Power plant-oriented carbon emission space-time sequence prediction method and system

Abstract

The invention discloses a power plant-oriented carbon emission space-time sequence prediction method and system, and the method comprises the steps: inputting a correlation network diagram of a power system and a data fragment of a carbon emission space-time sequence in a past window into a main branch model for carbon emission prediction; the main branch model comprises a multi-domain fusion coding unit, a cross-domain entanglement encoder and a prediction decoder which are connected in sequence; adding the high-dimensional representation obtained by mapping the data fragments with the multi-dimensional coding matrix by using a multi-domain fusion coding unit to obtain a fusion sequence; a cross-domain entanglement encoder is utilized to compress the fusion sequence along the time dimension to obtain potential features, then the potential features are extracted to obtain cross-region dependence features, and the cross-region dependence features are recovered to the original scale to obtain spatial-temporal features; processing the spatio-temporal characteristics by using a prediction decoder to obtain a carbon emission spatio-temporal sequence prediction value of a future window; according to the method, the accuracy of cross-regional carbon emission space-time sequence prediction can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent operation and maintenance of power systems and spatiotemporal data modeling, and in particular to a carbon emission spatiotemporal sequence prediction method and system for power plants. BACKGROUND

[0002] Under the impetus of the "double carbon" goal, as a key field of carbon emissions, real-time monitoring and accurate prediction of carbon emissions of power plants are the core link to realize low-carbon operation and maintenance, quota management, and dispatching optimization. Abnormal fluctuations in carbon emissions (such as the occurrence of threshold emissions during high-load operation of power plants, or deviation of the carbon emission benchmark of power plants from expectations) are prone to cause regional power grid carbon quota to exceed the standard, lag in emission reduction measures, and even affect the accuracy of carbon footprint accounting in the entire industry chain. Therefore, accurate prediction of carbon emissions in future time periods is a key prerequisite for supporting low-carbon regulation and optimization of the power system.

[0003] Traditional carbon emission prediction of power plants mostly uses statistical models such as historical mean, moving average, exponential smoothing, and ARIMA, which usually assume that data is stationary, linear, and independent and identically distributed, making it difficult to capture spatiotemporal coupling characteristics: in the time dimension, it is difficult to respond to non-stationary scenarios such as long-term cycles (such as carbon emission fluctuations caused by seasonal electricity consumption differences) and short-term sudden disturbances (such as carbon emission surges during peak electricity consumption when power plants are at full capacity); in the spatial dimension, static correlation modeling is mostly used, making it difficult to capture emission linkage patterns between power plants across regions and dynamic transmission rules of carbon emissions along energy transmission links. Therefore, such methods are sensitive to fluctuations in power plant operating conditions and environmental disturbances, making it difficult to achieve low-latency, high-precision rolling prediction in city-level or even regional power grids, and prone to the dual problems of long-term trend misjudgment and short-term peak omission, which cannot meet the real-time needs of low-carbon dispatching of the power system.

[0004] In recent years, the rapid development of deep learning technology has opened up new paths for spatio-temporal sequence prediction. However, existing methods have several significant limitations: (1) Long-distance, long-time similarity defects: Most models struggle to explicitly model the correlation between power plant carbon emissions across regions and cycles. For example, the same type of power plants distributed across provinces often exhibit similar daily emission curves during the winter heating period. Power plants in the same region show a high degree of consistent emission baseline during the seasonal transition period, such as the stable wind speed stage in spring and autumn. Existing models are often limited to short-term data from local power plants and cannot reliably identify and utilize these long-scale, long-distance similarity features, resulting in significant amplification of cross-regional prediction bias. (2) Static modeling makes it difficult to express "cross-domain entanglement": Existing methods often model the spatial correlation and temporal dynamics of power plants separately and use a static adjacency matrix to represent carbon emission relationships. However, the spatio-temporal dependence in actual power systems is highly intertwined. For example, the emission peak during the peak electricity consumption period mainly occurs in areas with high concentrations of thermal power plants and gradually spreads along the energy transmission link. When the load of a power plant increases sharply, its emission correlation with downstream regions immediately changes. Static modeling cannot capture this "cross-domain entanglement" characteristic, leading to prediction results that deviate from actual operating rules. (3) Efficiency and versatility of self-supervised methods are limited: Existing self-supervised prediction methods for power plant carbon emissions treat the spatio-temporal sequences of all power plants in a region as a single sample. As the size of the urban power grid and the time step increase, the computational load and memory consumption increase exponentially, making it difficult to meet the demand for low-latency online deployment in power dispatching. At the same time, many methods are highly dependent on specific prior knowledge and backbone architecture, and often require re-design of the network when migrating across tasks or domains, which is costly and unstable, limiting their versatility. (4) Insufficient characterization of spatio-temporal heterogeneity: Power plant carbon emissions exhibit strong heterogeneity. From a spatial perspective, different types of power plants have significantly different emission mechanisms. The same type of power plant in different regions (such as cold northern regions and warm southern regions) also exhibits different operating efficiency and emission levels due to differences in climate conditions. From a temporal perspective, the emission patterns during peak and off-peak electricity consumption periods are vastly different. Existing methods often use fixed model parameters, which cannot be dynamically adjusted according to the type of power plant, regional environment, and time period context, leading to "model mismatch" or "over-smoothing" and a significant decrease in prediction accuracy when migrating across regions and time periods.

[0005] In the related art, the patent application document with the publication number CN118627011A aims to address the problem of spatial heterogeneity in power carbon emission data. Based on structural information theory and graph neural network model, a new method for extracting power grid hierarchical structure features is proposed, which effectively captures the hierarchical relationship existing in power carbon emission data and avoids the loss of power grid structure information. This scheme is aimed at regional level power carbon emission prediction, not more granular power plant carbon emission prediction. It requires two types of heterogeneous data for data input: regional electric carbon data (including power carbon emission data and electricity consumption data) and power grid structure information (substation distribution, high-voltage distribution line, voltage relationship, etc.). In addition, the time series feature extraction in this scheme uses a GRU-Transformer combined model to capture short-term dependencies with GRU and long-term dependencies with Transformer. The structural feature extraction is based on structural information theory and graph neural network, which constructs the encoding tree of the power grid by minimizing the structural entropy to represent the physical hierarchy of the power grid. Finally, GCN is used to extract the features of this structure. By introducing the power carbon emission factor, the cross-prediction between electricity consumption and carbon emission is realized. SUMMARY

[0006] The technical problem to be solved by the present application is how to improve the accuracy of cross-regional carbon emission spatio-temporal sequence prediction while reducing computational overhead.

[0007] The present application solves the above technical problems by the following technical means: A carbon emission spatio-temporal sequence prediction method for power plants is proposed, which includes: inputting the associated network graph of the power system and the data segment of the carbon emission spatio-temporal sequence in the past window into a main branch model for carbon emission prediction, which includes a multi-domain fusion encoding unit, a cross-domain entanglement encoder, and a prediction decoder connected in turn; adding the high-dimensional representation obtained by mapping the input data segment to a multi-dimensional encoding matrix to obtain a fusion sequence, wherein the dimensions of the multi-dimensional encoding matrix include space, time, and position; using the cross-domain entanglement encoder to compress the fusion sequence along the time dimension to obtain latent features, then extracting cross-regional dependency features from the latent features, and restoring the cross-regional dependency features to the original scale to obtain spatio-temporal features; using the prediction decoder to process the spatio-temporal features to obtain the carbon emission spatio-temporal sequence prediction value of the future window.

[0008] Further, before inputting the associated network graph of the power system and the data segment of the carbon emission spatio-temporal sequence in the past window into the main branch model for carbon emission prediction, the method further includes: The associated network graph is constructed according to a topological structure and an operation correlation of a power system , wherein, is a set of power plant nodes, is a set of edges, is an adjacency matrix of , and is a total number of power plants.

[0009] Further, the high-dimensional representation obtained by mapping the input data segment by the multi-domain fusion encoding unit is added to a multi-dimensional encoding matrix to obtain a fusion sequence, including: projecting the input data segment through a full connection layer to obtain a high-dimensional representation; acquiring spatial information by using the associated network graph, and calculating a spatial graph Laplace code of the spatial information as a spatial encoding matrix; encoding each time step of the input data segment to obtain a time encoding matrix; encoding position information of the input data segment to obtain a position encoding matrix; adding the spatial encoding matrix, the time encoding matrix and the position encoding matrix to the high-dimensional representation after broadcasting in the corresponding dimensions to obtain the fusion sequence.

[0010] Further, the cross-domain entanglement encoder includes a time domain compressor, a spatial domain extractor and a time domain decompressor connected in sequence; fusing the time entanglement mask with the fusion sequence to obtain a mask fusion sequence; compressing the mask fusion sequence along the time dimension by using the time domain compressor to obtain a latent feature; converting the latent feature into a weight matrix of a region pair node by using the spatial domain extractor, and combining the latent feature, the weight matrix and an introduced spatial entanglement mask array to calculate an agent representation, and then performing multi-layer attention calculation on the agent representation to obtain a cross-region dependent feature; restoring the cross-domain dependent feature to the original time length by using the time domain decompressor, and then adding the restored cross-domain dependent feature to the fusion sequence in residual to obtain the spatiotemporal feature.

[0011] Further, the matrix representation of the time entanglement mask is:

[0012] In the formula, is a time entanglement mask, is a visible ratio control coefficient, is a fusion time mask obtained by weighting a plurality of time masks.

[0013] Furthermore, the various time masks include a first time mask, a second time mask, a third time mask, and a fourth time mask, wherein: The first-time mask is obtained by a masking strategy that sets some time points to 0 with a certain probability and sets the remaining time points to 1. The second time mask is obtained by a masking strategy that randomly samples several consecutive intervals of each node and sets them to 0; The third time mask is obtained by selecting a preset time period and setting it to 0; The fourth time mask is obtained by setting the values ​​before the current time to 1 and the values ​​before the future time to 0.

[0014] Furthermore, both the time-domain compressor and the time-domain decompressor include a first fully connected layer, a nonlinear activation function, and a second connected layer connected in sequence.

[0015] Furthermore, the spatial domain extractor includes a spatial proxy tensor, a structure-aware gated multi-head spatial entanglement attention mechanism, a standard multi-head attention layer, and a feedforward network; The node features of the latent features at each time step are extracted using the spatial proxy tensor, and the node features are converted into a weight matrix of regions to nodes; The structure-aware gated multi-head spatial entanglement attention mechanism is used to treat the weight matrix as the query vector and key vector, and the latent features as the value vector. The structure-aware gated multi-head spatial attention weight matrix is ​​calculated by combining the gating function. The introduced spatial entanglement mask is fused with the gated multi-head spatial attention weight matrix to obtain the multi-head fusion weight matrix. The multi-head fusion weight matrix is ​​then concatenated to obtain the surrogate representation. By using a standard multi-head attention layer, the surrogate representation is used as the key vector and value vector, and the latent features are used as the query vector to perform multi-layer attention computation, resulting in spatially restored features. The cross-regional dependent features are obtained by processing the spatially restored features using a feedforward network and then adding the residuals of the spatially restored features.

[0016] Furthermore, the formula for calculating the gated multi-head spatial attention weight matrix is ​​as follows:

[0017] In the formula, For the first Spatial attention weight matrix of size, Represents a spatial proxy tensor. For node features, For gated functions, This indicates that Hadamard multiplies element by element. This is the weight matrix of the region to the node. is a region-to-node mapping matrix, is a spatial encoding matrix, denotes a transpose of the spatial encoding matrix, , , , is a learnable parameter, denotes a scaling factor, is a transpose symbol.

[0018] Further, the spatial entanglement mask includes a first spatial mask, a second spatial mask, and a third spatial mask, and the first spatial mask, the second spatial mask, and the third spatial mask are fused with the gated multi-head spatial attention weight matrix respectively; The first spatial mask randomly selects nodes in each row with a certain probability and sets the nodes to 0, and the remaining nodes are set to 1 to obtain a mask strategy; The second spatial mask obtains a mask strategy by randomly selecting a number of agent rows and setting all columns of the agent rows to 0; The third spatial mask obtains a mask strategy by randomly determining a continuous interval of columns and setting all columns in the interval to 0.

[0019] Further, the prediction decoder includes a first linear layer, a nonlinear activation function, and a second linear layer connected in sequence.

[0020] Further, the main branch model is stacked a cross-domain entanglement encoder, the output of the previous layer cross-domain entanglement encoder is taken as the input of the next layer cross-domain entanglement encoder, the output of the last layer cross-domain entanglement encoder is added to the residual of the fusion sequence, and a spatiotemporal hidden representation is obtained; Correspondingly, the prediction decoder is used to process the spatiotemporal hidden representation, and a spatiotemporal sequence prediction value of the future window of carbon emissions is obtained.

[0021] In addition, the present aspect also proposes a carbon emission spatiotemporal sequence prediction system for a power plant, the system includes a data acquisition module and a pre-trained main branch model for carbon emission prediction, and the main branch model includes a multi-domain fusion encoding unit, a cross-domain entanglement encoder, and a prediction decoder connected in sequence, wherein: The data acquisition module is used to acquire the associated network graph of the power system and the data segment of the past window of carbon emission spatiotemporal sequence, and input the data segment to the main branch model; The main branch model is used to add the high-dimensional representation obtained by mapping the input data segment by the multi-domain fusion encoding unit to a multi-dimensional encoding matrix to obtain a fusion sequence, wherein the dimensions of the multi-dimensional encoding matrix include space, time, and position; The cross-region dependent feature is extracted from the potential feature after the fusion sequence is compressed along the time dimension by using the cross-region entanglement encoder, and the cross-region dependent feature is restored to the original scale to obtain the spatiotemporal feature; The spatiotemporal feature is processed by using the prediction decoder to obtain the prediction value of the carbon emission spatiotemporal sequence of the future window.

[0022] In addition, the application further provides a carbon emission spatiotemporal sequence prediction model training method for power plants, which is used for training the main branch model in the method. The carbon emission monitoring data set of the power plant is selected, and the carbon emission monitoring data of the power plant in the monitoring area is converted into a spatiotemporal sequence sample; Based on the given historical window and the prediction window, the spatiotemporal sequence sample is processed in a set step sliding manner to generate a training sample pair, and the training sample pair includes historical sequence data and future real sequence data; The historical sequence data and the associated network graph are enhanced to obtain enhanced sequence data and enhanced associated network graph; The enhanced sequence data and the enhanced associated network are used as the input of the enhanced branch model, and the historical sequence data and the associated network graph are used as the input of the main branch model for training to obtain the trained main branch model for carbon emission prediction, wherein the structure of the enhanced branch model and the main branch model is the same.

[0023] Further, the training method further comprises: The spatiotemporal feature output by the main branch model and the enhanced spatiotemporal feature output by the enhanced branch model are respectively enhanced in spatial heterogeneity to obtain spatial clustering allocation results and enhanced spatial clustering allocation results; The spatiotemporal feature output by the main branch model and the enhanced spatiotemporal feature output by the enhanced branch model are respectively enhanced in time heterogeneity to obtain positive sample pairs and negative sample pairs.

[0024] Further, the spatiotemporal feature output by the main branch model and the enhanced spatiotemporal feature output by the enhanced branch model are respectively enhanced in spatial heterogeneity to obtain spatial clustering allocation results and enhanced spatial clustering allocation results, comprising: The monitoring area is projected into a plurality of potential representation spaces, each representation space corresponding to a different operating condition; Based on each representation space, a group of cluster embeddings are generated, each cluster embedding representing a potential factor for distinguishing the difference between operating conditions; The first node in the spatiotemporal feature or the enhanced spatiotemporal feature is processed by using each cluster embedding in the time slice The embedding representation of the node at time step t and time slice s in the spatio-temporal feature is clustered to obtain a spatial clustering assignment result or an enhanced spatial clustering assignment result.

[0025] Further, the spatio-temporal feature output by the main branch model and the enhanced spatio-temporal feature output by the enhanced branch model are respectively subjected to time heterogeneity enhancement to obtain a positive sample pair and a negative sample pair, including: The embedding representation of the node at time step t and time slice s in the spatio-temporal feature is clustered to obtain a spatial clustering assignment result or an enhanced spatial clustering assignment result. The embedding representation of the node at time step t and time slice s in the spatio-temporal feature is clustered to obtain a spatial clustering assignment result or an enhanced spatial clustering assignment result. The embedding representation of the node at time step t and time slice s in the spatio-temporal feature is clustered to obtain a spatial clustering assignment result or an enhanced spatial clustering assignment result. The embedding representation of the node at time step t and time slice s in the spatio-temporal feature is clustered to obtain a spatial clustering assignment result or an enhanced spatial clustering assignment result. The embedding representation of the node at time step t and time slice s in the spatio-temporal feature is clustered to obtain a spatial clustering assignment result or an enhanced spatial clustering assignment result. The embedding representation of the node at time step t and time slice s in the spatio-temporal feature is clustered to obtain a spatial clustering assignment result or an enhanced spatial clustering assignment result. The embedding representation of the node at time step t and time slice s in the spatio-temporal feature is clustered to obtain a spatial clustering assignment result or an enhanced spatial clustering assignment result. The embedding representation of the node at time step t and time slice s in the spatio-temporal feature is clustered to obtain a spatial clustering assignment result or an enhanced spatial clustering assignment result.

[0026] Further, the joint learning objective function used when training the model includes a regional difference modeling loss based on the spatial clustering assignment result and the enhanced spatial clustering assignment result, a time sequence trend modeling loss based on the positive sample pair and the negative sample pair, a prediction loss for quantifying minimization between the spatio-temporal sequence prediction value of the future window and the future real sequence data, and a multi-scale consistency loss for quantifying consistency between the output result of the main branch model and the output result of the enhanced branch model.

[0027] Further, the multi-scale consistency loss includes a consistency loss between the prediction value of the main branch model and the future real sequence data, a consistency loss between the prediction value of the enhanced branch model and the future real sequence data, and a consistency loss between the prediction value of the main branch model and the prediction value of the enhanced branch model.

[0028] Further, the formula of the regional difference modeling loss is:

[0029]

[0030] In the formula, is an optimization objective function of the self-supervised enhancement task, is the spatial clustering assignment result, is the enhanced spatial clustering assignment result, is a parameter for controlling the smoothness of the softmax output, is the embedding representation of the node at time step t and time slice s in the spatio-temporal feature, is the embedding representation of the node at time step t and time slice s in the enhanced spatio-temporal feature, is the embedding representation of the node at time step t and time slice s in the spatio-temporal feature, is the number of cluster centers.

[0031] Further, the formula of the time sequence situation modeling loss is represented as:

[0032] wherein, represents the embedding similarity discrimination function, is an activation function; represents the regional level embedding at the same time step and the system level embedding , is the regional level embedding at different time steps and the system level embedding , represents the number of power plants.

[0033] Further, the formula of the multi-scale consistency loss is represented as:

[0034] wherein, is a balance coefficient, represents the square norm, is an average pooling operator, is the prediction value of the main branch model, is the future real sequence data, is the prediction value of the enhanced branch model, is the step length of the time scale average pooling, is the step length of the average pooling.

[0035] Further, the formula of the joint learning objective function is represented as:

[0036] wherein, , , , is an adjustable hyperparameter, is a prediction loss, is a regional difference modeling loss, is a time sequence situation modeling loss, is a multi-scale consistency loss.

[0037] In addition, the present application also proposes a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the power plant-oriented carbon emission space-time sequence prediction method or the power plant-oriented carbon emission space-time sequence prediction model training method.

[0038] The advantages of this invention are: (1) Under the constraints of power plant association topology and similarity, this invention constructs a dynamic mapping model, namely the main branch model, from the input spatiotemporal sequence (carbon emissions of each power plant) and the association network graph to the target variable (future carbon emissions). The main branch model includes a multi-domain fusion coding unit, a cross-domain entanglement encoder, and a prediction decoder connected in sequence. The high-dimensional representation obtained by mapping the input data segment using the multi-domain fusion coding unit is added to the multi-dimensional coding matrix. Through multi-domain feature fusion, multi-domain feature modeling of the input spatiotemporal sequence data segment is realized, providing a more comprehensive and robust representation for downstream prediction. The cross-domain entanglement encoder aims to coordinate... This method models the spatiotemporal dependence of carbon emissions from power plants and effectively mitigates the impact of noise disturbances. It condenses key information through time compression and aggregates high-level semantics to reduce redundant computation and thus lower computational costs. Cross-regional dependency features are extracted spatially and then decompressed to restore them to their original scale while maintaining dynamic consistency, thus stabilizing temporal semantics. Therefore, a cross-domain entanglement encoder can effectively compress and recover sequence information in the temporal dimension and fully capture cross-regional dependencies in the spatial dimension, enabling efficient capture of spatiotemporal entanglement relationships and providing support for real-time carbon emission prediction of large-scale power systems.

[0039] (2) In this invention, the spatial domain extractor uses a proxy tensor to reduce the complexity of large-scale node interactions when extracting cross-regional dependent features; and the cross-domain entanglement encoder stacks multiple layers to achieve spatiotemporal interleaving modeling through this alternating mechanism, thereby enhancing the model's expressive and fitting capabilities; therefore, the cross-domain entanglement encoder can not only efficiently capture spatiotemporal entanglement relationships, but also significantly reduce memory usage and computing power overhead.

[0040] (3) In view of the spatiotemporal characteristics of carbon emission data from power plants, an adaptive data augmentation scheme is applied to the carbon emission time series tensor and the associated network data of power plants to improve the model’s ability to perceive differences in power plant types, regional environmental differences and heterogeneity of operation modes.

[0041] (4) In view of the significant spatial heterogeneity of carbon emissions from power plants (e.g., differences in emission mechanisms among different types of power plants, and efficiency deviations of the same type of power plants in different regions due to different climate conditions and energy structures), the spatial heterogeneity modeling module can effectively enhance regional embedding, capture cross-regional spatial heterogeneity, and characterize complex power plant dependencies, thereby improving the ability of carbon emission prediction models to model spatial structures; through temporal heterogeneity modeling, the model can better identify the temporal heterogeneity patterns of carbon emissions, thereby improving the accuracy and robustness of power plant carbon emission prediction in cross-time period migration and multiple scenarios.

[0042] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0043] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation thereof. Figure 1 This is a flowchart illustrating a method for predicting the spatiotemporal sequence of carbon emissions from power plants, as proposed in an embodiment of the present invention. Figure 2 This is an overall network architecture diagram for training a spatiotemporal sequence prediction model for carbon emissions from power plants, according to one embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a cross-domain entanglement encoder according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating a method for predicting the spatiotemporal sequence of carbon emissions from power plants, as proposed in one embodiment of the invention. Figure 5 This is a schematic diagram of the structure of a spatiotemporal sequence prediction system for carbon emissions from power plants, proposed in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1 The training method for the spatiotemporal series prediction model of carbon emissions for power plants proposed in this embodiment generally includes the following steps: (2) Data preprocessing: Select the carbon emission monitoring dataset of power plants (such as the CEPE China Power Plant Carbon Emission Dataset), and organize the carbon emission observation data of power plants in different regions into spatiotemporal series data. (in This represents the number of power plant nodes. For the total time step, Simultaneously construct a power plant interconnection network graph (with feature dimensions for each node). (in For the set of power plant nodes, Let be the set of edges. yes (Adjacency matrix). Given a history window. With prediction window (2) Model building: A model is built that includes a multi-domain fusion coding unit, using a sliding window to generate training sample pairs; A hybrid network consisting of a cross-domain entanglement encoder, a spatiotemporal heterogeneity self-supervised enhancement module, and a predictive decoder; (3) Network training: using two parallel branches and one joint optimization. In the main branch, historical data is used to train the network. Self-supervised pre-training learning yields hidden features. (in (for temporary feature dimensions) The predictive decoder produces a future window prediction. Used for matching real future window sequences Calculate the predicted loss In the enhancement branch, for and Apply adaptive data augmentation to obtain and Hidden features are obtained through self-supervised pre-training. The predictive decoder produces future window predictions. ,Will The loss for modeling regional differences is obtained by feeding the spatiotemporal heterogeneity self-supervised enhancement module. Modeling loss for temporal situation differences At the same time through , and Calculate multi-scale consistency loss The final loss is calculated by weighted summation. , and jointly optimize. (4) Real-time prediction: In the inference stage, only the main branch is enabled. The spatiotemporal sequence of carbon emissions monitored by the power plant in the past 12 hours is input into the trained model, and the predicted carbon emissions for the next 12 hours are output, so as to realize the early warning of carbon emissions exceeding the threshold of the power plant and the decision support for low-carbon operation.

[0046] Specifically, such as Figures 1 to 2 As shown, the first embodiment of the present invention proposes a training method for a spatiotemporal series prediction model of carbon emissions from power plants, the method comprising the following steps: S11. Select the carbon emission monitoring dataset of power plants and convert the carbon emission monitoring data of power plants in the monitoring area into spatiotemporal sequence samples; Specifically, the power plant carbon emission monitoring dataset selected in this embodiment is the CEPE China Power Plant Carbon Emission Data Set. The carbon emissions from power plants within the monitoring area are organized into a spatiotemporal series. ,in This represents the number of power plant nodes. For the total time step, For the feature dimension of each node, here This refers to carbon emissions.

[0047] S12. Based on the given historical window and prediction window, the spatiotemporal sequence samples are processed using a set step size sliding method to generate training sample pairs, wherein the training sample pairs include historical sequence data and future real sequence data. Specifically, the input fragment is Recorded as The target is the future sequence. Recorded as Given a history window With prediction window Training sample pairs are generated using a sliding method with a step size of 1 (1-hour interval): .

[0048] S13. Perform augmentation processing on the historical sequence data and the association network graph to obtain augmented sequence data and augmented association network graph; Specifically, this embodiment constructs an interconnected network diagram based on the power system topology and operational relationships. ,in For the set of power plant nodes, Let be the set of edges. yes The adjacency matrix.

[0049] Specifically, this embodiment draws on existing data augmentation strategies to construct a carbon emission historical time-series tensor. An enhancement operator is introduced, which is an adaptive masking operator that probabilistically masks the input temporal tensor along the time-step-region dimension. This operator can adaptively adapt to the time-aware power pattern dependency of each region. Specifically, the goal of this embodiment is to enhance the time-aware power pattern dependency of each region. The Each time step masks observations with weak correlation to overall emission patterns to mitigate noise interference and anomalous fluctuations. This is based on a masking probability sampled from a Bernoulli distribution:

[0050] in, Let be a Bernoulli random variable, indicating whether the position is masked, where 1 indicates masking and 0 indicates that the original value is retained; To preserve frequency, higher The value means in the region The Carbon emissions at a given time step are more likely to be masked because they have a lower correlation with the overall carbon emission pattern in the region.

[0051] According to the sampling Masking is performed at the corresponding position (blocking the value, creating an obstruction), thereby transforming the original sequence into an enhanced sequence. The data after this carbon emission time series enhancement operation is denoted as:

[0052] This strategy can effectively suppress noise and abnormal fluctuations, and improve the model's attention to potential emission patterns.

[0053] Specifically, in addition to enhancing historical sequence data, this embodiment also enhances the power plant network diagram. To enhance the graph, the correlation between carbon emission patterns of two nodes is calculated based on the original interconnected network graph. Specifically, Dynamic Time Warping (DTW) is used to calculate the similarity of any node sequence, resulting in an edge weight matrix. Higher similarity corresponds to higher weights, and lower similarity to lower weights. Then, a threshold is set in the model, and all edges are iterated over; if the weight is below the threshold, the edge is discarded, resulting in the enhanced graph. This method not only mitigates the bias introduced in the regional connectivity of low-correlation power patterns but also helps the model capture long-range dependencies across regions within a global power system environment. The final enhanced carbon emission time-series data and power plant interconnected network data are as follows: , To enhance the network graph.

[0054] It should be noted that, in this embodiment, an adaptive data augmentation scheme is designed to address the spatiotemporal characteristics of carbon emission data from power plants. This scheme is applied to carbon emission time series tensors and power plant association network data to enhance the model's ability to perceive differences in power plant types, regional environmental differences, and heterogeneity in operating modes.

[0055] S14. The enhanced sequence data and enhanced correlation network are used as inputs to the enhanced branch model, and the historical sequence data and correlation network graph are used as inputs to the main branch model. The trained main branch model is used for carbon emission prediction. The enhanced branch model and the main branch model have the same structure.

[0056] Specifically, such as Figure 2 As shown, both the main branch model and the enhancement branch model include a multi-domain fusion coding unit, a cross-domain entangled encoder, and a predictive decoder connected in sequence, wherein: (1) A multi-domain fusion coding unit, used to add the high-dimensional representation obtained by mapping the input data segment to the multi-dimensional coding matrix to obtain the fusion sequence, specifically including: 1-1) Data projection: Projection of historical sequence data from the original input. A high-dimensional representation is obtained through a fully connected layer. The formula is expressed as:

[0057] in, This is a temporary feature dimension used to carry the information capacity for subsequent spatiotemporal modeling.

[0058] 1-2) Spatial coding: using relational network graphs The graph structure is used to obtain spatial information, calculate the symmetric normalized Laplace and take the antecedent. The eigenvectors are linearly mapped to... The dimension obtains the Laplace encoding of the spatial graph. .

[0059] 1-3) Time Coding: To characterize the strong periodicity of time series, time coding is constructed on two scales: intra-week and intra-day. Specifically, each time step... Mapped to intraday index Obtain the week code ;Will Mapped to intraday time slots Obtain intraday code .

[0060] 1-4) Location coding: using time-location coding Used to integrate the positional information of the input sequence.

[0061] The coding matrices obtained from the spatial coding, temporal coding, and position coding are broadcast in their respective dimensions to obtain... , and with data projection The sums are used to obtain the output of the multi-domain fusion coding unit as the fusion sequence. The formula is expressed as: .

[0062] (2) such as Figure 3 As shown, the cross-domain entanglement encoder includes a temporal domain compressor, a spatial domain extractor, and a temporal domain decompressor. It is used to compress the fused sequence along the temporal dimension to obtain latent features, extract cross-regional dependent features from these latent features, and restore the cross-regional dependent features to the original scale to obtain spatiotemporal features. 2-1) Time-domain compressor and time-domain decompressor First, the fusion sequence As input to the cross-domain entanglement encoder, a temporal entanglement mask is also introduced. First, perform a time masking on the fused sequence to obtain the masked fused sequence. The formula is expressed as:

[0063] in, express Element-wise multiplication, for The broadcast mask is obtained by broadcasting along the feature dimension.

[0064] Specifically, time entanglement mask It is obtained through the fusion of multiple strategies, and the formula is expressed as follows:

[0065]

[0066]

[0067] in, This is a fused time mask obtained by weighting multiple time masks. , , , These represent the weighting coefficients for the four masking strategies. This is a coefficient used to control the visibility ratio; This represents the first time mask obtained by setting a certain probability to 0 at some time points and 1 at the remaining time points, which can be used to simulate fine-grained random missing values. This represents the second time mask obtained by a masking strategy that randomly samples several consecutive intervals of each node and sets them to 0; This represents a third time mask obtained by selecting a longer time period and setting it to 0, which is used to reduce context and improve transferability and robustness. This represents the fourth time mask obtained by setting the time before the current time to 1 and the time before the future time to 0, which is used to indicate that only the past is visible.

[0068] Specifically, the time-domain compressor includes two fully connected layers in the time dimension, namely a first fully connected layer and a second fully connected layer, with a non-linear activation function used between the first and second fully connected layers. It acts as a compressor along the time dimension (the structure of a time-domain decompressor is the same as that of a time-domain compressor). The length is compressed in the compression layer. Time series mapping to length The latent time axis yields latent features After passing through the spatial extractor (described in detail in step 2-2), cross-regional dependency features are obtained. Finally, the decompressor separates the cross-regional dependency features. Restore to length and the fusion sequence output by the multi-domain fusion coding unit. Add the residuals to form a code representation As a spatiotemporal characteristic, the formula is as follows:

[0069]

[0070]

[0071] In the formula, This indicates a compression operation. This indicates a spatial feature extraction operation. This indicates a decompression operation.

[0072] It should be noted that by using compression operations to summarize the current time window into a small number of key states and aggregating high-level semantics, the computational overhead is reduced; the subsequent decompression operations restore these low-dimensional hidden states to their original time length, which helps to stabilize the temporal semantics.

[0073] 2-2) Spatial domain extractor, used to convert latent features into a weight matrix of region-to-node pairs, and to compute a surrogate representation by combining latent features, weight matrix, and an introduced spatial entanglement mask. Then, the surrogate representation is subjected to multi-layer attention computation to obtain cross-region dependent features: In the spatial domain extractor, this embodiment introduces a structure-aware gated multi-head spatial entanglement attention mechanism, a standard multi-head attention layer, and a learnable spatial surrogate tensor. To fully capture spatial dependencies, Represents the real number field. express The set of real matrices, Indicates the number of space proxies. This indicates the temporary feature dimension.

[0074] First, regarding latent features Each potential time step Extract the node features at that moment. Calculate the weight matrix of the region to the node. :

[0075] In the formula, For learnable spatial proxy tensors, express The transpose of .

[0076] And introduce a gating function The definition is as follows:

[0077] in, It is a learnable linear transformation used to adjust the scale of relationships between regions and nodes. For bias terms, for function.

[0078] At this point, the structure-aware gating spatial attention weight matrix is ​​calculated. The formula is as follows:

[0079] in, This is the mapping matrix from regions to nodes. , , , For learnable parameters, This represents the spatial encoding matrix.

[0080] At this point, the spatial entanglement mask matrices are introduced as follows: .in This represents the first spatial mask obtained by a masking strategy that randomly selects some nodes in each row with a certain probability, sets these nodes to 0, and sets the remaining nodes to 1. It is used to simulate fine-grained random missing nodes. This represents the second space mask obtained by randomly selecting a number of proxy rows and setting all columns of these rows to 0, used to simulate long-term proxy downtime; This represents a third-space mask obtained by randomly determining a column within a continuous interval and setting all proxies to 0 within that interval. This mask is used to reduce context, improve transitions, and enhance robustness.

[0081] For the three types of mask matrices mentioned above, the three attention methods are calculated respectively, as shown in the following formulas:

[0082]

[0083]

[0084] in, For learnable parameters, It is a learnable projection matrix. Indicating the attention of the bulls The number of masks is used for parallel modeling across multiple attention subspaces to improve the diversity and stability of the representation. The attention computation methods for the other two masks are also discussed. Similarly, simply put Replace with And by changing the corresponding learnable parameters, we obtain the following results. Then, these three outputs are concatenated and projected to obtain the proxy representation of the structure-aware gated multi-head spatially entangled attention layer. The formula is as follows:

[0085] in, It is a learnable projection matrix. Then, it passes through a standard multi-head attention layer to obtain the spatially reconstructed features. ,in As and , As The specific formula is as follows:

[0086]

[0087]

[0088] in, It is a learnable projection matrix. , , These are learnable parameters. Finally, cross-regional dependency features are obtained through a feedforward network and residual connections. For subsequent decompression, the formula is as follows:

[0089] In the formula, This represents a feedforward network.

[0090] It should be noted that the cross-domain entanglement encoder consists of three stages in sequence: "temporal domain compression - spatial domain extraction - temporal domain decompression," forming C. E The D-structure unit, a cross-domain entanglement encoder, aims to collaboratively model the spatiotemporal dependencies of power plant carbon emissions while effectively mitigating the impact of noise disturbances. This module consists of three stages: time compression, spatial extraction, and time decompression. Time compression condenses key information and reduces redundant computation; spatial extraction utilizes surrogate tensors to reduce the complexity of large-scale node interactions; and decompression and residual fusion restore the original scale and maintain dynamic consistency. Through this alternating mechanism, the encoder can not only efficiently capture spatiotemporal entanglement relationships but also significantly reduce memory usage and computational overhead, providing support for real-time carbon emission prediction of large-scale power systems.

[0091] As a further preferred technical solution, the cross-domain entanglement encoders in this embodiment can be stacked. Layers are added to enhance the model's expressive and fitting capabilities. Finally, the output of the last layer of the cross-domain entangled encoder is fused with the initial input sequence. Perform residual connections to obtain the final spatiotemporal hidden feature representation. The formula is as follows:

[0092] in, This indicates a complete encoding transformation of the input spatiotemporal features (including temporal compression, spatial extraction, and temporal decompression). This indicates that the output of one function is used as the input of the next function. Furthermore, for the first function in the time dimension... There are slices, including:

[0093] in, Indicates the first The embedding matrix for each time slice, then let:

[0094] That is, the first Each node in the time slice Embedded representation.

[0095] (3) Predictive decoder In order to obtain the future The predicted value at each time step is used in this embodiment. (Composed of two linear layers, with a middle band) (Activation) is used as the prediction head, outputting the spatiotemporal series prediction of carbon emissions for the future window. .

[0096] It should be noted that in this embodiment, the augmentation branch model and the main branch model have the same structure. The difference lies in that the input of the augmentation branch model is augmented sequence data and an augmented association network graph, and the output hidden features are represented as follows: .

[0097] As a further preferred technical solution, this embodiment also sets up a spatiotemporal heterogeneity self-supervised enhancement module during model training. The outputs of the time domain decompressor in the main branch model and the time domain decompressor in the enhanced branch model are both connected to the spatiotemporal heterogeneity self-supervised enhancement module. The spatiotemporal heterogeneity self-supervised enhancement module includes a spatiotemporal heterogeneity modeling unit and a spatial heterogeneity modeling unit.

[0098] In response to the significant spatial heterogeneity of carbon emissions from power plants (e.g., differences in emission mechanisms among different types of power plants, and efficiency deviations among similar power plants in different regions due to varying climate conditions and energy structures), this embodiment proposes a spatial heterogeneity modeling module based on soft clustering and self-supervised learning. This module represents the hidden features output by the main branch model. Hidden feature representation of the output of the augmented branch model The process yields spatial clustering assignment results and enhanced spatial clustering assignment results: First, each region is projected onto multiple potential representation spaces, each corresponding to a different operating condition (such as high-load thermal power, low-wind-speed wind power, and energy storage in different temperature zones). Then, a set of... There are clustering embeddings, denoted as . Each of them This represents the latent factors used to distinguish differences in operating conditions. The clustering process is defined as follows:

[0099] in, Indicates the first Within each time slice, the region encoded based on the enhanced carbon emission time series and power plant association network. Embedding; Indicates region embedding With the The relevance score of each cluster center.

[0100] area The clustering assignment results can be expressed as:

[0101] This vector reflects the degree of association between a region and each cluster.

[0102] To introduce self-supervised signals into the soft clustering framework, this embodiment further designs an auxiliary learning task: based on the original power plant association network. The obtained region embedding Predict its cluster assignment. The predicted cluster score is defined as:

[0103] in, Indicates the area Corresponding to the The prediction score of each cluster.

[0104] The optimization objective function for the self-supervised augmentation task is:

[0105] in, It is control The parameter determines the smoothness of the output. The overall self-supervised objective function is defined as the regional difference modeling loss:

[0106] Furthermore, to ensure that the clustering results reflect the true distribution of regional characteristics within the power system, this invention introduces a distribution regularization strategy during the training process. In this way, the spatial heterogeneity modeling module can effectively enhance region embedding. This allows us to capture cross-regional spatial heterogeneity and characterize complex power plant dependencies, thereby enhancing the ability of carbon emission prediction models to model spatial structures.

[0107] In response to the significant heterogeneity of carbon emissions from power plants over time (such as differences in carbon emission patterns between peak and off-peak electricity consumption, and differences between seasonal cycles and sudden load disturbances), this invention proposes a time heterogeneity modeling module based on self-supervised learning.

[0108] First, the time step embeddings encoded from the raw carbon emission time series data and the enhanced time series data are fused:

[0109] in, The first representing the spatiotemporal characteristics Time step region Embedded representation, The first character representing enhanced spatiotemporal features Time step region Embedded, ; , These are learnable parameters. Subsequently, by aggregating all power plants at time steps... The embedding generates a system-level representation at that moment. :

[0110] in, express Function. In self-supervised learning tasks, this invention integrates region-level embeddings at the same time step. With system-level embedding As positive sample pairs, embedding pairs at different time steps are regarded as negative sample pairs. This allows us to both align the carbon emission patterns of the entire system at the same time (such as multiple power plants emitting high emissions simultaneously during peak hours) through positive sample pairs to ensure temporal consistency, and effectively distinguish the system operating states at different time steps (such as peak load and valley load) through negative sample pairs, thereby capturing the temporal heterogeneity of carbon emissions.

[0111] Therefore, self-supervised learning with enhanced temporal heterogeneity can be optimized by minimizing the following loss function based on cross-entropy:

[0112] in, Indicates the region-level embedding at the same time step With system-level embedding , For regional embedding at different time steps With system-level embedding , This represents the embedding similarity discriminant function. for Activation function. This module enables the model to better identify temporal heterogeneity patterns in carbon emissions, thereby improving the accuracy and robustness of power plant carbon emission predictions across time periods and various scenarios.

[0113] Furthermore, during training, the main branch model obtains the predicted future sequence through the prediction decoder. ,Will With the real sequence Optimization is performed by minimizing the following loss function:

[0114] Simultaneously, the enhanced sequence is obtained through the predictive decoder in the augmentation branch model. Define a set of time scales. Each scale corresponds to an average pooling operator. This is used to align predicted and true sequences at different temporal resolutions. Multi-scale consistency loss. It consists of three parts: the consistency between the main branch prediction and the actual value, the consistency between the augmented branch prediction and the actual value, and the consistency between the main and augmented branch predictions. The formula is as follows:

[0115] in, For balance coefficient, Representing square Norm, The time scale is used to represent the step size of average pooling. This is the step size for average pooling.

[0116] Ultimately, the loss will be predicted. Regional difference modeling loss Loss in time-series situation modeling and multi-scale consistency loss Combined, this constitutes the joint learning objective loss:

[0117] in , , , It is an adjustable hyperparameter used to balance the importance of different loss terms.

[0118] As a further preferred technical solution, the process of training and testing the main branch model in this embodiment includes: (1) Model training 1-1) Parameter Initialization The learnable parameters in the model (including multi-domain fusion coding units, cross-domain entangled encoders, predictive decoders, self-supervised enhancement modules, etc.) are initialized. The optimizer uses... A learning rate decay strategy is set to ensure stable convergence.

[0119] 1-2) Forward propagation Given a history window input First, the basic representation is extracted through a multi-domain fusion coding unit, and then the spatiotemporal hidden features are obtained through a cross-domain entanglement encoder. The main branch generates future sequence predictions via the predictive decoder. The enhancement branch is generated after the adaptive enhancement operation. .

[0120] 1-3) Loss Calculation The training loss of the model is calculated using the joint learning objective loss.

[0121] 1-4) Backpropagation and Optimization Gradients are calculated through backpropagation, and the optimizer updates the model parameters. The gradients of each loss term work together on the encoder and decoder, optimizing the model simultaneously in time, space, and augmentation branches.

[0122] 1-5) Iterative Training Repeat steps 1-2) to 1-4) until the loss on the validation set converges or the preset number of iterations is reached, and the model training process is complete.

[0123] (2) Model testing 2-1) Input and Reasoning During the testing phase, the model only uses the main branch: it uses the spatiotemporal sequence data of carbon emissions from past windows. The trained model is input, and spatiotemporal features are extracted by a multi-domain fusion coding unit and a cross-domain entangled encoder. Then, the future window prediction result is output by the prediction decoder. This stage no longer uses enhanced branches and self-supervised modules to reduce computational overhead and ensure prediction efficiency.

[0124] 2-2) Predicted Output The model output is the future The carbon emission prediction sequence for each time step is denoted as:

[0125] The predicted value at each moment corresponds to the carbon emissions of each node in the power system, which can reflect the overall trend and local fluctuations.

[0126] 2-3) Evaluation Indicators To comprehensively evaluate the predictive performance of the model, this method introduces three commonly used error metrics. First, the mean absolute error (MAE) is used to characterize the average deviation between the predicted and actual values, and it is defined as:

[0127] in, Represents a node At any moment The actual carbon emissions, For the corresponding predicted value, For the number of nodes, To predict the step size.

[0128] Secondly, the root mean square error emphasizes the impact of large errors on the overall evaluation, and its formula is as follows:

[0129] This indicator is more sensitive to abnormal fluctuations and can be used to reflect the robustness of the prediction model in extreme scenarios.

[0130] Finally, the mean absolute percentage error is used to evaluate the relative error, which is defined as:

[0131] This indicator can reflect the proportional deviation of the predicted value from the actual value, and is more suitable for comparisons across regions or multiple types of power plants.

[0132] 2-4) Online Deployment and Application During the inference phase, this method retains only the main branch, significantly reducing computational complexity and enabling low-latency online prediction. Hourly rolling predictions effectively support real-time operation and dynamic dispatching of power systems. Furthermore, this method supports anomaly warning functionality. When predicted carbon emissions exceed a preset threshold, the system automatically triggers an alarm, assisting maintenance personnel in intervening and making adjustments before potential risks occur.

[0133] Furthermore, the forecast results can be directly used for decision support. For example, they can be used to assess the carbon reduction potential of the power grid in real time, dynamically adjust regional carbon emission quotas, or optimize the operation plans of high-carbon power plants. Thus, this method not only meets the needs of the power system for low-carbon operation and maintenance under the "dual carbon" objectives, but also provides important data support for quota management and dispatch optimization.

[0134] It should be noted that the main branch model trained in this embodiment has the following advantages when used for carbon emission prediction: (1) Spatiotemporal analysis capability: using C E The D-domain entanglement encoder introduces multiple spatiotemporal masks and attention mechanisms, simultaneously characterizing spatial dependence and temporal causality, taking into account both long-term cycles and short-term fluctuations, resulting in more accurate and stable predictions.

[0135] (2) Computational efficiency: Time dimension compression and decompression reduce redundancy, and with the help of proxy tensor routing, the memory and computing power consumption is significantly reduced; only the main branch from the encoder to the predictor decoder is retained in the inference stage, which meets the requirements of low latency and low resource online deployment.

[0136] (3) Adaptive Enhancement Robustness: An enhanced branch is constructed through adaptive data augmentation, which is consistent with the original branch at the feature and output levels, and spatial and temporal heterogeneous self-supervised signals are introduced. This makes the model more robust to noise and generalizes better across regions and time periods.

[0137] (4) Multi-loss collaborative optimization: In a single backpropagation, multiple losses are jointly minimized, which improves accuracy, stability and generalization without increasing inference cost.

[0138] (5) Cross-scenario versatility: The core module design does not rely on prior knowledge in a specific field and can be extended to various spatiotemporal sequence prediction tasks such as transportation, energy, and meteorology, and has broad industrial application value.

[0139] Example 2 like Figure 4 As shown, the second embodiment of the present invention proposes a spatiotemporal series prediction method for carbon emissions from power plants, the method comprising the following steps: S21. Input the data fragments of the power system's associated network diagram and the spatiotemporal sequence of carbon emissions within the past window into the main branch model for carbon emission prediction. The main branch model includes a multi-domain fusion coding unit, a cross-domain entangled encoder, and a prediction decoder connected in sequence. S22. The high-dimensional representation obtained by mapping the input data segment using the multi-domain fusion coding unit is added to the multi-dimensional coding matrix to obtain the fusion sequence, wherein the dimensions of the multi-dimensional coding matrix include space, time and position. S23. After compressing the fused sequence along the time dimension using a cross-domain entanglement encoder to obtain latent features, cross-regional dependency features are extracted from the latent features, and the cross-regional dependency features are restored to the original scale to obtain spatiotemporal features. S24. The spatiotemporal features are processed using a predictive decoder to obtain the spatiotemporal sequence prediction of carbon emissions for the future window.

[0140] It should be noted that the structure of the main branch model used for carbon emission prediction in this embodiment can be found in the training method of the spatiotemporal sequence prediction model for carbon emissions of power plants described in the first embodiment above.

[0141] As a further preferred technical solution, step S22: adding the high-dimensional representation obtained by mapping the input data segment using the multi-domain fusion coding unit to the multi-dimensional coding matrix to obtain the fusion sequence includes the following steps: S221. Project the input data fragment through a fully connected layer to obtain a high-dimensional representation; Specifically, for the input data fragment A high-dimensional representation is obtained through a fully connected layer. The formula is expressed as:

[0142] in This is a temporary feature dimension used to carry the information capacity for subsequent spatiotemporal modeling.

[0143] S222. Obtain spatial information using the aforementioned network graph, and calculate the spatial graph Laplace code of the spatial information as the spatial coding matrix; Specifically, utilizing the network of connections The graph structure is used to obtain spatial information, calculate the symmetric normalized Laplace and take the antecedent. The eigenvectors are linearly mapped to... The dimension obtains the Laplace encoding of the spatial graph. .

[0144] S223. Encode each time step of the input data segment to obtain the time encoding matrix; Specifically, to characterize the strong periodicity of time series, time coding is constructed on two scales: intra-week and intra-day. Specifically, each time step... Mapped to intraday index Obtain the week code ;Will Mapped to intraday time slots Obtain intraday code .

[0145] S224. Encode the position information of the input data segment to obtain the position encoding matrix; Specifically, using time-location coding Used to integrate the positional information of the input sequence.

[0146] S225. After broadcasting the spatial encoding matrix, temporal encoding matrix, and positional encoding matrix in their respective dimensions, add them to the high-dimensional representation to obtain the fused sequence.

[0147] Specifically, the above encoding matrix is ​​broadcast in the corresponding dimension to obtain , , and data projection Adding them together yields the fused sequence output by the multi-domain fusion coding unit. The formula is expressed as: .

[0148] As a further preferred technical solution, step S23: after compressing the fused sequence along the time dimension using a cross-domain entanglement encoder to obtain latent features, extracting cross-regional dependency features from the latent features, and restoring the cross-regional dependency features to the original scale to obtain spatiotemporal features, specifically includes the following steps: S231. Merge the time entanglement mask with the fusion sequence to obtain a mask fusion sequence; Specifically, with fusion sequence As input, a temporally entangled mask is also introduced. First, perform a time mask on the input to obtain... The formula is expressed as:

[0149] in express Element-wise multiplication, for Obtained by broadcasting along the feature dimension.

[0150] Furthermore, the various time masks include a first time mask, a second time mask, a third time mask, and a fourth time mask, wherein: The first-time mask is obtained by a masking strategy that sets some time points to 0 with a certain probability and sets the remaining time points to 1. The second time mask is obtained by a masking strategy that randomly samples several consecutive intervals of each node and sets them to 0; The third time mask is obtained by selecting a preset time period and setting it to 0; The fourth time mask is obtained by setting the values ​​before the current time to 1 and the values ​​before the future time to 0.

[0151] S232. The mask fusion sequence is compressed along the time dimension using a time-domain compressor to obtain latent features; Specifically, two fully connected layers are used in the time dimension, with a non-linear activation function in between. Acting as a compressor along the time dimension, the compressor will reduce the length Time series mapping to length The potential time axis is obtained The public notice indicates that: .

[0152] S233. The latent features are converted into a weight matrix of region-to-node using a spatial domain extractor. The surrogate representation is then calculated by combining the latent features, the weight matrix and the introduced spatial entanglement mask. The surrogate representation is then subjected to multi-layer attention calculation to obtain cross-region dependent features. Specifically, after spatial extraction, it is obtained The public notice indicates that .

[0153] S234. After restoring the cross-domain dependency features to their original time length using a time-domain decompressor, add the residuals of the fused sequence to obtain the spatiotemporal features.

[0154] Specifically, the time-domain decompressor uses two fully connected layers in the time dimension, with a non-linear activation function in between. After decompression, it is restored to its original state. and Add the residuals to form a code representation The formula is as follows: .

[0155] It should be noted that this embodiment summarizes the current time window into a small number of key states and aggregates high-level semantics, thereby reducing computational overhead; the subsequent decompression operation restores these low-dimensional hidden states to the original time length, which helps to stabilize the time semantics.

[0156] As a further preferred technical solution, the spatial domain extractor includes a spatial proxy tensor, a structure-aware gated multi-head spatial entanglement attention mechanism, a standard multi-head attention layer, and a feedforward network; Step S233: The latent features are converted into a weight matrix between regions and nodes using a spatial domain extractor. The surrogate representation is then calculated by combining the latent features, the weight matrix, and the introduced spatial entanglement mask. Finally, the surrogate representation is subjected to multi-layer attention calculation to obtain cross-region dependent features. This specifically includes the following steps: S2331. Extract the node features of the latent features at each time step using the spatial proxy tensor, and convert the node features into a weight matrix of regions to nodes; Specifically, this embodiment utilizes a learnable spatial proxy tensor. For latent features Each potential time step Extract the node features at that moment. Calculate the weight matrix of the region to the node. :

[0157] S2332. Using a structure-aware gated multi-head spatial entanglement attention mechanism, the weight matrix is ​​used as the query vector and key vector, and the latent features are used as the value vector. The structure-aware gated multi-head spatial attention weight matrix is ​​calculated by combining the gated function. The introduced spatial entanglement mask is fused with the gated multi-head spatial attention weight matrix to obtain the multi-head fusion weight matrix. The multi-head fusion weight matrix is ​​then concatenated to obtain the surrogate representation. Specifically, the formula for calculating the gated multi-head spatial attention weight matrix is ​​as follows:

[0158] In the formula, For the first Spatial attention weight matrix of size, For node features, For gated functions, This indicates that Hadamard multiplies element by element. This is the weight matrix of the region to the node. This is the mapping matrix from regions to nodes. For spatial encoding matrix, For learnable spatial proxy tensors, This represents the transpose of the spatial encoding matrix. , , , For learnable parameters, Indicates the scaling factor. This is the transpose symbol.

[0159] Among them, the gate function The public notice is as follows:

[0160] in, It is a learnable linear transformation used to adjust the scale of relationships between regions and nodes. For bias terms, for function.

[0161] Next, the introduced spatial entanglement mask is fused with the gated multi-head spatial attention weight matrix to obtain the multi-head fusion weight matrix. The spatial entanglement mask includes a first spatial mask, a second spatial mask, and a third spatial mask. The first spatial mask, the second spatial mask, and the third spatial mask are then fused with the gated multi-head spatial attention weight matrix respectively.

[0162]

[0163]

[0164] In the formula, Entanglement mask for space, For learnable parameters, express function, It is a learnable projection matrix.

[0165] Then, the multi-head fusion weight matrix is ​​concatenated to obtain: Attention calculation method for the other two masks Similarly, simply put Replace with That is, to obtain respectively Then, these three outputs are concatenated and projected to obtain the proxy representation of the structure-aware gated multi-head spatially entangled attention layer. The formula is as follows:

[0166] in, It is a learnable projection matrix.

[0167] S2333. Using a standard multi-head attention layer, the surrogate representation is used as a key vector and a value vector, and the latent features are used as query vectors to perform multi-layer attention calculations to obtain the spatially restored features. Specifically, the spatially reconstructed features are obtained after passing through a multi-head attention layer. ,in As and , As The specific formula is as follows:

[0168] ,

[0169] in, It is a learnable projection matrix. , , These are learnable parameters.

[0170] S2334. After processing the spatially restored features using a feedforward network, the residuals of the restored features are added to obtain the cross-regional dependent features.

[0171] Specifically, it is obtained through a feedforward network and residual connections. For subsequent decompression, the formula is as follows: .

[0172] As a further preferred technical solution, the main branch model stacking The output of the previous layer of the cross-domain entanglement encoder is used as the input of the next layer of the cross-domain entanglement encoder. The output of the last layer of the cross-domain entanglement encoder is added to the residual of the fusion sequence to obtain the spatiotemporal hidden representation. Accordingly, the spatiotemporal hidden representation is processed using a predictive decoder to obtain the spatiotemporal sequence prediction of carbon emissions for the future window.

[0173] Example 3 In addition, such as Figure 5 As shown, the third embodiment of the present invention also proposes a spatiotemporal sequence prediction system for carbon emissions from power plants, corresponding to the second embodiment described above. The system is characterized by including a data acquisition module 10 and a pre-trained main branch model 20 for carbon emission prediction. The main branch model 20 includes a multi-domain fusion coding unit, a cross-domain entanglement encoder, and a prediction decoder connected sequentially, wherein: The data acquisition module 10 is used to acquire data fragments of the power system's associated network diagram and the spatiotemporal sequence of carbon emissions within the past window, and input them into the main branch model; The main branch model 20 is used to add the high-dimensional representation obtained by mapping the input data fragments to the multi-dimensional coding matrix using the multi-domain fusion coding unit to obtain the fusion sequence, wherein the dimensions of the multi-dimensional coding matrix include space, time and position. After compressing the fused sequence along the time dimension using a cross-domain entanglement encoder to obtain latent features, cross-regional dependent features are extracted from the latent features, and the cross-regional dependent features are restored to the original scale to obtain spatiotemporal features. By using a predictive decoder to process the spatiotemporal features, we can obtain the spatiotemporal sequence prediction of carbon emissions for the future window.

[0174] It should be noted that other embodiments or specific implementation methods of the spatiotemporal sequence prediction system for carbon emissions of power plants described in this invention can refer to the above-mentioned method embodiments, and will not be repeated here.

[0175] Example 4 The fourth embodiment of the present invention proposes a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the spatiotemporal sequence prediction method for carbon emissions of power plants as described in the first embodiment above, or the training method for the spatiotemporal sequence prediction model for carbon emissions of power plants as described in the second embodiment above.

[0176] It should be noted that the computer-readable medium disclosed in this embodiment may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, and portable compact disk read-only memory (CD-ROM). ROM, optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0177] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform a zero-sample image anomaly detection method according to the above embodiments.

[0178] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server.

[0179] In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0180] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0181] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0182] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" or "several" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0183] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting the spatiotemporal series of carbon emissions from power plants, characterized in that, include: The data fragments of the spatiotemporal sequence of carbon emissions within the past window are input into the main branch model for carbon emission prediction. The main branch model includes a multi-domain fusion coding unit, a cross-domain entangled encoder, and a prediction decoder connected in sequence. The high-dimensional representation obtained by mapping the input data fragment using a multi-domain fusion coding unit is added to a multi-dimensional coding matrix to obtain a fusion sequence, wherein the dimensions of the multi-dimensional coding matrix include space, time, and position. After compressing the fused sequence along the time dimension using a cross-domain entanglement encoder to obtain latent features, cross-regional dependent features are extracted from the latent features, and the cross-regional dependent features are restored to the original scale to obtain spatiotemporal features. By using a predictive decoder to process the spatiotemporal features, we can obtain the spatiotemporal sequence prediction of carbon emissions for the future window.

2. The spatiotemporal series prediction method for carbon emissions from power plants as described in claim 1, characterized in that, Before inputting the data fragments of the power system's interconnected network diagram and the spatiotemporal sequence of carbon emissions within past windows into the main branch model for carbon emission prediction, the method further includes: The associated network diagram is constructed based on the topology and operational relationships of the power system. ,in, For the set of power plant nodes, Let be the set of edges. yes The adjacency matrix, This represents the total number of power plants.

3. The spatiotemporal series prediction method for carbon emissions from power plants as described in claim 1, characterized in that, The step of adding the high-dimensional representation obtained by mapping the input data segment using the multi-domain fusion coding unit to the multi-dimensional coding matrix to obtain the fused sequence includes: The input data fragment is projected through a fully connected layer to obtain a high-dimensional representation; Spatial information is obtained using the aforementioned network graph, and the spatial graph Laplace code of the spatial information is calculated as the spatial coding matrix; Encode each time step of the input data segment to obtain a time-coding matrix; The location information of the input data segment is encoded to obtain a location encoding matrix; The spatial encoding matrix, temporal encoding matrix, and positional encoding matrix are broadcast in their respective dimensions and then added to the high-dimensional representation to obtain the fused sequence.

4. The spatiotemporal series prediction method for carbon emissions from power plants as described in claim 1, characterized in that, The cross-domain entanglement encoder includes a time-domain compressor, a spatial-domain extractor, and a time-domain decompressor connected in sequence. The temporally entangled mask is fused with the fusion sequence to obtain a mask fusion sequence; The mask fusion sequence is compressed along the time dimension using a time-domain compressor to obtain latent features; The latent features are converted into a weight matrix of region-to-node using a spatial domain extractor. The surrogate representation is then calculated by combining the latent features, the weight matrix, and the introduced spatial entanglement mask. Finally, the surrogate representation is subjected to multi-layer attention calculation to obtain cross-region dependent features. The spatiotemporal features are obtained by restoring the cross-domain dependency features to their original time length using a time-domain decompressor and then adding the residuals of the fused sequence.

5. The spatiotemporal series prediction method for carbon emissions from power plants as described in claim 4, characterized in that, The matrix representation of the temporal entanglement mask is as follows: In the formula, Entangled in time masking, The visible proportion control coefficient, This is a fused time mask obtained by weighting multiple time masks.

6. The spatiotemporal series prediction method for carbon emissions from power plants as described in claim 5, characterized in that, The various time masks include a first time mask, a second time mask, a third time mask, and a fourth time mask, wherein: The first-time mask is obtained by a masking strategy that sets some time points to 0 with a certain probability and sets the remaining time points to 1. The second time mask is obtained by a masking strategy that randomly samples several consecutive intervals of each node and sets them to 0; The third time mask is obtained by selecting a preset time period and setting it to 0; The fourth time mask is obtained by setting the values ​​before the current time to 1 and the values ​​before the future time to 0.

7. The spatiotemporal series prediction method for carbon emissions from power plants as described in claim 4, characterized in that, Both the time-domain compressor and the time-domain decompressor include a first fully connected layer, a nonlinear activation function, and a second connected layer connected in sequence.

8. The spatiotemporal series prediction method for carbon emissions from power plants as described in claim 4, characterized in that, The spatial domain extractor includes a spatial proxy tensor, a structure-aware gated multi-head spatial entanglement attention mechanism, a standard multi-head attention layer, and a feedforward network. The node features of the latent features at each time step are extracted using the spatial proxy tensor, and the node features are converted into a weight matrix of regions to nodes; The structure-aware gated multi-head spatial entanglement attention mechanism is used to treat the weight matrix as the query vector and key vector, and the latent features as the value vector. The structure-aware gated multi-head spatial attention weight matrix is ​​calculated by combining the gating function. The introduced spatial entanglement mask is fused with the gated multi-head spatial attention weight matrix to obtain the multi-head fusion weight matrix. The multi-head fusion weight matrix is ​​then concatenated to obtain the surrogate representation. By using a standard multi-head attention layer, the surrogate representation is used as the key vector and value vector, and the latent features are used as the query vector to perform multi-layer attention computation, resulting in spatially restored features. The cross-regional dependent features are obtained by processing the spatially restored features using a feedforward network and then adding the residuals of the spatially restored features.

9. The spatiotemporal series prediction method for carbon emissions from power plants as described in claim 8, characterized in that, The formula for calculating the gated multi-head spatial attention weight matrix is ​​as follows: In the formula, For the first Spatial attention weight matrix of size, Represents a spatial proxy tensor. For node features, For gated functions, This indicates that Hadamard multiplies element by element. This is the weight matrix of the region to the node. This is the mapping matrix from regions to nodes. For spatial encoding matrix, This represents the transpose of the spatial encoding matrix. , , , For learnable parameters, Indicates the scaling factor. This is the transpose symbol.

10. The spatiotemporal series prediction method for carbon emissions from power plants as described in claim 8, characterized in that, The spatial entanglement mask includes a first spatial mask, a second spatial mask, and a third spatial mask, and the first spatial mask, the second spatial mask, and the third spatial mask are respectively fused with a gated multi-head spatial attention weight matrix; The first space mask is obtained by randomly selecting nodes in each row with a certain probability and setting them to 0, while setting the remaining nodes to 1. The second space mask is obtained by randomly selecting a number of proxy rows and setting all columns of the proxy rows to 0; The third space mask is obtained by randomly determining the columns of a continuous interval and setting all columns in that interval to 0.

11. The spatiotemporal series prediction method for carbon emissions from power plants as described in claim 1, characterized in that, The predictive decoder comprises a first linear layer, a nonlinear activation function, and a second linear layer connected in sequence.

12. The spatiotemporal series prediction method for carbon emissions from power plants as described in any one of claims 1-11, characterized in that, The main branch model stack The output of the previous layer of the cross-domain entanglement encoder is used as the input of the next layer of the cross-domain entanglement encoder. The output of the last layer of the cross-domain entanglement encoder is added to the residual of the fusion sequence to obtain the spatiotemporal hidden representation. Accordingly, the spatiotemporal hidden representation is processed using a predictive decoder to obtain the spatiotemporal sequence prediction of carbon emissions for the future window.

13. A spatiotemporal series prediction system for carbon emissions from power plants, characterized in that, This includes a data acquisition module and a pre-trained main branch model for carbon emission prediction. The main branch model comprises a multi-domain fusion coding unit, a cross-domain entangled encoder, and a prediction decoder connected in sequence, wherein: The data acquisition module is used to acquire data fragments of the power system's interconnected network diagram and the spatiotemporal sequence of carbon emissions within past windows, and input them into the main branch model; The main branch model is used to add the high-dimensional representation obtained by mapping the input data fragments to the multi-dimensional coding matrix using the multi-domain fusion coding unit to obtain the fusion sequence. The dimensions of the multi-dimensional coding matrix include space, time, and position. After compressing the fused sequence along the time dimension using a cross-domain entanglement encoder to obtain latent features, cross-regional dependent features are extracted from the latent features, and the cross-regional dependent features are restored to the original scale to obtain spatiotemporal features. By using a predictive decoder to process the spatiotemporal features, we can obtain the spatiotemporal sequence prediction of carbon emissions for the future window.

14. A training method for a spatiotemporal series prediction model of carbon emissions from power plants, characterized in that, Training the main branch model in the method as described in any one of claims 1-12 includes: Select a carbon emission monitoring dataset from power plants and convert the carbon emission monitoring data of power plants within the monitoring area into a spatiotemporal series sample; Based on a given historical window and a prediction window, a set step size sliding method is used to process spatiotemporal sequence samples to generate training sample pairs, which include historical sequence data and future real sequence data. Augmentation processing is performed on historical sequence data and association network graphs to obtain augmented sequence data and augmented association network graphs. The enhanced sequence data and enhanced correlation network are used as inputs to the enhanced branch model, and the historical sequence data and correlation network graph are used as inputs to the main branch model. The trained main branch model is used to predict carbon emissions. The enhanced branch model and the main branch model have the same structure.

15. The training method for a spatiotemporal series prediction model of carbon emissions for power plants as described in claim 14, characterized in that, The method further includes: Spatial heterogeneity enhancement is performed on the spatiotemporal features output by the main branch model and the enhanced spatiotemporal features output by the enhanced branch model, respectively, to obtain spatial clustering assignment results and enhanced spatial clustering assignment results; Temporal heterogeneity enhancement is performed on the spatiotemporal features output by the main branch model and the enhanced spatiotemporal features output by the enhanced branch model to obtain positive sample pairs and negative sample pairs.

16. The training method for a spatiotemporal series prediction model of carbon emissions for power plants as described in claim 15, characterized in that, The spatial heterogeneity enhancement is performed on the spatiotemporal features output by the main branch model and the enhanced spatiotemporal features output by the enhanced branch model to obtain spatial clustering assignment results and enhanced spatial clustering assignment results, including: The monitoring area is projected into multiple potential representation spaces, each representing space corresponding to a different operating condition; A set is generated based on each representation space. Each cluster embedding represents a latent factor used to distinguish differences in operating conditions; Using various clustering embeddings to analyze or enhance the spatiotemporal features of the first cluster. Each node in the time slice Clustering is performed on the embedded representation to obtain spatial clustering assignment results or enhanced spatial clustering assignment results.

17. The training method for a spatiotemporal series prediction model of carbon emissions for power plants as described in claim 15, characterized in that, The temporal heterogeneity enhancement is performed on the spatiotemporal features output by the main branch model and the enhanced spatiotemporal features output by the enhanced branch model to obtain positive sample pairs and negative sample pairs, including: spatiotemporal features at time step Embedded representations and enhanced spatiotemporal features at time steps The embedded representations of time are fused to obtain the time steps. Region-level embedding representation; Aggregate all power plants in time step The region-level embedding representation is used to obtain the time step. System-level representation; Region-level embeddings and system-level embeddings at the same time step are considered positive sample pairs, while region-level embeddings and system-level embeddings at different time steps are considered negative sample pairs.

18. The training method for a spatiotemporal series prediction model of carbon emissions for power plants as described in claim 15, characterized in that, The joint learning objective function used in model training includes the regional difference modeling loss based on spatial clustering assignment results and enhanced spatial clustering assignment results, the temporal situation modeling loss based on positive sample pairs and negative sample pairs, the prediction loss used to minimize the spatiotemporal sequence prediction of carbon emissions for quantifying future windows and the future real sequence data, and the multi-scale consistency loss used to quantify the consistency between the output of the main branch model and the output of the enhanced branch model.

19. The training method for a spatiotemporal series prediction model of carbon emissions for power plants as described in claim 18, characterized in that, The multi-scale consistency loss includes the consistency loss between the predictions of the main branch model and the future real sequence data, the consistency loss between the predictions of the augmented branch model and the future real sequence data, and the consistency loss between the predictions of the main branch model and the predictions of the augmented branch model.

20. The training method for a spatiotemporal series prediction model of carbon emissions for power plants as described in claim 18, characterized in that, The formula for the regional difference modeling loss is expressed as: In the formula, The objective function for optimizing the self-supervised augmentation task, Assign results to spatial clustering. To enhance spatial clustering assignment results, These are parameters used to control the smoothness of the softmax output. The first in the spatiotemporal features Each node in the time slice Embedded representation below, This represents the number of cluster centers.

21. The training method for a spatiotemporal series prediction model of carbon emissions for power plants as described in claim 18, characterized in that, The formula for the temporal situation modeling loss is expressed as: In the formula, This represents the embedding similarity discriminant function. for Activation function; Indicates the region-level embedding at the same time step With system-level embedding , For regional embedding at different time steps With system-level embedding , This indicates the number of power plants.

22. The training method for a spatiotemporal series prediction model of carbon emissions for power plants as described in claim 18, characterized in that, The formula for the multi-scale consistency loss is expressed as follows: In the formula, For balance coefficient, Representing square Norm, For average pooling operators, The predicted values ​​of the main branch model, For future real sequence data, To enhance the predictions of the branching model, The time scale is used to represent the step size of average pooling. This is the step size for average pooling.

23. The training method for a spatiotemporal series prediction model of carbon emissions for power plants as described in claim 18, characterized in that, The formula for the joint learning objective function is expressed as follows: In the formula, , , , It is an adjustable hyperparameter. To predict losses, Modeling loss for regional differences, For time-series situation modeling loss, This is a multi-scale consistency loss.

24. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the spatiotemporal sequence prediction method for carbon emissions of power plants as described in any one of claims 1-12 or the training method for the spatiotemporal sequence prediction model for carbon emissions of power plants as described in any one of claims 14-23.

Images