A source-load collaborative multi-modal spatio-temporal graph prediction method for deep sea microgrid

Abstract

The application discloses a source-load collaborative multi-modal space-time graph prediction method for deep-sea micro-grid, comprising: obtaining source end data and load end data and preprocessing; constructing an Ocean-STGF prediction model, including a source end space-time feature extraction module, a load end multi-modal feature fusion module and a source-load collaborative interaction prediction module; the source end space-time feature extraction module extracts source end space-time features by using an adaptive dynamic graph isomorphism network and an adaptive spectral filter; the load end multi-modal feature fusion module realizes fusion by using a cross-modal attention mechanism and a contrast prototype alignment mechanism, and outputs load end comprehensive features; the source-load collaborative interaction prediction module inputs the source end space-time features and the load end comprehensive features into a gated recurrent unit for training, learns the space-time coupling of source-load states, and outputs a predicted expected vector and a predicted standard deviation vector of the source end and the load end in a future period through a probability prediction head. The application can realize high-precision collaborative prediction of power fluctuation of the deep-sea micro-grid.

Description

Technical Field

[0001] This invention relates to the field of energy management and power prediction technology for offshore microgrids, and in particular to a source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids. Background Technology

[0002] With the continuous development of the marine economy and the ongoing development and utilization of deep-sea resources, deep-sea microgrids, as a crucial infrastructure supporting power supply for offshore facilities and energy consumption for marine engineering, have received widespread attention. These systems typically consist of a power generation side comprised of multi-energy units such as wind power, photovoltaics, energy storage, and diesel engines, and an energy consumption side comprised of energy-consuming facilities such as deep-sea aquaculture and seawater desalination. Their safe and stable operation is of great significance for ensuring the reliability of power supply to island power grids. Against this backdrop, achieving high-precision prediction of power on both the source and load sides has become a key prerequisite for carrying out microgrid energy management and optimized scheduling.

[0003] However, with the continuous expansion of the scale of deep-sea microgrids and the increasing complexity of application scenarios, their power prediction problem also faces many challenges. On the one hand, the marine operating environment is complex and changeable, with frequent extreme weather events, causing the output of multi-energy units to exhibit significant nonlinear and fluctuating characteristics. At the same time, affected by factors such as wind direction changes and wave propagation, the spatial coupling relationship between units has obvious dynamic time-varying characteristics, and is superimposed with a large amount of high-frequency meteorological noise interference, increasing the difficulty of prediction modeling. On the other hand, on the energy consumption side, the energy consumption behavior of loads such as deep-sea aquaculture is not only affected by environmental factors such as water temperature, salinity, and dissolved oxygen, but is also closely related to the rhythm of aquaculture operations and the operating status of equipment, forming a complex system of multi-source heterogeneous data coupling and evolution, with significant differences between different data modes in terms of time scale, data structure, and physical meaning.

[0004] In recent years, with the development of deep learning technology, spatiotemporal sequence prediction methods have made significant progress in the field of power systems. From early recurrent neural networks to the widely used graph neural networks in recent years, these methods have demonstrated good performance in characterizing complex temporal dependencies and non-Euclidean spatial structures. However, in the specific application scenario of deep-sea microgrids, existing mainstream methods still have certain limitations. In terms of capacity modeling, most methods rely on static adjacency relationships built based on physical distance to describe the spatial topology between units, which makes it difficult to effectively characterize the dynamic evolution of the topology under extreme weather conditions. Furthermore, they lack effective suppression mechanisms for high-frequency environmental noise, which can easily lead to information distortion during feature extraction.

[0005] In the area of ​​multimodal data fusion, existing studies typically integrate continuous meteorological and hydrological variables with discrete variables such as aquaculture operation status through direct splicing or simple weighting. This fails to establish cross-modal mapping relationships within a unified representation space, making it difficult to achieve deep semantic fusion of heterogeneous data and thus limiting further improvements in model performance. Furthermore, most existing prediction methods rely primarily on deterministic point predictions, lacking the ability to characterize the uncertainty of prediction results. Under extreme sea conditions, system operation exhibits strong randomness, and relying solely on point predictions is insufficient to provide reliable risk assessment basis for scheduling decisions.

[0006] In summary, existing technologies for power prediction in deep-sea microgrids still have the following shortcomings: First, they struggle to effectively characterize the dynamically changing spatiotemporal topology in complex marine environments; second, they lack unified modeling and efficient fusion methods for multi-source heterogeneous data; and third, prediction models generally lack uncertainty quantification capabilities, making it difficult to meet the requirements for system security and robustness in practical engineering. Therefore, it is necessary to propose a source-load collaborative prediction method for deep-sea microgrids to achieve unified optimization of dynamic topology modeling, multimodal feature fusion, and uncertainty quantification, thereby improving the accuracy and reliability of microgrid power prediction in complex marine environments. Summary of the Invention

[0007] In view of this, the purpose of this invention is to propose a source-load collaborative multimodal spatiotemporal map prediction method for deep-sea microgrids. This method fully considers factors such as the dynamic evolution of unit spatial topology caused by extreme sea conditions, high-frequency environmental noise interference, and deep coupling of multi-source heterogeneous data, and can achieve high-precision collaborative prediction of power fluctuations in deep-sea microgrids.

[0008] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows: This invention provides a source-load coordinated multimodal spatiotemporal graph prediction method for deep-sea microgrids, comprising the following steps: Step 1: Obtain source and payload data and perform preprocessing; Step 2: Construct the Ocean-STGF prediction model, which includes a source-end spatiotemporal feature extraction module, a load-end multimodal feature fusion module, and a source-load collaborative interaction prediction module. Step 3: The source-end spatiotemporal feature extraction module extracts source-end spatiotemporal features based on the source-end data using an adaptive dynamic graph isomorphism network and an adaptive spectral filter; Step 4: The load-end multimodal feature fusion module uses a cross-modal attention mechanism and a contrast prototype alignment mechanism to achieve deep fusion of load-end data and outputs comprehensive load-end features. Step 5: The source-load collaborative interaction prediction module inputs the spatiotemporal features of the source end and the comprehensive features of the load end into the gated recurrent unit for joint training, learns the spatiotemporal coupling mechanism of the source-load state, and outputs the prediction expectation vector and prediction standard deviation vector of the source end and the load end for future time periods through the probability prediction head.

[0009] Furthermore, step 1 specifically includes: Step 11: Obtain source-end data and load-end data, wherein the source-end data includes raw source-end meteorological power data and topology data of multi-energy units, and the load-end data includes discrete operating rhythm mode data and continuous environmental meteorological mode data; Step 12: Perform reference time alignment and outlier processing on the source data and payload data.

[0010] Furthermore, step 3 specifically includes: Step 31: Use an adaptive dynamic graph isomorphic network to dynamically generate a comprehensive adjacency matrix to capture the dynamic spatial topological dependencies of multi-energy units; Step 32: Based on the comprehensive adjacency matrix, use a multilayer perceptron to fit the high-order spatial interaction pattern and extract the high-order spatial features; Step 33: Use an adaptive spectral filter to transform the high-order spatial features to the frequency domain, suppress high-frequency meteorological noise through frequency domain gating mask, and output the denoised source spatiotemporal features.

[0011] Furthermore, the processing procedure for the adaptive dynamic graph isomorphic network in steps 31 and 32 is as follows: Step 31 specifically includes: Step 311: Extract source-end meteorological features from the original source-end meteorological power data, extract topological features from the topology data, and generate a query matrix by passing the real-time extracted source-end meteorological features and topological features through two mapping networks containing nonlinear activation functions. AND key matrix ; wherein, the query matrix AND key matrix The formula is:

[0012]

[0013] in, For time step The query matrix, For time step The key matrix, where N is the number of spatial nodes at the source end. For the latent space feature dimension, For time step Real-time extraction of source-end meteorological and topological features. As a dimension of meteorological characteristics at the source, To generate a nonlinear mapping network for query features, A nonlinear mapping network for generating value features; Step 312: Calculate the query matrix. AND key matrix The product of these features is divided by the square root of the latent space feature dimension and then normalized using Softmax to obtain the dynamic feature adjacency matrix. Step 313: Perform a weighted summation of the dynamic feature adjacency matrix and the normalized static spatial distance matrix, using learnable balance coefficients to obtain the comprehensive adjacency matrix for the current time step; the specific formula is:

[0014] in, For the latent space feature dimension, For time step The generated composite adjacency matrix, This is the normalized static spatial distance matrix. To adjust the balance coefficient between physical constraints and data-driven weights; Step 32 specifically includes: Step 321: After obtaining the comprehensive adjacency matrix, perform nonlinear information aggregation on the nodes of the multi-energy units along the spatial dimension; Step 322: To fit the high-order spatial interaction pattern using a multilayer perceptron, the Ocean-STGF prediction model employs an adaptive dynamic graph isomorphic network to iteratively evolve the hidden state features of nodes layer by layer. The final representation formed by the hidden state features as the network depth increases and wide-area topological information is continuously aggregated is used as the high-order spatial feature. ; The layer-by-layer update rule for the hidden state features is as follows: the hidden state feature of the next layer is equal to the hidden state feature of the current layer multiplied by the learnable coefficient, and then the result is added to the comprehensive adjacency matrix multiplied by the hidden state feature of the current layer. The result is then input into the multilayer perceptron for nonlinear transformation; the specific formula is:

[0015] in, For the first Hidden state characteristics of layer nodes. For the first Hidden state characteristics of layer nodes. For the first Layer feature dimension For time step The comprehensive adjacency matrix, For the first The learnable coefficients for the degree of information retention of the control center nodes within the layer. For the first Nonlinear mapping function of the layer.

[0016] Furthermore, the processing procedure for the adaptive spectral filter in step 33 is as follows: Step 331: Perform a one-dimensional fast Fourier transform on the higher-order spatial features along the time dimension, mapping them to the complex frequency domain space to obtain the frequency domain feature tensor; the formula is as follows:

[0017] in, For the frequency domain feature tensor mapped to the complex frequency domain space, For discrete frequency variables in the complex frequency domain, For discrete frequency numbers, For a one-dimensional fast Fourier transform mapping operator, These are high-order spatial features extracted via an adaptive dynamic graph isomorphism network; Step 332: Calculate the amplitude spectrum of the frequency domain feature tensor. Subtract the threshold parameter from the amplitude spectrum and modulate it with the sigmoid function and sharpness parameter to obtain the frequency domain gated mask tensor; the formula is as follows:

[0018] in, For frequency domain gated mask tensors, The amplitude spectrum of the frequency domain characteristic tensor. The threshold parameter is used to control the noise cutoff amplitude. To adjust the sharpness parameter for mask edge smoothness; Step 333: Perform a Hadamard product between the gate mask tensor and the frequency domain feature tensor to obtain the filtered frequency domain feature tensor; the formula is:

[0019] in, This represents the filtered frequency domain feature tensor. This represents the Hadama product operator; Step 334: Remap the filtered frequency domain feature tensor back to the time domain using inverse fast Fourier transform, and introduce residual connections and layer normalization to obtain the denoised source-end spatiotemporal features; the formula is:

[0020] in, The source spatiotemporal features after denoising. For layer normalization function, It is the inverse fast Fourier transform operator.

[0021] Furthermore, step 4 specifically includes: Step 41: Perform dimension alignment and high-dimensional mapping on continuous environmental meteorological modal data and discrete operational rhythm modal data to obtain environmental modal hidden features and operational modal hidden features; Step 42: Introduce a cross-modal attention mechanism to generate a query matrix based on the hidden layer features of the environmental modality. Generate key matrix based on hidden layer features of job mode Sum matrix And based on the query matrix Key matrix Sum matrix Calculate cross-modal interaction features; Step 43: Introduce a contrastive prototype alignment mechanism, construct a set of prototype vectors in the latent space, calculate the soft assignment probability of each prototype vector by the latent features of the environment modality and the cross-modal interaction features, and construct the contrastive alignment loss through KL divergence to achieve cross-modal semantic alignment. Step 44: The environmental modal hidden layer features after comparison prototype alignment constraints are concatenated with the cross-modal interaction features. The concatenated features are then input into a nonlinear feedforward layer composed of a multilayer perceptron, and the load-side integrated features are output.

[0022] Furthermore, step 41 specifically includes: Step 411: For continuous environmental meteorological modal data, a multilayer perceptron is used for feature space transformation to obtain environmental modal hidden layer features under a unified hidden space dimension; the formula is as follows:

[0023] in, The input feature tensor represents the continuous environmental meteorological modal data at the load end. Indicates the length of the time step. Representing the dimensions of meteorological variables, This represents the hidden layer features of the environmental modalities after processing by a multilayer perceptron. This represents the dimension of the unified implicit space. This represents a nonlinear mapping function for continuous environmental meteorological modal data; Step 412: For discrete job rhythm modal data, a learnable embedding dictionary is used to transform the feature space to obtain the job modal hidden layer features under a unified latent space dimension; the formula is as follows:

[0024] in, The input feature tensor represents the discrete operational rhythm modal data at the load end. Represents the discrete feature dimension. This represents the hidden layer features of the job modality after learning the learnable embedding dictionary mapping. This represents a learnable embedding dictionary mapping constructed for discrete job rhythm modal data; Step 42 specifically includes: Step 421: Generate a query matrix from the hidden features of the environmental modality through linear transformation. The formula is:

[0025] in, The query matrix generated for the hidden layer features of the environmental modality. This is a learnable linear mapping weight matrix for the hidden layer features of environmental modalities; Step 422: Generate the key matrix from the hidden layer features of the working mode through two sets of independent linear transformations. AND-value matrix The formula is:

[0026]

[0027] in, The key matrix generated for the hidden layer features of the job mode. The value matrix generated for the hidden layer features of the operation mode. This is a learnable linear mapping weight matrix for the hidden layer features of the task modality, used to generate state matching labels. This is a learnable linear mapping weight matrix for the hidden layer features of the operational modality, used to extract features related to actual energy consumption. The dimension of latent space features; Step 423: Calculate the query matrix AND key matrix The scaled dot product attention score, after Softmax normalization, is compared with the value matrix. Weighted summation yields the cross-modal interaction feature, which characterizes the dynamic nonlinear relationship between environmental variables and work rhythm; the formula is as follows:

[0028] in, These are cross-modal interaction features obtained through attention-weighted fusion. Step 43 specifically includes: Step 431: Construct a set of prototype vectors in the latent space, each prototype vector representing a typical coupling mode between sea state and operating condition. Step 432: Calculate the soft assignment probabilities of the environment modality hidden layer features and cross-modal interaction features to each prototype vector. The calculation formula is as follows:

[0029]

[0030] in, For the first The environmental modal hidden layer features at the time step are related to the _th The probability of belonging to each prototype vector. For the first Cross-modal interaction features at the time step for the The probability of belonging to each prototype vector. For the first Hidden layer features of environmental modalities at each time step For the first Cross-modal interaction features at each time step For the first One prototype vector, For the first One prototype vector, The number of prototype vectors. The temperature coefficient is used to adjust the smoothness of the probability distribution; Step 433: Calculate the KL divergence between the attribution probability distribution of the hidden layer features of the environment modality and the attribution probability distribution of the cross-modal interaction features. Use this KL divergence value as the contrastive alignment loss to constrain the semantic consistency of the two modalities in the prototype space. The formula for calculating the contrastive alignment loss is as follows:

[0031] in, For contrastive alignment loss used for cross-modal semantic alignment, For the first The probability of the hidden layer features of the environmental modality at each time step belonging to the complete prototype vector. For the first The probability of attribution of cross-modal interaction features to the complete prototype vector at each time step. The KL divergence between two attribution probability distributions. Indicates the length of the time step; The specific formula for the comprehensive feature of the load end in step 44 is as follows:

[0032] in, The combined features of the load end after joint mapping The output feature dimension of the fused features, This is a tensor concatenation operation along the feature dimension. This is a multimodal feature fusion network operator.

[0033] Furthermore, step 5 specifically includes: Step 51: Concatenate the spatiotemporal features of the source end with the comprehensive features of the load end, and use a gated cyclic unit to learn the dynamic interaction relationship between the source and load states; Step 52: After decoding the terminal hidden state feature vector output by the gated recurrent unit through the probability prediction head and the MLP, the expected prediction vector and the standard deviation prediction vector of the source end and the load end in the future time period are obtained. Step 53: Construct a global optimization objective function and optimize the model parameters of the Ocean-STGF prediction model based on the global optimization objective function.

[0034] Furthermore, step 51 specifically includes: Step 511: At each time step, the source-load collaborative interaction prediction module concatenates the spatiotemporal features of the source end and the comprehensive features of the load end along the feature dimension to obtain the global input feature tensor, with the formula as follows:

[0035] in, For time step The concatenated global input feature tensor For time step The spatiotemporal characteristics of the source end, For time step The comprehensive characteristics of the load end; Step 512: Input the global input feature tensor into the gated recurrent unit. The gating mechanism and the hidden state feature vector update process are calculated as follows:

[0036]

[0037]

[0038]

[0039] in, For time step Update the gate feature vector. For time step The reset gate feature vector, This represents the candidate hidden state vector. For the current time step The hidden state feature vector. This represents the hidden state feature vector from the previous time step. To update the gate weight matrix, To reset the weight matrix of the gate, The weight matrix is ​​the global input feature tensor. Let be the weight matrix from the hidden state to the update gate. Let be the weight matrix from the hidden state to the reset gate. Let be the weight matrix of the hidden states. To reset the gate's bias vector, To update the gate's bias vector, This is the bias vector for the hidden state. It is the Sigmoid activation function. The hyperbolic tangent nonlinear activation function; Step 52 specifically includes: Step 521: The probability prediction head adopts a dual-branch fully connected layer structure, with an expectation branch and a standard deviation branch; Step 522: Map the terminal hidden state feature vector output by the gated recurrent unit to the prediction expectation and uncertainty parameter space of the future time period; Step 523: Decode the terminal hidden state feature vector using two independent MLPs. The expectation branch maps the terminal hidden state feature vector output by the gated recurrent unit to the predicted expected vectors of source-side capacity and load-side energy consumption for multiple future time steps. The standard deviation branch maps identical terminal hidden state feature vectors to the predicted standard deviation vectors of the source and load sides. The calculation formula is as follows:

[0040]

[0041] in, The expected prediction vector at the source end. The source-side prediction standard deviation vector. Let the predicted expected vector be the vector at the load end. Let be the predicted standard deviation vector of the load end. This is the terminal hidden state feature vector output by the gated recurrent unit. For the decoding network of the desired branch, Decoding network with standard deviation branch, To ensure that the standard deviation is a positive real number, a nonlinear activation function is used; Step 53 specifically includes: Step 531: Using the joint loss function based on Gaussian negative log-likelihood loss, calculate the Gaussian negative log-likelihood loss of the source prediction task and the Gaussian negative log-likelihood loss of the load prediction task respectively. The Gaussian negative log-likelihood loss of the source prediction task The calculation formula is as follows:

[0042] in, For the source-end prediction task, use the Gaussian negative log-likelihood loss. To predict the time step length, For the source end in the future The time step and the first The actual production capacity of each node For the source end in the future The time step and the first The expected value of each node. For the source end in the future The time step and the first The standard deviation of the prediction for each node; The Gaussian negative log-likelihood loss of the load prediction task The calculation formula is as follows:

[0043] in, The Gaussian negative log-likelihood loss is used for the prediction task of the payload. To predict the time step length, For the future of the Netherlands The time step and the first The actual energy consumption of each node For the future of the Netherlands The time step and the first The expected value of each node. For the future of the Netherlands The time step and the first The standard deviation of the prediction for each node; Step 532: Construct a global optimization objective function containing a regularization term based on the Gaussian negative log-likelihood loss of the source-end prediction task and the Gaussian negative log-likelihood loss of the load-end prediction task. Obtain the globally optimal parameters of the Ocean-STGF prediction model by solving for the minimum value of this global optimization objective function. The specific optimization formula is as follows:

[0044] in, For the globally optimal parameter set, The set of learnable parameters for the Ocean-STGF prediction model. This is the balance coefficient between the source and payload task losses. for Regularization hyperparameters, For parameters Norm penalty term.

[0045] Furthermore, step 5 is followed by: The predicted expected vectors and predicted standard deviation vectors of the source and load sides are input into the upper-level scheduling and control module. Combined with the prediction uncertainty information, the power output allocation on the generation side, the power response on the load side, and the energy storage charging and discharging strategy are dynamically adjusted to achieve safe and stable operation of the microgrid.

[0046] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: 1. In the scenario of deep-sea microgrids, based on the original source-end meteorological power data, operational rhythm mode data, environmental meteorological mode data, and spatial topological relationships between units, the future evolution trend of source-end power generation and load-end energy consumption within the deep-sea microgrid is determined. A source-end spatiotemporal feature extraction module composed of an adaptive dynamic graph isomorphic network and a spectral filter is constructed. Feature-driven adaptive dynamic graph isomorphic network and adaptive spectral filter are introduced at the source end. Adjacency matrices are dynamically generated through feature-driven methods to capture the nonlinear spatial dependencies of multi-energy units, and a frequency domain truncation mechanism is used to effectively suppress high-frequency meteorological noise interference.

[0047] 2. Construct a multimodal feature fusion module at the load end that combines cross-modal attention and contrast prototype alignment. This module maps continuous environmental meteorological modal data and discrete operational rhythm modal data to a unified representation space. Furthermore, it promotes the consistency of cross-modal features through contrast alignment constraints, thereby achieving deep fusion of multi-source heterogeneous data. 3. A source-load collaborative prediction module based on a gated cyclic unit and a dual-branch probabilistic prediction head is constructed. This module jointly models the spatiotemporal characteristics of the source end and the comprehensive characteristics of the load end in the time series dimension. The probabilistic prediction head outputs the expected estimates and standard deviations of the source end capacity and load end energy consumption sequences for future periods, achieving high-precision collaborative prediction of microgrid capacity and load in complex marine environments. This significantly improves the power prediction accuracy in complex marine environments, provides reliable support for microgrid energy management and scheduling decisions, and obtains source-load collaborative prediction results that combine expected estimates and uncertainty quantification. It also solves the problem of spatial topological mutations and deep coupling of multi-source heterogeneous data under extreme sea conditions. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0049] Figure 1 This is an execution flowchart of a source-load coordinated multimodal spatiotemporal graph prediction method for deep-sea microgrids provided in an embodiment of the present invention.

[0050] Figure 2 This is a diagram of the Ocean-STGF prediction model architecture provided in an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of ADGIN dynamic graph isomorphic network spatial information aggregation provided in an embodiment of the present invention.

[0052] Figure 4 This is a schematic diagram of the ASF adaptive spectral filtering denoising mechanism provided in an embodiment of the present invention. Detailed Implementation

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Figure 1 This paper demonstrates the macroscopic operational mechanism of the Ocean-STGF (Ocean-Spatio Temporal Graph Forecasting, i.e., source-load collaborative multimodal spatiotemporal graph prediction for deep-sea scenarios) prediction model. This process mainly includes four stages: multi-source heterogeneous data preprocessing, multimodal spatiotemporal feature extraction and fusion, collaborative model training, and uncertainty quantification prediction. In the preprocessing stage, baseline time alignment and outlier handling are performed on the raw source-end meteorological power data and load-end data. Subsequently, the model independently extracts and fuses features from the source-end spatial topology and load-end multimodal features. Next, the model inputs the extracted and fused features into the source-load collaborative interaction prediction module for joint training, learning the spatiotemporal coupling mechanism of source-load states. Finally, the probability prediction head outputs the expected prediction vector and the standard deviation vector for future time periods.

[0055] The core structure of the Ocean-STGF prediction model consists of a rigorous three-module architecture: a source-end spatiotemporal feature extraction module, a payload-end multimodal feature fusion module, and a source-payload collaborative interaction prediction module. The specific topology is as follows: Figure 2 As shown.

[0056] The source-end spatiotemporal feature extraction module employs a feature-driven Adaptive Dynamic Graph Isomorphism Network (AD-GIN) and an Adaptive Spectral Filter (ASF) to learn the spatiotemporal features of multi-energy offshore generator sets. AD-GIN dynamically generates a comprehensive adjacency matrix from real-time input meteorological features, effectively extracting the complex nonlinear spatial dependencies between generator sets that vary with sea state. ASF suppresses high-frequency meteorological noise through a threshold gating mechanism in the frequency domain. This source-end spatiotemporal feature extraction module improves the smoothness of features in the temporal dimension while preserving spatial features.

[0057] The load-side multimodal feature fusion module introduces a cross-modal attention and contrastive prototype alignment cascade architecture to characterize and fuse multi-source heterogeneous data from the load end. This module first utilizes a cross-modal attention mechanism to capture the direct interaction between environmental and operational features at micro-time steps. Then, through learnable prototype vectors, the interacting features are mapped to a unified physical latent space, and the alignment and consistency of cross-modal feature distributions are enhanced through symmetric Kullback–Leibler Divergence (KL) constraints. Finally, the aligned features are fused through concatenation and a fully connected layer to generate comprehensive load-side features.

[0058] The source-load collaborative interaction prediction module utilizes a gated recurrent unit (GRU) to aggregate spatiotemporal features from the source end and comprehensive features from the load end, learning the dynamic interaction relationship between source and load states. Finally, the hidden state output of the GRU is passed through a two-branch fully connected layer to generate the expected value and standard deviation of the output force at future time steps, achieving uncertainty quantification prediction.

[0059] Through the aforementioned modular design and collaborative working mechanism, the Ocean-STGF prediction model combines AD-GIN's feature-driven dynamic spatial modeling capabilities, ASF's frequency domain denoising mechanism, cross-modal attention and contrastive prototype cascaded feature fusion, and GRU's spatiotemporal interaction, effectively capturing the complex spatiotemporal dependencies and cross-modal mapping features in offshore microgrid data. This design enables the model to simultaneously consider dynamic unit spatial topology and multi-dimensional source-load patterns, thereby improving the accuracy and robustness of source-load collaborative prediction under extreme sea conditions.

[0060] The present invention provides a source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids, comprising the following steps: Step 1: Multi-source heterogeneous data preprocessing: Acquire source data and payload data and perform preprocessing; In this embodiment, step 1 specifically includes: Step 11: Obtain source-end data and load-end data, wherein the source-end data includes raw source-end meteorological power data and topology data of multi-energy units, and the load-end data includes discrete operating rhythm mode data and continuous environmental meteorological mode data; Step 12: Perform reference time alignment and outlier processing on the source data and payload data.

[0061] Step 2, Model Construction: Construct the Ocean-STGF prediction model, which includes a source-end spatiotemporal feature extraction module, a load-end multimodal feature fusion module, and a source-load collaborative interaction prediction module. Step 3, Multimodal Spatiotemporal Feature Extraction: The source-end spatiotemporal feature extraction module extracts source-end spatiotemporal features based on source-end data using an adaptive dynamic graph isomorphism network and an adaptive spectral filter; The output characteristics of multi-energy units in offshore microgrids are deeply coupled with the non-stationary evolution of marine meteorology and the dynamic wake effect between units. To accurately extract the nonlinear spatial topology evolution law between units and suppress high-frequency meteorological noise, this invention designs a source-end spatiotemporal feature extraction module consisting of AD-GIN and ASF cascaded together.

[0062] In this embodiment, step 3 specifically includes: Step 31, Adaptive Dynamic Adjacency Matrix Construction: An adaptive dynamic graph isomorphic network is used to dynamically generate a comprehensive adjacency matrix to capture the dynamic spatial topological dependencies of multi-energy units; specifically including: Step 311: Extract source-end meteorological features from the original source-end meteorological power data, extract topological features from the topology data, and generate a query matrix by passing the real-time extracted source-end meteorological features and topological features through two mapping networks containing nonlinear activation functions. AND key matrix ; wherein, the query matrix AND key matrix The formula is:

[0063]

[0064] in, For time step The query matrix, For time step The key matrix, where N is the number of spatial nodes at the source end. For the latent space feature dimension, For time step Real-time extraction of source-end meteorological and topological features. As a dimension of meteorological characteristics at the source, To generate a nonlinear mapping network for query features, A nonlinear mapping network for generating key features; Step 312: Calculate the query matrix. AND key matrix The product of these features is divided by the square root of the latent space feature dimension and then normalized using Softmax to obtain the dynamic feature adjacency matrix. Step 313: Perform a weighted summation of the dynamic feature adjacency matrix and the normalized static spatial distance matrix, using learnable balance coefficients to obtain the comprehensive adjacency matrix for the current time step; the specific formula is:

[0065] in, For the latent space feature dimension, For time step The generated composite adjacency matrix, This is the normalized static spatial distance matrix. To adjust the balance coefficient between physical constraints and data-driven weights, this mechanism decouples model parameters from node size through feature mapping, thereby adapting to changes in the number of nodes. Step 32, Dynamic Graph Isomorphic Space Feature Extraction: Based on the comprehensive adjacency matrix, a multilayer perceptron is used to fit the high-order spatial interaction pattern and extract high-order spatial features; specifically including: Step 321: After obtaining the comprehensive adjacency matrix, perform nonlinear information aggregation on the nodes of the multi-energy units along the spatial dimension; Step 322: To fit the high-order spatial interaction pattern using a Multilayer Perceptron (MLP), the Ocean-STGF prediction model employs an adaptive dynamic graph isomorphic network to iteratively evolve the hidden state features of nodes layer by layer. The final representation formed by the hidden state features as the network depth increases and wide-area topological information is continuously aggregated is used as the high-order spatial feature. ; The layer-by-layer update rule for the hidden state features is as follows: the hidden state feature of the next layer is equal to the hidden state feature of the current layer multiplied by the learnable coefficient, and then the result is added to the comprehensive adjacency matrix multiplied by the hidden state feature of the current layer. The result is then input into the multilayer perceptron for nonlinear transformation; the specific formula is:

[0066] in, For the first Hidden state characteristics of layer nodes. For the first Hidden state characteristics of layer nodes. For the first Layer feature dimension For time step The comprehensive adjacency matrix, For the first The learnable coefficients for the degree of information retention of the control center nodes within the layer. For the first Nonlinear mapping function of the layer. Figure 3 The paper demonstrates the weighted aggregation process of node features of multi-energy units under the action of an adaptive dynamic adjacency matrix, and explains the process of nonlinear mapping and spatial feature dimensionality enhancement of aggregated features through a multilayer perceptron.

[0067] Step 33, Adaptive Spectral Filtering Noise Reduction Mechanism: Adaptive spectral filter is used to transform high-order spatial features to the frequency domain, and frequency domain gating mask is used to suppress high-frequency meteorological noise, outputting the denoised source spatiotemporal features; Meteorological and hydrological sensor data from offshore microgrids are typically accompanied by severe and high-frequency random fluctuations and measurement noise over time. Directly utilizing spatially aggregated features for time series modeling can easily lead to overfitting to these random disturbances, thus affecting the generalization performance of predictions. Therefore, this invention introduces ASF (Automatic Spatial Feature Extraction) after dynamic spatial feature extraction. Its denoising mechanism is as follows: Figure 4 As shown.

[0068] In this embodiment, step 33 specifically includes: Step 331: Perform a one-dimensional Fast Fourier Transform (FFT) on the higher-order spatial features along the time dimension, mapping them to the complex frequency domain space to obtain the frequency domain feature tensor; the formula is as follows:

[0069] in, For the frequency domain feature tensor mapped to the complex frequency domain space, For discrete frequency variables in the complex frequency domain, For discrete frequency numbers, For a one-dimensional fast Fourier transform mapping operator, These are high-order spatial features extracted via an adaptive dynamic graph isomorphism network; Step 332: To identify and suppress high-frequency noise components, calculate the amplitude spectrum of the frequency domain feature tensor. Subtract the threshold parameter from the amplitude spectrum and modulate it with the sigmoid function and sharpness parameter to obtain the frequency domain gated mask tensor; the formula is as follows:

[0070] in, For frequency domain gated mask tensors, The amplitude spectrum of the frequency domain characteristic tensor. The threshold parameter is used to control the noise cutoff amplitude. To adjust the sharpness parameter for mask edge smoothness; Step 333: Perform a Hadamard product between the gate mask tensor and the frequency domain feature tensor to obtain the filtered frequency domain feature tensor; the formula is:

[0071] in, This represents the filtered frequency domain feature tensor. This represents the Hadama product operator; Step 334: The filtered frequency domain feature tensor is remapped back to the time domain using Inverse Fast Fourier Transform (IFFT), and residual connections and layer normalization are introduced to prevent gradient vanishing in deep networks, resulting in the denoised source-end spatiotemporal features; the formula is:

[0072] in, The source spatiotemporal features after denoising. For layer normalization function, This is the inverse fast Fourier transform operator. The residual connection effectively preserves the original nonlinear spatial topological features, ensuring the training stability of the model under extreme sea conditions.

[0073] Step 4, Multimodal Fusion: The load-end multimodal feature fusion module uses a cross-modal attention mechanism and a contrast prototype alignment mechanism to achieve deep fusion of load-end data and outputs comprehensive load-end features. The energy consumption of deep-sea aquaculture facilities is influenced by both natural environmental factors and human operational rhythms. Environmental variables typically exhibit continuous numerical distributions, while operational rhythms display discrete categorical characteristics. Directly concatenating features at the raw data level can lead to semantic inconsistencies and feature interference between different modalities. Therefore, this invention designs a payload-side multimodal feature fusion module consisting of cross-modal attention and contrastive prototype alignment. The computational process of this payload-side multimodal feature fusion module includes four stages: modality-independent feature encoding, cross-modal attention interaction mechanism, latent space contrastive prototype alignment, and joint feature mapping.

[0074] In this embodiment, step 4 specifically includes: Step 41, Modality-Independent Feature Encoding: Dimensional alignment and high-dimensional mapping are performed on continuous environmental meteorological modal data and discrete operational rhythm modal data to obtain environmental modal hidden features and operational modal hidden features; specifically including: Step 411: To avoid mutual interference between heterogeneous features during the feature extraction stage, the model first performs independent dimensional alignment and high-dimensional mapping between continuous and discrete modes. For continuous environmental meteorological modal data, a multilayer perceptron (MLP) is used for feature space transformation to obtain environmental modal hidden layer features under a unified latent space dimension; the formula is as follows:

[0075] in, The input feature tensor represents the continuous environmental meteorological modal data at the load end. Indicates the length of the time step. Representing the dimensions of meteorological variables, This represents the hidden layer features of the environmental modalities after processing by a multilayer perceptron. This represents the dimension of the unified implicit space. This represents a nonlinear mapping function for continuous environmental meteorological modal data; Step 412: For discrete job rhythm modal data, a learnable embedding dictionary is used to transform the feature space to obtain the job modal hidden layer features under a unified latent space dimension; the formula is as follows:

[0076] in, The input feature tensor represents the discrete operational rhythm modal data at the load end. Represents the discrete feature dimension. This represents the hidden layer features of the job modality after learning the learnable embedding dictionary mapping. This represents a learnable embedding dictionary mapping constructed for discrete job rhythm modal data; Step 42, Cross-modal attention interaction mechanism: To model the interaction between environmental features and work rhythm, a cross-modal attention mechanism is introduced to generate a query matrix based on the hidden features of the environmental modality. Generate key matrix based on hidden layer features of job mode Sum matrix And based on the query matrix Key matrix Sum matrix Calculate cross-modal interaction features; this step enables the model to adaptively capture the non-linear relationship between environmental variables and job status over time. In this embodiment, step 42 specifically includes: Step 421: Generate a query matrix from the hidden features of the environmental modality through linear transformation. The formula is:

[0077] in, The query matrix generated for the hidden layer features of the environmental modality. This is a learnable linear mapping weight matrix for the hidden layer features of environmental modalities; Step 422: Generate the key matrix from the hidden layer features of the working mode through two sets of independent linear transformations. AND-value matrix The formula is:

[0078]

[0079] in, The key matrix generated for the hidden layer features of the job mode. The value matrix generated for the hidden layer features of the operation mode. A learnable linear mapping weight matrix for the hidden layer features of the task modality is used to generate state matching labels. This is a learnable linear mapping weight matrix for the hidden layer features of the operational modality, used to extract features related to actual energy consumption. The dimension of latent space features; Step 423: Calculate the query matrix AND key matrix The scaled dot product attention score, after Softmax normalization, is compared with the value matrix. Weighted summation yields the cross-modal interaction feature, which characterizes the dynamic nonlinear relationship between environmental variables and work rhythm; the formula is as follows:

[0080] in, These are cross-modal interaction features obtained through attention-weighted fusion. Step 43: Latent Space Contrastive Prototype Alignment: In non-stationary environments such as extreme sea states in the deep ocean, relying solely on cross-modal attention mechanisms can easily lead the model to overemphasize local abrupt changes, thereby reducing its generalization ability in long-tail conditions. To enhance the robustness of multimodal feature fusion, this invention introduces a Contrastive Prototype Alignment (CPA) mechanism as a regularization constraint after the attention interaction layer. A set of prototype vectors is constructed in the latent space, and the soft assignment probabilities of environmental modality latent layer features and cross-modal interaction features to each prototype vector are calculated. A contrastive alignment loss is constructed using KL divergence to achieve cross-modal semantic alignment; specifically including: Step 431: Construct a set of prototype vectors in the latent space, each prototype vector representing a typical coupling mode between sea state and operating condition. Step 432: Calculate the soft assignment probabilities of the environment modality hidden layer features and cross-modal interaction features to each prototype vector. The calculation formula is as follows:

[0081]

[0082] in, For the first The environmental modal hidden layer features at the time step are related to the _th The probability of belonging to each prototype vector. For the first Cross-modal interaction features at the time step for the The probability of belonging to each prototype vector. For the first Hidden layer features of environmental modalities at each time step For the first Cross-modal interaction features at each time step For the first One prototype vector, For the first One prototype vector, The number of prototype vectors. The temperature coefficient is used to adjust the smoothness of the probability distribution; Step 433: Calculate the KL divergence between the attribution probability distribution of the hidden layer features of the environment modality and the attribution probability distribution of the cross-modal interaction features. Use this KL divergence value as the contrastive alignment loss to constrain the semantic consistency of the two modalities in the prototype space, so that the probability distributions of different modal features in the prototype space remain consistent. The formula for calculating the contrastive alignment loss is as follows:

[0083] in, For contrastive alignment loss used for cross-modal semantic alignment, For the first The probability of the hidden layer features of the environmental modality at each time step belonging to the complete prototype vector. For the first The probability of attribution of cross-modal interaction features to the complete prototype vector at each time step. The KL divergence between two attribution probability distributions. The time step length is represented; after semantic alignment and macro-regularization constraints are completed in the latent space, the two types of modal features achieve consistency in the underlying physical representation.

[0084] Step 44, Feature Joint Mapping: The hidden features of the environmental modality after comparison prototype alignment constraints are concatenated with the cross-modal interaction features. The concatenated features are then input into a nonlinear feedforward layer composed of a multilayer perceptron, and the output is the load-end integrated feature. The specific formula for the load-end integrated feature is as follows:

[0085] in, The combined features of the load end after joint mapping The output feature dimension of the fused features, This is a tensor concatenation operation along the feature dimension. This is a multimodal feature fusion network operator. Combining cross-modal attention and prototype alignment mechanisms, the load-side multimodal feature fusion module effectively extracts high-order features reflecting multidimensional energy use behavior, providing structurally consistent load-side inputs for subsequent collaborative prediction.

[0086] Step 5, Model Training and Prediction: The source-load collaborative interaction prediction module inputs the spatiotemporal features of the source end and the comprehensive features of the load end into the gated recurrent unit for joint training, learns the spatiotemporal coupling mechanism of the source-load state, and outputs the prediction expectation vector and prediction standard deviation vector of the source end and the load end for future time periods through the probability prediction head.

[0087] In this embodiment, step 5 specifically includes: Step 51, Source-Load Global Spatiotemporal Feature Interaction: The source-end spatiotemporal features and the load-end comprehensive features are concatenated, and a gated recurrent unit is used to learn the dynamic interaction relationship between the source and load states; specifically including: Step 511: At each time step, the source-load collaborative interaction prediction module concatenates the spatiotemporal features of the source end and the comprehensive features of the load end along the feature dimension to obtain the global input feature tensor, with the formula as follows:

[0088] in, For time step The concatenated global input feature tensor For time step The spatiotemporal characteristics of the source end, For time step The comprehensive characteristics of the load end; Step 512: Input the global input feature tensor into the gated recurrent unit (GRU). The gating mechanism and the hidden state feature vector update process are calculated as follows:

[0089]

[0090]

[0091]

[0092] in, For time step Update the gate feature vector. For time step The reset gate feature vector, This represents the candidate hidden state vector. For the current time step The hidden state feature vector. This represents the hidden state feature vector from the previous time step. To update the gate weight matrix, To reset the weight matrix of the gate, The weight matrix is ​​the global input feature tensor. Let be the weight matrix from the hidden state to the update gate. Let be the weight matrix from the hidden state to the reset gate. Let be the weight matrix of the hidden states. To reset the gate's bias vector, To update the gate's bias vector, This is the bias vector for the hidden state. It is the Sigmoid activation function. The hyperbolic tangent nonlinear activation function; Step 52, Dual-branch probabilistic prediction head: The probabilistic prediction head decodes the terminal hidden state feature vector output by the gated recurrent unit using an MLP to obtain the expected prediction vector and standard deviation vector for the source and load ends in the future time period. This is used to quantify the prediction uncertainty under extreme sea states; specifically including: Step 521: The probability prediction head adopts a dual-branch fully connected layer structure, with an expectation branch and a standard deviation branch; Step 522: Map the terminal hidden state feature vector output by the gated recurrent unit to the prediction expectation and uncertainty parameter space of the future time period; Step 523: Decode the terminal hidden state feature vector using two independent MLPs. The expectation branch maps the terminal hidden state feature vector output by the gated recurrent unit to the predicted expected vectors of source-side capacity and load-side energy consumption for multiple future time steps. The standard deviation branch maps identical terminal hidden state feature vectors to the predicted standard deviation vectors of the source and load sides. The calculation formula is as follows:

[0093]

[0094] in, The expected prediction vector at the source end. The source-side prediction standard deviation vector. Let the predicted expected vector be the vector at the load end. Let be the predicted standard deviation vector of the load end. This is the terminal hidden state feature vector output by the gated recurrent unit. For the decoding network of the desired branch, Decoding network with standard deviation branch, To ensure that the standard deviation is a positive real number, a nonlinear activation function is used; Step 53, Jointly Optimize the Objective Function: Construct a global objective function and optimize the model parameters of the Ocean-STGF prediction model based on the global objective function; specifically including: Step 531: Using the joint loss function based on Gaussian negative log-likelihood (NLL) loss, calculate the Gaussian negative log-likelihood loss of the source prediction task and the Gaussian negative log-likelihood loss of the payload prediction task respectively. The Gaussian negative log-likelihood loss of the source prediction task The calculation formula is as follows:

[0095] in, For the source-end prediction task, use the Gaussian negative log-likelihood loss. To predict the time step length, For the source end in the future The time step and the first The actual production capacity of each node For the source end in the future The time step and the first The expected value of each node. For the source end in the future The time step and the first The standard deviation of the prediction for each node; The Gaussian negative log-likelihood loss of the load prediction task The calculation formula is as follows:

[0096] in, The Gaussian negative log-likelihood loss is used for the prediction task of the payload. To predict the time step length,

[0097] For the future of the Netherlands The time step and the first The actual energy consumption of each node For the future of the Netherlands The time step and the first The expected value of each node. For the future of the Netherlands The time step and the first The standard deviation of the prediction for each node; Step 532: To balance gradient backpropagation between the source and load tasks and suppress model overfitting, a global optimization objective function containing a regularization term is constructed based on the Gaussian negative log-likelihood loss of the source prediction task and the Gaussian negative log-likelihood loss of the load prediction task. The global optimal parameters of the Ocean-STGF prediction model are obtained by solving for the minimum value of this global optimization objective function. The specific optimization formula is as follows:

[0098] in, For the globally optimal parameter set, The set of learnable parameters for the Ocean-STGF prediction model. This is the balance coefficient between the source and payload task losses. for Regularization hyperparameters, For parameters Norm penalty term.

[0099] By learning the complex interactions between source and load characteristics, GRU can effectively capture the dynamic collaborative patterns of deep-sea microgrids in the spatiotemporal dimensions. (GRU's terminal hidden state feature vector) This method incorporates rich source-load coordination information learned from the global input sequence. This information is then input into a two-branch probabilistic prediction head for nonlinear mapping, and combined with joint optimization using a Gaussian negative log-likelihood loss function. Finally, a source-load coordination prediction result that balances prediction expectation and confidence standard deviation is obtained. This source-load coordination prediction result can be further applied to scenarios such as dynamic scheduling of source and load in deep-sea microgrids, optimized control of energy storage, and early warning of operational risks, providing decision-making support for energy management systems.

[0100] Step 6: Input the predicted expected vectors and predicted standard deviation vectors of the source and load sides into the upper-level scheduling and control module, and dynamically adjust the power output allocation on the generation side, the power response on the load side, and the energy storage charging and discharging strategy in combination with the prediction uncertainty information to achieve safe and stable operation of the microgrid.

[0101] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids, characterized in that, Includes the following steps: Step 1: Obtain source and payload data and perform preprocessing; Step 2: Construct the Ocean-STGF prediction model, which includes a source-end spatiotemporal feature extraction module, a load-end multimodal feature fusion module, and a source-load collaborative interaction prediction module. Step 3: The source-end spatiotemporal feature extraction module extracts source-end spatiotemporal features based on the source-end data using an adaptive dynamic graph isomorphism network and an adaptive spectral filter; Step 4: The load-end multimodal feature fusion module uses a cross-modal attention mechanism and a contrast prototype alignment mechanism to achieve deep fusion of load-end data and outputs comprehensive load-end features. Step 5: The source-load collaborative interaction prediction module inputs the spatiotemporal features of the source end and the comprehensive features of the load end into the gated recurrent unit for joint training, learns the spatiotemporal coupling mechanism of the source-load state, and outputs the prediction expectation vector and prediction standard deviation vector of the source end and the load end for future time periods through the probability prediction head.

2. The source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids as described in claim 1, characterized in that, Step 1 specifically includes: Step 11: Obtain source-end data and load-end data, wherein the source-end data includes raw source-end meteorological power data and topology data of multi-energy units, and the load-end data includes discrete operating rhythm mode data and continuous environmental meteorological mode data; Step 12: Perform reference time alignment and outlier processing on the source data and payload data.

3. The source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids as described in claim 2, characterized in that, Step 3 specifically includes: Step 31: Use an adaptive dynamic graph isomorphic network to dynamically generate a comprehensive adjacency matrix to capture the dynamic spatial topological dependencies of multi-energy units; Step 32: Based on the comprehensive adjacency matrix, use a multilayer perceptron to fit the high-order spatial interaction pattern and extract the high-order spatial features; Step 33: Use an adaptive spectral filter to transform the high-order spatial features to the frequency domain, suppress high-frequency meteorological noise through frequency domain gating mask, and output the denoised source spatiotemporal features.

4. The source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids as described in claim 3, characterized in that, The processing steps for the adaptive dynamic graph isomorphic network in steps 31 and 32 are as follows: Step 31 specifically includes: Step 311: Extract source-end meteorological features from the original source-end meteorological power data, extract topological features from the topology data, and generate a query matrix by passing the real-time extracted source-end meteorological features and topological features through two mapping networks containing nonlinear activation functions. AND key matrix ; wherein, the query matrix AND key matrix The formula is: in, For time step The query matrix, For time step The key matrix, where N is the number of spatial nodes at the source end. For the latent space feature dimension, For time step Real-time extraction of source-end meteorological and topological features. As a dimension of meteorological characteristics at the source, To generate a nonlinear mapping network for query features, A nonlinear mapping network for generating key features; Step 312: Calculate the query matrix. AND key matrix The product of these features is divided by the square root of the latent space feature dimension and then normalized using Softmax to obtain the dynamic feature adjacency matrix. Step 313: Perform a weighted summation of the dynamic feature adjacency matrix and the normalized static spatial distance matrix, using learnable balance coefficients to obtain the comprehensive adjacency matrix for the current time step; the specific formula is: in, For the latent space feature dimension, For time step The generated composite adjacency matrix, This is the normalized static spatial distance matrix. To adjust the balance coefficient between physical constraints and data-driven weights; Step 32 specifically includes: Step 321: After obtaining the comprehensive adjacency matrix, perform nonlinear information aggregation on the nodes of the multi-energy units along the spatial dimension; Step 322: To fit the high-order spatial interaction pattern using a multilayer perceptron, the Ocean-STGF prediction model employs an adaptive dynamic graph isomorphic network to iteratively evolve the hidden state features of nodes layer by layer. The final representation formed by the hidden state features as the network depth increases and wide-area topological information is continuously aggregated is used as the high-order spatial feature. ; The layer-by-layer update rule for the hidden state features is as follows: the hidden state feature of the next layer is equal to the hidden state feature of the current layer multiplied by the learnable coefficient, and then the result is added to the comprehensive adjacency matrix multiplied by the hidden state feature of the current layer. The result is then input into the multilayer perceptron for nonlinear transformation; the specific formula is: in, For the first Hidden state characteristics of layer nodes. For the first Hidden state characteristics of layer nodes. For the first Layer feature dimension For time step The comprehensive adjacency matrix, For the first The learnable coefficients for the degree of information retention of the control center nodes within the layer. For the first Nonlinear mapping function of the layer.

5. The source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids as described in claim 3, characterized in that, The processing procedure of the adaptive spectral filter in step 33 is as follows: Step 331: Perform a one-dimensional fast Fourier transform on the higher-order spatial features along the time dimension, mapping them to the complex frequency domain space to obtain the frequency domain feature tensor; the formula is as follows: in, For the frequency domain feature tensor mapped to the complex frequency domain space, For discrete frequency variables in the complex frequency domain, For discrete frequency numbers, For a one-dimensional fast Fourier transform mapping operator, These are high-order spatial features extracted via an adaptive dynamic graph isomorphism network; Step 332: Calculate the amplitude spectrum of the frequency domain feature tensor. Subtract the threshold parameter from the amplitude spectrum and modulate it with the sigmoid function and sharpness parameter to obtain the frequency domain gated mask tensor; the formula is as follows: in, For frequency domain gated mask tensors, The amplitude spectrum of the frequency domain characteristic tensor. The threshold parameter is used to control the noise cutoff amplitude. To adjust the sharpness parameter for mask edge smoothness; Step 333: Perform a Hadamard product between the gate mask tensor and the frequency domain feature tensor to obtain the filtered frequency domain feature tensor; the formula is: in, This represents the filtered frequency domain feature tensor. This represents the Hadama product operator; Step 334: Remap the filtered frequency domain feature tensor back to the time domain using inverse fast Fourier transform, and introduce residual connections and layer normalization to obtain the denoised source-end spatiotemporal features; the formula is: in, The source spatiotemporal features after denoising. For layer normalization function, It is the inverse fast Fourier transform operator.

6. The source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids as described in claim 1, characterized in that, Step 4 specifically includes: Step 41: Perform dimension alignment and high-dimensional mapping on continuous environmental meteorological modal data and discrete operational rhythm modal data to obtain environmental modal hidden features and operational modal hidden features; Step 42: Introduce a cross-modal attention mechanism to generate a query matrix based on the hidden layer features of the environmental modality. Generate key matrix based on hidden layer features of job mode Sum matrix And based on the query matrix Key matrix Sum matrix Calculate cross-modal interaction features; Step 43: Introduce a contrastive prototype alignment mechanism, construct a set of prototype vectors in the latent space, calculate the soft assignment probability of each prototype vector by the latent features of the environment modality and the cross-modal interaction features, and construct the contrastive alignment loss through KL divergence to achieve cross-modal semantic alignment. Step 44: The environmental modal hidden layer features after comparison prototype alignment constraints are concatenated with the cross-modal interaction features. The concatenated features are then input into a nonlinear feedforward layer composed of a multilayer perceptron, and the load-side integrated features are output.

7. The source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids as described in claim 6, characterized in that, Step 41 specifically includes: Step 411: For continuous environmental meteorological modal data, a multilayer perceptron is used for feature space transformation to obtain environmental modal hidden layer features under a unified hidden space dimension; the formula is as follows: in, The input feature tensor represents the continuous environmental meteorological modal data at the load end. Indicates the length of the time step. Representing the dimensions of meteorological variables, This represents the hidden layer features of the environmental modalities after processing by a multilayer perceptron. This represents the dimension of the unified implicit space. This represents a nonlinear mapping function for continuous environmental meteorological modal data; Step 412: For discrete job rhythm modal data, a learnable embedding dictionary is used to transform the feature space to obtain the job modal hidden layer features under a unified latent space dimension; the formula is as follows: in, The input feature tensor represents the discrete operational rhythm modal data at the load end. Represents the discrete feature dimension. This represents the hidden layer features of the job modality after learning the learnable embedding dictionary mapping. This represents a learnable embedding dictionary mapping constructed for discrete job rhythm modal data; Step 42 specifically includes: Step 421: Generate a query matrix from the hidden features of the environmental modality through linear transformation. The formula is: in, The query matrix generated for the hidden layer features of the environmental modality. This is a learnable linear mapping weight matrix for the hidden layer features of environmental modalities; Step 422: Generate the key matrix from the hidden layer features of the working mode through two sets of independent linear transformations. AND-value matrix The formula is: in, The key matrix generated for the hidden layer features of the job mode. The value matrix generated for the hidden layer features of the operation mode. This is a learnable linear mapping weight matrix for the hidden layer features of the task modality, used to generate state matching labels. This is a learnable linear mapping weight matrix for the hidden layer features of the operational modality, used to extract features related to actual energy consumption. The dimension of latent space features; Step 423: Calculate the query matrix AND key matrix The scaled dot product attention score, after Softmax normalization, is compared with the value matrix. Weighted summation yields the cross-modal interaction feature, which characterizes the dynamic nonlinear relationship between environmental variables and work rhythm; the formula is as follows: in, These are cross-modal interaction features obtained through attention-weighted fusion. Step 43 specifically includes: Step 431: Construct a set of prototype vectors in the latent space, each prototype vector representing a typical coupling mode between sea state and operating condition. Step 432: Calculate the soft assignment probabilities of the environment modality hidden layer features and cross-modal interaction features to each prototype vector. The calculation formula is as follows: in, For the first The environmental modal hidden layer features at the time step are related to the _th The probability of belonging to each prototype vector. For the first Cross-modal interaction features at the time step for the The probability of belonging to each prototype vector. For the first Hidden layer features of environmental modalities at each time step For the first Cross-modal interaction features at each time step For the first One prototype vector, For the first One prototype vector, The number of prototype vectors. The temperature coefficient is used to adjust the smoothness of the probability distribution; Step 433: Calculate the KL divergence between the attribution probability distribution of the hidden layer features of the environment modality and the attribution probability distribution of the cross-modal interaction features. Use this KL divergence value as the contrastive alignment loss to constrain the semantic consistency of the two modalities in the prototype space. The formula for calculating the contrastive alignment loss is as follows: in, For contrastive alignment loss used for cross-modal semantic alignment, For the first The probability of the hidden layer features of the environmental modality at each time step belonging to the complete prototype vector. For the first The probability of attribution of cross-modal interaction features to the complete prototype vector at each time step. The KL divergence between two attribution probability distributions. Indicates the length of the time step; The specific formula for the comprehensive feature of the load end in step 44 is as follows: in, The combined features of the load end after joint mapping The output feature dimension of the fused features, This is a tensor concatenation operation along the feature dimension. This is a multimodal feature fusion network operator.

8. The source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids as described in claim 1, characterized in that, Step 5 specifically includes: Step 51: Concatenate the spatiotemporal features of the source end with the comprehensive features of the load end, and use a gated cyclic unit to learn the dynamic interaction relationship between the source and load states; Step 52: After decoding the terminal hidden state feature vector output by the gated recurrent unit through the probability prediction head and the MLP, the expected prediction vector and the standard deviation prediction vector of the source end and the load end in the future time period are obtained. Step 53: Construct a global optimization objective function and optimize the model parameters of the Ocean-STGF prediction model based on the global optimization objective function.

9. The source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids as described in claim 8, characterized in that, Step 51 specifically includes: Step 511: At each time step, the source-load collaborative interaction prediction module concatenates the spatiotemporal features of the source end and the comprehensive features of the load end along the feature dimension to obtain the global input feature tensor, with the formula as follows: in, For time step The concatenated global input feature tensor For time step The spatiotemporal characteristics of the source end, For time step The comprehensive characteristics of the load end; Step 512: Input the global input feature tensor into the gated recurrent unit. The gating mechanism and the hidden state feature vector update process are calculated as follows: in, For time step Update the gate feature vector. For time step The reset gate feature vector, This represents the candidate hidden state vector. For the current time step The hidden state feature vector. This represents the hidden state feature vector from the previous time step. To update the gate weight matrix, To reset the weight matrix of the gate, The weight matrix is ​​the global input feature tensor. Let be the weight matrix from the hidden state to the update gate. Let be the weight matrix from the hidden state to the reset gate. Let be the weight matrix of the hidden states. To reset the gate's bias vector, To update the gate's bias vector, This is the bias vector for the hidden state. It is the Sigmoid activation function. The hyperbolic tangent nonlinear activation function; Step 52 specifically includes: Step 521: The probability prediction head adopts a dual-branch fully connected layer structure, with an expectation branch and a standard deviation branch; Step 522: Map the terminal hidden state feature vector output by the gated recurrent unit to the prediction expectation and uncertainty parameter space of the future time period; Step 523: Decode the terminal hidden state feature vector using two independent MLPs. The expectation branch maps the terminal hidden state feature vector output by the gated recurrent unit to the predicted expected vectors of source-side capacity and load-side energy consumption for multiple future time steps. The standard deviation branch maps identical terminal hidden state feature vectors to the predicted standard deviation vectors of the source and load sides. The calculation formula is as follows: in, The expected prediction vector at the source end. The source-side prediction standard deviation vector. Let the predicted expected vector be the vector at the load end. Let be the predicted standard deviation vector of the load end. This is the terminal hidden state feature vector output by the gated recurrent unit. For the decoding network of the desired branch, Decoding network with standard deviation branch, To ensure that the standard deviation is a positive real number, a nonlinear activation function is used; Step 53 specifically includes: Step 531: Using the joint loss function based on Gaussian negative log-likelihood loss, calculate the Gaussian negative log-likelihood loss of the source prediction task and the Gaussian negative log-likelihood loss of the load prediction task respectively. The Gaussian negative log-likelihood loss of the source prediction task The calculation formula is as follows: in, For the source-end prediction task, use the Gaussian negative log-likelihood loss. To predict the time step length, For the source end in the future The time step and the first The actual production capacity of each node For the source end in the future The time step and the first The expected value of each node. For the source end in the future The time step and the first The standard deviation of the prediction for each node; The Gaussian negative log-likelihood loss of the load prediction task The calculation formula is as follows: in, The Gaussian negative log-likelihood loss is used for the prediction task of the payload. To predict the time step length, For the future of the Netherlands The time step and the first The actual energy consumption of each node For the future of the Netherlands The time step and the first The expected value of each node. For the future of the Netherlands The time step and the first The standard deviation of the prediction for each node; Step 532: Construct a global optimization objective function containing a regularization term based on the Gaussian negative log-likelihood loss of the source-end prediction task and the Gaussian negative log-likelihood loss of the load-end prediction task. Obtain the globally optimal parameters of the Ocean-STGF prediction model by solving for the minimum value of this global optimization objective function. The specific optimization formula is as follows: in, For the globally optimal parameter set, The set of learnable parameters for the Ocean-STGF prediction model. This is the balance coefficient between the source and payload task losses. for Regularization hyperparameters, For parameters Norm penalty term.

10. The source-load coordinated multimodal spatiotemporal map prediction method for deep-sea microgrids as described in claim 1, characterized in that, Step 5 is followed by: The predicted expected vectors and predicted standard deviation vectors of the source and load sides are input into the upper-level scheduling and control module. Combined with the prediction uncertainty information, the power output allocation on the generation side, the power response on the load side, and the energy storage charging and discharging strategy are dynamically adjusted to achieve safe and stable operation of the microgrid.

Images