A multi-level power load prediction method, device and electronic equipment

Abstract

The disclosure provides a multi-level power load prediction method and device and electronic equipment, and relates to the technical field of power systems. The method comprises the following steps: constructing a multi-level power grid space-time data model according to the power grid topological structure of each level, the load data of each node of the power grid, and the physical parameters of each node of the power grid; determining the power load value of each level of the power grid according to the fusion space-time characteristics obtained by fusing the first space-time characteristics of the multi-level power grid space-time data model and the second space-time characteristics of the power grid space-time data model of a preset level.

Description

Technical Field

[0001] This disclosure relates to the field of power system technology, and more specifically, to a multi-level power load forecasting method, apparatus, and electronic equipment. Background Technology

[0002] In the relevant technologies of power load forecasting, when a complex two-layer optimization and multi-agent iterative architecture is adopted, continuous and large-scale exchange of scheduling instructions and feedback information is required between the upper and lower layers, and complex heuristic algorithms are relied upon to find iterative solutions. Therefore, it cannot meet the requirements of fast response and real-time performance for actual power load forecasting and scheduling, resulting in the defects of power load forecasting being complex to solve and having insufficient real-time computation. Summary of the Invention

[0003] In view of this, the present disclosure provides a multi-level power load forecasting method, apparatus and electronic equipment.

[0004] One aspect of this disclosure provides a multi-level power load forecasting method, comprising: constructing a multi-level power grid spatiotemporal data model based on the power grid topology of each level, load data of each node in the power grid, and physical parameters of each node in the power grid; and determining the power load value of each level of the power grid based on the fused spatiotemporal features obtained by fusing the first spatiotemporal features of the multi-level power grid spatiotemporal data model with the second spatiotemporal features of the preset-level power grid spatiotemporal data model.

[0005] According to embodiments of this disclosure, the third spatiotemporal features of the power grid spatiotemporal data models at each level are aggregated to obtain the first spatiotemporal features of the multi-level power grid spatiotemporal data models.

[0006] According to embodiments of this disclosure, the process of determining the third spatiotemporal feature of the power grid spatiotemporal data model at each level includes: extracting spatial and temporal features of the power grid spatiotemporal data model at each level; enhancing key features in the aforementioned spatial or temporal features to obtain enhanced spatial or temporal features; fusing the aforementioned spatial features, temporal features, and enhanced features to obtain the third spatiotemporal feature of the power grid spatiotemporal data model at each level; wherein the enhanced features are the enhanced spatial features or the enhanced temporal features.

[0007] According to embodiments of this disclosure, graph convolutional networks are applied to extract spatial features of the spatiotemporal data model of the power grid at each level.

[0008] According to embodiments of this disclosure, a multi-head temporal self-attention mechanism is used to extract the temporal features of the power grid spatiotemporal data model at each level.

[0009] According to embodiments of this disclosure, a compression and excitation module is applied to enhance key features in the aforementioned spatial features or temporal features to obtain enhanced spatial features or enhanced temporal features.

[0010] According to embodiments of this disclosure, the third spatiotemporal features of each level of power grid spatiotemporal data model are aggregated to obtain the first spatiotemporal features of the multi-level power grid spatiotemporal data model. This includes: using the aforementioned second spatiotemporal features as a query, and using the fusion results of the third spatiotemporal features of other levels of power grid spatiotemporal data models as keys and values ​​respectively, determining the correlation between each node in the aforementioned preset-level power grid spatiotemporal data model and all nodes in the aforementioned other-level power grid spatiotemporal data models; wherein, the aforementioned other-level power grid spatiotemporal data models are power grid spatiotemporal data models at levels other than the aforementioned preset level; determining a weight matrix based on the aforementioned correlation and the membership relationship between power grid nodes; determining the aggregated features of each node in the aforementioned preset-level power grid spatiotemporal data model based on the aforementioned weight matrix and the aforementioned values; and concatenating the aforementioned aggregated features to obtain the first spatiotemporal features of the multi-level power grid spatiotemporal data model.

[0011] According to embodiments of this disclosure, by performing residual connection and layer normalization operations, the first spatiotemporal features of a multi-level power grid spatiotemporal data model are fused with the second spatiotemporal features of a preset-level power grid spatiotemporal data model to obtain fused spatiotemporal features.

[0012] According to embodiments of this disclosure, based on the aforementioned fused spatiotemporal characteristics, multiple fully connected layers are applied to determine the power load values ​​of the power grid at multiple future time steps.

[0013] According to embodiments of this disclosure, the power load value of the power grid is determined by fusing the first spatiotemporal features of a multi-level power grid spatiotemporal data model with the second spatiotemporal features of a preset-level power grid spatiotemporal data model. This includes: inputting the multi-level power grid spatiotemporal data model into a load prediction model and outputting the power load value of the power grid; wherein the load prediction model includes a graph convolutional network, a multi-head temporal self-attention mechanism, a compression and excitation module, an attention module, residual connections, layer normalization, and a fully connected layer.

[0014] According to an embodiment of this disclosure, the training process of the load forecasting model includes: taking historical data of the multi-level power grid spatiotemporal data model as input and historical power load values ​​of the power grid as output, and applying a gradient descent algorithm to train the load forecasting model based on minimizing the total loss of the load forecasting model; the total loss of the load forecasting model includes prediction accuracy loss and physical constraint penalty loss.

[0015] According to embodiments of this disclosure, when the power grid includes distribution area level, busbar level, and substation level, based on the fact that the total load of a substation is equal to the sum of the loads of all distribution areas belonging to the substation, the aforementioned physical constraint penalty loss includes a first penalty term and a second penalty term; the first penalty term is expressed as: The second penalty item mentioned above is expressed as follows: ;in, This is the first penalty item; The number of substations; Let be the predicted total load of the j-th substation; This represents the number of distribution areas belonging to the j-th substation. The predicted load belongs to the i-th distribution area of ​​the j-th substation; This is the relationship matrix between the j-th substation and the i-th distribution area belonging to the j-th substation; This is the second penalty item; The predicted load for the k-th node; This represents the rated capacity of the k-th node.

[0016] Another aspect of this disclosure provides a multi-level power load forecasting device, the device comprising: a model building module for constructing a multi-level power grid spatiotemporal data model based on the power grid topology of each level, load data of each node in the power grid, and physical parameters of each node in the power grid; and a load value determination module for determining the power load value of each level of the power grid based on the fused spatiotemporal features obtained by fusing the first spatiotemporal features of the multi-level power grid spatiotemporal data model obtained based on the affiliation relationship between power grid nodes and the second spatiotemporal features of the preset level power grid spatiotemporal data model.

[0017] Another aspect of this disclosure provides an electronic device comprising: one or more processors; and a memory for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the methods described above.

[0018] Another aspect of this disclosure provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the methods described above.

[0019] Another aspect of this disclosure provides a computer program product including computer-executable instructions that, when executed, are used to implement the method as described above.

[0020] According to embodiments of this disclosure, by constructing a multi-level spatiotemporal data model of the power grid and integrating spatiotemporal features of different levels, the inherent correlation between the power grid topology, node loads, and physical parameters can be fully explored, thereby more accurately reflecting the spatiotemporal distribution characteristics of the power grid, significantly improving the accuracy and reliability of determining power load values ​​at each level, providing solid data support for the optimized scheduling, stable operation, and scientific planning of the power grid, and solving problems such as the lack of accurate expression of the physical space constraints of the power grid, insufficient global optimality due to the use of rigid local constraint rules, and computational complexity and poor real-time performance in related technical methods. Attached Figure Description

[0021] The above and other objects, features, and advantages of this disclosure will become clearer from the following description of embodiments of the present disclosure with reference to the accompanying drawings, in which:

[0022] Figure 1 An exemplary system architecture for which the multi-level power load forecasting method and apparatus of this disclosure can be applied is illustrated schematically; Figure 2 A flowchart illustrating a multi-level power load forecasting method according to an embodiment of the present disclosure is shown schematically. Figure 3 A flowchart illustrating a multi-level power load forecasting method according to another embodiment of the present disclosure is shown schematically; Figure 4 A flowchart illustrating the training process of a load prediction model according to an embodiment of the present disclosure is shown. Figure 5 A block diagram of a multi-level power load forecasting apparatus according to an embodiment of the present disclosure is shown schematically. Figure 6 A block diagram of an electronic device suitable for implementing a multi-level power load forecasting method according to an embodiment of the present disclosure is shown schematically. Detailed Implementation

[0023] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0024] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0025] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0026] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0027] In the embodiments disclosed herein, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to safeguard user personal information security, network security, and national security.

[0028] In the embodiments disclosed herein, user authorization or consent is obtained before acquiring or collecting user personal information.

[0029] Graph learning and spatiotemporal graph modeling are applied to relational data consisting of nodes and edges. Graph learning methods represent topological dependencies through message passing and aggregation operations. Typical forms include graph convolution, graph sampling and aggregation, and graph attention, which can perform prediction and classification tasks at the node, edge, and overall graph levels. In power scenarios, graphs are often constructed using geographical or electrical topology, with loads as node features, and spatiotemporal graph models are used to jointly characterize adjacency relationships and temporal evolution patterns.

[0030] Hierarchical time series forecasting and unification methods are designed for multi-level series with aggregation / decomposition relationships. Related methods include bottom-up (predicting the bottom level first and then summarizing), top-down (predicting the total first and then decomposing proportionally), and middle-outward strategies. In addition, there are unification methods that use a linear algebra framework to post-process the multi-level forecast vectors, such as optimal unification based on covariance estimation and generalized least squares, so that the results of each level algebraically satisfy the aggregation constraints.

[0031] In related technologies, based on the operating parameters of shallow and deep geothermal well groups, a well group output model is constructed to characterize the unit thermal output capacity, thermal storage capacity, and temperature decay characteristics of each well group. The hierarchical differences in geothermal resources are introduced during the formation of scheduling parameters, improving the matching accuracy of scheduling strategies for various types of geothermal resources. This method fully utilizes the advantages of multi-level time scales. By constructing a multi-scale load prediction model, based on historical load data and historical meteorological data, it generates prediction results for day-ahead load, mid-day load, and real-time load, respectively. This achieves comprehensive coverage of medium- and long-term operating trends and short-term load fluctuations by the scheduling strategy, overcoming the problem of delayed response to sudden load changes in traditional single-time-scale scheduling methods. The well group output model, multi-time-scale load prediction results, and a multi-objective optimization function based on formation thermal balance, operating cost, carbon emissions, and efficiency are jointly calculated. The optimal scheduling parameters are then obtained from the candidate scheduling parameter set using an objective genetic algorithm (such as a non-dominated sorting multi-objective genetic algorithm). This mechanism allows the generation of scheduling parameters to go beyond simple load matching, enabling dynamic and complementary adjustment of the output ratio between shallow and deep geothermal well groups. It also comprehensively considers the thermal storage capacity and temperature decay trend of the geothermal well groups, significantly improving the allocation efficiency, operational balance, and energy supply stability of geothermal resources. This effectively enhances the sustainability of the geothermal utilization system under different load scenarios and operating cycles. Furthermore, this method also has an emergency scheduling function, capable of responding to heat source unit failures by invoking backup output configuration strategies and replenishing the shallow geothermal well groups with heat from deep geothermal well groups and / or building heat during non-heating periods, ensuring the thermal stability and long-term operational efficiency of the geothermal system.

[0032] In another related technology, a multi-level bidirectional verification method for predicting the capacity of distributed renewable energy access in distribution networks is proposed. This method aims to solve the problem of insufficient prediction accuracy in existing technologies due to insufficient consideration of node differences and hierarchical coupling. Firstly, this method performs distribution transformer clustering based on a clustering scoring model, combined with a pre-set load prediction model, to achieve accurate prediction of the long-term annual load output of nodes in a multi-level distribution network (distribution transformers, medium-voltage lines, and substations). After obtaining load output information, multiple operational constraints such as power flow, voltage, current, economic efficiency, and environmental factors are added to each level to determine the threshold and actual access capacity of renewable energy. Most importantly, this mechanism performs bidirectional correctness verification of the actual access capacity of each level based on the hierarchical relationship of the distribution network, both from lower to upper levels (capacity superposition verification) and from upper to lower levels (capacity allocation / reduction verification). Based on the verification results, the capacity is adjusted to determine the final access capacity. This fully considers the coupling characteristics between levels, significantly improves the accuracy and rationality of the distributed renewable energy access capacity prediction results, and effectively ensures the safe and stable operation of the distribution network.

[0033] In another related technology, a multi-agent coordinated optimization control strategy for microgrids based on digital twins is proposed. This strategy aims to address the energy management and operation control of microgrid clusters composed of wind power, photovoltaics, diesel generators, fuel cells, energy storage, and loads, as the penetration rate of distributed energy increases. The core of this strategy lies in formulating the optimization problem as a two-layer optimization model. The upper layer, based on the microgrid cluster level, aims to minimize the total operating cost of the microgrid cluster, considering system balance constraints, microgrid output constraints, and tie-line power constraints, and determines the optimal solution through iterative iteration. The lower layer focuses on individual microgrids, establishing operating costs, power losses, pollutant emissions, and other factors while ensuring their own interests. A multi-objective mathematical model minimizing tie-line power fluctuations is employed, comprehensively considering constraints such as power balance within the microgrid, distributed power output, generator ramp-up, and energy storage charging and discharging. This mechanism allows the upper-level optimization results to be sent as scheduling instructions (including switching capacity and output) to the lower-level model. When the lower-level model cannot find a feasible solution, the results are fed back to the upper-level model to adjust the instructions, achieving iterative coordination between the upper and lower levels. Finally, an improved artificial bee colony algorithm (including fuzzy clustering control, Gaussian mutation, chaotic perturbation, etc.) is used to solve the problem, thereby effectively achieving optimized operation of the microgrid group, significantly improving the utilization rate of renewable energy, reducing microgrid reserve capacity and investment costs, and balancing the overall economic benefits of the microgrid group with the local interests of sub-microgrids.

[0034] Related technologies primarily focus on multi-timescale scheduling, lacking an expression of the physical and spatial constraints of the power network. They cannot integrate key power system constraints, such as the grid topology, into the load forecasting model for joint optimization. Although multi-level spatial structures such as distribution areas and substations are introduced, their core bidirectional verification mechanism uses capacity verification rules (such as finding the minimum value or uniform distribution). This often relies on pre-set local allocation rules rather than a globally optimal network physical model. It cannot perform flexible, end-to-end optimal adjustment and allocation based on actual load characteristics, power flow distribution, and network topology impedance, resulting in rigid constraint methods and insufficient global optimality. Furthermore, the use of complex two-layer optimization and multi-agent iterative architectures to balance the economic interests of microgrid clusters often requires continuous and extensive exchange of scheduling commands and feedback information between upper and lower layers. It also relies on complex heuristic algorithms (such as artificial bee colonies) to find iterative solutions, failing to meet the requirements of rapid response and real-time performance in actual power load forecasting and scheduling, resulting in complex solutions and insufficient real-time computation.

[0035] Based on this, embodiments of this disclosure provide a multi-level power load forecasting method, device, and electronic device. By introducing the physical topology and node constraints of the power grid, such as distribution areas and substations, into the load forecasting model, the load forecasting results are deeply coupled and optimized with the actual network space carrying capacity. Furthermore, an end-to-end automatic adjustment-based load forecasting constraint model is designed to achieve global optimal solution and flexible allocation of load forecasting across all levels, replacing the rigid verification method based on local rules (such as finding the minimum value or uniform allocation). In addition, a simplified interaction mechanism is established for lower levels (such as distribution areas) to obtain global information and scheduling instructions from higher levels (such as substations), thereby achieving efficient collaboration between multi-level systems and avoiding complex iterative calculations and large-scale information exchange between multiple agents.

[0036] Figure 1 The illustration schematically depicts an exemplary system architecture to which the multi-level power load forecasting method and apparatus of this disclosure can be applied. It should be noted that... Figure 1 The examples shown are merely examples of system architectures that can be applied to the embodiments of this disclosure, in order to help those skilled in the art understand the technical content of this disclosure, but do not mean that the embodiments of this disclosure cannot be used in other devices, systems, environments or scenarios.

[0037] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0038] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays, including but not limited to smartphones, tablets, laptops, and desktop computers, etc.

[0039] Server 105 can be a server that provides various services, such as a backend management server that supports users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process the received user requests and other data, and then feed the processing results back to the terminal devices.

[0040] It should be noted that the multi-level power load forecasting method provided in this embodiment can generally be executed by server 105. Correspondingly, the multi-level power load forecasting device provided in this embodiment can generally be located in server 105. The multi-level power load forecasting method provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the multi-level power load forecasting device provided in this embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Alternatively, the multi-level power load forecasting method provided in this embodiment can also be executed by the first terminal device 101, the second terminal device 102, or the third terminal device 103, or by other terminal devices different from the first terminal device 101, the second terminal device 102, or the third terminal device 103. Accordingly, the multi-level power load forecasting device provided in this embodiment of the present disclosure may also be installed in the first terminal device 101, the second terminal device 102 or the third terminal device 103, or in other terminal devices different from the first terminal device 101, the second terminal device 102 or the third terminal device 103.

[0041] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0042] Figure 2 A flowchart illustrating a multi-level power load forecasting method according to an embodiment of the present disclosure is shown schematically.

[0043] like Figure 2 As shown, the method includes operations S201~S202.

[0044] In operation S201, a multi-level spatiotemporal data model of the power grid is constructed based on the power grid topology at each level, the load data of each node in the power grid, and the physical parameters of each node in the power grid.

[0045] In operation S202, the power load values ​​of each level of the power grid are determined by fusing the first spatiotemporal features of the multi-level power grid spatiotemporal data model with the second spatiotemporal features of the preset-level power grid spatiotemporal data model.

[0046] According to embodiments of this disclosure, by constructing a multi-level spatiotemporal data model of the power grid and integrating spatiotemporal features of different levels, the inherent correlation between the power grid topology, node loads, and physical parameters can be fully explored, thereby more accurately reflecting the spatiotemporal distribution characteristics of the power grid, significantly improving the accuracy and reliability of determining power load values ​​at each level, providing solid data support for the optimized scheduling, stable operation, and scientific planning of the power grid, and solving problems such as the lack of accurate expression of the physical space constraints of the power grid, insufficient global optimality due to the use of rigid local constraint rules, and computational complexity and poor real-time performance in related technical methods.

[0047] The following is based on Figure 3 As one implementation method of this disclosure, the multi-level power load forecasting method provided in this disclosure is described.

[0048] Figure 3 A flowchart illustrating a multi-level power load forecasting method according to another embodiment of the present disclosure is shown schematically.

[0049] like Figure 3 As shown, the method includes operations S1 to S4.

[0050] Operation S1: Multi-level power grid data construction and preprocessing.

[0051] In the embodiments of this disclosure, multi-source heterogeneous data, including historical load data and multi-level power grid topology, are collected and integrated to construct a unified multi-level spatiotemporal data model. The multi-level power grid topology may include distribution area-level topology and substation-level topology.

[0052] For example, collect historical load data for nodes at various levels of the power grid, such as 15-minute or hourly active / reactive power data for distribution transformers, feeders, and substation buses; obtain the static topology of the power grid and construct a distribution transformer-level topology map. Substation hierarchical topology diagram It also establishes a hierarchical mapping relationship between the two levels; integrates the physical parameters of the nodes, such as line impedance, transformer rated capacity, and node voltage upper and lower limits, as implicit constraints or penalty terms of the model.

[0053] In operation S2: Multi-scale spatiotemporal feature extraction.

[0054] According to embodiments of this disclosure, the process of determining the third spatiotemporal feature of the power grid spatiotemporal data model at each level includes: extracting spatial and temporal features of the power grid spatiotemporal data model at each level; enhancing key features in the spatial or temporal features to obtain enhanced spatial features or enhanced temporal features; and fusing the spatial features, temporal features, and enhanced features to obtain the third spatiotemporal feature of the power grid spatiotemporal data model at each level; wherein the enhanced features are enhanced spatial features or enhanced temporal features.

[0055] In the embodiments of this disclosure, a graph convolutional network is applied to extract the spatial features of the power grid spatiotemporal data model at each level. A multi-head temporal self-attention mechanism is employed to extract the temporal features of the power grid spatiotemporal data model at each level. Compression and excitation modules are applied to enhance key features in the spatial or temporal features to obtain enhanced spatial or temporal features.

[0056] Specifically, deep feature extraction of the spatial topological dependence and temporal series dependence of power grid data is performed by graph convolutional network (GCN) and temporal self-attention network, respectively, to generate spatiotemporal feature representations at each level.

[0057] Feature extraction of power grid data is performed using graph convolutional networks and temporal self-attention networks. For nodes at each level (such as distribution areas and substations), graph convolutional networks are used to learn their spatial topological relationships with neighboring nodes to obtain spatial dependency features. For the historical load sequence of each node, temporal self-attention networks are used to capture its long-term and short-term temporal dependency patterns, and channel attention mechanism (SE Block) is combined to adaptively adjust the importance of different load features (such as active and reactive power) to obtain enhanced temporal features.

[0058] A graph convolutional network is used to aggregate spatial information of node features at each time step, learning the spatial topological relationships between nodes and their neighbors to obtain spatial dependency features. A multi-head temporal self-attention mechanism is employed to capture long-term and short-term temporal dependency patterns of the load sequence. Queries, keys, and values ​​are calculated for the temporal features of each node, and temporal features are obtained through attention weight aggregation. A channel attention mechanism is introduced to adaptively calibrate feature channels, learning the importance weights of different load features through global average pooling and a fully connected network to achieve feature recalibration. Spatial features, temporal features, and calibrated features are fused to obtain an initial multi-scale spatiotemporal feature representation. The calibrated features are either enhanced spatial features or enhanced temporal features. The initial multi-scale spatiotemporal feature representation serves as the third spatiotemporal feature of each level of the power grid spatiotemporal data model.

[0059] In Operation S3: Cross-level attention feature fusion.

[0060] According to embodiments of this disclosure, the third spatiotemporal features of each level of power grid spatiotemporal data model are aggregated to obtain the first spatiotemporal features of the multi-level power grid spatiotemporal data model. Specifically, this includes: using the second spatiotemporal feature as a query, and using the fusion results of the third spatiotemporal features of other levels of power grid spatiotemporal data models as keys and values, respectively, determining the correlation between each node in the preset-level power grid spatiotemporal data model and all nodes in other levels of power grid spatiotemporal data models; wherein, other levels of power grid spatiotemporal data models are power grid spatiotemporal data models at all levels other than the preset level; determining a weight matrix based on the correlation and the membership relationship between power grid nodes; determining the aggregated features of each node in the preset-level power grid spatiotemporal data model based on the weight matrix and values; and concatenating the aggregated features to obtain the first spatiotemporal features of the multi-level power grid spatiotemporal data model.

[0061] In the embodiments of this disclosure, taking the power grid hierarchy, which includes the distribution area level and the substation level, as an example, a cross-level attention mechanism is designed. The distribution area level is referred to as the fine-grained level, and the substation level is referred to as the coarse-grained level. Aggregating the third spatiotemporal features of the power grid spatiotemporal data model at each level involves aggregating the third spatiotemporal features of the fine-grained level with those of the coarse-grained level. This means adaptively transferring and fusing the global feature information of the coarse-grained level into the features of the fine-grained level, capturing the global influence of the coarse-grained level on the fine-grained level, and achieving refined feature enhancement guided by global information.

[0062] Specifically, the fine-level node features are used as queries, and the associated coarse-level node features are used as keys and values. In a power grid hierarchy that includes distribution area levels and substation levels, with the distribution area level referred to as the fine-level level and the substation level as the coarse-level level, the coarse-level nodes associated with the distribution area level nodes are substation nodes. The fine-level node features are used as the second spatiotemporal features, and the coarse-level node features are used as the third spatiotemporal features. By calculating the attention weight between the query and the key, each fine-level node can adaptively aggregate the feature information of its superior coarse-level nodes, thereby obtaining the first spatiotemporal feature of the multi-level power grid spatiotemporal data model. Specifically, a mask matrix is ​​generated using a membership mapping matrix to ensure that each fine-grained node only focuses on its superior coarse-grained node. A multi-head attention mechanism is employed to capture diverse cross-level dependencies from different subspaces. The second spatiotemporal features of the pre-defined level power grid spatiotemporal data model are consistent with the third spatiotemporal features of the pre-defined level power grid spatiotemporal data model.

[0063] According to embodiments of this disclosure, by performing residual connection and layer normalization operations, the first spatiotemporal features of a multi-level power grid spatiotemporal data model are fused with the second spatiotemporal features of a preset-level power grid spatiotemporal data model to obtain fused spatiotemporal features.

[0064] Specifically, by fusing residual connections and layer normalization with the original fine-grained features, updated fine-grained features are obtained, ensuring the stability and efficiency of information flow. In other words, the second spatiotemporal feature of the pre-defined layered power grid spatiotemporal data model is the original fine-grained feature; the fused spatiotemporal feature is the updated fine-grained feature.

[0065] In operation S4: End-to-end load forecasting.

[0066] In the embodiments of this disclosure, a unified end-to-end prediction model is constructed, and the load prediction values ​​of each node are directly generated by using the fused multi-level features. By designing a composite loss function that includes prediction accuracy and physical constraints, the model implicitly learns and satisfies the physical operation constraints of the power grid in an end-to-end manner, thereby achieving joint optimization of prediction accuracy and physical constraints to achieve globally optimal load allocation.

[0067] According to embodiments of this disclosure, multiple fully connected layers are applied based on the fused spatiotemporal characteristics to determine the power load values ​​of the power grid at multiple future time steps.

[0068] Specifically, a multilayer perceptron (MLP) is used as the prediction head to map the final fused features to load predictions for multiple future time steps. A composite loss function is designed, which includes not only a prediction accuracy loss based on the mean squared error (MSE) between the predicted and actual values, but also a regularization penalty term based on physical laws (such as Kirchhoff's laws and power flow equations), called the physical constraint penalty loss. The physical constraint penalty loss includes a penalty term based on Kirchhoff's current law and a capacity constraint penalty term. The Kirchhoff current law penalty term ensures that the total load of the substation equals the sum of the loads of all its subordinate distribution areas. The capacity constraint penalty term is activated when the predicted load exceeds the rated capacity of the transformer or line. The gradient descent algorithm is used to minimize the total loss function to train the model parameters. When the predicted load distribution violates physical constraints (such as causing line or transformer overload), the physical constraint penalty loss increases, thereby guiding the parameters to optimize in a direction that satisfies the physical constraints during model training.

[0069] In embodiments of this disclosure, operation S1 includes operation S1.1 and operation S1.2.

[0070] In operation S1.1: collect historical load, weather and other data to construct feature tensors.

[0071] Data from multiple levels of the power grid covering the research area is collected from the data platform, including distribution area level, feeder level, and substation level.

[0072] Historical load data from various levels of nodes in the power company's data platform is collected, processed, and used to form the input feature tensor. ,in, For batch size, The input time series length is... The total number of nodes. The feature dimension can be composed of features such as active power, reactive power, and weather information.

[0073] In operation S1.2: Obtain the static topology of the power grid, construct the adjacency matrix, and establish membership relationships.

[0074] Based on the geographic information data of the power grid, adjacency matrices are constructed for fine-grained networks (such as distribution area sets) and coarse-grained networks (such as substation sets). and And establish a fine-grained to coarse-grained membership mapping matrix. ,in, This represents the number of nodes in a fine-grained network. This represents the number of nodes in a coarse-grained network.

[0075] The embodiments of this disclosure achieve deep coupling between load forecasting and grid spatial topology constraints, resulting in physically feasible forecasts. Load forecasting in related technologies is often based on time series or regional totals, lacking precise perception of spatial levels such as grid distribution areas, buses, and substations. The embodiments of this disclosure overcome this limitation by directly incorporating spatial location information from each level into the load forecasting model. This allows the load forecasting model to refine the forecasts to spatially feasible load distributions, effectively improving the accuracy and precision of load level predictions for different spatial nodes.

[0076] In embodiments of this disclosure, operation S2 includes: capturing spatial dependencies between nodes through spatial feature extraction (GCN); capturing temporal dependencies of the sequence through temporal feature extraction; and enhancing key feature channels through channel feature calibration (SE Block).

[0077] Specifically, a hybrid neural network architecture is adopted to extract features in both time and space dimensions simultaneously.

[0078] Specifically, spatial feature extraction involves using a graph convolutional network (GCN) to aggregate spatial information of node features at each time step.

[0079] For the Layer GCN, its update rules are as follows: ; ; in, It is an adjacency matrix with self-loops added. It is an identity matrix. yes The degree matrix, It is the first The node feature matrix of the layer, This is the trainable weight matrix for this layer. It is a non-linear activation function, such as ReLU. Through stacking... Layered GCNs can obtain spatial features that capture high-order neighborhood information. ,in, For spatial feature dimensions.

[0080] Temporal feature extraction specifically involves employing a multi-head temporal self-attention mechanism to capture the dynamic dependencies of the load sequence. For the temporal features of each node... In its first The query, key, and value for each attention head are calculated as follows: ; ; ; in, It is the first Projection matrix of the head.

[0081] Attention output is: .in, This is the scaling factor.

[0082] The final time feature is obtained by concatenating the outputs of all heads and performing a linear transformation. .

[0083] ; ; in, It is the output projection matrix.

[0084] Channel feature calibration specifically involves introducing a Squeeze-and-Excitation (SE) module to adaptively enhance key feature channels.

[0085] First, a Squeeze operation is performed, which uses global average pooling to... or Compression in the spatial or temporal dimensions yields the channel descriptor. : .in, Let c be the channel descriptor for the c-th channel; This represents the original feature of the c-th channel; This represents the Squeeze function; This represents the original feature of the i-th spatial location or time step in the c-th channel.

[0086] Then, an excitation operation is performed, which uses a two-layer fully connected (FC) network to learn the non-linear relationship between channels and obtain the weights of each channel. : .in, W represents the excitation function; W represents the weights of the FC layer. These are the weights of the first fully connected layer; These are the weights of the second fully connected layer. It is the ReLU activation function. It is the Sigmoid activation function.

[0087] Finally, the learned weights With original features Multiplication enables feature recalibration: .in, This represents the Scale function, which is channel-wise multiplication. Let be the weight of the c-th channel; This refers to the characteristics of the c-th channel after recalibration.

[0088] The final characteristics obtained: Fusion can be a concatenation or addition operation; Features after recalibration; This represents the third spatiotemporal feature of the spatiotemporal data model of the power grid at each level.

[0089] In the embodiments of this disclosure, operation S3 includes: performing cross-level attention calculation using fine-grained features as Query values ​​and coarse-grained features as Key and Value values ​​respectively; and performing feature fusion and updating through residual connection and layer normalization operations as well as operations to enhance fine-grained features.

[0090] Among them, a cross-level attention module was designed to enable the effective utilization of coarse-level information by the fine-grained level.

[0091] Assumption and These are the fine-grained and coarse-grained features obtained after processing by operation S2.

[0092] Will As a query As keys and values. For each attention head The calculation is as follows: ; ; ; in, It is the first Projection matrix of the head.

[0093] Using the membership matrix Calculate the attention score matrix .

[0094] ; ; in, For the first The similarity between the i-th fine-grained node and the j-th coarse-grained node of a given node. For the first Similarity in size It is by The generated mask matrix is ​​used to ensure that each fine-grained node only focuses on its parent coarse-grained node. This indicates element-wise multiplication.

[0095] The aggregated features are: .

[0096] Feature fusion and updating specifically involves concatenating the outputs of all heads and fusing them with the original fine-grained features through residual connections and layer normalization to obtain the updated fine-grained features. .

[0097] ; ; in, It is the output projection matrix. It is a learnable fusion coefficient.

[0098] The process of feature fusion and updating can be stacked in multiple layers to achieve deeper cross-level information interaction.

[0099] This disclosure addresses the problems of low solution efficiency and poor real-time performance in the prediction process of complex multi-level power grid information interaction. It proposes a hierarchical collaborative prediction mechanism based on efficient information flow, abandoning complex, multi-round bidirectional iterative calculations and establishing a simplified and efficient unidirectional or local collaborative information flow. By allowing lower levels (such as distribution areas) to directly or simply obtain key global load information or trends from higher levels (such as substations) as their own prediction input, it effectively avoids a large amount of computation and information exchange, significantly improving the prediction collaboration efficiency between multi-level systems and the real-time output capability of the model.

[0100] In embodiments of this disclosure, operation S4 includes: mapping fused features to future loads via an MLP decoder, ultimately outputting: high-precision, multi-level load predictions that meet physical constraints.

[0101] Specifically, a prediction head (MLP decoder) consisting of multiple fully connected layers is used to process the final features. Mapping to the future Predicted values ​​at each time step : .

[0102] According to embodiments of this disclosure, the power load value of the power grid is determined by fusing the first spatiotemporal features of a multi-level power grid spatiotemporal data model with the second spatiotemporal features of a preset-level power grid spatiotemporal data model. This includes: inputting the multi-level power grid spatiotemporal data model into a load prediction model and outputting the power load value of the power grid; wherein the load prediction model includes a graph convolutional network, a multi-head temporal self-attention mechanism, a compression and excitation module, an attention module, residual connections, layer normalization, and a fully connected layer.

[0103] To address the shortcomings of existing technologies in terms of accuracy and flexibility when handling complex multi-level coupling relationships, embodiments of this disclosure construct a high-precision load forecasting model based on an end-to-end neural network. This model takes multi-level power grid data, physical constraints, and historical loads as overall input, and leverages the powerful nonlinear fitting and feature extraction capabilities of neural networks to automatically learn and reflect the complex coupling relationships between each level, directly outputting refined load forecasting results for each level. This end-to-end automated learning mechanism replaces computational methods that rely on local rules or linear models, thereby significantly improving the accuracy and flexibility of load change forecasting in complex power grid environments.

[0104] Figure 4 A flowchart illustrating the training process of a load prediction model according to an embodiment of the present disclosure is shown.

[0105] like Figure 4As shown, the training process of the load forecasting model includes: taking historical data from a multi-level spatiotemporal data model of the power grid as input and historical power load values ​​of the power grid as output, and applying the gradient descent algorithm to train the load forecasting model based on minimizing the total loss of the load forecasting model; the total loss of the load forecasting model includes prediction accuracy loss and physical constraint penalty loss.

[0106] Specifically, the training process for the load prediction model, after training begins, includes the following operations: Step 1: Input training data.

[0107] Step 2: Perform operations S2, S3 and S4 sequentially on the training data to output the historical power load values ​​of the power grid.

[0108] Step 3: Calculate the total loss of the load forecasting model.

[0109] Step 4: Based on the total loss of the load forecasting model, calculate the gradient of the parameters of the load forecasting model through backpropagation.

[0110] Step 5: Based on the gradient obtained in Step 4, update the parameters of the load forecasting model and use the optimizer to update the weights of the load forecasting model.

[0111] Step 6: Determine if the updated load forecasting model has met the stopping condition; if the updated load forecasting model has not met the stopping condition, return to step 2; if the updated load forecasting model has met the stopping condition, end the training and save the trained load forecasting model.

[0112] In embodiments of this disclosure, the total loss of the load forecasting model Due to prediction accuracy loss and physical constraint penalty loss Weighted composition.

[0113] ;in, It is a hyperparameter that balances the two.

[0114] Prediction accuracy loss is typically expressed as mean squared error (MSE) or mean absolute error (MAE).

[0115] Where B is the batch size, i.e. the number of independent samples processed in one forward propagation; The output time step is the length of the time series predicted by the model. The number of zones represents the number of node features in each sample, which is the number of nodes. For the historical power load value of the nth node at the t-th time step of the b-th sample, Let n be the predicted power load value of the nth node at the tth time step for the bth sample.

[0116] In the case of power grid levels including distribution area level, bus level and substation level, the total load of a substation is equal to the sum of the loads of all distribution areas belonging to the substation, and the physical constraint penalty loss includes the first penalty term and the second penalty term.

[0117] Physical constraint penalty loss This is used to penalize predictions that violate grid operation constraints. In a grid structure including distribution area, busbar, and substation levels, based on Kirchhoff's Current Law (KCL), the total load of a substation equals the sum of the loads of all distribution areas belonging to that substation. The total load of a substation is expressed as: ;in, Let J be the total load of the j-th substation; For the load belonging to the i-th distribution area of ​​the j-th substation; This represents the set of distribution areas belonging to the j-th substation.

[0118] Taking the penalty term based on Kirchhoff's current law as the first penalty term, the first penalty term is expressed as: ; in, This is the first penalty item; The number of substations; Let be the predicted total load of the j-th substation; This represents the number of distribution areas belonging to the j-th substation. The predicted load belongs to the i-th distribution area of ​​the j-th substation; Let be the relationship matrix between the j-th substation and the i-th distribution area belonging to the j-th substation.

[0119] To address capacity constraints, a capacity penalty term is designed to apply when the predicted load exceeds the rated capacity of the transformer or line. At this time, the capacity penalty term is activated. If the capacity penalty term is used as the second penalty term, then the second penalty term is expressed as: ;in, This is the second penalty item; The predicted load for the k-th node; This represents the rated capacity of the k-th node.

[0120] Minimize the total loss using the gradient descent algorithm. To train the model parameters, the gradient descent algorithm can use the Adam optimizer.

[0121] After training, real-time power grid data from the multi-level spatiotemporal power grid data model is input into the model to generate high-precision multi-level load forecasting results that meet physical constraints. Through this end-to-end approach, the model can automatically learn how to perform optimal load allocation among different levels, avoiding suboptimal solutions based on rigid rules found in traditional methods.

[0122] The embodiments disclosed herein achieve deep coupling between load forecasting and power grid spatial topology constraints, ensuring the physical feasibility of the forecast results. For the first time, Graph Convolutional Networks (GCNs) are applied to modeling multi-level power grid topologies, directly encoding the physical connections of distribution areas, substations, and other levels into the model. Through a designed composite loss function, physical constraints such as Kirchhoff's laws and equipment capacity are used as intrinsic objectives for model optimization, ensuring that the final load forecast results are not only numerically accurate but also spatially consistent with the physical operating rules of the power grid. This addresses the fundamental deficiency of existing technologies that focus only on time-series forecasting while neglecting spatial physical constraints.

[0123] The embodiments disclosed herein abandon the rigid approach of relying on hard local rules such as "capacity superposition" and "uniform distribution" for hierarchical verification in existing technologies, and construct an end-to-end global optimal solution framework, resulting in more flexible and accurate load allocation. By constructing an end-to-end neural network model from multi-level spatiotemporal data input to refined load forecast output, its powerful nonlinear fitting capability is utilized to automatically learn and adjust the load allocation relationship between each level. This data-driven global optimization method can flexibly optimize based on the actual power flow distribution and network impedance characteristics, significantly improving the accuracy and global optimality of load allocation across the entire power grid hierarchy.

[0124] The embodiments disclosed herein employ an efficient cross-level information fusion mechanism that balances high accuracy and real-time performance. Addressing the computational complexity and slow response of multi-agent iterative or bi-layer optimization models in related technologies, a novel cross-level attention mechanism is proposed. This mechanism, through an efficient, feedforward-like information flow, allows lower-level (transformer area) nodes to adaptively "pay attention to" and aggregate global feature information from the upper level (substation), achieving precise guidance of local predictions based on global trends. This approach replaces time-consuming bidirectional information exchange and iterative solutions, significantly reducing the computational complexity of the model and dramatically improving prediction speed, enabling it to meet the requirements of real-time dispatch and rapid response in power systems.

[0125] Figure 5 A block diagram of a multi-level power load forecasting apparatus according to an embodiment of the present disclosure is shown schematically.

[0126] like Figure 5 As shown, the multi-level power load forecasting device 500 includes a model building module 510 and a load value determination module 520.

[0127] The model building module 510 is used to build a multi-level spatiotemporal data model of the power grid based on the power grid topology at each level, the load data of each node in the power grid, and the physical parameters of each node in the power grid.

[0128] The load value determination module 520 is used to determine the power load value of each level of the power grid based on the fused spatiotemporal features obtained by fusing the first spatiotemporal features of the multi-level power grid spatiotemporal data model obtained based on the affiliation relationship between power grid nodes and the second spatiotemporal features of the preset-level power grid spatiotemporal data model.

[0129] Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure, or at least part of the functions of any one or more of them, can be implemented in one module. Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be implemented by dividing them into multiple modules. Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be at least partially implemented as hardware circuitry, such as a Field-Programmable Gate Array (FPGA), a Programmable Logic Array (PLA), a System-on-Chip, a System-on-a-Substrate, a System-on-Package, an Application-Specific Integrated Circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be at least partially implemented as computer program modules, which, when run, can perform corresponding functions.

[0130] For example, any plurality of the model building module 510 and the load value determination module 520 can be combined into one module / unit / subunit, or any one of the modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least part of the functionality of one or more of these modules / units / subunits can be combined with at least part of the functionality of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of this disclosure, at least one of the model building module 510 and the load value determination module 520 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the model building module 510 and the load value determination module 520 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0131] It should be noted that the data processing system part in the embodiments of this disclosure corresponds to the data processing method part in the embodiments of this disclosure. The specific description of the data processing system part is referred to in the data processing method part, and will not be repeated here.

[0132] Figure 6 A block diagram of an electronic device suitable for implementing a multi-level power load forecasting method according to an embodiment of the present disclosure is shown schematically. Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0133] like Figure 6 As shown, an electronic device 600 according to an embodiment of this disclosure includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage portion 608 into a random access memory (RAM) 603. The processor 601 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this disclosure.

[0134] RAM 603 stores various programs and data required for the operation of electronic device 600. Processor 601, ROM 602, and RAM 603 are interconnected via bus 604. Processor 601 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 602 and / or RAM 603. It should be noted that programs may also be stored in one or more memories other than ROM 602 and RAM 603. Processor 601 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in one or more memories.

[0135] According to embodiments of this disclosure, the electronic device 600 may further include an input / output (I / O) interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output (I / O) interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output (I / O) interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0136] According to embodiments of this disclosure, the method flow according to embodiments of this disclosure can be implemented as a computer software program. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via communication section 609, and / or installed from removable medium 611. When the computer program is executed by processor 601, it performs the functions defined in the system of embodiments of this disclosure. According to embodiments of this disclosure, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0137] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0138] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0139] For example, according to embodiments of this disclosure, a computer-readable storage medium may include the ROM 602 and / or RAM 603 described above and / or one or more memories other than ROM 602 and RAM 603.

[0140] Embodiments of this disclosure also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of this disclosure. When the computer program product is run on an electronic device, the program code enables the electronic device to implement the multi-level power load forecasting method provided in the embodiments of this disclosure.

[0141] When the computer program is executed by the processor 601, it performs the functions defined in the system / apparatus of this disclosure embodiments. According to embodiments of this disclosure, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0142] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0143] According to embodiments of this disclosure, program code for executing the computer programs provided in embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can execute entirely on a user's computing device, partially on a user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device 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 computing device (e.g., via the Internet using an Internet service provider).

[0144] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of the present disclosure can be combined and / or combined in various ways, even if such combinations are not explicitly described in the present disclosure. In particular, the features described in the various embodiments of this disclosure may be combined and / or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0145] The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A multi-level power load forecasting method, characterized in that, The method includes: Based on the power grid topology at each level, the load data of each node in the power grid, and the physical parameters of each node in the power grid, a multi-level spatiotemporal data model of the power grid is constructed. The power load values ​​at each level of the power grid are determined by fusing the first spatiotemporal features of the multi-level power grid spatiotemporal data model with the second spatiotemporal features of the preset-level power grid spatiotemporal data model.

2. The multi-level power load forecasting method according to claim 1, characterized in that, The third spatiotemporal features of the power grid spatiotemporal data models at each level are aggregated to obtain the first spatiotemporal features of the multi-level power grid spatiotemporal data models.

3. The multi-level power load forecasting method according to claim 2, characterized in that, The process of determining the third spatiotemporal feature of the spatiotemporal data model of the power grid at each level includes: Extract the spatial and temporal features of the spatiotemporal data models of the power grid at each level; The key features in the spatial features or the temporal features are enhanced to obtain enhanced spatial features or enhanced temporal features; The spatial features, the temporal features, and the enhanced features are fused to obtain the third spatiotemporal features of the power grid spatiotemporal data model at each level; wherein the enhanced features are either the enhanced spatial features or the enhanced temporal features.

4. The multi-level power load forecasting method according to claim 3, characterized in that, Graph convolutional networks are used to extract spatial features from spatiotemporal data models of power grids at various levels.

5. The multi-level power load forecasting method according to claim 3, characterized in that, A multi-head temporal self-attention mechanism is used to extract the temporal features of the power grid spatiotemporal data model at each level.

6. The multi-level power load forecasting method according to claim 3, characterized in that, The compression and excitation modules are used to enhance key features in the spatial or temporal features to obtain enhanced spatial or temporal features.

7. The multi-level power load forecasting method according to claim 3, characterized in that, The third spatiotemporal features of the power grid spatiotemporal data models at each level are aggregated to obtain the first spatiotemporal features of the multi-level power grid spatiotemporal data models, including: Using the second spatiotemporal feature as the query, and the fusion results of the third spatiotemporal features of other levels of power grid spatiotemporal data models as the key and value respectively, the correlation between each node in the preset level of power grid spatiotemporal data model and all nodes in the other levels of power grid spatiotemporal data models is determined; wherein, the other levels of power grid spatiotemporal data models are power grid spatiotemporal data models of all levels other than the preset level. The weight matrix is ​​determined based on the correlation and the membership relationships between power grid nodes; Based on the weight matrix and the values, the aggregated features of each node in the preset-level spatiotemporal data model of the power grid are determined; The aggregated features are then spliced ​​together to obtain the first spatiotemporal feature of the multi-level power grid spatiotemporal data model.

8. The multi-level power load forecasting method according to claim 7, characterized in that, By performing residual connection and layer normalization operations, the first spatiotemporal features of the multi-level power grid spatiotemporal data model are fused with the second spatiotemporal features of the preset-level power grid spatiotemporal data model to obtain the fused spatiotemporal features.

9. The multi-level power load forecasting method according to claim 1, characterized in that, Based on the fused spatiotemporal characteristics, multiple fully connected layers are applied to determine the power load values ​​of the power grid at multiple future time steps.

10. The multi-level power load forecasting method according to claim 1, characterized in that, Based on the fused spatiotemporal features obtained by fusing the first spatiotemporal features of a multi-level power grid spatiotemporal data model with the second spatiotemporal features of a preset-level power grid spatiotemporal data model, the power load value of the power grid is determined, including: A multi-level spatiotemporal data model of the power grid is input into a load prediction model, which outputs the power load value of the power grid. The load prediction model includes a graph convolutional network, a multi-head temporal self-attention mechanism, a compression and excitation module, an attention module, residual connections, layer normalization, and a fully connected layer.

11. The multi-level power load forecasting method according to claim 10, characterized in that, The training process of the load prediction model includes: Using historical data from the multi-level spatiotemporal data model of the power grid as input and historical power load values ​​of the power grid as output, the load prediction model is trained by applying a gradient descent algorithm based on minimizing the total loss of the load prediction model; the total loss of the load prediction model includes prediction accuracy loss and physical constraint penalty loss.

12. The multi-level power load forecasting method according to claim 1, characterized in that, In the case of power grid levels including distribution area level, busbar level and substation level, the total load of a substation is equal to the sum of the loads of all distribution areas belonging to the substation, and the physical constraint penalty loss includes a first penalty term and a second penalty term. The first penalty term is represented as: ； The second penalty item is represented as follows: ； in, This is the first penalty item; The number of substations; Let be the predicted total load of the j-th substation; This represents the number of distribution areas belonging to the j-th substation. The predicted load belongs to the i-th distribution area of ​​the j-th substation; This is the relationship matrix between the j-th substation and the i-th distribution area belonging to the j-th substation; This is the second penalty item; The predicted load for the k-th node; This represents the rated capacity of the k-th node.

13. A multi-level power load forecasting device, characterized in that, The device includes: The model building module is used to construct multi-level spatiotemporal data models of the power grid based on the power grid topology at each level, the load data of each node in the power grid, and the physical parameters of each node in the power grid. The load value determination module is used to determine the power load value of each level of the power grid based on the fused spatiotemporal features obtained by fusing the first spatiotemporal features of the multi-level power grid spatiotemporal data model obtained based on the affiliation relationship between power grid nodes and the second spatiotemporal features of the preset-level power grid spatiotemporal data model.

14. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 12.

15. A computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 12.

16. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-12.

Images