CN-121998448-A - Federal physical driving unit scheduling method considering information barriers and regulation rights

CN121998448ACN 121998448 ACN121998448 ACN 121998448ACN-121998448-A

Abstract

The invention relates to the technical field of power system operation optimization, and discloses a federal physical drive unit scheduling method considering information barriers and regulation rights, which comprises the steps of constructing a plurality of scheduling agent models, wherein each scheduling agent model corresponds to one scheduling main body, and the scheduling agent model is a decision agent model based on a space-time diagram convolutional network; based on personalized federal learning framework, cooperatively training a plurality of scheduling agent models, inputting the real-time acquired power grid running state into the trained scheduling agent models, and directly outputting a cooperative scheduling scheme meeting physical constraint and economic targets through forward reasoning. The invention has the advantages that the scheduling method can effectively cope with multi-province information barriers, can efficiently process complex control authorities, and meanwhile, the weak supervision learning strategy fused with the physical laws obviously reduces the dependence on large-scale tag data, and improves the generalization capability and decision security of the model under complex working conditions.

Inventors

QIU GAO
LIU SHUHANG
WANG ZIXU
XIANG XIANMING
XU DUO

Assignees

四川大学

Dates

Publication Date: 20260508
Application Date: 20251217

Claims (9)

1. The federal physical drive unit scheduling method considering information barriers and regulation authorities is characterized by comprising the following steps of, Constructing a plurality of scheduling agent models, wherein each scheduling agent model corresponds to a scheduling main body, and the scheduling agent model is a decision agent model based on a space-time diagram convolution network and comprises a STGCN network and a micro-projectable layer which are connected with each other; in the training process, local data of each scheduling subject is kept locally, and only a shared parameter layer in the model is uploaded to a coordination server for aggregation and update, and the personalized parameter layer is kept locally for update; and inputting the power grid running state acquired in real time into a trained scheduling agent model, and directly outputting a cooperative scheduling scheme meeting physical constraint and economic targets through forward reasoning.
2. The federal physical drive unit scheduling method for accounting for information barriers and regulating authority of claim 1, wherein the plurality of scheduling agent models are a network-level scheduling agent model and a plurality of provincial scheduling agent models, the network-level scheduling agent model is used for generating and issuing boundary conditions to each provincial scheduling agent model according to the whole network information, and the provincial scheduling agent model is used for generating unit output plans in each jurisdiction according to the received boundary conditions.
3. The federal physical drive unit scheduling method for accounting for information barriers and regulating authority according to claim 2, wherein STGCN network comprises a plurality of sequentially stacked space-time convolution blocks, the space-time convolution blocks adopt a time-space-time series structure, and the space-time series structure comprises a time gating convolution layer, a space diagram convolution layer and a time gating convolution layer which are sequentially connected; The space diagram convolution layer adopts spectrogram convolution based on Chebyshev polynomial approximation, and is defined as: ; in the formula, A learnable parameter representing a graph convolution kernel; for the graph convolution operator, For scaled Laplace matrix A kind of electronic device An order chebyshev polynomial; Is a feature matrix; A learnable vector that is a polynomial coefficient; is a Laplace matrix Is the maximum eigenvalue of (2); Neighborhood order for graph convolution kernel aggregation; Integrating residual connections in each spatial map convolution operation, input features for the map convolution layers Its final output Is defined as the sum of the input features and the graph convolution result: 。
4. the federal physical drive unit scheduling method for accounting for information barriers and regulating authority according to claim 1, wherein the time feature extraction is performed by the time-gated convolution layer, and is defined as: ; in the formula, Is a time convolution operation; And The input of the gating linear unit GLU is two parts obtained by dividing the convolution output in the channel dimension; is Hadamard product; activating a function for Sigmoid; for time series input of any node in the grid, ; Is the sequence length; The number of the characteristic channels is input; the final output characteristic channel number of the time gating convolution layer is obtained; Is a convolution kernel; is nuclear wide.
5. The federal physical drive unit scheduling method for accounting for information barriers and regulating authority according to claim 4, wherein the output of the space-time convolution block is: ; in the formula, Is the first An input of a plurality of space-time convolution blocks; Time-gated convolution kernels at the upper and lower portions, respectively; is the spectrum kernel of the space diagram convolution, and the ReLU is the activation function after the diagram convolution.
6. The federal physical drive train dispatching method considering information barriers and regulation authority according to claim 4, wherein the micro-projectable layer projects the original output of STGCN networks to a feasible region meeting linear equality constraint and boundary constraint by solving a quadratic programming projection problem introducing relaxation variables, and the physical constraint comprises upper and lower limit constraint of active output, node active power balance constraint and new energy output constraint of a traditional generator set.
7. The federal physical driver set scheduling method for accounting for information barriers and regulatory authorities according to claim 4, wherein training the scheduling agent model with a physical information enhanced weakly supervised learning strategy is aimed at minimizing a complex augmented Lagrange function: ; in the formula, Is an economic cost loss; Is the first Learning parameters of the individual provincial intelligent agent neural network; To save the power Including the active output of the neural network And start-stop state ; To save the power A set of all physical constraints inside; Is that Province of the first kind Violation of the individual inequality constraint; is a Lagrangian dual variable corresponding to the constraint; The second order penalty coefficient; A non-negative relaxation variable vector introduced in a micro-projectable layer at the tail of the model; Penalty weights for relaxation variables; Cost of economy loss The calculation is as follows: ; in the formula, The method is the operation cost of the traditional thermal power generating unit; Is a unit At the moment of (3) An actual active force plan value; The starting and stopping cost is the cost generated by starting and stopping the thermal power unit; Is a unit At the position of A start-stop state at a moment; to punish unnecessary wind and light; is a new energy unit i Active force at the moment.
8. The federal physical driver set scheduling method for accounting for information barriers and regulating authority according to claim 2, wherein the collaborative training of a plurality of said scheduling agent models based on a personalized federal learning framework specifically comprises: S11, generating boundary conditions according to the global state by the network-level scheduling agent model, and issuing the boundary conditions to each provincial-level scheduling agent model; S12, forward calculation is carried out on each provincial scheduling agent model according to the received boundary conditions, so that a local output plan is obtained, and local loss is calculated; S13, calculating the gradient of the local loss to the network dispatching boundary condition by each provincial dispatching agent model, injecting differential privacy noise into the gradient to mask information, and transmitting the processed privacy gradient back to the network dispatching agent model; S14, the network-level dispatching agent model receives the privacy gradient returned by each province, calculates the gradient of the network-level dispatching agent model parameter by utilizing a chain rule in combination with the loss of the network-level dispatching agent model and updates the gradient; s15, uploading the updated shared parameter layer of each provincial scheduling agent model to a coordination server, and the coordination server aggregates all uploaded shared parameters, updates a global shared model and transmits the updated shared parameters to each provincial scheduling agent model; S16, repeating the steps S11 to S15 until the model converges.
9. The federal physical drive unit scheduling method for accounting for information barriers and regulating authority of claim 6, wherein the upper and lower limits of active output constraint of the conventional generator set are: ; in the formula, Is a unit At the position of The starting and stopping state at moment, 1 is started and 0 is stopped; , Respectively the units The minimum/maximum active set output of (2); Generating set At the moment of time Is an active force of (a); The node active power balance constraint is: ; in the formula, The total active injection for node j; Is the total active load of node j; active power flow for the branch from node j to node k; the new energy output constraint is as follows: ; in the formula, Is a new energy unit At the position of The maximum available predicted force at the moment; Is a new energy unit Is an active force of the (c).

Description

Federal physical driving unit scheduling method considering information barriers and regulation rights Technical Field The invention relates to the technical field of operation optimization of power systems, in particular to a federal physical drive unit scheduling method considering information barriers and regulation rights. Background As a key infrastructure for maintaining operation of modern society, the chinese dispatch control system is evolving towards refinement, marketization and intellectualization. In a typical multi-level dispatching system formed by a regional power grid dispatching mechanism and a plurality of provincial power grid dispatching mechanisms, cross-regional resource optimization configuration and whole-network operation cost minimization are realized, and the method is a core target of dispatching work. However, current technology paths are generally limited by inherent contradictions between data privacy, computational complexity, and model accuracy when dealing with this complex multi-subject co-optimization problem. Traditionally, the problem is solved mainly by an analytical model driven method based on classical optimization theory. The basis of the centralized global optimization or the distributed decomposition coordination optimization is to construct an accurate power grid operation mathematical programming model. As a core, the unit combination (Unit Commitment, UC) problem is essentially a large-scale, non-convex mixed integer nonlinear programming (MINLP) problem, as it includes 0-1 integer variables for the start and stop of the generator unit, as well as the nonlinear characteristics due to fuel costs (typically quadratic or higher order functions). In the centralization mode, although the global and unified MINLP model can be theoretically achieved, the dependence on the detailed model parameters (such as unit cost curve, output plan and the like) in each province forms a business privacy barrier which is difficult to surmount. Meanwhile, solving the large-scale MINLP problem constitutes huge calculation burden, is difficult to solve and long in solving time, and is difficult to meet the timeliness requirement of a real-time scheduling scene. The distributed mode alleviates the problems to a certain extent, the method (such as an algorithm based on Lagrange relaxation or ADMM) decomposes the global problem into a plurality of sub-problems, convergence solutions are sought among scheduling subjects through iterative coordination of multiple rounds, each subject optimizes own sub-problems according to boundary information of the previous round, and then new boundary information is fed back to a coordinator to do so. However, such iterative coordination mechanisms have inherent limitations, and in the face of inherent non-convex, nonlinear and discrete decision characteristics of the grid, the convergence performance of such algorithms tends to be poor, making it difficult to guarantee stable convergence to high quality solutions within a limited decision time. Further, the gensets that are partially dispatched by the network topology are physically located within the geographical and electrical scope of the provincial dispatching. This interleaving of rights with home further worsens the convergence of this approach. With the development of artificial intelligence technology, the data driving method provides a new view for solving the problems, but the application of the data driving method in the field of collaborative scheduling is still in a preliminary exploration stage, and a plurality of challenges are faced. Firstly, the traditional neural network model has insufficient expressive force, complex space-time coupling characteristics contained in power grid data are difficult to effectively capture, secondly, the traditional data driving method is mostly supervised learning, the strong dependence on high-quality label data is extremely difficult to obtain a large-scale optimal scheduling scheme covering multiple working conditions, the generalization capability of the model is limited, finally, high-quality power grid operation data are scattered in independent scheduling mechanisms, so that a data island which is difficult to surmount is formed, and the data cannot be intensively trained due to the limitation of data privacy and safety, so that the upper performance limit of the data driving method is greatly limited. Disclosure of Invention The invention aims to overcome the defects of the prior art and provides a federal physical drive unit scheduling method considering information barriers and regulation authorities. The invention aims at realizing the technical scheme that the federal physical drive unit scheduling method considering information barriers and regulation authorities comprises the following steps of, Constructing a plurality of scheduling agent models, wherein each scheduling agent model corresponds to a scheduling main body, and the scheduling