Urban power load time sequence prediction method and system

Abstract

The invention discloses an urban power load time sequence prediction method and system, and the method comprises the steps: obtaining multi-dimensional time sequence data related to a power load, and constructing a power knowledge graph based on the multi-dimensional time sequence data; utilizing a relational graph convolutional neural network to carry out structured embedding representation learning on nodes of the electric power knowledge graph to obtain an embedding vector used for representing a structural dependency relationship between regions; splicing the embedded vector with a plurality of time sequence fragments obtained by dividing the multi-dimensional time sequence data to obtain a plurality of structure enhanced sequence fragments; inputting the plurality of structure enhancement sequence fragments into an FEDform model for modeling processing to obtain a time sequence load prediction value; according to the method, high-precision power load prediction is realized by explicitly depicting the spatial dependency relationship and fusing the periodic trend and uncertainty modeling, and the method has higher adaptability, stability and interpretability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load prediction and intelligent scheduling, and particularly relates to a city power load time series prediction method and system. BACKGROUND

[0002] With the continuous growth of urban power grid load and the deepening of smart grid construction, how to realize large-scale and fine power load prediction has become a key link to ensure the safe and stable operation and scientific scheduling of the power grid. Especially under the background of "double carbon", the large-scale access of new energy, the concentrated charging of electric vehicles and the increasing uncertainty of distributed energy have made the spatio-temporal evolution characteristics of urban power grid load increasingly complex. Among them, distributed photovoltaic as a typical distributed renewable energy, its output is significantly affected by weather, irradiance and temperature and other factors, with strong time-varying and volatility. With the large-scale access of distributed photovoltaic and other new energy, the urban power grid faces the challenges of increased output volatility at the source side and complex spatio-temporal distribution of load, bringing new challenges to city-level load prediction. Traditional load prediction methods relying on single time series modeling have been difficult to meet the demand for accurate description and response prediction of the coupling evolution of multi-regional and multi-dimensional factors.

[0003] Currently, city-level power load prediction still faces the following key challenges:

[0004] (1) Lack of effective modeling of structural dependence between regions: Existing methods are mostly based on modeling of single-point or independent regional load time series, ignoring the complex topological structure of the power grid, such as the natural existence of device power supply links, regional adjacency relationships, and meteorological influence range. Especially in the power grid containing distributed photovoltaic, there is often geographical proximity and output-load interaction between distributed photovoltaic sites and user loads. Without explicitly modeling the graph structure dependence between regions, it is difficult to capture the disturbance transmission effect of synchronous fluctuation of distributed output on local load, affecting the overall prediction accuracy and robustness.

[0005] (2) Weak fusion capability of heterogeneous multi-source data: Urban power grid operation involves multi-dimensional factors such as substation load monitoring, meteorological data, user behavior, price mechanism, holiday information, real-time output of distributed photovoltaic, component state and weather forecast. The data types are diverse and the semantics are complex. Traditional modeling methods often have difficulty in integrating data from different sources and different granularities, lack of structured expression ability at the knowledge level, resulting in low information utilization rate and difficulty in forming a knowledge-driven prediction reasoning framework.

[0006] (3) Long-term modeling capability is insufficient and the calculation bottleneck is prominent: City-level load forecasting often involves long window historical data (such as 7 days, 30 days) and long period rolling prediction (such as 72 hours), and also needs to consider the characteristics of daily cycle and weekly rhythm of photovoltaic output, which puts higher requirements on the long-term dependent modeling capability of the model. Traditional time series models (such as LSTM, GRU) have problems of gradient disappearance and memory decay when the sequence is long.

[0007] (4) The fusion of graph structure and time series model is insufficient: Some studies try to combine graph neural networks with time series prediction models, but mostly use simple splicing or concatenation, which fails to deeply integrate graph structure information and time series attention mechanism, resulting in the inability to model the regional topology and load evolution law simultaneously, making it difficult to effectively capture patterns such as "photovoltaic-load" coupling fluctuations, "region-neighborhood" linkage responses, and limiting the ability to describe complex evolution behaviors.

[0008] (5) The prediction system lacks distributed processing and knowledge reasoning capability: Current city load prediction platforms are mostly based on centralized databases and isolated model deployment, which is difficult to meet the storage, calculation and real-time update needs of super large scale graph data. Especially after integrating a large number of distributed photovoltaic access information, the size of the graph rapidly expands and the frequency of prediction continues to increase, and the centralized structure is prone to become a bottleneck in data writing, query response and task scheduling, restricting the scalability and response efficiency of the system. At the same time, the existing model lacks reasoning capability based on graph structure, making it difficult to make intelligent judgments beyond experience for sudden events and rare combination patterns.

[0009] Therefore, there are problems such as insufficient multi-source data fusion capability, lack of structure-dependent modeling, low efficiency of long-time series modeling, and weak system feedback loop capability in current urban-level power grid load prediction. Especially in the complex scenario where the source-side volatility is significantly enhanced under the background of large-scale access of distributed photovoltaic and other renewable energy sources, how to effectively fuse the distributed photovoltaic output characteristics at the power grid level and realize the collaborative prediction of load and source-side volatility has become an important research direction in current load modeling. In the related technology, the patent application with the publication number CN116070804A proposes a power system load prediction method based on knowledge graph and data driving. This method focuses on the change trend modeling and network selection problem of power load under the influence of uncertain factors, and does not involve complex topological structure. It extracts the correlation of change rate by constructing a knowledge graph and selects a matching LSTM sub-model for prediction, which belongs to the paradigm of classification first and then prediction. Although it has a certain flexibility in processing the selection of prediction network, the model updating capability is weak and the adaptability is poor. The patent application with the publication number CN118036471A proposes a power grid load modeling and analysis method based on knowledge graph and digital twinning. This method supplements data integrity through digital twinning and knowledge graph, and relies on model integration to improve precision. However, it does not depict the structure-dependent relationship of the power system. Due to the lack of modeling of the physical structure of the equipment, it is difficult to support actual power grid topology perception. Moreover, it uses LSTM and XGBoost parallel modeling, and then uses weighted fusion to improve precision, but neither of the two models utilizes the structural features between entities. In the literature "Application of Graph Neural Networks in Power Systems from Perception-Prediction-Optimization: A Review", China Electric Power, Vol. 57, No. 12, Li Zhuo et al., graph neural networks are reviewed. This research focuses on the theoretical adaptability and evolution trend of graph neural networks in power systems, lacks specific problem-oriented and engineering landing analysis, and mainly stays at the theoretical level. SUMMARY

[0010] The technical problem to be solved by the present application is how to accurately predict the power load of a complex urban power grid.

[0011] The present application solves the above technical problems by the following technical means:

[0012] A city power load time series prediction method is proposed, which comprises the following steps:

[0013] Obtain multi-dimensional time series data related to power load, and construct a power knowledge graph based on the multi-dimensional time series data;

[0014] Use a relational graph convolutional neural network to perform structured embedding representation learning on the nodes of the power knowledge graph, to obtain embedding vectors for representing the structure-dependent relationship between regions;

[0015] The embedding vector is spliced with a plurality of time sequence segments obtained by dividing the multi-dimensional time sequence data, to obtain a plurality of structure-enhanced sequence segments.

[0016] The plurality of structure-enhanced sequence segments are input into a FEDformer model for modeling processing, to obtain a time sequence load prediction value.

[0017] Further, the multi-dimensional time sequence data related to the power load includes a regional power load sequence in a previous time period, distributed photovoltaic operation data, environmental meteorological variables, and holiday and behavior labels.

[0018] Further, the power knowledge graph is constructed based on the multi-dimensional time sequence data, including:

[0019] The multi-dimensional time sequence data is structured and processed to obtain standardized multi-dimensional time sequence data;

[0020] The standardized multi-dimensional time sequence data is associated and mapped with a power grid topology to construct a power knowledge graph.

[0021] Further, the power knowledge graph is represented as G=(V,E,R,N r ), wherein V is a node set, the node represents different types of power entities, and each node v has an attribute vector E is an edge set, the edge represents a connection between nodes; R is a set of edge types, the edge type represents a connection relationship of different relationship types; N r represents a set of neighbor nodes under the relationship r∈R;

[0022] The types of graph nodes include regional nodes, device nodes, user nodes, meteorological factor nodes, and photovoltaic system nodes.

[0023] The semantic association includes power supply relationship, device ownership, meteorological influence, photovoltaic access, and output influence.

[0024] Further, a graph database is used as a storage and access platform for storing the power knowledge graph.

[0025] Further, the relationship type graph convolutional neural network is used to structure and embed the nodes of the power knowledge graph to learn an embedding vector for representing the structure-dependent relationship between regions, including:

[0026] The relationship type graph convolutional neural network is used to aggregate the upstream and downstream neighbor node states and relationship types of the nodes in the power knowledge graph to generate an embedding vector for representing the structure-dependent relationship between regions.

[0027]

[0028] In the formula: denotes the embedding vector of node v at the (l+1)-th layer; R is a set of edge types, r e R denotes the semantic relationship between nodes; N r (v) denotes the set of neighbor nodes connected with node v under the relationship r; denotes the transformation matrix for relationship r in the l-th layer; denotes the hidden representation vector of node u at the l-th layer, node u is a neighbor node of node v under the relationship r; c v,r denotes the number of neighbor nodes of node v under the relationship r; σ denotes the activation function; denotes the self-loop weight matrix.

[0029] Further, the embedding vector is spliced with the plurality of time sequence segments obtained by dividing the multi-dimensional time sequence data to obtain a plurality of structure-enhanced sequence segments, comprising:

[0030] The normalized multi-dimensional time sequence data is divided according to a window with a length of p to obtain N time sequence segments, wherein the j-th time sequence segment is:

[0031]

[0032] In the formula: denotes the j-th time sequence segment, Y p(j-1)+1:pj denotes the division of the multi-dimensional time sequence data from p(j-1)+1 to pj, and L is the total length of the normalized multi-dimensional time sequence data;

[0033] The embedding vector is spliced with the j-th time sequence segment to obtain a corresponding structure-enhanced sequence segment:

[0034]

[0035] In the formula: t j denotes the j-th structure-enhanced sequence segment, InputRSB denotes a residual block mapping function, denotes the mask corresponding to the j-th time sequence segment, and indicates that the missing value is 0, denotes the graph embedding vector, PE j denotes the position encoding information.

[0036] Further, the position encoding information is obtained by using a sine-cosine encoding, and the formula is represented as:

[0037]

[0038] In the formula: PE (pos,2i) denotes the sine position encoding value of the even dimension corresponding position pos in the encoding vector, pos denotes the position index carried by the current element, d model denotes the dimension size of the model vector, PE(pos,2i+1) denotes the sine position encoding value corresponding to the odd dimension position pos in the vector, and i denotes half of the current position index.

[0039] Further, the plurality of structure-enhanced sequence segments are input into the FEDformer model for modeling processing to obtain a time series load prediction value, including:

[0040] Each structure-enhanced sequence segment is input into the FEDformer model for modeling processing to obtain a time series load output feature, and the formula is represented as:

[0041]

[0042] In the formula, o j denotes the load prediction feature output by the FEDformer model on the jth time window, denotes that the jth time series segment is valid, otherwise the jth time series segment is invalid, t1,...,t j denotes the jth time window, and FEDformer denotes a deep time series prediction model based on a frequency domain modeling mechanism;

[0043] The time series load output feature is subjected to residual mapping to obtain a time series load prediction value.

[0044] Further, the point prediction result is subjected to linear mapping function f point obtained:

[0045]

[0046] In the formula: denotes the point prediction load value of the jth time window, W p , b p are respectively a weight and a bias parameter of point prediction.

[0047] The probability prediction result is output by using a quantile loss function:

[0048]

[0049] In the formula: denotes the prediction result of the jth window at the quantile point q, are respectively parameters of corresponding quantile prediction

[0050] Further, the relational graph convolutional neural network and the FEDformer model are both models pre-trained through end-to-end training, and the loss function used for training is:

[0051] L total = αL MSE +(1-α)L q

[0052]

[0053] In the formula: L total represents the total loss function, L MSE represents the point prediction loss function, L q represents the probability prediction loss function, alpha is in [0, 1] and is used to control the weight of the point prediction loss function and the probability prediction loss function, y j represents the real future load sequence value in the jth prediction window, represents the predicted future load sequence value in the jth prediction window, N represents the number of time sequence segments, (·) + = max (0, ·), q is in (0, 1) and represents the predicted quantile value, represents the prediction result of the jth window at the quantile point q.

[0054] Further, after the plurality of structure-enhanced sequence segments are input into the FEDformer model for modeling processing to obtain the time sequence load prediction value, the method further comprises:

[0055] Based on the residual error and the error between the time sequence load prediction value and the actual load value, the related nodes in the power knowledge graph are updated as new attributes.

[0056] Further, the time sequence load prediction value comprises a point prediction result and a probability prediction result, and the method further comprises:

[0057] Scheduling the power equipment according to the point prediction result;

[0058] Comparing the probability prediction result with the real-time load, and performing risk warning when the real-time load exceeds the probability prediction result.

[0059] In addition, the present application also proposes a city power load time sequence prediction system, comprising:

[0060] A data acquisition layer is used to acquire multi-dimensional time sequence data related to power load;

[0061] A data processing layer is used to construct a power knowledge graph based on the multi-dimensional time sequence data;

[0062] A modeling and reasoning layer is used to perform structured embedding representation learning on the nodes of the power knowledge graph by using a relational graph convolutional neural network, to obtain embedding vectors for representing the structural dependency relationship between regions; the embedding vectors are spliced with a plurality of time sequence segments divided from the multi-dimensional time sequence data to obtain a plurality of structure-enhanced sequence segments for inputting into a FEDformer model for modeling processing to obtain a time sequence load prediction value.

[0063] The present application has the advantages that:

[0064] (1) The present application takes power grid operation topology and regional relationship as the core skeleton, integrates multi-source heterogeneous data to form multi-dimensional time sequence data, constructs a power knowledge graph, and after the power knowledge graph is constructed, a relational graph convolutional neural network is used to fully capture the structural dependency relationship and semantic information between different types of entity nodes in the power knowledge graph, and an embedding vector for representing the structural dependency relationship between regions is obtained, which explicitly describes the spatial structural dependency relationship between regions. Then the embedding vector is spliced with a plurality of time sequence segments obtained by dividing the multi-dimensional time sequence data, to obtain a plurality of structure-enhanced sequence segments. The time sequence segment is helpful for the FEDformer model to learn the time evolution mode (such as periodicity, trend change, etc.), so as to facilitate the FEDformer model to extract the periodic signal through frequency domain modeling, and the structural embedding of the node in the embedding vector in the power knowledge graph is taken as an additional input feature to participate in the modeling process of the FEDformer, which is helpful for the model to learn more comprehensive information expression, and realizes high-precision power load prediction. Compared with the traditional prediction method which only depends on single-region time sequence, the present application not only explicitly describes the spatial dependency relationship, but also integrates periodic trend and uncertainty modeling, and has stronger adaptability, stability and interpretability.

[0065] (2) In the generation process of the structure-enhanced sequence segment, a mask mechanism is introduced to realize automatic filtering of missing values and invalid data, so that the model can skip missing or abnormal points and only train and predict on reliable data, further improving the prediction accuracy.

[0066] (3) The relational graph convolutional neural network used for graph embedding learning is not trained or used alone, but as part of the input token of the FEDformer model, which participates in training and prediction together with time sequence information. This design allows the prediction error to be back propagated to the relational graph convolutional neural network through the loss function, so as to dynamically adjust the graph embedding representation to better fit the prediction target.

[0067] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 is a flowchart of a city power load time sequence prediction method according to an embodiment of the present application;

[0069] Figure 2 is a schematic diagram of the principle of city power load time sequence prediction in an embodiment of the present application;

[0070] Figure 3 is an example diagram of a power knowledge graph in an embodiment of the present application;

[0071] Figure 4 is a structural schematic diagram of a city power load time sequence prediction system according to an embodiment of the present application. DETAILED DESCRIPTION

[0072] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in a clear and complete manner in conjunction with the embodiments of the present application. Obviously, the described embodiments are only some, but not all of the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0073] As shown in Figures 1 to 2 , the present application provides a city power load time sequence prediction method, which comprises the following steps:

[0074] S10, acquiring power load related multi-dimensional time sequence data, and constructing a power knowledge graph based on the multi-dimensional time sequence data;

[0075] Specifically, the present embodiment can use a multi-source perception module to collect power load related multi-dimensional time sequence data from a city power grid system in real time.

[0076] S20, using a relational graph convolutional neural network to perform structured embedding representation learning on nodes of the power knowledge graph, to obtain embedding vectors for representing inter-regional structural dependency relationships;

[0077] It should be noted that in a power system, regional load is often affected by the states of upstream substations, surrounding user loads and distributed photovoltaic devices, and these influences form structural dependencies through physical topology and power supply links. The relational graph convolutional neural network R-GCN can model these dependencies using the power knowledge graph, perform embedding representation learning on the graph nodes, and explicitly model the inter-regional structural dependency relationships, thereby explicitly capturing the disturbance conduction paths between regions. This enables the final load fluctuation prediction to comprehensively consider changes in other entities related to the current node, thereby improving the physical rationality and context awareness of the modeling.

[0078] S30, splicing the embedding vectors with a plurality of time sequence segments divided from the multi-dimensional time sequence data, to obtain a plurality of structure enhanced sequence segments;

[0079] It should be noted that the embodiment divides the long time sequence of multi-dimensional time series data into shorter time sequence segments (such as a window of 12 hours) that can be processed by the FEDformer model, and splices the embedding vectors to form a structure-enhanced sequence segment as the input token of the FEDformer model. The time sequence segment helps the FEDformer model capture local periodic characteristics and trend patterns (for example, daytime load rises and nighttime drops), and each token corresponds to an "event window", facilitating the FEDformer model to extract periodic signals through frequency domain modeling.

[0080] It should be noted that the traditional time series model usually only relies on a single point time series when training or load forecasting, and lacks structural context knowledge. After introducing the R-GCN in the embodiment, each node carries its structural embedding in the graph as an additional input feature to participate in the modeling process of the subsequent FEDformer model. These graph embeddings encode the node's position, adjacency relationship and propagation path in the power grid, helping the model to learn more comprehensive information representation and have stronger generalization ability, especially in scenarios such as new regions, device access or partial data missing, which can maintain prediction accuracy and stability.

[0081] S40, inputting a plurality of structure-enhanced sequence segments into the FEDformer model for modeling processing to obtain time series load prediction values.

[0082] The embodiment realizes the deep coupling modeling of structure + time by constructing a power knowledge graph and structurally embedding the nodes of the power knowledge graph using a relational graph convolutional neural network R-GCN to learn the representation of the nodes, extract graph structure features, and then combine the FEDformer model with strong frequency domain modeling capability to jointly predict the load time series. Thus, the embodiment realizes accurate modeling and prediction of regional load disturbance caused by photovoltaic output fluctuation, has high generalization, self-adaptability and structural explainability, and is particularly suitable for urban-level complex power grid dispatching systems.

[0083] As a further preferred technical solution, the acquisition of the multi-dimensional time series data related to the power load includes but is not limited to the regional power load sequence in the previous time window, device operating state, user behavior, environmental meteorological factors, holiday information, and key variables such as distributed photovoltaic output, irradiance, component state, and distributed photovoltaic operation data. Among them, the regional load historical sequence: represents the active power load of region r at time t; environmental meteorological variables: w t = [T t , H t , W t ], T t , H t , W tTemperature, humidity, wind speed corresponding to time t, respectively; holiday and behavior label: c t ∈{0,1} d Holiday flag, season type, etc.; Topology information: provided by SCADA / GIS system, representing the connection relationship of regions, devices and links; Distributed photovoltaic operation data Actual output of distributed photovoltaic in region r at time t Irradiance G of a representative photovoltaic assembly in the region t Temperature T of the assembly mode .

[0084] As a further preferred technical solution, in step S10, the power knowledge graph is constructed based on the multi-dimensional time series data, specifically including the following steps:

[0085] S11, structurally process the multi-dimensional time series data to obtain standardized multi-dimensional time series data;

[0086] Specifically, the multi-dimensional time series data is structurally processed to form a standardized representation that can be used for graph modeling and time series modeling, specifically including missing value processing and mask marking, feature normalization and scaling processing, wherein:

[0087] (1) Missing value processing and mask marking

[0088] To deal with missing and abnormal data, a binary mask m is constructed 1:L ∈{0,1} L Wherein:

[0089]

[0090] In the formula: m t Indicates whether the t-th time step is a valid observation point, m 1:L Indicates a binary mask sequence with a time series segment length of L, composed of a mask m t At each time step, Indicates the original observation value of the r-th physical quantity or feature at the t-th time step, and L indicates the time step number of the time series segment, which is used to unify the input length of the model.

[0091] It should be noted that the mask runs through the entire modeling process and is used to control the participation of valid data in training and reasoning.

[0092] (2) Feature normalization and scaling processing

[0093] This embodiment adopts a grouping normalization method to scale different features to a relatively uniform interval:

[0094]

[0095] wherein: is the original data, is the global maximum and minimum value of the feature,

[0096] The embodiment scales the temperature and humidity features to within [0, 1] through feature normalization and scaling processing, and other data is scaled to within [0, 10].

[0097] It should be noted that, in order to improve the overall quality of the data and ensure that the key information is retained, through this series of processing, the usability and effectiveness of the data are enhanced, providing accurate and reliable data input for the training and prediction tasks of the model layer.

[0098] S12, the standardized multi-dimensional time series data is associated and mapped with the power grid topology structure to construct a power knowledge graph.

[0099] It should be noted that, in the embodiment, the structured multi-dimensional time series data is associated and mapped with the power grid topology structure to construct a power knowledge graph to express the structured semantic knowledge of various power entities and their relationships in the city power grid. The graph integrates traditional power operation elements and key node information in the new energy scenario, and has good scalability and semantic expression ability.

[0100] As a further preferred technical solution, the power knowledge graph is represented as G=(V, E, r, N r ), wherein V is a node set, the node represents different types of power entities and each node v has an attribute vector E is an edge set, the edge represents the connection between nodes; R is an edge type set, the edge type represents the connection relationship of different relationship types; N r represents the neighbor node set under the relationship r e R;

[0101] The types of graph nodes include but are not limited to regional nodes, device nodes, user nodes, meteorological factor nodes, and photovoltaic system nodes, wherein the regional node represents a power grid division region or a transformer station; the device node includes transformer, circuit breaker, power transmission line and other core power facilities; the user node represents different categories of power users such as industry, commerce, and residents; temperature, humidity, wind speed, and irradiance are modeled as graph nodes through the abstract entity type of "meteorological factor node"; the photovoltaic system node represents each distributed photovoltaic station and its characteristics (installation location, rated capacity, output state, etc.).

[0102] The semantic association includes but is not limited to power supply relationship, device ownership, meteorological influence, photovoltaic access and output influence, wherein the power supply relationship represents that the edge connects the transformer substation and the user node as a power supply path, the device ownership represents which region or site a certain device belongs to, the meteorological influence represents the effect of meteorological factors on a certain region or node, the photovoltaic access represents the connection between the distributed photovoltaic node and the accessed region or feeder, and the output influence represents the influence relationship of photovoltaic output on regional load or voltage stability.

[0103] Specifically, each node also records attribute values such as device type, installation location, real-time state, etc. As shown in the following table: Figure 3 As shown: CB-23 circuit breaker and bus B1, downstream charging terminal U-9 build a three-node power supply graph, the system marks the type of CB-23 as "circuit breaker", B1 as "bus", U-9 as "user terminal", the edge relationship as "power supply connection", and sets the weight as "current value", which is used for subsequent graph modeling.

[0104] As a further preferred technical solution, the embodiment adopts a graph database as the storage and access platform for storing the power knowledge graph.

[0105] Specifically, the embodiment adopts a graph database (such as Neo4j, TigerGraph or JanusGraph, etc.) as the storage and access platform for the power knowledge graph, which has the following functions:

[0106] Multi-entity type modeling: different node types such as regions, devices, time, and weather are distinguished by label (Label) mechanism;

[0107] Multi-relation modeling and attribute labeling: support directed edge representation of power supply relationship, spatial adjacency, influence relationship, etc., and can add weight, timestamp, type, etc. attributes to the edge;

[0108] Subgraph and path query capability: quickly support K-hop neighbor retrieval, path matching and subgraph filtering through Cypher statement, which is used to build input graph or local graph convolution substructure;

[0109] Graph computing engine linkage: the graph database can be integrated with Spark GraphX or Flink Gelly to realize batch graph embedding calculation and high concurrency access;

[0110] Real-time writing and dynamic updating mechanism: when the model prediction result or new event occurs, the node state attribute can be updated back to the graph database to realize the adaptive closed loop of "prediction-graph evolution-re-prediction".

[0111] As a further preferred technical solution, the step S20: utilizing a relational graph convolutional neural network to perform structured embedding representation learning on nodes of the power knowledge graph, to obtain embedding vectors for representing inter-regional structural dependency relationships, specifically comprising:

[0112] Utilizing the relational graph convolutional neural network, aggregating the upstream and downstream neighbor node states and relationship types of the nodes in the power knowledge graph, to generate embedding vectors for representing inter-regional structural dependency relationships:

[0113]

[0114] In the formula: represents the embedding vector of node v at the l+1 layer; R is a set of edge types, r e R represents the semantic relationship between nodes; N r (v) represents a set of neighbor nodes connected to node v under relationship r; represents a transformation matrix for relationship r in the l layer; represents the hidden representation vector of node u at the l layer, node u being a neighbor node of node v under relationship r; c v,r = |N r (v) | is a normalization coefficient, representing the number of neighbors of node v under relationship r, for balancing the information aggregation influence of different degree points; sigma represents an activation function; represents a self-loop weight matrix.

[0115] It should be noted that after the power knowledge graph is constructed, the relational graph convolution network R-GCN is used for structured embedding representation learning of the graph nodes, so as to fully capture the structural dependency relationship and semantic information between different types of entities, and explicitly model the structural dependency relationship between regions. In the power system, the regional load is often affected by the states of the upstream substations, the surrounding user loads and the distributed photovoltaic devices, and these influences form a structural dependency through the physical topology and the power supply link. The R-GCN can model these dependencies by using the graph structure, and extract the relationship of "who influences who" through the aggregation operation of the adjacent node states, so as to explicitly capture the disturbance conduction path between regions. For example, the entity types and edge relationships in the power knowledge graph are very diverse, including regional nodes, meteorological nodes, device nodes, and various edge types such as "power supply", "meteorological influence" and "photovoltaic access". The R-GCN natively supports modeling of such multi-type graph structure, and different relationships transmit information through independent transformation matrices, thereby retaining the structural characteristics of different semantics. This enables the model to understand complex semantics such as "the influence of meteorological changes on photovoltaic output" or "the disturbance of a certain device state to the regional load", which enables the model to consider the changes of other entities related to the current node when predicting load fluctuations, improves the physical rationality and context awareness of the model, and improves the representation ability and interpretability of the model to the real system.

[0116] In addition, the traditional time series model usually only depends on a single point time series during training, and lacks structural context knowledge. After introducing the R-GCN in the embodiment, each node carries its structural embedding in the graph as an additional input feature to participate in the modeling process of the FEDformer, thereby introducing structural prior information and improving the generalization ability of the model. These graph embeddings encode the position, adjacency relationship and propagation path of the node in the power grid, which helps the model to learn more comprehensive information representation and has stronger generalization ability, especially in scenarios such as new regions, device access or partial data missing, which can maintain prediction accuracy and stability.

[0117] As a further preferred technical solution, the step S30 of splicing the embedding vector with the plurality of time sequence segments obtained by dividing the multi-dimensional time sequence data to obtain a plurality of structure-enhanced sequence segments comprises the following steps:

[0118] S31, the standardized multi-dimensional time sequence data is divided according to a window with a length of p to obtain N time sequence segments, wherein the jth time sequence segment is:

[0119]

[0120] In the formula: Y p(j-1)+1:pjdenotes the division of the multi-dimensional time series data from p(j-1)+1 to pj, L seq is the total length of the standardized multi-dimensional time series data;

[0121] S32, splicing the embedding vector and the jth time sequence segment to obtain a corresponding structure-enhanced sequence segment is:

[0122]

[0123] In the formula: t j denotes the jth structure-enhanced sequence segment, InputRSB denotes a residual block mapping function, denotes the mask corresponding to the jth time sequence segment, and denotes that the missing value is set to 0, denotes the graph embedding vector, PE j denotes the position encoding information.

[0124] It should be noted that the embodiment divides the multi-dimensional time series data into shorter segments (such as a window of 12 hours) that can be processed by the model, and encodes the input token of the FEDformer model in the form of “segment + mask”. Each time sequence segment is beneficial to capture local periodic characteristics and trend patterns (for example, daytime load rises and nighttime drops), and each token corresponds to an “event window”, which facilitates the FEDformer model to extract periodic signals through frequency domain modeling.

[0125] And the mask mechanism is introduced to automatically filter missing values and invalid data, which allows the model to skip missing or abnormal points and only train and predict on reliable data.

[0126] As a further preferred technical solution, the position encoding information is obtained by sine-cosine encoding the time sequence segment, and the formula is:

[0127]

[0128] In the formula: PE (pos,2i) denotes the sine position encoding value of the even dimension corresponding to the position pos in the encoding vector, pos denotes the position index carried by the current element, d model denotes the dimension size of the model vector, PE (pos,2i+1) denotes the sine position encoding value of the odd dimension corresponding to the position pos in the vector, and i denotes half of the current position index.

[0129] It should be noted that the embodiment encodes the even position of the time sequence segment by using the sine encoding mode, and encodes the odd position of the time sequence segment by using the cosine encoding mode, thereby obtaining the position encoding information corresponding to the time sequence segment.

[0130] As a further preferred technical solution, the step S40: inputting the plurality of structure-enhanced sequence fragments into the FEDformer model for modeling processing to obtain the time series load prediction value, specifically comprising the following steps:

[0131] S41, input each structure-enhanced sequence fragment into the FEDformer model for modeling processing to obtain the time series load output feature, which is expressed by the formula:

[0132]

[0133] In the formula: o j represents the load prediction feature output by the FEDformer model on the jth time window, represents that the jth time series fragment is valid, otherwise the jth time series fragment is invalid, t1,...,t j represents the jth time window, and FEDformer represents a deep time series prediction model based on a frequency domain modeling mechanism.

[0134] S42, residual mapping is performed on the time series load output feature to obtain the time series load point prediction value and the quantile prediction value.

[0135] Specifically, the point prediction result is obtained by a linear mapping function f point Obtain:

[0136]

[0137] In the formula: represents the point prediction load value of the jth time window, W p , b p are the weight and bias parameters of point prediction, respectively.

[0138] The probability prediction result is output by a quantile loss function:

[0139]

[0140] In the formula: represents the prediction result of the jth window at the quantile point q, are the parameters corresponding to the quantile prediction

[0141] The method further comprises: scheduling the power equipment according to the point prediction result.

[0142] The probability prediction result is compared with the real-time load, and a risk warning is performed when the real-time load exceeds the probability prediction result.

[0143] It should be noted that by transmitting the load prediction value to the application layer, it is used to guide power resource scheduling, early warning, etc. Among them, point prediction is used for daily fine scheduling tasks, such as load curve tracking and power balance calculation; probability prediction provides a prediction interval and risk perception ability, such as early warning perception of rare peak or sudden abnormality.

[0144] Specifically, point prediction aims to generate an accurate numerical estimate for each future time point for the rigid scheduling needs of power grid operation, for example, if the model predicts that the load of a certain area at 14:00 is 120.5kW, the corresponding power distribution strategy can be arranged. The goal of probability prediction is to provide a confidence interval reflecting the uncertainty of the predicted value, and to enhance the risk perception ability of the system, for example: if the model outputs 0.1, 0.5, 0.9 quantile prediction as [115.0, 120.5, 127.0]kW, and the actual load is 130.2kW, it means that a rare high load situation has occurred, and the system will trigger an overrun alarm.

[0145] Specifically, the EDformer model is composed of a multi-layer causal attention mechanism and a frequency domain decomposition module. When the FEDformer model performs power load prediction, it not only utilizes the change trend of historical time series data, but also fuses the structural information of regional nodes in the power grid. By introducing graph embedding representation, it can perceive the spatial topological dependence and upstream and downstream relationship between regions, and simultaneously consider time evolution and structural correlation in time series modeling, thereby more accurately depicting the periodic fluctuations of the load and improving the accuracy and robustness of the prediction. The structure-enhanced sequence fragments are input into the FEDformer model for modeling processing, and the FEDformer model will efficiently model these structure-enhanced sequences through frequency domain attention mechanism, trend and seasonal decomposition module, etc., to predict the future power load value. In the prediction process, the EDformer model can learn the joint pattern of local load change and global structural features within each time window, which helps the model identify complex "time series-structure" coupling rules, such as whether the change of the load of a certain region is affected by the fluctuation of upstream photovoltaic or the state of neighboring equipment.

[0146] As a further preferred technical solution, after the plurality of structure-enhanced sequence fragments are input into the FEDformer model for modeling processing to obtain the time series load prediction value, it further includes:

[0147] Based on the residual error and error between the time series load prediction value and the actual load value, the related nodes in the power knowledge graph are updated as new attributes.

[0148] It should be noted that the prediction result of the embodiment can be written back to the graph database in real time, and the node attributes (such as photovoltaic module state, prediction error, load offset, etc.) in the power knowledge graph are updated, and the system has a closed loop capability of “knowledge-prediction-feedback”. At the same time, the prediction error and the key mismatch can be automatically converted into new knowledge facts through root cause tracking, improving the evolution ability of the graph and the continuous adaptability of the model, and effectively supporting robust prediction in a dynamic environment (such as photovoltaic installed capacity adjustment, frequent changes in operating state, etc.).

[0149] As a further preferred technical solution, the relational graph convolutional neural network and the FEDformer model are both pre-trained end-to-end models, and the loss function used in the training is:

[0150] L total =αL MSE +(1-α)L q

[0151]

[0152] In the formula, L total represents the total loss function, L MSE represents the point prediction loss function, L q represents the probability prediction loss function, and α∈[0,1] is used to control the weight of the point prediction loss function and the probability prediction loss function, y j represents the real future load sequence value in the jth prediction window, represents the predicted future load sequence value in the jth prediction window, N represents the number of time series segments, and (·) + =max(0,·), q∈(0,1) represents the predicted quantile value, represents the prediction result of the jth window at the quantile point q.

[0153] It should be noted that in the embodiment, the graph embedding is not trained or used separately, but is part of the FEDformer model input token, and participates in training and prediction together with time series information. This design allows the prediction error to be back propagated to the graph neural network through the loss function, thereby dynamically adjusting the graph embedding representation to better fit the prediction target. Compared with the hard coding method of static graph features, end-to-end joint optimization can continuously enhance the synergy effect of structure representation and time series prediction, significantly improving the overall prediction accuracy and system adaptability in multi-region, multi-scale prediction tasks.

[0154] In addition, as shown in Figure 4 Another embodiment of the present application also provides a city power load time series prediction system, which comprises:

[0155] a data collection layer configured to acquire multi-dimensional time-series data related to power load;

[0156] a data processing layer configured to construct a power knowledge graph based on the multi-dimensional time-series data;

[0157] a modeling and reasoning layer configured to perform structured embedding representation learning on nodes of the power knowledge graph by using a relational graph convolutional neural network, to obtain embedding vectors for representing structural dependency relationships between regions; and to splice the embedding vectors and a plurality of time-series segments divided from the multi-dimensional time-series data to obtain a plurality of structure-enhanced sequence segments for inputting into a FEDformer model for modeling processing, to obtain time-series load prediction values.

[0158] As a further preferred technical solution, the embodiment integrates multiple types of sensors in the data collection layer and deploys them at key equipment points (such as power distribution rooms, transformer nodes, charging piles, feeder breakers, regional busbars, electric energy meters, and distributed photovoltaic access points) of urban charging networks and power distribution sides, to collect operating variables including active load, current, voltage, frequency, power factor, photovoltaic output, light intensity, component temperature, switch state, and external weather. The multiple types of sensors work based on principles such as electromagnetic induction, Hall elements, thermistors, and photoelectric conversion, and the collected analog signals are uploaded to a data platform through analog-to-digital conversion. All collected multi-source data are converged into a central database through communication links such as 5G, LoRa, and industrial Ethernet, and are efficiently stored by device number and time stamp.

[0159] The collected multi-source data mainly include:

[0160] (1) Load history sequence: automatically records the active load of each region and each terminal, for trend analysis;

[0161] (2) Weather and behavior variables: synchronously collects the temperature, humidity, and wind speed of the charging region, and combines non-power factors such as holidays and seasonal types;

[0162] (3) Topological structure information: obtains the device connection relationship of the regional power grid by means of SCADA / GIS systems, to provide support for graph structure modeling.

[0163] For example, at a charging station in a core business district, the system deploys a current monitoring unit at the main feeder line, installs a switch state sensor at the access side breaker, and configures a weather station to collect environmental parameters such as wind speed and temperature. All sensors send data to the edge node in real time through the 5G communication module, and the edge node synchronizes the structured information to the platform. The GIS module provides physical connection information of the station and adjacent sub-stations, which is used for subsequent construction of graph edge structure.

[0164] As a further preferred technical solution, the data processing layer needs to go through multiple rounds of processing on the collected multi-source data to ensure model availability and stability. Mainly includes:

[0165] (1) Abnormality rejection and missing mark: The system has a built-in data quality detection mechanism, which uses historical distribution + short-term fluctuation strategy to identify abnormalities. For example, if the voltage value of a terminal is continuously 0V for 5 minutes, the system will judge it as a failure state and perform mask processing. All missing or invalid data are automatically shielded before entering the model, ensuring that training is only based on valid observations.

[0166] (2) Normalization and feature standardization: In order to eliminate the dimensional differences between different sensing features, the system compresses each type of feature value to a unified interval. Environmental variables such as temperature and humidity are standardized to [0, 1], while voltage, current, etc. are compressed to the [0, 10] interval, ensuring that the model has equal learning weights for each dimension during training.

[0167] (3) Graph structure and knowledge graph generation: The system maps normalized device and environmental data to the power topology, building a unified power knowledge network. The graph nodes cover regions, terminals, substations, meteorological factors, etc., and the edges represent power supply, ownership, adjacency, etc. Each node also records attribute values such as device type, installation location, real-time state, etc.

[0168] Specifically, after completing multi-source data cleaning and topology mapping, the system abstracts charging devices, transformers, weather nodes, and photovoltaic sites as nodes in the graph, and defines the types of edges based on power supply links, weather influence ranges, and device ownership relationships, to build a multi-relation power knowledge graph. The historical time series data of each node (such as photovoltaic output, voltage fluctuation, and load change) are input as features, combined with the node's position in the graph and its adjacency structure, into the graph neural network for spatial dependence modeling.

[0169] As a further preferred technical solution, the modeling and reasoning layer uses a structure-enhanced FEDformer model to achieve high-precision time series prediction of node states, and integrates graph embedding results to improve prediction accuracy and interpretability, specifically including:

[0170] Graph embedding generation: The system first extracts graph structure features through the graph convolution mechanism R-GCN. For each device node, the system aggregates its upstream and downstream neighbor states and relationship types to generate an embedding vector. For example, the graph embedding of CB-23 will integrate the voltage of B1 bus, the power state of U-9 terminal, and the environmental influence of the weather node, forming a structured feature representation.

[0171] Temporal data windowing and Token construction: The system divides the historical running data of the node (such as the past 12 hours) into multiple fixed-length temporal segments and splices them with the corresponding graph embedding to form the input Token of the FEDformer model. For example: for a charging pile node in a certain area, the system extracts its past 48 15-minute granularity load data (i.e. 12 hours), splices its graph structure features to form the model input Token.

[0172] Model training and prediction output: The FEDformer model learns through the causal attention mechanism and the frequency domain decomposition module, fuses the structural embedding representation extracted by the graph neural network (such as regional topological dependence, photovoltaic-load coupling strength), and jointly models with the normalized multi-dimensional time series data (such as historical load, price, meteorological factors, photovoltaic output), and finally outputs the prediction value at future time points. For example, the system can continuously predict the load sequence in the next 4 hours, with a granularity of 15 minutes, a total of 16 prediction points.

[0173] FEDformer identifies key cycles and trend signals through frequency domain attention mechanism, and can model both long-term trends (such as load differences between weekdays and holidays) and short-term disturbances (such as sudden changes in photovoltaic output caused by sudden weather). In addition, the model introduces a position encoding mechanism to enhance the perception of time sequence and prevent prediction bias caused by incorrect input time slices. The prediction results output by the model can not only be used for short-term load scheduling, but also can be superimposed into the graph to support time series visualization, trend evolution tracking and structural risk assessment under the graph structure.

[0174] As a further preferred technical solution, the system further includes an application service layer, which specifically implements the following functions:

[0175] (1) Point prediction and rigid scheduling: The system uses point prediction results as the basis for rigid scheduling, for example, the model predicts that the load of a certain area at 14:00 is 122.4kW, and the system can arrange transformer load switching, current limiting instructions and other operations accordingly.

[0176] (2) Probabilistic prediction and confidence interval output: The model supports multiple quantile point predictions, such as 0.1, 0.5, and 0.9, to generate upper and lower boundary intervals and reflect the fluctuation range of the prediction. For example, the system predicts that the load interval in the next 30 minutes is [115.0kW, 127.5kW], and the real-time load is 130.2kW, indicating that the actual value exceeds the prediction confidence interval, and the system will automatically trigger an abnormal warning and record the node as "potential load overload".

[0177] It is assumed that the system is deployed in a charging station group in Dongcheng District: the system samples the load and current data every 15 minutes, and the edge gateway synchronously reports; the FEDformer model predicts the future 4-hour trend based on the past 12-hour load sequence and topology; the model prediction shows that the load continues to rise from 23:00, exceeding the upper limit of 0.9 quantile, and the system issues a risk prompt; the platform pushes the alarm information to the operation and maintenance center, and the on-duty personnel dispatch engineers for on-site review; if it is confirmed that there is an overcurrent trend on site, the platform automatically adjusts the regional power supply scheme to reduce the load of the main line.

[0178] (3) Real-time update and atlas evolution: the system supports dynamic writing of prediction results into the graph database. For example, the model determines that CB-23 has abnormal fluctuations within 30 minutes in the future, and the system synchronously updates this state to the graph node attribute, while recording its abnormal score and diagnosis time in the prediction process, realizing the cycle of “prediction-atlas evolution-re-prediction”.

[0179] It should be noted that, compared with existing prediction technologies that only rely on single-region time series modeling or static feature splicing, the present embodiment has achieved a systematic breakthrough in the following aspects:

[0180] (1) Stronger semantic modeling capability: the power knowledge graph models the structural and semantic relationships between entities such as regions, devices, users, weather, and photovoltaic output, enabling explicit expression and graph-based reasoning of load influencing factors, and improving the understanding and utilization of key context information by the model; (2) Better multi-region collaborative prediction capability: graph neural networks enhance the model's ability to model the topology of the power grid and the dependence between regions, making it suitable for multi-point concurrent prediction tasks in complex urban scenarios, effectively addressing the problem of data silos and unmodeled mutual influence between regions; (3) Higher efficiency of long sequence modeling: the FEDformer model based on frequency domain modeling significantly reduces the reasoning complexity, and compared with traditional LSTM and Transformer, it performs more stably in modeling long sequences and multi-periodic trends, making it suitable for large-scale deployment; (4) More comprehensive and controllable prediction results: point prediction results can support precise control strategies for dispatching systems, while probability prediction provides interval confidence judgment, which helps to improve the fault tolerance of peak value misjudgment caused by photovoltaic disturbance in power grid operation, providing double protection for safe and stable operation, especially improving the fault tolerance of peak value misjudgment caused by photovoltaic disturbance; (5) Enhanced closed-loop feedback capability for model evolution: through the feedback and evolution mechanism driven by the graph, the system can continuously absorb prediction errors and abnormal event information, improving its adaptability to sudden scenarios and non-stationary patterns, and realizing the “self-evolution” of the intelligent system.

[0181] In summary, the system improves the load prediction accuracy, intelligent scheduling efficiency and abnormal response capability of the city-level power grid as a whole, and is particularly suitable for complex operation scenarios with large-scale access of distributed photovoltaic and diversified power supply structure

[0182] It should be noted that other embodiments or specific implementation methods of the city power load time sequence prediction system described in the present application can refer to the above-mentioned method embodiments, which will not be described here.

[0183] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a list of executable instructions for implementing logical functions, which can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, device or apparatus, such as a computer-based system, a system including a processor or other system that can fetch instructions from an instruction execution system, device or apparatus and execute them. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by or in connection with an instruction execution system, device or apparatus, or in conjunction with these instruction execution systems, devices or apparatus. More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connections having one or more wires (electronic devices), portable computer diskettes (magnetic devices), random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memories), fiber optic devices, and portable compact disc read-only memories (CDROMs). In addition, computer-readable media can even be paper or other suitable media on which the program can be printed, as the program can be electronically obtained, for example, by optical scanning of the paper or other media, followed by editing, interpretation or processing, if necessary, in other suitable ways, and then stored in a computer memory.

[0184] It should be understood that parts of the present application can be implemented in hardware, software, firmware or a combination thereof. In the above-described embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following technologies known in the art or their combinations can be used: discrete logic circuit with logic gate circuit for implementing logical functions on data signals, application specific integrated circuit with suitable combination of logic gate circuit, programmable gate array (PGA), field programmable gate array (FPGA) and the like.

[0185] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the description of the specification, the illustrative description of the above terms does not necessarily mean the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0186] In addition, the terms "first", "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0187] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and cannot be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. A method for time-series forecasting of urban electricity load, characterized in that, include: Acquire multidimensional time-series data related to power load and construct a power knowledge graph based on the multidimensional time-series data; By using a relational graph convolutional neural network to learn structured embedding representations for nodes in an electric power knowledge graph, embedding vectors are obtained to represent structural dependencies between regions. The embedded vector is concatenated with multiple time segments obtained from the multidimensional time series data to obtain multiple structure-enhanced sequence segments. Multiple structure-enhanced sequence fragments are input into the FEDformer model for modeling processing to obtain time-series load predictions.

2. The urban power load time-series forecasting method as described in claim 1, characterized in that, The acquired multidimensional time-series data related to power load includes regional power load sequences, distributed photovoltaic operation data, environmental meteorological variables, and holiday and behavioral tags from the previous time period.

3. The urban power load time-series forecasting method as described in claim 1, characterized in that, The construction of the power knowledge graph based on multidimensional time-series data includes: The multidimensional time series data is subjected to structured processing to obtain standardized multidimensional time series data; A power knowledge graph is constructed by mapping standardized multidimensional time-series data with power grid topology.

4. The urban power load time-series forecasting method as described in claim 1, characterized in that, The power knowledge graph is represented as G = (V, E, R, N) r V is a set of nodes, where nodes represent different types of power entities and each node v has an attribute vector. E is the set of edges, where edges represent connections between nodes; R is the set of edge types, where edge types represent connection relationships of different relation types; N r Represents the set of neighboring nodes under the relation r∈R; The types of map nodes include regional nodes, equipment nodes, user nodes, meteorological factor nodes, and photovoltaic system nodes; The semantic associations include power supply relationships, equipment ownership, meteorological impacts, photovoltaic access, and power output impacts.

5. The urban power load time-series forecasting method as described in claim 1, characterized in that, A graph database is used as the storage and access platform for the power knowledge graph.

6. The urban power load time-series forecasting method as described in claim 1, characterized in that, The method of using a relational graph convolutional neural network to learn structured embedding representations for nodes in the power knowledge graph, resulting in embedding vectors representing structural dependencies between regions, includes: Using a relational graph convolutional neural network, for each node in the power knowledge graph, the states and relationship types of its upstream and downstream neighbor nodes are aggregated to generate embedding vectors representing structural dependencies between regions: In the formula: Represents the embedding vector of node v at level l+1; R is the set of edge types, r∈R represents the semantic relationship between nodes; N r (v) represents the set of neighboring nodes connected to node v under relation r; This represents the transformation matrix for relation r in the l-th layer; c represents the hidden representation vector of node u at level l, where node u is a neighbor node of node v under relation r; v,r σ represents the number of neighboring nodes of node v under relation r; σ represents the activation function. This represents the self-loop weight matrix.

7. The urban power load time-series forecasting method as described in claim 1, characterized in that, The process involves concatenating the embedded vector with multiple time-series segments obtained from multi-dimensional time-series data to obtain multiple structure-enhanced sequence segments, including: The standardized multidimensional time series data is divided into N time series segments according to a window of length p, where the j-th time series segment is: In the formula: Y represents the j-th time segment. p(j-1)+1:pj This represents the partitioning of multidimensional time series data from p(j-1)+1 to pj, where L is the total length of the standardized multidimensional time series data. Concatenating the embedding vector with the j-th time segment yields the corresponding structure-enhanced sequence segment: In the formula: t j Let j represent the structure-enhanced sequence segment, and let InputRSB represent the residual block mapping function. This represents the mask corresponding to the j-th time segment, and ⊙ indicates that missing values ​​are set to 0. Represents the graph embedding vector, PE j This indicates location encoding information.

8. The urban power load time-series forecasting method as described in claim 7, characterized in that, The location encoding information is obtained using sine and cosine encoding, expressed by the following formula: Where: PE (pos,2i) This represents the sinusoidal positional encoding value of position pos in the even-numbered dimension of the encoding vector, where pos represents the position index of the current element, and d model PE represents the dimension of the model vector. (pos,2i+1) This represents the sinusoidal position code value of position pos corresponding to the odd-numbered dimension in the vector, where i represents half of the current position index.

9. The urban power load time-series forecasting method as described in claim 1, characterized in that, The process of inputting multiple structure-enhanced sequence fragments into the FEDformer model for modeling to obtain time-series load prediction values ​​includes: Each structure-enhanced sequence fragment is input into the FEDformer model for modeling processing, yielding the time-series load output features, expressed by the formula: In the formula: o j Let represent the load forecast characteristics output by the FEDformer model in the j-th time window. This indicates that the j-th time segment is valid; otherwise, the j-th time segment is invalid. t1,...,t j Let j represent the j-th time window, and FEDformer represent a deep time series prediction model based on frequency domain modeling mechanism; The time-series load output characteristics are subjected to residual mapping to obtain the time-series load prediction value.

10. The urban power load time-series forecasting method according to any one of claims 1 to 9, characterized in that, Both the relational graph convolutional neural network and the FEDformer model are pre-trained end-to-end, and the loss function used for training is: L total =αL MSE +(1-α)L q In the formula: L total Let L represent the total loss function. MSE L represents the point prediction loss function. q Let α represent the probabilistic prediction loss function, α∈[0,1] be used to control the weights of the point prediction loss function and the probabilistic prediction loss function, and y represent the weights of the point prediction loss function and the probabilistic prediction loss function. j This represents the actual future load sequence value in the j-th forecast window. Let N represent the predicted future load sequence value in the j-th forecast window, and let N represent the number of time segments. + =max(0,·), where q∈(0,1) represents the predicted quantile value. This represents the prediction result of the j-th window at quantile q.

11. The urban power load time-series forecasting method according to any one of claims 1 to 9, characterized in that, After inputting multiple structure-enhanced sequence fragments into the FEDformer model for modeling processing to obtain time-series load prediction values, the process further includes: The residuals and errors between the predicted and actual load values ​​are used as new attributes to update the relevant nodes in the power knowledge graph.

12. A time-series forecasting system for urban power load, characterized in that, include: The data acquisition layer is used to acquire multi-dimensional time-series data related to power load. The data processing layer is used to construct an electricity knowledge graph based on multi-dimensional time-series data; The modeling and reasoning layer is used to learn the structured embedding representation of the nodes in the power knowledge graph using a relational graph convolutional neural network, thereby obtaining embedding vectors that represent the structural dependencies between regions. The embedded vector is concatenated with multiple time series segments obtained from the multidimensional time series data to obtain multiple structure-enhanced sequence segments, which are then input into the FEDformer model for modeling processing to obtain the time series load prediction value.

Images