical. The output membership function uses a single-point fuzzy set, meaning each level corresponds to a specific value: Alert = 1, Warning = 2, Critical = 3. The activation strength of each rule is calculated by taking the minimum value of the membership degrees of each condition.

[0101]

[0102] in, For the first The activation strength of the rule, For the first Rule No. The membership degree of the fuzzy subset corresponding to each input variable. arrive Corresponding in sequence , , , , In the defuzzification stage, the conclusions of all activated rules are fuzzy set and combined into a single clear output value. A weighted average method is then used to calculate the warning level value. :

[0103]

[0104] in, For the first The single-point value (1, 2, or 3) corresponding to the conclusion of each rule. Calculated. For continuous values, the final warning level is determined by threshold division: if This is the prompt level. Then it is classified as severe. This is classified as an emergency level. When the warning level reaches the severe level or above, the edge computing node also needs to output root cause information. Root cause information is obtained by analyzing the contribution of each input variable to the activation intensity of the fuzzy rule; the variable with the largest contribution factor is considered the primary root cause. The edge computing node encapsulates the warning level and root cause information into a warning signal, displays it through the power plant monitoring system's human-machine interface, and simultaneously uploads it to the station control system via the Modbus TCP protocol. It can also execute corresponding operations according to preset linkage strategies.

[0105] Furthermore, embodiments of this application also include a model training step: based on variational inference, maximizing the lower bound of logarithmic evidence for the observed data sequence, and performing end-to-end joint training on the encoded neural network, the learnable parameters in the physical constraint neural stochastic differential equation, and the decoded neural network, wherein the observed data sequence includes voltage, current, and temperature observations at multiple time points.

[0106] In this step, the cloud-side devices undergo end-to-end joint training based on a variational inference framework. Observation data sequence It consists of voltage, current, and temperature observations collected at multiple time points in step S1. Specifically, a length of [length missing] is extracted from the historical operating data of each energy storage power station. A continuous time window, each containing 100 sampling points, with a sampling interval of... Milliseconds, with a total time length of 1 second. For each time window, the observed data sequence is represented as follows: Each of them Includes all Each monitoring unit at time The voltage, current, and temperature observations were used. These observation data sequences were extracted from historical databases of multiple energy storage power stations to form a training dataset, which served as training samples. ,in The total number of samples is taken in this embodiment. .

[0107] The loss function for variational inference is the lower bound of the logarithmic evidence for the observed data sequence, and its mathematical expression is:

[0108]

[0109] in, This is the set of all learnable parameters, including the parameters of the encoding neural network, the drift and spread function parameters in the physically constrained neural stochastic differential equations, and the parameters of the decoding neural network. The first term is the reconstruction loss, representing the loss given the sequence of hidden states. Under these conditions, observation data The expected value of the log-likelihood. Among them, The decoding neural network defined by step S4 is typically assumed to follow a Gaussian distribution:

[0110]

[0111] in, To decode the observed reconstructed values ​​output by the neural network, To observe the standard deviation of the noise, it is set to 0.01 in this embodiment. The second term is the KL divergence term, used to constrain the approximate posterior distribution. Do not deviate excessively from the prior distribution. The prior distribution employs a drift-free Wiener process. During training, the Adam optimizer is used for parameter updates, with an initial learning rate of 0.001, decreasing to 0.9 times the original rate every 10 epochs, a batch size of 32, and 200 training epochs. In each training iteration, data is generated from the training dataset. A batch of data is randomly sampled, and forward propagation is performed on each sample to calculate the reconstruction loss and KL divergence term. The negative log evidence lower bound is obtained as the loss value. The gradient is calculated through backpropagation to update all learnable parameters. After training is complete, the cloud-based device packages the updated encoding neural network parameters, drift function network parameters, diffusion function network parameters, and decoding neural network parameters into a model file, and distributes it to each intelligent early warning gateway via the 5G private network. The gateway then loads the file and replaces the original model, completing the model update and iteration.

[0112] This application embodiment also includes a cloud-side update step: using the compressed hidden state or model gradient uploaded by the edge computing nodes of multiple energy storage power stations, updating the global model parameters through distributed training, and distributing the updated model to the edge computing nodes.

[0113] In this step, cloud-side devices connect to all intelligent early warning gateways via a 5G private network, receiving compressed hidden state sequences uploaded by each gateway weekly. The compressed hidden state sequences are dimensionality-reduced using principal component analysis, retaining the first 16 principal components. The hidden state data at each time point is compressed to 16 dimensions, with a time window length of 1000 sampling points. Specifically, each edge computing node locally accumulates a mean hidden state sequence for 1000 consecutive time points. To form a data matrix subscript Indicates the first An energy storage power station. On the cloud side, first collect the data matrix of all nodes and calculate the global covariance matrix:

[0114]

[0115] in, To determine the total number of edge computing nodes participating in the update, this embodiment assumes... For the covariance matrix Perform eigenvalue decomposition:

[0116]

[0117] in, The eigenvector matrix, This is an eigenvalue diagonal matrix. The projection matrix is ​​formed by taking the eigenvectors corresponding to the first 16 largest eigenvalues. Then the compressed hidden state sequence uploaded by each node Calculated using the following formula:

[0118]

[0119] Cloud-side devices utilize the compressed hidden state uploaded by each node Alternatively, the model gradients can be directly uploaded for distributed training to update the global model parameters. This embodiment uses the federated averaging algorithm to implement distributed training. Assume the global model parameters are... Each edge computing node calculates the loss function locally using its private data. And calculate the gradient After collecting the gradients from all nodes, the cloud-side device calculates the weighted average gradient:

[0120]

[0121] in, For the first The number of local samples per node is uniformly set to 1000 in this embodiment. Then, the cloud side updates the model parameters using global gradients:

[0122]

[0123] in, The learning rate is set to 0.001 in this embodiment. If the node uploads local model parameters instead of gradients, the cloud side directly performs a weighted average of the parameters. After the update is complete, the cloud-side device will upload the new global model parameters. The model is packaged into a model file and distributed to all edge computing nodes via a 5G private network. Each node receives the file, loads the model, replaces its original local model, and uses it for subsequent real-time monitoring and early warning.

[0124] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0125] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system, characterized in that, Includes the following steps: Real-time observation data of each monitoring unit in a grid-type energy storage system are acquired. The monitoring unit includes cells or modules, and the observation data includes at least voltage, current, and temperature. The real-time observation data is input into an encoding neural network to obtain the initial probability distribution of the system's hidden state. The system's hidden state is used to characterize the internal state of each monitoring unit and the coupling relationship between monitoring units. Based on the physical constraint neural stochastic differential equation, the hidden state of the system is continuously evolved over time to obtain the distribution of the hidden state of the system at the current time or in the future time. The drift term of the physical constraint neural stochastic differential equation incorporates electrochemical physical constraints and thermodynamic physical constraints, and the drift term includes an adaptive coupling term dynamically generated by the hidden state of the system. The adaptive coupling term is used to characterize the coupling relationship between monitoring units that evolves over time. The system's hidden state distribution obtained through evolution is input into a decoding neural network to obtain multi-dimensional state monitoring indicators and their uncertainty measures for each monitoring unit. The multi-dimensional state monitoring indicators include at least intrinsic health indicators and comprehensive risk scores. Based on the aforementioned multi-dimensional state monitoring indicators and their uncertainty measures, hierarchical early warning decisions are implemented.

2. The multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system according to claim 1, characterized in that, The drift term is represented as the sum of the electrochemical physical constraint term, the thermodynamic physical constraint term, and the adaptive coupling term; The electrochemical physical constraint term is used to calculate the rate of state change based on the electrochemical parameters decoded from the hidden state, the thermodynamic physical constraint term is used to calculate the rate of state change based on the temperature field decoded from the hidden state, and the adaptive coupling term is used to calculate the mutual influence between monitoring units based on the adjacency matrix dynamically generated from the hidden state.

3. The multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system according to claim 2, characterized in that, The adaptive coupling term is implemented through a graph neural network. The adjacency matrix of the graph neural network is dynamically generated by the system's hidden state at the current moment, and the adjacency matrix is ​​a time-varying matrix that is updated as the system's hidden state evolves.

4. The multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system according to claim 1, characterized in that, The physical constraint neural stochastic differential equation also includes a diffusion term, which is represented by the product of the diffusion function and the Wiener process differential. The diffusion function is parameterized by the neural network to characterize the effects of unmodeled dynamics and external random perturbations.

5. The multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system according to claim 1, characterized in that, The initial probability distribution is a Gaussian distribution, and the encoding neural network outputs the mean and variance of the Gaussian distribution.

6. The multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system according to claim 1, characterized in that, The decoding neural network includes an intrinsic health indicator decoder, a support capability indicator decoder, and a comprehensive risk score decoder. The intrinsic health index decoder is used to decode the intrinsic health index and its uncertainty measure of each monitoring unit from the system latent state distribution. The support capability index decoder is used to decode the support capability index of the grid-type converter from the system latent state distribution. The support capability index includes at least virtual inertia response speed and voltage support strength. The comprehensive risk score decoder is used to decode the comprehensive risk score and its uncertainty measure of each monitoring unit from the system latent state distribution.

7. The multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system according to claim 6, characterized in that, The tiered early warning decision-making includes: Based on the mean of the comprehensive risk score, the standard deviation of the comprehensive risk score, the mean of the intrinsic health index, the standard deviation of the intrinsic health index, and the temperature change rate, a fuzzy rule base is used to determine the warning level, which includes alert level, severe level, and emergency level. Output early warning signals and root cause information.

8. The multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system according to claim 1, characterized in that, It also includes the model training step: Based on variational inference, the lower bound of logarithmic evidence for the observed data sequence is maximized, and the learnable parameters in the encoding neural network, the physical constraint neural stochastic differential equation, and the decoding neural network are jointly trained end-to-end. The observed data sequence includes voltage, current, and temperature observations at multiple time points.

9. The multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system according to claim 1, characterized in that, The method is executed by an edge computing node, which acquires the real-time observation data collected by a sensor array deployed in an energy storage power station.

10. The multi-dimensional state monitoring and early warning method for a grid-type integrated energy storage system according to claim 9, characterized in that, It also includes cloud-side update steps: By utilizing the compressed hidden states or model gradients uploaded by edge computing nodes of multiple energy storage power stations, the global model parameters are updated through distributed training, and the updated model is then distributed to the edge computing nodes.

Images