Grid-connected fault comprehensive evaluation strategy for network-constructed hybrid energy storage power station

Abstract

The application discloses a grid-connected fault comprehensive evaluation strategy for a network-constructed hybrid energy storage power station. By adopting a hybrid evaluation framework that fuses an analytic hierarchy process, a convolutional neural network, a long short-term memory network and an attention mechanism, the evaluation precision of a complex fault scene is improved while the model interpretability is maintained, and accurate quantification of the grid-connected fault severity of the network-constructed hybrid energy storage power station is realized.

Description

Technical Field

[0001] This invention relates to the field of grid connection fault comprehensive assessment technology, and in particular to a grid-connected hybrid energy storage power station grid connection fault comprehensive assessment strategy. Background Technology

[0002] As a core component of the power system, new energy sources require a long-term dynamic evolution process, during which grid-based hybrid energy storage technology will play a crucial supporting role. Grid-based energy storage technology can provide voltage and frequency support, alleviating the problems of insufficient short-circuit capacity and lack of inertia in grids with a high proportion of new energy sources. Hybrid energy storage technology, through its complementary advantages, can improve system economy and operational efficiency, effectively mitigating power fluctuations from distributed sources. However, grid-based hybrid energy storage systems face certain fault risks in practical engineering applications: grid-side faults can easily trigger grid-connected overcurrent and reactive power imbalance problems; voltage drops caused by three-phase faults may induce converter inrush current and loss of synchronization; and DC-side short-circuit faults threaten the safety of inductors and energy storage devices. Therefore, it is urgent to establish a rapid fault assessment mechanism.

[0003] Currently, research on the quantitative assessment of grid-connected fault severity in grid-connected hybrid energy storage power stations remains scarce. Existing fault assessment methods are mostly developed based on fault simulation sample data from other systems, and can be broadly categorized into three types: analytical methods based on physical models, multi-criteria decision-making methods, and data-driven methods. Analytical methods based on physical models assess the impact of faults by establishing precise mathematical models, such as hierarchical fault tree methods, accident chain models, and transient energy function methods. Multi-criteria decision-making methods, based on the entropy-weighted fusion analytic hierarchy process (AHP), have the advantages of flexible weight allocation and effective integration of qualitative and quantitative indicators, and are widely used. However, analytical methods based on physical models and multi-criteria decision-making methods require a large amount of domain-specific knowledge or prior system information, and this reliance on expert intervention limits their practical application in industrial environments. Data-driven methods can overcome these limitations. This type of method can utilize machine learning to mine fault features from massive amounts of data, with typical applications such as convolutional neural networks (CNN) and long short-term memory neural networks (LSTM). However, data-driven methods still face challenges in practical applications, including significant data dependence, high computational costs, and insufficient model interpretability. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a comprehensive assessment strategy for grid-connected faults in grid-connected hybrid energy storage power stations. This strategy aims to integrate the advantages of data-driven and multi-criteria decision-making, thereby improving the assessment accuracy for complex fault scenarios in grid-connected hybrid energy storage power stations while maintaining model interpretability.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A comprehensive assessment strategy for grid-connected faults in a grid-connected hybrid energy storage power station includes:

[0007] S1. Based on the grid connection faults of grid-connected hybrid energy storage power stations, establish a quantitative evaluation index system;

[0008] S2, The initial weights of all characteristic indicators in the evaluation index system are calculated using the analytic hierarchy process (AHP).

[0009] S3 uses a convolutional neural network to extract spatial features of grid-connected fault data;

[0010] S4, A long short-term memory network is used to model the temporal sequence of the spatial features and output the hidden state sequence;

[0011] S5, The hidden state sequence is dynamically weighted using an attention mechanism to obtain the final weights;

[0012] S6. Based on the final weight and the evaluation index system, perform dynamic weight quantification scoring on grid connection faults of grid-connected hybrid energy storage power stations.

[0013] In some implementations, in S1, the process of establishing the evaluation index system includes:

[0014] Multi-dimensional indicators are extracted to identify the characteristics of each grid-connected fault, resulting in multiple feature indicators. Then, a correlation analysis is performed on the fault type and characteristics of the grid-connected fault, and the feature indicators are constructed into a three-layer evaluation indicator system. In this evaluation indicator system, there is only one primary indicator, which is defined as the severity of the grid-connected fault. The secondary indicators include DC characteristic dimension, energy characteristic dimension, and AC characteristic dimension. The tertiary indicators are the feature indicators and are attached to the corresponding secondary indicators.

[0015] In some implementations, the types of grid-connected faults include at least: lithium battery short-circuit fault, lithium battery open-circuit fault, supercapacitor short-circuit fault, supercapacitor open-circuit fault, flywheel energy storage short-circuit fault, hydrogen fuel cell short-circuit fault, hydrogen fuel cell open-circuit fault, hydrogen electrolyzer short-circuit fault, hydrogen electrolyzer open-circuit fault, DC bus short-circuit fault, DC bus open-circuit fault, system grid-connected to islanded mode, three-phase grid frequency drop, three-phase grid voltage drop, three-phase grid voltage imbalance, single-phase ground fault, two-phase ground fault, three-phase short-circuit fault, and phase-to-phase short-circuit fault.

[0016] In some implementations, in S2, the calculation of the initial weights of all characteristic indicators in the evaluation index system using the analytic hierarchy process includes:

[0017] S201, the first-level indicator, the second-level indicator and the third-level indicator are respectively defined as target-level indicator, criterion-level indicator and scheme-level indicator;

[0018] S202, Based on the Saaty 1-9 scaling method, compare any two of the criteria layer indicators and construct a 3×3 judgment matrix to determine the relative importance of each of the criteria layer indicators;

[0019] S203, based on the relative importance, the eigenvector method is used to extract the criterion layer weights and weight vectors of each criterion layer indicator from the judgment matrix;

[0020] S204, check the consistency of the judgment matrix according to the criterion layer weights and the weight vector, and determine whether to accept the criterion layer weights according to the test results. If not, adjust the elements in the judgment matrix and repeat steps S203 and S204 until the consistency test is passed.

[0021] S205, for each scheme layer indicator attached to the criterion layer indicator, repeat steps S203 and S204 to calculate the local weight of the scheme layer indicator. The local weight is multiplied by the corresponding criterion layer weight to obtain the global weight of the third-level indicator, until the initial weight matrix of all the third-level indicators is obtained.

[0022] In some implementations, in S3, the grid connection fault data includes: the voltage and current output by the DC side lithium-ion battery, the voltage and current output by the supercapacitor, the voltage and current output by the flywheel energy storage, the voltage and current output by the hydrogen fuel cell, the voltage and current output by the hydrogen electrolyzer, the DC bus voltage and current, the AC side voltage and current, and the AC side active and reactive power when a grid connection fault occurs in the grid-connected hybrid energy storage power station.

[0023] In some implementations, in S3, the extraction of spatial features from grid-connected fault data using a convolutional neural network includes:

[0024] S301, Perform filtering preprocessing on the grid connection fault data;

[0025] S302, obtain the variable data column related to each feature indicator, define the mapping relationship between the feature indicator and the variable data column, and extract the CNN features specific to the feature indicator;

[0026] S303, construct a convolutional neural network that omits pooling layers and fully connected layers, and then extract the spatial features of the CNN features within the local time window;

[0027] S304, the Flatten layer of the convolutional neural network is invoked to flatten the spatial features, and the feature map obtained after flattening the spatial features is converted into a feature vector; then the time steps after convolution of the spatial features are reorganized.

[0028] In some implementations, in S4, the temporal modeling of the spatial features using a long short-term memory network includes:

[0029] A long short-term memory network containing an input gate, a forget gate, an output gate, and memory cells is constructed. The spatial features are input into the long short-term memory network, and the temporal sequence of the spatial features is modeled to output the hidden state sequence.

[0030] In some implementations, in S5, the dynamic weighting of the hidden state sequence using an attention mechanism includes:

[0031] S501, the hidden state sequence is used as a query vector;

[0032] S502, calculate the attention score of the hidden state and the corresponding query vector at each time step;

[0033] S503, the attention score is converted into attention weights using the Softmax function;

[0034] S504, the attention weights are weighted and summed with all the hidden states to generate a context vector;

[0035] S505, the context vector is mapped to a dynamic weight correction factor;

[0036] S506, calculate the final weights for the evaluation index system based on the dynamic weight correction factor.

[0037] The beneficial effects of this invention are as follows: by adopting a hybrid evaluation framework that integrates the analytic hierarchy process, convolutional neural network, long short-term memory network and attention mechanism, the evaluation accuracy of complex fault scenarios is improved while maintaining the interpretability of the model, and the severity of grid-connected faults of grid-connected hybrid energy storage power stations is accurately quantified. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the comprehensive assessment strategy for grid-connected faults of a grid-connected hybrid energy storage power station disclosed in an embodiment of the present invention.

[0039] Figure 2 This is a system architecture diagram of the grid-connected fault comprehensive assessment strategy for a grid-connected hybrid energy storage power station disclosed in an embodiment of the present invention;

[0040] Figure 3This is a bar chart showing the weight value changes of a grid-connected hybrid energy storage power station under 19 typical grid-connected fault conditions, as disclosed in an embodiment of the present invention.

[0041] Figure 4 This is a quantitative scoring comparison chart of the grid-connected hybrid energy storage power station disclosed in the embodiments of the present invention under 19 typical grid-connected fault conditions. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the content of this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to this invention are shown in the accompanying drawings, not all of them.

[0043] This embodiment proposes a comprehensive assessment strategy for grid-connected faults in grid-connected hybrid energy storage power stations. It integrates a data-driven and multi-criteria decision-making AHP+CNN-LSTM-Attention comprehensive assessment strategy. AHP provides prior weights that conform to the laws of electrical physics, and combined with the spatiotemporal features of faults extracted by the CNN-LSTM-Attention network, a weight correction factor is dynamically generated. This strategy quantifies and ranks the severity of grid-connected fault types in different grid-connected hybrid energy storage power stations, improving the assessment accuracy of complex fault scenarios while maintaining model interpretability.

[0044] like Figure 1 As shown, it includes:

[0045] S1. Based on the grid connection faults of grid-connected hybrid energy storage power stations, establish a quantitative evaluation index system.

[0046] In one example, the process of establishing the evaluation indicator system includes:

[0047] Multi-dimensional indicators are extracted to identify the characteristics of various grid-connected faults, resulting in multiple feature indicators. Then, a correlation analysis is performed on the fault type and characteristics of the grid-connected faults, and the feature indicators are constructed into a three-layer evaluation indicator system. In this evaluation indicator system, there is only one primary indicator, which is defined as the severity of the grid-connected fault. The secondary indicators include DC characteristic dimension, energy characteristic dimension, and AC characteristic dimension. The tertiary indicators are feature indicators and are attached to the corresponding secondary indicators.

[0048] In this case, the types of grid connection faults include at least the following 19 types, denoted as F1-F19: lithium battery short circuit fault, lithium battery open circuit fault, supercapacitor short circuit fault, supercapacitor open circuit fault, flywheel energy storage short circuit fault, hydrogen fuel cell short circuit fault, hydrogen fuel cell open circuit fault, hydrogen electrolyzer short circuit fault, hydrogen electrolyzer open circuit fault, DC bus short circuit fault, DC bus open circuit fault, system grid connection to islanding, three-phase grid frequency drop, three-phase grid voltage drop, three-phase grid voltage imbalance, single-phase ground fault, two-phase ground fault, three-phase short circuit fault, and phase-to-phase short circuit fault. Specifically, this plan proposes to construct a hierarchical evaluation index system that includes one primary index (quantitative assessment of the severity of grid connection faults of grid-connected hybrid energy storage power stations A), three secondary indexes (DC characteristic dimension B1, energy characteristic dimension B2, AC characteristic dimension B3), and ten tertiary quantitative indexes (DC voltage fluctuation entropy C1, DC voltage drop rate C2, bus voltage deviation rate C3, maximum change rate of DC current C4, energy loss ratio C5, power deviation rate C6, frequency deviation rate C7, three-phase voltage distortion change rate C8, negative sequence voltage ratio C9, and three-phase voltage deviation rate C10).

[0049] S2 uses the Analytic Hierarchy Process (AHP) to calculate the initial weights of all characteristic indicators in the evaluation index system.

[0050] In one example, the initial weights of all feature indicators in the evaluation index system calculated using the analytic hierarchy process (AHP) include:

[0051] S201, Construct a hierarchical structure: Define the first-level indicators, second-level indicators, and third-level indicators as target-level indicators, criterion-level indicators, and alternative-level indicators, respectively. That is, select the first-level indicator (A) as the target level, the second-level indicators (B1-B3) as the criterion level, and the third-level indicators (C1-C10) as the alternative level.

[0052] S202, Constructing the Judgment Matrix: Based on the Saaty 1-9 scaling method, compare any two criterion-level indicators and construct a 3×3 judgment matrix to determine the relative importance of each criterion-level indicator. The specific formula is as follows:

[0053] A ahp =|a ij |n ahp ×n ahp

[0054] In the formula, A ahp For the judgment matrix; a ij This represents the importance of the i-th indicator relative to the j-th indicator; it is a quantitative evaluation value for pairwise comparisons and is determined by expert analysis and discussion. ahp To construct the number of indicators in the judgment matrix.

[0055] S203, Calculate the weight vector: Based on the relative importance calculated above, the eigenvector method is used to extract the criterion layer weights and weight vectors of each criterion layer indicator from the judgment matrix. Specifically, the geometric mean of each row in the judgment matrix is ​​first calculated, and then the weight vector is obtained through normalization. The specific formula is as follows:

[0056]

[0057] In the formula, To determine the geometric mean of each row in the matrix; The weights are the normalized values ​​of the criterion layer weights; w ahp This is used to determine the weight vector corresponding to the matrix; T is the transpose symbol.

[0058] S204, Consistency Check: The consistency of the judgment matrix is ​​checked based on the criterion layer weights and weight vectors. The criterion layer weights are then accepted based on the check results. If not, the elements in the judgment matrix are adjusted, and steps S203 and S204 are repeated until the consistency test passes. Specifically, when constructing the judgment matrix, pairwise comparisons based on expert experience require a consistency check (CR < 0.1) to ensure the rationality of the weight allocation. First, the largest eigenvalue of the judgment matrix is ​​calculated, then the consistency index (CI) is calculated, and finally the consistency ratio (CR) is calculated. The specific formula is as follows:

[0059]

[0060] In the formula, λ ahp,max To determine the largest eigenvalue of the matrix; CI is the consistency index; CR is the consistency ratio; RI is the value based on the index dimension n. ahp The random consistency index is obtained by looking up the table. The calculated weights are then accepted based on the CR value. If CR < 0.1, the decision is made; otherwise, the decision matrix A is adjusted. ahp The elements in the form are processed, and steps S203 and S204 are recalculated until the consistency test is passed.

[0061] S205, Recursively calculate the scheme layer weights: For each scheme layer indicator attached to a criterion layer indicator, repeat steps S203 and S204 to calculate the local weights of the scheme layer indicators. Multiply the local weights by the corresponding criterion layer weights to obtain the global weights of the tertiary indicators. Continue in this manner until the initial weight matrix w of all tertiary indicators is obtained. ahp .

[0062] S3 uses a convolutional neural network (CNN) to extract spatial features of grid-connected fault data.

[0063] In one example, the grid connection fault data described in this scheme includes: the voltage and current output by the DC side lithium-ion battery, the voltage and current output by the supercapacitor, the voltage and current output by the flywheel energy storage, the voltage and current output by the hydrogen fuel cell, the voltage and current output by the hydrogen electrolyzer, the DC bus voltage and current, the AC side voltage and current, and the AC side active and reactive power when a grid connection fault occurs in the grid-connected hybrid energy storage power station.

[0064] The specific steps for extracting spatial features from grid-connected fault data using a convolutional neural network include:

[0065] S301, Grid Connection Fault Data Preprocessing: Grid connection fault data undergoes filtering preprocessing. Specifically, Adaptive Wavelet Thresholding Denoising (AWTD) is used to process noise in the grid connection fault operation data of the grid-connected hybrid energy storage power station. The denoised coefficients are then used to perform wavelet reconstruction on the operation data samples.

[0066]

[0067] In the formula, For the reconstructed monitoring and operation data; cA J [k] represents the lowest level of approximation coefficients; φ represents the detail coefficients of each layer after denoising. j,k,t and ψ j,k,t Here, represents the scaling function and wavelet function of the Daubechies 4 wavelet basis function, respectively; j and J represent the current decomposition level and the total number of decomposition levels, respectively; and k is the translation parameter. Then, the monitoring operation data reconstructed by AWTD is subjected to Min-Max normalization to generate data in the following format: A multivariate time series matrix, where T total N represents the sampling time of the data samples. cla As a variable dimension, C cla For the number of channels, this solution selects a single channel, and the specific formula is as follows:

[0068]

[0069] In the formula, The data is the normalized monitoring and operation data; The minimum and maximum values ​​in the monitoring data after AWTD reconstruction are respectively used; finally, the normalized multivariate time series is reconstructed into a pseudo-image of a 2D matrix by directly stacking variables, which is then processed by CNN.

[0070] S302, Indicator-Specific Feature Extraction: Obtain the variable data column related to each feature indicator, define the mapping relationship between the feature indicator and the variable data column, and extract the CNN features specific to the feature indicator;

[0071] S303, CNN Network Architecture Design: Construct a convolutional neural network that omits pooling and fully connected layers, and then extract the spatial features of the CNN features within a local time window. Specifically, in the 1D convolutional layer, the spatial features of the CNN features within a local time window are extracted using the following formula:

[0072]

[0073] In the formula: For time t, the kth... cnn The output activation values ​​of each convolutional kernel; For time t+1, the nth time cnn The cth variable cnn The input values ​​for each channel; f cnn The kernel size; For the kth cnn The weights of each convolutional kernel at position i; σ() is the activation function, and this scheme uses the Rectified Linear Unit (ReLU) function; For the kth cnn The bias term of each convolutional kernel. It should be noted that the CNN architecture used in this scheme omits pooling layers and fully connected layers in order to preserve the millisecond-level spatiotemporal features of the fault waveform.

[0074] S304, Output and Downstream Connection: The Flatten layer of the convolutional neural network is called to flatten the spatial features, and the resulting feature map is converted into a feature vector; then the time steps after convolution of the spatial features are reassembled, and finally the feature shape data is input into the downstream LSTM, as shown in the following formula:

[0075]

[0076] In the formula, The feature data is the final output of the CNN network; K cnn The total number of convolution kernels; T′ = T total -f cnn +1.

[0077] S4 uses a long short-term memory network to model the temporal sequence of spatial features and outputs a sequence of hidden states.

[0078] In this scheme, the aforementioned use of Long Short-Term Memory (LSTM) networks to model the temporal sequence of spatial features specifically includes:

[0079] A Long Short-Term Memory (LSTM) network is constructed, comprising an input gate, a forget gate, an output gate, and memory cells. Spatial features are input into the LTM network, and the temporal sequence of these features is modeled to output a sequence of hidden states. The main steps include:

[0080] (1) Inputting the CNN output data into the LSTM unit: Inputting the CNN output data into the LSTM unit. Input LSTM cells.

[0081] (2) LSTM gating mechanism processing: LSTM uses three gating units and one memory cell to realize the time series modeling of grid-connected fault data. The specific formula is as follows:

[0082]

[0083]

[0084] In the formula, F t lstm , and These are the output gates for forgetting, input, candidate memory, memory update state, output, and hidden state, respectively. This refers to grid-connected fault data at time t; This represents the hidden layer state at time t-1, and also the cell output of the LSTM at the previous time step. Candidate memories at time t-1; These are the weight matrices obtained from model training for three gating units and one memory cell, respectively. , respectively, are the bias vectors obtained from model training for three gating units and one memory cell; is the element-wise multiplication symbol; σ() is the activation function, which is the sigmoid function used in this scheme; tanh() is the tanh activation function.

[0085] (3) Output and downstream connection: The hidden state sequence output by the LSTM is used as the input to the Attention layer, as shown in the following formula:

[0086]

[0087] In the formula: This refers to the final output data of the LSTM layer; D lstm is the hidden layer dimension of the LSTM layer.

[0088] S5 uses an attention mechanism to dynamically assign weights to the hidden state sequence to obtain the final weights.

[0089] In one example, dynamically weighting the hidden state sequence using an attention mechanism includes:

[0090] S501: Input the output data of the LSTM into the Attention layer in time steps: hidden state sequence As query vector q att .

[0091] S502, Calculate the attention score: Calculate the attention score between the hidden state and the corresponding query vector at each time step, using the following formula:

[0092]

[0093] In the formula, For relevance score; The weight matrix is ​​a learnable weight matrix; q is a learnable weight vector; att For query vector; b att This is used to compute the learnable bias vector in the attention score.

[0094] S503, Normalized Attention Weights: Attention scores are converted into attention weights using the Softmax function, with the specific formula as follows:

[0095]

[0096] In the formula, The normalized attention weight values ​​reflect the importance of each time step.

[0097] S504, Generating the Context Vector: The attention weights are weighted and summed with all hidden states to generate the context vector. The specific formula is as follows:

[0098]

[0099] In the formula, This serves as a context vector, summarizing information from key time periods.

[0100] S505, Generating Dynamic Correction Factors: Context vectors are mapped to dynamic weight correction factors, as shown in the following formula:

[0101]

[0102] In the formula, Δw i W is a dynamic weight correction factor used to adjust the static weights of AHP; c att As a learnable weight matrix, it determines how the input features affect the correction of the weights of each indicator; To calculate the learnable bias vector in the dynamic correction factor, a default adjustment tendency for the index weights is provided.

[0103] S506, Calculate dynamic weights: The final weights for the evaluation index system are calculated based on the dynamic weight correction factor, using the following formula:

[0104]

[0105] In the formula, The dynamic weight of the i-th indicator; is the normalized dynamic weight of the i-th indicator; M is the total number of indicators.

[0106] S6, based on the final weight and evaluation index system, performs dynamic weight quantification scoring of grid connection faults of grid-connected hybrid energy storage power stations.

[0107] Specifically, based on the dynamic weight allocation method of AHP+CNN-LSTM-Attention in steps S1-S5 above, a quantitative score of the severity of grid connection faults in a grid-connected hybrid energy storage power station can be obtained. The specific formula is as follows:

[0108]

[0109] In the formula, The score for the f-th grid-connected fault of a grid-connected hybrid energy storage power station; This represents the i-th normalized index value. The framework diagram for the comprehensive assessment strategy of grid connection fault severity for grid-connected hybrid energy storage power stations is shown below. Figure 2 As shown.

[0110] In this embodiment, the weight values ​​of the grid-connected hybrid energy storage power station under 19 typical grid-connected fault conditions obtained according to steps S1-S5 are as follows: Figure 3 As shown in the figure, the data indicates that, overall, the energy loss ratio (C5) and the DC voltage sag rate (C2) both play a prominent and crucial role in various types of faults. Specifically, the dynamic weight of C5 generally reaches 34.52%-36.03%, an increase of 37.0%-42.9% compared to its fixed weight of 25.21% in AHP. This significant change confirms the core position of energy loss in the comprehensive assessment strategy for grid-connected faults of hybrid energy storage power stations proposed in this scheme. Meanwhile, the dynamic weight of C2 remains in a high range of 13.10%-13.84%, and its correction factor is stable in the range of 0.45-0.49, reflecting the continued importance of voltage dynamic characteristics in judging the severity of faults.

[0111] Furthermore, by comparing and analyzing the static and dynamic weight correction results of AHP, the mean value of the dynamic weight correction factor is 0.47±0.02, which effectively alleviates the subjective limitations of the traditional AHP method and significantly improves the objectivity and adaptability of the evaluation system. Taking the three-phase voltage distortion (C8) and the three-phase voltage deviation rate (C10) as examples, for typical faults such as three-phase voltage drop and frequency drop under grid-side fault conditions, under the F15 voltage imbalance fault condition, the dynamic weight correction factors of the two key indicators are stably maintained in the range of 0.47 to 0.49, which is more prominent than the fixed weight of AHP in terms of the special sensitivity of the proposed algorithm to grid parameter fluctuations. Secondly, the weight of the bus voltage deviation rate (C3) is reduced from the fixed 3.27% of AHP to a reasonable range of 1.61% to 1.85% after dynamic correction. This adjustment is more consistent with the actual impact of this indicator in the measured fault data, verifying the scientific optimization effect of the dynamic correction mechanism on the weight distribution.

[0112] In terms of fault characteristic differentiation, dynamic weights show significant advantages. For example, in the DC bus short-circuit and open-circuit fault scenarios, the dynamic weights of the C2 voltage drop rate index are 13.31% and 13.45% respectively, with a difference of 1.3%. Compared with the zero difference of the fixed weights of AHP, this differentiated adjustment more accurately reflects the essential difference between the two fault types in terms of the impact on the system, and effectively enhances the ability of the evaluation model to distinguish the severity of the fault.

[0113] In this embodiment, the quantitative total score comparison chart of the grid-type hybrid energy storage power station under 19 typical grid-connected fault conditions, obtained according to steps S1-S6, is shown below. Figure 4 As shown in the figure, the differences in assessment between the two methods under different fault types can be clearly observed through a quantitative scoring bar chart visualization comparison.

[0114] From the perspective of power supply-side faults, taking the F5 flywheel energy storage short-circuit fault as an example, the total score of the AHP+CNN-LSTM-Attention integrated evaluation strategy was 0.53684, a decrease of 10.15% compared to the 0.59746 of the AHP method. Similarly, the total score of the F10 DC bus short-circuit fault decreased from 0.40945 of AHP to 0.37833 of the integrated evaluation strategy, a decrease of 7.60%. These data indicate that the AHP+CNN-LSTM-Attention integrated evaluation strategy is more conservative in evaluating severe short-circuit faults, and it captures the dynamic characteristics of the fault transient process more accurately, avoiding the overestimation phenomenon caused by the static weight allocation of the traditional AHP method. In other power supply-side faults, the integrated evaluation strategy generally showed an upward trend in scores. For example, the short-circuit fault score of the F3SCP increased from 0.15215 to 0.16084, an increase of 5.71%; the open-circuit fault score of the F7 hydrogen fuel cell increased from 0.09453 to 0.12152, an increase of 28.55%. This difference highlights the improved sensitivity of the proposed AHP+CNN-LSTM-Attention integrated evaluation strategy to specific faults of coupled energy storage units, especially for devices with significant dynamic response characteristics such as supercapacitors and hydrogen fuel cells.

[0115] In grid-side faults, the advantages of the AHP+CNN-LSTM-Attention integrated evaluation strategy are further highlighted. The fault severity score for the F12 system transitioning from grid connection to islanding improved by 5.84%, while the frequency sag score for the F13 three-phase grid increased by 20.90%. This significant improvement of over 20% verifies the evaluation advantage of the integrated strategy for frequency-related faults, as its use of LSTM network-captured frequency time-series features compensates for the shortcomings of traditional AHP in quantifying dynamic processes. Notably, the F14 three-phase grid voltage sag and F15 voltage imbalance sag faults showed reductions of 14.43% and 11.72%, respectively, indicating that the fusion algorithm can more accurately distinguish the essential differences between voltage amplitude changes and energy-related faults.

[0116] In summary, the comprehensive evaluation strategy for grid-connected faults in grid-connected hybrid energy storage power stations proposed in this invention, namely the AHP+CNN-LSTM-Attention comprehensive evaluation strategy, achieves more accurate quantification of fault severity by combining the structured weight allocation of the analytic hierarchy process (AHP) with the temporal feature extraction capabilities of deep learning. For faults with significant dynamic characteristics, such as the specific faults of coupled energy storage units F1-F4 and F6-F9, the system transition from grid connection to islanding (F12), and the three-phase grid frequency drop (F13), the average score is improved by 15.32%. For faults with predominantly steady-state characteristics, such as those involving voltage amplitude changes (F14-F17), the average score is reduced by 11.25%. These results statistically verify that the AHP+CNN-LSTM-Attention comprehensive evaluation strategy can adaptively adjust the evaluation weights according to the physical nature of the fault, achieving a more scientific quantification of severity.

[0117] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A comprehensive assessment strategy for grid-connected faults in a grid-connected hybrid energy storage power station, characterized in that, include: S1. Based on the grid connection faults of grid-connected hybrid energy storage power stations, establish a quantitative evaluation index system; S2, The initial weights of all characteristic indicators in the evaluation index system are calculated using the analytic hierarchy process (AHP). S3 uses a convolutional neural network to extract spatial features of grid-connected fault data; S4, A long short-term memory network is used to model the temporal sequence of the spatial features and output the hidden state sequence; S5, The hidden state sequence is dynamically weighted using an attention mechanism to obtain the final weights; the dynamic weighting of the hidden state sequence using an attention mechanism includes: S501, the hidden state sequence is used as a query vector; S502, the attention score of the hidden state and the corresponding query vector at each time step is calculated; S503, the attention score is converted into attention weights using a Softmax function; S504, the attention weights are weighted and summed with all the hidden states to generate a context vector; S505, the context vector is mapped to a dynamic weight correction factor; the dynamic weight correction factor is used to adjust the initial weights; S506, the final weights for the evaluation index system are calculated based on the initial weights adjusted by the dynamic weight correction factor. S6. Based on the final weight and the evaluation index system, perform dynamic weight quantification scoring on grid connection faults of grid-connected hybrid energy storage power stations.

2. The comprehensive assessment strategy for grid-connected faults of a grid-connected hybrid energy storage power station as described in claim 1, characterized in that, In S1, the process of establishing the evaluation index system includes: Multi-dimensional indicators are extracted to identify the characteristics of each grid-connected fault, resulting in multiple feature indicators. Then, a correlation analysis is performed on the fault type and characteristics of the grid-connected fault, and the feature indicators are constructed into a three-layer evaluation indicator system. In this evaluation indicator system, there is only one primary indicator, which is defined as the severity of the grid-connected fault. The secondary indicators include DC characteristic dimension, energy characteristic dimension, and AC characteristic dimension. The tertiary indicators are the feature indicators and are attached to the corresponding secondary indicators.

3. The comprehensive assessment strategy for grid-connected faults of a grid-connected hybrid energy storage power station as described in claim 2, characterized in that, The types of grid connection faults include: lithium battery short circuit fault, lithium battery open circuit fault, supercapacitor short circuit fault, supercapacitor open circuit fault, flywheel energy storage short circuit fault, hydrogen fuel cell short circuit fault, hydrogen fuel cell open circuit fault, hydrogen electrolyzer short circuit fault, hydrogen electrolyzer open circuit fault, DC bus short circuit fault, DC bus open circuit fault, system grid connection to islanding, three-phase grid frequency drop, three-phase grid voltage drop, three-phase grid voltage imbalance, single-phase ground fault, two-phase ground fault, three-phase short circuit fault, and phase-to-phase short circuit fault.

4. The comprehensive assessment strategy for grid-connected faults of a grid-connected hybrid energy storage power station as described in claim 2, characterized in that, In S2, the calculation of the initial weights of all characteristic indicators in the evaluation index system using the analytic hierarchy process includes: S201, the first-level indicator, the second-level indicator and the third-level indicator are respectively defined as target-level indicator, criterion-level indicator and scheme-level indicator; S202, Based on the Saaty 1-9 scaling method, compare any two of the criteria layer indicators and construct a 3×3 judgment matrix to determine the relative importance of each of the criteria layer indicators; S203, based on the relative importance, the eigenvector method is used to extract the criterion layer weights and weight vectors of each criterion layer indicator from the judgment matrix; S204, check the consistency of the judgment matrix according to the criterion layer weights and the weight vector, and determine whether to accept the criterion layer weights according to the test results. If not, adjust the elements in the judgment matrix and repeat steps S203 and S204 until the consistency test is passed. S205, for each scheme layer indicator attached to the criterion layer indicator, repeat steps S203 and S204 to calculate the local weight of the scheme layer indicator. The local weight is multiplied by the corresponding criterion layer weight to obtain the global weight of the third-level indicator, until the initial weight matrix of all the third-level indicators is obtained.

5. The comprehensive assessment strategy for grid-connected faults of a grid-connected hybrid energy storage power station as described in claim 1, characterized in that, In S3, the grid connection fault data includes: the voltage and current output of the DC side lithium-ion battery, the voltage and current output of the supercapacitor, the voltage and current output of the flywheel energy storage, the voltage and current output of the hydrogen fuel cell, the voltage and current output of the hydrogen electrolyzer, the DC bus voltage and current, the AC side voltage and current, and the AC side active and reactive power when a grid connection fault occurs in the grid-connected hybrid energy storage power station.

6. The comprehensive assessment strategy for grid-connected faults of a grid-connected hybrid energy storage power station as described in claim 1, characterized in that, In S3, the spatial features extracted from grid-connected fault data using a convolutional neural network include: S301, Perform filtering preprocessing on the grid connection fault data; S302, obtain the variable data column related to each feature indicator, define the mapping relationship between the feature indicator and the variable data column, and extract the CNN features of the feature indicator; S303, construct a convolutional neural network that omits pooling layers and fully connected layers, and then extract the spatial features of the CNN features within the local time window; S304, the Flatten layer of the convolutional neural network is invoked to flatten the spatial features, and the feature map obtained after flattening the spatial features is converted into a feature vector; then the time steps after convolution of the spatial features are reorganized.

7. The comprehensive assessment strategy for grid-connected faults of a grid-connected hybrid energy storage power station as described in claim 1, characterized in that, In S4, the temporal modeling of the spatial features using a long short-term memory network includes: A long short-term memory network containing an input gate, a forget gate, an output gate, and memory cells is constructed. The spatial features are input into the long short-term memory network, and the temporal sequence of the spatial features is modeled to output the hidden state sequence.

Images