Source network load storage routing adjustment method based on AI control strategy

Abstract

The invention discloses a source network load storage route adjusting method based on an AI control strategy, and relates to the technical field of power system control, and the method comprises the steps: collecting a source side signal, a network side signal, a load side signal, a storage side signal and an exogenous signal, and generating a unified system state of confidence weight and topology embedding; the state is used for digital twinning online calibration, a feasible range and a constraint set are extracted, and a projection operator and guaranteed operation curve set is generated; generating a unified energy routing vector in the feasible range by a hierarchical strategy; performing simulation checking calculation in the digital twinning and compiling into a final instruction and execution plan table; performing closed-loop issuing and receipt checking, and performing event write-back to update guardrails and strategies; the method realizes collaboration, improves on-site consumption of new energy, reduces comprehensive cost, enhances boundary compliance and traceability, is suitable for parks, micro-grids and the like, and is convenient for deployment, operation, maintenance, expansion and upgrading.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power system control, in particular to a method for source-grid-load-storage routing adjustment based on an AI control strategy. BACKGROUND

[0002] Industrial parks, urban microgrids and new power systems integrate various types of renewable energy power generation and storage such as distributed photovoltaic / wind power, gas combined cooling heating and power, battery / thermal energy storage, various types of industrial and building and transportation loads, and are connected to distribution network feeders, switch stations and substations on the source, load, storage and network sides.

[0003] With the increasing proportion of new energy, the peak-valley time-of-use electricity price and carbon emission indicators are gradually implemented, and the operation scenarios have high uncertainty and multi-time scale characteristics: the source-side power generation output fluctuates greatly with weather changes; the load side has a certain amount of power shift or interruptible process power window period; the energy storage is limited by power-energy and power-life coupling; and the network side has the risk of cross-section flow limit or voltage limit.

[0004] The existing technology mainly adopts decentralized energy management and control: source-side power tracking, independent storage scheduling, load shifting according to rules, or deterministic model predictive control for intraday rolling; network topology and safety constraints are often simplified, and external factors such as electricity price, carbon intensity and equipment health are often split for management. This results in insufficient cross-side coordination, delayed response, poor strategy feasibility, and easy occurrence of wind and light curtailment, peak demand exceeding, feeder overload or voltage limit, and other situations. Due to information island and poor data quality, the control system is frequently re-adjusted or even triggered to reduce the capacity, and the overall energy efficiency and local consumption of new energy are limited.

[0005] In addition, the industry mainly uses offline simulation for evaluation, and it is difficult to form a unified constraint set that can be called in real time in advance. After the strategy is issued, it is found that it is not executable or needs to be manually corrected; the data layer lacks credibility and time alignment management, and the drift, packet loss and abnormal value amplify the decision deviation; the storage scheduling is biased towards electricity arbitrage or peak shaving, ignoring the life and thermal constraints, which easily leads to excessive charging and discharging; the load shifting relies on static rules, and it is difficult to coordinate with process windows, comfort and demand constraints; the energy management, monitoring and control system protocols are heterogeneous, and lack of unified instruction semantics and closed-loop feedback, which exacerbates the uncertainty and lag of landing.

[0006] Based on the above scenario, under the conditions of considering the power distribution network topology and operation safety boundary, energy storage power-energy and life coupling, load process time window and comfort, time-varying constraints of electricity price and carbon emission, and measurement data reliability difference, how to make unified energy routing decision on source side output, energy storage charging and discharging and load time shifting in multiple time scales, so that it is in the physical feasible region at any time and has real-time adaptive ability to new energy rapid fluctuation and working condition mutation; if it cannot be solved in the long term, it will continue to cause increased curtailment, rising maximum demand and electricity cost, accelerated equipment life decay and rising risk of power distribution network out-of-limit, thereby causing difficulty in achieving energy efficiency and low carbon goals. SUMMARY

[0007] (I) Technical problems solved

[0008] In view of the deficiencies of the prior art, the present application provides a method for source network load storage routing adjustment based on AI control strategy, comprising collecting source side, network side, load side, storage side and exogenous signals to generate a unified system state with confidence weight and topology embedding; the state is used for digital twin online calibration, the feasible range and constraint set are extracted and the projection operator and the bottom line operation curve set are generated; the unified energy routing vector is generated in the feasible range with a hierarchical strategy; simulation and calculation are carried out in the digital twin and compiled into final instructions and execution schedule; closed loop issuing and receipt checking, event backwriting updating guardrail and strategy; the method realizes cooperation, improves local consumption of new energy, reduces comprehensive cost, enhances boundary compliance and traceability, and is suitable for parks and microgrids, etc., solving the technical problems described in the background art.

[0009] (II) Technical solutions

[0010] In order to achieve the above purpose, the present application is realized by the following technical solutions:

[0011] The method for source network load storage routing adjustment based on AI control strategy comprises collecting source side, network side, load side and storage side data, generating confidence weight according to measurement point reliability, and encoding and aligning with electrical topology to form a unified system state with confidence weight and topology embedding and a double scale cache;

[0012] The unified system state is input into digital twin, the equipment and network parameters are calibrated online according to the measurement, the time-related feasible range and constraint set are extracted, and the corresponding projection operator and bottom line operation curve set are generated;

[0013] In the feasible range, a hierarchical strategy of reinforcement learning and digital twin is adopted, the energy quota and routing direction are determined in the slow time scale, the correction is generated in the fast time scale, and the unified energy routing vector is obtained after correction by the projection operator;

[0014] The unified energy routing vector is simulated in the digital twin, compiled into instructions and plans for power output, energy storage charging and discharging, and load shifting, and corrected by a scaled model to form final instructions.

[0015] The final instructions are issued and feedback is received, and the constraints and deviations are monitored and checked. When the limit or deviation occurs, the backup operation curve set is switched, and the feedback and events are written back to update the feasible range and strategy.

[0016] Further, the source side, network side, load side, and storage side measurement points are established with a credibility feature set, a confidence weight is generated based on sensor health, time deviation, packet loss, and abnormal density, and the measurement point naming, unit, and sampling period are unified.

[0017] The measurement points are mapped to topological embeddings based on electrical topology, and the state entries corresponding to the node and line positions are output and written into a dual-scale cache with minute and second levels and source identification information.

[0018] Further, the topology edge weight is constructed according to line reactance, geographical connectivity, and historical tidal flow, and the edge weight source is the rated parameter and operation archive; when time alignment is performed, the unified time source is used to unify the timestamp, unit, and measurement caliber, and the alignment log and difference threshold are established; the system state record version number, source identification, and change description are unified.

[0019] Further, in the digital twin, the unified system state and field measurement are compared to calibrate the equipment and network parameters online; the time-dependent feasible range and constraint set is extracted, covering section limits, node voltage, unit start-stop and climbing, energy storage power and state of charge, load process time window, and carbon budget, and a projection operator is generated; the set and operator use a unified version number and establish a mapping relationship with the previous input.

[0020] Further, when the calibration residual exceeds the set threshold, a backup operation curve set is generated and the strategy incremental update that needs to be triggered is identified, and the switching reason, trigger time, and responsible person identification are recorded;

[0021] The backup operation curve set and the feasible range and constraint set are versioned and archived together, including file verification digest and dependency list, as well as interface adaptation information and storage path.

[0022] Further, within the feasible range, a hierarchical strategy is used, and the slow time scale is based on cost, carbon budget deviation, energy storage life, and wind and light rejection entries to determine energy quota and routing direction.

[0023] The fast time scale constructs a disturbance correction, and the feasible range limits the candidate action space; the candidate vector is corrected by the projection operator to form a unified energy routing vector and establish a source label, constraint reference, and simulation

[0024] Further, tail risk control is introduced in generating perturbation correction to limit exposure of high cost scenarios, and the correction is mapped to power supply output, energy storage power and load time-shifting channel through inter-layer influence matrix and superimposed on slow time scale results;

[0025] The matrix determines the non-zero structure according to the unit response bandwidth, energy storage available capacity and load whitelist, and is consistent with the channel order of the unified system state and records the mapping entries.

[0026] Further, a multi-scenario set of weather, load and communication packet loss is constructed in the digital twin, the unified energy routing vector is simulated and calculated, and a modified component table is output; the vector is compiled into AGC given, energy storage active and reactive power instructions and load task list, and a plan execution table and a reply template are prepared, and a scenario sorting view and a source label are generated on the check page.

[0027] Further, compliance correction is performed on the compiled instructions, and constraints include at least minimum start-stop time, ramp rate, cross-sectional flow limit, node voltage range and demand limit;

[0028] The energy management, monitoring and building control interface is connected, downlink message samples are generated, field mapping relationship and verification method, message length and time scale are recorded, and the index of the model and the set is established.

[0029] Further, according to the final instruction issued and the device reply collected, the deviation is aggregated in a fixed window and bound to the device caliber; combined with the active constraint list, consistency check is performed, a deviation source list is generated, and the corresponding relationship with the final instruction is recorded in the version file; the window slides on the time scale of the execution plan table, and the device attribution list is generated on the reply check page according to the output entries of the channel and the information is archived.

[0030] Further, when the joint threshold criterion meets any triggering condition, the corresponding entry of the guaranteed operation curve set is executed and a switching event sheet is generated; the criterion includes a reply deviation threshold, an active constraint limit mark and a distribution deviation of perturbation statistics;

[0031] The event sheet records the switching time, trigger source and positioning index and is written back to the previous step for updating the calibration residual and the warning band, and for adjusting the risk weight and the candidate action space.

[0032] Further, in the parallel contrast framework, key performance indicators are calculated for the main strategy and the shadow strategy respectively, and contrast entries are generated; the weight variables are updated according to the preset weight to adjust the inter-layer influence matrix; the key performance indicators are calculated in the order of energy consumption cost, carbon emission standard, safety margin, device health loss and wind and light suppression system, and the source reference is established with the unified energy routing vector to generate daily and weekly reports.

[0033] Further, the unified system state, the feasible range and constraint set, the projection operator, the unified energy routing vector and the final instruction are established version number, source and timestamp archive record;

[0034] The archive contains field mapping, active constraint list, source label and instruction index; when cross-step reference is made, object identification and version number are located and kept consistent with the strategy page, audit page and interface page and a rollback point is set.

[0035] (Three) beneficial effects

[0036] The present application provides a method for source network load storage routing adjustment based on AI control strategy, which has the following beneficial effects:

[0037] Through the establishment of a unified system state, information such as source side, network side, load side, storage side and exogenous signals is collected, and the system online time and space mapping of online confidence weight and electrical topology fusion is realized to achieve minute / second scale dual-scale buffering and time alignment, online input caliber is ensured to ensure data extraction and backtracking of original state; at the same time, based on online calibration of digital twin, device and network parameters are updated online according to field measurement, and then the operating procedures and device boundary values are extracted into time and state value related feasible range and constraint set corresponding in time, and projection operator and bottom line operation curve set are generated to form front guardrail and fallback path.

[0038] In the strategy generation of hierarchical and risk sharing, energy quota and routing direction are calculated in minutes, and correction disturbance is calculated in seconds, and then the unified energy routing vector is output under the projection operator correction on the basis of power output, energy storage power and load time shifting channel by means of interlayer influence matrix; in the strategy simulation verification and instruction compiling process, the modified component table is obtained on the premise of replaying weather, load, communication packet loss and other scenes in digital twin, and is converted into automatic generation control given value, energy storage active / reactive power instruction and load task list, and the final instruction, execution plan and reply template are obtained after correction by means of hybrid model.

[0039] According to closed loop execution and online self-adaptation: according to the final instruction and the reply information of the device, the online active constraint list and the deviation and distribution offset are verified to trigger the switching of the bottom line operation curve set, and the event is structured and written back to the digital twin and the strategy layer, and the closed loop is realized by means of parallel comparison of strategy and key performance indicator management. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The present application provides a method for source network load storage routing adjustment based on AI control strategy, which has the following beneficial effects: DETAILED DESCRIPTION

[0041] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work belong to the protection scope of the present application.

[0042] Please refer to Figure 1 The present application provides a method for source network load storage routing adjustment based on an AI control strategy, comprising,

[0043] Step one, based on the park / micro-grid scene, in the source side, network side, load side, storage side and exogenous signal measurement values, the measurement points are gated and fused according to the confidence, and the time-space coding and time alignment are completed combined with the electrical topology, forming a unified system state with confidence weight and topology embedding, and completing the caching of minute-level and second-level states, unifying the name, unit, time stamp, and fixing the source identification as the unified input caliber of subsequent constraint extraction and strategy generation;

[0044] The field measurement sources are different, the precision is different, and the time calibration is not unified. If directly spliced, it will distort the state and affect the subsequent constraint extraction and decision calculation. At this time, according to the confidence, the measurement is gated and fused, and based on the electrical adjacency relationship of the network topology, the state is structured and coded to keep consistent with the input of the physical system.

[0045] A confidence characteristic is established for each test point, and a confidence weight is calculated. Then, the original measurement and the model completion measurement are fused together by using the interior point method. By using the method of smooth embedding based on network topology, the states contaminated by noise or close in position are brought together by using the electrical adjacency property, and a topologically consistent state representation is given, obtaining the state of the unified system. It is illustrated that the electricity price / carbon intensity is obtained from the power market interface / supervisor platform bulletin / enterprise internal account; the process time window is obtained from the work order of the production execution system; the device curve and rated value are obtained from the account + nameplate + recent test.

[0046] At each measurement point, according to four types of observable characteristics such as sensor health, time deviation, packet loss and abnormal density, the confidence weight of the measurement point is calculated. Under the confidence weight gating, the original measurement and the completion measurement value are fused to form a measurement synthesis quantity that is not sensitive to noise and missing measurement. The specific method is as follows:

[0047]

[0048] In the formula: single-point confidence weight , the value range , is used for gating; the confidence gating function , monotonic sigmoid mapping defined as , control the smoothness of weight; sensor health score , value , generated by check record and self-check result, reflecting hardware status; time synchronization deviation , value range , derived from comparison with time server; packet loss rate , value range , derived from communication statistics; abnormal density , value , derived from out-of-bound count and shape check; weight coefficient , non-negative real number, used to characterize the relative influence of four types of features, which can be obtained by on-site calibration.

[0049] At the vector level, the confidence weight of each measuring point is combined into a confidence weight vector, and the original measurement vector and the missing measurement completion vector are fused element by element to obtain a measurement synthesis vector:

[0050]

[0051] In the formula: the measurement synthesis vector , the dimension is consistent with the measuring point, which is used as the subsequent embedding input; the confidence weight vector , each element is stacked to form a value range element by element ; the original measurement vector , the on-site reading at the sampling time , the unit is unified; the missing measurement completion vector , the alternative reading obtained by history segment, device model and neighborhood regression;

[0052] The unit vector , the same dimension as the confidence weight vector , is used to complete element by element; the element-by-element multiplication operator , an operator that multiplies element by element of vectors of the same dimension;

[0053] In application, the measurement of unstable or short-time missing on site will not enter the state, but will be replaced by homologous / similar information; different weights will be obtained at different measuring points, which can also reduce manual rule intervention, and the measurement stability curve will be smoothed in the situation page.

[0054] According to the power grid topology network, a topological Laplace matrix is established using the node-line electrical adjacency relationship, and the measurement synthesis vector is mapped into the topological embedding matrix using Laplace smoothing to mine the cooperative changes between nodes that are physically close and weaken the influence of isolated point noise:

[0055]

[0056] topological embedding matrix , dimension consistent with , as an important component of the unified system state; identity matrix , dimension consistent with ; balance coefficient , positive real number, controls the strength of topological smoothing, taking values in ; topological Laplacian matrix , defined by the difference between the degree matrix and the adjacency matrix , where edge weights are weighted by line reactance and geographical proximity.

[0057] By using topological smoothing, the node measurements connected on the same feeder or connected on the same voltage level are considered simultaneously in the application, and the random jitter of those isolated measuring points is suppressed, which improves the spatial consistency of the entire system. On the visual curve, it can be seen that the changes of node power on the same feeder are phase-locked.

[0058] The data on site has time error and sampling interval. If they are mixed directly, the actual state and the time corresponding relationship of the equipment will be lost. The values obtained by sampling different physical quantities of equipment also need to follow the power conservation law, capacity boundary, and voltage level qualification conditions, and can only be used as prior conditions for subsequent screening of feasible range. Therefore, we first align the time base, and then use the mode fitting method based on physical consistency projection correction for fitting.

[0059] First, construct a time alignment function to map different sampling sequences to the same time base, and establish a minute-level buffer-second-level buffer double buffer architecture. Then, project the topological embedding results and exogenous signals based on physical consistency, eliminate system state components that violate conservation and capacity boundaries, and obtain the integrated system state for downstream use.

[0060] To solve the problem that the reporting time of multi-source equipment may not be synchronized, a time delay function is introduced to minimize the maximum offset between source data and source curves and optimize the smoothness of the curves, and the aligned results are stored in two-level cache to meet the calling needs of different time scales. The method is as follows:

[0061]

[0062] time alignment function , strictly increasing time mapping, satisfying , used to correct the misalignment of reporting time; continuous time observation function ​The energy calculation is aligned with the continuous curve obtained by spline interpolation of the original discrete sequence; this serves as the first step in the normalization and unification of the multi-channel discrete measurement. The vector function in the corresponding interval is used to input the continuous curve required for digital twin calibration, time and policy calculation;

[0063] The default method uses piecewise cubic spline interpolation: each sampling interval is represented by a cubic polynomial, with adjacent intervals being continuous at the function value and its first and second derivatives, and natural boundaries at the endpoints (i.e., the second derivative is 0 at the endpoints). The spline coefficients are solved by a tridiagonal linear equation system. The input only needs to provide the time series and measurements of the same caliber, and the output is a multi-channel observation vector at any time. When the samples are sparse or there are step changes, the method can be changed to conformal Hermitian splines or zero-order preservation according to the caliber. Matching this is the interface of dual-scale buffer, feasible range extraction and projection operator. In this way, relevant technical personnel can realize a partial understanding and application of the alignment algorithm.

[0064] Observation window , a positive real number, representing the length of the time interval used for alignment; smoothing weight A non-negative real number that controls the curvature penalty of the mapping; a typical value is... to Norm symbol The first norm is used to reduce the impact of abnormal spikes on alignment.

[0065] When in use, the time mapping pulls the electrical energy and power curves from different devices back to the same time base; the two-level buffer is used to support minute-level and second-level calculations, reducing misaligned operations caused by cross-scale calls.

[0066] For the time-aligned topology embedding matrix, the topology embedding matrix and exogenous signals are used to construct the state to be projected. Under constraints such as power conservation, branch limits, node voltage acceptable range, energy storage state of charge, power coupling, process time window, and comfort, a consistency set is formed. The projection operator is then used to push the state back to the vicinity of the consistency set boundary, eliminating physical violations caused by noise or alignment errors. This projection can be performed sequentially according to priority: first, projection of power conservation and branch limits; second, projection of energy storage coupling and process time window; and finally, projection of voltage acceptable range. Correction residuals are left for each projection step for later statistical analysis.

[0067] When applied to training, the interpretation of state parameters of different systems at the physical level is unified under the consistent projection state after the consistent projection transformation. This allows the downstream feasible range and constraint set to directly reuse the above consistency conditions. The projection problem can be solved by sequential quadratic programming and constrained least squares, and the solver can use sparse decomposition to reduce the amount of computation.

[0068] Step 2: Generate a digital twin based on the unified system state-driven approach, complete the online calibration of equipment and network parameters, extract the relevant feasible range and constraint set of the time series online in combination with the operation procedures and device boundaries, generate the corresponding projection operator and the set of guaranteed operation curves, and establish an active constraint list and calibration residual index table, so that subsequent strategies can be invoked within a certain physical range and emergency channel, and establish a correlation on the dual-scale cache.

[0069] Due to aging equipment, drifting line parameters, and environmental changes, existing models may become disconnected from the field, and without correction, these errors will propagate into the feasible range or constraint set. This will affect the implementation of subsequent decisions.

[0070] Therefore, by utilizing the sequential logic of energy coupling assimilation and power flow voltage consistency convergence, the total residual is first reduced, and then converged to the engineering-acceptable range under electrical constraints; this is achieved through confidence weight vectors. Selective Trust - Using Digital Twin Parameter Vectors Correction - by residual vector The given link emphasizes the traceability of parameter write-to-disk, the verifiability of power flow consistency, and the visible final state that can be rolled back in case of errors.

[0071] First, a unified system state vector is used. Trigger the energy balance calculation of the twin model to obtain the model measurement mapping. And verified on-site synthetic measurement vector Compare these values ​​to form a calibration residual vector. Based on this, a residual weight matrix that considers the importance and reliability of the measurement points is added. And through the parameter update formula of the first-order structural damping, To make corrections, the update should both incorporate information and avoid excessive deviation. The specific methods are as follows:

[0072]

[0073] Wherein: digital twin parameter vector This includes: line reactance, transformer short-circuit impedance, unit efficiency coefficient, energy storage internal resistance, and capacity correction coefficient, the values ​​of which are determined by the equipment nameplate and historical records; updated parameter vectors. With digital twin parameter vector Same dimension, write it back into the model.

[0074] Jacobian matrix This refers to model measurement mapping. For digital twin parameter vectors The partial derivatives can be obtained through automatic differentiation or finite difference, and their size is the measurement dimension × the parameter dimension; the residual weight matrix It is a diagonal matrix whose diagonal elements are composed of the confidence weight vector. All values ​​obtained by mapping to the importance of the measuring points are positive; damping coefficient Non-negative real numbers control the update step size, with a typical range of [missing information]. to ;

[0075] identity matrix ,and Same dimension; matrix transpose symbol calibrate residual vector The definition is shown in the following formula;

[0076]

[0077] Where: calibration residual vector Consistent with the measurement dimension, it is used to characterize the difference between the model output and the on-site synthetic measurement;

[0078] Model Measurement Mapping ,Will and The vector is mapped to comparable measurements, including bus active and reactive power, branch power, transformer load rate, and energy storage state of charge estimates. The functional expression is obtained by using equivalent circuits or equipment characteristic curves.

[0079] Unify system status With twin parameters The mapping is to comparable measurement vectors (bus reactive power / reactive power, branch power, transformer load rate, and estimated state of charge of energy storage). The construction method can use the equivalent circuit power flow equations combined with the superposition of equipment efficiency / loss curves. In this process, the equipment curves are selected as piecewise affine or logarithmic-exponential fitting.

[0080] Parameter source / range: Includes line reactance, transformer short-circuit impedance, unit efficiency coefficient, energy storage internal resistance / capacity correction, etc., derived from nameplate, type test, and maintenance records of the past three months; parameter ranges start from nameplate ± engineering tolerance (e.g., ±20%), exceeding limits requires manual confirmation. Jacobian matrix Automatic differentiation can be used; finite difference can be used when conditions are not met.

[0081] In-situ synthetic measurement vector Obtained by gating fusion in step one, and Align them one by one in terms of dimensions and position;

[0082] The processing method is as follows: First, use the confidence weight vector Set the residual weight matrix The values ​​of the diagonal elements ensure that high-confidence measurement points contribute significantly to parameter updates; then, the parameter increment is solved and the parameter vector is updated. At the same time, update the parameter vector Save the data and record the data source, time, and responsible person for the updated parameter vector to form a parameter update record.

[0083] When applying the technology, if the parameters are updated, the technology is more sensitive at high-confidence measurement points and less sensitive at low-confidence measurement points. You can see the convergence of the residuals by checking the parameter change record page or by looking at the energy balance residual trend curve.

[0084] After convergence in terms of energy coupling, engineering consistency still needs to be achieved at the power flow voltage level. This involves updating the parameter vector. Substitute the load into the network equivalent model to calculate the power flow. Compare the load rate and voltage deviation with the previous data. If the deviation exceeds the allowable range of the project, select and adjust the subset of parameters that affect the sensitivity (such as line reactance temperature correction, transformer equivalent impedance short-time correction, and the segment coefficient of the unit efficiency curve shift). Then, perform a process of inspection, fine-tuning, and re-inspection until all monitoring points meet the requirements.

[0085] Furthermore, sensitivity scanning is used to determine the influence matrix of the parameter-monitoring point. First, parameters with significant influence are adjusted, followed by cascaded fine-tuning. Using the on / off flag and zero offset from the power flow consistency verification report as reference indicators, sparse Newton iteration is employed for power flow solution. Convergence criteria consistent with the scheduling procedure are set, and the updated parameter vector is obtained after convergence at both the energy and electrical layers. It has physical significance and can be used in the field.

[0086] To ensure that subsequent strategies include elements such as grid security boundaries, equipment boundaries, process time windows, and carbon budgets, forming a set of limits, inaccuracies could lead to boundary violations or execution failures. Therefore, a linearization-constraint shaping-projector generation-safety curve binding chain is used to obtain the set.

[0087] First, the equivalent power grid is linearized at the operating point to obtain the injection-branch-voltage mapping relationship, and all subsequent constraints are converted into unified variables. Then, the cross-sectional power flow limit, node voltage acceptable range, energy storage power-state of charge coupling, unit ramping and minimum start-up and shutdown, load process time window and comfort, carbon budget and carbon intensity curve are combined to form the feasible range and constraints. .

[0088] Furthermore, the linearization mapping can be written as:

[0089]

[0090] Where: Branch power vector It consists of active and reactive power, with the dimension being the number of branches; node voltage vector. Phase voltage amplitude or voltage offset, dimension is the number of nodes; sensitivity matrix In the system state vector With the updated parameter vector The determined operating point is linearized, with real numbers as elements, and can be derived from the Jacobian block method of the Newton-Raphson method; the injected active vector Injecting reactive vector Composed of power source, energy storage and load The input / output is composed of; bias vector The intercept term reflects the offset between the linearization point and the current working point;

[0091] Under this mapping, the constraint is shaped as follows: branch limits are applied to the branch power vector as interval constraints. Voltage qualification is applied to the node voltage vector by interval constraints. Energy storage power-state of charge is applied using the power-energy coupling inequality. The corresponding components; unit ramp-up and minimum start-up / shutdown are applied to the injection sequence using the adjacent time period difference inequality; process time windows and comfort are applied to the corresponding load injections using binary window functions; carbon budget is applied to the injected active power vector using time period cumulative constraints. The cumulative emissions are obtained by convolving the carbon intensity curve with the carbon intensity curve.

[0092] According to minute-level time index set The reactive vector to be injected Mapping to the correct time granularity, and linearly or piecewise linearly generating a series of time-varying constraints over time periods that satisfy the power injection characteristics, then incorporating and labeling a 15-minute segment within a certain interval and adding it to the set. A version number is assigned. Within the feasible range and constraint set archive, each type of constraint's corresponding interval set and whole-time constraints can be viewed item by item, and the feasible range version number for the main station-side task is implemented; the power grid model, device nameplate, etc., are taken from... In this study, sparse matrix decomposition and sensitivity calculation steps are used as the solution procedure.

[0093] Forming a feasible range and a set of constraints Subsequently, a directly callable projection operator is needed to push any candidate state decision vector back to the vicinity of the set boundary; simultaneously, when calibration residuals or constraint conflicts intensify, it is necessary to switch to an engineering-executable set of baseline running curves. The projection can be defined by the following formula:

[0094]

[0095] Where: Projection solution It has the same dimension as the candidate state decision vector and is used as the input to the subsequent stages as the candidate vector to be corrected; the candidate state decision vector The dimensions are derived from prior actions in step three or extrapolations from historical feasible points, with power output, energy storage power-state of charge, and load injection; the projected metric matrix... A symmetric positive definite matrix, used to highlight the deviation costs of different components, with all values ​​being positive real numbers, and set according to the importance of the equipment and the cost of the action; feasible range and constraint set. It has already been provided.

[0096] The projection problem is given to the solver in the form of a quadratic programming problem to obtain the projection solution. And an active constraint list; if there are branch limits, voltage qualification high-risk items, etc., that are in an active constraint state, and have been in this state for multiple consecutive periods, then synthesize a set of guaranteed operation curves based on the guaranteed strategy library. (Unit fixed output curve, energy storage limiting curve, load shifting whitelist) and push it to the main interface for display, a set of guaranteed operation curves. Available as a state to be switched.

[0097] The projection solution can be found on the projection calibration page. and candidate state decision vector Component comparison bar, active constraint list, and set of guaranteed operation curves. After the switch occurs, a notification will appear on the operation mode page indicating that the safety curve has taken effect.

[0098] The quadratic programming solver employs either the interior-point method or the active-set method, with its derivatives provided by the linearization mapping block, allowing for a set of solution times in minutes. Injecting reactive vector The time index set remains consistent.

[0099] Step 3: Within feasible limits, apply a hierarchical strategy of reinforcement learning and digital twins to determine energy quotas and routing directions at the minute level, and generate corrections based on perturbations at the second level. Project the mapping matrix between layers onto channels such as power output, energy storage power, and load time shift. After transformation by the projection operator, a unified energy routing vector is obtained, which is then aligned with the minute-level and second-level caches. The source labels and candidate constraint references are retained, and the source data is then used as parameters for compilation and simulation.

[0100] The electricity price, carbon intensity, and process window in the industrial park all have obvious time-based and sequential characteristics. On the slow time scale, the quota and direction should be determined. Otherwise, the fast time scale will not be able to take into account the goals of cost, carbon, and lifespan coordination.

[0101] Therefore, following the multi-objective benefit configuration-safe projection correction as the sequential logic, the long-term dimensions at the minute level and the network boundary are moved to the skeleton layer of the candidate vectors to ensure the degrees of freedom of fast time-scaled correction under support; system state vector Given the state, feasible range, and set of constraints. Given the boundary, the time-period projection operator To ensure terminal convergence; finally, the main station's slow timescale can be observed and read, the projection verification page can trace the source of constraints through querying, and the version archive can record the source of parameters.

[0102] To avoid neglecting one aspect for another, multiple objectives, such as electricity purchase and fuel costs, carbon budget deviations, energy storage lifespan losses, and wind and solar curtailment penalties, are compressed into a single minute-level objective functional. Appropriate weights are assigned to each, and the range of values ​​for each weight is determined, enabling a coupled solution that addresses various factors including costs, carbon emissions, lifespan, and grid integration.

[0103]

[0104] Among them: slow-timescaled objective functional is a real number not less than 0, used to be minimized, corresponding to a set of time indices on the order of minutes. Then according to the buffer The time scale is an ordered finite set; cost weight vector It is a decision vector related to the candidate state. A column vector of equal dimensions, where each component is a non-negative real number, reflecting the electricity price and marginal cost of fuel in the current time period.

[0105] Candidate state decision vector To group power output, energy storage capacity, state of charge, and time-shiftable load injection together into a single column, corresponding to a candidate state decision value vector under each command channel; carbon budget weights. , is a positive real number, representing the severity of the penalty for defaulting on the carbon budget; when carbon emissions during that period... Non-negative means that the sum of the area under the carbon intensity curve of the electricity purchased and injected during the time period and the area under the curve of the carbon intensity curve of the time period.

[0106] Carbon Budget for the Period A scalar value, either externally committed or internally allocated, and non-negative; energy storage lifetime weight. Positive real numbers control the intensity of the loop depth penalty; lifetime loss function. , which is a nonlinear increasing function of energy storage power and energy storage state of charge components, employs a piecewise logarithmic function to enhance the effect of heavier penalties for deeper cycles, and is continuous within different piecewise intervals:

[0107]

[0108] For the first Lifetime weighting of energy storage units; This is a depth index approximated by rainflow count or window range; its value range corresponds to the manufacturer's lifespan curve. Wind and solar curtailment weight. Positive real number; amount of wind and solar power curtailment A nonnegative scalar obtained from the difference between the predicted source-side power output and the load and energy storage absorption capacity; superscript This indicates transpose.

[0109] As a supplement, carbon emissions are calculated:

[0110]

[0111] The carbon intensity on the grid side at that time (platform bulletin or enterprise ledger). To purchase power components from the public grid, The time step is in minutes.

[0112] Calculate abandoned electricity:

[0113]

[0114] The three sources are the source-side prediction curve, the process / building load whitelist, and the energy storage available capacity assessment, respectively.

[0115] system state vector Minute-level component generation , and The input entries, and the carbon budget for the given time period. The upper limit is used as the input item, and the feasible range and constraint set are combined to construct the slow time-scaled objective functional. Subsequently, a convex programming solver with boundaries is used to solve for the feasible range and constraint set under linear-piecewise affine constraints. The candidate state decision vector obtained by solving The skeleton solution is obtained; after the solution is obtained, a minute-level energy quota list is generated using a slow timescale view and associated with the carbon budget usage curve.

[0116] In use, the intraday distribution of minute-level electricity purchases, energy storage charging and discharging, and portable loads is displayed in an ordered segmented manner using a slow timescale view; carbon budgets are checked item by item using curves and a list of abandoned electricity items; the solver is set to the interior point method, with a tolerance set to... The derivative is generated by the linearization mapping block completed in step two, and is transformed into a solution of the equivalent unified functional of multi-source constraints through single-objective transformation, which is conducive to the lower-level development of inter-layer coordination work and improves coordination efficiency.

[0117] Because of the slow time scale objective functional The solution may exhibit sharp turns near the boundary, so a gentle correction is used to pull the skeleton solution back into the feasible range, and then a certain degree of freedom is granted to allow the fast timescale correction to pass:

[0118]

[0119] Where: slow-timescaled safe candidate vector Column vectors; slow-timescale skeleton solution The temperature coefficient is obtained by solving the aforementioned steps. The value is in The real number controls the injection ratio for projection correction; the time-period projection operator. The projection operator defined in step two is in the time index Instances at that location.

[0120] Projection operator for each time period The system applies the projection vector, active constraints, and existing projection verification pages to the time period, respectively. Then, it performs a linear mixture of the existing active constraints according to the above formula and incorporates them into the projection verification page. If there are high-risk items such as branch limit or voltage qualification active constraints in a certain time period, the system will mark the cluster of this time period in yellow on the strategy page to prompt the fast time scale to pay attention.

[0121] The projection verification page contains difference bars for the skeleton solution and the projection solution, which are gentle and smooth. To ensure that the difference under projection is not too large and that slow timescale candidates will never again touch high-risk boundaries; a gentle projection method is adopted to ensure boundary consistency while leaving some room for fast timescale to be coherent and adjustable.

[0122] Short-cycle disturbances caused by cloud shadows, load fluctuations, and communication gaps, if directly applied to the grid skeleton solution at that point, may cause the output or voltage of branches passing through that point to exceed limits. Therefore, according to the sequential logic of tail risk control-interlayer synthesis mapping, the fast timescale correction is limited to the direction most sensitive to tail events, and the correction is mapped to the three types of tail execution quantities using a fixed interlayer influence matrix; a slow timescale safety candidate vector is used. As a baseline, inject reactive vector Provide high-frequency perturbations within a feasible range and constraint set. Provide boundaries; the key is to ensure that the corrected channels can be viewed separately on the monitoring page, and that the synthesized vectors do not contain an active constraint list that exceeds the boundaries.

[0123] In the face of second-level disturbances, high-cost events are controlled through conditional tail risk measurement, ensuring that corrections are only executed when mitigating over-limit and power curtailment:

[0124]

[0125] Among them: Conditional tail risk measurement Risk assessment operator, taking values ​​of non-negative real numbers; confidence level The value is in For real numbers, pay more attention to extreme tails when they are close to 1; auxiliary variables Real numbers, used to characterize cutoff points; mathematical expectation operator Take the expectation of the sample distribution with second-level perturbation; positive part function Take the positive part of the input; tail cost random variable. The non-negative random quantity is composed of the weighted sum of branch load over-limit penalties, node voltage over-limit penalties, and wind / solar curtailment penalties, and its weights inherit... And add a safety weight to items that exceed the limit;

[0126] Injecting reactive vector The high-frequency components are mapped to the linearized components in step two to generate perturbation samples, and the tail cost random quantity is constructed. The sample set; given a confidence level Calculate under the condition The subgradient direction limits the correction amount as it decreases. In the direction of the direction, and through amplitude threshold mapping, ensure that the maximum step size of each execution channel does not exceed the device's allowed step size; form a fast timescale correction vector. The source of each component is then output on the corrected channel page.

[0127] The second-level correction component correction channel page shows a trend consistent with the disturbance, and the active constraint table of related entries that exceed the time limit decreases sharply; by using tail risk measurement, without affecting the slow time scale, the phenomenon of prioritizing resources to correct the situation that may cause great costs is mitigated, thereby mitigating the occurrence of extreme events.

[0128] To ensure consistency across different execution types, an inter-layer influence matrix is ​​constructed, mapping fast timescale corrections to three channels: power output, energy storage power, and load time shift. These are then superimposed and projected onto the boundary to obtain candidate state decision vectors and safe candidates.

[0129]

[0130] Where: projection operator Any candidate Projected onto a set consisting of linear-segmental-affine-cone constraints superior:

[0131]

[0132] Feasible range Specifically, this includes cross-sectional power flow limit constraints and node voltage range constraints (composed of sensitivity blocks). (Representation with L2 norm constraints), minimum start / stop / ramp time (linear differential), energy storage power-state-of-charge coupling (affine inequality), load process time window (interval / indication constraints), carbon budget (time period cumulative). Sensitivity block: The linearization of the exchange current at the working point is obtained by dividing the block by the Newton-Raphson Jacobian method.

[0133] Complete candidate state decision vector Column vector, with dimensions consistent with the instruction channel in step four; slow timescale safe candidate vector The interlayer influence matrix is ​​obtained by concatenating the outputs at the corresponding time points. A three-block matrix, where elements within each block are non-negative real numbers, calibrated according to the unit response bandwidth, energy storage power coupled with state of charge, and availability within the load shift window; fast time-scaled correction vector. The output is from the previous text.

[0134] Furthermore, let's first look at the interlayer influence matrix. The complete candidate state decision vector is obtained by superposition. Then, the time-period projection operator is executed once on the projection verification page. Obtain a safe candidate The active constraint list, overlay source, and projection deviation are archived. If projection fails or a high-risk term with active constraints is generated at a certain time, the strategy will then implement a safety net running curve set at that moment. The corresponding curve is displayed and highlighted in red on the strategy page.

[0135] The hierarchical view allows you to see the three-channel overlay curve and its source label simultaneously; the projection verification page displays safety candidates. All components are within a feasible range; through a fixed mapping, the slow and fast layers can be synthesized, and the two variables complete a closed loop within a single space, ensuring a one-to-one correspondence between the output and the structure of the instructions in step four.

[0136] Step 4: In the digital twin scenario, use the aforementioned unified energy routing vector as the basis for calculation verification under various conditions, and form a parameter table to be repaired; on this basis, compile the automatic generation control values, energy storage active and reactive power control instructions and load task sheets, and perform constraint corrections on the joint model in terms of start-up, ramping, cross-section and voltage, generate the final instructions, execution plan and reply template, prepare relevant message examples and interface mappings, and form the corresponding version number and effective constraint list.

[0137] Security Candidate Candidates from the previous stage may collide with physical boundaries or device limits due to combined disturbances; coverage verification needs to be completed before distribution.

[0138] Based on the perturbation construction-response replay-consistency assessment-modification suggestion chain: firstly, utilize minute-level buffers respectively. and second-level buffer The perturbation set is generated in a certain way, and then according to... Playback of the defined isonetwork and device curves, based on... After reviewing each out-of-bounds and critical item, a list of suggested modifications is generated. Then, the modified data points are processed simultaneously in the next step using the compilation method described in this step.

[0139] For three types of disturbances—weather, load, and communication—a minute-level buffer is implemented. and second-level buffer A perturbation sample set is constructed using a time scale, and a coverage weight is used to represent the frequency of occurrence and the importance of the process; in digital twins, the ... To obtain unified verification metrics for input playback responses, which are used for sorting and screening key scenarios:

[0140]

[0141] Among them: unified verification indicators A non-negative real number used to rank the overall risk of a scenario; the total number of samples. Positive integers, determined by the scene library and time window; coverage weight. Non-negative real numbers that satisfy The weighting table is derived from historical statistics and operational experience; the loss function... This is a function that applies piecewise convex penalties for power flow exceeding limits, node voltage exceeding limits, device saturation, and load defaults. The parameters are derived from the operating procedures; details are as follows:

[0142] Segmented convex penalties are applied to quantities exceeding limits / boundaries / saturation / violations to standardize verification metrics. For any quantity exceeding limits... :

[0143]

[0144] coefficient Penalty intensity; inflection point Control the continuous slope of the two-segment structure; This represents the extent of the violation (a positive value indicates a breach). The parameter is set by the operating procedures and a weighted table based on maintenance experience.

[0145] Response Operator , will security candidate With perturbation samples Input a twin model and output vectors of branch power, node voltage, device power and energy state, and load execution state;

[0146] In parameters Next, security candidates With perturbation samples The mapping is a response vector of branch power, node voltage, device power-energy state, and load execution state.

[0147] By Flow Solver With device dynamic integrator Cascading: First, select safe candidates As the injection boundary, solve for the power flow, and then integrate it in the device power / energy equation (energy storage). (Unit ramp-up, task start-stop sequence), and synthesized response.

[0148] Perturbation samples Includes cloud cover-irradiance curves, fluctuating load curves, and packet loss sequences; taken from historical or scenario databases, with time steps and... Alignment. Power flow can be achieved using sparse Newton iteration; device state integration can be achieved using backward Euler or trapezoidal rule.

[0149] For voltage-sensitive but low-capacity applications, A linearized sensitivity mapping can also be used as an approximation.

[0150] Perturbation samples A vector consisting of three parts: cloud cover-irradiance curve, pulsating load curve, and packet loss sequence; digital twin parameter vector. The parameter set that takes effect after calibration in step two.

[0151] Extract disturbance samples from the scene library by category and time window. and assign coverage weight Input the event into the response operator The playback yields a vector of engineering quantities, which is then input into the loss function. The unified verification index is obtained by summing the results. Output to the scene sorting view on the verification page; and automatically generate a list of modification suggestions based on the scenes ranked higher, which includes annotations for restrictive branches, critical nodes, and device limits.

[0152] The verification conclusions are arranged in descending order of the unified verification indicators in the above scenario sorting view. Engineers can replay the trajectory by selecting the verification results. The unified verification indicators are used to achieve normalized sorting of multi-source risks, which facilitates centralized processing of key issues.

[0153] If several high-order scenarios are selected from the scenario sorting view, the view is browsed according to the scenario sorting method. Each selected item is viewed individually, including the branch limits, voltage compliance ranges, energy storage power-state-of-charge coupling curves, unit start-up-ramp curves, and load process time windows. Modification suggestions are given according to a three-tier threshold system: out-of-bounds, critical, and safe. For out-of-bounds items, the priority order is given to reduce output components, transfer load components, and adjust charging and discharging power components. For critical items, suggestions are given to reduce step size and delay execution. For safe items, the original values ​​are retained.

[0154] For each suggestion, establish a security candidate. The component matching list is generated, and the proposed component table in the version file is used as input for the next sub-step. The threshold stratification statistics bar can be viewed on the calibration page, and the relationship between specific components and corresponding devices can be found in the component table to be calibrated. Furthermore, based on the procedure thresholds and device curves, structured suggestions that can be compiled and directly called by the standard process are directly generated.

[0155] There are semantic and constraint differences between policy variables and field devices. These differences need to be eliminated through explicit mapping and backend correction, following the sequence of semantic mapping, format arrangement, constraint correction, and protocol encapsulation:

[0156] First, the variables are converted into channel quantities that match the semantics and dimensions of the device; then, deterministic corrections are made to the merging model; finally, electrical frames are generated according to the specification, and a receipt template is provided. To ensure consistency across the three channels, we constructed a device mapping matrix-bias vector, distributing the three strategy variables into the unit given values, energy storage active / reactive power, and load task sheets, respectively.

[0157]

[0158] Where: instruction vector The column vector is composed of three parts: generator channel, energy storage channel, and load channel. The units of the elements are megawatts, volts, or dimensionless work order opening and closing flags.

[0159] Device mapping matrix A block-based sparse matrix, where each block is constructed from unit response bandwidth, energy storage power-state-of-charge coupling, load whitelist, and process time window; elements are real numbers; safety candidates. The projected safe candidate vector; the bias vector It is synthesized from the device baseline, external scheduling limit and historical work order status.

[0160] First, the components selected in the component table to be modified are mapped to safe candidates that contain the same feature attribute. After overwriting the corresponding components, the mapping is performed to obtain the instruction vector. Next, the AGC setpoint table, the energy storage active-reactive power instruction table, and the load task sheet table are generated, and source tags are added to the components (from the original candidate or from the modified component). Finally, the message generation function is used to output downlink frame samples of the IEC60870-5-104 protocol and IEC61850 (MMS / GOOSE / SV standard).

[0161] On the compilation page, you can view the three types of tables and source tags one by one; on the interface page, you can capture downlink frame samples for frame structure verification; device mapping matrix. The message generator is jointly generated from curves and interface manuals for digital twin devices, and fills in the message according to a fixed template in accordance with the specification. Through unified mapping, a one-to-one correspondence between policy variables and device semantics is ensured, laying the foundation for reasonable rule modification in the later stage.

[0162] To ensure that the instructions meet various constraints such as minimum start / stop time, ramp rate, cross-sectional limits, voltage compliance range, and demand, the given instruction vector is used. As a reference value, the final issueable command is calculated based on the deterministic correction model constrained by the metric matrix. :

[0163]

[0164] Among them: final instructions , column vector, and Same dimension, representing the distribution volume after compliance correction; metric matrix A symmetric positive definite matrix used for component weighting, with elements being positive real numbers, set according to device cost and adjustment sensitivity; an inequality matrix. With upper bound vector Linear constraint set, encoding minimum start / stop time, ramp rate, process time window, and upper limit of demand; second-order cone parameters. This is used to represent the L2 norm constraints of cross-sectional power flow and nodal voltage, and is generated by linearization mapping and sensitivity block generation in step two; superscript Indicates transpose and linear form;

[0165] This type of problem is solved using a quadratic-second-order cone solver to obtain the optimal solution and the corresponding set of active constraints. Then, all constraints in the active constraint set of high-risk terms (i.e., rule violations) are recorded in the violation details page, and a predetermined execution sequence—reserved step size and delayed execution steps—is added to the execution plan table. Finally, the sample information for the next frame, the execution plan table, and the receipt template are output, and the solver type, tolerance, and current set of active constraints are added to the version file archive.

[0166] The compliance page displays a bar chart showing instruction deviations and a list of active constraints. The interface page verifies that the downlink frame sample has passed inspection. The solver uses either the interior-point method or the active set method, with a fixed tolerance; the derivative is derived from the sensitivity generation process in step two. Deterministic correction: The strategy output is corrected deterministically to ensure consistent distribution, allowing deviations within a certain range.

[0167] Step 5: Issue and receive receipts to the field devices according to the instruction, aggregate the deviations according to the fixed window and compare them with the list of active constraints. Replace the original guaranteed operation curve set under the premise of meeting the joint threshold, and write back the replacement event, cause and distribution offset information to the digital twin and strategy layer, update the feasible scope and strategy weight, and generate governance reports and comparison records of key performance indicators. Finally, archive them according to object identifier and version number.

[0168] Even after strategy compilation and compliance adjustments, the feasibility of operation still needs to be verified in the power grid and equipment. Without quantitative tracking criteria or closed-loop conditions that trigger rollback, the closed loop lacks boundary control and is prone to dead loops.

[0169] Therefore, by focusing on a single chain encompassing instruction execution, receipt aggregation, anomaly detection, and switchover write-back, the connection between execution and security boundaries is made tighter, ultimately culminating in the instruction vector. As input, use the receipt measurement vector Active constraint set Set of guaranteed operating curves This serves as the basis for verification and switching.

[0170] On-site, the final instruction vector is determined according to the execution plan. After distribution, collect the measurement vectors on the receipt according to the device diameter. and with the final instruction vector To align the data and avoid occasional jitter from interfering with the interpretation, an exponential memory aggregation tracking metric is introduced to weight and summarize the deviations within the nearest window.

[0171]

[0172] Among them: closed-loop tracking indicators Non-negative real numbers are used for quantization. Weighted bias within a given time period; window length Positive integer, determined by the granularity of the execution plan table; memory factor value range When the value is closer to 1, it emphasizes long-term information; metric matrix A symmetric positive definite matrix, set as a constant or piecewise constant according to the adjustment costs and error tolerance of the three channels of generator, energy storage, and load; the final command vector. Column vectors, each corresponding one-to-one with the sent frames; receipt measurement vectors. , column vector, and Consistent dimensions, aligned according to device diameter;

[0173] L2 norm weighted operator Defined as The first operation is the sum of the squared errors of a vector by weights; the second norm weighted operator refers to the operation of summing the squared errors of a vector by weights, denoted as... .in: These are real vectors of the same dimension (such as the difference between instructions and acknowledgments, projection bias, etc.). To and Symmetric positive definite weight matrices of the same dimension; when When it degenerates to the square of the ordinary second norm; when Time equals It can be set according to channel importance, unit dimension, or cost coefficient. This operator is used in the scheme to track closed-loop indicators and compliance correction objective functions, ensuring consistent metrics, convex objectives, and direct solvability by readily available quadratic programming / second-order cone solvers.

[0174] The closed-loop tracking indicators are continuously calculated on the receipt verification page. And bind the device-channel-time index; when the closed-loop tracking indicator When the deviation increases and its source is concentrated in a certain device, a list of deviation sources is generated and pushed to the on-duty engineer; if the closed-loop tracking indicator... If the price falls rapidly within the window, it indicates that the fluctuation has converged.

[0175] The deviation curve and device attribution diagram can be viewed on the receipt verification page, and the window length can be written in the version file. Memory factors With metric matrix The configuration uses weighted aggregation of stable interpretation results with limited memory to distinguish between occasional noise and persistent mismatch. Based on aggregation tracking, it combines three types of signals—distribution offset, active constraints, and acknowledgment timeout—to form a backoff switch variable. Joint threshold criterion; when the switch variable is deactivated When selecting one option, immediately switch to the guaranteed operating curve set. The corresponding entries are then used to structure and write back the events to steps two and three:

[0176]

[0177] Wherein: Back off switch variable The value is or This determines whether to switch to a guaranteed minimum; threshold. , Positive real numbers, defined by the operation and maintenance procedures; distribution offset. The relative entropy metric calculated on the distribution monitoring page is compared with the current disturbance statistics and the calibration period statistics; active constraint set. From the projection operator Active constraint list; Over-limit margin , on constraints The out-of-limit range, a positive value indicates that the limit has been exceeded; indicator function Take when the condition is met Otherwise take .

[0178] When the back switch variable When the value is 1, the system switches to the guaranteed minimum operating curve set on the operating mode page. A switching event ticket is generated on the alarm and pending page. At the same time, the out-of-limit constraint-deviation source-distribution offset is written as an event triple into the version file for step two to update the calibration residual vector. Step 3 updates the tail risk weights; if the rollback switch variable... If the value is zero, it will only be recorded on the receipt verification page.

[0179] When in use, for on-duty engineers, the usage instructions page should clearly indicate the guaranteed activation status and that the version file should be searchable for event triplets, etc. Measurement conditions should be uniformly set according to synchronization rules to ensure that network latency is within an acceptable range and alarms are triggered according to procedures. By combining thresholds, execution deviation, distribution offset, and over-limit signals are unified into a binary decision, which facilitates the judgment and confirmation of switching actions and has a write-back function.

[0180] The operational objectives include multiple dimensions such as cost, carbon, lifespan, and safety, and different dimensions under different circumstances need to be calibrated based on facts. Without establishing a comparison and adjusting the weights, the strategy cannot reach a steady state with the environment.

[0181] Therefore, a parallel comparison-weight adjustment-indicator governance-source tracing and archiving chain ensures that strategies and governance rules mutually check and balance each other and evolve together. The main strategy output... With shadow strategy output The parallel receipts are used as input, with key performance indicator vectors. For accounting purposes.

[0182] Output the shadow policy. Non-issued reference instructions are generated using the same compilation and compliance process as the main strategy. Hypothetical receipt metrics are calculated on the receipt verification page to create a consistent comparison with the main strategy. Weighting variables are introduced. Reflecting the degree of confidence in the shadow strategy, and adjusting it with mirror updates:

[0183]

[0184] Where: weight variables value range This is used for the interlayer influence matrix in step three. Medium-weighted shadow correction; step size Positive real number, adjustment range; shadow strategy indicator vector Main strategy indicator vector A column vector of the same dimension, with components in a fixed order: energy cost - carbon emission compliance - safety margin - equipment health deterioration - wind and solar curtailment suppression; a weighted vector of indicators. A non-negative column vector, weighted according to corporate objectives; function Crop the input to Interval operators are of the form: ; superscript This indicates transpose.

[0185] Display the main-image control bar on the comparison evaluation page, and update the weight variables according to the above formula. When weight variables If the value remains high for several consecutive periods, the system generates a suggestion on the strategy page to include it as a candidate, which will be used in step three for the next round of inter-layer influence matrix. Add a shadow component to the weight variable; If a value remains low for an extended period, it will be marked as a discarded reference.

[0186] The comparison evaluation page displays weight curves and component source labels, which help explain the strategy evolution path; the measurement conditions are the same metric and the same execution plan. Interpretable weight adjustments directly translate the comparison results into visible adjustments to the strategy structure for the next round.

[0187] To ensure consistent management and auditing standards, key performance indicator vectors are aggregated using exponential smoothing to generate two governance curves: daily and weekly reports. Four categories of metadata—source, threshold, event, and correction—are then traced and archived.

[0188]

[0189] Where: Aggregate index vector ,and Same dimension, used for governance reports; smoothing coefficient value range Configured separately according to daily or weekly reporting standards; Key Performance Indicator Vector Generates indicators in the following order: energy consumption cost, carbon emission compliance, safety margin, equipment health deterioration, and wind and solar curtailment suppression; aggregated index vector from the previous time step. Its initial value is set by the baseline scheme.

[0190] The aggregation results are presented side-by-side in the curve chart on the indicator governance page and processed. Daily or weekly reports are generated based on the format requirements. At the same time, the threshold of the joint threshold criterion, the list of active constraints, the list of deviation sources, the minimum switching event form, and the historical trajectory of the weight variable are written into the archive page as traceability metadata, thereby realizing a traceable link from source to threshold to event to correction. Finally, the data is packaged and sent to the management users in an encapsulated manner.

[0191] When in use, the indicator management page shows continuous curves and consistent definitions, while the archive page allows you to find the corresponding data tables and thresholds by time / device. The fixed exit consolidates the operational facts into a reconcilable management document, ensuring consistency between previous and subsequent steps and using the same terminology and definitions.

[0192] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0193] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0194] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0195] The units described 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 can be selected to achieve the purpose of this embodiment according to actual needs.

[0196] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for source-grid-load-storage routing regulation based on AI control strategy, characterized in that: include, Data from the source side, network side, load side, and storage side are collected, confidence weights are generated according to the confidence level of the measurement points, and the electrical topology is used for encoding and alignment to form a unified system state and dual-scale cache containing confidence weights and topology embedding. The unified system status is input into the digital twin, and the online calibration equipment and network parameters are measured. The time-related feasible range and constraint set are extracted, and the corresponding projection operator and the set of guaranteed operation curves are generated. Within feasible limits, a hierarchical strategy combining reinforcement learning and digital twins is adopted. The slow timescale determines the energy quota and routing direction, while the fast timescale generates corrections. After correction by the projection operator, a unified energy routing vector is obtained. In the digital twin, a unified energy routing vector is simulated and verified in multiple scenarios, compiled into instructions and plans for power output, energy storage charging and discharging, and load shifting, and then corrected by a compliance model to form the final instructions; Based on the final instruction, issue and receive the receipt, and verify according to the constraints and deviation monitoring; when the limit is exceeded or the deviation occurs, switch the set of guaranteed operation curves, and write back the receipt and event to update the feasible range and strategy.

2. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 1, characterized in that: A confidence feature set is established for the measurement points on the source side, grid side, load side, and storage side. Confidence weights are generated based on sensor health, time deviation, packet loss, and anomaly density. The naming of measurement points, units, and sampling periods are also unified. The measurement points are mapped to the electrical topology as a topology embedding, and the status entries corresponding to the node and line positions are output and written to a dual-scale buffer at the minute and second level, as well as source identification information.

3. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 2, characterized in that: Topological edge weights are constructed collaboratively based on line reactance, geographical connectivity, and historical power flow, with the edge weights sourced from rated parameters and operational records. During time alignment, a unified timestamp, unit, and metering caliber are used for the time source, and an alignment log and difference threshold are established. The system status records version number, source identifier, and change description are unified.

4. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 3, characterized in that: In the digital twin, the online calibration equipment and network parameters are compared with the unified system status and on-site measurements; the time-related feasible range and constraint set are extracted, covering cross-sectional limits, node voltage, unit start-up and ramp-up, energy storage power and state of charge, load process time window and carbon budget, and projection operators are generated; the set and operators adopt a unified version number and establish a mapping relationship with the previous input.

5. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 4, characterized in that: When the calibration residual exceeds the set threshold, a set of guaranteed operation curves is generated and the incremental update of the strategy that needs to be triggered is identified. At the same time, the switching reason, trigger time and responsible person identification are recorded. The set of guaranteed operating curves, along with the set of feasible ranges and constraints, are versioned and archived together, including file verification summaries, dependency lists, interface adaptation information, and storage paths.

6. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 5, characterized in that: Within feasible limits, a tiered strategy is adopted, with the slow timescale determining energy quotas and routing directions based on cost, carbon budget deviation, energy storage lifetime, and wind and solar curtailment items. The perturbation correction is constructed using fast timescales, and the candidate action space is limited by a feasible range. After candidate vectors are corrected by projection operators, a unified energy routing vector is formed, and source labels and constraint references are established for simulation.

7. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 6, characterized in that: Tail risk control is introduced when generating disturbance corrections to limit the exposure of high-cost scenarios, and the correction amount is mapped to power output, energy storage power and load time shift channels through inter-layer influence matrix and then superimposed with slow time-scaled results; The matrix determines the non-zero structure based on the unit response bandwidth, available energy storage capacity, and load whitelist, and ensures that the channel order is consistent with the unified system status and records the mapping entries.

8. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 7, characterized in that: In the digital twin, a set of multiple scenarios including weather, load, and communication packet loss is constructed. The unified energy routing vector is simulated and verified, and the component table to be modified is output. The vector is compiled into AGC given, energy storage active and reactive power instructions and load task sheets. An execution plan table and a receipt template are prepared, and a scenario sorting view and source labels are generated on the verification page.

9. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 8, characterized in that: Compliance corrections are performed on the compiled instructions, and the constraints include at least the minimum start and stop time, ramp rate, cross-sectional power flow limit, node voltage range, and demand limit. The system interfaces with energy management, monitoring, and building control to generate downlink message samples and record field mapping relationships, verification methods, message length, and time identifier compliance models and sets to establish corresponding indexes.

10. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 9, characterized in that: Based on the final instruction, collect device feedback, aggregate deviations according to a fixed window and bind them to device specifications; perform consistency verification in conjunction with the active constraint list, generate a list of deviation sources and record the correspondence with the final instruction in the version archive; The window slides across the timescale of the execution plan table and generates a device attribution list and archives information on the receipt verification page by channel output item.

11. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 10, characterized in that: When any trigger condition is met by the joint threshold criterion, the corresponding entry of the guaranteed operation curve set is executed and a switching event ticket is generated; the criterion includes the acknowledgment deviation threshold, the active constraint over-limit flag, and the distribution offset of the disturbance statistics; The event log records the switching time, trigger source, and location index and writes it back to the previous step write-back, which is used to update the calibration residuals and warning bands, and to adjust the risk weights and candidate action space.

12. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 11, characterized in that: Under the parallel comparison framework, key performance indicators are calculated and comparison bars are generated for the main strategy and shadow strategy respectively. Weight variables are updated according to preset weights to adjust the inter-layer influence matrix. Key performance indicators are calculated in the order of energy consumption cost, carbon emission compliance, safety margin, equipment health loss and wind and solar curtailment suppression degree, and source references are established with the unified energy routing vector to generate daily and weekly reports.

13. The method for source-grid-load-storage routing adjustment based on AI control strategy as described in claim 12, characterized in that: Establish archive records with version numbers, sources, and timestamps for the unified system state, feasible range and constraint set, projection operators, unified energy routing vectors, and final instructions; The file contains field mappings, a list of active constraints, source tags, and instruction indexes; when referenced across steps, it is located by object identifier and version number and kept consistent across the policy page, verification page, and interface page, with rollback points set.

Images