Micro-grid island operation control method and system

Abstract

This invention relates to the field of power system technology and discloses a method and system for controlling islanded operation of a microgrid. The method includes: acquiring multi-source heterogeneous data under islanded operation mode of the microgrid to obtain an islanded operation dataset; based on the islanded operation dataset, using a pre-constructed deep neural network model to predict the load change trend under the islanded operation mode within a preset time window of the future, obtaining the load probability distribution of the microgrid; solving a multi-objective optimization function based on the load probability distribution according to predefined constraints to generate optimal output commands for distributed power sources and optimal charging / discharging commands for energy storage systems within the microgrid; and sending the optimal output commands and optimal charging / discharging commands to the microgrid control system, so that the microgrid control system can adjust the islanded operation state of the microgrid in real time according to the optimal output commands and optimal charging / discharging commands. This invention can improve the stability and power supply reliability of microgrid islanded operation.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a microgrid islanded operation control method and system. Background Technology

[0002] With the development of energy transition and smart grid technology, microgrids, as independent power supply units integrating distributed power sources and multiple loads, play a crucial role in improving power supply reliability and energy efficiency. Especially in islanded operation mode, microgrids operate independently from the main grid, requiring precise balancing of internal distributed power output and load demand, posing severe challenges to operational stability, economy, and security.

[0003] Traditional microgrid islanded operation control typically relies on preset strategies or static prediction models based on limited historical data. In existing technologies, load forecasting in microgrid islanded operation modes often depends on single historical data points or simple statistical models, whose robustness is insufficient to cope with complex dynamic disturbances such as weather fluctuations and changes in user behavior. This limitation of the prediction mechanism leads to a poor match between distributed generation output and actual load demand, resulting in power quality problems such as voltage fluctuations and frequency drift. In extreme cases, these problems may even trigger protection device tripping, severely impacting the microgrid's power supply continuity and operational reliability. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a microgrid islanded operation control method and system that can accurately adapt to dynamic operating environments, significantly improve supply and demand matching accuracy, and effectively ensure power supply reliability.

[0005] In a first aspect, embodiments of the present invention provide a microgrid islanding operation control method, comprising: Obtain multi-source heterogeneous data under the islanded operation mode of the microgrid to obtain the islanded operation dataset; Based on the islanded operation dataset, a pre-constructed deep neural network model containing a temporal feature extraction layer, a spatial feature learning layer, a hypergraph attention fusion layer, and a probability prediction output layer is used to predict the load change trend of the microgrid under the islanded operation mode within a future preset time window, thereby obtaining the load probability distribution of the microgrid. A multi-objective optimization function is constructed with the objectives of minimizing energy storage lifetime loss and maximizing carbon emission reduction. The multi-objective optimization function is then solved based on the load probability distribution under predefined uncertainty robust constraints and scenario elastic constraints to generate the optimal output command of the distributed power source and the optimal charging and discharging command of the energy storage system within the microgrid. The optimal output command and the optimal charge / discharge command are sent to the microgrid control system so that the microgrid control system can adjust the islanded operation status of the microgrid in real time according to the optimal output command and the optimal charge / discharge command.

[0006] Preferably, after sending the optimal output command and the optimal charge / discharge command to the microgrid control system so that the microgrid control system can adjust the islanded operation state of the microgrid in real time according to the optimal output command and the optimal charge / discharge command, the method further includes: The system monitors the operational status data of each node within the microgrid in real time, and dynamically corrects the deep neural network model and the multi-objective optimization function based on the deviation analysis results of the operational status data.

[0007] Preferably, the step of acquiring multi-source heterogeneous data under the microgrid islanded operation mode to obtain an islanded operation dataset includes: Collect node electrical parameters, source-storage operation data, regional environmental data, and historical source-load operation data under the islanded operation mode of the microgrid to obtain multi-source heterogeneous data; The preprocessed multi-source heterogeneous data is injected into the island spatiotemporal features and source-storage operation response features to obtain the island operation dataset.

[0008] Preferably, based on the islanded operation dataset, the method of predicting the load change trend of the microgrid under the islanded operation mode within a future preset time window using a pre-constructed deep neural network model that includes a temporal feature extraction layer, a spatial feature learning layer, a hypergraph attention fusion layer, and a probability prediction output layer, to obtain the load probability distribution of the microgrid, includes: The isolated dataset is divided to obtain an initial training set, and a generative adversarial network is introduced to augment the initial training set to obtain the target training set. Based on the target training set, the deep neural network model is trained using the power task joint loss function to obtain the load prediction model; The isolated operation dataset is input into the load prediction model, and the data is processed sequentially through the time series feature extraction layer, the spatial feature learning layer, the hypergraph attention fusion layer, and the probability prediction output layer to obtain the load probability distribution under the isolated operation mode within a preset time window of the microgrid.

[0009] Preferably, the joint loss function for the power task includes mean error loss and negative log-likelihood loss.

[0010] Preferably, the step of inputting the islanded operation dataset into the load prediction model and processing the data sequentially through the time-series feature extraction layer, the spatial feature learning layer, the hypergraph attention fusion layer, and the probability prediction output layer to obtain the load probability distribution under the islanded operation mode within a future preset time window of the microgrid includes: The isolated operation dataset is input into the load prediction model, and the isolated operation dataset is subjected to time series feature extraction and cross-domain fusion processing through the time series feature extraction layer to obtain a cross-domain time series feature matrix. The cross-domain temporal feature matrix is ​​input into the spatial feature learning layer. The spatial feature learning layer performs heterogeneous graph spatial feature learning and cross-type node association aggregation on the cross-domain temporal feature matrix to obtain the node spatial feature matrix. The cross-domain temporal feature matrix and the node spatial feature matrix are input into the hypergraph attention fusion layer to perform dynamic hypergraph construction and temporal-spatial feature attention fusion to obtain the hypergraph feature matrix; The hypergraph feature matrix is ​​input into the probability prediction output layer for Bayesian inference and adaptation to the risk threshold constraint of the islanding scenario, so as to obtain the load probability distribution under the islanding operation mode within the future preset time window of the microgrid.

[0011] Preferably, the process of constructing a multi-objective optimization function with at least the objectives of minimizing energy storage lifetime loss and maximizing carbon emission reduction, and solving the multi-objective optimization function based on the load probability distribution according to predefined uncertainty robust constraints and scenario elastic constraints, to generate the optimal output command of the distributed power source and the optimal charging and discharging command of the energy storage system within the microgrid, includes: With the goals of maximizing power supply reliability, minimizing operating costs, maximizing voltage and frequency stability, minimizing energy storage lifespan loss, and maximizing carbon emission reduction, a multi-objective optimization function is constructed, using the output of distributed power sources and the charging and discharging power of energy storage systems within the microgrid as optimization variables. Based on the load probability distribution, the particle swarm optimization algorithm is used to solve the multi-objective optimization function under predefined uncertainty robust constraints and scenario elastic constraints to generate the optimal output command of the distributed power source and the optimal charging and discharging command of the energy storage system in the microgrid.

[0012] Preferably, the uncertainty robust constraint conditions include islanded power balance constraints and frequency voltage stability constraints, and the scenario elastic constraint conditions include distributed power output constraints, energy storage charge / discharge rate-lifetime coupling constraints, and line power flow constraints.

[0013] Preferably, the real-time monitoring of the operating status data of each node within the microgrid, and the dynamic correction of the deep neural network model and the multi-objective optimization function based on the deviation analysis results of the operating status data, includes: The system monitors the operating status data of each node in the microgrid in real time, and obtains the operating status deviation value by comparing the operating status data with a preset benchmark value. The preset benchmark value includes the load prediction result output by the deep neural network model and the optimal command result obtained by solving the multi-objective optimization function. The parameters of the deep neural network model are corrected in real time based on the operational state deviation value, and the multi-objective optimization function is optimized in a rolling manner based on the operational state data.

[0014] Secondly, embodiments of the present invention provide a microgrid islanding operation control system, comprising: The data acquisition module is used to acquire multi-source heterogeneous data in the islanded operation mode of the microgrid to obtain the islanded operation dataset; The load forecasting module is used to predict the load change trend of the microgrid under the islanded operation mode within a preset time window in the future, based on the islanded operation dataset, through a pre-constructed deep neural network model that includes a time-series feature extraction layer, a spatial feature learning layer, a hypergraph attention fusion layer, and a probability prediction output layer, and to obtain the load probability distribution of the microgrid. The instruction generation module is used to construct a multi-objective optimization function with at least the goal of minimizing energy storage lifetime loss and maximizing carbon emission reduction, and to solve the multi-objective optimization function based on the load probability distribution according to predefined uncertainty robust constraints and scenario elastic constraints, so as to generate the optimal output instruction of the distributed power source in the microgrid and the optimal charging and discharging instruction of the energy storage system in the microgrid. The execution module is used to send the optimal output command and the optimal charge / discharge command to the microgrid control system, so that the microgrid control system can adjust the islanded operation status of the microgrid in real time according to the optimal output command and the optimal charge / discharge command.

[0015] Compared with the prior art, the microgrid islanding operation control method and system of this invention have the following advantages at least one point: (1) The four-layer deep neural network architecture accurately captures the long-term and short-term fluctuation patterns of island loads through adaptive fusion of temporal and spatial features. The output load probability distribution can quantify the uncertainty of prediction, providing a scientific basis for risk-avoidance scheduling. Compared with conventional models, the long-term prediction accuracy is significantly improved, and the adaptability to extreme scenarios is better. (2) The multi-objective optimization function takes into account both the minimum energy storage life loss and the maximum carbon emission reduction. Combining uncertainty robust constraints and scenario elastic constraints, it can dynamically adapt to different operating conditions such as extreme weather and low energy storage charge state. Under the premise of ensuring island power balance and voltage / frequency stability, it can significantly extend the life of energy storage system, improve the absorption rate of clean energy such as wind and solar, and reduce operating costs and carbon emission intensity. (3) By accurately generating the optimal output command of the distributed power source and the optimal charging and discharging command of the energy storage system, the refined scheduling of island energy resources can be realized. After the command is issued, it can drive the control system to respond and adjust quickly, effectively avoid risks such as power shortage and voltage collapse, reduce the probability of protection device malfunction, and comprehensively improve the safety, stability and control reliability of microgrid island operation. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a microgrid islanding operation control method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the process of obtaining the probability distribution of microgrid load through load forecasting in an embodiment of the present invention; Figure 3 This is a schematic diagram of the process for generating the optimal output command and the optimal charge / discharge command in an embodiment of the present invention; Figure 4 This is another schematic flowchart of a microgrid islanding operation control method according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a microgrid islanded operation control system according to an embodiment of the present invention; Figure label: 01. Data Acquisition Module; 02. Load Forecasting Module; 03. Instruction Generation Module; 04. Execution Module. Detailed Implementation

[0017] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by those skilled in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0019] Microgrid islanding operation refers to the independent operation of a microgrid, detached from the main grid. In this mode, the microgrid relies on internal distributed power sources (such as solar and wind power), energy storage systems, and various loads to form an autonomous power supply system, completing a closed loop of power generation, transmission, distribution, and consumption. This mode is often used in situations such as power outages due to main grid failures, remote areas without main grid coverage, and independent power supply needs in special scenarios. Compared to grid-connected operation, islanding operation places higher demands on the precision of energy resource scheduling, control response speed, and system robustness. Therefore, constructing scientific and efficient operation control methods to achieve dynamic power balance and multi-objective collaborative optimization in islanding scenarios is crucial to ensuring the safe and stable operation of microgrid islands.

[0020] like Figure 1 The diagram shown is a flowchart illustrating a microgrid islanding operation control method according to an embodiment of the present invention. (Refer to...) Figure 1 An embodiment of the present invention provides a microgrid islanding operation control method, comprising the following steps: S1. Obtain multi-source heterogeneous data under the islanded operation mode of the microgrid to obtain the islanded operation dataset; Specifically, step S1 includes: 1) Collect node electrical parameters, source-storage operation data, regional environmental data, and historical source-load operation data under the islanded operation mode of the microgrid to obtain multi-source heterogeneous data; The system collects voltage, current, power, and system frequency data at key nodes of the microgrid island using a power parameter sensor array, focusing on capturing transient data such as voltage dips and frequency surges during island switching. It also collects real-time island output data for photovoltaic and wind power generation using a distributed power source monitoring unit, as well as island charging and discharging power, state of charge (SOC), and state of health (SOH) data for the energy storage system. Furthermore, it utilizes an environmental and meteorological sensor array to collect environmental data such as temperature, humidity, wind speed, and solar irradiance in the island area. Finally, it extracts historical source-load data, including island load curves, distributed power source island output curves, energy storage island operation data, and holiday load characteristics, through a historical data storage and interface unit.

[0021] During the data collection process, interactive data related to the main grid (such as the power of the main grid tie line) are removed, and electrical impact data at the time of islanding switching and steady-state characteristic data of islanding operation are added. Finally, the data are integrated to form multi-source heterogeneous data covering four dimensions: electrical, source and storage, environment, and history.

[0022] 2) Inject the preprocessed multi-source heterogeneous data into the island spatiotemporal features and source-storage operation response features to obtain the island operation dataset.

[0023] Preprocessing of the acquired multi-source heterogeneous data: Specifically, this embodiment adopts Outlier detection is performed, missing data is filled using linear interpolation, and denoising is achieved using wavelet transform to ensure data quality. Time alignment is performed on data with different sampling frequencies (e.g., 1 min / time for sensors, 5 min / time for meteorology), and resampling is performed to obtain a unified time series of 15 min / steps to eliminate differences in time dimension. Z-score normalization is used to standardize data with different dimensions such as voltage and power to the [0,1] interval to avoid the influence of dimensions on model training. At the same time, a microgrid island topology diagram is constructed, and spatial features such as node dependencies and line impedance are extracted to lay the foundation for subsequent feature injection.

[0024] After data preprocessing, target feature injection is achieved through quantization representation and feature fusion: a) Spatiotemporal characteristics of isolated islands: The spatial-temporal characteristics of islanding refer to the dynamic coupling characteristics of the spatial layout of core elements such as loads, nodes, and lines in the islanding operation mode of a microgrid (such as the power supply dependence between nodes, line impedance distribution, and regional load aggregation patterns), as well as the fluctuation patterns of loads in the time dimension (such as short-term load fluctuations and long-term periodic changes).

[0025] To address the spatiotemporal characteristics of isolated areas, on the one hand, we calculate the spatiotemporal clustering characteristics of loads, such as the regional load coupling coefficient and the load correlation of nodes; on the other hand, we transform the extracted spatial features, such as node dependencies and line impedance, into structured feature vectors, and fuse them with the time-series dimension of the preprocessed data step by step.

[0026] b) Source-storage operation response characteristics: The source-storage operation response characteristics refer to the output fluctuation patterns of distributed power sources (photovoltaics, wind power, etc.) in a microgrid under islanded conditions (such as output stability and instantaneous fluctuation amplitude), as well as the dynamic response capabilities exhibited by the energy storage system to adapt to islanded power balance and support stable system operation (such as charge and discharge response speed, the correlation characteristics between charge and discharge rate and its own lifespan, and the adjustment capability under different SOC states).

[0027] To address the operational response characteristics of energy storage, the power output volatility of distributed power sources in islanded regions is quantified by calculating the 15-minute volatility of wind power and the standard deviation of photovoltaic power output. Simultaneously, the energy storage charging and discharging response time and the coupled correlation between charging and discharging rate and lifetime are extracted to characterize the energy storage support response capability. These two types of features are constructed into a dedicated feature matrix, which is then concatenated with the preprocessed multi-source heterogeneous data matrix according to the feature dimensions. Finally, an islanded operation dataset is formed that integrates electrical parameters, environmental data, historical source-load data, as well as the spatiotemporal characteristics of islanded regions and the operational response characteristics of energy storage.

[0028] S2. Based on the islanded operation dataset, a pre-constructed deep neural network model containing a temporal feature extraction layer, a spatial feature learning layer, a hypergraph attention fusion layer, and a probability prediction output layer is used to predict the load change trend of the microgrid under the islanded operation mode within a future preset time window, and obtain the load probability distribution of the microgrid. like Figure 2 As shown, this is a flowchart illustrating step S2. (Refer to...) Figure 2 Step S2 includes: S201. The isolated dataset is divided to obtain an initial training set, and a generative adversarial network is introduced to augment the initial training set to obtain the target training set. In this embodiment, the isolated dataset is divided into an initial training set, a validation set, and a test set in a ratio of 7:2:1. The initial training set is used for model parameter learning, the validation set is used for hyperparameter adjustment and overfitting monitoring during the iterative optimization process, and the test set is used for final model performance evaluation.

[0029] To address the issue of sparse real-world samples in isolated operation scenarios, a generative adversarial network based on isolated scenario constraints is introduced. Using a small amount of real isolated operation data as a foundation, virtual samples that conform to the coupling relationship of "electrical parameters-meteorological conditions-load demand" are generated and added to the initial training set. At the same time, load fluctuation variance labels are calculated and added to the training set samples based on the historical load fluctuation range, providing a supervision signal for subsequent probability prediction. Finally, a target training set with balanced data distribution and complete feature information is formed.

[0030] To facilitate understanding, the process of generative adversarial networks generating virtual samples will be explained in detail below: Generative Adversarial Networks (GANs) consist of a generator and a discriminator, and incorporate island scenario constraints to ensure the validity of virtual samples. The generator employs a deep convolutional neural network (CNN), taking a random noise vector and island scenario constraint parameters (such as typical meteorological intervals, load level ranges, and transient impact intensity thresholds) as input. Through multi-layer convolution and deconvolution operations, it learns the coupling mapping relationship between "electrical parameters-meteorological conditions-load demand" in real island data, outputting virtual samples that conform to the island's operational patterns (containing corresponding time-series load data, meteorological data, and electrical parameters). The discriminator also uses a convolutional neural network, taking either real island samples or virtual samples output by the generator as input. Its core task is to distinguish the source of the samples, optimizing parameters through gradient descent to improve discrimination accuracy.

[0031] During training, the generator and discriminator engage in an adversarial game. The generator iterates and optimizes with the goal of "deceiving the discriminator," while the discriminator updates its parameters with the goal of "accurately distinguishing between real and fake data." At the same time, a scenario constraint loss term is introduced (to ensure that the electrical parameters of the virtual samples, such as voltage and frequency, are within the stable operating range of the island, and that the correlation between load and weather conforms to actual laws). Training stops when the game reaches Nash equilibrium (the discriminator's discrimination accuracy is close to 50%). At this point, the generator can stably output high-quality virtual samples that are consistent with the distribution of real island data and meet the scenario constraints, effectively supplementing the sample size of the initial training set.

[0032] S202. Based on the target training set, train the deep neural network model using the power task joint loss function to obtain the load prediction model; The training process is optimized by adopting a transfer pre-training strategy. First, the deep neural network model is pre-trained using data from similar scenarios of grid-connected microgrids to initialize the core parameters of the temporal feature extraction layer and the spatial feature learning layer. Then, the target training set is input into the model for fine-tuning, which effectively alleviates the model overfitting problem caused by insufficient island sample size.

[0033] During training, the model parameters are optimized using a joint loss function for power tasks. This joint loss function includes mean error loss and negative log-likelihood loss. The expression for this function is: in, This represents the joint loss function for power tasks. This represents the weighting coefficient, used to balance the accuracy of mean prediction with the accuracy of probability distribution; it is typically set to 0.5. This represents the loss due to mean error. This represents the predicted average load. This represents the actual load value. This represents the negative log-likelihood loss. This represents the predicted load probability distribution.

[0034] In this embodiment, the optimizer uses AdamW with a learning rate decay. The initial learning rate is set to 0.001, and it decays by 10% every 10 training rounds. Overfitting is suppressed by weight decay, which improves the model's convergence stability. After training, the performance is verified using the validation set and the test set, respectively. The requirements are: basic prediction accuracy MAPE ≤ 5%, the proportion of actual load falling within the 95% confidence interval ≥ 90%, and MAPE ≤ 8% in extreme scenarios. Once the requirements are met, a qualified load prediction model is obtained.

[0035] S203. Input the islanded operation dataset into the load prediction model, and process the data sequentially through the time series feature extraction layer, spatial feature learning layer, hypergraph attention fusion layer and probability prediction output layer to obtain the load probability distribution under the islanded operation mode within the future preset time window of the microgrid.

[0036] Specifically, step S203 includes: 1) Input the isolated operation dataset into the load prediction model, and perform time series feature extraction and cross-domain fusion processing on the isolated operation dataset through the time series feature extraction layer to obtain the cross-domain time series feature matrix; The temporal feature extraction layer consists of a reversible temporal convolutional network and a meteorological-load attention embedding module. The input data consists of three types of core temporal data from the isolated operation dataset, including 15-minute / step historical time-series data of isolated load, environmental meteorological time-series data (temperature / wind speed / solar irradiance), and transient time-series data of isolated switching (voltage / frequency / power impact curves at the moment of switching).

[0037] Specifically, the input data is extracted dimensionally by stacking 8 layers of reversible residual convolutional blocks. The first 4 layers focus on capturing local time-series patterns such as load fluctuation patterns within 1 to 3 hours. The last 4 layers expand the receptive field through expansion factors to explore long-distance time dependencies such as periodic load correlation within 24 hours. At the same time, the reversible structure is used to trace the feature flow direction to locate the temporal root cause of load abrupt changes (such as transient shocks and extreme weather effects).

[0038] Furthermore, a correlation matrix between meteorological factors and load changes is constructed through a meteorological-load attention embedding module. The attention weights of each meteorological time series data are calculated using the Softmax function (e.g., the weight is increased to 0.7 for extreme wind speeds and set to 0.3 for regular meteorological time series). The weighted meteorological time series features are then fused with the load time series features extracted by the reversible temporal convolutional network step by step. The final output is a cross-domain time series feature matrix with a dimension of [time step × feature dimension] (e.g., 72h × 128 dimensions). This matrix simultaneously contains the load's own time series patterns and the coupled influence of meteorological factors.

[0039] 2) Input the cross-domain temporal feature matrix into the spatial feature learning layer. The spatial feature learning layer performs heterogeneous graph spatial feature learning and cross-type node association aggregation on the cross-domain temporal feature matrix to obtain the node spatial feature matrix. The spatial feature learning layer adopts a weighted heterogeneous graph attention network. The input consists of two parts: one is the cross-domain temporal feature matrix (as the initial temporal feature embedding of each node in the heterogeneous graph), and the other is the pre-constructed island heterogeneous graph topology (including three types of nodes: load nodes, distributed power generation nodes, and distribution network nodes, as well as two types of edges: electrical connection edges and source-load association edges, with the edge weight being the physical association strength between nodes).

[0040] The temporal features of different node types are embedded using a node type mapping matrix (the matrix dimensions for load nodes, power supply nodes, and distribution network nodes are all 128×64) to perform type-aware transformation, projecting them onto a unified subspace to form an initial node feature representation of N×64 dimensions (N being the number of nodes). Then, weighted heterogeneous attention calculation is performed. First, a basic weight of 0.6 is assigned to electrical connection edges and 0.4 is assigned to source-load association edges. Then, dynamic weight adjustment is learned through the linear transformation of the LeakyReLU activation function to capture the differences in the contribution of different edge types to information transmission. At the same time, an 8-head multi-head attention mechanism is used for each node to calculate its attention coefficient with neighboring nodes, and a node type mask is introduced to ensure that only nodes of the same type or related type participate in the attention calculation (e.g., load nodes only establish associations with power supply nodes and distribution network nodes).

[0041] Finally, iterative aggregation is performed through three heterogeneous attention layers. Each layer sums the features of neighboring nodes according to their attention coefficients, concatenates them with its own features, and updates the node features through a linear layer and the GELU activation function. This process gradually captures spatial dependencies such as the power supply impact of distributed generation on surrounding loads and the regional regulation role of distribution network nodes, ultimately outputting an N×128-dimensional node spatial feature matrix. Each row vector in this matrix corresponds to the comprehensive features of an isolated node (load node, distributed generation node, or distribution network node), which includes both the initial temporal feature embedding of the node after type-aware transformation and the spatial dependencies between nodes obtained through weighted heterogeneous attention calculation (such as the power supply support of distributed generation to surrounding loads and the regional regulation association of distribution network nodes).

[0042] 3) Input the cross-domain temporal feature matrix and the node spatial feature matrix into the hypergraph attention fusion layer to perform dynamic hypergraph construction and temporal-spatial feature attention fusion to obtain the hypergraph feature matrix; The input data for the hypergraph attention fusion layer includes cross-domain temporal feature matrices, node spatial feature matrices, and real-time operating status data of isolated nodes (voltage fluctuation sequences, frequency change curves, power surge timing, etc.).

[0043] First, a dynamic hypergraph is constructed, where the temporal features corresponding to the cross-domain temporal feature matrix and the spatial features corresponding to the node spatial feature matrix are respectively used as two types of hypernodes in the hypergraph. Hyperedges connect different types of hypernodes (i.e., temporal-spatial hyperedges), and the hyperedge weights are initialized based on the island's operating state (e.g., when voltage fluctuations are severe, the hyperedge weights associated with temporal and spatial features are temporarily increased to 0.8). Then, a deep deterministic policy gradient reinforcement learning algorithm is introduced, with the prediction error minimization as the reward function. In each iteration, the reward is calculated based on the feature fusion effect of the current hypergraph (positive reward for reduced prediction error, negative reward for increased prediction error), and the hyperedge weight allocation strategy is optimized in real time to achieve dynamic adaptation of weights to the island's state. Finally, a multi-layer hypergraph attention mechanism is used to aggregate the hypernode features, calculate the attention coefficients between hypernodes based on the hyperedge weights, and combine the weighted fusion of associated hypernode features with its own features to update the hypernode representation. Ultimately, this captures the deep correlation between temporal load fluctuations and distributed power supply spatial layout, outputting an M×256-dimensional (M is the number of hypernodes) hypergraph feature matrix.

[0044] 4) Input the hypergraph feature matrix into the probability prediction output layer and perform Bayesian inference and adapt it to the risk threshold constraint of the islanding scenario to obtain the load probability distribution under the islanding operation mode within the future preset time window of the microgrid.

[0045] The probabilistic prediction output layer uses a Bayesian fully connected network, and the input data is a hypergraph feature matrix (multimodal features that fuse temporal and spatial correlation information).

[0046] Specifically, the network weights are probabilistically modeled using a Bayesian inference mechanism, replacing the deterministic weights of conventional fully connected layers. During the forward propagation of the network, multiple weight combinations are generated through Monte Carlo sampling (e.g., 10 samplings), and each combination is used to calculate the input hypergraph feature matrix. This comprehensively captures the sources of uncertainty in the model prediction (such as data noise and feature coupling bias).

[0047] Furthermore, the reliability requirements of islanded power supply are transformed into risk thresholds (such as the upper limit of allowable load forecasting error and the probability threshold of over-predicted load) and embedded in the output calculation process. By comparing the fit between different Monte Carlo sampling results and risk thresholds, the output load probability distribution is adjusted to ensure that the probability of over-predicted load (extreme loads that exceed the normal forecast range) is controlled within an acceptable range (such as the probability of the load exceeding the 95th percentile ≤ 5%).

[0048] Finally, the load probability distribution of the microgrid within a future preset time window (e.g., 24h, 15min / step) is output. This distribution includes the load mean and variance for each time step. At the same time, the values ​​and probability thresholds corresponding to over-predictable load (over-predictable load refers to the extreme load value that exceeds the upper limit of the conventional prediction interval (e.g., 95% confidence interval) of the load probability distribution within the future preset time window) and under-predictable load (under-predictable load refers to the load value that is lower than the lower limit of the conventional prediction interval (e.g., 95% confidence interval) of the load probability distribution within the future preset time window are output, providing uncertainty quantification basis for subsequent multi-objective optimization scheduling.

[0049] S3. Construct a multi-objective optimization function with the objectives of minimizing energy storage lifetime loss and maximizing carbon emission reduction. Solve the multi-objective optimization function based on the load probability distribution and predefined uncertainty robust constraints and scenario elastic constraints to generate the optimal output command of the distributed power source in the microgrid and the optimal charging and discharging command of the energy storage system in the microgrid. like Figure 3 As shown, this is a flowchart illustrating step S3. (Refer to...) Figure 3 Step S3 includes: S301. With the goals of maximizing power supply reliability, minimizing operating costs, maximizing voltage and frequency stability, minimizing energy storage lifespan loss, and maximizing carbon emission reduction, a multi-objective optimization function is constructed, using the output of distributed power sources in the microgrid and the charging and discharging power of the energy storage system as optimization variables. When constructing the multi-objective optimization function, a scenario-adaptive dynamic weighting mechanism is first introduced. This mechanism identifies isolated operating conditions (extreme weather, low SOC of energy storage, and normal operating conditions) in real time and assigns target weights according to preset rules (extreme weather: , , , , Low SOC for energy storage: , , , , Normal operating conditions: , , , , ,in The weights for power supply reliability, operating cost, voltage frequency stability, energy storage life loss, and carbon emission reduction are respectively, and the sum is 1).

[0050] Then, quantitative modeling is performed on each objective: Power supply reliability A tiered quantification method is adopted, consisting of "critical load satisfaction rate × 0.8 + non-critical load satisfaction rate × 0.2", with the formula as follows: in, Indicates the first The actual power supply of time-step critical loads (such as hospital ICUs and emergency communication loads). Indicates the first Power demand of critical loads at time step Indicates the first The actual power supplied to non-critical loads (such as general lighting and non-essential industrial loads) at the time step. Indicates the first Power demand of non-critical loads at time step This indicates the total number of time steps within the preset time window. This represents the time step. The formula quantifies the degree of load grading satisfaction through weighted summation, with the weight allocation reflecting the principle of prioritizing power supply to critical loads. The value range is [0,1], and the closer it is to 1, the higher the power supply reliability.

[0051] Operating costs The formula encompasses the operation and maintenance costs of distributed power sources, energy storage loss costs, and load shortage penalty costs, and is as follows: in, This represents the total number of distributed generation (DG) sources. Indicates the first The unit operation and maintenance cost of distributed power sources (such as photovoltaic and wind power), Indicates the first Distributed power generation The output of the time step, This indicates the unit charge / discharge loss cost of an energy storage system. Indicates the first energy storage system The absolute value of the charging and discharging power at each time step (charging is negative and discharging is positive; the absolute value is used to calculate the loss cost uniformly). This represents the penalty cost per unit load due to power shortage. Indicates the first Total load demand power at time step Indicates the first The total actual power supply at each time step (the difference between the sum of the output of distributed power sources and the discharge power of energy storage and the charging power of energy storage). The formula ensures that power outage penalties are calculated only when power supply is insufficient, thus comprehensively covering the economic costs of microgrid islanding operations.

[0052] Voltage frequency stability The formula, obtained by weighting and quantizing the absolute values ​​of frequency deviation and voltage deviation, is as follows: in, This represents the frequency deviation weighting coefficient (set according to power quality priority). Indicates the first The actual frequency of the time-step system Indicates the rated frequency. This represents the voltage deviation weighting coefficient. Indicates the total number of microgrid nodes. Indicates the first The node The actual voltage at the time step Indicates the rated voltage. Indicates the first The relative voltage deviation (dimensionless) of each node is taken as the average value of all nodes to characterize the overall voltage stability of the system. The smaller the value, the closer the system voltage frequency is to the rated value, and the higher the stability.

[0053] Energy storage lifespan loss Based on the "charge / discharge depth-cycle count" coupled model, the formula is: in, Indicates the first Depth of charge and discharge of time-step energy storage systems Indicates the first The time step represents the energy storage charge / discharge power, with discharge being positive and charging being negative. Coefficients 0.02 and 0.005 are weighted coefficients obtained from lithium battery cycle life tests, representing the coupled contributions of charge / discharge depth and charge / discharge power to lifespan loss, respectively. This formula accurately quantifies the cumulative effect of lifespan loss of energy storage systems under islanded dynamic conditions by coupling the two influencing factors of charge / discharge depth and charge / discharge power, avoiding the limitations of single-index evaluation.

[0054] carbon emission reduction Combining the power output ratio of wind and solar power, the formula is: in, Indicates the first Time-step photovoltaic power output Indicates the first Time step wind power output The time step is indicated, and 0.85 represents the emission reduction coefficient per unit of wind and solar power generation.

[0055] Finally, they are integrated to form a comprehensive multi-objective optimization function. As shown below: The optimization variables are the output of each distributed power source and the charging and discharging power of the energy storage system.

[0056] S302. Based on the load probability distribution, the particle swarm optimization algorithm is used to solve the multi-objective optimization function according to the predefined uncertainty robust constraint conditions and scenario elastic constraint conditions, so as to generate the optimal output command of the distributed power source in the microgrid and the optimal charging and discharging command of the energy storage system in the microgrid.

[0057] Before solving, first consider the load probability distribution (including the mean). With fluctuation value To quantify uncertainty, a weighted summation method is used to transform the multi-objective optimization function into a single-objective function, and then the particle swarm optimization algorithm is started to solve it: the population size is set to 50, the number of iterations is 100, the particle dimension corresponds to the total number of optimization variables (output of each distributed power source and energy storage charging and discharging power), and the particle position boundary is set according to the rated operating range of the equipment.

[0058] During the iteration process, uncertainty robust constraints and scenario elastic constraints are embedded into the fitness function calculation. Specifically, the uncertainty robust constraints include islanded power balance constraints and frequency voltage stability constraints, while the scenario elastic constraints include distributed power output constraints, energy storage charge / discharge rate-lifetime coupling constraints, and line power flow constraints.

[0059] The islanded power balance constraint adopts a robust form, as shown below: in, This represents the total output of the distributed power source. This represents the maximum output fluctuation value of the distributed power source. Indicates the discharge power of the energy storage system. This indicates the charging power of the energy storage system.

[0060] Frequency and voltage stability constraints employ tiered robust ranges. Specifically, for critical loads: frequency 50±0.1Hz, voltage ±3% of rated value; for non-critical loads: frequency 50±0.3Hz, voltage ±6% of rated value. In other words, critical loads (such as hospital ICUs and emergency communication loads) are highly sensitive to power quality, and their operation is strictly limited to a frequency range of 50±0.1Hz and a voltage range of ±3% of rated value to prevent voltage and frequency fluctuations from affecting or damaging the equipment. Non-critical loads (such as general lighting and non-essential industrial loads) have a higher tolerance for power quality, and a relatively relaxed constraint range (frequency 50±0.3Hz, voltage ±6% of rated value) is adopted to reduce the increase in operating costs caused by excessive constraints while ensuring power supply reliability.

[0061] In the scenario-based elastic constraints, the output constraints of distributed power sources are adjusted according to the state of health (SOH) of the equipment (e.g., when the SOH of a wind turbine is less than 80%, the upper limit of output is reduced to 80% of the rated value); the energy storage charge / discharge rate-lifetime coupling constraint is set as "charging rate ≤ 0.5C when SOC is less than 30%, and discharging rate ≤ 0.3C when SOC is greater than 80%", where C represents the charge / discharge rate of the energy storage system, which is the ratio of the actual charge / discharge power of the energy storage system to the rated capacity. Its physical meaning is the time required for the energy storage system to fully charge or discharge its rated capacity when charging and discharging at the current power; the line power flow constraint takes into account the line temperature (when the temperature is greater than 60℃, the upper limit of power flow is reduced to 80% of the rated capacity) to avoid line overheating and tripping.

[0062] Each iteration calculates the particle fitness value, optimizing the particle trajectory by updating the global and individual optimal positions, ensuring a solution time of ≤10s to meet real-time requirements; if the iteration result violates rigid constraints, i.e., uncertainty robustness constraints, then triggering... - Constraint method ( (Set to 0.1Hz) and solve again, finally outputting the optimal output command for each distributed power source and the optimal charging and discharging command for the energy storage system.

[0063] S4. Send the optimal output command and the optimal charge / discharge command to the microgrid control system so that the microgrid control system can adjust the islanded operation status of the microgrid in real time according to the optimal output command and the optimal charge / discharge command.

[0064] The generated optimal output command of the distributed power source and the optimal charge and discharge command of the energy storage system are encapsulated into GOOSE messages according to the IEC61850 standard protocol and sent to the corresponding device controllers of the microgrid control system via industrial Ethernet or power-specific communication protocols.

[0065] Specifically, it transmits output adjustment commands to the distributed power controller to precisely regulate the output power of photovoltaic inverters and wind turbine converters, ensuring that the distributed power supply outputs stably according to the optimization target; it issues charging and discharging commands and SOC control targets to the energy storage system controller to coordinate the energy storage charging and discharging rate and state of charge, maintaining islanded power balance; and it transmits load regulation commands to the controllable load management device to orderly reduce or transfer non-critical loads when necessary, ensuring continuous power supply to critical loads.

[0066] Furthermore, after receiving the command, the microgrid control system compares the actual operating parameters of each device with the command value in real time, and dynamically adjusts the operating status of the device through closed-loop control logic. If the output deviation of the device is detected to exceed 5% or the voltage and frequency fluctuate slightly, a secondary fine-tuning is immediately triggered to ensure that the islanded operation status of the microgrid always conforms to the optimization target, and maintains power balance, voltage and frequency stability and multi-objective collaborative optimization effect.

[0067] likeFigure 4 The diagram shown is another flowchart illustrating a microgrid islanding operation control method according to an embodiment of the present invention. (Refer to...) Figure 4 According to an embodiment of the present invention, a microgrid islanding operation control method further includes the following step after step S4: S5. Monitor the operating status data of each node in the microgrid in real time, and dynamically correct the deep neural network model and multi-objective optimization function based on the deviation analysis results of the operating status data.

[0068] Specifically, step S5 includes: 1) Monitor the operating status data of each node in the microgrid in real time, and obtain the operating status deviation value by comparing the operating status data with the preset benchmark value; The system collects real-time data on voltage, frequency, and power flow at key nodes within the microgrid, as well as operational data such as actual output of distributed power sources, charging and discharging power of energy storage systems, state of charge, and health status. The collected operational data is then compared step-by-step with preset benchmark values. These benchmark values ​​include load prediction results from a deep neural network model and optimal command results obtained by solving a multi-objective optimization function.

[0069] Specifically, the absolute and relative deviations of load forecasting are calculated by comparing the load forecast results (including the mean of the load probability distribution and the confidence interval) output by the deep neural network model; and the command execution deviation is calculated by comparing the results with the optimal command results (optimal output command of distributed power source and optimal charging and discharging command of energy storage system) obtained by solving the multi-objective optimization function. The two types of deviation results are then combined to form an operating state deviation value that encompasses both load forecasting deviation and command execution deviation.

[0070] 2) Based on the deviation value of the running status, the parameters of the deep neural network model are corrected in real time, and the multi-objective optimization function is optimized in a rolling manner according to the running status data.

[0071] If the relative deviation of load prediction (MAPE) is greater than 8%, the online fine-tuning mechanism of the deep neural network model is triggered. Specifically, the weights of the reversible residual convolutional blocks in the temporal feature extraction layer are focused, and the parameters are updated using gradient descent (with the learning rate set to 1 / 10 of the initial training rate). At the same time, the core parameters of the spatial feature learning layer and the hypergraph attention fusion layer are fixed to avoid over-correction that would lead to a decrease in the model's generalization ability.

[0072] Every 15 minutes, the real-time monitored operating status data (actual load curve, equipment SOH, line temperature, and real-time meteorological data) is fed back to the multi-objective optimization function. The current operating conditions are re-identified (such as changes in energy storage SOC and the persistence of extreme weather). The scenario adaptive weights are dynamically adjusted (e.g., when the energy storage SOC drops below 30%, the weight of the energy storage life loss target is increased). The uncertainty robust constraint conditions and scenario elastic constraint conditions are updated (e.g., when the line temperature rises to 60℃, the upper limit of the power flow constraint is lowered). Then, the particle swarm optimization algorithm is used to solve the problem again and generate new optimization instructions adapted to the real-time operating conditions.

[0073] It should be noted that if emergency scenarios such as islanding frequency exceeding the limit (greater than 50.2Hz) or voltage collapse (less than 0.9pu) are detected, the rapid correction logic is immediately triggered. Temporary instructions such as emergency load reduction and emergency charging and discharging of energy storage are used to stabilize the system state. Parameter correction and rolling optimization are then performed after the system returns to steady state.

[0074] This invention discloses a microgrid islanding operation control method. A four-layer deep neural network architecture, through adaptive fusion of temporal and spatial features, accurately captures the long- and short-term fluctuation patterns of islanded loads. The output load probability distribution can quantify and predict uncertainties, providing a scientific basis for risk-averse scheduling. Compared to conventional models, the long-term prediction accuracy is significantly improved, and its adaptability to extreme scenarios is superior. The multi-objective optimization function balances the requirements of minimizing energy storage lifespan loss and maximizing carbon emission reduction. Combining uncertainty robustness constraints and scenario elastic constraints, it can dynamically adapt to different operating conditions such as extreme weather and low energy storage charge states. While ensuring islanded power balance and voltage / frequency stability, it significantly extends the lifespan of the energy storage system, improves the absorption rate of clean energy such as wind and solar power, and reduces operating costs and carbon emission intensity. By accurately generating optimal output commands for distributed power sources and optimal charging and discharging commands for the energy storage system, it achieves refined scheduling of islanded energy resources. After the commands are issued, the control system can respond and adjust rapidly, effectively avoiding risks such as power shortages and voltage collapses, reducing the probability of malfunctions of protection devices, and comprehensively improving the safety, stability, and control reliability of microgrid islanding operation.

[0075] like Figure 5 The diagram shown is a structural schematic of a microgrid islanding operation control system according to an embodiment of the present invention. (Refer to...) Figure 5 An embodiment of the present invention provides a microgrid islanding operation control system, comprising: Data acquisition module 01 is used to acquire multi-source heterogeneous data in the islanded operation mode of microgrids to obtain the islanded operation dataset; The load forecasting module 02 is used to predict the load change trend of the microgrid under the islanded operation mode within a preset time window based on the islanded operation dataset. It uses a pre-built deep neural network model that includes a time-series feature extraction layer, a spatial feature learning layer, a hypergraph attention fusion layer, and a probability prediction output layer to obtain the load probability distribution of the microgrid. The instruction generation module 03 is used to construct a multi-objective optimization function with at least the goal of minimizing energy storage lifetime loss and maximizing carbon emission reduction, and solve the multi-objective optimization function based on the load probability distribution according to predefined uncertainty robust constraints and scenario elastic constraints to generate the optimal output command of the distributed power source in the microgrid and the optimal charging and discharging command of the energy storage system in the microgrid. The execution module 04 is used to send the optimal output command and the optimal charge / discharge command to the microgrid control system, so that the microgrid control system can adjust the islanded operation status of the microgrid in real time according to the optimal output command and the optimal charge / discharge command.

[0076] It should be noted that the various modules in the aforementioned microgrid islanding operation control system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module. For specific limitations regarding the microgrid islanding operation control system, please refer to the limitations regarding the microgrid islanding operation control method above; both have the same function and role, and will not be repeated here.

[0077] In summary, the present invention provides a microgrid islanding operation control method and system. A four-layer deep neural network architecture, through adaptive fusion of temporal and spatial features, accurately captures the long- and short-term fluctuation patterns of islanded loads. The output load probability distribution can quantify and predict uncertainties, providing a scientific basis for risk-averse scheduling. Compared to conventional models, the long-term prediction accuracy is significantly improved, and the adaptability to extreme scenarios is superior. The multi-objective optimization function balances the requirements of minimizing energy storage lifespan loss and maximizing carbon emission reduction. Combined with uncertainty robustness constraints and scenario elastic constraints, it can dynamically adapt to different operating conditions such as extreme weather and low energy storage charge states. While ensuring islanded power balance and voltage / frequency stability, it significantly extends the lifespan of the energy storage system, improves the absorption rate of clean energy such as wind and solar power, and reduces operating costs and carbon emission intensity. By accurately generating optimal output commands for distributed power sources and optimal charging and discharging commands for the energy storage system, it achieves refined scheduling of islanded energy resources. After the commands are issued, the control system can respond and adjust rapidly, effectively avoiding risks such as power shortages and voltage collapses, reducing the probability of malfunctions of protection devices, and comprehensively improving the safety, stability, and control reliability of microgrid islanding operation.

[0078] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0079] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A microgrid islanded operation control method, characterized in that, include: Obtain multi-source heterogeneous data under the islanded operation mode of the microgrid to obtain the islanded operation dataset; Based on the islanded operation dataset, a pre-constructed deep neural network model containing a temporal feature extraction layer, a spatial feature learning layer, a hypergraph attention fusion layer, and a probability prediction output layer is used to predict the load change trend of the microgrid under the islanded operation mode within a future preset time window, thereby obtaining the load probability distribution of the microgrid. A multi-objective optimization function is constructed with the objectives of minimizing energy storage lifetime loss and maximizing carbon emission reduction. The multi-objective optimization function is then solved based on the load probability distribution under predefined uncertainty robust constraints and scenario elastic constraints to generate the optimal output command of the distributed power source and the optimal charging and discharging command of the energy storage system within the microgrid. The optimal output command and the optimal charge / discharge command are sent to the microgrid control system so that the microgrid control system can adjust the islanded operation status of the microgrid in real time according to the optimal output command and the optimal charge / discharge command.

2. The microgrid islanding operation control method according to claim 1, characterized in that, After issuing the optimal output command and the optimal charge / discharge command to the microgrid control system so that the microgrid control system can adjust the islanded operation state of the microgrid in real time according to the optimal output command and the optimal charge / discharge command, the method further includes: The system monitors the operational status data of each node within the microgrid in real time, and dynamically corrects the deep neural network model and the multi-objective optimization function based on the deviation analysis results of the operational status data.

3. The microgrid islanding operation control method according to claim 1, characterized in that, The acquisition of multi-source heterogeneous data under the microgrid islanding operation mode to obtain the islanding operation dataset includes: Collect node electrical parameters, source-storage operation data, regional environmental data, and historical source-load operation data under the islanded operation mode of the microgrid to obtain multi-source heterogeneous data; The preprocessed multi-source heterogeneous data is injected into the island spatiotemporal features and source-storage operation response features to obtain the island operation dataset.

4. The microgrid islanding operation control method according to claim 1, characterized in that, Based on the islanded operation dataset, a pre-constructed deep neural network model, including a temporal feature extraction layer, a spatial feature learning layer, a hypergraph attention fusion layer, and a probability prediction output layer, is used to predict the load change trend of the microgrid under the islanded operation mode within a preset time window in the future, obtaining the load probability distribution of the microgrid, including: The isolated dataset is divided to obtain an initial training set, and a generative adversarial network is introduced to augment the initial training set to obtain the target training set. Based on the target training set, the deep neural network model is trained using the power task joint loss function to obtain the load prediction model; The isolated operation dataset is input into the load prediction model, and the data is processed sequentially through the time series feature extraction layer, the spatial feature learning layer, the hypergraph attention fusion layer, and the probability prediction output layer to obtain the load probability distribution under the isolated operation mode within a preset time window of the microgrid.

5. The microgrid islanding operation control method according to claim 4, characterized in that, The joint loss function for power tasks includes mean error loss and negative log-likelihood loss.

6. The microgrid islanding operation control method according to claim 4, characterized in that, The process of inputting the islanded operation dataset into the load prediction model and sequentially processing the data through the time-series feature extraction layer, the spatial feature learning layer, the hypergraph attention fusion layer, and the probability prediction output layer to obtain the load probability distribution under the islanded operation mode within a future preset time window of the microgrid includes: The isolated operation dataset is input into the load prediction model, and the isolated operation dataset is subjected to time series feature extraction and cross-domain fusion processing through the time series feature extraction layer to obtain a cross-domain time series feature matrix. The cross-domain temporal feature matrix is ​​input into the spatial feature learning layer. The spatial feature learning layer performs heterogeneous graph spatial feature learning and cross-type node association aggregation on the cross-domain temporal feature matrix to obtain the node spatial feature matrix. The cross-domain temporal feature matrix and the node spatial feature matrix are input into the hypergraph attention fusion layer to perform dynamic hypergraph construction and temporal-spatial feature attention fusion to obtain the hypergraph feature matrix; The hypergraph feature matrix is ​​input into the probability prediction output layer for Bayesian inference and adaptation to the risk threshold constraint of the islanding scenario, so as to obtain the load probability distribution under the islanding operation mode within the future preset time window of the microgrid.

7. The microgrid islanding operation control method according to claim 1, characterized in that, The process involves constructing a multi-objective optimization function with the objectives of minimizing energy storage lifetime loss and maximizing carbon emission reduction. Based on the load probability distribution, the multi-objective optimization function is solved according to predefined uncertainty robustness constraints and scenario elasticity constraints to generate the optimal output command for distributed power sources within the microgrid and the optimal charging and discharging command for the energy storage system within the microgrid. This includes: With the goals of maximizing power supply reliability, minimizing operating costs, maximizing voltage and frequency stability, minimizing energy storage lifespan loss, and maximizing carbon emission reduction, a multi-objective optimization function is constructed, using the output of distributed power sources and the charging and discharging power of energy storage systems within the microgrid as optimization variables. Based on the load probability distribution, the particle swarm optimization algorithm is used to solve the multi-objective optimization function under predefined uncertainty robust constraints and scenario elastic constraints to generate the optimal output command of the distributed power source and the optimal charging and discharging command of the energy storage system in the microgrid.

8. The microgrid islanding operation control method according to claim 7, characterized in that, The uncertainty robust constraints include islanded power balance constraints and frequency voltage stability constraints, while the scenario elastic constraints include distributed power output constraints, energy storage charge / discharge rate-lifetime coupling constraints, and line power flow constraints.

9. The microgrid islanding operation control method according to claim 2, characterized in that, The real-time monitoring of the operating status data of each node within the microgrid, and the dynamic correction of the deep neural network model and the multi-objective optimization function based on the deviation analysis results of the operating status data, includes: The system monitors the operating status data of each node in the microgrid in real time, and obtains the operating status deviation value by comparing the operating status data with a preset benchmark value. The preset benchmark value includes the load prediction result output by the deep neural network model and the optimal command result obtained by solving the multi-objective optimization function. The parameters of the deep neural network model are corrected in real time based on the operational state deviation value, and the multi-objective optimization function is optimized in a rolling manner based on the operational state data.

10. A microgrid islanded operation control system, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous data in the islanded operation mode of the microgrid to obtain the islanded operation dataset; The load forecasting module is used to predict the load change trend of the microgrid under the islanded operation mode within a preset time window in the future, based on the islanded operation dataset, through a pre-constructed deep neural network model that includes a time-series feature extraction layer, a spatial feature learning layer, a hypergraph attention fusion layer, and a probability prediction output layer, and to obtain the load probability distribution of the microgrid. The instruction generation module is used to construct a multi-objective optimization function with at least the goal of minimizing energy storage lifetime loss and maximizing carbon emission reduction, and to solve the multi-objective optimization function based on the load probability distribution according to predefined uncertainty robust constraints and scenario elastic constraints, so as to generate the optimal output instruction of the distributed power source in the microgrid and the optimal charging and discharging instruction of the energy storage system in the microgrid. The execution module is used to send the optimal output command and the optimal charge / discharge command to the microgrid control system, so that the microgrid control system can adjust the islanded operation status of the microgrid in real time according to the optimal output command and the optimal charge / discharge command.

Images