Federal physical drive unit scheduling method considering information barrier and regulation authority
By constructing a federated physical-driven unit scheduling method based on spatiotemporal graph convolutional networks, the problems of data privacy and computational complexity in cross-regional power grid scheduling are solved, achieving efficient and accurate power grid collaborative scheduling and meeting real-time scheduling requirements.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SICHUAN UNIV
- Filing Date
- 2025-12-17
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies struggle to achieve globally optimal scheduling in cross-regional power grid dispatching due to the conflict between data privacy, computational complexity, and model accuracy. Furthermore, distributed coordination methods exhibit poor convergence performance, failing to meet real-time dispatching requirements.
A federated physical-driven unit scheduling method based on spatiotemporal graph convolutional networks is adopted. By constructing multiple scheduling agent agent models and co-training them using a personalized federated learning framework, combined with weakly supervised learning with differentiable projection layers and physical information enhancement, cross-regional resource optimization is achieved.
It enables the accurate capture of the spatiotemporal characteristics of the power grid while protecting data privacy, reduces reliance on high-quality labeled data, improves the model's generalization ability, and provides an efficient cross-regional power grid collaborative scheduling solution.
Smart Images
Figure CN121998448A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system operation optimization technology, and in particular to a federal physical drive unit scheduling method that takes into account information barriers and control authority. Background Technology
[0002] As a critical infrastructure for maintaining the operation of modern society, China's power system dispatch and control system is evolving towards refinement, marketization, and intelligence. In a typical multi-level dispatch system composed of regional power grid dispatching agencies and multiple provincial power grid dispatching agencies, achieving optimal cross-regional resource allocation and minimizing overall network operating costs are the core objectives of dispatching work. However, current technological approaches are generally limited by the inherent contradiction between data privacy, computational complexity, and model accuracy when dealing with this complex multi-stakeholder collaborative optimization problem.
[0003] Traditionally, this problem has been solved primarily through analytical model-driven methods based on classical optimization theory. Whether using centralized global optimization or distributed decomposition and coordination optimization, the foundation lies in constructing an accurate mathematical programming model for power grid operation. At its core, the Unit Commitment (UC) problem, due to its inclusion of 0-1 integer variables for generator start-up and shutdown, and the nonlinearity arising from fuel costs (typically a quadratic or higher-order function), is essentially a large-scale, non-convex mixed-integer nonlinear programming (MINLP) problem. In a centralized model, while theoretically achieving global optimum by constructing a globally unified MINLP model, its dependence on detailed model parameters within each provincial dispatch center (such as unit cost curves and output plans) constitutes a significant barrier to commercial privacy. Furthermore, solving large-scale MINLP problems inherently imposes a huge computational burden, is difficult and time-consuming, and fails to meet the timeliness requirements of real-time dispatch scenarios.
[0004] Distributed approaches alleviate these problems to some extent. These methods (such as those based on Lagrange relaxation or ADMM) decompose the global problem into multiple sub-problems and seek a convergent solution through multiple rounds of iterative coordination among scheduling agents. Each agent optimizes its sub-problem based on the boundary information from the previous round and then feeds back the new boundary information to the coordinator, repeating this process. However, this iterative coordination mechanism itself has inherent limitations. When faced with the inherent non-convex, nonlinear, and discrete decision-making characteristics of the power grid, the convergence performance of these algorithms often performs poorly, making it difficult to guarantee stable convergence to a high-quality solution within a finite decision-making time. Furthermore, some generator units directly dispatched by the grid dispatch center are physically located within the geographical and electrical boundaries of the provincial dispatch center. This overlap of authority and location further deteriorates the convergence of this method.
[0005] With the development of artificial intelligence technology, data-driven methods offer a new perspective for solving the aforementioned problems. However, their application in the field of collaborative scheduling is still in the initial exploratory stage and faces many challenges. First, traditional neural network models lack expressive power and struggle to effectively capture the complex spatiotemporal coupling characteristics inherent in power grid data. Second, existing data-driven methods are mostly supervised learning, heavily reliant on high-quality labeled data, making it extremely difficult to obtain optimal scheduling schemes covering large scales and multiple operating conditions, thus limiting the model's generalization ability. Finally, high-quality power grid operation data is scattered across independent scheduling agencies, forming insurmountable data silos. Due to data privacy and security restrictions, it is impossible to train these datasets centrally, significantly limiting the performance ceiling of data-driven methods. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a federal physical drive unit scheduling method that takes into account information barriers and control authority.
[0007] The objective of this invention is achieved through the following technical solution: a federated physical drive unit scheduling method that takes into account information barriers and control authority, the method comprising,
[0008] Multiple scheduling agent proxy models are constructed, each corresponding to a scheduling agent. The scheduling agent proxy model is a decision agent model based on a spatiotemporal graph convolutional network, including a connected STGCN network and a differentiable projection layer. The spatiotemporal features of power grid operation data are extracted through the STGCN network and a preliminary decision is output. The differentiable projection layer is used to map the preliminary decision to a feasible region that satisfies predetermined physical constraints.
[0009] Based on a personalized federated learning framework, multiple scheduling agent agent models are trained collaboratively. During the training process, the local data of each scheduling agent is kept locally, and only the shared parameter layer in the model is uploaded to the coordination server for aggregation and update, while the personalized parameter layer is kept locally for updating.
[0010] The real-time power grid operation status is input into the trained scheduling agent model, and through forward reasoning, a collaborative scheduling scheme that meets physical constraints and economic objectives is directly output.
[0011] Specifically, multiple scheduling agent agent models correspond to one network-level scheduling agent model and multiple provincial-level scheduling agent models; the network-level scheduling agent agent model is used to generate and distribute boundary conditions to each provincial-level scheduling agent agent model based on the information of the entire network; the provincial-level scheduling agent agent models are used to generate the unit output plan within their respective jurisdictions based on the received boundary conditions.
[0012] Specifically, the STGCN network includes multiple spatiotemporal convolutional blocks stacked sequentially. The spatiotemporal convolutional blocks adopt a time-space-time serial structure, including a time-gated convolutional layer, a spatial graph convolutional layer, and a time-gated convolutional layer connected sequentially.
[0013] The spatial graph convolutional layer employs spectral graph convolution based on Chebyshev polynomial approximation, defined as:
[0014] ;
[0015] In the formula, The learnable parameters of the graph convolution kernel; For graph convolution operators; The scaled Laplace matrix of Chebyshev polynomial; The characteristic matrix; is a learnable vector of polynomial coefficients; Laplace matrix The largest eigenvalue; The order of the neighborhood clustering of the graph convolution kernel;
[0016] Integrating residual connections in each spatial graph convolution operation, for the input features of the graph convolutional layer. Its final output Defined as the sum of the input features and the graph convolution result:
[0017] .
[0018] Specifically, temporal feature extraction is performed through the temporally gated convolutional layer, defined as follows:
[0019] ;
[0020] In the formula, This is a temporal convolution operation; and is the input to the gated linear unit (GLU), and is the two parts obtained by splitting the convolution output along the channel dimension; For Hadamah accumulation; Use the Sigmoid activation function; For any node in the power grid, the time series input is... ; The sequence length; The number of input feature channels; This represents the number of feature channels in the final output of the time-gated convolutional layer. For convolution kernel; For kernel width.
[0021] Specifically, the output of the spatiotemporal convolutional block is:
[0022] ;
[0023] In the formula, For the first The input to a spatiotemporal convolutional block; These are the temporally gated convolution kernels located at the top and bottom, respectively; is the spectral kernel for spatial graph convolution; ReLU is the activation function after graph convolution.
[0024] Specifically, the differentiable projection layer projects the original output of the STGCN network to a feasible region that satisfies linear equality constraints and boundary constraints by solving a quadratic programming projection problem that introduces slack variables; the physical constraints include at least the upper and lower limits of active power output of traditional generator sets, active power balance constraints of nodes, and new energy output constraints.
[0025] Specifically, the scheduling agent model is trained using a physically-enhanced weakly supervised learning strategy, with the training objective being to minimize a composite augmented Lagrangian function:
[0026] ;
[0027] In the formula, This results in economic cost losses. For the first Learnable parameters of the neural network of the intelligent agent in each province; For provincial transfer The system state variable set, including the active power output of the neural network. and start / stop status ; For provincial transfer The set of all internal physical constraints; for The first in the province The quantity of violation of an inequality constraint; The Lagrange dual variable corresponding to this constraint; This is the secondary penalty coefficient; It is a vector of non-negative relaxation variables introduced in the differentiable projection layer at the tail of the model; The penalty weights for slack variables.
[0028] Specifically, the collaborative training of multiple scheduling agent models based on a personalized federated learning framework includes:
[0029] S11. The network-level scheduling agent model generates boundary conditions based on the global state and distributes them to the provincial-level scheduling agent models.
[0030] S12. Each provincial-level scheduling agent model performs forward calculations based on the received boundary conditions to obtain the local output plan and calculate the local loss.
[0031] S13. Each provincial-level scheduling agent model calculates the gradient of the local loss with respect to the network scheduling boundary conditions, injects differential privacy noise into the gradient to mask the information, and then feeds back the processed privacy gradient to the network-level scheduling agent model.
[0032] S14. The network-level scheduling agent model receives the privacy gradients returned by each province, combines them with its own loss, and uses the chain rule to calculate and update the gradients of the network scheduling model parameters. At the same time, each provincial scheduling agent model updates its own model parameters using its local loss.
[0033] S15. Each provincial-level scheduling agent model uploads its updated shared parameter layer to the coordination server. The coordination server aggregates all uploaded shared parameters, updates the global shared model, and distributes the updated shared parameters to each provincial-level scheduling agent model.
[0034] S16. Repeat steps S11 to S15 until the model converges.
[0035] The present invention has the following advantages:
[0036] 1. This invention proposes a federated multi-agent collaborative scheduling framework that can simultaneously address information barriers between multiple provinces and complex unit control authority restrictions. This framework transforms the originally highly coupled and difficult-to-decompose scheduling problem caused by overlapping permissions into a well-structured and efficiently solvable physical driving problem by constructing and collaboratively training decision agent models for each scheduling entity. This provides a new and feasible technical path for solving power grid collaborative scheduling under specific complex constraints.
[0037] 2. This invention achieves a unified approach to accurate modeling of the spatiotemporal dynamic characteristics of the power grid and privacy protection for scheduling entities. By introducing a federated personalized spatiotemporal graph convolutional network as the decision agent for each scheduling entity, it can not only accurately capture the complex topological constraints and temporal evolution patterns contained in the power grid operation data, but also ensure, through the federated learning paradigm, that the core operational data of each entity (such as topological data, predicted load data, and unit output plans) does not leave the local area, thus fundamentally solving the problems of data silos and commercial privacy protection.
[0038] 3. This invention uses a physically-informed weakly supervised learning strategy, which significantly reduces data dependence. By directly embedding the physical laws of the power grid (in the form of physical loss function and projection layer) into the model training process, the dependence on large-scale, high-quality labeled data is greatly reduced. This enables the model to learn self-consistently from a wider range of more readily available unlabeled data, which greatly improves the model's generalization ability and feasibility in practical engineering. Attached Figure Description
[0039] Figure 1 This is a schematic diagram of the spatiotemporal graph convolutional network structure of the present invention;
[0040] Figure 2 This is a schematic diagram of the multi-agent joint optimization framework of the present invention;
[0041] Figure 3 This is a schematic diagram of the basic framework of the personalized federated learning FedPer protocol of the present invention. Detailed Implementation
[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0043] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0044] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0045] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.
[0046] like Figures 1 to 3 As shown, a federal physical drive unit scheduling method considering information barriers and control authority is described. This method includes:
[0047] Multiple scheduling agent proxy models are constructed, each corresponding to a scheduling agent. Each scheduling agent proxy model is a decision-making agent model based on a spatiotemporal graph convolutional network, including a connected STGCN network and a differentiable projection layer. The STGCN network extracts spatiotemporal features of power grid operation data and outputs preliminary decisions. The differentiable projection layer maps the preliminary decisions to feasible regions that satisfy predetermined physical constraints. The multiple scheduling agent proxy models correspond to one grid-level scheduling agent model and multiple provincial-level scheduling agent models. The grid-level scheduling agent model generates and distributes boundary conditions to each provincial-level scheduling agent model based on the network-wide information. The provincial-level scheduling agent models generate unit output plans within their respective jurisdictions based on the received boundary conditions.
[0048] To accurately simulate the internal optimal economic dispatch response of each provincial dispatch center under given external boundary conditions and effectively capture the inherent complex spatiotemporal coupling characteristics in power grid operation data, this invention proposes a specially designed decision agent model based on a spatiotemporal graph convolutional network (STGCN). Its overall architecture is a federated personalized spatiotemporal graph convolutional network (FedPer-STGCN). Structurally, this network consists of an STGCN network acting as the main body for feature extraction and preliminary decision-making, and a differentiable projection layer responsible for ensuring physical feasibility, connected in series. The STGCN network includes multiple sequentially stacked spatiotemporal convolutional blocks and an output block. The spatiotemporal convolutional blocks adopt a time-space-time serial structure and integrate residual connections to enhance gradient flow. Each spatiotemporal convolutional block includes a sequentially connected time-gated convolutional layer, a spatial graph convolutional layer, and a time-gated convolutional layer.
[0049] To capture the topological connectivity and global physical characteristics of the provincial power grid, this invention employs graph convolution directly applied to the graph structure data composed of power grid nodes. For a node in the graph, its state can be represented by a multi-dimensional feature vector. Therefore, the state of the entire provincial power grid can be represented by a feature matrix. The description, in which It is a provincial transfer The number of nodes, It is the feature dimension of each node, and the goal of graph convolution is to transform this feature matrix. To extract the spatial features of the system state, the operation is first performed to construct the spatial features of each provincial survey. Weighted adjacency matrix Based on this, the normalized graph Laplace matrix is calculated. Traditional spectral convolution relies on... The eigenvalue decomposition has a computational complexity of up to [missing information]. However, due to the limitations of traditional methods, they are difficult to apply to large-scale power grids. Therefore, this invention employs Chebyshev polynomials as an approximation strategy to approximate the spectral domain convolution kernel. The graph convolution kernel can be defined as a diagonal matrix of eigenvalues of the graph Laplacian matrix. A function This function can be derived from truncated Chebyshev polynomial of order Perform an approximate expansion:
[0050] ;
[0051] In the formula, is a learnable vector of polynomial coefficients; It is the order of the neighborhood aggregation of the graph convolution kernel; It is a scaled eigenvalue diagonal matrix, the purpose of which is to constrain the spectral radius to a specific size. To ensure numerical stability, among which It is the Laplace matrix The largest eigenvalue; applying graph convolution to the feature matrix. The process can be performed efficiently and is defined as follows:
[0052] ;
[0053] In the formula, The learnable parameters of the graph convolution kernel; For graph convolution operators; The scaled Laplace matrix of Chebyshev polynomial; The characteristic matrix; is a learnable vector of polynomial coefficients; Laplace matrix The largest eigenvalue; Let be the order of the neighborhood aggregated by the graph convolution kernel; through this approximation, the learning complexity of graph convolution is reduced from... Reduced to a value proportional to the number of edges. This enables it to be efficiently applied to large-scale power grid topologies.
[0054] To further enhance the model's gradient propagation capability and alleviate the degradation problem in deep network training, residual connections are integrated into each spatial graph convolution operation for the input features of the graph convolutional layer. Its final output Defined as the sum of the input features and the graph convolution result:
[0055] ;
[0056] This structure allows the model to retain the original feature information while learning new spatial features, greatly improving the model's performance and training stability.
[0057] To capture the temporal dependencies of various state variables in power system dispatching, such as load and unit output, a temporal feature extraction mechanism based on gated convolutional networks (Gated CNNs) is adopted. Compared to recurrent neural networks (RNNs) and their variants, Gated CNNs offer advantages such as faster training speed and simpler structure. Its core component, the gated linear unit (GLU), effectively controls the flow of information in the temporal dimension through a gating mechanism, thereby mitigating the gradient vanishing problem in deep networks. The temporal gated convolution layer consists of a kernel with a width of... It consists of a one-dimensional causal convolution and a gated linear unit (GLU) as a nonlinear activation. For any node in the power grid, the time-series input... , For sequence length, Design a convolution kernel to represent the number of feature channels. , will input Mapped to Temporal feature extraction is performed through the temporally gated convolutional layer, defined as follows:
[0058] ;
[0059] In the formula, This is a temporal convolution operation; and is the input to the gated linear unit (GLU), and is the two parts obtained by splitting the convolution output along the channel dimension; For Hadamah accumulation; Use the Sigmoid activation function; For any node in the power grid, the time series input is... ; The sequence length; The number of input feature channels; This represents the number of feature channels in the final output of the time-gated convolutional layer. For convolution kernel; For kernel width; Sigmoid gate Its function is to transform the feature map Each element is mapped to the interval (0,1), thus generating a gating coefficient vector for... The information flow within is finely and non-linearly adjusted.
[0060] To deeply model the spatiotemporal coupling characteristics of power grid operation, temporal convolution and spatial graph convolution are combined to construct a spatiotemporal convolution block with a temporal-spatial-temporal structure. For the th Input of each spatiotemporal convolutional block Its output The calculation process can be formally represented as:
[0061] ;
[0062] In the formula, For the first The input to a spatiotemporal convolutional block; These are the temporally gated convolution kernels located at the top and bottom, respectively; is the spectral kernel of the spatial graph convolution; ReLU is the activation function after graph convolution. This structure enables deep integration and propagation of temporal and spatial features.
[0063] The differentiable projection layer projects the original output of the STGCN network to a feasible region that satisfies linear equality constraints and boundary constraints by solving a quadratic programming projection problem that introduces slack variables; the physical constraints include at least the upper and lower limits of active power output of traditional generator sets, active power balance constraints of nodes, and new energy output constraints.
[0064] The output of the scheduling agent proxy model must strictly satisfy a series of physical constraints. This invention adds a differentiable projection layer to the output of the main network, constructing a least-squares projection problem. For the original proposal output of the neural network... The aim is to find a way to achieve the desired result while satisfying the core physical constraints. Feasible solution with minimum Euclidean distance Meanwhile, to address the unsolvable problem caused by conflicting boundary conditions in the early stages of training, non-negative relaxation variables are introduced. This can be expressed as the following optimization problem:
[0065] ;
[0066] In the formula, This is the corrected active power output matrix of all units in the province at all times, as output from the projection layer. The original active power output suggestion matrix of the unit is output by the neural network; The introduced slack variable vector is used to ensure that the optimization problem always has a solution; The penalty coefficient for the slack variables is set to a large value to suppress physical violations; The first output of the projection layer The generator set at time The effective output value; This indicates the total number of generator sets within the jurisdiction of the provincial dispatch center; For the purpose of online surveys Boundary conditions for the switching power of the time-period tie line and the output of the grid-dispatch and direct-dispatch units; for The slack variable components corresponding to the power balance constraint at any given time; for The total active power load of all nodes within the province at any given time; This is the start-stop status matrix of the unit output by STGCN; For Hadamah accumulation; and These represent the minimum and maximum active power outputs of each generator unit, respectively. This projection layer allows the final output to meet some physical constraints, significantly reducing training time and difficulty.
[0067] Node active power balance constraint: This constraint requires that at each node j and at each time t, the algebraic sum of the active power injected into that node is zero;
[0068] ;
[0069] In the formula, For the total active power injection of node j; Let be the total active load of node j; This represents the active power flow of the branch from node j to node k.
[0070] Traditional generator set active power output upper and lower limit constraints:
[0071] ;
[0072] In the formula, For the unit exist The start / stop status at any time, 1 for power on, 0 for power off; , The units Minimum / maximum active power output; generator set At any moment Those who have made contributions.
[0073] Constraints on new energy output:
[0074] ;
[0075] In the formula, For new energy units exist Maximum available predictable output at any given time; For new energy units Those who have made contributions.
[0076] For system-level constraints that couple all state variables of the province, reflect the overall system balance and safety margin, and include discrete logic constraints, this invention transforms them into physical loss functions during the training process, guiding model learning through gradient optimization, as follows:
[0077] Unit ramp rate constraint:
[0078] ;
[0079] ;
[0080] In the formula, and The units The upward and downward ramp rate limits; the unit ramp rate constraint is used to limit the maximum rate of change of the unit's active power output between adjacent time points.
[0081] Minimum continuous start / stop time constraint for the unit:
[0082] ;
[0083] ;
[0084] In the formula, and They are the generator sets The minimum continuous start-up and shutdown time. The minimum continuous start-up / shutdown time constraint is used to ensure that once the unit is started up or shut down, it will maintain this state for at least a certain period of time.
[0085] Maximum transmission capacity constraint of the line:
[0086] ;
[0087] In the formula, For any route within the province The trend; For any line The transmission limit.
[0088] Provincial system reserve constraint: This constraint requires that the total available capacity provided by all operating traditional generator units within the province is not less than the minimum reserve requirement of a certain system. ;
[0089] ;
[0090] In the formula, The minimum reserve capacity requirement determined by the provincial dispatch center at time t based on factors such as load and new energy forecast uncertainties; For the unit The maximum active power output limit; For the unit At the moment Actual planned value of active power output.
[0091] To avoid dependence on large-scale, high-quality optimal solution labeled data, this invention employs a physically-informed, weakly supervised learning strategy. First, preliminary federated pre-training is performed using a small amount of labeled data. Then, the learning strategy minimizes a composite augmented Lagrangian function. For local training objectives: ;
[0092] In the formula, This results in economic cost losses. For the first Learnable parameters of the neural network of the intelligent agent in each province; For provincial transfer The system state variable set, including the active power output of the neural network. and start / stop status ; For provincial transfer The set of all internal physical constraints; for The first in the province The quantity of violation of an inequality constraint; The Lagrange dual variable corresponding to this constraint; This is the secondary penalty coefficient; This is a vector of non-negative relaxation variables introduced in the differentiable projection layer at the tail of the model; These are the penalty weights for slack variables. During training, the model parameters... and dual variables It updates alternately, thus guiding the model to converge to a feasible solution more effectively than a simple quadratic penalty; the function consists of four parts: the economic cost objective, the quadratic penalty term for the physical constraints, the output cost of the grid-connected units, and the Lagrange dual variables (multipliers) corresponding to the physical constraints.
[0093] Economic cost loss Defined as the sum of three factors, the provincial dispatch center aims to minimize power generation costs and reduce the abandonment of renewable energy sources.
[0094] ;
[0095] In the formula, The operating cost of a traditional thermal power unit is a quadratic function; start-up and shutdown costs are also included. Costs incurred for starting and stopping thermal power units; To penalize unnecessary wind and solar power curtailment, this cost item aims to minimize the amount of renewable energy that is wasted, that is, to maximize the consumption of renewable energy.
[0096] By minimizing this composite augmented Lagrangian function and combining it with the forced mapping effect of the projection layer, the model is driven to find a good decision strategy that is both economical and fully satisfies various complex physical constraints.
[0097] During the training phase, the grid dispatch agent acts as a federated learning coordinator and a global goal definer. It is also a learnable decision-making agent. To fully utilize the inherent spatiotemporal characteristics of grid data, the grid dispatch agent's model maintains consistency with the provincial dispatch agent model in terms of network architecture, employing a Spatiotemporal Graph Convolutional Network (STGCN). This model is responsible for mapping macroscopic information at the grid level (such as total grid load forecasts and initial states) into specific boundary conditions distributed to each provincial dispatch center, namely, the tie-line power plans and the output plans of directly dispatched units in each province.
[0098] Unlike the provincial survey model, the online survey agent model, as a centralized, single entity, has different parameters. It does not distinguish between shared and personalized layers, nor does it participate in the personalized federated learning process. Instead, it directly updates the data through a gradient collaboration mechanism based on differential privacy.
[0099] Simultaneously, the decision variables—namely, the output plans of the grid-dispatch and directly dispatched generating units and the exchange power of inter-provincial tie lines—must also satisfy a series of physical constraints. These constraints, such as the self-operational constraints of the grid-dispatch and directly dispatched generating units (output upper and lower limits, ramp-up, minimum start-up and shutdown times, etc.), grid-dispatch system reserve constraints, and inter-provincial tie line power transmission limit constraints, correspond in type to the constraints at the provincial dispatch level. Therefore, this invention adopts a method similar to the provincial dispatch model, using a differentiable projection layer or setting a Lagrangian function to handle these constraints, which will not be elaborated here.
[0100] Based on a personalized federated learning framework, multiple scheduling agent agent models are trained collaboratively. During the training process, the local data of each scheduling agent is kept locally, and only the shared parameter layer in the model is uploaded to the coordination server for aggregation and update, while the personalized parameter layer is kept locally for updating.
[0101] To achieve joint optimization of the entire multi-agent system while protecting information barriers between multiple provinces, the training process of this invention follows a federated, physically driven collaborative framework. The core of this framework lies in simulating real scheduling hierarchies and information flows during offline training to achieve joint optimization of the multi-agent system.
[0102] Specifically, it includes:
[0103] S11. The network-level scheduling agent model generates boundary conditions based on the global state (global decision-making, boundary conditions for all provincial scheduling systems). (and distribute it to the corresponding provincial-level scheduling intelligent agent models);
[0104] S12. Each provincial-level dispatching intelligent agent model performs forward calculations based on the received boundary conditions to obtain the output scheme of the internal units and calculate local losses.
[0105] S13. Each provincial-level scheduling agent model calculates the gradient of the local loss with respect to the network scheduling boundary conditions, injects differential privacy noise into the gradient to mask the information, and then feeds back the processed privacy gradient to the network-level scheduling agent model.
[0106] S14. The network-level scheduling agent model receives the privacy gradients returned by each province, combines them with its own loss, and uses the chain rule to calculate and update the gradients of the network scheduling model parameters. At the same time, each provincial scheduling agent model updates its own model parameters using its local loss.
[0107] S15. Each provincial-level scheduling agent model uploads its updated shared parameter layer to the coordination server. The coordination server aggregates all uploaded shared parameters and updates the global shared model. The updated shared parameters are then distributed to the provincial-level dispatching intelligent agent models; the updated personalized parameter layers of each provincial dispatching agency are also distributed. The model parameters are always kept locally and do not participate in any uploading or aggregation. The model parameters themselves do not participate in the federated learning process.
[0108] S16. Repeat steps S11 to S15 until the model converges.
[0109] like Figure 3 As shown, to ensure privacy protection for provincial survey data within the aforementioned training framework, the collaborative parameter update process strictly adheres to the personalized federated learning FedPer protocol; under this framework, each provincial survey proxy model... Network parameters It is clearly divided into two parts: one part is a shared parameter layer consisting of multiple ST-Conv Blocks at the bottom of the model. The other part is the personalized parameter layer, which consists of the Output Block and projection layer at the top of the model. .
[0110] The real-time acquired power grid operating status is input into the trained scheduling agent model. Through forward inference, a collaborative scheduling scheme that satisfies both physical constraints and economic objectives is directly output. In the online application phase, a trained distributed collaborative decision-making system model is used for efficient forward inference. The higher-level dispatching agency (grid dispatcher) determines its final decision variables based on the real-time predictions and operating status. This refers to the actual boundary conditions of all provincial surveys. Subsequently, these determined boundary conditions are fed in parallel and at once into the pre-trained decision-making agent models of each provincial survey. In the process, forward calculations are performed to directly obtain the recommended power output scheme for generating units within the province. The entire online decision-making process is an online, non-iterative computational process, providing technical feasibility for achieving real-time economic scheduling.
[0111] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technology of the present invention without departing from the scope of the present invention are within the protection scope of the present invention.
Claims
1. A federalized physical drive unit scheduling method that takes into account information barriers and control authority, characterized in that: The method includes, Multiple scheduling agent proxy models are constructed, each corresponding to a scheduling agent. The scheduling agent proxy model is a decision agent model based on a spatiotemporal graph convolutional network, including a connected STGCN network and a differentiable projection layer. The spatiotemporal features of power grid operation data are extracted through the STGCN network and a preliminary decision is output. The differentiable projection layer is used to map the preliminary decision to a feasible region that satisfies predetermined physical constraints. Based on a personalized federated learning framework, multiple scheduling agent agent models are collaboratively trained. During training, each scheduling entity retains its local data locally, and only the shared parameter layer in the model is uploaded to the coordination server for aggregation and update, while the personalized parameter layer is kept locally for updates. The real-time power grid operation status is input into the trained scheduling agent model, and through forward reasoning, a collaborative scheduling scheme that meets physical constraints and economic objectives is directly output.
2. The federated physical drive unit scheduling method considering information barriers and control authority as described in claim 1, characterized in that: Multiple scheduling agent agent models correspond to one network-level scheduling agent model and multiple provincial-level scheduling agent models; the network-level scheduling agent agent model is used to generate and distribute boundary conditions to each provincial-level scheduling agent agent model based on the information of the entire network; the provincial-level scheduling agent agent models are used to generate the unit output plan within their respective jurisdictions based on the received boundary conditions.
3. The federated physical drive unit scheduling method considering information barriers and control authority according to claim 2, characterized in that: The STGCN network includes multiple spatiotemporal convolutional blocks stacked sequentially. The spatiotemporal convolutional blocks adopt a time-space-time serial structure, including a time-gated convolutional layer, a spatial graph convolutional layer, and a time-gated convolutional layer connected sequentially. The spatial graph convolutional layer employs spectral graph convolution based on Chebyshev polynomial approximation, defined as: ； In the formula, Learnable parameters representing the graph convolution kernel; For graph convolution operators, The scaled Laplace matrix of Chebyshev polynomial; The characteristic matrix; is a learnable vector of polynomial coefficients; Laplace matrix The largest eigenvalue; The order of the neighborhood clustering of the graph convolution kernel; Integrating residual connections into each spatial graph convolution operation, for the input features of the graph convolutional layer. Its final output Defined as the sum of the input features and the result of graph convolution: 。 4. The federated physical drive unit scheduling method considering information barriers and control authority as described in claim 1, characterized in that: Temporal feature extraction is performed through the temporally gated convolutional layer, defined as follows: ； In the formula, This is a temporal convolution operation; and is the input to the gated linear unit (GLU), and is the two parts obtained by splitting the convolution output along the channel dimension; For Hadamah accumulation; Use the Sigmoid activation function; For any node in the power grid, the time series input is... ; The sequence length; The number of input feature channels; This represents the number of feature channels in the final output of the time-gated convolutional layer. For convolution kernel; For kernel width.
5. The federated physical drive unit scheduling method considering information barriers and control authority as described in claim 4, characterized in that: The output of the spatiotemporal convolutional block is: ； In the formula, For the first The input to a spatiotemporal convolutional block; These are the temporally gated convolution kernels located at the top and bottom, respectively; For the spectral kernel of spatial graph convolution; ReLU is the activation function after graph convolution.
6. The federated physical drive unit scheduling method considering information barriers and control authority according to claim 4, characterized in that: The differentiable projection layer projects the original output of the STGCN network to a feasible region that satisfies linear equality constraints and boundary constraints by solving a quadratic programming projection problem that introduces slack variables. The physical constraints include upper and lower limits of active power output of traditional generator sets, active power balance constraints of nodes, and new energy output constraints.
7. The federated physical drive unit scheduling method considering information barriers and control authority according to claim 4, characterized in that: The scheduling agent model is trained using a physically-injected weakly supervised learning strategy, with the training objective being to minimize a composite augmented Lagrangian function: ； In the formula, This results in economic cost losses. For the first Learnable parameters of the neural network of the intelligent agent in each province; For provincial transfer The system state variable set, including the active power output of the neural network. and start / stop status ; For provincial transfer The set of all internal physical constraints; for The first in the province The quantity of violation of an inequality constraint; The Lagrange dual variable corresponding to this constraint; This is the secondary penalty coefficient; This is a vector of non-negative relaxation variables introduced in the differentiable projection layer at the tail of the model; The penalty weights for slack variables; Economic cost loss The calculation is as follows: ； In the formula, The operating cost of traditional thermal power units; For the unit At the moment Actual active power output (Planned value); Start-up and shutdown costs are the costs incurred in starting and stopping thermal power units. For the unit exist The start / stop status at any given moment; To punish unnecessary wind and light waste; For new energy units i in Contributing to the cause at all times.
8. The federated physical drive unit scheduling method considering information barriers and control authority according to claim 2, characterized in that: The collaborative training of multiple scheduling agent models based on a personalized federated learning framework specifically includes: S11. The network-level scheduling agent model generates boundary conditions based on the global state and distributes them to the provincial-level scheduling agent models. S12. Each provincial-level scheduling agent model performs forward calculations based on the received boundary conditions to obtain the local output plan and calculate the local loss. S13. Each provincial-level scheduling agent model calculates the gradient of the local loss with respect to the network scheduling boundary conditions, injects differential privacy noise into the gradient to mask the information, and then feeds back the processed privacy gradient to the network-level scheduling agent model. S14. The network-level scheduling agent model receives the privacy gradients returned by each province, combines them with its own loss, and uses the chain rule to calculate and update the gradients of the network scheduling model parameters. At the same time, each provincial scheduling agent model updates its own model parameters using its local loss. S15. Each provincial-level scheduling agent model uploads its updated shared parameter layer to the coordination server. The coordination server aggregates all uploaded shared parameters, updates the global shared model, and distributes the updated shared parameters to each provincial-level scheduling agent model. S16. Repeat steps S11 to S15 until the model converges.
9. The federated physical drive unit scheduling method considering information barriers and control authority according to claim 6, characterized in that: The upper and lower limits of the active power output of the conventional generator set are as follows: ； In the formula, For the unit exist The start / stop status at any time, 1 for power on, 0 for power off; , The units Minimum / maximum active power output; generator set At any moment Those who have made meritorious contributions; The active power balance constraint of the node is: ； In the formula, For the total active power injection of node j; Let be the total active load of node j; For the active power flow of the branch from node j to node k; The constraint on the output of the new energy source is: ； In the formula, For new energy units exist Maximum available predictable output at any given time; For new energy units Those who have made contributions.