Micro-energy network intelligent regulation and control method, system and device based on multi-mode perception and medium

Abstract

The invention relates to the technical field of power grid intelligent regulation and control, and discloses a micro-energy grid intelligent regulation and control method, system, device and medium based on multi-mode perception, and the method comprises the steps: introducing a dynamic graph mechanism, and achieving the dynamic perception of the access state of distributed energy equipment and the change of a communication link, the problem that a static micro-energy network topological structure cannot adapt to environmental changes is solved. A multi-modal space-time attention mechanism is utilized, meteorological images, historical micro-energy network supply and demand data and distributed energy equipment operation states are fused, the prediction precision of photovoltaic power output and load demands is effectively improved, unified modeling and collaborative optimization of discrete and continuous control actions are achieved based on a hierarchical reinforcement learning framework, and the optimal control effect is achieved. The control response speed of the micro-energy network is improved, efficient structure sensing, accurate power prediction and intelligent control of the micro-energy network in a dynamic complex operation environment are achieved, and the overall operation efficiency, control flexibility and overall stability of the micro-energy network are improved.

Description

Technical Field

[0001] This invention relates to the field of smart grid control technology, and in particular to a method, system, device and medium for smart control of microgrids based on multimodal sensing. Background Technology

[0002] With the continuous development of distributed energy technology, various types of distributed energy sources, such as photovoltaic power generation, energy storage, and controllable loads, are increasingly being connected to microgrids, bringing richer energy supply options to microgrids. However, the rapid increase in the number and diversification of distributed energy devices has led to an increasingly complex operating environment for microgrids. At the same time, frequent device connection and disconnection, uncertainties in communication links, fluctuations in operating conditions, and meteorological factors place higher demands on the regulation capabilities of microgrids.

[0003] Existing microgrid regulation methods suffer from several key technical limitations. First, existing methods generally rely on static topology models, using fixed adjacency matrices to describe the connections between distributed energy devices. This lack of adaptive topology update mechanisms to handle dynamic access or communication interruptions by distributed energy devices makes it difficult for microgrid regulation to dynamically adjust to changes in the operating environment, severely restricting its flexibility. Second, the photovoltaic power output and load demand forecasting technologies used in existing methods mostly rely on single historical power data or load sequences, failing to effectively integrate multi-source heterogeneous information such as meteorological images and microgrid device operating status. This results in significantly larger prediction errors when facing sudden climate changes, abnormal microgrid device operation, or non-stationary states, making it difficult to adapt to complex and changing operating environments. Finally, existing methods often model and optimize discrete and continuous control separately, lacking a collaborative control mechanism under a unified strategy space. This makes it difficult to achieve global optimum control decisions, leading to delayed response, low control efficiency, and insufficient stability in microgrid regulation. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method and system for intelligent control of microgrids based on multimodal sensing to solve the problems of insufficient structural self-adaptation capability, lack of multi-source information fusion capability, and poor coordination of control strategies in current microgrids.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for intelligent control of microgrids based on multimodal sensing, comprising: acquiring and preprocessing to generate a multi-source dynamic operation dataset;

[0008] A microgrid intelligent control model based on multimodal perception is constructed. A portion of the data from the running dataset is input into the spatiotemporal feature extraction module in the model to generate a microgrid node fusion feature sequence X and a microgrid dynamic adjacency matrix A.

[0009] The fused feature sequence X, the microgrid dynamic adjacency matrix A, and the power data in the running dataset are input into the attention feature extraction module in the model to generate the microgrid power supply and demand prediction value T.

[0010] The current distributed energy equipment operating status dataset, the fusion feature sequence X, and the power prediction value T are input into the HHACRL-CM module in the model to generate the microgrid hybrid strategy feature sequence E.

[0011] The feature sequence of the microgrid hybrid strategy is input into the decision generation module in the model, and the microgrid regulation decision is output to achieve precise control and global optimization of the operating status of various devices in the microgrid.

[0012] As a preferred embodiment of the intelligent control method for microgrids based on multimodal sensing described in this invention, the multi-source dynamic operation dataset includes a microgrid environmental parameter dataset, a distributed energy equipment operation status dataset, a microgrid communication link status dataset, and a microgrid historical power supply and demand dataset.

[0013] As a preferred embodiment of the intelligent control method for microgrids based on multimodal sensing described in this invention, the method includes: inputting a portion of the running dataset into the spatiotemporal feature extraction module in the model to generate a microgrid node fusion feature sequence X and a microgrid dynamic adjacency matrix A, including:

[0014] Input the microgrid environmental parameter dataset, the distributed energy equipment operation status dataset, and the microgrid communication link status dataset into the three branches of the spatiotemporal feature extraction module, respectively.

[0015] The features output from the three branches are concatenated to generate the fusion feature sequence X14 of the fourteenth microgrid node.

[0016] X14 is input into the Transformer encoder to generate the 15th microgrid node fusion feature sequence X15; X15 is input into the cosine similarity calculation submodule to generate the 16th microgrid node fusion feature sequence X16; X16 is input into the Softmax normalization submodule to generate the 17th microgrid node fusion feature sequence X17; X17 is subjected to Top-K filtering operation and then input into the mask processing submodule to finally generate the microgrid dynamic adjacency matrix A.

[0017] Simultaneously, the fusion feature sequence X15 of the fifteenth microgrid node is processed through a fully connected layer to generate the fusion feature sequence X18 of the eighteenth microgrid node; X18 is input into a batch normalization layer to generate the fusion feature sequence X19 of the nineteenth microgrid node; X19 is subjected to a random deactivation operation to finally generate the fusion feature sequence X of the microgrid node.

[0018] As a preferred embodiment of the intelligent control method for microgrids based on multimodal perception described in this invention, the method further includes: inputting the microgrid environmental parameter dataset into the first branch of the spatiotemporal feature extraction module, processing it through a fully connected layer to complete the initial feature transformation, then entering the fully connected layer again through the GELU activation function for further mapping, and finally performing feature normalization through the layer normalization submodule to generate the first microgrid node fusion feature sequence X1.

[0019] The first microgrid node fusion feature sequence X1 is input into three convolutional layers for fusion, and the corresponding three fusion feature sequences are output respectively. Then, the sequences are spliced ​​together to output the fifth microgrid node fusion feature sequence X5.

[0020] It also includes inputting the distributed energy equipment operation status dataset into the second branch of the spatiotemporal feature extraction module, extracting local perception features through the convolutional layer, and then generating the sixth microgrid node fusion feature sequence X6 through the ReLU activation function;

[0021] The fusion feature sequence X6 of the sixth microgrid node is processed by a multi-head self-attention mechanism, random deactivation dropout operation, and then processed by a fully connected layer to complete the initial feature transformation. It is then further mapped by the GELU activation function and finally normalized by the layer normalization submodule to output the fusion feature sequence X9 of the ninth microgrid node.

[0022] It also includes inputting the microgrid communication link state dataset into the third branch of the spatiotemporal feature extraction module, enhancing the location features through the location encoding submodule, and then processing it through a fully connected layer to generate the 10th microgrid node fusion feature sequence X10; inputting X10 into a graph convolutional network to generate the 11th microgrid node fusion feature sequence X11; concatenating X10 and X11 to generate the 12th microgrid node fusion feature sequence X12; and passing X12 through a layer normalization submodule to output the 13th microgrid node fusion feature sequence X13.

[0023] As a preferred embodiment of the intelligent control method for microgrids based on multimodal sensing described in this invention, the method involves inputting the fused feature sequence X, the microgrid dynamic adjacency matrix A, and power data from the running dataset into the attention feature extraction module of the model to generate predicted power supply and demand values ​​for the microgrid, including:

[0024] The fused feature sequence X, the microgrid dynamic adjacency matrix A, and the power data are respectively input into the three branches of the attention feature extraction module;

[0025] The final power characteristic sequences output from the three branches are spliced ​​together to generate the fourteenth microgrid supply and demand power characteristic sequence T14.

[0026] T14 is input into the Transformer encoder for global time series modeling to generate the power supply and demand characteristic sequence T15 of the fifteenth microgrid.

[0027] T15 is input into the fully connected layer for deep nonlinear transformation, and then the ReLU activation function is used to generate the power supply and demand characteristic sequence T16 of the sixteenth microgrid.

[0028] T16 is input into a random deactivation dropout layer for feature suppression processing to generate the power supply and demand feature sequence T17 of the seventeenth microgrid.

[0029] T17 is input into the fully connected layer for feature mapping to generate the power supply and demand feature sequence T18 of the 18th microgrid.

[0030] Input T18 into the prediction head to generate the predicted power supply and demand value T for the microgrid.

[0031] As a preferred embodiment of the intelligent control method for microgrids based on multimodal sensing described in this invention, the following steps are taken: the current distributed energy equipment operating status dataset, the fused feature sequence X, and the power prediction value T are input into the HHACRL-CM module in the model to generate a microgrid hybrid strategy feature sequence E, including:

[0032] Input the current distributed energy equipment operating status dataset, the fused feature sequence X, and the power prediction value T into the three branches of the HHACRL-CM module, respectively;

[0033] The fourth microgrid hybrid strategy feature sequence E4 output from the second branch in the HHACRL-CM module and the fifth microgrid hybrid strategy feature sequence E5 initially output from the third branch are spliced ​​together to generate the sixth microgrid hybrid strategy feature sequence E6.

[0034] E6 is processed through a layer normalization submodule and a multi-head self-attention mechanism to perform deep global feature modeling, generating the eighth microgrid hybrid strategy feature sequence E8.

[0035] The third microgrid hybrid strategy feature sequence E3 and E8, output from the first branch of the HHACRL-CM module, are concatenated and then mapped and normalized through a fully connected layer to generate the eleventh microgrid hybrid strategy feature sequence E11.

[0036] E11 is simultaneously input into two fully connected layers, and the first and second processing are performed respectively to output the fifteenth microgrid hybrid strategy feature sequence E15 and the sixteenth microgrid hybrid strategy feature sequence E16. These are then spliced ​​together to obtain the seventeenth microgrid hybrid strategy feature sequence E17.

[0037] E17 is sequentially input into a fully connected layer and a hyperbolic tangent activation function for nonlinear transformation to generate the feature sequence E19 of the nineteenth microgrid hybrid strategy.

[0038] By concatenating E15 and E19, the final microgrid hybrid strategy feature sequence E is generated.

[0039] As a preferred embodiment of the intelligent control method for microgrids based on multimodal sensing described in this invention, the following steps are taken: the microgrid hybrid strategy feature sequence is input into the decision generation module in the model, and the microgrid control decision is output, wherein:

[0040] The microgrid control decision is a vector of command sequences arranged in chronological order:

[0041] I = [I1, I2, ..., I t ,…,I N ]

[0042] Where N represents the total number of time steps within the future time window; I t This represents the device control parameter vector at time step t, which is the control command for all critical device units in the microgrid at time step t.

[0043] I t =[p1,p2,…,p j ,…,p m ]

[0044] Where m represents the number of microgrid devices; p j This represents the specific control parameters of the j-th device at time step t.

[0045] Secondly, the present invention provides a microgrid intelligent control system based on multimodal sensing, comprising:

[0046] The acquisition module is used to acquire and preprocess the generated multi-source dynamic running dataset;

[0047] The first generation module is used to construct a microgrid intelligent control model based on multimodal perception. It inputs a portion of the data from the running dataset into the spatiotemporal feature extraction module in the model to generate a microgrid node fusion feature sequence X and a microgrid dynamic adjacency matrix A.

[0048] The second generation module is used to input the fused feature sequence t, the microgrid dynamic adjacency matrix A, and the power data in the running dataset into the attention feature extraction module in the model to generate the microgrid power supply and demand prediction value T.

[0049] The third generation module is used to input the current distributed energy equipment operating status dataset, the fusion feature sequence X, and the power prediction value T into the HHACRL-CM module in the model to generate the microgrid hybrid strategy feature sequence E.

[0050] The strategy output module is used to input the microgrid hybrid strategy feature sequence into the decision generation module in the model, and output the microgrid regulation decision to achieve precise control and global optimization of the operating status of various devices in the microgrid.

[0051] Thirdly, the present invention provides a computer device, comprising:

[0052] Memory and processor;

[0053] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of a microgrid intelligent control method based on multimodal sensing.

[0054] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the aforementioned intelligent control method for microgrids based on multimodal sensing.

[0055] Compared with existing technologies, the beneficial effects of this invention are as follows: By introducing a dynamic graph mechanism, this invention achieves dynamic perception of the access status of distributed energy devices and changes in communication links, solving the problem that static microgrid topologies cannot adapt to environmental changes; by utilizing a multimodal spatiotemporal attention mechanism, it integrates meteorological images, historical microgrid supply and demand data, and the operating status of distributed energy devices, effectively improving the prediction accuracy of photovoltaic power output and load demand, overcoming the problem that traditional microgrid control methods, which rely on single historical load or meteorological data, suffer from insufficient response to nonlinear fluctuation factors and accumulated prediction biases; and based on a hierarchical reinforcement learning framework, it achieves unified modeling and collaborative optimization of discrete and continuous control actions, improving the microgrid control response speed, realizing efficient structural perception, accurate power prediction, and intelligent control of microgrids in dynamic and complex operating environments, and improving the overall operating efficiency, control flexibility, and overall stability of microgrids. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a schematic diagram of the overall process of a microgrid intelligent control method based on multimodal sensing, according to an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of a multimodal sensing-based intelligent control model for microgrids, as described in an embodiment of the present invention.

[0059] Figure 3 This is a schematic diagram of the overall structure of the PDTM module in a microgrid intelligent control method based on multimodal sensing, as described in an embodiment of the present invention.

[0060] Figure 4 This is a schematic diagram of the overall structure of the MSAF module in a microgrid intelligent control method based on multimodal sensing, as described in one embodiment of the present invention.

[0061] Figure 5 This is a schematic diagram of the overall structure of the HHACRL-CM module in a microgrid intelligent control method based on multimodal sensing, as described in an embodiment of the present invention. Detailed Implementation

[0062] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0063] Example 1, referring to Figure 1 As an embodiment of the present invention, a smart control method for microgrids based on multimodal sensing is provided, comprising:

[0064] S100: Acquire and preprocess to generate a multi-source dynamic running dataset;

[0065] S200: Construct a microgrid intelligent control model based on multimodal perception, input a portion of the running dataset into the spatiotemporal feature extraction module in the model, and generate a microgrid node fusion feature sequence X and a microgrid dynamic adjacency matrix A;

[0066] S300: Input the fused feature sequence X, the microgrid dynamic adjacency matrix A, and the power data in the running dataset into the attention feature extraction module in the model to generate the microgrid power supply and demand prediction value T;

[0067] S400: Input the current distributed energy equipment operating status dataset, the fusion feature sequence X, and the power prediction value T into the HHACRL-CM module in the model to generate the microgrid hybrid strategy feature sequence E;

[0068] S500: Input the microgrid hybrid strategy feature sequence into the decision generation module in the model, output the microgrid regulation decision, and realize precise control and global optimization of the operating status of various devices in the microgrid.

[0069] It should be noted that existing technologies largely rely on static topology and fixed adjacency matrices, lacking real-time perception and adaptive updates for changes such as dynamic access of distributed energy devices and communication interruptions, resulting in insufficient flexibility in microgrid regulation. Furthermore, they often depend on single historical power or load sequences, failing to effectively integrate multi-source heterogeneous data such as meteorological images and equipment operating status. This leads to decreased prediction accuracy when microgrids face sudden climate changes and abnormal operation of distributed energy devices, making them ill-suited to complex environments. Moreover, discrete and continuous control are often modeled and optimized separately, lacking a unified collaborative control mechanism, making it difficult to achieve globally optimal control, resulting in delayed response and low efficiency in microgrid regulation.

[0070] To address the aforementioned main issues, steps S100-S500 utilize multimodal sensing to dynamically identify changes in the microgrid's topology and accurately predict photovoltaic power output and load demand. This provides data support and decision-making basis for subsequent rapid and stable microgrid regulation and response, thereby effectively enhancing the microgrid's regulation and control capabilities in complex environments.

[0071] Example 2, refer to Figures 1-5 As an embodiment of the present invention, based on the above embodiment, a microgrid intelligent control method based on multimodal sensing is provided.

[0072] In this embodiment of the application, step S100 involves acquiring and preprocessing to generate a multi-source dynamic operation dataset; wherein, the multi-source dynamic operation dataset includes a microgrid environmental parameter dataset, a distributed energy device operation status dataset, a microgrid communication link status dataset, and a microgrid historical supply and demand power dataset.

[0073] Specifically, first obtain the original microgrid environmental parameter dataset D. 1_1 This includes: temperature, humidity, and light intensity;

[0074] Original Distributed Energy Device Operation Status Dataset D 1_2This includes: voltage, current, and the state of charge of the energy storage device;

[0075] Original microgrid communication link state dataset D 1_3 This includes: signal latency, data packet loss rate, and link stability;

[0076] Original microgrid historical power supply and demand dataset D 1_4 This includes: photovoltaic power output and load demand;

[0077] In one alternative implementation, the preprocessing operation can be changed depending on the type of the original data;

[0078] For example, for the original microgrid environmental parameter dataset D 1_1 After smoothing, standardization, and outlier processing, the microgrid environmental parameter dataset D is obtained. 2_1 ; Targeting the original distributed energy device operating status dataset D 1_2 After time axis alignment, interpolation, and normalization, the distributed energy device operating status dataset D is obtained. 2_2-std ; Targeting the original microgrid communication link state dataset D 1_3 After missing value completion, normalization, and stability label construction, the microgrid communication link state dataset D is obtained. 2_3 ; Targeting the original microgrid historical power supply and demand dataset D 1_4 After time series completion, denoising, and normalization, the historical power supply and demand dataset D of the microgrid is obtained. 2_4 D 2_1 D 2_2-std D 2_3 and D 2_4 Merge the data to generate a preprocessed multi-source dynamic running dataset D2.

[0079] In this embodiment of the application, step S200 involves inputting a portion of the running dataset into the spatiotemporal feature extraction module in the model to generate a microgrid node fusion feature sequence X and a microgrid dynamic adjacency matrix A, including the following steps A1-A3:

[0080] A1: The microgrid environmental parameter dataset D 2_1 Distributed energy equipment operation status dataset D 2_2-std Microgrid communication link state dataset D 2_3 Input the three branches of the Spatiotemporal Feature Extraction Module (PDTM) respectively;

[0081] In the embodiments of this application, reference is made to Figure 2As shown, in step S200, A1 inputs the microgrid environmental parameter dataset, the distributed energy device operating status dataset, and the microgrid communication link status dataset into the three branches of the spatiotemporal feature extraction module, respectively, specifically including:

[0082] A1-1: The microgrid environmental parameter dataset D 2_1 The first branch of the input spatiotemporal feature extraction module is processed by the fully connected layer FC to complete the initial feature transformation. Then, it is further mapped by the GELU activation function and then enters the fully connected layer FC again. Finally, the feature is normalized by the layer normalization submodule LN to generate the first microgrid node fusion feature sequence X1.

[0083] The first microgrid node fusion feature sequence X1 is input into three convolutional layers for fusion, and the corresponding three fusion feature sequences are output respectively. Then, the sequences are spliced ​​together to output the fifth microgrid node fusion feature sequence X5.

[0084] For details, please refer to Figure 3 As shown, X1 can be input into a one-dimensional convolutional layer Conv1D with a kernel size of 3×1 for local feature extraction, then enhanced with a ReLU activation function, and then input into a batch normalization submodule BN for feature normalization to generate the second microgrid node fusion feature sequence X2; X1 can be input into a one-dimensional convolutional layer Conv1D with a kernel size of 5×1 for local feature extraction, then enhanced with a ReLU activation function, and then input into a batch normalization submodule BN for feature normalization to generate the third microgrid node fusion feature sequence X3; X1 can be input into a one-dimensional convolutional layer Conv1D with a kernel size of 7×1 for local feature extraction, then enhanced with a ReLU activation function, and then input into a batch normalization submodule BN for feature normalization to generate the fourth microgrid node fusion feature sequence X4; X2, X3 and X4 are concatenated using the concat operation to generate the fifth microgrid node fusion feature sequence X5.

[0085] A1-2: Data set D of the operating status of distributed energy devices 2_2-std The second branch of the input spatiotemporal feature extraction module extracts local perceptual features through a convolutional layer, and then generates the sixth microgrid node fusion feature sequence X6 through the ReLU activation function;

[0086] The fusion feature sequence X6 of the sixth microgrid node is processed by a multi-head self-attention mechanism, random deactivation dropout operation, and then processed by a fully connected layer to complete the initial feature transformation. It is then further mapped by the GELU activation function and finally normalized by the layer normalization submodule to output the fusion feature sequence X9 of the ninth microgrid node.

[0087] For details, please refer to Figure 3 As shown, D 2_2-std Local perceptual feature extraction is performed on a 3×1 one-dimensional convolutional layer Conv1D. Then, the nonlinear expression capability is enhanced by the ReLU activation function, generating the sixth microgrid node fusion feature sequence X6. X6 is then processed by a multi-head self-attention mechanism to generate the seventh microgrid node fusion feature sequence X7. X7 undergoes a random dropout operation to generate the eighth microgrid node fusion feature sequence X8. X8 is first processed by a fully connected layer (FC) for preliminary feature transformation, then enhanced with a GELU activation function for nonlinear expression, and further mapped by the fully connected layer (FC). Finally, feature normalization is performed through a layer normalization submodule (LN) to generate the ninth microgrid node fusion feature sequence X9.

[0088] A1-3: The microgrid communication link state dataset D 2_3 The third branch of the spatiotemporal feature extraction module is input, and the position feature is enhanced by the position encoding submodule. Then, it is processed by a fully connected layer to generate the fusion feature sequence X10 of the tenth microgrid node. X10 is input into the graph convolutional network to generate the fusion feature sequence X11 of the eleventh microgrid node. X10 and X11 are concatenated to generate the fusion feature sequence X12 of the twelfth microgrid node. X12 is passed through the layer normalization submodule to output the fusion feature sequence X13 of the thirteenth microgrid node.

[0089] A2: The features X5, X9 and X13 output from the three branches are concatenated to generate the fusion feature sequence X14 of the fourteenth microgrid node.

[0090] A3: Input X14 into the Transformer encoder to generate the 15th microgrid node fusion feature sequence X15; input X15 into the cosine similarity calculation submodule to generate the 16th microgrid node fusion feature sequence X16; input X16 into the Softmax normalization submodule to generate the 17th microgrid node fusion feature sequence X17; perform Top-K filtering on X17, and then input it into the mask processing submodule to finally generate the microgrid dynamic adjacency matrix A.

[0091] Simultaneously, the fusion feature sequence X15 of the fifteenth microgrid node is processed through a fully connected layer to generate the fusion feature sequence X18 of the eighteenth microgrid node; X18 is input into a batch normalization layer to generate the fusion feature sequence X19 of the nineteenth microgrid node; X19 is subjected to a random deactivation operation to finally generate the fusion feature sequence X of the microgrid node.

[0092] It should be noted that the Top-K filtering operation mentioned above refers to retaining only the top K most relevant connections at each time step in the similarity matrix or attention weight matrix, while setting the remaining connections to zero or masking them, thus forming sparse adjacency relationships. This operation can highlight key associations, reduce noise interference, and effectively reduce computational complexity. In microgrid modeling, Top-K filtering ensures that each node establishes connections only with its K most important neighbors, thereby enhancing the expressive power of the graph structure and the robustness of the model.

[0093] It should also be noted that the Spatiotemporal Feature Extraction Module (PDTM) is used to extract multi-scale, globally correlated spatiotemporal features from heterogeneous data such as environmental parameters, equipment operating status, and communication link status.

[0094] In the PDTM module, a multi-branch deep feature extraction structure is adopted, introducing a graph convolutional network (GCN) into the location-enhanced communication link feature branch to capture the topological relationships between microgrid nodes. A multi-head self-attention mechanism is integrated into the device state feature branch to model global dependencies across time and nodes. Multi-scale one-dimensional convolution is introduced into the environmental feature branch to extract local patterns of external disturbances such as weather. Through a feature selection mechanism, PDTM can adaptively update the attention level to different features as the input data changes, achieving sensitive capture of the dynamic evolution of the microgrid topology. This module ultimately outputs a fused feature sequence X of microgrid nodes and a dynamic adjacency matrix A. X represents the unified feature representation of each node at the current moment after fusing multimodal spatiotemporal information; physically, it can be understood as a node state vector simultaneously containing device operating conditions, communication status, and environmental influences, serving as the basis for subsequent policy generation and control decisions. A represents the dynamic relationship between nodes at the current moment, reflecting the communication and energy interaction structure under actual microgrid operation, providing the most influential node connection reference for subsequent model decision-making.

[0095] In this embodiment of the application, step S300 involves fusing the feature sequence X, the microgrid dynamic adjacency matrix A, and the power data D from the running dataset. 2_4 The input is fed into the attention feature extraction module (MSAF) in the model to generate the power supply and demand prediction values ​​for the microgrid, including the following steps B1-B7:

[0096] B1: Input the fused feature sequence X, the microgrid dynamic adjacency matrix A, and the power data into the three branches of the attention feature extraction module respectively;

[0097] Specifically, refer to Figure 4 In the first branch of MSAF, D 2_4Local spatial features are extracted from a 3×3 2D convolutional layer (Conv2D), and then input into the batch normalization submodule (BN) to generate the first microgrid power supply and demand feature sequence T1. T1 is then input into a 3×1 1D convolutional layer (Conv1D), which extracts local temporal features. This is followed by input into a ReLU activation function to generate the second microgrid power supply and demand feature sequence T2. T2 is then input into a 5×1 1D convolutional layer (Conv1D), which extracts deep local contextual features. This is followed by ReLU activation to generate the third microgrid power supply and demand feature sequence T3. T2 and T3 are concatenated (concat) to generate the fourth microgrid power supply and demand feature sequence T4. T4 is then input into the batch normalization submodule (BN) to generate the fifth microgrid power supply and demand feature sequence T5.

[0098] In the second branch of MSAF, the spatial topology information of the input graph convolutional network GraphConv is fused, and then the ReLU activation function is used to generate the power supply and demand feature sequence T6 of the sixth microgrid. The power supply and demand feature sequence T7 of the seventh microgrid is generated in the LayerNorm input layer normalization submodule. The deep graph structure modeling process of the input graph convolutional network GraphConv of T7 is performed, and then the power supply and demand feature sequence T8 of the eighth microgrid is generated in the LayerNorm input layer normalization submodule.

[0099] In the third branch of MSAF, X is input into a one-dimensional convolutional layer Conv1D with a kernel size of 3×1 for shallow temporal feature extraction. Then, the ReLU activation function is used to generate the power supply and demand feature sequence T9 of the ninth microgrid. T9 is input into a multi-head self-attention mechanism for temporal dependency modeling to generate the power supply and demand feature sequence T10 of the tenth microgrid. T10 is input into a fully connected layer FC for feature mapping to generate the power supply and demand feature sequence T11 of the eleventh microgrid. T9 and T11 are concatenated to generate the power supply and demand feature sequence T12 of the twelfth microgrid. T12 is input into the batch normalization submodule BN to generate the power supply and demand feature sequence T13 of the thirteenth microgrid.

[0100] B2: The final power characteristic sequences (i.e., T5, T8 and T13) output by the three branches are spliced ​​together to generate the fourteenth microgrid power supply and demand characteristic sequence T14;

[0101] B3: Input T14 into the Transformer encoder for global time series modeling to generate the power supply and demand characteristic sequence T15 of the fifteenth microgrid;

[0102] B4: Input T15 into the fully connected layer for deep nonlinear transformation, and then use the ReLU activation function to generate the power supply and demand characteristic sequence T16 of the sixteenth microgrid;

[0103] B5: Input T16 into the random deactivation dropout layer for feature suppression processing to generate the power supply and demand feature sequence T17 of the seventeenth microgrid;

[0104] B6: Input T17 into the fully connected layer for feature mapping processing to generate the power supply and demand feature sequence T18 of the 18th microgrid;

[0105] B7: Input T18 into the prediction head to generate the predicted power supply and demand value T for the microgrid.

[0106] It should be noted that the attention feature extraction module (MSAF) adopts a multi-branch structure, combining multi-scale convolution, graph convolution, and multi-head self-attention to integrate the multi-scale spatial features and temporal dependencies of microgrid power supply and demand data. After multimodal feature extraction, dynamic fusion, and global modeling, the predicted power supply and demand value T of the microgrid is output. This predicted value comprehensively reflects the current and historical trends in supply and demand, spatial coordination relationships, and network topology characteristics, providing a crucial basis for subsequent microgrid regulation and control decisions.

[0107] In this embodiment of the application, step S400 involves storing the current distributed energy device operating status dataset D. 2_2-cur The fusion feature sequence X and the power prediction value T are input into the HHACRL-CM module in the model to generate the microgrid hybrid strategy feature sequence E, including the following steps C1-C7:

[0108] C1: Reference Figure 5 The current distributed energy equipment operating status dataset D 2_2-cur The fused feature sequence X and the power prediction value T are respectively input into the three branches of the HHACRL-CM module;

[0109] Specifically, the current distributed energy device operating status dataset D 2_2-cur This data originates from the voltage, current, and state of charge of distributed energy devices and energy storage devices during the operation of the microgrid. (D) 2_2-cur Time resolution and D 2_2-std Consistent, but D 2_2-cur The data coverage is limited to the current control cycle and is used for real-time operational status analysis and decision generation. 2_2-cur Without prior historical batch preprocessing, the real-time and accurate control response is ensured.

[0110] Specifically, in the first branch of HHACRL-CM, the one-dimensional convolutional layer Conv1D with a kernel size of 3×1 input T is processed for local temporal feature extraction, and then the nonlinear expression ability is enhanced by the ReLU activation function to generate the first micro-energy network hybrid strategy feature sequence E1; the feature normalization is performed in the normalization submodule BN of the input layer of E1 to generate the second micro-energy network hybrid strategy feature sequence E2; E2 is input into the multi-head self-attention mechanism for global temporal dependency modeling to generate the third micro-energy network hybrid strategy feature sequence E3.

[0111] In the second branch of HHACRL-CM, shallow local temporal feature extraction is performed on the one-dimensional convolutional layer Conv1D with a kernel size of 3×1 input X. Then, the nonlinear expression capability is enhanced by the ReLU activation function to generate the fourth microgrid hybrid strategy feature sequence E4.

[0112] C2: Concatenate the fourth microgrid hybrid strategy feature sequence E4 output from the second branch of the HHACRL-CM module and the fifth microgrid hybrid strategy feature sequence E5 initially output from the third branch to generate the sixth microgrid hybrid strategy feature sequence E6.

[0113] C3: E6 is processed through a layer normalization submodule and a multi-head self-attention mechanism to perform deep global feature modeling and generate the eighth microgrid hybrid strategy feature sequence E8.

[0114] Specifically, steps C2-C3 refer to 5. In the third branch of HHACRL-CM, the current distributed energy device operating status dataset D is... 2_2-cur The input is processed by linear transformation in the fully connected layer FC, followed by nonlinear feature enhancement through the ReLU activation function to generate the fifth microgrid hybrid strategy feature sequence E5. E4 and E5 are concatenated to generate the sixth microgrid hybrid strategy feature sequence E6. E6 is input into the layer normalization submodule LN for feature normalization to generate the seventh microgrid hybrid strategy feature sequence E7. E7 is input into a multi-head self-attention mechanism for deep global feature modeling to generate the eighth microgrid hybrid strategy feature sequence E8.

[0115] C4: The third microgrid hybrid strategy feature sequence E3 and E8 output from the first branch in the HHACRL-CM module are concatenated and then processed by fully connected layer mapping and normalization to generate the eleventh microgrid hybrid strategy feature sequence E11.

[0116] Specifically, C4 reference Figure 5E9 is input into the fully connected layer FC for feature mapping processing to generate the feature sequence E10 of the tenth microgrid hybrid strategy; E10 is input into the layer normalization submodule LN for feature normalization processing to generate the feature sequence E11 of the eleventh microgrid hybrid strategy.

[0117] C5: Input E11 into two fully connected layers simultaneously, perform the first and second processing respectively, output the fifteenth microgrid hybrid strategy feature sequence E15 and the sixteenth microgrid hybrid strategy feature sequence E16 respectively, and splice them to obtain the seventeenth microgrid hybrid strategy feature sequence E17.

[0118] C6: Input E17 sequentially into the fully connected layer and the hyperbolic tangent activation function for nonlinear transformation processing to generate the feature sequence E19 of the nineteenth microgrid hybrid strategy;

[0119] C7: Concatenate E15 and E19 to generate the microgrid hybrid strategy feature sequence E.

[0120] Specifically, C5-C7 refer to Figure 5 E11 is input into the fully connected layer FC for feature transformation processing to generate the twelfth microgrid hybrid strategy feature sequence E12; E12 is input into the gated linear unit submodule GLU for gated feature processing to generate the thirteenth microgrid hybrid strategy feature sequence E13; E13 is input into the fully connected layer FC for output mapping processing to generate the fourteenth microgrid hybrid strategy feature sequence E14; E14 is input into the normalized output submodule (SoftmaxOut) for normalization processing to generate the fifteenth microgrid hybrid strategy feature sequence E15. Simultaneously, E11 is input into the fully connected layer FC for feature supplementation processing to generate the sixteenth microgrid hybrid strategy feature sequence E16; E15 and E16 are concatenated using the concat operation to generate the seventeenth microgrid hybrid strategy feature sequence E17; E17 is input into the fully connected layer FC for feature fusion processing to generate the eighteenth microgrid hybrid strategy feature sequence E18; E18 is input into the hyperbolic tangent activation function TanhAct for nonlinear transformation processing to generate the nineteenth microgrid hybrid strategy feature sequence E19; E15 and E19 are concatenated to finally generate the microgrid hybrid strategy feature sequence E.

[0121] It should be noted that the HHACRL-CM (Hybrid Hierarchical Attention and Cooperative Representation Learning for ControlModule) in the above scheme is designed for the operation and control of microgrids in complex multi-source dynamic environments. It employs a branched hierarchical feature modeling strategy, combining local convolutional extraction, global multi-head self-attention modeling, and cooperative representation learning to achieve collaborative perception and deep fusion of multi-source operational state data, dynamic topological features, and external environmental information. This module outputs a hybrid strategy feature sequence E for the microgrid, reflecting the multi-dimensional state cooperative mode of the microgrid at the current moment and its response tendency to future operational trends. This sequence can directly drive the generation of control decisions, improving the adaptability, stability, and response efficiency of the microgrid in dynamic environments.

[0122] The hyperbolic tangent activation function (TanhAct) is a non-linear activation function. This function maps the input to a range of -1 to 1, and compared to the sigmoid function, it has a stronger gradient representation capability, helping to alleviate the gradient vanishing problem. Therefore, it is often used for non-linear transformations in intermediate layers of deep neural networks. SoftmaxOut is an activation function commonly used in the output layer of neural networks, particularly suitable for multi-class classification tasks. It transforms the model output into predicted probabilities for each class by performing an exponential operation on the value of each output node and normalizing it to a probability distribution that sums to 1, facilitating subsequent supervised learning using loss functions such as cross-entropy.

[0123] In this embodiment of the application, in step S500, the microgrid hybrid strategy feature sequence E is input to the decision generation module in the model, and the microgrid regulation decision I is output, wherein:

[0124] The microgrid control decision is a vector of command sequences arranged in chronological order:

[0125] I = [I1, I2, ..., I t ,…,I N ]

[0126] Where N represents the total number of time steps within the future time window; I t This represents the device control parameter vector at time step t, which is the control command for all critical device units in the microgrid at time step t.

[0127] I t =[p1,p2,…,p j ,…,p m ]

[0128] Where m represents the number of microgrid devices, such as distributed photovoltaic inverters, energy storage systems, load management units, communication equipment, etc.; p j This indicates the specific control parameters of the j-th device at time step t, such as the maximum active power output limit and power factor setting of the photovoltaic inverter, the charging and discharging power and target state of charge (SOC) of the energy storage system, the start / stop status and power regulation ratio of the load unit, and the transmission power and retransmission count of the communication equipment.

[0129] It should be noted that the decision generation module is a key module for microgrid regulation. Its function is to receive operating status data from various devices in the microgrid and generate specific control commands based on preset strategies or optimization algorithms, thereby guiding device operation and power allocation and realizing basic system regulation.

[0130] Finally, the microgrid intelligent control model based on multimodal perception is embedded into the microgrid scheduling system. It accepts real-time multi-source dynamic operation data from the actual operating environment as input, generates microgrid control decisions, and outputs them directly to the scheduling execution module to realize real-time control and optimized scheduling of distributed energy devices.

[0131] In summary, this invention constructs a microgrid intelligent control model based on multimodal sensing. The overall structure of the model is as follows: Figure 2 As shown, by fusing heterogeneous data from multiple sources, including environmental parameters, the operating status of distributed energy devices, communication link status, and historical power supply and demand, the model achieves accurate perception and dynamic decision-making regarding the operational status of microgrids. This model, with multimodal data fusion at its core, uses three key steps—feature extraction, temporal modeling, and policy generation—to map real-time and historical information from different dimensions into a unified high-dimensional feature space. Combined with reinforcement learning and optimization decision-making mechanisms, it generates adaptive and robust control strategies, thereby enabling efficient, stable, and economical operation of microgrids in complex and ever-changing operating environments.

[0132] Overall, the PDTM module constructs a multi-branch deep network that integrates environmental data, device status, and communication status, combining convolution, graph convolution, and self-attention mechanisms to achieve joint modeling of multi-scale spatial features and temporal dependencies. PDTM dynamically fuses multimodal features, effectively enhancing the integrity and robustness of feature representation, and providing a solid foundation for the perception and prediction of dynamic changes in microgrids.

[0133] The MSAF module integrates convolutional neural networks and graph convolutional networks to fully exploit the spatial topology and multi-scale temporal features of microgrids. Through multi-scale branching design and dynamic weighting with attention mechanisms, the module can capture key change information at different temporal and spatial scales, enhance its ability to represent multi-source heterogeneous data, and improve the model's responsiveness and generalization performance to key supply and demand changes. This provides more accurate and stable supply and demand forecasting support for the generation of subsequent regulation strategies.

[0134] The HHACRL-CM module employs a multi-branch parallel structure, integrating layer normalization, multi-head self-attention mechanisms, and hybrid activation functions. Through multi-layer feature fusion and dynamic filtering mechanisms, it achieves efficient perception and policy generation of the multi-source dynamic environment of microgrids. This module combines multi-source state perception results with supply and demand forecast information, and models the correlation features between key states through a hybrid attention mechanism, enhancing the policy generation's ability to perceive and adapt to dynamic disturbances. Simultaneously, HHACRL-CM adopts a global-local collaborative modeling approach, fusing discrete and continuous control actions within a unified policy space to improve the stability and efficiency of control response. This module enables the generation of microgrid control decisions under complex operating scenarios, providing microgrids with globally superior dynamic scheduling capabilities.

[0135] Overall, it has improved the adaptability and execution stability of the strategy network in complex and dynamic environments, providing strong technical support for the intelligent regulation of distributed energy in microgrids.

[0136] Example 3: Based on the above implementation methods, this example provides an illustrative scheme of a microgrid intelligent control method based on multimodal sensing to verify the feasibility and beneficial effects of the method.

[0137] S100: Acquire and preprocess to generate a multi-source dynamic running dataset;

[0138] Taking a microgrid in a coastal city as an example, multi-source data on the operation of the microgrid from January 2023 to June 2024 were collected to construct a multi-source dynamic operation dataset D1. Each sub-dataset was preprocessed to form a standardized multi-source dynamic operation dataset D2.

[0139] (1) Microgrid environmental parameter dataset D 1_1

[0140] Meteorological sensors deployed around the microgrid photovoltaic array and energy storage sites collected temperature, humidity, and light intensity data at 5-minute intervals, totaling over 470,000 data records. First, a moving average method was used to smooth the environmental parameters with a window size of 3 to reduce instantaneous fluctuations. Then, the Z-score method was used to standardize each feature, ensuring a mean of 0 and a standard deviation of 1. Finally, box plots were used to identify outliers and replaced them with the mean of adjacent time points. The resulting standardized microgrid environmental parameter dataset D was obtained after processing. 2_1 Its characteristic size is 87648×3.

[0141] (2) Distributed energy equipment operation status dataset D 1_2

[0142] Voltage, current, and State of Charge (SOC) data of distributed energy devices were collected from the inverter monitoring system and the energy storage management system (BMS) at 1-minute intervals, totaling over 2.36 million records. First, data from different devices were aligned along a unified timeline. For data with inconsistent collection intervals, cubic spline interpolation was used to fill in missing points. Then, each feature was normalized to the [0,1] interval. After processing, a standardized distributed energy device operating status dataset D was obtained. 2_2-std Its characteristic size is 87648×3.

[0143] (3) Microgrid communication link state dataset D 1_3

[0144] Communication latency, packet loss rate, and link stability indicators were collected from the microgrid communication management system at a 1-minute sampling interval, totaling over 2.36 million records. First, missing values ​​were filled using forward imputation. Then, min-max normalization was used to map each feature to [0,1]. Finally, binary stability labels were constructed based on the historical fluctuation range of communication link stability, with stable states corresponding to a value of 1 and unstable states corresponding to a value of 0. After processing, a standardized microgrid communication link state dataset D was obtained. 2_3 Its feature size is 87648×4, where the first three columns are normalized features and the last column is a stability label.

[0145] (4) Historical power supply and demand dataset of microgrids D 1_4

[0146] Data on photovoltaic power output and grid load demand were collected from the SCADA system of the energy dispatch center at 15-minute intervals, totaling over 175,000 records. First, the time series data was padded to ensure no missing time points. Then, wavelet thresholding was used to reduce short-term spike noise. Finally, interval normalization was performed to [0,1]. The resulting standardized microgrid historical power supply and demand dataset D was obtained. 2_4 Its characteristic size is 35056×2.

[0147] S200: Construct a microgrid intelligent control model based on multimodal perception, input a portion of the running dataset into the spatiotemporal feature extraction module in the model, and generate a microgrid node fusion feature sequence X and a microgrid dynamic adjacency matrix A;

[0148] In the first branch of the spatiotemporal feature extraction module, D 2_1 First, the feature sequence is processed by a fully connected layer (FC) to complete the initial feature transformation. Then, the nonlinear expression is enhanced by the GELU activation function, followed by further mapping in the fully connected layer (FC). Finally, the feature is normalized by the layer normalization submodule (LN) to generate the first microgrid node fusion feature sequence X1, with a feature size of 48×256. X1 is then input into a one-dimensional convolutional layer (Conv1D) with a kernel size of 3×1 for local feature extraction, generating the second microgrid node fusion feature sequence X2, with a feature size of 48×256. Subsequently, the nonlinear expression is enhanced by the ReLU activation function, and the sequence is then input into the batch normalization submodule (BN) for feature normalization. Finally, X1 is input into a one-dimensional convolutional layer (Conv1D) with a kernel size of 5×1 for local feature extraction, generating the third microgrid node fusion feature sequence X3, with a feature size of 48×256. Subsequently, the nonlinear expression is enhanced by the ReLU activation function, and the sequence is then input into the batch normalization submodule (BN) for feature normalization. X1 is input into a 7×1 one-dimensional convolutional layer Conv1D for local feature extraction, generating the fourth microgrid node fusion feature sequence X4, with a feature size of 48×256. Subsequently, the nonlinear expression is enhanced using the ReLU activation function, and the features are normalized by the batch normalization (BN) module. X2, X3, and X4 are concatenated along the channel dimension to generate the fifth microgrid node fusion feature sequence X5, with a feature size of 48×768.

[0149] In the second branch of the spatiotemporal feature extraction module, D 2_2-stdA one-dimensional convolutional layer Conv1D with a kernel size of 3×1 is input for local perceptual feature extraction, generating the sixth microgrid node fusion feature sequence X6, with a feature size of 48×256. This is then enhanced with a ReLU activation function to improve the nonlinear representation. X6 is input into a multi-head self-attention mechanism to generate the seventh microgrid node fusion feature sequence X7, with a feature size of 48×256 and 8 attention heads. X7 is randomly deactivated at a deactivation rate of 0.1, generating the eighth microgrid node fusion feature sequence X8, with a feature size of 48×256. X8 is processed through a fully connected layer FC, then enhanced with a GELU activation function, and then mapped again through a fully connected layer FC. Finally, feature normalization is performed using a layer normalization submodule LN to generate the ninth microgrid node fusion feature sequence X9, with a feature size of 48×256.

[0150] In the third branch of the spatiotemporal feature extraction module, D 2_3 The input location encoding submodule PE performs location feature enhancement, which is then processed by the fully connected layer FC to generate the fused feature sequence X10 for the tenth microgrid node, with a feature size of 48×256. Here, D... 2-3 A temporal feature matrix of 87648×4 is constructed, where 87648 represents the total number of samples and 4 represents the feature dimension. A single window of samples has 48 time steps, with 4 features per time step. X10 is input into the Graph Convolutional Network (GCN) to generate the eleventh microgrid node fusion feature sequence X11, with a feature size of 10×256 and 10 nodes. X10 and X11 are concatenated along their feature dimensions to generate the twelfth microgrid node fusion feature sequence X12, with a feature size of 10×512. X12 is input into the layer normalization submodule LN to generate the thirteenth microgrid node fusion feature sequence X13, with a feature size of 10×512.

[0151] After time alignment with X5, X9, and X13, they are concatenated along the feature dimension to generate the fourteenth microgrid node fusion feature sequence X14, with a feature size of 48×1536, where X5 is 48×768, X9 is 48×256, and X13 is mapped to 48×512. X14 is input into the Transformer Encoder to generate the fifteenth microgrid node fusion feature sequence X15, with a feature size of 48×256. The encoder has 4 layers, 8 attention heads, and 512 hidden dimensions. X15 is input into the CosSim cosine similarity calculation submodule to generate the sixteenth microgrid node fusion feature sequence X16, with a feature size of 48×48. X16 is input into the Softmax Norm normalization submodule to generate the seventeenth microgrid node fusion feature sequence X17, with a feature size of 48×48. X17 is filtered by Top-K and input into the mask processing submodule Mask to finally generate the microgrid dynamic adjacency matrix A, which has a feature size of 48×48 and a K value of 5.

[0152] X15 is processed through a fully connected layer (FC) to generate the 18th microgrid node fusion feature sequence X18, with a feature size of 48×256. X18 is then input into a batch normalization layer (BN) to generate the 19th microgrid node fusion feature sequence X19, also with a feature size of 48×256. X19 is then randomly deactivated at a deactivation rate of 0.1, finally generating the microgrid node fusion feature sequence X, with a feature size of 48×256. Here, 48 represents that each sample window contains 48 consecutive time steps, each corresponding to the state of the microgrid node at that moment; 256 represents the feature dimension of each time step, with each dimension carrying multi-source feature information after node fusion, including temporal features, spatial features, and nonlinear mapping results.

[0153] S300: Input the fused feature sequence X, the microgrid dynamic adjacency matrix A, and the power data in the running dataset into the attention feature extraction module in the model to generate the microgrid power supply and demand prediction value T;

[0154] In the first branch of the attention feature extraction module, D 2_4Local spatial feature extraction is performed on a 3×3 2D convolutional layer (Conv2D), resulting in an output feature size of 48×10×64. This output feature size is then fed into the batch normalization (BN) module to generate the first microgrid power supply and demand feature sequence T1, maintaining the same 48×10×64 feature size. T1 is then fed into a 3×1 1D convolutional layer (Conv1D), performing local temporal feature extraction. The output feature size is 48×10×128, which is then fed into a ReLU activation function to generate the second microgrid power supply and demand feature sequence T2. T2 is then fed into a 5×1 1D convolutional layer (Conv1D), performing deep local context feature extraction. The output feature size is 48×10×256, which is then used with a ReLU activation function to generate the third microgrid power supply and demand feature sequence T3. T2 and T3 are concatenated along the channel dimension to generate the fourth microgrid power supply and demand feature sequence T4, with a feature size of 48×10×384. T4 is then input into the batch normalization submodule BN to generate the fifth microgrid power supply and demand feature sequence T5, with a feature size of 48×10×384.

[0155] In the second branch of the attention feature extraction module, the input graph convolutional network GraphConv is processed by spatial topological structure information fusion. Combining the input node feature dimension of 256, the output feature size is 48×10×256. Then, the ReLU activation function is applied to generate the power supply and demand feature sequence T6 of the sixth microgrid. The input layer normalization submodule LayerNorm of T6 generates the power supply and demand feature sequence T7 of the seventh microgrid, maintaining its feature size of 48×10×256. The input graph convolutional network GraphConv of T7 is processed by deep graph structure modeling, outputting a feature size of 48×10×256. Then, the input layer normalization submodule LayerNorm generates the power supply and demand feature sequence T8 of the eighth microgrid.

[0156] In the third branch of the attention feature extraction module, X is input to a 3×1 one-dimensional convolutional layer (Conv1D) for shallow temporal feature extraction, resulting in an output feature shape of 48×256. Then, a ReLU activation function is applied to generate the power supply and demand feature sequence T9 for the ninth microgrid. T9 is then input to a multi-head self-attention mechanism (Multi-Head Self-Attention) for temporal dependency modeling, with 8 attention heads and an output dimension maintained at 256, generating the power supply and demand feature sequence T10 for the tenth microgrid. T10 is then input to a fully connected layer (FC) for feature mapping, expanding the number of channels to 384, generating the power supply and demand feature sequence T11 for the eleventh microgrid. T9 and T11 are concatenated along the channel dimension to generate the power supply and demand feature sequence T12 for the twelfth microgrid, with a concatenated shape of 48×640. T12 is then input to the batch normalization (BN) submodule to generate the power supply and demand feature sequence T13 for the thirteenth microgrid, maintaining its 48×640 feature shape.

[0157] Concatenate T5, T8, and T13 along the channel dimension to generate the fourteenth microgrid power supply and demand feature sequence T14, with a feature size of 48×10×1280. Input T14 into a Transformer encoder for global temporal modeling (4 encoder layers, 8 attention heads, 512 hidden dimensions), maintaining the output feature size of 48×10×1280, generating the fifteenth microgrid power supply and demand feature sequence T15. Input T15 into a fully connected (FC) layer for deep nonlinear transformation, mapping the number of channels to 512, and then apply the ReLU activation function to generate the sixteenth microgrid power supply and demand feature sequence T16. Input T16 into a dropout layer for feature suppression (0.1 dropout rate), generating the seventeenth microgrid power supply and demand feature sequence T17. T17 is input into the fully connected layer FC for feature mapping, mapping the number of channels to 1, generating the 18th microgrid power supply and demand feature sequence T18, with a feature size of 48×10×1. Finally, T18 is input into the prediction head to generate the microgrid power supply and demand prediction value T, with a feature size of 48×10. Here, 48 represents the time step of the prediction sequence, i.e., the prediction results covering 48 consecutive future time points; 10 represents the number of nodes in the microgrid, i.e., the power supply and demand predictions corresponding to 10 typical distributed energy units.

[0158] S400: Input the current distributed energy equipment operating status dataset, the fusion feature sequence X, and the power prediction value T into the HHACRL-CM module in the model to generate the microgrid hybrid strategy feature sequence E;

[0159] In the first branch of the HHACRL-CM module, a one-dimensional convolutional layer Conv1D with a size of 48×10 T, a kernel size of 3×1, and 64 output channels is processed for local temporal feature extraction to obtain a feature sequence of size 48×64. Then, the nonlinear expression capability is enhanced by the ReLU activation function to generate the first microgrid hybrid strategy feature sequence E1 with a size of 48×64. The feature normalization is performed in the input layer normalization submodule BN of E1 with a size of 48×64 to generate the second microgrid hybrid strategy feature sequence E2 with a size of 48×64. The E2 with a size of 48×64 is input into the multi-head self-attention mechanism MHSA for global temporal dependency modeling to generate the third microgrid hybrid strategy feature sequence E3 with a size of 48×64.

[0160] In the second branch of the HHACRL-CM module, shallow local temporal feature extraction is performed on a one-dimensional convolutional layer Conv1D with an input convolution kernel size of 3×1 and an output channel number of 64, which has a size of 48×32, to obtain a feature sequence with a size of 48×64. Then, the nonlinear expression capability is enhanced by the ReLU activation function to generate the fourth microgrid hybrid strategy feature sequence E4 with a size of 48×64.

[0161] In the third branch of the HHACRL-CM module, the current distributed energy device operating status dataset D with a size of 48×20 is used. 2_2-std Linear transformation is performed on a fully connected layer (FC) with input and output dimensions of 64 to obtain a feature sequence of size 48×64. This is then processed by a ReLU activation function for non-linear feature enhancement, generating the fifth microgrid hybrid strategy feature sequence E5, also of size 48×64. The 48×64 E4 and E5 are concatenated to generate the sixth microgrid hybrid strategy feature sequence E6, of size 48×128. Feature normalization is performed on the input layer normalization submodule LN of the 48×128 E6 to generate the seventh microgrid hybrid strategy feature sequence E7, also of size 48×128. Finally, the 48×128 E7 is input to a multi-head self-attention mechanism for deep global feature modeling, generating the eighth microgrid hybrid strategy feature sequence E8, also of size 48×128.

[0162] Subsequently, the E3 (48×64) and E8 (48×128) are concatenated to generate the ninth microgrid hybrid strategy feature sequence E9 (48×192). Feature mapping is then performed on the fully connected layer FC (128 input-output dimensions) of E9 (48×192) to generate the tenth microgrid hybrid strategy feature sequence E10 (48×128). Feature normalization is then performed on the input layer normalization submodule LN (48×128) of E10 to generate the eleventh microgrid hybrid strategy feature sequence E11 (48×128).

[0163] The feature transformation process is performed on the fully connected layer FC with a size of 48×128 and an input-output dimension of 128 in E11 to generate the 12th microgrid hybrid strategy feature sequence E12 with a size of 48×128. The gating feature processing is performed on the input-gated linear unit submodule GLU of E12 with a size of 48×128 to generate the 13th microgrid hybrid strategy feature sequence E13 with a size of 48×128. The output mapping process is performed on the fully connected layer FC with a size of 48×128 and an input-output dimension of 64 in E13 to generate the 14th microgrid hybrid strategy feature sequence E14 with a size of 48×64. The normalization process is performed on the input-normalized output submodule SoftmaxOut of E14 with a size of 48×64 to generate the 15th microgrid hybrid strategy feature sequence E15 with a size of 48×64.

[0164] Simultaneously, feature supplementation is performed on the fully connected layer FC of E11 (size 48×128, input-output dimension 64) to generate the sixteenth microgrid hybrid strategy feature sequence E16 (size 48×64); the 48×64 E15 and E16 are concatenated to generate the seventeenth microgrid hybrid strategy feature sequence E17 (size 48×128); and the 48×128 fully connected layer F of E17 (size 48×128, input-output dimension 64) is then processed. In step C, feature fusion processing is performed to generate the 18th microgrid hybrid strategy feature sequence E18 with a size of 48×64. The 48×64 E18 is input into the hyperbolic tangent activation function TanhAct for nonlinear transformation processing to generate the 19th microgrid hybrid strategy feature sequence E19 with a size of 48×64. The 48×64 E15 and the 48×64 E19 are concatenated to finally generate the microgrid hybrid strategy feature sequence E with a size of 48×128.

[0165] S500: Input the microgrid hybrid strategy feature sequence into the decision generation module in the model, output the microgrid regulation decision, and realize precise control and global optimization of the operating status of various devices in the microgrid.

[0166] Input the microgrid hybrid strategy feature sequence E into the decision generation module, and output the microgrid regulation decision I.

[0167] This decision is represented as a vector of instructions arranged in chronological order:

[0168] I = [I1, I2, ..., I 48 ]

[0169] The time window length is 48, meaning 48 consecutive time steps in the future, with each I... t This represents the set of control commands for all critical equipment units within the microgrid at time step t.

[0170] I t =[p1,p2,…,p j ,…,p 10 ]

[0171] Among them, 10 represents the number of key devices for microgrids, p j This represents the specific control parameters of the j-th device at time step t. For example:

[0172] For distributed photovoltaic unit inverters, p j Includes the upper limit of active power output (unit: kW) and the power factor setting value; for energy storage systems, p j Includes charge / discharge power (unit: kW) and target state of charge (SOC) (%); for load cells, p j Indicates the start / stop status (0 / 1) and power regulation ratio (%); for communication equipment, p j This indicates the transmission power (dBm) and the number of retransmissions.

[0173] A microgrid intelligent control model based on multimodal sensing was deployed in a coastal city's microgrid area. The model received real-time dynamic operational data from multiple sources as input, including environmental parameter data (temperature, humidity, light intensity, and wind speed, with a time resolution of 30 seconds and a data size of 90,000×4); distributed energy device operating status data (voltage, current, and energy storage device SOC, with a time resolution of 30 seconds and a data size of 90,000×3); communication link status data (communication delay, packet loss rate, link stability label, and signal-to-noise ratio, with a data size of 90,000×5); and historical power supply and demand data (photovoltaic power output, load demand, and regional grid frequency, with a data size of 40,000×3). The model fused these real-time data to generate a hybrid strategy feature sequence E, and further output a microgrid control decision sequence D with a dimension of 96×12, corresponding to 96 consecutive time steps and 12 key equipment units.

[0174] In the first time step, the scheduling execution submodule sends the control commands in the decision sequence to each device: the photovoltaic inverter unit 1 sets the maximum active power limit to 900kW and adjusts the power factor to 0.99; the energy storage system 1 sets the charging power limit to -150kW and the target SOC to 62%; the load management unit 1 starts and adjusts the power to 90%; the communication equipment 1 transmits power to 18dBm and retransmits 4 times; other devices make corresponding adjustments according to the model commands.

[0175] In the second time step, the maximum power of photovoltaic inverter unit 1 is increased to 920kW, and the power factor is maintained at 0.99; the energy storage system 1 increases the charging power to -180kW, and the SOC target is increased to 64%; the power of load management unit 1 is reduced to 85%; the transmission power of communication equipment 1 is reduced to 15dBm, and the number of retransmissions is 3; other devices execute the instructions at the corresponding time.

[0176] In the third time step, the power limit of photovoltaic inverter unit 1 is finely adjusted to 930kW, and the power factor is slightly reduced to 0.985; the energy storage system 1 switches from charging to discharging, with the discharge power limited to 20kW and the SOC target reduced to 60%; the power of load management unit 1 is adjusted to 80%; the transmission power of communication equipment 1 is adjusted to 16dBm, and the number of retransmissions is maintained at 3; each device continues to receive and execute the dynamic control commands generated by the model.

[0177] The control strategy for subsequent time steps is the same as described above, dynamically adjusting the operating status of each device;

[0178] By embedding a multimodal sensing-based intelligent control model into the microgrid and combining it with real-time multi-source dynamic data, the model can dynamically generate refined control decisions for distributed energy devices, enabling real-time optimized management of the microgrid and effectively improving its stability, economy, and operational efficiency.

[0179] Example 4 illustrates a schematic scheme of a microgrid intelligent control method based on multimodal sensing. It should be noted that the technical solution of this microgrid intelligent control system based on multimodal sensing belongs to the same concept as the technical solution of the aforementioned microgrid intelligent control method based on multimodal sensing. Details not described in detail in this embodiment of the microgrid intelligent control system based on multimodal sensing can be found in the description of the aforementioned microgrid intelligent control method based on multimodal sensing.

[0180] This embodiment also provides another intelligent control system for microgrids based on multimodal sensing, including:

[0181] The acquisition module is used to acquire and preprocess the generated multi-source dynamic running dataset;

[0182] The first generation module is used to construct a microgrid intelligent control model based on multimodal perception. It inputs a portion of the data from the running dataset into the spatiotemporal feature extraction module in the model to generate a microgrid node fusion feature sequence X and a microgrid dynamic adjacency matrix A.

[0183] The second generation module is used to input the fused feature sequence X, the microgrid dynamic adjacency matrix A, and the power data in the running dataset into the attention feature extraction module in the model to generate the microgrid power supply and demand prediction value T.

[0184] The third generation module is used to input the current distributed energy equipment operating status dataset, the fusion feature sequence X, and the power prediction value T into the HHACRL-CM module in the model to generate the microgrid hybrid strategy feature sequence E.

[0185] The strategy output module is used to input the microgrid hybrid strategy feature sequence into the decision generation module in the model, and output the microgrid regulation decision to achieve precise control and global optimization of the operating status of various devices in the microgrid.

[0186] This embodiment also provides a computer device applicable to a microgrid intelligent control method based on multimodal sensing, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the microgrid intelligent control method based on multimodal sensing as proposed in the above embodiment.

[0187] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a microgrid intelligent control method based on multimodal sensing as proposed in the above embodiments.

[0188] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for implementing a microgrid intelligent control based on multimodal sensing proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0189] Based on the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0190] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-modal perception-based micro-energy network intelligent regulation method, characterized in that, The method comprises the following steps: acquiring and preprocessing a multi-source dynamic operation data set; constructing a micro-energy network intelligent regulation and control model based on multi-modal perception, inputting part of the data in the operation data set into a space-time feature extraction module in the model to generate a micro-energy network node fusion feature sequence X and a micro-energy network dynamic adjacency matrix A; inputting the fusion feature sequence X, the micro-energy network dynamic adjacency matrix A and power data in the operation data set into an attention feature extraction module in the model to generate a micro-energy network supply and demand power prediction value T; inputting a current distributed energy equipment operation state data set, the fusion feature sequence X and the power prediction value T into an HHACRL-CM module in the model to generate a micro-energy network hybrid strategy feature sequence E; inputting the micro-energy network hybrid strategy feature sequence into a decision generation module in the model to output a micro-energy network regulation and control decision, thereby realizing precise control and global optimization of the operation state of each type of equipment in the micro-energy network.

2. The multi-modal perception based microgrid intelligent regulation method of claim 1, wherein, The multi-source dynamic operation data set comprises a micro-energy network environment parameter data set, a distributed energy equipment operation state data set, a micro-energy network communication link state data set and a micro-energy network historical supply and demand power data set.

3. The multi-modal perception based microgrid intelligent regulation method of claim 2, wherein, Inputting part of the data in the operation data set into the space-time feature extraction module in the model to generate the micro-energy network node fusion feature sequence X and the micro-energy network dynamic adjacency matrix A comprises the following steps: inputting the micro-energy network environment parameter data set, the distributed energy equipment operation state data set and the micro-energy network communication link state data set into three branches of the space-time feature extraction module respectively; performing a splicing operation on the features finally output by the three branches to generate a fourteenth micro-energy network node fusion feature sequence X14; inputting X14 into a Transformer encoder to generate a fifteenth micro-energy network node fusion feature sequence X15; inputting X15 into a cosine similarity calculation submodule to generate a sixteenth micro-energy network node fusion feature sequence X16; inputting X16 into a Softmax normalization submodule to generate a seventeenth micro-energy network node fusion feature sequence X17; performing a Top-K screening operation on X17 and then inputting it into a mask processing submodule to finally generate the micro-energy network dynamic adjacency matrix A; Meanwhile, the fifteenth micro-energy network node fusion feature sequence X15 is processed through a fully connected layer to generate an eighteenth micro-energy network node fusion feature sequence X18; X18 is input into a batch normalization layer to generate a nineteenth micro-energy network node fusion feature sequence X19; X19 is subjected to a random inactivation operation to finally generate the micro-energy network node fusion feature sequence X.

4. The multi-modal perception based microgrid intelligent regulation method of claim 3, wherein, Furthermore, the micro-energy network environment parameter data set is input into the first branch of the space-time feature extraction module, is processed through a fully connected layer to complete preliminary feature transformation, is then input into a GELU activation function and then is input into a fully connected layer for further mapping, and finally is subjected to feature normalization through a layer normalization submodule to generate a first micro-energy network node fusion feature sequence X1; the first micro-energy network node fusion feature sequence X1 is input into three convolutional layers respectively for fusion, three corresponding fusion feature sequences are output respectively, and a splicing operation is performed to output a fifth micro-energy network node fusion feature sequence X5; It also includes inputting the distributed energy equipment operation status dataset into the second branch of the spatiotemporal feature extraction module, extracting local perception features through the convolutional layer, and then generating the sixth microgrid node fusion feature sequence X6 through the ReLU activation function; The fusion feature sequence X6 of the sixth microgrid node is processed by a multi-head self-attention mechanism, random deactivation dropout operation, and then processed by a fully connected layer to complete the initial feature transformation. It is then further mapped by the GELU activation function and finally normalized by the layer normalization submodule to output the fusion feature sequence X9 of the ninth microgrid node. It also includes inputting the microgrid communication link state dataset into the third branch of the spatiotemporal feature extraction module, enhancing the location features through the location encoding submodule, and then processing it through a fully connected layer to generate the 10th microgrid node fusion feature sequence X10; inputting X10 into a graph convolutional network to generate the 11th microgrid node fusion feature sequence X11; concatenating X10 and X11 to generate the 12th microgrid node fusion feature sequence X12; and passing X12 through a layer normalization submodule to output the 13th microgrid node fusion feature sequence X13.

5. The multi-modal perception based microgrid intelligent regulation method of claim 4, wherein, The fused feature sequence X, the microgrid dynamic adjacency matrix A, and the power data from the running dataset are input into the attention feature extraction module of the model to generate predicted power supply and demand values ​​for the microgrid, including: The fused feature sequence X, the microgrid dynamic adjacency matrix A, and the power data are respectively input into the three branches of the attention feature extraction module; The final power characteristic sequences output from the three branches are spliced ​​together to generate the fourteenth microgrid supply and demand power characteristic sequence T14. T14 is input into the Transformer encoder for global time series modeling to generate the power supply and demand characteristic sequence T15 of the fifteenth microgrid. T15 is input into the fully connected layer for deep nonlinear transformation, and then the ReLU activation function is used to generate the power supply and demand characteristic sequence T16 of the sixteenth microgrid. T16 is input into a random deactivation dropout layer for feature suppression processing to generate the power supply and demand feature sequence T17 of the seventeenth microgrid. T17 is input into the fully connected layer for feature mapping to generate the power supply and demand feature sequence T18 of the 18th microgrid. Input T18 into the prediction head to generate the predicted power supply and demand value T for the microgrid.

6. The multi-modal perception based microgrid intelligent regulation method of claim 5, wherein, The current distributed energy device operating status dataset, the fused feature sequence X, and the power prediction value T are input into the HHACRL-CM module in the model to generate the microgrid hybrid strategy feature sequence E, including: Input the current distributed energy equipment operating status dataset, the fused feature sequence X, and the power prediction value T into the three branches of the HHACRL-CM module, respectively; The fourth microgrid hybrid strategy feature sequence E4 output from the second branch in the HHACRL-CM module and the fifth microgrid hybrid strategy feature sequence E5 initially output from the third branch are spliced ​​together to generate the sixth microgrid hybrid strategy feature sequence E6. The E6 is subjected to deep global feature modeling processing through a layer normalization submodule and a multi-head self-attention mechanism to generate an eighth micro-energy network hybrid strategy feature sequence E8; The third micro-energy network hybrid strategy feature sequence E3 output from the first branch of the HHACRL-CM module is spliced with the E8, and subjected to full connection layer mapping and normalization processing to generate an eleventh micro-energy network hybrid strategy feature sequence E11; The E11 is input into two full connection layers for first processing and second processing, respectively, to output a fifteenth micro-energy network hybrid strategy feature sequence E15 and a sixteenth micro-energy network hybrid strategy feature sequence E16, which are spliced to obtain a seventeenth micro-energy network hybrid strategy feature sequence E17; The E17 is sequentially input into a full connection layer and a hyperbolic tangent activation function for nonlinear transformation processing to generate a nineteenth micro-energy network hybrid strategy feature sequence E19; The E15 and the E19 are spliced to finally generate a micro-energy network hybrid strategy feature sequence E.

7. The multi-modal perception based microgrid intelligent regulation method of claim 6, wherein, The micro-energy network hybrid strategy feature sequence is input into a decision generation module in the model to output a micro-energy network regulation and control decision, wherein: The micro-energy network regulation and control decision is an instruction sequence vector arranged in time sequence: I = [I1, I2,..., In] T t ,…,In] T N ​ where N denotes the total number of time steps within the future time window; I t denotes the device control parameter vector at the t-th time step, i.e., the control instructions for all the key device units in the microgrid at the t-th time step: I t = [p1, p2,..., p j ,..., p m ] wherein m is the number of microgrid devices; p j represents the specific control parameter of the jth device at the tth time step.

8. A multi-modal perception based micro-grid intelligent regulation system, applying the method of any one of claims 1-7, characterized in that, including: an acquisition module for acquiring and preprocessing a multi-source dynamic operation data set; a first generation module for constructing a micro-energy network intelligent regulation and control model based on multi-modal perception, inputting part of data in the operation data set into a spatio-temporal feature extraction module in the model to generate a micro-energy network node fusion feature sequence X and a micro-energy network dynamic adjacency matrix A; a second generation module for inputting the fusion feature sequence t, the micro-energy network dynamic adjacency matrix A and power data in the operation data set into an attention feature extraction module in the model to generate a micro-energy network supply-demand power prediction value T; a third generation module for inputting a current distributed energy equipment operation state data set, the fusion feature sequence X and the power prediction value T into a HHACRL-CM module in the model to generate a micro-energy network hybrid strategy feature sequence E; a strategy output module for inputting the micro-energy network hybrid strategy feature sequence into a decision generation module in the model to output a micro-energy network regulation and control decision, realizing precise control and global optimization of the operation states of various devices in the micro-energy network.

9. A computer device, comprising: including: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions, which realize the steps of the micro-energy network intelligent regulation and control method based on multi-modal perception according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer executable instructions, which realize the steps of the micro-energy network intelligent regulation and control method based on multi-modal perception according to any one of claims 1 to 7 when executed by a processor.

Images