A reactive power aware kernel-driven spatiotemporal learning and lightweight direct current optimal power flow projection scheduling method

Abstract

A reactive power-aware graph kernel driven spatio-temporal learning and light DC optimal power flow projection scheduling method, comprising: constructing a reactive power-aware graph kernel based on the admittance matrix, extracting the reactive power-aware spatial dependence relationship of the grid nodes, and organizing the multi-period node feature tensor; using a spatio-temporal learner containing a graph convolutional encoder and a time series encoder to model the node features in space and time, and learn the cross-period evolution law of the node marginal price and the voltage phase angle; in the training process, the mechanism consistent physical regularization term of power flow safety and climbing consistency is introduced, and the influence of physical constraints is gradually enhanced through weight scheduling; the phase angle trajectory is converted into the equivalent initial value of unit output through physical mapping, and a light quadratic programming projection model with minimum deviation initial value is constructed to generate a feasible day-ahead scheduling plan under the constraints of power balance and other constraints. The method connects the full link from prediction to scheduling, balances speed and reliability, and provides an engineering usable path for day-ahead multi-period clearing and large-scale scenario analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to machine learning and power system optimization computing technology, in particular to a reactance-aware graph kernel driven spatiotemporal learning and lightweight direct current optimal power flow projection scheduling method. BACKGROUND

[0002] Under the background of high proportion of renewable energy access, deepening of power market mechanism and coexistence of large-scale scenario simulation demand, the day-ahead clearing needs to determine the unit output plan and node marginal price (LMP) on the cross-period constrained power grid model at the same time, and to meet the power balance, unit output boundary, line flow limit and climbing constraints. This task requires near real-time computing response, and must ensure the physical feasibility and cross-period consistency of the solution, which has become a key bottleneck for dispatch-market integrated simulation and strategy evaluation.

[0003] Engineering practice has long followed the optimization modeling route based on power transfer distribution factor (PTDF) and injection shift factor (ISF): through PTDF / ISF, the sensitivity of line flow to node injection is linearized, and then the direct current optimal power flow (DC-OPF) is converted into a large-scale linear / quadratic programming problem; then relying on industrial-level commercial solvers or open-source quadratic programming / linear programming (QP / LP) solvers, it is solved by time period or block. The advantage of this route is the maturity of feasibility and interpretability, clear constraint expression and strict verification. However, when cross-period constraints are coupled, a large number of scenarios are rolled, and the network size is hundreds to thousands of nodes, the solution frequency and problem size increase multiplicatively, resulting in a significant increase in computing load and time delay pressure, making it difficult to meet the needs of batch evaluation, market sensitivity analysis and strategy simulation with "near real-time + high frequency".

[0004] To alleviate the time delay bottleneck, in recent years, there have been multiple data-driven learning alternatives, mainly focusing on single-period DC approximate power flow, LMP estimation, feasibility discrimination or constraint softening related sub-problems, often using graph neural networks (GNN) to describe network topology correlation, or using shallow time series / statistical methods to handle short-term correlation, to shorten the evaluation time and assist operation decision-making.

[0005] However, implementing alternative learning methods in engineering still faces key challenges: First, relying solely on topological adjacency graphs is insufficient to accurately reflect the reactance sensing intensity and power distribution patterns characterized by the susceptance matrix. Second, physical constraints during training are often loosely applied using empirical penalties, making it difficult to connect them systematically along the "phase angle → injection → power flow" mechanism chain, and they are not adequately integrated with consistency constraints for cross-time-period ramping. Third, excessively heavy post-processing optimization can negate the speed advantage of the inference end, while excessively lightweight optimization makes it difficult to guarantee strict feasibility and cross-time-period consistency. Under the comprehensive requirements of "high-frequency evaluation + strict feasibility + multi-time-period consistency," existing solutions still lack a lightweight, closed-loop, and engineering-usable end-to-end process.

[0006] Existing numerical optimization schemes, after introducing cross-time constraint coupling, massive scenarios, and medium-to-large-scale networks, experience a simultaneous increase in both the scale of a single optimization and the frequency of solutions, leading to significant computational latency and resource consumption. Meanwhile, learning-based alternatives, lacking a strict physical constraint loop, are prone to issues such as power flow exceeding limits, balance deviations, and inconsistent ramping across time periods, impacting direct execution on the engineering side. Therefore, the industry urgently needs a hybrid end-to-end process: the front end uses a reactance-aware spatiotemporal learner to provide initial LMP / phase angle values ​​consistent with DC mechanisms at the millisecond level, and suppresses exceeding limits and abrupt changes during training through link regularization ("phase angle → injection → power flow") and ramping regularization; the back end then uses lightweight QP projection to make minimal corrections under power balance, unit boundaries, and line current limits, utilizing hot start and rapid acceptance channels to ensure throughput and stability. This collaborative process must simultaneously achieve high speed, strict feasibility, cross-time consistency, and ease of verification and maintenance to meet the dual requirements of batch evaluation and online application.

[0007] In addition, the existing solutions also have the following problems:

[0008] Insufficient reactance sensing representation: Conventional GNNs that construct graphs using topological adjacency do not explicitly encode the susceptance matrix. The power distribution pattern and node correlation strength implied by the equivalent Laplace method limit the physical consistency and generalization ability of the "state → phase angle / power flow / LMP" mapping; a graph kernel / propagation operator based on reactance sensing (such as based on...) is needed. Symmetric normalization and multi-order MixHop are used to improve the efficiency of message passing.

[0009] The physical regularization lacks a close connection with the learning objectives: empirical penalty terms are difficult to constrain DC safety along the mechanism chain of "phase angle → node injection → line flow", and their integration with cross-time-period ramping is often loose. It is necessary to adopt consistent regularization and progressive weight scheduling during the training period to significantly reduce potential limit violations and mutations without compromising the convergence of the main task.

[0010] Lack of a strictly feasible light-weight projector: Overly heavy post-processing optimization will swallow the reasoning speed bonus, and heuristic patching is difficult to guarantee strict satisfaction of power balance, unit commitment boundaries, and line flow limits. A light-weight QP projector that can be hot-started, with fast acceptance and failure fallback mechanisms, is needed to generate executable plans with minimal deviation under strict DC constraints and maintain robust throughput in batch scenarios.

[0011] True multi-period coupled day-ahead learning: Independent learning per period ignores ramping and inter-period consistency, making it difficult to directly form an engineering usable daily plan. The model needs to explicitly depict ramping and smoothness through time-dependent modeling modules and inter-period physical regularity inside the model, making the front-end initial value naturally "more feasible" and reducing the burden of back-end projection.

[0012] It should be noted that the information disclosed in the above background section is only for understanding the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0013] The main purpose of the present application is to overcome the defects existing in the background art, and to provide a reactance-aware graph kernel driven spatio-temporal learning and light-weight DCOP projection scheduling method.

[0014] To achieve the above purpose, the present application adopts the following technical solutions:

[0015] A reactance-aware graph kernel driven spatio-temporal learning and light-weight DCOP projection scheduling method, comprising the following steps:

[0016] S1, construct a reactance-aware graph kernel based on the admittance matrix, for extracting the reactance-aware spatial dependence relationship between nodes of the power grid, and organizing multi-period node feature tensors as input;

[0017] S2, use a spatio-temporal learner to model the node features jointly in space and time, the spatio-temporal learner includes a graph convolutional encoder and a time series encoder, for learning the cross-period evolution law of the node marginal price and the voltage phase angle;

[0018] S3, introduce mechanism-consistent physical regularization terms in the training process, including power flow safety regularization and ramp consistency regularization, and gradually enhance the influence of physical constraints through weight scheduling strategy;

[0019] S4, convert the phase angle trajectory output by the learner into equivalent unit output initial values through physical mapping, and construct a light-weight quadratic programming projection model with the minimum deviation initial value as the target, to generate a feasible day-ahead scheduling plan under the conditions of satisfying power balance, line flow limit, unit output boundary and cross-period ramping constraints.

[0020] A computer readable storage medium stores a computer program, the computer program is implemented when executed by a processor to realize the reactance-aware graph kernel driven spatiotemporal learning and lightweight DC optimal power flow projection scheduling method.

[0021] A computer program product comprises a computer program, the computer program is implemented when executed by a processor to realize the reactance-aware graph kernel driven spatiotemporal learning and lightweight DC optimal power flow projection scheduling method.

[0022] The present application has the following beneficial effects:

[0023] The present application proposes a reactance-aware graph kernel driven spatiotemporal learning and lightweight DC optimal power flow projection scheduling method, which constructs a set of "reactance-aware graph kernel spatiotemporal learner + mechanism-consistent physical regularization + lightweight strict DC-OPF projection" collaborative scheme, which is a hybrid end-to-end method for day-ahead multi-period clearing high-speed evaluation and strict feasible plan generation. The method of the present application embeds a lightweight DC-OPF projection link between the learner output and the scheduling executable solution, which connects the whole link process from prediction to scheduling, and takes into account the speed and reliability, and provides an engineering usable technical path for day-ahead multi-period clearing and large-scale scenario analysis.

[0024] Specifically, the method uses a susceptance matrix driven graph kernel to depict the reactance-aware characteristics, uses a graph neural network combined with a strong time sequence structure of TCN and BiLSTM to learn the cross-period evolution of price and phase angle, constructs a physical regularization through a DC mechanism chain in the training period to obtain a "physically friendly" prediction, and then projects it to an executable plan under strict constraints through a lightweight quadratic programming after inference. The framework has a structured understanding and automated processing capability for day-ahead clearing and network constraints, can automatically analyze multiple types of constraints such as power balance, line flow limit, unit output and ramping, directly generate candidate trajectories of price / phase angle / output / power flow and quickly correct them under strict DC constraints without human intervention, significantly reduce modeling and solving complexity; under the synergistic action of strong time sequence modeling and physical regularization, gradually approach the executable scheduling interval, balance speed and consistency under the premise of guaranteeing safety constraints, and provide an engineering usable technical path for multi-scenario high-frequency evaluation and market strategy research and judgment.

[0025] The hybrid end-to-end method of "spatial-temporal learner of reactance-aware graph kernel + mechanism-consistent physical regularization + lightweight strict DC-OPF projection" proposed in the present application has achieved a significant breakthrough in technical effect. First, the present application introduces a graph kernel based on a reactance-aware matrix, which expands the traditional adjacency relationship based on topology to a propagation mechanism consistent with the strong correlation between reactance and admittance, ensuring that the message passing between nodes is closer to the real power distribution law of the power grid. This improvement significantly improves the physical consistency of the learning model in the "state -> phase angle / power flow / LMP" mapping, enabling it to maintain strong generalization ability in medium and large-scale systems and unseen disturbance scenarios. Further, the present application uses a strong time series modeling structure of TCN + BiLSTM at the learning end, explicitly depicting the load evolution at the day-ahead scale and the cross-period coupling characteristics, thereby performing outstandingly in capturing short-term ramping and long-term smoothness, making the initial values generated by the model in the multi-period scenario naturally more consistent with the physical law. On this basis, the present application embeds physical regularization along the mechanism chain of "phase angle - injection - power flow" during training, and closely combines it with the cross-period ramping constraint through a weight increasing mechanism, thereby effectively suppressing power flow violations and ramping mutations without affecting the convergence of the main task. This design not only enhances the physical friendliness of the learning end output, but also significantly reduces the correction burden of the subsequent strict projection link.

[0026] In the post-processing link, the present application no longer relies on heuristics or heavy optimization, but constructs a lightweight strict DC-OPF projection module. This module takes the learning output as the hot start initial value, the objective function is only the minimum deviation, the constraint conditions cover power balance, line flow limit and unit boundary, and combines with the fast acceptance and failure bottom-up mechanism. In this way, a strictly feasible daily scheduling plan can be generated within milliseconds of time scale, realizing the integration of "high-speed candidate generation + strict constraint projection". Finally, the present application not only solves the bottleneck of existing numerical optimization methods in terms of scale and time delay, but also overcomes the problem of insufficient physical consistency of pure learning methods, achieving high-speed, strictly feasible, cross-period consistent low-cost scheduling. More importantly, this method takes into account the reviewability and operability of the engineering side, providing a new technical path with accuracy, explainability and deployability for batch scenario evaluation and online application of power systems.

[0027] Other benefits of the embodiments of the present application will be further described below. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The hybrid end-to-end scheduling method flow chart based on reactance-aware graph kernel, spatial-temporal learning and strict DC-OPF projection of the embodiments of the present application. DETAILED DESCRIPTION

[0029] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0030] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0031] This invention aims to address the technical bottlenecks in existing power system day-ahead clearing, such as high computational latency and resource consumption in numerical optimization schemes under cross-time period constraints, massive scenarios, and medium-to-large-scale networks, as well as insufficient physical consistency of pure learning-based alternatives and the tendency for power flow exceeding limits and inconsistent ramping across time periods. It aims to meet the engineering requirements of "high-frequency evaluation + strict feasibility + multi-time period consistency" by proposing a spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by a reactance-sensing graph kernel. By embedding a lightweight DC-OPF projection link between the learner output and the executable scheduling solution, it streamlines the entire process from prediction to scheduling, balancing speed and reliability, and providing an engineering-usable technical path for day-ahead multi-time period clearing and large-scale scenario analysis.

[0032] See Figure 1 This invention provides a spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by reactance sensing graph kernel, comprising the following steps:

[0033] Step S1: Construct a reactance sensing kernel based on the susceptance matrix to extract the reactance sensing spatial dependency between power grid nodes, and organize multi-time period node feature tensors as input.

[0034] In some embodiments, the construction of the reactance sensing kernel in step S1 includes: constructing a reactance sensing weight matrix based on the absolute values ​​of the off-diagonal elements of the susceptance matrix, with the diagonal elements set to zero; adding self-loops to the weight matrix, calculating the degree matrix and performing symmetric normalization to obtain a numerically stable reactance sensing kernel; pre-calculating the multi-order propagation matrix of the kernel to expand the spatial receptive field, and employing a learnable weighted multi-channel parallel fusion mechanism to combine the outputs of the local non-propagation path, single-hop propagation path, and multi-hop propagation path.

[0035] Step S2: Use a spatiotemporal learner to perform spatiotemporal joint modeling of node features. The spatiotemporal learner includes a graph convolutional encoder and a temporal encoder, which are used to learn the cross-time period evolution of node marginal electricity price and voltage phase angle.

[0036] In some embodiments, the graph convolutional encoder of the spatio-temporal learner in step S2 includes: performing linear transformation on the node feature matrix of each time period along a local path, a single-hop propagation path, and a multi-hop propagation path, respectively; weighting and fusing the outputs of the three paths through learnable gating coefficients to obtain a hidden representation; applying a nonlinear activation function, random deactivation, and layer normalization processing in sequence after fusion, and introducing a residual connection to maintain gradient flow; gradually increasing the feature dimension through a multi-layer stacking structure, and adding cross-layer projection between layers to alleviate information degradation.

[0037] In some embodiments, the time series encoder in step S2 adopts a TCN and BiLSTM joint structure, and includes: performing a dilated convolution operation on the time series sequence output by graph encoding using a time convolutional network, covering multi-scale time receptive fields through convolution layers with different dilation coefficients; feeding the output of the time convolutional network into a bidirectional long short-term memory network to capture long-term cross-time period dependencies in the forward and backward directions; dynamically weighting and fusing the outputs of the time convolutional network and the bidirectional long short-term memory network through a learnable gating mechanism, and adaptively adjusting the contribution ratio of medium and short-term features and long-term features during the training process.

[0038] Step S3, introducing mechanism-consistent physical regularization terms including power flow safety regularization and ramp consistency regularization, and gradually enhancing the influence of physical constraints through weight scheduling strategies.

[0039] In some embodiments, the mechanism-consistent physical regularization term in step S3 includes: linearly mapping the predicted voltage phase angle to node injection power through a susceptance matrix, and then mapping it to line power flow through a power transmission distribution factor matrix; applying a distance-based convex function penalty to the power flow component that exceeds the line thermal stability limit to guide the predicted result to approach the feasible region; deriving equivalent unit output based on the predicted phase angle, and applying a regularization penalty to the part of the output change between adjacent time periods that exceeds the ramp rate limit; adopting a staged weight scheduling strategy, with the prediction task loss being dominant in the early stage of training, and gradually linearly increasing the weight coefficient of the physical regularization term as the training round increases.

[0040] Step S4, converting the phase angle trajectory output by the learner into equivalent unit output initial values through physical mapping, and constructing a lightweight quadratic programming projection model with the goal of minimizing the deviation from the initial values, generating a feasible day-ahead scheduling plan under the conditions of satisfying power balance, line power flow limit, unit output boundary, and cross-time period ramp constraint.

[0041] In some embodiments, the lightweight quadratic programming projection model in step S4 includes: converting the voltage phase angle predicted by the learner into node injection power through the admittance matrix, and combining the load prediction to synthesize the initial value of the equivalent unit output; taking the minimum norm deviation between the actual unit output and the initial value as the objective function to construct a multi-period joint optimization problem; the constraint conditions include: system power balance constraint, upper and lower limit constraint of unit output, line flow limit constraint based on power transmission distribution factor, and unit ramp rate constraint between adjacent time periods; a hot start mechanism based on the solution of adjacent time periods is used to accelerate the solving process, and the initial value that meets the constraint tolerance is directly adopted, and the fast quadratic programming solver is started for projection correction in the case of out-of-bound.

[0042] In some embodiments, the fast solving mechanism of the projection model includes: performing constraint satisfaction pre-check before projection, if the equivalent unit output initial value of a certain period is within the tolerance range of all constraints, then skip the optimization solving of this period; for the period that needs to be solved, the optimization solution of the previous adjacent period is used as the hot start initial value to reduce the number of iterations; in the case of extreme out-of-bound, a rebalancing adjustment algorithm based on line flow sensitivity is started to pull the power flow back within the limit value while maintaining power balance.

[0043] In some embodiments, the method further includes: uniformly determining a reference bus in the data preprocessing stage, and maintaining the consistency of the reference bus throughout the whole process; re-anchoring the voltage phase angle output by the learner before performing physical mapping, subtracting the phase angle value of the reference bus to eliminate the zero-point uncertainty; using the same unit system and power reference in the training, inference and projection stages to ensure the system consistency of physical quantity calculation.

[0044] In the embodiment of the application, the front-end learner is used to The symmetric normalized graph kernel derived and the MixHop multi-order propagation explicitly encode the reactance perception characteristics, the temporal convolutional network (TCN) and the bidirectional long short-term memory network (BiLSTM) capture the cross-period dependence and the ramp behavior; the training period applies physical regularization along the DC mechanism chain of “phase angle→injection→power flow”, and uses gradual weight scheduling to improve stability and physical friendliness; the inference backend uses a hot start lightweight QP projection to make the minimum correction under the power balance, unit boundary and line flow limit, and quickly obtains a strictly feasible multi-period plan, thereby achieving a synergistic optimization between throughput speed and engineering executability.

[0045] The technical advantages of the reactive-aware graph kernel driven spatio-temporal learning and lightweight DC optimal power flow projection scheduling method are mainly that a hybrid end-to-end architecture of "reactive-aware graph kernel spatio-temporal learner + mechanism-consistent physical regularization + lightweight strict DC-OPF projection" is constructed, which breaks through the traditional technical bottleneck and also considers engineering practicability: firstly, the reactive-aware graph kernel driven by the admittance matrix replaces the pure topological adjacency modeling, so that the node message passing is more consistent with the real power distribution law of the power grid, and the physical consistency and generalization ability of the "state→phase angle / power flow / LMP" mapping under the large-scale system and unknown disturbance scenario are significantly improved; secondly, TCN and BiLSTM are combined for strong time sequence modeling, so as to accurately capture the short-term climbing and long-term smoothing characteristics of the day-ahead scheduling, and the initial value generated by the learning end naturally conforms to the physical law; thirdly, the physical regularization is applied along the mechanism chain of "phase angle→injection→power flow" during the training period, and the weight is adjusted in stages, so that the power flow limit and climbing mutation are suppressed without affecting the convergence of the main task, and the correction burden of the subsequent projection link is greatly reduced; fourthly, the lightweight DC-OPF projection module takes the learning output as the hot start initial value, and combines the fast acceptance and boundary rebalancing mechanism to realize millisecond-level solving, so as to ensure that the generated full-day scheduling plan strictly meets the power balance, line flow limit and other constraints; finally, the architecture not only solves the time delay problem of traditional numerical optimization in large-scale and multi-period scenarios, but also overcomes the defects of the physical inconsistency of pure learning methods, and considers high speed, strict feasibility and cross-period consistency, and has the characteristics of engineering review and operation, which provides an accurate, interpretable and easy-to-deploy technical path for power system day-ahead multi-period clearing, batch scenario evaluation and online application.

[0046] The specific embodiments of the present application and the algorithm implementation examples thereof are further described below.

[0047] The present application proposes a reactive-aware graph kernel driven spatio-temporal learning and lightweight DC optimal power flow projection scheduling method, which designs a hybrid end-to-end method of "reactive-aware graph kernel spatio-temporal learner + mechanism-consistent physical regularization + lightweight strict DC-OPF projection".

[0048] Specifically, the method of the present application comprises the following steps:

[0049] (1) DC-OPF modeling and reactive-aware graph kernel

[0050] (1-1) variable and data organization

[0051] The present application is aimed at day-ahead multi-period (such as 96 15-minute periods) clearing and large-scale scenario evaluation. Let the period index set be , the node (bus) index set be , and the line index set be . In each period , the node The core state and the decision variable include the unit active power output and the node phase angle .

[0052] The whole process of the application takes the power with the radian system phase angle and MW dimension as the standard: the phase angle is uniformly expressed in the radian system, and the unit is kept consistent in all aspects related to the admittance matrix and the power flow calculation, if the input data source provides the angle system, it will be converted into the radian system in the data preprocessing stage ; in the calculation related to , the system reference power (uniformly taken as 100 MVA) is introduced to map from the per unit system to the MW unit, so that , and the same power reference constant is kept in the sample generation, training and projection stage, which ensures strict consistency with the DC-OPF modeling.

[0053] Under the modeling framework of DC-OPF, the basic physical quantities required by the input end are derived from the system static parameters or the day-ahead prediction module: the active load prediction of the node side is generated by the day-ahead prediction process (historical alignment, abnormality elimination and time period modeling), as an exogenous quantity in the period; the network structure is summarized as the admittance matrix through the line parameters and topology, which is essentially a Laplace type admittance matrix under the DC approximation, reflecting the reactance perception relationship between nodes; the power flow transfer factor can be derived by and the line parameter in the modeling stage, which is used to quickly represent the linear influence of node injection on the line power flow; the line thermal limit comes from the system static parameters, which is given by the operation and maintenance parameters, and limits the maximum power flow that each branch can withstand; the output upper bound of the generator unit or the node level comes from the unit operation constraint, when a unit is shut down, the upper bound can be set to zero to represent it.

[0054] To ensure the complete consistency of the physical interface of the "data end-learning end-projection end", the application uniformly sets the node number and reference bus convention in the data organization and processing link, avoiding the systematic deviation caused by the inconsistency of index or reference.

[0055] The samples are organized into tensors according to days: the input , wherein = 6, and the six-dimensional characteristics are as follows:

[0056] The difference between the node available output upper bound and the load ,

[0057] The difference between the node available output lower bound and the load ,

[0058] Linear cost coefficient ,

[0059] Secondary cost coefficient ,

[0060] Node active load forecasting ,

[0061] The indicator mask genmask for the bus where the unit is located.

[0062] Corresponding output These are the nodal marginal price (LMP) and the voltage phase angle, respectively. During training and deployment, a uniform z-score standardization based on training set statistics is adopted to ensure the numerical stability of features with different dimensions; the phase angle is always based on the reference bus. Anchoring ( Eliminate zero-point uncertainty from the source.

[0063] (1-2) Core Relationships and Constraints

[0064] First, a standard DC-OPF mathematical expression for each time period is given to clarify the objective and feasible region.

[0065] For any time period Let the active power output vector of the node be... Load vector Line power flow vector .

[0066] DC-OPF aims to minimize the total system power generation cost.

[0067]

[0068] in For the assembly of units, , , This is the cost coefficient.

[0069] The constraints include:

[0070] (i) Active power balance at nodes

[0071]

[0072] in To connect to the node A collection of generator units, For nodes The set of adjacent nodes, This represents the branch reactance. This constraint can also be written in vector form: and and satisfy .

[0073] (ii) upper and lower limits of generator active power output

[0074]

[0075] (iii) branch power flow capacity constraints

[0076]

[0077] or written as a vector inequality: where .

[0078] (iv) reference bus constraints

[0079]

[0080] To eliminate the non-uniqueness of phase angle solutions, the reference bus is determined once in the data generation stage (preferably the system balance bus; if there is no unified marking, it is automatically selected according to historical phase angle statistics), and is consistent in / the definition of PTDF and all training-reasoning-projection links. The phase angle output by the model is uniformly re-anchored before entering any physical mapping or loss .

[0081] Therefore, the mathematical model corresponds to the equivalent vector-matrix DC-OPF form as follows:

[0082]

[0083] Wherein:

[0084] (3-1) ensures the active power balance of the system;

[0085] (3-2), (3-3) describe the line power flow calculation and thermal limit constraints based on PTDF;

[0086] (3-4) is the upper and lower limit constraint of the unit output;

[0087] (3-5a), (3-5b) describe the linear relationship between node injection and voltage phase angle under DC approximation, and ensure that the injection is consistent with the unit output / load;

[0088] (3-6) eliminates the zero-point ambiguity by fixing the reference bus phase angle.

[0089] In addition, when coupled across time periods, a ramping constraint needs to be added to reflect the smoothness of the output change:​

[0090]

[0091] Based on the above standard model, the application adopts a hybrid process of "learning end generates physically friendly initial value + light quadratic programming (QP) projection strict feasibility": the learning end directly predicts the space-time trajectory of ; the phase angle is mapped to the node injection by formula (3-5a), and then combined with the load to obtain:

[0092]

[0093] as the "center point" of projection. In the projection stage, instead of introducing price or phase angle variables, the "minimum deviation from the initial value" is used as the criterion to jointly optimize the unit output of the whole day in each time period, taking the center point derived by the learning end as the benchmark, minimizing the total deviation and explicitly considering the ramping constraint, so as to inevitably ensure the consistency across time periods:

[0094]

[0095] The solved strictly satisfies power balance, unit upper and lower limits, line flow limits and cross-period ramping, and serves as an executable daily plan.

[0096] In engineering implementation, if the center point of some time period meets all the constraints within the tolerance, it enters the fast acceptance channel to skip numerical solution; the rest of the time periods are solved by the QP solution process supporting hot start, and in rare boundary cases, the flow-aware rebalancing bottom is triggered to ensure that the output is always feasible and close to the structure information of the learner.

[0097] (1-3) Construction of reactance-aware graph kernel

[0098] Using only the topological adjacency matrix (0 / 1 edge), graph modeling will ignore the information of the strength of the association caused by the difference in different branch reactance; and in the DC framework, the near-end power-phase angle relationship is dominated by , and the sensitivity of injection and power flow to phase angle disturbance is closely related to the size of the reactance. The reactance-aware graph constructed by taking the susceptance value as the edge weight is closer to the real message transmission strength, which is the key to moving from a pure topological graph to a physically aware graph.

[0099] First, define the reactance-aware weight matrix as a non-diagonal weight:

[0100]

[0101] Its physical meaning is: ​​Reflecting bus With the interaction strength between the smaller, the line reactance the greater, the stronger the coupling of power disturbance.

[0102] Then add a unit ring and do symmetric normalization, get numerical stable graph kernel.

[0103] Let 、 , defined:

[0104]

[0105] The normalization balances the scales of nodes with different degrees to the same spectral domain, which is conducive to stable forward propagation and gradient transmission, and suppresses degree bias.

[0106] To cover further dependencies without deepening the network layers, the two-hop propagation kernel is pre-computed. The physical intuition is that the series reactance effect of "adjacent-neighbor" corresponds to the two-hop influence, which is close to the indirect effect of multi-bus injection in PTDF on far-end power flow.

[0107] In message passing, multi-channel parallel fusion is adopted, and unit channel , single-hop and double-hop are used jointly. Once forward propagation can be written as:

[0108]

[0109] Where is the learnable fusion coefficient, is the linear mapping of the corresponding channel.

[0110] Compared with the method containing only , the above "reactance-aware graph kernel + multi-scale parallel" directly injects the strength relationship depicted by line reactance into the graph structure, making the message propagation strength more consistent with power physics, while enhancing the joint expression ability of "one-hop-two-hop-local" at the same depth.

[0111] (1-4) Robustness and scalability

[0112] To ensure numerical stability, computational scalability, and smooth connection with feasible region constraints under large-scale networks and long-period samples, the invention makes a systematic design from three aspects.

[0113] First, numerical stability. The symmetric normalized kernel ​The spectral radius of the graph operator is constrained in an effective range by the graph structure and the size of the self-loop, which helps stabilize the gradient propagation and convergence; the explicitly reserved unit channel in the parallel channel The residual path is provided, which can alleviate the over-smoothing and gradient attenuation in deep propagation. During training, the main operator of the graph forward is several times of "sparse matrix Dense feature" multiplication The numerical jitter is much smaller than the explicit one based on High-order propagation is performed, so as to maintain the predictability of convergence.

[0114] Secondly, the scalability. In the spatial dimension, for a graph with hundreds to thousands of nodes, The sparse multiplication complexity is approximately proportional to the number of edges; The off-line generation or caching of "sparse multiplication of sparse" is adopted to avoid the densification caused by explicit squaring; if online updating is needed, the dominant associated edges can also be pruned under the sparse threshold. In the time dimension, for a long sequence of T = 96 and above, the learner adopts "time-graph kernel decoupling": the graph kernel only handles spatial correlation, and the time dimension is modeled by TCN and BiLSTM to model the cross-period coupling with shared parameters; the smoothing and climbing regularization is added to the loss to further stabilize the learning of long dependence. In terms of overall throughput, the complexity of graph multiplication and convolution / cyclic modules is linear or nearly linear.

[0115] Thirdly, the training-inference interface with feasible region constraints. During the training period, soft constraints are adopted: convex penalties are introduced for violations of power balance, line limits, output upper and lower limits, and climbing constraints.

[0116] Taking the line limit as an example, by adding a loss term:

[0117]

[0118] Among them,

[0119]

[0120] The mapping from the current prediction ;

[0121] The climbing can be used in the same way:

[0122]

[0123] After the inference post-processing stage, for the samples that still have slight out-of-bound, the least squares projection is adopted to return to the feasible region, while strictly maintaining the previous constraint conditions. In terms of unit consistency, the conversion between p.u. and MW is performed through the system power reference, so as to avoid systematic deviation caused by mixed use.

[0124] (1-5) Reference bus consistency

[0125] DC framework of zero, resulting in phase angle is determined only under constant translation, i.e. the zero point is not unique; The definition of also depends on the reference bus. If the reference used in the data set, training, inference, projection stage is inconsistent, there will be systematic errors of "phase angle zero drift" and "power flow mismatch". For this reason, the present application formulates and implements a unified strategy throughout the whole process.

[0126] First, the reference bus is fixed in the data generation stage , when constructing , and labels, the anchor is unified: , and the generated is consistent with the reference.

[0127] Second, align the reference at the model input / output level: if the model directly outputs , do "re-anchor" before entering any physical mapping or loss:

[0128]

[0129] According to formula (3-5), calculate the injection and power flow; if the model is changed to output injection or output , directly use the pre-calculated mapping under the same reference, without explicit angle anchoring.

[0130] Third, in the evaluation and visualization stage, all angle / power flow-based indicators use the same reference; across data sets and across days, evaluation is strictly prohibited with different references.

[0131] Finally, set up a consistency unit test: check whether it is true; the consistency error between the , calculated by and the , , calculated by is whether it is within the preset threshold.

[0132] The above conventions ensure that the interface of "learned phase angle → physical injection / power flow → optimized projection" is strictly aligned, and maintains stable and reproducible physical consistency under multi-period, multi-scenario batch evaluation.

[0133] (2) Spatio-temporal learner design and output strategy (GCN + TCN + BiLSTM)

[0134] (2-1) Graph encoder

[0135] This part uses the reactance-aware graph kernel constructed in Section (1-3) for spatial feature extraction of each time period, so that the message passing strength between nodes is consistent with the strength of reactance awareness.

[0136] Let the input daily sample tensor be , where denotes the node feature matrix of time period . For any time period , make linear transformation to along the "three propagation paths" in parallel and then perform weighted fusion:

[0137] 1) Local path (no propagation): ;

[0138] 2) One-hop propagation (one-time diffusion of reactance awareness): ;

[0139] 3) Two-hop propagation (two-time diffusion of reactance awareness): .

[0140] where is the symmetric normalized reactance-aware graph kernel defined in Section (1-3), is a learnable channel mapping. The three-path results are weighted and fused through a learnable gating coefficient to obtain the first-layer hidden representation:

[0141]

[0142] To improve numerical stability and convergence, a nonlinearity and normalization are applied after fusion, and a residual is added:

[0143]

[0144] where LN is LayerNorm, for dimension matching.

[0145] In actual implementation, 3 layers of the above structure are stacked, and linear cross-layer projection is added between layers to alleviate over-smoothing and information degradation:

[0146] The first layer has a channel dimension ; the second layer has a channel dimension ; and the third layer maintains a channel dimension . After the fusion of the three channels, ReLU activation, Dropout (p=0.1) and LayerNorm are sequentially connected to enhance non-linear expression, suppress overfitting and maintain numerical stability. The three-layer graph convolution projects the input dimension F to 64, 128 and 128 dimensions respectively, and the final graph encoder output is , which is stacked into a time-series embedding for all time periods.

[0147] This "three-path parallelism + reactive sensing kernel" design essentially captures the three spatial dependencies of "one-hop - two-hop - local" at the same network depth. Compared with single topology diffusion, it can more accurately express the correlation strength distribution dominated by electrical susceptance, while maintaining sparse efficiency and training stability.

[0148] (2-2) Strong Temporal Modeling

[0149] Significant cross-period coupling exists in day-ahead power dispatching (such as ramp rate, consistency, and intraday load evolution), with obvious superposition of short-term and long-term dependencies. To simultaneously cover multi-scale local correlations and long-range dependencies, this invention feeds the graph-encoded time-series embedding into a "TCN + BiLSTM" joint structure and sets a gated residual stabilizer between the two.

[0150] First, Considered to be of length The feature sequence is used to apply a one-dimensional dilated convolutional network (TCN) to each node independently. Let the sequence of a node be denoted as . , No. Layer convolution uses kernel length Expansion coefficient The causal relationship or convolution with "receptive field alignment" padding:

[0151]

[0152] By layering different expansion coefficients (such as...) The TCN output is obtained by covering the time correlation from nearest neighbor to medium span. TCN is computationally efficient and convergently stable in modeling hill climbs and surrounding interactions on short to medium timescales.

[0153] Secondly, Feed the data into a bidirectional LSTM (BiLSTM) to model long-range cross-time dependencies and bidirectional interactions between key time periods. For each time step... The BiLSTM outputs forward and backward hidden states concatenated. After linear compression, To avoid the gradient degradation and overfitting risks caused by concatenated convolution and loop operations, gated residual fusion is introduced:

[0154]

[0155] in It is Sigmoid. It is a learnable scalar or channel vector.

[0156] The stabilizer is in the early training smaller, more robust short-to-medium term patterns of TCN are preserved; as training progresses, adaptive upscaling, gradually enhancing the weight of long-range dependencies. Finally, all nodes are computed in parallel to obtain the time-series output of the joint structure .

[0157] In practice, the TCN part adopts one-dimensional dilated convolution to cover the time correlation from the near neighbor to the medium span. In principle, the receptive field can be flexibly controlled by changing the kernel size and dilation factor. In this invention, two layers of convolution are actually used, with a kernel length of 5 and a dilation factor of 1 and 2, so that the effective receptive field covers about 9-13 time steps (corresponding to a 2-3 hour window), which can accurately depict the medium-term dependence of load changes and ramping constraints in day-ahead scheduling. The BiLSTM part is used to capture long-range inter-period dependencies and bidirectional interactions. In principle, multiple layers can be stacked to increase modeling capacity, but to control computational complexity, this invention uses 1 layer of bidirectional LSTM, with a hidden layer dimension of 128 in each direction, and a linear compression to 128 dimensions after output as the final time series representation.

[0158] This "TCN + BiLSTM" deep strong time series modeling is one of the important innovations of this invention: it realizes fast modeling of short-to-medium time scales and accurate capture of long-range dependencies under the stabilizing effect of gated residuals, thus fully reflecting the load evolution and inter-period ramping constraints in day-ahead scheduling. Compared with traditional single GNN or single RNN, this combination significantly enhances the inter-period consistency and endogenous expression ability of physical constraints without increasing the inference complexity.

[0159] (2-3) Output, standardization and anchoring

[0160] The learning end of the invention takes the period-node embedding tensor output by the graph-time series encoder as input, and directly regresses the standardized values of price and phase angle by node and period through a linear regression head; to achieve this, the output head is a multivariate linear mapping of Linear( ).

[0161] Let the time series representation of a certain period-node be , then the output relationship is written as:

[0162]

[0163] The training and deployment use the same z-score standardization:

[0164] For any real physical quantity (e.g. price or phase angle), its standardization and de-standardization are:

[0165]

[0166] where , are obtained from the training set statistics and fixed throughout the training, validation, testing, and deployment pipeline. At inference, the model outputs , are first de-normalized to obtain the physical-domain LMPs and .

[0167] The phase angles must be anchored to a uniform reference bus at each time period to eliminate the bias caused by the non-unique zero point of the DC model. The same anchoring strategy is applied at both the training and evaluation stages:

[0168] For any time period , the network output is de-normalized to obtain , and then "re-anchored":

[0169]

[0170] to ensure strict consistency with the reference bus convention in Section (1-5). Similarly, the phase angles in the true labels are also anchored in the same way before participating in the loss calculation and evaluation.

[0171] The unified normalization and anchoring ensure the unit, zero point, and interface consistency across the three stages of learning-end output, physical mapping, and projection repair; the anchored will be directly used for the physical mapping and DC safety regularization during training, and also for the construction of the projection initial value in the next section.

[0172] (2-4) Feasible daily schedule from learning output to strict DC-OPF projection

[0173] After the learning-end output of the full-day trajectory , the prices are used for market analysis and evaluation, and the phase angles enter the two-step process of physical mapping and feasibility projection: the first step projects the anchored phase angles to the injection / power flow space, and the second step synthesizes the equivalent output initial value from the injections to solve the constrained quadratic programming (QP) to generate the strictly feasible unit output and line power flow schedule.

[0174] (i) Physical mapping:

[0175] According to the DC mechanism in Section (1-2), the node injections

[0176]

[0177] ​Subsequently, the line flow estimate is quickly obtained using PTDF .

[0178] (ii) Construct the projection initial value:

[0179] Combine the injection and load into the "equivalent unit output" initial value:

[0180]

[0181] This step is physically equivalent to: if the potential is used, the group that is most consistent with the predicted phase angle in the unit output vector that satisfies the power balance is . .

[0182] (iii) Strict DC-OPF projection: multi-period, with ramp

[0183] The projection side aims to minimize the two-norm deviation from the initial value, and solves the quadratic programming of the T=96 time periods in groups at one time under the constraints of power balance, line limits, unit box constraints, and, if necessary, ramp constraints (corresponding to (3-1)-(3-4), (3-7)): ensure cross-period ramp consistency:

[0184]

[0185] In this way, the physically friendly high-quality initial value obtained by the front-end "reactance-aware graph kernel + TCN / BiLSTM strong time series modeling" can be quickly projected into a strictly feasible daily plan, realizing the integration of "high-speed candidate generation + strict DC constraint projection"; the two-end interface fully complies with the reference bus consistency and unit consistency requirements in section (1-5), facilitating stable access and reproduction in engineering systems.

[0186] (3) Mechanism-consistent training target and weight scheduling

[0187] In the design of the front-end learner, if only the conventional supervised regression loss is used for training, the model can often fit the trajectories of node prices and phase angles, but there are significant defects in physical consistency and cross-period constraints. In order to make the results generated by the learner not only approximate the true labels, but also gradually approach the feasible solution space during the training phase, the present invention designs a set of mechanism-consistent multi-objective loss systems, and cooperates with the weight scheduling strategy to dynamically adjust the importance of each sub-target. This avoids the non-convergence caused by too strong physical constraints in the early training stage, and ensures the physical friendliness of the model under cross-period and network security constraints.

[0188] (3-1) Uncertainty-weighted multi-task regression

[0189] First, in the core prediction task, the present application adopts uncertainty-weighted regression (UWL) to simultaneously learn the spatio-temporal trajectories of nodal marginal price and nodal voltage phase angle.

[0190] Specifically, let the prediction output be

[0191]

[0192] where is the node At time period , the price prediction, is the phase angle prediction;

[0193] The true label is

[0194]

[0195] The design of UWL is based on a learnable log-variance parameter , whose loss function is

[0196]

[0197] Here denotes the uncertainty of the prediction task . By automatically adjusting during the training process, the model can dynamically allocate the relative weights of the two tasks. To improve the robustness to abnormal samples and spikes, the present application introduces a smooth L1 term as a supplement in addition to the mean square term.

[0198] (3-2) DC security regularization

[0199] Optimizing the prediction error alone is not enough to ensure the safety of power system operation, so the present application introduces DC mechanism constraints as regularization terms. Specifically, the phase angle predicted by the model will first be converted to injected power via the node admittance matrix

[0200] Then it is mapped to line power flow via the PTDF matrix:

[0201]

[0202] For all lines

[0203] , if the predicted power flow exceeds the limit , a convex penalty is imposed on the part exceeding the limit:

[0204]

[0205] ​The design is equivalent to directly embedding the power flow safety constraint into the loss function during the training stage, thereby guiding the prediction result to gradually approach the feasible region.

[0206] (3-3) Cross-period ramping regularization

[0207] To further reflect the multi-period coupling, the application is based on the equivalent unit active power output derived from the phase angle , and imposes constraints on the output change between adjacent periods. Specifically, for all units :

[0208]

[0209] where is the upper limit of the ramping of the unit .

[0210] This constraint ensures that the learning output not only meets the power and flow conditions at a single time, but also reflects smoothness and executability in the time sequence dimension, avoiding inconsistency or excessive fluctuations across periods.

[0211] (3-4) Weight scheduling strategy

[0212] Considering that directly optimizing multiple physical regularizations with the main task may cause instability at the beginning of training, the application designs a phased weight scheduling strategy:

[0213] 1. In the warm-up stage, only is optimized, allowing the model to quickly converge to a reasonable prediction baseline;

[0214] 2. As the training round increases, gradually increase the weights of and to enhance their influence in a linear or segmented manner;

[0215] 3. In the later stage, the loss function is:

[0216]

[0217] where , gradually increases with the training round.

[0218] Through the above weight scheduling mechanism, the model training process gradually transitions from "learning label mapping" to "learning physically consistent mapping", achieving a balance between convergence efficiency and physical feasibility. The final prediction result can take into account both accuracy and engineering executability.

[0219] (4) Lightweight strict DC-OPF projection and multi-period planning generation

[0220] Although the front-end learner can generate high-quality predictions of prices and phase angles, these outputs are essentially regression results and cannot guarantee to strictly meet the flow balance, line limit, and ramp constraints. Therefore, the present invention introduces a lightweight DC-OPF projection module after the learning end, takes the learning output as the initial value, and maps it to the strictly feasible solution domain through a fast optimization process, thereby realizing the integration of "high-speed candidate generation + strict constraint repair".

[0221] (4-1) Projection problem construction

[0222] The unit active power initial value obtained by the learner prediction As the center, a small-scale quadratic programming (QP) problem is constructed at each time period:

[0223]

[0224] The constraint conditions include:

[0225] 1. Unit output boundary constraint

[0226]

[0227] Ensure that the power of all units is within the physical range;

[0228] 2. Power balance constraint

[0229]

[0230] Ensure that the power generation and load are strictly balanced at each time period;

[0231] 3. Line thermal stability constraint (based on PTDF)

[0232]

[0233] Ensure that the power flow does not exceed the limit;

[0234] 4. Cross-period ramp constraint

[0235]

[0236] Through the above construction, the projection problem of each time period is a convex QP, which can quickly converge and has a clear physical interpretation.

[0237] (4-2) Fast solution and numerical strategy

[0238] In order to meet the near-real-time operation requirements, the present invention adopts multiple acceleration and numerical stabilization strategies in the projection stage:

[0239] 1. Hot start mechanism: use the solution of the adjacent time period as the initial value, thereby significantly reducing the number of iterations.

[0240] 2. Fast corridor identification: if the learning output has already satisfied all constraints within tolerance, directly accept the result and skip the optimization solution to save computational overhead.

[0241] 3. Boundary rebalancing mechanism: in the rare case of severe boundary violation, a rebalancing adjustment based on line flow sensitivity is adopted to bring the flow back within the limit while maintaining power balance and boundary feasibility.

[0242] These strategies ensure that the projection module guarantees strict feasibility while keeping the computation time within milliseconds, meeting the batch solution demand even in the 96-period day-ahead panoramic scenario.

[0243] (4-3) Result aggregation and interface

[0244] After completing the per-period projection, the invention aggregates the unit output and line flow trajectory of all periods to form a complete intra-day plan. The result is consistent with the price and phase angle output by the learner, which can be directly interfaced to the market simulation and dispatch interface. The final overall process can not only achieve fast batch evaluation of a large number of scenarios, but also ensure that the output solution strictly satisfies the physical and safety constraints of DC-OPF.

[0245] In summary, the invention proposes a hybrid end-to-end method of "reactance-aware graph kernel spatio-temporal learner + mechanism-consistent physical regularization + lightweight strict DC-OPF projection", which realizes high-speed evaluation and strict feasible plan generation of day-ahead multi-period clearing. The method uses a graph kernel driven by a susceptance matrix to depict the reactance-aware characteristics, a strong time sequence structure of a graph neural network combined with TCN and BiLSTM to learn the cross-period evolution of price and phase angle, a DC mechanism chain to construct a physical regularization during the training period to obtain a "physically friendly" prediction, and a lightweight quadratic programming to project an executable plan under strict constraints after inference. This framework has a structured understanding and automated processing capability for day-ahead clearing and network constraints, can automatically analyze power balance, line flow limit, unit output, and ramping, directly generate candidate trajectories of price / phase angle / output / flow, and quickly correct them under strict DC constraints without human intervention, significantly reducing modeling and solving complexity; under the synergistic effect of strong time sequence modeling and physical regularization, the executable dispatch interval is gradually approached, ensuring safety constraints while balancing speed and consistency, providing an engineering usable technical path for multi-scenario high-frequency evaluation and market strategy research.

[0246] Overall, the invention embeds a lightweight DC-OPF projection link between the learner output and the dispatch executable solution, bridging the full-link process from prediction to dispatch, balancing speed and reliability, and providing an engineering usable technical path for day-ahead multi-period clearing and large-scale scenario analysis.

[0247] The embodiment of the present application further provides a storage medium for storing a computer program, which is executed to perform at least the method described above.

[0248] The embodiment of the present application further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute the computer program to perform at least the method described above.

[0249] The embodiment of the present application further provides a processor, which executes a computer program to perform at least the method described above.

[0250] The storage medium can be implemented by any type of nonvolatile storage device, or a combination thereof. The nonvolatile storage device can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a ferromagnetic random access memory (FRAM), a flash memory, a magnetic surface storage, an optical disc or a compact disc read-only memory (CD-ROM). The magnetic surface storage can be a disk storage or a tape storage. The storage medium described in the embodiment of the present application is intended to include, but is not limited to, these and any other suitable types of memory.

[0251] In several embodiments provided by the present application, it should be understood that the disclosed system and method can be implemented in other manners. The embodiments described above are merely exemplary, for example, the division of units is only a logical function division, and there can be another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling or direct coupling or communication connection between each component can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0252] The units described as separate components above can or can not be physically separate, and the components shown as units can or can not be physical units, that is, can be located in one place or distributed on multiple network units; part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0253] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be realized in the form of hardware or in the form of hardware plus software functional unit.

[0254] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware, and the foregoing program can be stored in a computer readable storage medium, and the program executes the steps including the above-mentioned method embodiments when executed; and the foregoing storage medium includes: mobile storage device, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and various storage program codes.

[0255] Alternatively, the integrated unit of the present application, if implemented in the form of a software function module and sold or used as an independent product, can also be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of the present application. The foregoing storage medium includes: mobile storage device, ROM, RAM, magnetic disk or optical disk and various storage program codes.

[0256] The methods disclosed in the several method embodiments of the present application can be combined arbitrarily without conflict to obtain new method embodiments.

[0257] The features disclosed in the several product embodiments of the present application can be combined arbitrarily without conflict to obtain new product embodiments.

[0258] The features disclosed in the several method or device embodiments of the present application can be combined arbitrarily without conflict to obtain new method embodiments or device embodiments.

[0259] The above is further detailed description of the present application in combination with specific preferred embodiments, and cannot be deemed as limitation of the specific implementation of the present application to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of equivalent substitutions or obvious variations can be made, and the performance or use is the same, which should be deemed as falling within the protection scope of the present application.

Claims

1. A reactive power aware kernel driven spatio-temporal learning and light-weighted DC optimal power flow projection scheduling method, characterized in that, The method comprises the following steps: S1, constructing an electric reactance perception graph kernel based on admittance matrix, for extracting electric reactance perception spatial dependence between nodes of power grid and organizing multi-period node feature tensor as input; In step S1, the electric reactance perception weight matrix is constructed based on the absolute value of the non-diagonal elements of the admittance matrix; after adding a self-loop to the weight matrix, a degree matrix is calculated and symmetric normalization processing is performed to obtain an electric reactance perception graph kernel with stable numerical value; a multi-order propagation matrix of the graph kernel is calculated in advance to expand the spatial receptive field, and the output results of local non-propagation path, single-hop propagation path and multi-hop propagation path are combined; S2, using a space-time learner to perform space-time joint modeling on the node features, the space-time learner comprising a graph convolution encoder and a time series encoder, for learning the cross-period evolution law of node marginal price and voltage phase angle; the graph convolution encoder of the space-time learner in step S2 comprises: performing linear transformation on the node feature matrix of each period along the local path, the single-hop propagation path and the multi-hop propagation path respectively; the outputs of the three paths are weighted and fused through learnable gating coefficients to obtain hidden representation; after the fusion, nonlinear activation function, random inactivation and layer normalization processing are applied in sequence, and residual connection is introduced to maintain gradient flow; S3, introducing mechanism-consistent physical regularization terms in the training process, including power flow safety regularization and ramp consistency regularization, and gradually enhancing the influence of physical constraints through weight scheduling strategy; S4, converting the phase angle trajectory output by the learner into equivalent unit initial output through physical mapping, and constructing a lightweight quadratic programming projection model with the minimum deviation initial value as the target, to generate a feasible day-ahead scheduling plan under the conditions of satisfying power balance, line power flow limit, unit output boundary and cross-period ramp constraint.

2. The reactance-aware-kernel-driven spatiotemporal learning and light-weighted DC optimal power flow projection dispatching method according to claim 1, wherein, In step S1, the diagonal elements of the electric reactance perception weight matrix are set to zero. A learnable weighted multi-channel parallel fusion mechanism is used to combine the output results of local non-propagation path, single-hop propagation path and multi-hop propagation path.

3. The reactance-aware-kernel-driven spatiotemporal learning and light-weighted DC optimal power flow projection dispatching method according to claim 1, wherein, The graph convolution encoder of the space-time learner in step S2 further comprises: The feature dimension is gradually improved through a multi-layer stacking structure, and cross-layer projection is added between layers to alleviate information degradation.

4. The reactance-aware-kernel-driven spatiotemporal learning and light-weighted DC optimal power flow projection dispatching method of claim 1, wherein, The time series encoder in step S2 adopts a TCN and BiLSTM joint structure, and comprises: The time series convolution network is used to perform a cavity convolution operation on the time series sequence output by the graph encoding, and different expansion coefficient convolution layers are stacked to cover multi-scale time receptive fields; The output of the time series convolution network is sent to the bidirectional long short-term memory network to capture the long-term cross-period dependence relationship of the forward and backward directions; The outputs of the time series convolution network and the bidirectional long short-term memory network are dynamically weighted and fused through a learnable gating mechanism, and the contribution proportion of medium and long-term features and long-term features is adaptively adjusted during the training process.

5. The reactance-aware-kernel-driven spatiotemporal learning and light-weighted DC optimal power flow projection dispatching method of claim 1, wherein, The mechanism-consistent physical regularization terms in step S3 comprise: The predicted voltage phase angle is linearly mapped to node injection power through the admittance matrix, and then mapped to line power flow through the power transmission distribution factor matrix; A distance-based convex function penalty is applied to the power flow component exceeding the line thermal stability limit to guide the predicted result to approach the feasible region. Based on the predicted phase angle, the equivalent unit output is derived, and a regular penalty is imposed on the part of the output change between adjacent time periods that exceeds the ramp rate limit; A phased weight scheduling strategy is adopted, with the prediction task loss as the main focus in the early training stage, and the weight coefficient of the physical regularization term gradually increased linearly with the increase of training rounds.

6. The reactance-aware-kernel-driven spatiotemporal learning and light-weighted DC optimal power flow projection dispatching method of claim 1, wherein, The lightweight quadratic programming projection model in step S4 includes: The voltage phase angle predicted by the learner is converted into node injection power through the admittance matrix, and combined with the load prediction to synthesize the initial value of the equivalent unit output; A multi-time period joint optimization problem is constructed with the objective function being the minimum two-norm deviation between the actual unit output and the initial value; The constraint conditions include: system power balance constraint, upper and lower limit constraint of unit output, line flow limit constraint based on power transmission distribution factor, and unit ramp rate constraint between adjacent time periods; A hot start mechanism based on the solution of adjacent time periods is adopted to accelerate the solving process, and the initial value that meets the constraint tolerance is directly adopted, and the fast quadratic programming solver is started for projection correction in the case of out-of-bound.

7. The reactance-aware-kernel-driven spatiotemporal learning and light-weighted DC optimal power flow projection dispatching method according to claim 6, wherein, The fast solving mechanism of the projection model includes: Constraint satisfaction pre-check is performed before projection, if the initial value of the equivalent unit output of a certain time period is within the tolerance range of all constraints, the optimization solving of this time period is skipped; For the time period that needs to be solved, the optimization solution of the previous adjacent time period is used as the hot start initial value to reduce the number of iterations; In the case of extreme out-of-bound, a rebalancing adjustment algorithm based on line flow sensitivity is started to pull the power flow back within the limit value while maintaining power balance.

8. The reactance-aware-kernel-driven spatiotemporal learning and light-weighted DC optimal power flow projection dispatching method of claim 1, wherein, The method further includes: The reference bus is uniformly determined in the data preprocessing stage, and the consistency of the reference bus is maintained throughout the whole process; The voltage phase angle output by the learner is re-anchored before performing physical mapping, and the phase angle value of the reference bus is subtracted to eliminate the zero-point uncertainty; The same unit system and power reference are used in the training, inference and projection stages to ensure the system consistency of physical quantity calculation.

9. A computer readable storage medium storing a computer program, characterized in that, The computer program, when executed by the processor, implements the reactance-aware graph kernel driven spatiotemporal learning and lightweight direct current optimal power flow projection scheduling method according to any one of claims 1 to 8.

10. A computer program product comprising a computer program, characterized in that, The computer program, when executed by the processor, implements the reactance-aware graph kernel driven spatiotemporal learning and lightweight direct current optimal power flow projection scheduling method according to any one of claims 1 to 8.

Images