A method and system for monitoring energy of an energy storage cabinet based on multi-source data

Abstract

This invention relates to the field of intelligent control technology and discloses a method and system for energy monitoring of energy storage cabinets based on multi-source data. The method includes acquiring multi-source heterogeneous data, standardizing it to obtain an initial dataset, performing feature extraction and pattern recognition to obtain a vector dimension mapping; performing feature fusion and feature dimensionality reduction to obtain an aligned feature set; integrating multi-source heterogeneous data to judge abnormal fluctuations and obtain a dynamic state representation sequence; calculating the load fluctuation probability in the sequence, and if it is higher than a threshold, performing load prediction to obtain a predicted load value; acquiring real-time energy monitoring data, optimizing equipment parameter configuration, and generating an energy allocation scheme; extracting equipment operating parameters from the scheme, and generating an allocation scheme if it meets a preset load threshold; extracting similarity information, adjusting grid parameters, and generating a demand prediction set; updating the operating parameters of the energy storage cabinet based on the demand prediction set, and generating a monitoring output report. This method can improve the accuracy of energy monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to an energy monitoring method and system for energy storage cabinets based on multi-source data. Background Technology

[0002] Currently, energy storage cabinets, as core equipment in new energy systems, are crucial for ensuring the stable operation of the power system and optimizing energy utilization through energy monitoring. With the widespread application of renewable energy, energy storage cabinets need to monitor and manage various energy data in real time to cope with complex grid demands. However, current monitoring methods often struggle to achieve efficient integration and accurate prediction when processing multi-source data, resulting in low energy management efficiency and failing to meet the real-time needs of a dynamic grid environment. This makes developing a monitoring method capable of comprehensively integrating multi-source data and achieving dynamic prediction a critical issue that urgently needs to be addressed in the energy storage field.

[0003] In existing technologies, energy monitoring methods for energy storage cabinets based on multi-source data typically employ a single processing flow to handle diverse and formatted data such as battery status, grid load, and environmental parameters. For example, existing technologies use simple data aggregation and normalization methods to convert data at different time granularities into a unified format, and then utilize basic time-series analysis models for feature extraction and status prediction. Specifically, existing methods may first perform time interpolation and smoothing on battery voltage and load power sequences to align the data on the time axis, then extract key features using traditional statistical methods, and finally perform load prediction and energy allocation adjustments based on these features. However, due to the inherent differences in time dimension and physical characteristics of multi-source data, such as the rapid fluctuations in battery voltage and the slow changes in grid load, existing technologies often fail to effectively unify the feature representations of these heterogeneous data, resulting in a lack of consistency and comparability in the extracted features. This lack of feature alignment makes it difficult for the system to form a comprehensive and accurate status characterization when analyzing the operating status of the energy storage cabinet, thereby introducing noise and reducing the reliability of the prediction algorithm, ultimately affecting the accuracy and real-time performance of energy monitoring.

[0004] In summary, existing technologies struggle to achieve efficient integration and accurate feature alignment of multi-source heterogeneous data, resulting in low accuracy in energy monitoring. Summary of the Invention

[0005] This invention provides a method, device, electronic equipment, and storage medium for energy monitoring of energy storage cabinets based on multi-source data, in order to solve the problem of low accuracy in energy monitoring.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an energy monitoring method for energy storage cabinets based on multi-source data, comprising: Acquire multi-source heterogeneous data and standardize it to obtain an initial dataset; perform feature extraction and pattern recognition on the initial dataset to obtain a vector dimension mapping; If the similarity between vectors in the vector dimension mapping exceeds a preset similarity threshold, feature fusion and feature dimensionality reduction are performed to obtain an aligned feature set; The temporal dependencies of the aligned feature sets are analyzed, and multi-source heterogeneous data are integrated to determine whether abnormal fluctuations occur, thus obtaining a dynamic state representation sequence. Calculate the load fluctuation probability in the dynamic state characterization sequence. If it is higher than the preset probability threshold, optimize the operating parameters and perform load prediction to obtain the predicted load value. Real-time energy monitoring data is acquired and fused with the predicted load value. Through load balancing verification, the parameter configuration of the equipment is optimized, and an optimized energy allocation scheme is generated. Extract equipment operating parameters from the optimized energy allocation scheme. If the load balance index in the equipment operating parameters meets the preset load threshold, generate a verified allocation scheme. Extract and integrate feature set similarity information from the verified allocation scheme. If the integrated information result matches the power grid demand, adjust the power grid demand parameters and generate a demand forecast set. Based on the aforementioned demand forecast set, update the operating parameters of the energy storage cabinet, and collect and generate monitoring output reports through real-time feedback data.

[0007] Secondly, the present invention provides an energy monitoring device for an energy storage cabinet based on multi-source data, comprising: The data acquisition module is used to acquire multi-source heterogeneous data, standardize it to obtain an initial dataset, and perform feature extraction and pattern recognition on the initial dataset to obtain a vector dimension mapping. The feature analysis module is used to perform feature fusion and feature dimensionality reduction to obtain an aligned feature set if the similarity between vectors in the vector dimension mapping exceeds a preset similarity threshold. The temporal integration module is used to analyze the temporal dependencies of the aligned feature sets and integrate multi-source heterogeneous data to determine whether abnormal fluctuations occur, thereby obtaining a dynamic state representation sequence. The parameter optimization module is used to calculate the load fluctuation probability in the dynamic state characterization sequence. If it is higher than the preset probability threshold, the operating parameters are optimized and the load is predicted to obtain the predicted load value. The load balancing module is used to acquire real-time energy monitoring data and integrate it with the predicted load value. Through load balancing verification, the parameter configuration of the equipment is optimized to generate an optimized energy distribution scheme. The scheme generation module is used to extract equipment operating parameters from the optimized energy distribution scheme. If the load balance index in the equipment operating parameters meets the preset load threshold, a verified distribution scheme is generated. The demand matching module is used to extract and integrate feature set similarity information from the verified allocation scheme. If the integrated information result matches the power grid demand, the power grid demand parameters are adjusted and a demand prediction set is generated. The results output module is used to update the operating parameters of the energy storage cabinet based on the demand forecast set, and to collect and generate a monitoring output report through real-time feedback data.

[0008] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the energy monitoring method for energy storage cabinets based on multi-source data as described above.

[0009] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the energy monitoring method for energy storage cabinets based on multi-source data described above.

[0010] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention effectively solves the problem of inconsistency in time and dimension of multi-source data by acquiring multi-source heterogeneous data from battery management systems and power grid dispatching systems and converting it into a unified time-scale format through standardization. By performing temporal convolution and feature extraction on the initial dataset, it accurately captures the periodic fluctuation characteristics of battery voltage sequences and the multi-scale patterns of load change sequences, generating vector dimension mappings with clear physical meaning. This process realizes data preprocessing and feature optimization, reduces data noise and redundant information through standardization and feature extraction, improves the accuracy and comparability of feature representation, provides a reliable data foundation for subsequent feature alignment, and significantly improves the system's processing accuracy and adaptability for multi-source heterogeneous data.

[0011] (2) This invention calculates the similarity between vector dimension mappings. When the similarity exceeds a preset threshold, feature fusion and dimensionality reduction are performed. Attention weights are used to adjust the weights of features at different time scales, thereby achieving feature alignment between battery state sequences and load fluctuation sequences. This process effectively integrates multi-source features, eliminates differences and conflicts between features through dynamic weight allocation and dimensionality reduction, retains key dynamic features, and improves the consistency and representativeness of the feature set. This enhances the accuracy and reliability of subsequent state analysis and strengthens the system's ability to fuse heterogeneous data features.

[0012] (3) This invention analyzes the temporal dependencies of aligned feature sets, uses a long short-term memory network to capture the long-term dependency features of battery state sequences, and combines an attention mechanism to weight the different time-scale features of load fluctuation sequences to achieve abnormal fluctuation detection of multi-source data. This process realizes temporal modeling and anomaly identification, comprehensively captures the system operation trend through long-term dependency analysis and dynamic weight adjustment, accurately identifies abnormal load fluctuations, generates precise dynamic state representation sequences, improves the real-time performance and accuracy of state monitoring, and provides a reliable basis for load prediction.

[0013] (4) This invention calculates the load fluctuation probability in a dynamic state characterization sequence, optimizes operating parameters based on probability thresholds, and uses a weighted average method to calculate the contribution of features at each time scale, thereby achieving accurate prediction of load changes. This process realizes probability assessment and parameter optimization, dynamically assesses the risk of load changes through fluctuation probability calculation, and fits the mapping relationship of operating parameters with historical data to ensure a high degree of matching between predicted load values ​​and actual demand, thereby improving the accuracy and foresight of load prediction and enhancing the system's responsiveness to energy dispatch.

[0014] (5) This invention verifies load balancing by integrating real-time energy monitoring data with predicted load values. It uses a data fusion method to integrate multi-source data streams, weights the features of each data source, and optimizes the configuration of equipment parameters. This process achieves data fusion and load balancing, ensures information consistency by integrating real-time data with predicted values, optimizes energy allocation schemes by combining historical trends and periodic characteristics, improves system stability and allocation accuracy, and enhances the accuracy and reliability of energy dispatch.

[0015] (6) This invention extracts equipment operating parameters from the optimized energy distribution scheme, verifies whether the load balance index meets the threshold, and uses smoothing processing and feature extraction to analyze voltage change characteristics to verify the applicability of the distribution scheme. This process realizes scheme verification and parameter calibration. The feasibility of the load balance index is ensured by threshold judgment, and the operating parameters are calibrated by combining voltage fluctuation characteristics to generate a reliable verification scheme, improve the accuracy and stability of scheme implementation, and ensure that the system operates in the optimal state.

[0016] (7) This invention extracts and integrates feature set similarity information, uses a clustering algorithm to classify the similarity information, and generates a demand prediction set after weighted fusion, thereby achieving matching and adjustment of power grid demand parameters. This process realizes demand prediction and parameter optimization, ensures the consistency between demand prediction and actual power grid demand through similarity analysis and weighted fusion, and improves prediction accuracy and power grid synchronization by dynamically calibrating and adjusting power grid parameters, thereby enhancing the foresight and coordination of system scheduling.

[0017] (8) This invention updates the operating parameters of the energy storage cabinet based on the demand forecast set, collects real-time feedback data to generate a monitoring output report, and integrates real-time load data through feature data classification and processing to achieve closed-loop optimization of system operation. This process realizes real-time monitoring and continuous optimization, adapts to changes in demand through parameter updates, and ensures continuous system adjustment by combining real-time data feedback, thereby improving operational accuracy and monitoring comprehensiveness, and enhancing the reliability and efficiency of system response. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a method for monitoring the energy of an energy storage cabinet based on multi-source data, provided in the first embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an energy monitoring device for an energy storage cabinet based on multi-source data, provided in the second embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1 The first embodiment of the present invention provides an energy monitoring method for energy storage cabinets based on multi-source data, comprising the following steps: S11, acquire multi-source heterogeneous data and standardize it to obtain an initial dataset; perform feature extraction and pattern recognition on the initial dataset to obtain a vector dimension mapping; S12, analyze the temporal dependency of the aligned feature set, and integrate multi-source heterogeneous data to determine whether abnormal fluctuations occur, and obtain the dynamic state representation sequence. S13, analyze the temporal dependency of the alignment feature set, and integrate multi-source heterogeneous data to determine whether abnormal fluctuations occur, and obtain the dynamic state representation sequence. S14, calculate the load fluctuation probability in the dynamic state characterization sequence. If it is higher than the preset probability threshold, optimize the operating parameters and perform load prediction to obtain the predicted load value. S15: Obtain real-time energy monitoring data and integrate it with the predicted load value. Through load balancing verification, optimize the parameter configuration of the equipment and generate an optimized energy allocation scheme. S16, extract equipment operating parameters from the optimized energy distribution scheme; if the load balance index in the equipment operating parameters meets the preset load threshold, generate a verified distribution scheme. S17. Extract and integrate feature set similarity information from the verified allocation scheme. If the integrated information result matches the power grid demand, adjust the power grid demand parameters and generate a demand prediction set. S18. Based on the demand forecast set, update the operating parameters of the energy storage cabinet, collect and generate a monitoring output report through real-time feedback data.

[0021] In step S11, multi-source heterogeneous data is acquired and standardized to obtain an initial dataset; feature extraction and pattern recognition are performed on the initial dataset to obtain a vector dimension mapping, including: Acquire multi-source heterogeneous data, standardize the data to unify the time scale format, and obtain the initial dataset; The multi-source heterogeneous data includes sequences of minute fluctuations in battery voltage and sequences of minute changes in load within the microgrid. Based on the initial dataset, the battery voltage sequence and load change sequence are extracted, and the fluctuation features of the battery voltage sequence are extracted to obtain the first feature vector set; The load change sequence is decomposed into a multi-scale decomposition to obtain a second set of feature vectors; If the time scales of the first feature vector set and the second feature vector set are consistent, the first feature vector set and the second feature vector set are integrated to obtain a multi-source heterogeneous sequence; The multi-source heterogeneous sequences are subjected to dimensionality reduction processing to extract and identify key patterns, thereby obtaining a vector dimension mapping.

[0022] It should be noted that the step of acquiring multi-source heterogeneous data and standardizing the data to unify the time scale format to obtain the initial dataset specifically targets high-frequency voltage data (e.g., second-level sampling) from the battery management system and low-frequency load data (e.g., minute-level sampling) from the power grid dispatching system. Time interpolation is achieved through linear aggregation. For the second-level voltage data from the battery management system, the time axis is first divided into minute intervals. Then, the arithmetic mean of all second-level sampling points within each minute is calculated to generate a minute-level aggregated value, thereby downsampling the high-frequency data to a uniform minute granularity. For the minute-level load data from the power grid dispatching system, since the time scale is already consistent, the original values ​​are directly retained without processing. Simultaneously, a minimum-maximum scaling algorithm is used to normalize data with different dimensions, eliminating dimensional differences and generating a standardized initial dataset. This dataset contains time-aligned and normalized sequences, facilitating subsequent processing by machine learning models.

[0023] Based on the initial dataset, the battery voltage sequence and load change sequence are extracted, and the fluctuation features of the battery voltage sequence are extracted to obtain the first feature vector set. This step uses a deep architecture of Temporal Convolutional Neural Network (TCN) to process the battery voltage sequence. Its core adopts a causal convolutional structure with an dilation mechanism: First, the input voltage sequence is mapped to a 128-dimensional high-dimensional feature space through an embedding layer; then, 8 layers of residual blocks are stacked (each residual block contains two dilated causal convolutional layers), the kernel size is fixed at 3, and the dilation coefficient d increases exponentially with the number of layers (the first layer is the first layer). layer , The network operates from 0 to 7, maintaining the output time series length using symmetric zero-padding. Each convolutional layer is followed by a ReLU activation function to extract nonlinear features, and layer normalization (LayerNorm) is used to stabilize the training process. Finally, global average pooling (GAP) compresses the time series features into a 256-dimensional fixed vector. The resulting first feature vector set contains key pattern information such as voltage fluctuation amplitude (represented by the maximum amplitude |max-min|) and frequency characteristics (extracted by the dominant frequency component using fast Fourier transform). The network employs an end-to-end training approach, using the Adam optimizer (learning rate 0.001) and mean squared error loss function. After 200 training rounds, it converges, and the final output feature vector set provides a structured input foundation containing time-frequency domain characteristics for subsequent multi-source heterogeneous data fusion. The step of performing multi-scale decomposition of the load change sequence to obtain the second set of eigenvectors is achieved through wavelet transform. Specifically, continuous wavelet transform (CWT) is performed using the Daubechies wavelet basis (db4 wavelet): First, the load sequence is input into the wavelet transform system, and then processed through the mother wavelet function... with sequence The convolution calculation of wavelet coefficients is given by the following formula: Where 'a' is the scale parameter (corresponding to the frequency sub-band) and 'b' is the translation parameter; the scale selection covers multiple resolutions, from high-frequency details (smaller 'a', such as a=1-10 corresponding to minute-level fluctuations) to low-frequency approximations (larger 'a', such as a=50-100 corresponding to hourly trends), calculating the coefficient matrix by discretizing the scale parameter 'a' (e.g., linear or logarithmic intervals); subsequently, coefficient features are extracted, and statistical processing (e.g., calculating energy values) is performed on the coefficients at each scale. (or variance), generating feature vectors representing different time scales; finally, these feature vectors are spliced ​​and integrated to form a second feature vector set, which includes short-term fluctuation details (such as rapid changes at the minute level) and long-term trend components (such as slow changes at the hour level), providing multi-scale load feature inputs for subsequent multi-source data fusion.

[0024] If the time scales of the first feature vector set and the second feature vector set are consistent, the step of integrating the first feature vector set and the second feature vector set to obtain the multi-source heterogeneous sequence is achieved through vector concatenation. The specific process is as follows: First, the first feature vector set (voltage feature vector...) ) and the second set of feature vectors (load feature vectors) Time alignment verification is performed to ensure that the two sets of vectors have the same time granularity (e.g., one feature point per minute) and vector dimension n; then, horizontal concatenation is used to connect the corresponding dimensions of the two feature vectors in sequence to generate a joint feature vector. During the splicing process, the Z-score normalization method was used to normalize the feature values, and the final multi-source heterogeneous sequence is a two-dimensional time series matrix. Where T represents the time step, each row contains the voltage-load joint features at a specific time point. This sequence fully preserves the time-domain features and cross-domain correlations of the original data, providing structured input for subsequent dimensionality reduction and pattern recognition.

[0025] The step of dimensionality reduction of multi-source heterogeneous sequences, extraction and identification of key patterns, and obtaining vector dimension mapping is implemented using principal component analysis (PCA) algorithm. The specific process is as follows: First, the covariance matrix C of the multi-source heterogeneous sequence matrix is ​​calculated. Then, eigenvalue decomposition is performed on C to solve the characteristic equation. , to obtain eigenvalues and the corresponding feature vector The number of principal components, k, is selected by calculating the variance explained rate, so that the cumulative variance contribution rate is... (Retaining 95% of the original information); finally, the original data is projected onto the feature subspace formed by the first k feature vectors. The projection formula is ,in The vector dimension mapping is a low-dimensional feature set after dimensionality reduction. This mapping preserves the key patterns of the original data and significantly reduces redundant information. The vector dimension mapping is a low-dimensional feature representation. It is obtained by retaining the principal component directions with the largest variance in the original data and projecting the high-dimensional features onto these directions to form a compact representation. This mapping is used for subsequent feature alignment and state representation. It evaluates feature consistency by calculating the cosine similarity between vectors and provides input to the attention mechanism to weightedly fuse multi-source features.

[0026] In step S12, if the similarity between vectors in the vector dimension mapping exceeds a preset similarity threshold, feature fusion and feature dimensionality reduction are performed to obtain an aligned feature set, including: The battery state sequence and load fluctuation sequence are obtained from the multi-source heterogeneous data; Extract the dynamic change features of the battery state sequence over time to obtain the first time-series feature set; The weights of features at different time scales in the load fluctuation sequence are adjusted to obtain a weighted feature set; If the similarity between the weighted feature set and the first temporal feature set exceeds a preset similarity threshold, the weighted feature set and the first temporal feature set are merged to generate a unified feature representation set; The unified feature representation set is subjected to dimensionality reduction processing to obtain the aligned feature set.

[0027] It should be noted that the step of obtaining battery state sequences and load fluctuation sequences from multi-source heterogeneous data first establishes a unified time reference system. Using the minute-level timestamps of the power grid dispatch system as a benchmark, timestamp matching is performed on the high-frequency data of the battery management system. The nearest neighbor interpolation algorithm is then used to aggregate the second-level data of the battery management system to the minute-level granularity. Specifically, using the hourly timestamp of each minute (e.g., 10:00:00) as the benchmark, all second-level sampled data within that minute (e.g., 60 voltage sample values ​​between 10:00:00 and 10:00:59) are selected, and their arithmetic mean is calculated as the representative value for that minute. For cases where timestamps are not perfectly matched, a time window sliding matching method is used. Specifically, a sliding window is established with a tolerance range of ±30 seconds. When the power grid data timestamp is T, all battery data within the window [T-30s, T+30s] are collected. If multiple sampling points exist within the window, a weighted average is taken. Sampling points closer to T have higher weights. The weight calculation uses a Gaussian kernel function, and the specific formula is as follows: ,in These are the timestamps of the sampling points, where T is the target time. Set the bandwidth parameter (to half the window width, i.e., 15 seconds) and normalize the weights. This ensures that sampling points closer to T contribute more; simultaneously, data integrity is verified by calculating the missing rate for each time period. %,in It represents the number of missing time periods. It is the total number of time periods, when When the value is greater than 5%, interpolation is performed to ensure that the two sequences have exactly the same number of valid time points and data length; finally, the time-synchronized battery status sequence (including parameters such as state of charge and voltage) and load fluctuation sequence (including parameters such as active power and reactive power) are extracted. The step of extracting the dynamic change features of the battery state sequence over time to obtain the first temporal feature set is implemented using a temporal convolutional neural network (TCN). Specifically, it adopts a deep residual architecture containing 8 dilated causal convolutional blocks, each block consisting of two dilated causal convolutional layers, using a one-dimensional convolutional kernel of size 3. The dilation coefficient increases exponentially with the number of layers (the first layer being the first layer). layer , From 0 to 7), its convolution operation formula is: Symmetric zero-padding is used to keep the output sequence length consistent with the input to ensure causality. After each convolutional layer, a ReLU activation function is applied to introduce nonlinearity, and layer normalization (LayerNorm) is performed to stabilize the training process. After stacking multiple convolutional layers, global average pooling (GAP) is used to compress the temporal features into a 256-dimensional fixed vector. The final output is the first temporal feature set containing dynamic features of charge state changes (such as fluctuation amplitude and frequency characteristics), which is used for subsequent multi-source data fusion.

[0028] The step of adjusting the weights of features at different time scales in the load fluctuation sequence to obtain a weighted feature set is achieved through an attention mechanism, specifically using a scaled dot product attention calculation method: First, the load fluctuation sequence is input into a linear transformation layer to generate a query matrix Q, a key matrix K, and a value matrix V, where Q, K, and V are calculated using learnable weight matrices W_Q, W_K, and W_V with the input sequence X (Q=XW_Q, K=XW_K, V=XW_V); then, the attention score is calculated using the formula: Where d_k is the dimension of the key vector (usually set to 128), the softmax function ensures weight normalization; the weights are set based on the temporal importance of the features and are automatically assigned by the parameters learned during training. Minute-level fluctuation features (such as rapidly changing parts) have a significant impact on real-time scheduling, so they are assigned higher weights (0.7-0.9), while hourly-level trend features (such as slowly changing parts) have a smaller impact and are assigned lower weights (0.1-0.3). These weight values ​​are determined through backpropagation optimization to maximize task performance; finally, the weighted feature set highlights short-term fluctuation patterns and is used for subsequent fusion processing.

[0029] If the similarity between the weighted feature set and the first time-series feature set exceeds a preset similarity threshold, the step of fusing the weighted feature set and the first time-series feature set to generate a unified feature representation set is achieved through vector concatenation and feature fusion techniques. The specific process is as follows: First, the cosine similarity between the weighted feature set A (load fluctuation features) and the first time-series feature set B (battery state features) is calculated. Then, the two feature sets are horizontally concatenated according to their dimensions to generate a joint vector. Where m and n are the dimensions of the two feature sets, respectively; finally, feature fusion is performed through a two-layer fully connected neural network. The first layer uses 256 neurons and the ReLU activation function, and the second layer uses 128 neurons and linear activation, ultimately generating a unified feature representation set containing joint information on battery state and load fluctuation. This set retains key cross-domain features and eliminates redundant information; the similarity threshold is determined through historical data statistics and experimental verification. The specific setting process is as follows: during the training phase, the grid search method is used to test different thresholds in the range of 0.5-1.0 with a step size of 0.05, and the best balance point is selected by combining the precision-recall curve (usually 0.7-0.9).

[0030] The step of reducing the dimensionality of the unified feature representation set to obtain the aligned feature set is implemented using the Principal Component Analysis (PCA) algorithm. The specific process is as follows: First, the input feature matrix... (m is the number of samples, n is the feature dimension) Centering is performed so that the mean of each feature dimension is 0. The calculation formula is: , where μ is the mean vector of each feature dimension; then the covariance matrix C is calculated, reflecting the linear correlation between features; next, eigenvalue decomposition is performed on C to solve the characteristic equation. We obtain the eigenvalues ​​and corresponding eigenvectors; by calculating the variance explained rate, we select the number of principal components k such that the cumulative variance contribution rate... (Retain 95% of the original information); then take the eigenvectors corresponding to the first k largest eigenvalues ​​to construct the projection matrix. Finally, a linear transformation is used to project the high-dimensional data into a low-dimensional space, resulting in a dimension-reduced aligned feature set. The dimensionality of this set is reduced but the main feature information of the original data is retained, which can be used for subsequent time series analysis and pattern recognition. The aligned feature set is a low-dimensional feature representation obtained through the above multi-step processing. Its acquisition process includes feature extraction, weight adjustment, similarity fusion and dimensionality reduction, which can be used for subsequent time series dependency analysis, anomaly detection and load prediction tasks, providing accurate feature input for energy storage cabinet energy monitoring.

[0031] In step S13, it is necessary to analyze the temporal dependencies of the aligned feature sets and integrate multi-source heterogeneous data to determine whether abnormal fluctuations occur, thereby obtaining a dynamic state representation sequence, including: The battery state sequence and load fluctuation sequence are extracted from the alignment feature set. The battery state sequence is time-series modeled, and long-term dependency features in the time dimension are extracted to obtain a first time-series dependency set. The weights of features at different time scales in the load fluctuation sequence are adjusted to generate a weighted time-series feature set. If the similarity between the weighted temporal feature set and the first temporal dependency set exceeds a preset feature similarity threshold, the weighted temporal feature set and the first temporal dependency set are fused to obtain a unified temporal feature representation set; Based on the unified time-series feature representation set, determine whether there are abnormal fluctuations in the load fluctuation sequence. If an anomaly is detected, generate a dynamic state representation sequence. The dynamic state characterization sequence includes the time points and amplitudes of load fluctuations.

[0032] It should be noted that the step of extracting the battery state sequence and load fluctuation sequence from the alignment feature set, performing time-series modeling on the battery state sequence, and extracting long-term dependency features in the time dimension to obtain the first time-series dependency set is implemented through a Long Short-Term Memory (LSTM) network. Specifically, a two-layer bidirectional LSTM architecture is adopted, with each layer containing 128 hidden units. The sequence modeling process is implemented through forget gate, input gate, and output gate mechanisms, and its calculation formula is as follows: (The forget gate controls the degree to which historical information is retained.) The output of the forget gate is represented by the sigmoid function. Calculations are used to control the memory units from the previous time step. The retention ratio, whose value ranges from 0 to 1, indicates that more historical information is retained as the value is closer to 1. (Input gate controls new information updates) Represents the input gate output, also through Function calculation determines the current input. The degree of updating of memory units; (Candidate memory units) These are candidate memory units, generated using the tanh function, which provides new candidate values ​​to update the memory state. (Memory unit update) It represents the state of the memory unit at the current time step, determined by the forget gate. and input gate Together; (Output gate controls output information). Represents the output gate output, through Function calculation; (Currently hidden state), where For the sigmoid function, It is currently in a hidden state, controlled by the output gate. and The product of the two is used as the output of the LSTM and passed to the next time step or subsequent layers; W is the weight matrix, which corresponds to the trainable parameters of each gate and is optimized through backpropagation; b is the bias vector, which is used to adjust the activation threshold of the gate; the network is trained for 200 epochs using the Adam optimizer with a learning rate of 0.001 and the parameters are optimized using the mean squared error loss function; finally, the hidden state outputs of all time steps are extracted and a 256-dimensional fixed-length feature vector is generated by global average pooling. This vector fully preserves the long-term trend (such as charge-discharge cycle mode) and short-term fluctuation (such as instantaneous voltage change) features of the battery's state of charge. The generated first temporal dependency set provides structured temporal feature input for subsequent multi-source feature fusion. The step of adjusting the weights of features at different time scales in the load fluctuation sequence to generate a weighted time-series feature set is achieved through an attention mechanism, specifically using the scaled dot product attention calculation method: First, the load fluctuation sequence is input into three different linear transformation layers to generate a query matrix Q, a key matrix K, and a value matrix V, where Q = XW_Q, K = XW_K, and V = XW_V (where X is the input sequence and W is the trainable weight matrix); then, the attention score is calculated using the formula: ,in The key vector dimension (usually set to 64) is determined by dividing by To prevent gradient vanishing, the softmax function normalizes the attention weights, automatically assigning higher weights (0.7-0.9) to minute-level fluctuation features and lower weights (0.1-0.3) to hourly-level trend features. The resulting weighted time-series feature set highlights short-term load change patterns, providing important input for subsequent multi-source data fusion.

[0033] If the similarity between the weighted temporal feature set and the first temporal dependency set exceeds a preset feature similarity threshold, the step of fusing the weighted temporal feature set and the first temporal dependency set to obtain a unified temporal feature representation set is achieved through cosine similarity calculation and vector concatenation. The specific process is as follows: First, calculate the cosine similarity between the weighted temporal feature set A (load fluctuation sequence) and the first temporal dependency set B (battery state sequence), using the formula: ,in Represents the vector dot product. and These are the L2 norms of the vectors; then, the two sets are horizontally concatenated along their dimensions to generate a joint vector. (m and n are the dimensions of A and B, respectively); finally, the data is fused through two fully connected layers. The first layer uses 256 neurons and the ReLU activation function, and the second layer uses 128 neurons and linear activation, ultimately generating a unified temporal feature representation set. This set integrates multi-source features and retains key temporal information; the feature similarity threshold is determined through historical data statistics and experimental verification. The grid search method is used to test different thresholds in the range of 0.5-1.0 with a step size of 0.05. The optimal balance point (usually 0.7-0.9) is selected by combining the precision-recall curve. Fusion is triggered when sim(A,B) is greater than or equal to the threshold. Based on the unified temporal feature representation set, the algorithm determines whether there are abnormal fluctuations in the load fluctuation sequence. If an anomaly is detected, the step of generating a dynamic state representation sequence is implemented using an isolation forest algorithm. This algorithm first constructs multiple isolation trees (usually 100 trees), and each tree recursively splits the data by randomly selecting features and splitting values ​​until a sample is isolated or the tree depth limit is reached. The path length of each sample is then calculated. This is the number of edges from the root node to isolated nodes, and the expected length of each path in the tree. ; The average path length is used for normalization, and the formula is: Where H is the harmonic number Abnormal scores The closer the value is to 1, the greater the likelihood of an anomaly. Related anomaly detection is determined by setting an anomaly threshold. Specifically, this is done through training on historical data, calculating the distribution of anomaly scores using a validation set, and selecting the 95th percentile as the threshold (e.g., ...). If the threshold is exceeded, an anomaly is identified; the dynamic state representation sequence contains the time point, amplitude and type information of abnormal fluctuations, which is used for real-time monitoring and scheduling decisions; the dynamic state representation sequence is a time series data set obtained through the above multi-step processing, the acquisition process of which includes feature extraction, weight adjustment, fusion and anomaly detection, which can be used for real-time early warning and optimized scheduling of energy storage systems, such as adjusting the battery charging and discharging strategy to maintain grid stability when a sudden increase in load is detected.

[0034] In step S14, it is necessary to calculate the load fluctuation probability in the dynamic state characterization sequence. If it is higher than a preset probability threshold, the operating parameters are optimized and load prediction is performed to obtain the predicted load value, including: The temporal feature weights of load fluctuations are extracted from the dynamic state characterization sequence, and the contribution of each time scale feature in the load fluctuations is calculated. The load fluctuation probability is obtained by weighted averaging. If the load fluctuation probability is higher than a preset probability threshold, obtain the battery state sequence, fit the battery state sequence with the operating parameters, obtain the mapping relationship between the two, predict the load value change, and obtain the optimized set of operating parameters. Based on the optimized set of operating parameters, the charging and discharging parameters are adjusted to generate a set of charging and discharging parameters that matches the predicted load value. Extract the feature fusion weights of the battery and load from the set of charging and discharging parameters to generate a fused feature set; If the load surge trend in the fused feature set is consistent with the battery state decline trend, then the predicted load value is reasonable, and a definite predicted load value is obtained.

[0035] It should be noted that the step of extracting the temporal feature weights of load fluctuations from the dynamic state representation sequence, calculating the contribution of each time scale feature in the load fluctuations, and obtaining the load fluctuation probability after weighted averaging is achieved through multi-scale feature decomposition and weighted fusion. The specific process is as follows: First, the Daubechies wavelet basis (db4) is used to perform continuous wavelet transform on the load sequence, decomposing the original signal into components at different time scales, including minute-level components (scale a = ...). to Capture short-term fluctuations, hourly components (scale a = to Capture long-term trends; calculate the variance contribution rate for each component, using the formula: ,in Let n be the variance of the i-th component and n be the total number of components; weight Based on historical data analysis, different weight combinations were tested within the range of 0.1-0.9 using a grid search method and 5-fold cross-validation with a step size of 0.1. The optimal weights were selected based on the prediction accuracy on the validation set (e.g., the weight of the minute-level component was set to 0.7, and the weight of the hour-level component was set to 0.3). Finally, the load fluctuation probability was calculated using a weighted average. This probability value comprehensively reflects the fluctuation characteristics across multiple time scales and is used for subsequent threshold determination. If the load fluctuation probability is higher than a preset probability threshold, obtain the battery state sequence, fit the battery state sequence with the operating parameters, obtain the mapping relationship between the two, predict the load value change, and obtain the optimized set of operating parameters. This step is achieved through multiple linear regression, and the specific process is as follows: First, extract battery state feature sequences (such as state of charge and voltage fluctuation amplitude) and corresponding operating parameter data (such as charge / discharge rate and power distribution) from historical data to construct a feature matrix X and a response vector Y; then, estimate the regression coefficients using the least squares method, as shown in the formula. Solving for coefficients through matrix operations To ensure the minimization of the sum of squared residuals; to predict operating parameter values ​​based on the current battery state characteristics input into the model. Subsequently, based on the predicted changes in operating parameters... Combined with sensitivity coefficient Calculate load value changes The sensitivity coefficient It is obtained by fitting historical data, specifically using linear regression to analyze the correlation between changes in historical operating parameters ΔY and load changes ΔL. The formula is as follows: Where cov is the covariance and var is the variance, calculated based on the past 365 days of operational data, with a confidence interval of 95%, and determined through iterative optimization. =0.8, indicating that for every unit change in operating parameters, the load changes by 0.8 units accordingly. Finally, the operating parameters are adjusted based on the predicted load value to generate an optimized set of operating parameters, such as adjusting the charging and discharging rate and power allocation strategy to match grid demand. The model formula is as follows: Where Y is the running parameter, This refers to battery state characteristics. The parameters are estimated using the least squares method, with regression coefficients used to predict changes in load values. , The sensitivity coefficient is used. The preset probability threshold is determined by a combination of historical data analysis and experimental verification. The specific setting process is as follows: First, the load fluctuation probability sequence is extracted from the historical operating data, and its statistical distribution characteristics, such as mean, standard deviation, and quantiles, are calculated to understand the typical range of probability. Then, a grid search method is used to test different threshold candidate values ​​in the range of 0.5 to 1.0 with a step size of 0.05. The performance indicators of each threshold on the validation set, such as precision, recall, and F1 score, are evaluated by combining cross-validation. The point that achieves the best balance between precision and recall is selected (for example, when the threshold is 0.8, the precision reaches 90% and the recall is 85%, which meets the system's tolerance requirements for false positives and false negatives).

[0036] Based on the optimized set of operating parameters, the step of adjusting the charging and discharging parameters to generate a set of charging and discharging parameters that matches the predicted load value is achieved through a parameter mapping function. The specific process is as follows: First, based on historical operating data, establish the SOC (State of Charge) and... The piecewise linear interpolation model for (predicted load value) divides the State of Charge (SOC) into multiple intervals (e.g., 0-20%, 20-80%, 80-100%), and defines baseline charge and discharge parameters for each interval; for each interval, based on... The value is calculated using linear interpolation, and the formula is: ),in and For the corresponding load value and The preset parameter values ​​are used; the interpolation parameters are optimized through grid search and cross-validation to ensure adaptability under different operating scenarios; finally, a set of charging and discharging parameters that accurately matches the predicted load value is generated, including adjusted charging and discharging rates, power allocation strategies, etc., to optimize the energy storage system response.

[0037] The step of extracting the feature fusion weights of the battery and load from the set of charge and discharge parameters to generate the fused feature set is achieved through principal component analysis (PCA). The specific process is as follows: First, extract battery state features (such as state of charge, charge and discharge rate) and load features (such as power distribution, fluctuation amplitude) from the set of charge and discharge parameters to construct a feature matrix. Where m is the number of samples and n is the feature dimension; X is centered so that the mean of each feature dimension is 0, and the covariance matrix is ​​calculated. Then, eigenvalue decomposition is performed on C to solve the characteristic equation. , to obtain eigenvalues and the corresponding feature vector The number of principal components, k, is selected by calculating the cumulative variance explained rate, such that... (Retain 95% of the original information); take the first k eigenvectors to construct the projection matrix. Projecting the original data into a low-dimensional space Generate a fusion feature set This set integrates key features of the battery and load for subsequent energy distribution optimization. If the load surge trend in the fused feature set is consistent with the battery state decline trend, then the predicted load value is reasonable. This step of obtaining a definite predicted load value is achieved through trend correlation analysis, specifically as follows: First, extract the load time series X (representing a load surge trend, such as a short-term power increase) and the battery state time series Y (representing a battery state decline trend, such as a decrease in charge percentage) from the fused feature set, ensuring consistent sequence lengths and aligned timestamps; then calculate the Pearson correlation coefficient r. (Indicating a strong negative correlation), it is determined that the sudden increase in load is consistent with the downward trend of battery state, and the predicted load value is reasonable. Finally, based on the verification of this correlation, the confirmed predicted load value is output for optimizing the energy storage system scheduling decision. The predicted load value is the estimated future load power obtained through the above multi-step processing. It is obtained based on historical data training model and combined with real-time state input. The calculation results are used to optimize the charging and discharging strategy of energy storage cabinet, such as adjusting the charging and discharging rate and power allocation to balance grid demand and improve energy utilization efficiency.

[0038] In step S15, real-time energy monitoring data needs to be acquired and fused with the predicted load value. Through load balancing verification, the device parameter configuration is optimized to generate an optimized energy allocation scheme, including: Energy data is acquired from real-time monitoring, multi-source data streams are extracted and fused from the energy data, the feature weights of each data source are calculated, and the fused first energy dataset is obtained. The predicted load value is corrected by combining historical trends and periodic features in the first energy dataset. If the predicted load value is higher than a preset prediction threshold, the operating parameters are extracted and optimized from the battery state sequence to obtain an adjusted first set of operating parameters. Based on the first set of operating parameters, the correlation between the device's state of charge and charge / discharge rate is analyzed. Combined with the predicted load value, an optimized energy allocation scheme is generated. It should be noted that the step of acquiring energy data from real-time monitoring, extracting and fusing multi-source data streams from the energy data, calculating the feature weights of each data source, and obtaining the fused first energy dataset is achieved through multi-source data fusion technology. The specific process is as follows: First, heterogeneous data streams such as battery voltage, load power, and ambient temperature are acquired from the real-time monitoring system, with sampling frequencies at the second, minute, and hour levels, respectively. Min-max normalization is performed on each data source to eliminate dimensional differences. Subsequently, principal component analysis (PCA) is used to calculate the feature weights of each data source. The specific steps include: constructing the feature matrix of all normalized data and calculating the covariance matrix. Eigenvalues ​​are obtained by performing eigenvalue decomposition. According to the formula The variance contribution rate of each principal component is calculated as a weight (e.g., battery voltage weight 0.5, load power weight 0.3, ambient temperature weight 0.2); finally, the multi-source data is fused by weighted average, using the following formula: This generates the first fused energy dataset, which integrates multi-source information and retains key features. The predicted load value is corrected by combining historical trends and periodic characteristics in the first energy dataset. If the predicted load value is higher than a preset prediction threshold, operating parameters are extracted and optimized from the battery state sequence to obtain an adjusted first set of operating parameters. This step is achieved through real-time data fusion and dynamic calibration. Specifically, a sliding window mechanism is used to collect actual grid operating data (including parameters such as voltage, current, and power) and calculate the deviation between the predicted load value and the actual measured value. ;when When the predicted value exceeds a preset threshold, a correction mechanism is triggered: First, a weighted moving average algorithm is used to smooth the predicted value, as shown in the formula below. The weight Data decays exponentially based on its timeliness, with more recent data receiving higher weight. The correction coefficients were determined through regression analysis of historical data. Simultaneously, key indicators such as state of charge (SOC) and state of health (SOH) were extracted from the battery state sequence, and operating parameters were optimized using a multivariate linear regression model. The objective function was... Regularized gradient descent was used (learning rate 0.01, 1000 iterations, L2 regularization coefficient). =0.1) Solve for the optimal parameters Finally, the adjusted first set of operating parameters is generated, including the corrected load expectation value, the optimized charge and discharge rate (e.g., adjusted from 1.0 kW / min to 1.2 kW / min), and the dynamic power allocation ratio; the preset prediction threshold is determined through historical data analysis, and different thresholds are tested in the range of 0.5 to 1.0 using the grid search method. The precision and recall at each threshold are calculated by combining cross-validation, and the best balance point is selected (e.g., the threshold is set to 0.8, corresponding to a precision of 90% and a recall of 85%).

[0039] Based on the first set of operating parameters, the step of analyzing the correlation between the device's state of charge (SOC) and charge / discharge rate, and generating an optimized energy allocation scheme by combining the predicted load value, is implemented through association rule mining and optimization algorithms. Specifically, the Apriori algorithm is used to analyze the historical association rules between SOC and charge / discharge rate: First, discrete transaction data of SOC and charge / discharge rate are extracted from historical operating data (e.g., SOC is divided into intervals [0-20%], [20-80%], [80-100%], and charge / discharge rate is divided into [0-1kW / min], [1-2kW / min], etc.). The minimum support min_support=0.1 and the minimum confidence min_confidence=0.7 are set. Frequent itemsets are generated by scanning the dataset multiple times, and association rules are derived based on the frequent itemsets (e.g., the rule "SOC∈[80-100%]→discharge rate=1.5kW / min"). The reliability of the rule is evaluated by calculating the confidence level, where the confidence level represents the value of the SOC when the SOC is within a certain range. The probability that the charge / discharge rate meets expectations when the OC is within a specific range is calculated. Then, combined with load forecasts, a linear programming model is used to solve for the optimal allocation scheme. The objective function is to maximize energy utilization, i.e., maximize the ratio of energy output to energy input. Constraints include SOC range limits, upper limits for charge / discharge rates, load demand matching, and battery health status. The simplex method is used to iteratively optimize parameters, ultimately generating an optimized energy allocation scheme, including specific charge / discharge rate adjustment values ​​and power allocation strategies. The optimized energy allocation scheme is a set of charge / discharge strategies obtained through the above multi-step processing, including charge / discharge rate adjustment values, power allocation ratios, and timing scheduling instructions. It is obtained based on multi-source data fusion, time series prediction, and optimization algorithm calculations, and is used to guide the energy storage cabinet to adjust its operating parameters in real time, such as dynamically adjusting the charge / discharge rate (e.g., increasing it from 1.0 kW / min to 1.2 kW / min) and power allocation (e.g., allocating 70% of the power to peak load periods) based on load forecasts, to improve energy utilization efficiency and ensure stable grid operation.

[0040] In step S16, equipment operating parameters need to be extracted from the optimized energy allocation scheme. If the load balancing index in the equipment operating parameters meets the preset load threshold, a verified allocation scheme is generated, including: Obtain the battery voltage fluctuation sequence from real-time monitoring and smooth it to obtain the smoothed first fluctuation sequence dataset; Extract the voltage change features from the first fluctuation sequence dataset. If the voltage change feature data is higher than a preset feature data threshold, extract and calibrate the equipment operating parameters from the optimized energy distribution scheme to obtain the first set of calibrated operating parameters. Based on the first set of operating parameters, a load power balance value is obtained by fitting. If the load power balance value meets the preset load threshold, a verified allocation scheme is generated.

[0041] It should be noted that the step of obtaining the battery voltage fluctuation sequence from real-time monitoring and smoothing it to obtain the smoothed first fluctuation sequence dataset is achieved through a mean filtering algorithm. The specific process is as follows: First, the original battery voltage fluctuation sequence is obtained from the real-time monitoring system. The sliding window size was set to 5 data points; zero-padding was applied to the sequence to mitigate boundary effects, i.e., two zero values ​​were added to the beginning and end of the sequence; then, the sliding window arithmetic mean was calculated using the following formula: The algorithm iterates through all valid positions from 3 to n-2. After each window is calculated, the sum of five adjacent voltage values ​​is divided by 5 to obtain the smoothed value at the center point of the window. Finally, the smoothed first fluctuation sequence dataset is generated. This process effectively suppresses random noise (such as measurement errors and transient disturbances) while preserving the overall trend characteristics of voltage changes.

[0042] The step of extracting voltage change features from the first fluctuation sequence dataset, and extracting and calibrating equipment operating parameters from the optimized energy allocation scheme to obtain the calibrated first set of operating parameters, is achieved through time series analysis and linear regression. The specific process is as follows: First, the voltage change rate is calculated using the difference method, with time intervals... Set to 1 minute (based on sampling frequency), the formula is: ,in and The voltage values ​​are given at continuous time points; then the calculated rate of change is compared with a preset characteristic data threshold of 0.3V / min. If the voltage is >0.3V / min, then the current operating parameters, such as the charge / discharge rate, are extracted from the optimized energy allocation scheme. Then, using a linear regression model... Calibration is performed, where the coefficients and It is trained using historical datasets, which contain ideal operating parameters. and actual parameters The objective function is The optimal solution is obtained using the gradient descent method. and The learning rate is set to 0.01, and the iteration is performed 1000 times until the loss function converges. Finally, the first set of calibrated operating parameters is generated, including the adjusted charge and discharge rates and other relevant parameters. The feature data threshold is determined through historical data statistics and experimental verification. Specifically, the grid search method is used to test different thresholds in the range of 0.1 to 0.5 V / min with a step size of 0.05. The optimal balance point is selected by combining the precision-recall curve. It is usually set to 0.3 V / min to ensure high sensitivity and low false alarm rate.

[0043] Based on the first set of operating parameters, a load power balance value is fitted. If the load power balance value meets a preset load threshold, the step of generating a verified allocation scheme is achieved through a power balance equation, the specific calculation formula of which is as follows: ,in Extracted in real time from the power generation unit. Obtained in real time from load monitoring. The energy storage efficiency coefficient, η, is obtained from the battery status data of the energy storage cabinet and is fitted to historical operating data using the least squares method. The objective function is... Gradient descent is used to iteratively optimize with a learning rate of 0.001. Value, when | |≤0.05· The system determines whether the load threshold is met and generates a verified allocation scheme, including adjusted charge / discharge rates, power allocation ratios, and timing scheduling instructions. The load threshold is determined through statistical analysis of historical data. Specifically, a grid search method is used to test different thresholds in 0.5% steps within a range of 1% to 10% of the rated power. The optimal balance point is selected by combining system stability indicators such as frequency deviation and voltage fluctuation. Typically, it is set to ±5% of the rated power based on historical load fluctuation data within a 95% confidence interval. The verified allocation scheme is a set of optimized operating parameters obtained through the above multi-step processing. It includes calibrated charge / discharge rates, power allocation ratios, and timing scheduling strategies. It is obtained based on real-time voltage feature extraction, parameter calibration, and power balance verification. It is used to guide the actual operation of the energy storage system. For example, the charge / discharge rate is adjusted to the calibrated value of 1.2P_current, and critical loads are prioritized for power supply according to the allocation ratio during load fluctuations to ensure optimal system stability and energy efficiency.

[0044] In step S17, it is necessary to extract and integrate feature set similarity information from the verified allocation scheme. If the integrated information result matches the grid demand, the grid demand parameters are adjusted and a demand forecast set is generated, including: Feature set similarity information is extracted from the verified allocation scheme, and the similarity information is classified to obtain the first similarity dataset after classification. The feature set similarity information includes similarity measures of voltage, current, and power distribution; The first similarity dataset is weighted and fused to generate the first demand prediction set; Determine the matching status between the first demand forecast set and the grid demand. If the matching is consistent, verify the grid demand parameters of the first demand forecast set to obtain the verified first forecast parameter set. The power grid demand parameters include the allocated power of the equipment and the load priority. The first set of estimated parameters after verification is dynamically calibrated, and the power grid demand parameters are adjusted according to the verification results to generate a demand estimate set.

[0045] It should be noted that the step of extracting feature set similarity information from the verified allocation scheme, classifying the similarity information, and obtaining the first similarity dataset after classification is implemented using the K-means clustering algorithm. The specific process is as follows: First, extract multi-dimensional feature sets (including indicators such as voltage stability, load matching degree, and timing consistency) from the verified allocation scheme, and calculate the Euclidean distance between feature vectors as a similarity measure; then initialize cluster centers, and use the k-means++ algorithm to select three initial centroids (corresponding to high, medium, and low demand scenarios), set the maximum number of iterations to 100, and the convergence threshold to 0.001; update the cluster allocation through an iterative optimization process: for each feature vector, calculate its Euclidean distance to all centroids and assign it to the cluster containing the nearest centroid, and then recalculate the centroid position of each cluster (taking the mean of all points within the cluster); stop iterating when the centroid movement distance is less than the threshold or the maximum number of iterations is reached, and finally generate the first similarity dataset after classification, which contains feature sets of three clusters and their centroid coordinates. The step of weighted fusion of the first similarity dataset to generate the first demand prediction set uses the entropy weight method to determine the weight of each feature. The specific process is as follows: First, the first similarity dataset is subjected to min-max normalization to eliminate the influence of units. Then, the proportion of the i-th sample value under the j-th feature is calculated. Next, the information entropy of each feature is calculated. ; Calculate the difference coefficient based on the entropy value And finally determine the feature weights. Finally, the first demand forecast set is generated through weighted fusion. ,in Let be the j-th feature vector. This weight allocation method reflects the degree of data dispersion through information entropy. The smaller the entropy value, the greater the degree of feature variation, and the higher the weight assigned, thereby ensuring that key features dominate the fusion process.

[0046] The first step involves determining the match between the first demand forecast set and the grid demand. If a match is found, the first demand forecast set is validated using grid demand parameters to obtain a validated first forecast parameter set. This validation process employs a logistic regression model, and the specific steps are as follows: First, grid demand features and matching labels (y=1 indicates a match, y=0 indicates a mismatch) are extracted from historical data to construct a training set; then, a hypothesis function is defined. Where x is the feature vector, Given a parameter vector; solve for the parameters using maximum likelihood estimation. The likelihood function is The gradient descent method is used for iterative optimization, and the update rule is as follows: Learning rate Set the threshold to 0.01 and iterate 1000 times until convergence. Matching conditions with grid demand are set through historical data statistical analysis. A grid search method is used to test different probability thresholds within the range of 0.5 to 1.0 with a step size of 0.05. The AUC value is calculated using the ROC curve, and the optimal balance point (usually set to 0.9) is selected. A match is determined when the parameter value is greater than 0.9; finally, the first set of estimated parameters after verification is generated, including the optimized parameters. Matching probability and matching status flags. The step of dynamically calibrating the first set of verified predicted parameters and adjusting the power grid demand parameters according to the verification results to generate a demand prediction set is implemented using an adaptive Kalman filter algorithm. The specific process is as follows: First, initialize the state vector. (Including key parameters such as load power and battery state of charge), error covariance matrix Set as diagonal matrices, the process noise covariance matrix Q and the measurement noise covariance matrix R are determined through statistical analysis of historical data (e.g., ...). The prediction phase uses state transition equations. (Where A is the state transition matrix, and B is the control input matrix,) (For external control variables), and simultaneously update the prior error covariance. The correction phase calculates the Kalman gain. (H is the observation matrix), combined with actual measured values Update state estimation and posterior error covariance The parameters are dynamically calibrated through recursive iteration, and finally a demand forecast set is generated, which includes the calibrated load power sequence, battery state parameters and confidence intervals. The demand forecast set is a set of grid demand prediction results obtained through multi-step processing, which includes load power, time series distribution and priority parameters. It is obtained based on feature clustering, weighted fusion and dynamic calibration, and is used to guide the scheduling of energy storage system, such as adjusting the charging and discharging strategy according to the forecast results to match the actual grid demand.

[0047] In step S18, the operating parameters of the energy storage cabinet need to be updated based on the demand forecast set, and a monitoring output report needs to be generated by collecting and transmitting real-time feedback data, including: Feature data of microgrid demand is extracted from the demand forecast set, and the feature data is classified to obtain the first microgrid demand dataset after classification. The microgrid demand characteristic data includes load power, energy storage cabinet status, and grid frequency; Real-time load data is extracted from the monitoring network sensors and integrated with the demand data of the first microgrid to generate the parameter set of the first energy storage cabinet; If the deviation between the first energy storage cabinet parameter set and the expected value of the power grid is within the preset deviation threshold, the operating parameters are updated to obtain the updated first operating parameter set; The system monitors and feeds back the power grid's operating data in real time, compares it with the updated first set of operating parameters, and generates a monitoring output report if the result is within the preset operating threshold.

[0048] It should be noted that the step of extracting feature data of microgrid demand from the demand forecast set and classifying the feature data to obtain the classified first microgrid demand dataset is implemented using the K-means clustering algorithm. The specific process is as follows: First, multi-dimensional feature vectors (including indicators such as load power, voltage fluctuation, and frequency deviation) are extracted from the demand forecast set, and Euclidean distance is used to measure feature similarity. Then, three cluster centers (corresponding to high, medium, and low load scenarios) are initialized using the k-means++ algorithm, with a maximum iteration count of 100 and a convergence threshold of 0.001. In each iteration, the Euclidean distance from each feature vector to each cluster center is calculated, and the feature vector is assigned to the nearest cluster. Then, the centroid position of each cluster is recalculated (taking the mean of all points within the cluster). When the centroid movement distance is less than the threshold or the maximum iteration count is reached, the iteration stops. Finally, the classified first microgrid demand dataset is generated, containing feature sets of the three clusters and their centroid coordinates, providing a scenario classification basis for subsequent energy allocation. The step of extracting real-time load data from monitoring network sensors and integrating the demand data of the first microgrid to generate the parameter set of the first energy storage cabinet is achieved using a weighted average fusion method. The specific process is as follows: First, real-time load data (such as current, voltage, and power values) is extracted from the monitoring network sensors, and at the same time, the demand dataset of the first microgrid (including load forecast values ​​and time-series characteristics) is obtained; the variance of each data source is calculated. Based on historical data or sliding window statistics (window size set to 30 sample points), among which The standard deviation of the i-th data source is represented; then the weights are calculated based on the variance. The smaller the variance, the greater the weight, emphasizing the importance of high-precision data sources; through the fusion formula A weighted average is performed on the multi-source data, where The feature vector of the i-th data source is used to generate the first energy storage cabinet parameter set, which includes the power allocation ratio (e.g., 60% for the main load area and 40% for the auxiliary load area), the charging and discharging rate (e.g., 1.5kW / min) and other operating parameters, ensuring that the parameter set reflects the optimized integration of real-time load and microgrid demand. If the deviation between the parameter set of the first energy storage cabinet and the expected value of the power grid is within a preset deviation threshold, the operating parameters are updated to obtain the updated first operating parameter set. This step is implemented using a proportional-integral-derivative (PID) control algorithm, and the specific process is as follows: First, the actual power of the power grid is collected in real time. Compared with expected power Calculate instantaneous deviation Subsequently, a control quantity is generated through a PID controller, and the calculation formula is used. Among them, the proportional term (Take 0.8) Quick response to the current deviation, integral term (Take 0.2) Eliminate steady-state error, differential term (Setting the value to 0.1) suppresses overshoot; parameter adjustment employs a trial-and-error method combined with the Ziegler-Nichols tuning rule. The specific process is as follows: First, the proportional gain is manually adjusted using a trial-and-error method. (Maintain the integral term) and differential terms (From zero), gradually increase Measure and observe the system's step response until the output exhibits constant-amplitude oscillations (critical oscillation state), and record the critical gain at this point. and oscillation period Then, the Ziegler-Nichols formula is applied to calculate the PID parameters. For a standard PID controller, the following values ​​are taken: , , Finally, the parameters were fine-tuned and optimized, and adjustments were made based on the actual response curve. , , Fine-grained values ​​(e.g., ±0.05 step size) are used to balance response speed, overshoot, and steady-state error, ensuring stable system operation under load fluctuations. The preset deviation threshold is determined through historical operating data analysis, specifically by using a sliding window method to collect power deviation sequences over the past N time points (e.g., N=1000 sampling points) and calculating the mean deviation. and standard deviation Set the threshold range to Where k is selected according to the system stability requirements (e.g., k=2 corresponds to the 95% confidence interval), or directly set as a percentage of the rated power (e.g., ±5%). Different thresholds are tested in the range of 3% to 10% using the grid search method. The false alarm rate and false negative rate are calculated by combining cross-validation, and the optimal threshold is selected to balance the system sensitivity and stability. The step of real-time monitoring and feedback of power grid operation data, compared with the updated first set of operation parameters, and generating a monitoring output report if the result is within the preset operation thresholds (voltage fluctuation ±2%, frequency deviation ±0.1Hz), employs a data consistency check algorithm. The specific process is as follows: First, extract key indicator sequences such as voltage and frequency from the real-time monitoring data stream, and perform time alignment and normalization with the predicted values ​​in the first set of operation parameters; then, calculate the Pearson correlation coefficient. ,in Let covariance be the variance of the two sets of data. and The data points are their standard deviations. The dynamic correlation coefficient is calculated in real time using a sliding window (e.g., a 60-second window). Specifically, the window size is set to 60 seconds, and the sliding step size is 1 second (based on the data sampling frequency). A circular buffer is maintained within each window to store the latest 60 data points, and the mean of the data within the window is updated in real time. and , sum of squares , and product sum Covariance is calculated incrementally. and standard deviation To avoid full recalculation and improve efficiency, the oldest data point is removed and a new point is added with each swipe step. After updating the statistics, the correlation coefficient r of the current window is calculated, achieving low-latency dynamic monitoring. At the same time, the mean absolute error is calculated. As an auxiliary indicator, consistency judgment must simultaneously meet the following criteria: (The threshold is obtained through training with historical data, and the optimal balance point is determined using an ROC curve.) ( The error tolerance is set to 0.05 per unit. Operating thresholds are set based on historical operating data statistical analysis, collecting long-term voltage and frequency data (e.g., the past 365 days), calculating the distribution characteristics of daily volatility, and using the mean ± 2 standard deviations for voltage fluctuation thresholds (covering a 95% confidence interval). Frequency deviation thresholds are determined through spectrum analysis to identify typical fluctuation ranges (e.g., ±0.1Hz corresponds to a 99% probability of stable operation). Different threshold combinations are tested within the ranges of ±1% to ±3% and ±0.05Hz to ±0.15Hz using a grid search method. The optimal value is selected through cross-validation (precision > 90%, recall > 85%), ultimately generating a monitoring output report containing consistency scores, deviation analysis, and early warning suggestions. The monitoring output report is a structured document containing system status assessment, parameter deviation analysis, and early warning levels, obtained through real-time data acquisition, algorithm analysis, and threshold judgment. It is used for operation and maintenance decisions and system optimization, such as guiding adjustments to energy storage cabinet charging and discharging strategies or providing fault warnings.

[0049] Reference Figure 2 The second embodiment of the present invention provides an energy monitoring device for an energy storage cabinet based on multi-source data, comprising: The data acquisition module is used to acquire multi-source heterogeneous data, standardize it to obtain an initial dataset, and perform feature extraction and pattern recognition on the initial dataset to obtain a vector dimension mapping. The feature analysis module is used to perform feature fusion and feature dimensionality reduction to obtain an aligned feature set if the similarity between vectors in the vector dimension mapping exceeds a preset similarity threshold. The temporal integration module is used to analyze the temporal dependencies of the aligned feature sets and integrate multi-source heterogeneous data to determine whether abnormal fluctuations occur, thereby obtaining a dynamic state representation sequence. The parameter optimization module is used to calculate the load fluctuation probability in the dynamic state characterization sequence. If it is higher than the preset probability threshold, the operating parameters are optimized and the load is predicted to obtain the predicted load value. The load balancing module is used to acquire real-time energy monitoring data and integrate it with the predicted load value. Through load balancing verification, the parameter configuration of the equipment is optimized to generate an optimized energy distribution scheme. The scheme generation module is used to extract equipment operating parameters from the optimized energy distribution scheme. If the load balance index in the equipment operating parameters meets the preset load threshold, a verified distribution scheme is generated. The demand matching module is used to extract and integrate feature set similarity information from the verified allocation scheme. If the integrated information result matches the power grid demand, the power grid demand parameters are adjusted and a demand prediction set is generated. The results output module is used to update the operating parameters of the energy storage cabinet based on the demand forecast set, and to collect and generate a monitoring output report through real-time feedback data.

[0050] It should be noted that the energy monitoring device for energy storage cabinet based on multi-source data provided in this embodiment of the invention is used to execute all the process steps of the energy monitoring method for energy storage cabinet based on multi-source data in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0051] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a voltage output control program. When the processor executes the computer program, it implements the steps in the above-described embodiments of the energy monitoring method for energy storage cabinets based on multi-source data, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the demand matching module.

[0052] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0053] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0054] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0055] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0056] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0057] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0058] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for monitoring the energy of an energy storage cabinet based on multi-source data, characterized in that, include: Acquire multi-source heterogeneous data and standardize it to obtain an initial dataset. Perform feature extraction and pattern recognition on the initial dataset to obtain a vector dimension mapping, including: A multi-source heterogeneous data set is acquired, and its time scale format is standardized to obtain an initial dataset. This multi-source heterogeneous data includes sequences of minute fluctuations in battery voltage and sequences of minute load changes within a microgrid. Based on the initial dataset, battery voltage sequences and load change sequences are extracted, and fluctuation features are extracted from the battery voltage sequences to obtain a first feature vector set. The load change sequences are decomposed using multi-scale methods to obtain a second feature vector set. If the time scales of the first and second feature vector sets are consistent, they are integrated to obtain a multi-source heterogeneous sequence. The multi-source heterogeneous sequence is then subjected to dimensionality reduction processing to extract and identify key patterns, resulting in a vector dimension mapping. If the similarity between vectors in the vector dimension mapping exceeds a preset similarity threshold, feature fusion and feature dimensionality reduction are performed to obtain an aligned feature set; The temporal dependencies of the aligned feature sets are analyzed, and multi-source heterogeneous data are integrated to determine whether abnormal fluctuations occur, thus obtaining a dynamic state representation sequence. Calculate the load fluctuation probability in the dynamic state characterization sequence. If it is higher than the preset probability threshold, optimize the operating parameters and perform load prediction to obtain the predicted load value. Real-time energy monitoring data is acquired and fused with the predicted load value. Through load balancing verification, the parameter configuration of the equipment is optimized, and an optimized energy allocation scheme is generated. Extract equipment operating parameters from the optimized energy allocation scheme. If the load balance index in the equipment operating parameters meets the preset load threshold, generate a verified allocation scheme. Extract and integrate feature set similarity information from the verified allocation scheme. If the integrated information result matches the power grid demand, adjust the power grid demand parameters and generate a demand forecast set. Based on the aforementioned demand forecast set, update the operating parameters of the energy storage cabinet, and collect and generate monitoring output reports through real-time feedback data.

2. The energy monitoring method for energy storage cabinets based on multi-source data according to claim 1, characterized in that, If the similarity between vectors in the vector dimension mapping exceeds a preset similarity threshold, feature fusion and feature dimensionality reduction are performed to obtain an aligned feature set, including: The battery state sequence and load fluctuation sequence are obtained from the multi-source heterogeneous data; Extract the dynamic change features of the battery state sequence over time to obtain the first time-series feature set; The weights of features at different time scales in the load fluctuation sequence are adjusted to obtain a weighted feature set; If the similarity between the weighted feature set and the first temporal feature set exceeds a preset similarity threshold, the weighted feature set and the first temporal feature set are merged to generate a unified feature representation set; The unified feature representation set is subjected to dimensionality reduction processing to obtain the aligned feature set.

3. The energy monitoring method for energy storage cabinets based on multi-source data according to claim 1, characterized in that, The analysis aligns the temporal dependencies of the feature sets, integrates multi-source heterogeneous data, determines whether abnormal fluctuations occur, and obtains a dynamic state representation sequence, including: Battery state sequences and load fluctuation sequences are extracted from the alignment feature set. Time-series modeling is performed on the battery state sequences to extract long-term dependency features in the time dimension, resulting in a first time-series dependency set. Adjust the weights of features at different time scales in the load fluctuation sequence to generate a weighted time series feature set; If the similarity between the weighted temporal feature set and the first temporal dependency set exceeds a preset feature similarity threshold, the weighted temporal feature set and the first temporal dependency set are fused to obtain a unified temporal feature representation set; Based on the unified time-series feature representation set, determine whether there are abnormal fluctuations in the load fluctuation sequence. If an anomaly is detected, generate a dynamic state representation sequence. The dynamic state characterization sequence includes the time points and amplitudes of load fluctuations.

4. The energy monitoring method for energy storage cabinets based on multi-source data according to claim 1, characterized in that, The calculation of the load fluctuation probability in the dynamic state characterization sequence, if higher than a preset probability threshold, optimizes the operating parameters and performs load prediction to obtain the predicted load value, including: The temporal feature weights of load fluctuations are extracted from the dynamic state characterization sequence, and the contribution of each time scale feature in the load fluctuations is calculated. The load fluctuation probability is obtained by weighted averaging. If the load fluctuation probability is higher than a preset probability threshold, obtain the battery state sequence, fit the mapping relationship between the battery state sequence and the operating parameters, predict the load value change, and obtain the optimized set of operating parameters. Based on the optimized set of operating parameters, the charging and discharging parameters are adjusted to generate a set of charging and discharging parameters that matches the predicted load value. Extract the feature fusion weights of the battery and load from the set of charging and discharging parameters to generate a fused feature set; If the load surge trend in the fused feature set is consistent with the battery state decline trend, then the predicted load value is reasonable, and a definite predicted load value is obtained.

5. The energy monitoring method for an energy storage cabinet based on multi-source data according to claim 4, characterized in that, The process of acquiring real-time energy monitoring data, fusing it with the predicted load value, verifying load balancing, optimizing equipment parameter configurations, and generating an optimized energy allocation scheme includes: Energy data is acquired from real-time monitoring, multi-source data streams are extracted and fused from the energy data, the feature weights of each data source are calculated, and the fused first energy dataset is obtained. The predicted load value is corrected by combining historical trends and periodic features in the first energy dataset. If the predicted load value is higher than a preset prediction threshold, the operating parameters are extracted and optimized from the battery state sequence to obtain an adjusted first set of operating parameters. Based on the first set of operating parameters, the correlation between the device's state of charge and charge / discharge rate is analyzed, and combined with the predicted load value, an optimized energy allocation scheme is generated.

6. The energy monitoring method for an energy storage cabinet based on multi-source data according to claim 1, characterized in that, The step of extracting equipment operating parameters from the optimized energy allocation scheme, and generating a verified allocation scheme if the load balancing index in the equipment operating parameters meets a preset load threshold, includes: Obtain the battery voltage fluctuation sequence from real-time monitoring and smooth it to obtain the smoothed first fluctuation sequence dataset; Extract the voltage change features from the first fluctuation sequence dataset. If the voltage change feature data is higher than a preset feature data threshold, extract and calibrate the equipment operating parameters from the optimized energy distribution scheme to obtain the first set of calibrated operating parameters. Based on the first set of operating parameters, a load power balance value is obtained by fitting. If the load power balance value meets the preset load threshold, a verified allocation scheme is generated.

7. The energy monitoring method for energy storage cabinets based on multi-source data according to claim 1, characterized in that, The step of extracting and integrating feature set similarity information from the verified allocation scheme, and adjusting the grid demand parameters and generating a demand forecast set if the integrated information matches the grid demand, includes: Feature set similarity information is extracted from the verified allocation scheme, and the similarity information is classified to obtain the first similarity dataset after classification; the feature set similarity information includes similarity measures of voltage, current, and power allocation; The first similarity dataset is weighted and fused to generate the first demand prediction set; Determine the matching status between the first demand forecast set and the grid demand. If the matching is consistent, verify the grid demand parameters of the first demand forecast set to obtain the verified first forecast parameter set. The grid demand parameters include the allocated power of the equipment and the load priority. The first set of estimated parameters after verification is dynamically calibrated, and the power grid demand parameters are adjusted according to the verification results to generate a demand estimate set.

8. The energy monitoring method for an energy storage cabinet based on multi-source data according to claim 1, characterized in that, The process of updating the operating parameters of the energy storage cabinet based on the demand forecast set and collecting and generating a monitoring output report through real-time feedback data includes: Feature data of microgrid demand is extracted from the demand forecast set, and the feature data is classified to obtain the first microgrid demand data after classification; the first microgrid demand data includes load power, energy storage cabinet status, and grid frequency. Real-time load data is extracted from the monitoring network sensors and integrated with the demand data of the first microgrid to generate the parameter set of the first energy storage cabinet; If the deviation between the first energy storage cabinet parameter set and the expected value of the power grid is within the preset deviation threshold, the operating parameters are updated to obtain the updated first operating parameter set; The system monitors and feeds back the power grid's operating data in real time, compares it with the updated first set of operating parameters, and generates a monitoring output report if the result is within the preset operating threshold.

9. An energy monitoring device for an energy storage cabinet based on multi-source data, used to implement the energy monitoring method for an energy storage cabinet based on multi-source data as described in any one of claims 1-8, comprising: The data acquisition module is used to acquire heterogeneous data from multiple sources and perform standardization processing to obtain an initial dataset; Feature extraction and pattern recognition are performed on the initial dataset to obtain a vector dimension mapping; The feature analysis module is used to perform feature fusion and feature dimensionality reduction to obtain an aligned feature set if the similarity between vectors in the vector dimension mapping exceeds a preset similarity threshold. The temporal integration module is used to analyze the temporal dependencies of the aligned feature sets and integrate multi-source heterogeneous data to determine whether abnormal fluctuations occur, thereby obtaining a dynamic state representation sequence. The parameter optimization module is used to calculate the load fluctuation probability in the dynamic state characterization sequence. If it is higher than the preset probability threshold, the operating parameters are optimized and the load is predicted to obtain the predicted load value. The load balancing module is used to acquire real-time energy monitoring data and integrate it with the predicted load value. Through load balancing verification, the parameter configuration of the equipment is optimized to generate an optimized energy distribution scheme. The scheme generation module is used to extract equipment operating parameters from the optimized energy distribution scheme. If the load balance index in the equipment operating parameters meets the preset load threshold, a verified distribution scheme is generated. The demand matching module is used to extract and integrate feature set similarity information from the verified allocation scheme. If the integrated information result matches the power grid demand, the power grid demand parameters are adjusted and a demand prediction set is generated. The results output module is used to update the operating parameters of the energy storage cabinet based on the demand forecast set, and to collect and generate a monitoring output report through real-time feedback data.

Images