Reactance perception graph kernel-driven space-time learning and lightweight DC optimal power flow projection scheduling method

Abstract

A space-time learning and lightweight direct-current optimal power flow projection scheduling method driven by a reactance perception graph kernel comprises the steps that the reactance perception graph kernel based on a susceptance matrix is constructed, a power grid node reactance perception spatial dependency relationship is extracted, and multi-period node feature tensors are organized; a space-time learning device comprising a graph convolution encoder and a time sequence encoder is adopted to carry out space-time joint modeling on the node features, and the cross-time evolution law of the node marginal electricity price and the voltage phase angle is learned; mechanism consistent physical regular terms of load flow safety and climbing consistency are introduced in training, and physical constraint influences are gradually enhanced by means of weight scheduling; and converting the phase angle trajectory into an equivalent unit output initial value through physical mapping, constructing a lightweight quadratic programming projection model with the minimum deviation initial value, and generating a feasible day-ahead scheduling plan under the constraint of meeting power balance and the like. According to the method, the whole link from prediction to scheduling is opened, the speed and the reliability are both considered, and an engineering available path is provided for day-ahead multi-period clearing and large-scale scene analysis.

Description

Technical Field

[0001] This invention relates to machine learning and power system optimization computing techniques, and in particular to a spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by reactance sensing graph kernel. Background Technology

[0002] Against the backdrop of high-proportion renewable energy integration, deepening electricity market mechanisms, and the simultaneous need for large-scale scenario simulations, day-ahead clearing requires determining both unit output plans and locational marginal prices (LMPs) on a grid model with cross-time constraints, while satisfying constraints such as power balance, unit output boundaries, line current limits, and ramping. This task demands near-real-time computational response while ensuring the physical feasibility and cross-time consistency of the solution, becoming a key bottleneck in integrated dispatch-market simulation and strategy evaluation.

[0003] Engineering practice has long followed an optimization modeling approach based on the Power Transfer Distribution Factor (PTDF) and Injection Shift Factor (ISF): PTDF / ISF linearizes the sensitivity of line power flow to node injection, transforming the Direct Current Optimal Power Flow (DC-OPF) problem into a large-scale linear / quadratic programming problem; then, it relies on industrial-grade commercial solvers or open-source quadratic / linear programming (QP / LP) solvers to solve the problem on a time-period or block-period basis. The advantage of this approach lies in its mature feasibility and interpretability, clear constraint expression, and rigorous verification. However, when introducing cross-time-period constraint coupling, rolling solutions to numerous scenarios, and network scales of hundreds to thousands of nodes, the solution frequency increases exponentially with the problem size, leading to a significant increase in computational load and latency pressure, making it difficult to meet the requirements of "near real-time + high-frequency" batch evaluation, market sensitivity analysis, and strategy simulation.

[0004] To alleviate latency bottlenecks, various data-driven learning alternatives have emerged in recent years, mainly focusing on sub-problems such as single-period DC approximation power flow, LMP estimation, feasibility judgment, or constraint softening. They often use graph neural networks (GNNs) to characterize network topology relationships or employ shallow time series / statistical methods to handle short-term correlations, in order to shorten evaluation time and assist in operation and maintenance decisions.

[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: 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.

[0008] 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.

[0009] The lack of a robust and feasible lightweight projector is a significant issue: excessive post-processing optimization can negate inference speed gains, while heuristic patching struggles to guarantee strict compliance with power balance, unit boundaries, and line current limits. A lightweight QP projector with hot-start capability, rapid acceptance testing, and failover mechanisms is needed to generate executable plans with minimal deviation under strict DC constraints and maintain robust throughput in batch scenarios.

[0010] True multi-period coupled day-ahead learning: Independent learning for each period ignores the sloping effect and cross-period consistency, making it difficult to directly generate a daily plan usable in engineering. It is necessary to explicitly characterize the sloping effect and smoothness within the model through a time-series dependency modeling module and cross-period physical regularization, making the initial values ​​at the front end naturally "more feasible" and reducing the burden of back-end projection.

[0011] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0012] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology and provide a spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by reactance sensing graph kernel.

[0013] To achieve the above objectives, the present invention adopts the following technical solution: A spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by reactance sensing graph kernel includes the following steps: 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; S2. A spatiotemporal learner is used 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 law of node marginal electricity price and voltage phase angle. S3. During the training process, introduce physical regularization terms with consistent mechanisms, including power flow safety regularization and ramp consistency regularization, and gradually enhance the influence of physical constraints through weight scheduling strategy. S4. Convert the phase angle trajectory output by the learner into the initial value of the equivalent unit output through physical mapping, and construct a lightweight quadratic programming projection model with the goal of minimizing the deviation from the initial value. Under the conditions of satisfying power balance, line power flow limit, unit output boundary and cross-time ramping constraint, generate a feasible day-ahead scheduling plan.

[0014] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned reaction sensing kernel-driven spatiotemporal learning and lightweight DC optimal power flow projection scheduling method.

[0015] A computer program product includes a computer program that, when executed by a processor, implements the spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by the reactance sensing graph kernel.

[0016] The present invention has the following beneficial effects: This invention proposes a spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by a reactance-sensing graph kernel. It constructs a collaborative scheme of "a spatiotemporal learner based on the reactance-sensing graph kernel + a mechanistically consistent physical regularization + a lightweight, rigorous DC-OPF projection," which is a hybrid end-to-end method for high-speed evaluation and rigorously feasible plan generation for day-ahead multi-period clearing. By embedding a lightweight DC-OPF projection link between the learner output and the executable solution for scheduling, this invention streamlines the entire process from prediction to scheduling, balancing speed and reliability, and providing an engineering-ready technical path for day-ahead multi-period clearing and large-scale scenario analysis.

[0017] Specifically, the method uses a graph kernel driven by the susceptance matrix to characterize reactance sensing characteristics, and a strong temporal structure combining graph neural networks with TCN and BiLSTM to learn the cross-period evolution of price and phase angle. During the training period, physical regularization is constructed through DC mechanism chains to obtain "physically friendly" predictions. After inference, lightweight quadratic programming is used to project the predictions into executable plans under strict constraints. This framework has a structured understanding and automated processing capability for day-ahead clearing and network constraints. It can automatically resolve various constraints such as power balance, line current limit, unit output, and ramping, directly generate candidate trajectories for price / phase angle / output / power flow, and quickly correct them under strict DC constraints without manual intervention, significantly reducing the complexity of modeling and solving. With the synergistic effect of strong temporal modeling and physical regularization, it gradually approaches the executable scheduling range, balancing speed and consistency while ensuring safety constraints. This provides an engineering-ready technical path for multi-scenario high-frequency assessment and market strategy judgment.

[0018] The hybrid end-to-end method proposed in this invention, which combines a "spatiotemporal learner based on a reactance-sensing graph kernel, a mechanistically consistent physical regularization, and a lightweight, rigorous DC-OPF projection," achieves significant breakthroughs in technical performance. First, this invention introduces a graph kernel based on a reactance-sensing matrix, extending the traditional adjacency relationship based solely on topology to a propagation mechanism consistent with strong correlations with reactance and susceptance, ensuring that message transmission between nodes more closely reflects the actual power distribution patterns of the power grid. This improvement significantly enhances the physical consistency of the learning model's mapping from "state to phase angle / power flow / LMP," enabling it to maintain strong generalization capabilities even in medium-to-large-scale systems and unseen disturbance scenarios. Furthermore, this invention employs a strong temporal modeling structure of TCN+BiLSTM at the learning end, explicitly characterizing load evolution and cross-period coupling characteristics at the day-ahead scale. This results in outstanding performance in capturing both short-term ramp-ups and long-term smoothness, making the initial values ​​generated by the model in multi-period scenarios naturally more consistent with physical laws. Building upon this foundation, the present invention embeds physical regularization along the "phase angle—injection—current flow" mechanism chain during training, and tightly integrates it with cross-time-segment ramping constraints through a weighted incremental mechanism. This effectively suppresses current flow exceeding limits and ramping abrupt changes 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 subsequent rigorous projection stages.

[0019] In the post-processing stage, this invention no longer relies on heuristics or heavyweight optimization, but instead constructs a lightweight, rigorous DC-OPF projection module. This module uses the learned output as the initial value for hot-start, with the objective function being simply minimum deviation. Constraints cover power balance, line current limits, and unit boundaries, and incorporate rapid acceptance and fail-safe mechanisms. In this way, rigorously feasible all-day scheduling plans can be generated within millisecond timescales, achieving the integration of "high-speed candidate generation + rigorously constrained projection." Ultimately, this invention not only solves the bottlenecks of existing numerical optimization methods in terms of scale and latency, but also overcomes the problem of insufficient physical consistency in pure learning methods, achieving high-speed, rigorously feasible, and cross-time-period consistent low-cost scheduling. More importantly, this method balances engineering-side verifiability and maintainability, providing a novel technical path for batch scenario assessment and online applications of power systems that combines accuracy, interpretability, and deployability.

[0020] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0021] Figure 1 Flowchart of the hybrid end-to-end scheduling method based on reactance sensing graph kernel, spatiotemporal learning and strict DC-OPF projection in this embodiment of the invention. Detailed Implementation

[0022] 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.

[0023] 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.

[0024] 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.

[0025] 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: 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.

[0026] 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.

[0027] 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.

[0028] In some embodiments, the graph convolutional encoder of the spatiotemporal learner in step S2 includes: performing linear transformations on the node feature matrices of each time period along the local path, single-hop propagation path, and multi-hop propagation path, respectively; weighting and fusing the outputs of the three paths using learnable gating coefficients to obtain a hidden representation; sequentially applying nonlinear activation functions, random deactivation, and layer normalization after fusion, and introducing residual connections to maintain gradient flow; gradually increasing the feature dimension through a multi-layer stacked structure, and adding cross-layer projection between layers to alleviate information degradation.

[0029] In some embodiments, step S2 employs a joint TCN and BiLSTM structure for the temporal encoder, and includes: performing dilated convolution operations on the temporal sequence output by the graph encoding using a temporal convolutional network, covering a multi-scale temporal receptive field by stacking convolutional layers with different dilation coefficients; feeding the output of the temporal convolutional network into a bidirectional long short-term memory network to capture long-term cross-temporal dependencies between forward and backward directions; and dynamically weighting and fusing the outputs of the temporal convolutional network and the bidirectional long short-term memory network through a learnable gating mechanism, adaptively adjusting the contribution ratio of short-term features and long-term features during training.

[0030] Step S3: Introduce physical regularization terms with consistent mechanisms during the training process, including power flow safety regularization and ramp consistency regularization, and gradually enhance the influence of physical constraints through weight scheduling strategy.

[0031] In some embodiments, the physical regularization term in step S3 with consistent mechanism includes: linearly mapping the predicted voltage phase angle to node injected power through the susceptance matrix, and then mapping it to line power flow through the 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 prediction result closer to the feasible region; deriving the 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; and adopting a phased weight scheduling strategy, with the prediction task loss as the main focus in the early stage of training, and gradually linearly increasing the weight coefficient of the physical regularization term as the training rounds increase.

[0032] Step S4: Convert the phase angle trajectory output by the learner into the initial value of the equivalent unit output through physical mapping, and construct a lightweight quadratic programming projection model with the goal of minimizing the deviation from the initial value. Under the conditions of satisfying power balance, line power flow limit, unit output boundary and cross-time ramping constraint, generate a feasible day-ahead scheduling plan.

[0033] In some embodiments, the lightweight quadratic programming projection model in step S4 includes: converting the voltage phase angle predicted by the learner into nodal injected power through the susceptance matrix, and then combining it with the load prediction to synthesize the initial value of the equivalent unit output; constructing a multi-time period joint optimization problem with the objective function of minimizing the L2 deviation between the actual unit output and the initial value; the constraints include: system power balance constraints, upper and lower limit constraints of unit output, line power flow limit constraints based on power transmission distribution factor, and unit ramp rate constraints between adjacent time periods; and using a hot-start mechanism based on solutions of adjacent time periods to accelerate the solution process, directly adopting the initial value that meets the constraint tolerance, and starting a fast quadratic programming solver for projection correction in the case of out-of-bounds situations.

[0034] In some embodiments, the fast solution mechanism of the projection model includes: performing a constraint satisfaction pre-check before projection; if the initial value of the equivalent unit output for a certain period is within the tolerance range of all constraints, then skipping the optimization solution for that period; for the period that needs to be solved, using the optimization solution of the previous adjacent period as the initial value for hot start to reduce the number of iterations; in extreme out-of-bounds situations, starting a rebalancing adjustment algorithm based on line flow sensitivity to pull the power flow back within the limit while maintaining power balance.

[0035] In some embodiments, the method further includes: uniformly determining a reference bus during the data preprocessing stage and maintaining the consistency of the reference bus throughout the entire 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 zero-point uncertainty; and using the same unit system and power reference in each stage of training, inference, and projection to ensure the system consistency of physical quantity calculation.

[0036] In this embodiment of the invention, the front-end learner uses The derived symmetric normalized graph kernel and MixHop multi-stage propagation explicitly encode reactance sensing characteristics, and superimpose a Temporal Convolutional Network (TCN) and a Bidirectional Long Short-Term Memory (BiLSTM) network to capture cross-period dependencies and ramping behavior. During training, physical regularization is applied along the DC mechanism chain of "phase angle → injection → power flow", and progressive weight scheduling is used to improve stability and physical friendliness. The inference backend uses a lightweight QP projection with hot start to make minimum corrections under power balance, unit boundary and line current limit, and quickly obtain a strictly feasible multi-period plan, thereby achieving synergistic optimization between throughput speed and engineering feasibility.

[0037] The proposed method for spatiotemporal learning and lightweight DC optimal power flow projection scheduling driven by reactance sensing graph kernel has the following main technical advantages: It constructs a hybrid end-to-end architecture of "reactance sensing graph kernel spatiotemporal learner + mechanistic consistent physical regularization + lightweight strict DC-OPF projection," which overcomes traditional technical bottlenecks while also considering engineering practicality. First, by using a reactance sensing graph kernel driven by the susceptance matrix to replace pure topological adjacency modeling, node message passing more closely matches the real power distribution of the power grid, significantly improving the physical consistency of the "state → phase angle / power flow / LMP" mapping and its generalization ability under medium-to-large-scale systems and unknown disturbance scenarios. Second, it employs joint strong temporal modeling of TCN and BiLSTM to accurately capture the short-term ramp-up and long-term smoothing characteristics of day-ahead scheduling, ensuring that the initial values ​​generated by the learner naturally conform to physical laws. Third, during the training period, the method follows the "phase angle / power flow / LMP" mapping. The "angle-injection-flow" mechanism chain is subject to physical regularization and phased weighted scheduling. This suppresses flow overruns and ramp-up abrupt changes without affecting the convergence of the main task, significantly reducing the correction burden of subsequent projection stages. Fourth, the lightweight DC-OPF projection module uses the learned output as the initial value for hot start. Combined with rapid acceptance and boundary rebalancing mechanisms, it achieves millisecond-level solutions, ensuring that the generated daily scheduling plan strictly meets constraints such as power balance and line current limits. Ultimately, this architecture solves the latency problem of traditional numerical optimization in large-scale, multi-period scenarios and overcomes the deficiency of insufficient physical consistency in pure learning methods. It balances high speed, strict feasibility, and cross-period consistency, and has engineering-side verifiability and maintainability characteristics. It provides an accurate, interpretable, and easily deployable technical path for day-ahead multi-period clearing, batch scenario assessment, and online application of power systems.

[0038] The following further describes specific embodiments of the present invention and examples of its algorithm implementation.

[0039] This invention proposes a spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by reactance sensing graph kernel. It designs a hybrid end-to-end method consisting of "spatiotemporal learner of reactance sensing graph kernel + physical regularization with consistent mechanism + lightweight strict DC-OPF projection".

[0040] Specifically, the method of the present invention includes the following steps: (1) DC-OPF modeling and reactance sensing kernel (1-1) Variables and Data Organization This invention addresses pre-holiday multi-period clearing and large-scale scenario assessment (e.g., a 15-minute period at T=96). The period index set is denoted as... The node (bus) index set is The line index set is In each time period ,node Above, the core state and decision variables include the unit's active power output. Phase angle with node .

[0041] The entire process of this invention is based on radian phase angles and power in MW units: phase angle A uniform radian unit is used, and the units are kept consistent in all stages involving susceptance matrix and power flow calculations. If the input data source is provided in degrees, it is converted to radians during the data preprocessing stage. Involving In the calculation, a system reference power is introduced. (Uniformly set to 100 MVA) Mapping from per-unit to MW units, Furthermore, the same power baseline constant is maintained during the sample generation, training, and projection stages to ensure strict consistency with DC-OPF modeling.

[0042] Within the DC-OPF modeling framework, the basic physical quantities required for input all originate from system static parameters or day-ahead forecasting modules: active power load forecasting at the node side. The data is generated from the current forecast process (historical alignment, anomaly removal, and time-segmented modeling) and serves as the exogenous quantity within the time period; the network structure is summarized into an susceptance matrix based on line parameters and topology. Essentially, it is a Laplace-type admittance matrix under the DC approximation, reflecting the reactance sensing relationship between nodes; power flow transfer factor During the modeling stage, it can be achieved through Derived from line parameters, it is used to quickly represent the linear impact of node injection on line power flow; line thermal limit. Derived from system static parameters, given by operation and maintenance parameters, limiting the maximum power flow that each branch can withstand; upper bound of output at the generator set or node level. Derived from unit operation constraints, when a unit is shut down, this can be represented by setting the upper bound to zero.

[0043] To ensure complete consistency of the physical interfaces of "data terminal - learning terminal - projection terminal", this invention unifies the node numbering and reference bus convention in the data organization and processing stages to avoid systematic deviations caused by inconsistencies in indexes or references.

[0044] Samples organized into tensors by day: Input ,in = 6, and its six-dimensional features are as follows: The difference between the upper limit of the available output and the load at the node. , The difference between the lower limit of the available output at the node and the load. , Linear cost coefficient , Secondary cost coefficient , Node active load forecasting , The indicator mask genmask for the bus where the unit is located.

[0045] 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.

[0046] (1-2) Core Relationships and Constraints First, a standard DC-OPF mathematical expression for each time period is given to clarify the objective and feasible region.

[0047] For any time period Let the active power output vector of the node be... Load vector Line power flow vector .

[0048] DC-OPF aims to minimize the total system power generation cost.

[0049] in For the assembly of generator units, , , This is the cost coefficient.

[0050] The constraints include: (i) Active power balance at nodes

[0051] in To connect to the node A collection of 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 .

[0052] (ii) Upper and lower limits of generator active power output

[0053] (iii) Branch power flow capacity constraints

[0054] Or use Rewritten as a vector inequality: ,in .

[0055] (iv) Reference bus constraint

[0056] To eliminate the non-uniqueness of the phase angle solution, the reference busbar is used. The system's balanced bus is determined once during the data generation phase (preferably selected during the data generation phase; if no unified standard is specified, it is automatically selected based on historical phase angle statistics). / The definition of / PTDF remains consistent throughout all training-inference-projection stages. The phase angle of the model output is uniformly re-anchored before entering any physical mapping or loss. .

[0057] Therefore, the equivalent vector-matrix DC-OPF form of this mathematical model is as follows:

[0058] in: (3-1) Ensure the active power balance of the system; (3-2) and (3-3) describe the line power flow calculation and thermal limit constraints based on PTDF; (3-4) are the upper and lower limits of the unit's output; (3-5a) and (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; (3-6) Zero-point uncertainty is eliminated by fixing the phase angle of the reference bus.

[0059] In addition, when coupling across time periods, a ramp constraint needs to be added to reflect the smoothness of force changes:

[0060] Based on the aforementioned standard model, this invention adopts a hybrid process of "generating physically friendly initial values ​​at the learning end + rigorous feasibility of lightweight quadratic programming (QP) projection": the learning end directly predicts... The spatiotemporal trajectory; the phase angle is mapped to the node injection using equation (3-5a), and then combined with the load to form:

[0061] As the "center point" of the projection. During the projection phase, price or phase angle variables are no longer introduced; instead, the "minimum deviation from the initial value" criterion is used for the entire day. Unit output during a given time period Joint optimization is performed, with the learning end as the central point. Using this as a baseline, minimizing the total deviation and explicitly incorporating it into the ramp constraint will necessarily guarantee consistency across time periods:

[0062]

[0063] The solution Strictly meet power balance, unit upper and lower limits, line current limit, and cross-time ramping requirements as an executable daily plan.

[0064] In engineering implementation, if the center point in certain time periods has satisfied all constraints within the tolerance, it enters the fast acceptance channel to skip numerical solution; in other time periods, it is solved by the QP solution process that supports hot start, and in rare boundary cases, it triggers flow-aware rebalancing as a fallback to ensure that the output is always feasible and close to the structural information of the learner.

[0065] (1-3) Construction of reactance sensing map kernel Using only the topological adjacency matrix Graph modeling using (0 / 1 connections) ignores the correlation strength information caused by differences in reactance between different branches; however, in a DC framework, the near-end power-phase relationship is determined by... The dominance, injection, and power flow sensitivity to phase angle disturbances are closely related to the reactance magnitude. Constructing a "reactance-aware graph" using susceptance as edge weight more closely reflects the actual message transmission intensity and is key to moving from a pure topological graph to a physically-aware graph.

[0066] First, define the reactance sensing weight matrix. Off-diagonal weights:

[0067] Its physical meaning is: Reflecting the busbar and Interaction strength between lines, line reactance The smaller, the better The larger the value, the stronger the coupling of the power disturbance.

[0068] Then, a unit self-loop is added and symmetric normalization is performed to obtain a numerically stable graph kernel.

[0069] make 、 ,definition:

[0070] This normalization balances the scale of nodes with different degrees to the same spectral domain, which is beneficial for stabilizing forward propagation and gradient transfer, and suppressing degree bias.

[0071] To cover more distant dependencies without increasing the number of network layers, the two-hop propagation kernel is pre-computed. The physical intuition is that the series reactance of "adjacent-neighbor" corresponds to the effect of the second jump, which is close to the indirect effect of multi-bus injection on the power flow at the far end in PTDF.

[0072] Multi-channel parallel fusion is used in message passing, and the unit channel is used in combination. Single jump With double jump A single forward propagation can be written as:

[0073]

[0074] in For learnable fusion coefficients, This is a linear mapping for the corresponding channel.

[0075] Compared to only containing The above-mentioned approach of "reactor perception graph kernel + multi-scale parallelism" directly injects the strength relationship depicted by the line reactance into the graph structure, making the message propagation strength more in line with power physics, while enhancing the joint expression capability of "one hop - two hops - local" at the same depth.

[0076] (1-4) Robustness and scalability To ensure numerical stability, computational scalability, and smooth integration with feasible region constraints under large-scale networks and long-term samples, this invention has made system designs in three aspects.

[0077] First, numerical stability. Symmetric normalized kernel. Constraining the spectral radius of graph operators within an effective range based on graph structure and self-loop size helps stabilize gradient propagation and convergence; the unit channel is explicitly reserved in parallel channels. This is equivalent to providing a residual path, which can alleviate oversmoothing and gradient decay in deep propagation. During training, the master operator of the graph forward pass is a sparse matrix of several orders. Multiplication of "dense features" Its numerical jitter is much smaller than that of explicit data based on... Perform high-order propagation to maintain the predictability of convergence.

[0078] Secondly, scalability. In terms of spatial dimension, for graphs with hundreds to thousands of nodes, The complexity of sparse multiplication is approximately proportional to the number of edges; Offline generation or caching using a "sparse multiplication of sparse" approach avoids the densification caused by explicit squaring. For online updates, dominant edges can be retained by pruning under a sparsity threshold. In the temporal dimension, for long sequences T = 96 and above, the learner employs a "temporal-graph kernel decoupling": the graph kernel only handles spatial dependencies, while the temporal dimension is modeled by TCN and BiLSTM with shared parameters, coupling across time periods. Smoothing and ramp-up regularization are added to the loss function to further stabilize the learning of long dependencies. In terms of overall throughput, the complexity of both graph multiplication and convolution / recurrent modules is linear or near-linear.

[0079] Third, the training-inference interface with feasible region constraints. Soft constraints are adopted during the training period: convex penalties are introduced for violations of power balance, line limits, output upper and lower limits, and ramp constraints.

[0080] Taking line limits as an example, by adding a loss term:

[0081] in,

[0082] From current forecasts Mapping; The same principle applies to climbing hills:

[0083] In the post-inference processing stage, for samples that still slightly exceed the limits, least squares projection back to the feasible region is used, while strictly maintaining the previous constraints. Regarding unit consistency, a unified conversion between pu and MW is performed using a system power benchmark to avoid systemic biases caused by mixing different units.

[0084] (1-5) Reference bus consistency DC framework The sum of the rows is zero, resulting in a phase angle of zero. It is determined only by the translation of the additive constant, that is, the zero point is not unique; The definition also depends on the reference bus. If the references used in the dataset, training, inference, and projection stages are inconsistent, systematic errors such as "phase angle zero-point drift" and "power flow mismatch" will occur. To address this, this invention formulates and implements a unified strategy throughout the entire process.

[0085] First, a fixed reference bus is used during the data generation phase. In construction , Unified anchoring with tags: And generate a reference consistent with that reference. .

[0086] Secondly, model input / output alignment reference: if the model directly outputs... Before entering any physical mapping or loss, perform "re-anchoring":

[0087] Then calculate the injection and power flow according to equation (3-5); if the model is changed to output injection... Or exert effort Then, the pre-calculated value under the same reference is used directly. Mapping without explicit angle anchoring.

[0088] Furthermore, during the evaluation and visualization phases, all angle / trend-based indicators should use the same reference; different references must not be mixed for cross-dataset and cross-day evaluations. .

[0089] Finally, set up consistency unit tests: check Whether it is valid; by , Recalculation with by , , Recalculation Whether the consistency error between them is within a preset threshold.

[0090] The above conventions ensure that the interfaces of "learning phase angle → physical injection / power flow → optimization projection" are strictly aligned, maintaining stable and reproducible physical consistency under batch evaluation in multiple time periods and scenarios.

[0091] (2) Spatiotemporal learner design and output strategy (GCN + TCN + BiLSTM) (2-1) Graphic encoder This section uses the reactance sensing kernel constructed in Section (1-3) to extract spatial features for each time period, so that the message transmission strength between nodes is consistent with the reactance sensing strength.

[0092] Let the input daily sample tensor be... ,in Indicates time period The node feature matrix. For any time period Parallel propagation along the "three propagation paths" Perform a linear transformation followed by weighted fusion:

[0093] 1) Local path (not propagated): ; 2) One-hop propagation (reactance sensing first-step diffusion): ; 3) Two-hop propagation (reactance sensing spreads twice): .

[0094] in The symmetric normalized reactance sensing kernel defined in Sections (1-3), This is a learnable channel mapping. The three results are weighted and fused using learnable gating coefficients to obtain the first-layer hidden representation:

[0095]

[0096] To improve numerical stability and convergence, nonlinearity and normalization are incorporated after fusion, and residuals are added:

[0097] Where LN stands for LayerNorm. Perform dimensional matching.

[0098] In the actual implementation, three layers of the above structure are stacked, and linear cross-layer projection is added between layers to alleviate oversmoothing and information degradation: Level 1 channel dimension ; second floor ; The third layer remains After fusing the three channels, ReLU activation, Dropout (p=0.1), and LayerNorm are sequentially applied to enhance nonlinear expression, suppress overfitting, and maintain numerical stability. The three graph convolutional layers project the input dimension F to 64, 128, and 128 dimensions respectively, resulting in the final graph encoder output. Stack all time periods into a time-series embedding. .

[0099] 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.

[0100] (2-2) Strong Temporal Modeling 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.

[0101] 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:

[0102]

[0103] 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.

[0104] 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:

[0105]

[0106] in It is Sigmoid. It is a learnable scalar or channel vector.

[0107] The stabilizer is used in the early stages of training. Smaller, retaining more of the robust short-to-medium-term pattern of TCN; as training progresses, The weights of long-range dependencies are adaptively increased. Finally, parallel computation is performed on all nodes to obtain the time-series output of the joint structure. .

[0108] In practice, the TCN part uses one-dimensional dilated convolutions to cover temporal dependencies ranging from nearest neighbors to medium spans. In principle, the receptive field can be flexibly controlled by changing the kernel size and dilation coefficients. In this invention, two convolutional layers are actually used, with a kernel length of 5 and dilation coefficients of 1 and 2 respectively, resulting in an effective receptive field covering approximately 9–13 time steps (corresponding to a 2–3 hour window), accurately characterizing the short- to medium-term dependencies of load changes and ramp constraints in day-ahead scheduling. The BiLSTM part is used to capture long-range cross-temporal dependencies and bidirectional interactions. In principle, multiple layers could be stacked to increase modeling capabilities, but to control computational complexity, this invention uses a single-layer bidirectional LSTM with 128 hidden dimensions in each direction, and the output is linearly compressed to 128 dimensions as the final temporal representation.

[0109] This "TCN + BiLSTM" deep strong temporal modeling is one of the key innovations of this invention: under the stabilizing effect of gated residuals, it achieves rapid modeling at short to medium time scales and accurate capture of long-range dependencies, thus fully reflecting load evolution and cross-period ramping constraints in day-ahead scheduling. Compared with traditional single GNNs or single RNNs, this combination significantly enhances cross-period consistency and endogenous representation of physical constraints without increasing inference complexity.

[0110] (2-3) Output, Standardization and Anchoring The learning endpoint of this invention uses the time-node embedding tensor output by the graph-time encoder. As input, the standardized values ​​of price and phase angle are directly regressed node-by-node and time-by-time using a linear regression header; in practice, the output header is a Linear( A multivariate linear mapping of ).

[0111] Let the time sequence representation of a certain time period - node be: The output relationship can then be written as:

[0112] Training and deployment use consistent z-score standardization: For any real physical quantity (For example, price or phase angle), its standardization and destandardization are as follows:

[0113] in , The values ​​are obtained from statistics on the training set and remain constant throughout the training, validation, testing, and deployment processes. During inference, the model outputs... , First, the LMP of the physical domain is obtained through denormalization. .

[0114] The phase angle must be anchored to a uniform reference bus in each time period to eliminate the offset caused by the non-uniqueness of the DC model's zeros. This invention implements the same anchoring strategy in both the training and evaluation phases:

[0115] For any time period The network output After destandardization, we get Then perform "re-anchoring":

[0116] Ensure strict consistency with the reference bus convention in Sections (1-5). Similarly, the phase angle in the truth label should be anchored in the same way before being used in loss calculation and evaluation.

[0117] Unified standardization and anchoring ensure consistency in units, zero points, and interfaces across the three stages of learning terminal output, physical mapping, and projection restoration; after anchoring... Will be used directly for " The physical mapping and DC safety regularization during training are also used in the construction of initial projection values ​​in the next section.

[0118] (2-4) Feasible day plan from learning output to rigorous DC-OPF projection Learning platform generates daily trajectory Afterwards, the price is used for market analysis and evaluation, and the phase angle enters a two-step process of physical mapping and feasibility projection: the first step is to anchor the phase angle. Projecting onto the injection / power flow space, the second step uses the initial value of the injected synthetic equivalent output to solve a constrained quadratic programming (QP) problem to generate a strictly feasible unit output and line power flow plan.

[0119] (i) Physical mapping:

[0120] According to the DC mechanism in Section (1-2), the injection is first performed by the anchor phase angle calculation node.

[0121] Then, PTDF was used to quickly obtain the line power flow estimate. .

[0122] (ii) Construct initial projection values: The injection and load are combined to form the initial value of the "equivalent unit output":

[0123] This step is physically equivalent to: if this... Given the potential, among the unit output vectors that satisfy power balance, the one that best matches the predicted phase angle is... .

[0124] (iii) Strict DC-OPF projection: multi-time period, with slope The projection side aims to minimize the L2 deviation from the initial value. Under the constraints of power balance, line limits, unit box constraints, and ramping constraints when necessary (corresponding to (3-1)-(3-4), (3-7)), a quadratic programming approach is used to solve the problem in groups for all 96 time periods of the day (T=96) at once (ensuring consistent ramping across time periods).

[0125] This allows for the rapid projection of the physically-friendly, high-quality initial values ​​obtained from the front-end "reactor sensing kernel + TCN / BiLSTM strong temporal modeling" into rigorously feasible values. The full-day plan achieves the integrated implementation of "high-speed candidate generation + strict DC constraint projection"; the two-end interfaces fully comply with the reference bus consistency and unit consistency conventions in Section (1-5), which facilitates stable access and reproduction in the engineering system.

[0126] (3) Training objectives and weight scheduling with consistent mechanisms In the design of front-end learners, if training is based solely on conventional supervised regression loss, the model can often fit the trajectory of node price and phase angle, but it has significant shortcomings in terms of physical consistency and cross-time constraints. To ensure that the learner's generated results approximate the true labels while gradually approaching the feasible solution space during training, this invention designs a multi-objective loss system with consistent mechanisms, and dynamically adjusts the importance of each sub-objective using a weight scheduling strategy. This avoids non-convergence caused by excessively strong physical constraints in the early stages of training, and also ensures the physical friendliness of the model under cross-time and network security constraints in later stages.

[0127] (3-1) Uncertainty-weighted multi-task regression First, in the core prediction task, this invention uses uncertainty-weighted regression (UWL) to simultaneously learn the spatiotemporal trajectories of the node marginal electricity price and the node voltage phase angle.

[0128] Specifically, let the predicted output be:

[0129] in For nodes During the period Price forecast, For phase angle prediction; The actual tags are:

[0130] UWL's design is based on learnable log-variance parameters. Its loss function is:

[0131] here Indicates prediction task The uncertainty. Through automatic adjustment during training. The model can dynamically allocate the relative weights of the two tasks. To improve robustness to outliers and spikes, this invention introduces a smoothing L1 term in addition to the mean square term.

[0132] (3-2) DC security regularity Optimizing prediction errors alone is insufficient to guarantee the safety of power system operation; therefore, this invention introduces DC mechanism constraints as regularization terms. Specifically, the phase angle predicted by the model... The first step will be through the nodal susceptance matrix. Converted to injected power:

[0133]

[0134] Then, the power flow is mapped to the line using the PTDF matrix:

[0135] For all lines If you predict the trend Exceeding the limit Then, a convex penalty is applied to the portion exceeding the limit:

[0136] This design is equivalent to directly embedding power flow safety constraints into the loss function during the training phase, thereby guiding the prediction results to gradually approach the feasible region.

[0137] (3-3) Cross-time climbing regularity To further demonstrate multi-time-period coupling, this invention derives the equivalent active power output of the unit based on phase angle. And it imposes constraints on output variations between adjacent time periods. Specifically, for all units... :

[0138]

[0139] in For the unit The upper limit of the climbing slope.

[0140] This constraint ensures that the learned output not only meets the power and power flow conditions at a single moment, but also reflects smoothness and executability in the time-series dimension, avoiding inconsistencies or excessive fluctuations across time periods.

[0141] (3-4) Weighted scheduling strategy Considering that directly optimizing multiple physical regularizations simultaneously with the main task may cause instability in the early stages of training, this invention designs a phased weight scheduling strategy: 1. During the warm start phase, only optimization is performed. This enables the model to converge quickly to a reasonable prediction baseline; 2. Gradually increase the intensity as the number of training rounds increases. and The weights are used to increase their influence in a linear or piecewise manner; 3. In the later stages, the loss function is as follows:

[0142] in , The number of training rounds gradually increases.

[0143] Through the aforementioned 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 results can balance accuracy and engineering feasibility.

[0144] (4) Lightweight and rigorous DC-OPF projection and multi-time period plan generation While the front-end learner can generate high-quality predictions of price and phase angle, these outputs are essentially regression results and cannot guarantee strict compliance with power flow balance, line limits, and ramp constraints. Therefore, this invention introduces a lightweight DC-OPF projection module after the learning end, using the learning output as initial values ​​and mapping it to a strictly feasible solution domain through a rapid optimization process, thereby achieving an integrated implementation of "high-speed candidate generation + strict constraint repair".

[0145] (4-1) Construction of the projection problem Initial value of unit active power predicted by the learner Centered on this problem, construct a small-scale quadratic programming (QP) problem for each time period:

[0146] Its constraints include: 1. Unit output boundary constraints

[0147] Ensure that the power output of all units is within the physical limits; 2. Power balance constraints

[0148] Ensure that power generation and load are strictly balanced in every period; 3. Line thermal stability constraints (based on PTDF)

[0149] Ensure that trends do not exceed limits; 4. Cross-time ramping constraints

[0150] With the above construction, the projection problem for each time period is a convex QP, which can converge quickly and has a clear physical interpretation.

[0151] (4-2) Fast solution and numerical strategy To meet near real-time operation requirements, this invention employs multiple acceleration and numerical stabilization strategies during the projection phase: 1. Hot start mechanism: Use the solutions from adjacent time periods as initial values, thereby significantly reducing the number of iterations.

[0152] 2. Fast path discrimination: If the learning output If all constraints have been satisfied within the tolerance, the result is accepted directly and the optimization solution is skipped to save computational costs.

[0153] 3. Boundary rebalancing mechanism: In rare cases of severe predicted out-of-bounds situations, a rebalancing adjustment based on line flow sensitivity is adopted to bring the power flow back within the limits while maintaining power balance and boundary feasibility.

[0154] These strategies ensure that the projection module maintains millisecond-level computation time while ensuring strict feasibility, meeting the batch solution requirements even in day-to-day panoramic scenes with 96 time periods.

[0155] (4-3) Result Aggregation and Interface After completing the time-period projection, this invention aggregates the unit output and line power flow trajectories for all time periods to form a complete intraday plan. This result maintains consistency with the price and phase angle output by the learner and can be directly interfaced with market simulation and scheduling. The final overall process enables rapid batch evaluation of numerous scenarios while ensuring that the output solution strictly meets the physical and safety constraints of DC-OPF.

[0156] In summary, this invention proposes a hybrid end-to-end method combining a "spatiotemporal learner of reactance-sensing graph kernel + mechanistically consistent physical regularization + lightweight strict DC-OPF projection," achieving high-speed evaluation and rigorously feasible plan generation for day-ahead multi-period clearing. The method uses a graph kernel driven by the susceptance matrix to characterize reactance-sensing characteristics, and a graph neural network combining TCN and BiLSTM with a strong temporal structure to learn the cross-period evolution of price and phase angle. During the training phase, physical regularization is constructed through a DC mechanism chain to obtain "physically friendly" predictions. Finally, after inference, lightweight quadratic programming is used to project the results into an executable plan under strict constraints. This framework possesses a structured understanding and automated processing capability for day-ahead clearing and network constraints. It can automatically resolve various constraints such as power balance, line current limit, unit output, and ramping, directly generate candidate trajectories for price / phase angle / output / power flow, and quickly correct them under strict DC constraints without manual intervention, significantly reducing the complexity of modeling and solving. With the synergistic effect of strong time-series modeling and physical regularization, it gradually approaches the executable scheduling range, balancing speed and consistency while ensuring safety constraints. It provides an engineering-ready technical path for multi-scenario high-frequency assessment and market strategy analysis.

[0157] In summary, this invention, by embedding a lightweight DC-OPF projection link between the learner output and the scheduled executable solution, opens up the entire link process from prediction to scheduling, balancing speed and reliability, and provides an engineering-usable technical path for day-ahead multi-period clearing and large-scale scenario analysis.

[0158] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0159] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0160] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0161] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc or CD-ROM; magnetic surface memory can be disk storage or magnetic tape storage. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0162] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0163] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0164] In addition, in the various embodiments of the present invention, each functional unit 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 implemented in hardware or in the form of hardware plus software functional units.

[0165] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0166] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0167] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0168] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0169] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0170] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by reactance sensing graph kernel, characterized in that, Includes the following steps: 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; S2. A spatiotemporal learner is used 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 law of node marginal electricity price and voltage phase angle. S3. During the training process, introduce physical regularization terms with consistent mechanisms, including power flow safety regularization and ramp consistency regularization, and gradually enhance the influence of physical constraints through weight scheduling strategy. S4. Convert the phase angle trajectory output by the learner into the initial value of the equivalent unit output through physical mapping, and construct a lightweight quadratic programming projection model with the goal of minimizing the deviation from the initial value. Under the conditions of satisfying power balance, line power flow limit, unit output boundary and cross-time ramping constraint, generate a feasible day-ahead scheduling plan.

2. The method for spatiotemporal learning and lightweight DC optimal power flow projection scheduling driven by reactance sensing kernel according to claim 1, characterized in that, The construction of the reactance sensing kernel in step S1 includes: The reactance sensing weight matrix is ​​constructed based on the absolute values ​​of the off-diagonal elements of the susceptance matrix, with the diagonal elements set to zero. After adding self-loops to the weight matrix, the degree matrix is ​​calculated and symmetric normalization is performed to obtain a numerically stable reactance sensing kernel. The multi-order propagation matrix of the graph kernel is pre-calculated to expand the spatial receptive field. A learnable weighted multi-channel parallel fusion mechanism is adopted to combine the output results of local non-propagation paths, single-hop propagation paths and multi-hop propagation paths.

3. The method for spatiotemporal learning and lightweight DC optimal power flow projection scheduling driven by reactance sensing kernel according to claim 1, characterized in that, The graph convolutional encoder of the spatiotemporal learner in step S2 includes: The node feature matrix for each time period is linearly transformed along the local path, single-hop propagation path, and multi-hop propagation path, respectively. The hidden representation is obtained by weighting and fusing the outputs of the three paths using learnable gating coefficients. After fusion, nonlinear activation functions, random deactivation, and layer normalization are applied sequentially, and residual connections are introduced to maintain gradient flow. The feature dimension is gradually increased by using a multi-layer stacked structure, and cross-layer projection is added between layers to alleviate information degradation.

4. The method for spatiotemporal learning and lightweight DC optimal power flow projection scheduling driven by reactance sensing graph kernel according to claim 1, characterized in that, In step S2, the timing encoder adopts a joint structure of TCN and BiLSTM, and includes: Temporal convolutional networks are used to perform dilated convolution operations on the temporal sequences output by graph encoding, and multi-scale temporal receptive fields are covered by stacking convolutional layers with different dilation coefficients. The output of the temporal convolutional network is fed into a bidirectional long short-term memory network to capture long-term cross-temporal dependencies between the forward and backward directions. The outputs of the temporal convolutional network and the bidirectional long short-term memory network are dynamically weighted and fused through a learnable gating mechanism, which adaptively adjusts the contribution ratio of short-term and long-term features during training.

5. The method for spatiotemporal learning and lightweight DC optimal power flow projection scheduling driven by reactance sensing graph kernel according to claim 1, characterized in that, The physical regularization terms with consistent mechanisms in step S3 include: The predicted voltage phase angle is linearly mapped to node injected power through the susceptance 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 components that exceed the line thermal stability limit to guide the prediction results toward the feasible region; Based on the predicted phase angle, the equivalent unit output is derived, and a regularized penalty is applied to the portion of the output change that exceeds the ramp rate limit between adjacent time periods. A phased weight scheduling strategy is adopted. In the early stage of training, the focus is on predicting task loss, and the weight coefficient of the physical regularization term is gradually increased linearly with the increase of training rounds.

6. The method for spatiotemporal learning and lightweight DC optimal power flow projection scheduling driven by reactance sensing graph kernel according to claim 1, characterized in that, The lightweight quadratic programming projection model in step S4 includes: The voltage phase angle predicted by the learner is converted into nodal injection power through the susceptance matrix, and then combined with the load prediction to synthesize the initial value of the equivalent unit output. A multi-period joint optimization problem is constructed with the objective function of minimizing the L2 deviation between the actual unit output and the initial value. The constraints include: system power balance constraints, unit output upper and lower limit constraints, line power flow limit constraints based on power transmission distribution factor, and unit ramp rate constraints between adjacent time periods. A hot-start mechanism based on solutions from adjacent time periods is adopted to accelerate the solution process. Initial values ​​that meet the constraint tolerance are directly adopted, and a fast quadratic programming solver is started for projection correction in the case of out-of-bounds situations.

7. The method for spatiotemporal learning and lightweight DC optimal power flow projection scheduling driven by reactance sensing graph kernel according to claim 6, characterized in that, The fast solution mechanism for the projection model includes: Before projection, a constraint satisfaction pre-check is performed. If the initial value of the equivalent unit output for a certain period is within the tolerance range of all constraints, the optimization solution for that period is skipped. For the time period that needs to be solved, the optimized solution of the previous adjacent time period is used as the initial value for hot start to reduce the number of iterations; In extreme cases of exceeding limits, a rebalancing adjustment algorithm based on line flow sensitivity is activated to bring the power flow back within the limits while maintaining power balance.

8. The method for spatiotemporal learning and lightweight DC optimal power flow projection scheduling driven by reactance sensing graph kernel according to claim 1, characterized in that, The method further includes: A reference bus is uniformly determined during the data preprocessing stage, and the consistency of the reference bus is maintained throughout the entire process; The phase angle of the voltage output by the learner is re-anchored before physical mapping is performed, and the phase angle value of the reference bus is subtracted to eliminate zero-point uncertainty. The same unit system and power reference are used in all stages of training, inference, and projection to ensure system consistency in the calculation of physical quantities.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by the reactance sensing graph kernel as described in any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the spatiotemporal learning and lightweight DC optimal power flow projection scheduling method driven by the reactance sensing graph kernel as described in any one of claims 1 to 8.

Images