A power distribution network batched recovery control method based on cold load starting feature mapping and power supply capacity occupation release

Abstract

The application provides a power distribution network batch recovery control method based on cold load starting feature mapping and power supply capacity occupation release, and belongs to the technical field of power distribution networks. The method collects the operation state and load outage record after fault isolation of the power distribution network, generates a segmented linear cold load starting power curve according to the mapping relationship between the outage duration and the cold load starting feature, calculates the additional power demand and the remaining additional energy demand of each load area to be recovered, releases the temporary power supply capacity when the additional demand decreases to a preset threshold, and enters the next round of power transmission until all the loads are recovered. The method solves the technical problems that the additional power demand of the cold load starting cannot be accurately quantified and dynamically tracked during the batch recovery power transmission of the power distribution network, and that the power supply capacity evaluation is inaccurate, which leads to power transmission delay or network over-limit. The method makes the judgment of power supply capacity occupation and release based on the physically interpretable prediction.

Description

Technical Field

[0001] This invention belongs to the field of power distribution network technology, and specifically relates to a method for batch recovery control of power distribution networks based on cold load start-up feature mapping and power supply capacity occupancy release. Background Technology

[0002] The phased restoration of loads after distribution network fault isolation is a critical control step in ensuring power supply reliability. Existing technologies typically use the normal operating power of the load as the basis for power estimation during restoration, combining static power flow calculations to determine the scale of each restoration batch, and relying on dispatcher experience or fixed delay strategies to determine whether conditions for continued power restoration are met after actual restoration. In scenarios of prolonged power outages, cold loads such as temperature-controlled loads generate power peaks far exceeding normal operating levels at the moment of power restoration. Existing methods lack systematic modeling of the correlation between the cold load initiation power characteristics and the duration of the outage. They cannot predict the peak multiple and duration before restoration, nor can they distinguish between the cold load initiation attenuation process and random load fluctuations in real time after restoration. This leads to a long-term state of either overly conservative or overly aggressive power supply capacity assessments. In other words, existing technologies suffer from the technical problem that the additional power demand for cold load initiation during phased power restoration cannot be accurately quantified and dynamically tracked, resulting in inaccurate power supply capacity assessments and causing power transmission delays or network overreach. Summary of the Invention

[0003] In view of this, the present invention provides a distribution network batch restoration control method based on cold load start-up feature mapping and power supply capacity occupancy release, which can solve the technical problems in the prior art where the additional power demand for cold load start-up during the batch restoration of power supply in the distribution network cannot be accurately quantified and dynamically tracked, leading to inaccurate power supply capacity assessment and resulting in power transmission delays or network overruns.

[0004] This invention is implemented as follows: This invention provides a batch recovery control method for distribution networks based on cold load start-up characteristic mapping and power supply capacity occupancy release, including the following steps:

[0005] Collect data on line connections, switch status, power supply and energy storage equipment operation status, line capacity, node voltage limits and load outage records after the distribution network fault isolation. Divide the power-loss load into multiple load areas to be restored according to the control range of the load switch, and determine the current outage duration and normal operating power of each load area to be restored.

[0006] Determine the planned power supply time for this round, calculate the expected power outage duration for each load area to be restored, and determine the peak multiple, peak duration, and attenuation duration of the cold load based on the expected power outage duration and the pre-established mapping relationship between the cold load start characteristics, thereby generating a piecewise linear cold load start power curve.

[0007] Calculate the additional power demand and remaining additional energy demand of each load area to be restored, and use the graph topology incremental update mechanism and sparse matrix local correction operator to perform power flow differential reconstruction on local branches to update the power flow of the distribution network.

[0008] According to the preset power supply sequence, the load areas to be restored are added to the current power supply batch one by one. The additional power demand and the remaining additional energy demand are calculated and decoupled by using a high-order spatiotemporal tensor product decoupling algorithm. An explicit safety envelope manifold is constructed to check the node voltage and line capacity constraints. The comprehensive load rate is calculated. When the comprehensive load rate is greater than 1, the last added load area to be restored is canceled, and the current power supply batch is determined.

[0009] At the planned power transmission time, a closing command is issued to the load switch corresponding to the power transmission batch in this round, and a power adjustment command is issued to the power source and energy storage equipment. The dynamic power supply capacity occupancy and release state transition network is used to track and predict the dynamic changes in the additional power demand for cold load start-up and the remaining additional energy demand, and update the power and energy available for subsequent power transmission.

[0010] Continuously collect operational data and update the additional power demand and remaining additional energy demand for cold load startup in real time. When the additional power demand and remaining additional energy demand drop to preset thresholds and the distribution network has the conditions to continue power supply, the temporary power supply capacity is re-included in the subsequent power supply availability, and the next round is entered and the above steps are repeated until all loads to be restored are energized or the remaining loads do not have the conditions to be energized.

[0011] The calculation of the expected power outage duration is specifically based on the formula... Determined, in the formula For the current moment, The scheduled power transmission time for this round of operations. Area to be restored The duration of the power outage at the current moment. This indicates the expected duration of the power outage.

[0012] Specifically, the establishment of the cold load start-up characteristic mapping relationship involves using a linear relationship between the cold load start-up characteristics and the power outage duration, reaching a saturation value. , These correspond to the peak multiple of the cold load start-up, the peak duration, and the attenuation duration, respectively. The parameters of the relationship function are corrected based on at least one of the following: ambient temperature, season, power supply period, temperature-controlled load ratio, and building type.

[0013] Specifically, the piecewise linear cold load start-up power curve refers to the curve when the attenuation duration is... At that time, the curve successively includes peak stages. Linear decay stage and normal operation phase When the decay duration When the peak power is 0, the system will directly return to normal operating power after the peak phase ends.

[0014] Specifically, the calculation of the additional power demand and the remaining additional energy demand is as follows: , In the formula For additional power requirements exceeding normal operating power, To experience time Additional electrical energy required from the start of the cold load startup process until its completion.

[0015] Specifically, the graph topology incremental update mechanism and the sparse matrix local correction operator are implemented by maintaining the adjacency list of the network topology graph structure and performing topology refresh only on the affected local subgraphs. The matrix rank-1 correction method is used to locally update the admittance matrix and its triangular decomposition factor. The power flow linearization differential calculation is performed with the corrected local branch power as the boundary condition, and the change in local power is mapped to the change in node voltage and branch current.

[0016] Specifically, the spatiotemporal higher-order tensor product decoupling algorithm constructs a set of multi-zone recovery power curves into a third-order tensor. The three dimensions correspond to the index of the load zone to be restored, the index of the time discrete point, and the index of the cold load start-up characteristic component, respectively. The third-order tensor is decomposed into the product of the time basis matrix, the spatial basis matrix, and the characteristic kernel tensor through higher-order singular value decomposition. The superimposed additional power demand of any candidate batch combination is directly calculated through the linear combination of the basis matrices.

[0017] Specifically, the construction of the explicit safety envelope manifold is achieved by performing a second-order Taylor expansion of the power flow equations of the distribution network at the current operating point, constructing a feasible domain boundary surface in the power injection parameter space that satisfies node voltage constraints and line capacity constraints. The boundary surface is represented by a quadratic algebraic equation. When checking newly added load areas to be restored, the power injection vector is projected onto the algebraic surface, and the power injection point is determined to be inside the feasible domain boundary by solving the algebraic equation.

[0018] Specifically, the comprehensive load factor is the maximum value of three factors: the proportion of additional power demand to available power during the entire cold load start-up process, the proportion of remaining additional energy demand to available energy, and the distribution network operating load factor. .

[0019] The dynamic power supply capacity occupancy release state transition network is constructed based on deep learning. The input layer receives the actual collected load power time series, power supply remaining capacity time series and expected attenuation curve residual time series. After being smoothed by a streaming sliding window, it enters the multi-scale self-attention coding layer, nonlinear state evolution layer and control flow branch layer with embedded graph theory shortest path operator. The output layer outputs the predicted cold load start-up additional power demand and remaining additional energy demand.

[0020] The multi-scale self-attention coding layer consists of three parallel self-attention heads. The attention window lengths of the three attention heads are set to short-scale, medium-scale, and long-scale window values, respectively. The outputs of the three attention heads are concatenated in the channel dimension and then compressed to a unified hidden layer dimension by a linear mapping layer.

[0021] The nonlinear state evolution layer comprises sequentially stacked nonlinear transformation modules, with long-span residual dense jump matrices between each module. The input of each nonlinear transformation module is the previous one. The weighted summation of the outputs of each module is stored in the form of a diagonal sparse matrix with weighted coefficients.

[0022] The loss function of the dynamic power supply capacity occupancy release state transition network is embedded with a state entropy increase dynamic penalty term. The penalty term is constructed based on the thermodynamic model of temperature-controlled load and constrains the entropy growth rate of the predicted additional power sequence on the time axis.

[0023] Among them, the power supply capacity release gating adjustment function is based on the normalized decay rate. The shortest path length output by the shortest path operator and the proportion of additional power demand to available power Calculate the comprehensive gating index When the comprehensive gating index is not lower than the aggressive release threshold, an aggressive release strategy is adopted; when the comprehensive gating index is between the gradual release threshold and the aggressive release threshold, a gradual release strategy is adopted; when the comprehensive gating index is lower than the gradual release threshold, a conservative maintenance strategy is adopted.

[0024] The sliding window length ranges from 16 to 32 sampling points; the attention window lengths for the three attention heads are 4 time steps for the short-scale, 8 time steps for the medium-scale, and 16 time steps for the long-scale, respectively; the memory blocks for the three attention heads are allocated within the layer in a 2:4:8 ratio; there are 4 nonlinear transformation modules; the aggressive release threshold is 0.8, and the gradual release threshold is 0.5; the initial learning rate is... ~ The learning rate decay coefficient is 0.5–0.8. The penalty coefficient is ~ The weighting coefficient of the state entropy increase dynamic penalty term is 0.01 to 0.1; the preset threshold corresponds to the 10% to 20% quantile of the distribution of additional power demand from peak value to 5% to 15% time point.

[0025] This invention establishes a feature mapping relationship between the power outage duration and the peak multiple, peak duration, and attenuation duration of the cold load. Before power restoration, it generates a piecewise linear cold load initiation power curve. After power restoration, it utilizes a deep learning-driven dynamic power supply capacity occupancy and release state transition network to continuously track the cold load initiation attenuation process. Physical thermodynamic constraints are injected into the network training, making the power attenuation prediction physically interpretable. This separates the additional power demand caused by cold load initiation from random load fluctuations, achieving accurate occupancy and timely release of power supply capacity. Before power restoration, this invention transforms the additional demand into a calculable quantity through feature mapping. After power restoration, the state transition network eliminates the uncertainty of the cold load attenuation process. This ensures that the comprehensive load factor verification for each power restoration batch is based on physically reasonable power predictions, avoiding both overly conservative assessments that delay subsequent load restoration and overly aggressive assessments that lead to network overruns. In summary, this invention solves the technical problem mentioned in the background art: the additional power demand for cold load initiation during the phased power restoration process of the distribution network cannot be accurately quantified and dynamically tracked, leading to inaccurate power supply capacity assessments and resulting in power restoration delays or network overruns. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a schematic diagram showing the relationship between cold load start-up characteristics and power outage duration.

[0028] Figure 3 A schematic diagram of the segmented linear cold load starting power curve after power is supplied to the load area to be restored. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0030] like Figure 1 The diagram shown is a flowchart of a batch recovery control method for a distribution network based on cold load start-up feature mapping and power supply capacity occupancy release provided by the present invention. This method includes the following steps:

[0031] S01. Collect the line connection relationship, switch status, power supply and energy storage equipment operation status, line capacity, node voltage limit and load outage records after the distribution network fault isolation; divide the power loss load into multiple load areas to be restored according to the control range of the load switch, and determine the current outage duration and normal operating power of each load area to be restored.

[0032] S02. Based on the switching operations, power supply startup, and energy storage preparations required before this round of power transmission, determine the planned power transmission time for this round; calculate the expected outage duration for each load area to be restored at the planned power transmission time; based on the expected outage duration and the pre-established cold load startup characteristic mapping relationship, determine the cold load startup peak multiple, peak duration, and attenuation duration for each load area to be restored; generate a segmented linear cold load startup power curve for each load area to be restored after power transmission.

[0033] S03. Based on the piecewise linear cold load starting power curves of each load area to be restored, calculate the additional power demand and remaining additional energy demand of each load area to be restored; use the graph topology incremental update mechanism and sparse matrix local correction operator to perform power flow differential reconstruction on the local branches affected by switching actions, and update the power flow of the distribution network.

[0034] S04. According to the preset power supply sequence, add the load areas to be restored to the current power supply batch in sequence; for each load area to be restored added, calculate the additional power demand and remaining additional energy demand of the current power supply batch; use a high-order spatiotemporal tensor product decoupling algorithm to decouple the additional demand of multiple areas; construct an explicit safety envelope manifold of the distribution network boundary voltage and power limit, and check the node voltage and line capacity constraints through the geometric boundary projection of the algebraic manifold; calculate the comprehensive load rate of the current power supply batch; when the comprehensive load rate is greater than 1, cancel the last added load area to be restored, and determine the load areas to be restored that were added before the cancellation as the current power supply batch;

[0035] S05. At the planned power transmission time of this round, a closing command is issued to the load switch corresponding to the power transmission batch of this round, and a power adjustment command is issued to the power source and energy storage equipment; after confirming that the load has been restored to power supply, the normal operating power of the restored load and the additional demand for cold load start-up are included in the current power supply demand; a dynamic power supply capacity occupancy release state transfer network is adopted, and the actual collected load power, power supply remaining capacity and expected attenuation curve residual are input to track and predict the dynamic changes of the additional power demand for cold load start-up and the remaining additional energy demand, and update the power and energy available for subsequent power transmission;

[0036] S06. After this round of power transmission, continuously collect operational data such as actual load power, node voltage, line power flow, power output, and energy storage charge status; based on the output of the dynamic power supply capacity occupancy and release state transfer network, update the additional power demand for cold load startup and the remaining additional energy demand in real time; when the additional power demand and the remaining additional energy demand decrease to the corresponding preset thresholds, and the distribution network has the conditions to continue power transmission, the temporary power supply capacity that is no longer needed will be re-included in the subsequent power transmission available capacity, enter the next round, and repeat steps S02 to S05 until all loads to be restored are energized or the remaining loads do not have the conditions to be energized.

[0037] Specifically, the calculation method for the estimated power outage duration in step S02 is as follows: In the formula, For the current moment, The scheduled power transmission time for this round of operations. Area to be restored The duration of the power outage at the current moment. Area to be restored The estimated duration of the power outage corresponding to the planned power supply time in this round.

[0038] Specifically, the method for establishing the cold load start-up characteristic mapping relationship in step S02 is as follows: the relationship between the cold load start-up characteristic and the power outage duration is established according to the load type, and the data sources include at least one of historical power outage recovery measurement data, load monitoring data, building thermal simulation data, temperature-controlled load simulation data, and field test data; the relationship function is represented by at least one of analytical function, piecewise function, interpolation function, fitting function, or discrete data table; the parameters of the relationship function are corrected according to at least one of ambient temperature, season, power supply period, temperature-controlled load ratio, and building type; the relationship between the cold load start-up characteristic and the power outage duration adopts a linear relationship that reaches a saturation value, specifically: In the formula, The duration of the power outage. This indicates three characteristics: peak load multiple, peak duration, and attenuation duration. , The parameters are for the corresponding relationship. The saturation value corresponding to the cold load start-up characteristics; the cold load start-up peak multiple. Peak duration and decay duration The specific method for determining it is as follows: , , In the formula, , , These are the relationship functions between the peak multiple of cold load start-up, peak duration, attenuation duration, and power outage duration, respectively.

[0039] Specifically, the method for determining the piecewise linear cold load start-up power curve in step S02 is as follows: when the attenuation duration... When the value is greater than 0, the load area to be restored The segmented linear cold load starting power curve after power supply is determined according to the following formula: When hour, ;when hour, ;when hour, In the formula, Area to be restored Normal operating power This refers to the time elapsed since the planned power transmission time in this round, calculated from the planned power transmission time in this round. When the decay duration When the value is 0, the load power will be directly restored to the normal operating power after the peak phase ends; the power curve includes the peak phase, the linear decay phase and the normal operating phase in sequence.

[0040] Specifically, the calculation methods for the additional power demand and the remaining additional energy demand in step S03 are as follows: In the formula, Area to be restored The time elapsed after power is supplied is The additional power demand exceeding normal operating power caused by cold load startup. To experience time Load area to be restored from the start of the cold load start-up process to the end of the cold load start-up process Additional power supply still needed This is the time variable for integration.

[0041] Specifically, the principle of the graph topology incremental update mechanism and the sparse matrix local correction operator in step S03 is as follows: During the batch power transmission process of the distribution network, each switching action only changes the connection state of local branches in the network, and the change in the overall admittance matrix is ​​concentrated in a few node rows and columns adjacent to the switch; the graph topology incremental update mechanism maintains the graph structure adjacency list of the network topology, and only performs topology updates on the affected local subgraphs after each switching action, identifying the set of local branches where electrical connections have changed; the sparse matrix local correction operator, based on the set of local branches, uses the matrix rank-1 correction method to locally update the admittance matrix and its triangular decomposition factors, performing forward back substitution elimination only on the affected matrix rows and columns, rather than... The admittance matrix of the entire network is re-decomposed using Gaussian elimination. After the local admittance matrix is ​​corrected, the power flow differential reconstruction uses the corrected local branch power as the boundary condition to perform power flow linearization differential calculation on the affected local branches, mapping the local power change to the changes in node voltage and branch current. This method, while ensuring the accuracy of power flow calculation, compresses the power flow update calculation after each switching action from the entire network scale to the local branch scale, significantly reducing the computational load of admittance matrix reconstruction and power flow solution caused by frequent switching operations during batch power transmission. This allows the distribution network operating status to be updated in a timely manner during continuous power transmission, providing accurate network operation basic data for the comprehensive load factor verification of subsequent power transmission batches.

[0042] Specifically, the method for determining the preset power supply sequence in step S04 is as follows: the preset power supply sequence is determined based on at least one of the following: load importance level, power outage duration, normal operating power, number of power supply users, and manually set sequence; the load areas to be restored are added to the current power supply batch one by one according to the preset power supply sequence; after each load area to be restored is added, the additional power demand, remaining additional energy demand, and comprehensive load rate of the current power supply batch are recalculated; when the comprehensive load rate is greater than 1 after adding the next load area to be restored, the addition of the load area to be restored to the current power supply batch is cancelled, and the load areas to be restored that were added before cancellation are determined as the current power supply batch.

[0043] Specifically, the calculation method for the superposition of the additional power demand and the remaining additional energy demand in this round of power transmission in step S04 is as follows: In the formula, This refers to the set of load areas that have been added to this round of power supply and are awaiting restoration. and These are the time elapsed after the current batch of power transmissions was completed. The total additional power demand and remaining additional energy demand; the maximum duration of cold load start-up in this round of power transmission is determined as the calculation period length: In the formula, This refers to the calculation period length of this round of power transmission batches.

[0044] Specifically, the principle of the spatiotemporal higher-order tensor product decoupling algorithm in step S04 is as follows: When multiple load areas to be restored are simultaneously powered on, the cold load start-up power curves of each load area together constitute a high-dimensional coupled data structure in both the time axis and spatial region dimensions. If all candidate batch combinations are directly enumerated and superimposed, the number of combinations increases exponentially with the number of load areas. The spatiotemporal higher-order tensor product decoupling algorithm constructs the set of multi-area restoration power curves into a third-order tensor, with the three dimensions corresponding to the index of the load area to be restored, the index of the time discrete point, and the index of the cold load start-up feature components, respectively. The third-order tensor is decomposed into the product of the time basis matrix, the spatial basis matrix, and the feature kernel tensor through higher-order singular value decomposition. The time-based matrix extracts the common change pattern of the power curves of each load area on the time axis, and the spatial-based matrix extracts the power superposition coupling relationship between each load area. Under the decoupled basis, the superposition additional power demand of any candidate batch combination can be directly calculated by linear combination of the basis matrices without performing full-time domain power superposition independently for each candidate combination. The method compresses the computational complexity of multi-zone power superposition from an order of magnitude that grows exponentially with the number of combinations to an order of magnitude that grows linearly with the dimension of the basis, effectively avoiding the combination explosion problem in the process of superposition of additional demand in multiple zones, and enabling the comprehensive load rate verification to be completed within a reasonable calculation time when the number of load areas is large.

[0045] Specifically, the construction and verification method of the explicit safety envelope manifold in step S04 is as follows: the explicit safety envelope manifold of the distribution network boundary voltage and power limit is constructed by performing a second-order Taylor expansion of the distribution network power flow equation at the current operating point, and constructing a feasible domain boundary surface in the power injection parameter space that satisfies the node voltage constraint and the line capacity constraint; the boundary surface is represented by a quadratic algebraic equation, and its coefficients are pre-calculated offline and stored by the current network parameters and the Jacobian matrix and Hessian matrix at the operating point; when verifying the newly added load area to be restored, the power injection vector of the current power supply batch is projected onto the algebraic surface, and the algebraic equation is used to verify the boundary surface. The process determines whether the power injection point is located inside the feasible region boundary, thus transforming the original complete power flow calculation, which required multiple iterations for convergence, into a single algebraic projection operation. The verification method continuously uses the established safety envelope manifold when the distribution network topology and operating point do not change significantly, and triggers a local recalculation of the safety envelope manifold after topology switching or operating point offset exceeding a preset range. This method enables the safety verification of node voltage and line capacity to be completed quickly when each candidate load area is added, providing computational efficiency assurance for continuous comprehensive load factor verification during batch power transmission, while ensuring the effective approximation accuracy of the verification results for the actual power flow constraints.

[0046] Specifically, the method for determining the distribution network operating load rate in step S04 is as follows: the distribution network operating load rate is determined by the maximum allowable load amplification ratio of this round of power transmission batches; after power transmission, a certain period of time is elapsed. Under the premise of keeping the distribution network topology and loads outside the current power transmission batch unchanged, the active and reactive power of the current power transmission batch will be proportionally adjusted according to the power factor, while keeping the power factor unchanged. Adjustments are made to determine the operating load factor of the distribution network through power flow calculations: In the formula, This refers to the load power adjustment ratio for this round of power transmission batches. Corresponding to the actual expected load power of this round of power transmission batches, To be based on the proportional coefficient Adjusting the load nodes of this round of power transmission batches voltage amplitude, and These are the lower and upper limits of the allowable node voltage, respectively. For the line Apparent power For the line Permissible transport capacity and These are the set of energized nodes and the set of operational lines after this round of power transmission, respectively; the overall load factor of this round of power transmission is determined according to the following formula: In the formula, and The time elapsed after power is supplied It can be used to handle the additional power and energy required for cold load startup. The comprehensive load factor for this round of power transmission is the maximum value of the following three factors: the proportion of additional power demand to available power, the proportion of remaining additional energy demand to available energy, and the operating load factor of the distribution network during the entire cold load start-up process of this round.

[0047] The power available to meet the additional demand for cold load startup is determined by deducting the expected total power of the restored load at the corresponding time, the normal operating power of the current power transmission batch, and the reserved power, based on the available output of the power source and energy storage equipment. The energy available to meet the additional demand for cold load startup is determined by deducting the electrical energy that the power source can provide and the electrical energy that the energy storage equipment can release during the calculation period, as well as the electrical energy required for the normal operation of the restored load and the current power transmission batch, the additional electrical energy still required by the previous power transmission batch, and the reserved electrical energy. When the available capacity of the power source or energy storage equipment, the load power, or the additional demand for cold load startup changes, the available power and available energy are recalculated.

[0048] The specific structure of the dynamic power supply capacity occupancy and release state transition network in steps S05 and S06 is as follows: The dynamic power supply capacity occupancy and release state transition network is constructed based on deep learning. The input layer receives the actual collected load power time series, power supply remaining capacity time series, and expected attenuation curve residual time series. All three input signals are smoothed in the time domain through a streaming sliding window with a length of 16 to 32 sampling points. The feature vectors output by the three sliding windows are merged into a unified input feature tensor after orthogonalization linear transformation in the input layer. The weight matrix of the orthogonalization linear transformation is orthogonally allocated between neurons during the training phase through Schmitt orthogonalization constraints to ensure that the three heterogeneous features do not generate redundant coupling in the input layer, and the gradient channels of each input feature are independent of each other; the input features The feature tensor enters a multi-scale self-attention encoding layer, which consists of three parallel self-attention heads. The attention window lengths of the three attention heads are set to 4 time steps for short-scale, 8 time steps for medium-scale, and 16 time steps for long-scale, respectively. The memory blocks for the query matrix, key matrix, and value matrix of the three attention heads are allocated within the layer according to the window length in a ratio of 2:4:8. The outputs of the three attention heads are concatenated in the channel dimension and then compressed to a unified hidden layer dimension through a linear mapping layer. The hidden layer features enter a nonlinear state evolution layer, which contains four sequentially stacked nonlinear transformation modules. Each nonlinear transformation module contains a fully connected layer and a nonlinear activation function. A long-span residual dense jump matrix is ​​set between the nonlinear transformation modules, where the first... The input of each nonlinear transformation module is the previous one. The weighted summation of the outputs of each module, with the weighting coefficient matrix stored in a diagonal sparse matrix form between layers to reduce memory bandwidth consumption during data loops between layers, and the long-span residual dense jump matrix used to retain the accumulated energy occupancy traces generated by historical batch power supply, enabling the network to perceive the accumulated additional energy state of multiple historical batches in the current prediction; the output of the nonlinear state evolution layer enters the control flow branch layer with an embedded graph theory shortest path operator, which uses the current batch power supply node set and the power supply capacity release node set as graph vertices and the energy transfer path between nodes as graph edges, calculates the shortest energy transfer path traversed by the current power supply capacity from the occupancy state to the release state using the Dijkstra algorithm, and injects the shortest path length as an additional feature into the gating unit of the control flow branch layer; the gating unit calculates the shortest energy transfer path traversed by the current power supply capacity from the occupancy state to the release state based on the shortest path length and the actual The actual power decay rate triggers conditional branch jumps: if the actual power decay rate is higher than expected, the gating unit issues an energy release signal and jumps to the new batch capacity calculation subnet; if a power rebound occurs, the hidden layer state backtracking is triggered and the subsequent recovery scale is corrected; the output layer outputs the predicted additional power demand for cold load startup and the remaining additional energy demand, which are directly used for the dynamic occupancy and release judgment of power supply capacity in step S06; the loss function of the dynamic power supply capacity occupancy and release state transition network embeds a state entropy increase dynamic penalty term, which is constructed based on the thermodynamic model of temperature-controlled load, constrains the entropy growth rate of the predicted additional power sequence on the time axis, so that the power decay process output by the network conforms to the physical thermodynamic law, effectively removes the pseudo-peak components caused by environmental temperature disturbances in the actual load power, and enhances the robustness of state prediction.

[0049] The steps for establishing the training dataset for the dynamic power supply capacity occupancy release state transition network specifically include: extracting load power acquisition sequences, power supply remaining capacity records, and temperature-controlled load distribution data for each power outage event from historical power outage recovery events in the distribution network; performing stratified sampling of the historical data according to the power outage duration interval to ensure that the training dataset covers the power outage duration range of 1 to 720 minutes; for each historical record, using the actually recorded load power time series after power restoration as the output label, and using the average load power before power restoration, the power supply remaining capacity, and the residual of the expected attenuation curve generated based on the thermal simulation model as input features to construct an input-output pair; for power outage duration intervals with insufficient historical data, using building thermal simulation models and temperature-controlled load simulation models to supplement and generate simulation data for the corresponding intervals, with simulation parameters based on... The local ambient temperature statistical distribution, building type ratio, and temperature control load density survey data were set. The training dataset was divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The training set was used for parameter updates, the validation set for hyperparameter adjustment, and the test set for final performance evaluation. The preset threshold was determined by statistical analysis of the additional power demand decline curve on the validation set. Specifically, the additional power demand time series of all samples in the validation set were normalized, and the distribution of additional power demand from peak to 5% to 15% was calculated. The additional power demand ratio corresponding to the 10% to 20% quantile of the distribution was used as the preset threshold for triggering power supply capacity release. The quantile range was determined by multiple iterative experiments on the validation set, with the comprehensive index of load recovery efficiency and the number of times the distribution network exceeded its limits as the optimization objective.

[0050] The specific steps of training the dynamic power supply capacity occupancy and release state transition network include: updating the network parameters using an adaptive moment estimation optimization algorithm, with the initial learning rate set at... ~ Within the training range; a step-decay strategy for the learning rate is adopted during training, with the learning rate multiplied by a decay factor of 0.5 to 0.8 every 50 to 100 training epochs; independent weight decay factors are applied to the query matrix, key matrix, and value matrix of the three attention heads in the multi-scale self-attention encoding layer. The weight decay factors of the three attention heads are set according to the window length, following the principle of strong constraints for short scales and weak constraints for long scales. ~ Within the range; apply sparsity constraints to the diagonal sparse weight coefficients of the long-span residual dense jump matrix by adding to the loss function. The norm penalty term controls the amount of historical batch cumulative energy usage traces retained. The penalty coefficient is determined by performing a grid search on the validation set with the goal of balancing prediction accuracy and sparsity, and the search range is set to... ~ The weight coefficients of the state entropy increase dynamic penalty term are determined by a comprehensive evaluation of the pseudo-peak removal effect and the overall prediction error on the validation set. The search range is set between 0.01 and 0.1. Each experiment uses the weighted sum of the root mean square error and the number of misjudged pseudo-peaks as the evaluation index. The final weight coefficients are determined through 20 to 50 iterations. The network training is terminated when the validation set loss no longer decreases significantly for 10 to 20 consecutive cycles. The network parameters corresponding to the lowest validation set loss are saved as the final deployment parameters.

[0051] Specifically, a power supply capacity release gating function is designed for the dynamic power supply capacity occupancy and release state transition network. This gating function calculates a comprehensive gating index based on three inputs: the actual power attenuation rate, the shortest path length, and the proportion of additional power demand to available power. , In the formula, The normalized decay rate is the additional power demand within the current sampling window. The power supply capacity output by the shortest path operator is transferred to the shortest path length. The proportion of current additional power demand to available power. , , These are weighting coefficients, and the sum of the three is 1. The weighting coefficients are determined through iterative experiments on the validation set, targeting the accuracy of power supply release timing, with 30–60 iterations. When this happens, the gating unit adopts an aggressive release strategy, including temporarily reserved power supply capacity that is no longer needed in the subsequent power supply availability, and triggers the calculation of the next round of power supply batches; when At this time, the gating unit adopts a gradual release strategy, according to the current... The difference between the value and 0.5 linearly releases a portion of the temporary power supply capacity and maintains continuous monitoring of additional power demands; when At this time, the gating unit adopts a conservative maintenance strategy, temporarily not releasing the temporary power supply capacity, and continues to wait for the status update of the next sampling cycle.

[0052] The dynamic power supply capacity occupancy and release state transition network continuously tracks the occupancy and release process of power supply capacity, separating the random fluctuations of actual load power from the physical laws of the cold load start-up attenuation process. This enables the distribution network to accurately determine the amount of temporary power supply capacity that can be released at each moment during continuous batch power supply, avoiding delays in subsequent load recovery due to overly conservative power supply capacity assessment, and preventing network overreach due to aggressive assessment. The multi-scale self-attention mechanism enables the network to simultaneously perceive short-term power fluctuations and long-term attenuation trends. The long-span residual dense jump matrix retains the cumulative energy occupancy state of multiple historical batches, ensuring prediction continuity in multiple batch continuous recovery scenarios. The embedded state entropy increase dynamic penalty term injects the thermophysical process constraints of temperature-controlled loads into the network training, making the power attenuation prediction output by the network physically interpretable. This enhances the ability to identify load pseudo-peaks caused by environmental disturbances and improves the accuracy and robustness of power supply capacity release judgment.

[0053] Specifically, the method for determining whether the distribution network has the conditions to continue supplying power in step S06 is as follows: at the current sampling time, when the additional power demand for starting the cold load of the current power supply batch drops below a preset threshold, and the remaining additional energy demand of the current power supply batch drops below the corresponding preset threshold, and at the same time the power supply and energy storage equipment have sufficient remaining output and releaseable energy to support the next batch of power supply, it is determined that the distribution network has the conditions to continue supplying power; the source and acquisition method of the preset threshold are detailed in the step of establishing the training dataset of the dynamic power supply capacity occupancy and release state transition network.

[0054] Optionally, the present invention also provides a computer-based distribution network batch restoration control system. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are executed in the computer, they call the status acquisition module to collect the distribution network operating status and load outage records, call the cold load start-up calculation module to perform cold load start-up feature mapping and additional demand calculation, call the power supply batch determination module to complete the comprehensive load rate verification and batch construction, and call the power supply control module to perform power supply operation and drive the dynamic power supply capacity occupancy release state transfer network to dynamically update and release power supply capacity, thereby executing the above-mentioned distribution network batch restoration control method.

[0055] The specific implementation of step S01 is as follows: After the distribution network fault isolation is completed, the current network line connection relationship, the on / off status of each switch, the power output and energy storage device charge status, the thermal stability capacity of each line, and the upper and lower limits of node voltage are collected through the distribution automation system. The outage start time records for each outage section are also obtained from the load management system. Based on the control range of the load switches, the power outage area is divided into several load areas to be restored. Each load area to be restored corresponds to a set of power consumption nodes controlled by the same switch. For each load area to be restored, the current outage duration is calculated based on the difference between the current time and the outage start time, and the average active power within the most recent statistical period before the outage is used as the normal operating power. The above data provides the basic input for subsequent cold load start-up characteristic calculation and comprehensive load rate verification.

[0056] The specific implementation of step S02 is as follows: Based on the number of switching operations to be completed before this round of power transmission, the power supply startup preparation time, and the energy storage device pre-charge and discharge preparation time, the planned power transmission time for this round is comprehensively determined. And according to the formula Calculate the estimated outage duration for each load area to be restored during the planned power restoration time in this round. Substitute the estimated outage duration into the pre-established cold load start-up characteristic mapping relationship, and then apply the relationship function. Calculate the peak starting multiple of the cold load for each load area to be restored. Peak duration and decay duration The parameters of the above mapping relationship are determined by fitting historical power outage recovery measurement data, building thermal simulation data, and temperature-controlled load simulation data, and can be corrected according to factors such as ambient temperature and season. After obtaining the start-up characteristics of the three types of cold loads, piecewise linear start-up power curves for each load area to be restored are generated based on the piecewise linear model. The curves sequentially include the peak stage, the linear decay stage, and the normal operation stage. When the decay duration is 0, the power directly recovers to the normal operation stage after the peak stage ends, and there is no linear decay stage.

[0057] The specific implementation of step S03 is as follows: Based on the piecewise linear cold load start-up power curve, calculate the additional power demand exceeding the normal operating power for each load area to be restored. And the remaining additional energy demand from the current elapsed time until the cooling load starts and ends. The latter is obtained by integrating the additional power demand over time. For power flow updates, a graph topology incremental update mechanism and a sparse matrix local correction operator are employed. After each switching action, by maintaining the adjacency list of the network topology graph structure, topology updates are performed only on the local subgraphs affected by the switching state changes, identifying the set of local branches where electrical connections have changed, avoiding repeated traversal of the entire network topology. After the set of local branches is determined, a matrix rank-1 correction method is used to locally update the admittance matrix and its triangular decomposition factors, performing forward-backward substitution elimination only on the affected rows and columns, rather than re-performing a complete Gaussian elimination decomposition of the entire network admittance matrix. Subsequently, using the corrected local branch power as boundary conditions, power flow linearization differential calculations are performed on the affected local branches, mapping the local power change to the changes in node voltage and branch current, thus completing the power flow update.

[0058] The specific implementation of step S04 is as follows: According to the preset power restoration sequence, the load areas to be restored are added one by one to the current power restoration batch. The preset power restoration sequence is based on a comprehensive ranking of factors such as load importance level, power outage duration, and number of power supply users. For each load area to be restored added, the total additional power demand for this round of power restoration is calculated using an additive method. and additional energy demand The maximum value of the cold load start-up duration for each load zone is taken as the length of the calculation period. For the superposition of additional demand from multiple zones, a spatiotemporal high-order tensor product decoupling algorithm is adopted. The set of power recovery curves from multiple zones is constructed as a third-order tensor, which is decomposed into a time basis matrix, a spatial basis matrix, and an eigenvalue tensor through high-order singular value decomposition. The superposition demand of any candidate batch combination can be directly obtained through a linear combination of the basis matrices, without the need to perform full-time domain superposition independently for each candidate combination, effectively avoiding combination explosion. In terms of safety verification, based on the Jacobian matrix and Hessian matrix at the current operating point, a second-order Taylor expansion of the power flow equation is performed to construct an explicit safety envelope manifold in the power injection parameter space that satisfies the node voltage constraint and the line capacity constraint, and the boundary of the power flow feasible domain is represented as a quadratic algebraic surface. For each newly added load zone to be restored, the power injection vector of the current power supply batch is projected onto this algebraic surface, and the constraint is determined by solving the algebraic equation, without the need for multiple iterations of the complete power flow calculation. The comprehensive load factor is the maximum value of the ratio of additional power demand to available power, the ratio of remaining additional energy demand to available energy, and the distribution network operating load factor over the entire calculation period. When the overall load factor exceeds 1, the last added load area to be restored is cancelled, and the set that was added before cancellation is determined as the current power supply batch.

[0059] The specific implementation of step S05 is as follows: At the planned power transmission time of this round, a closing command is issued to the load switch corresponding to the power transmission batch of this round, and at the same time, a power adjustment command is issued to the power source and energy storage equipment to include the normal operating power of the restored load and the additional demand for cold load start-up in the current power supply demand. Subsequently, the dynamic power supply capacity occupancy release state transition network is started, and the actual collected load power time series, power supply remaining capacity time series and expected attenuation curve residual time series are input. After smoothing by a streaming sliding window, the data is passed through a multi-scale self-attention coding layer, a nonlinear state evolution layer and a control flow branch layer with embedded graph theory shortest path operator in sequence, and the predicted additional power demand for cold load start-up and the remaining additional energy demand are output, and the power and energy available for subsequent power transmission are updated in real time.

[0060] The specific implementation of step S06 is as follows: After the current power transmission, continuous collection of operational data such as actual load power, node voltage, line power flow, power output, and energy storage charge status is performed. This data is then input into the dynamic power supply capacity occupancy and release state transition network in real time to update the predicted values ​​of additional power demand for cold load startup and remaining additional energy demand. When the additional power demand drops below a preset threshold, the remaining additional energy demand drops below the corresponding preset threshold, and the power supply and energy storage equipment have sufficient remaining output to support the next batch of power transmission, the gating unit, based on the comprehensive gating index... Implement a power supply capacity release strategy. A preset threshold corresponds to the quantile of the additional power demand decreasing from its peak to 5%–15%, specifically within the quantile range of 10%–20%. When the comprehensive threshold index is not lower than 0.8, an aggressive release strategy is implemented, immediately adding the temporary power supply capacity that is no longer needed after updates to the available power supply capacity. When the comprehensive threshold index is between 0.5 and 0.8, a gradual release strategy is implemented, releasing a portion of the temporary power supply capacity linearly and proportionally. When the comprehensive threshold index is lower than 0.5, a conservative maintenance strategy is implemented, temporarily not releasing temporary power supply capacity and waiting for the next sampling period update. After the release conditions are met, proceed to the next round and repeat steps S02 to S05 until all loads to be restored are energized or the remaining loads are no longer energized.

[0061] It should be noted that the key technologies of this invention include: a cold load start-up characteristic mapping relationship, which establishes a saturated linear mapping between the power outage duration and the peak multiple, the peak duration and the attenuation duration, transforming the physical laws of thermodynamic relaxation of temperature-controlled loads into a calculable quantitative model, giving a physical basis for the prediction of additional power demand before power transmission, and making up for the fundamental defect of traditional methods that ignore the dynamic characteristics of cold load start-up; a combination of a spatiotemporal high-order tensor product decoupling algorithm and an explicit safety envelope manifold, the former compressing the exponential combination search of multi-zone power curve superposition into a linear combination of basis matrices, and the latter transforming multiple iterative power flow verifications into a single algebraic projection operation, the two working together to enable the batch construction process to be completed efficiently even when there are many load zones; and a dynamic power supply capacity occupancy release state transition network, which perceives short-term fluctuations and long-term attenuation trends through a multi-scale self-attention mechanism, and injects thermodynamic physical constraints into the network training through a state entropy increase dynamic penalty term, realizing the effective separation of the cold load start-up attenuation process and random fluctuations, providing a reliable judgment basis for the accurate release of power supply capacity. The three key technologies work together to form a closed-loop control system from prediction and verification to tracking, enabling the power supply capacity management during the phased restoration of the distribution network to simultaneously meet the requirements of physical interpretability and algorithmic efficiency.

[0062] It should be noted that in the scenario of large-scale power outage restoration in the distribution network, when there are many load areas to be restored and the cold load starting power curves of each area have complex overlap on the time axis, the number of candidate combinations for power transmission batches increases exponentially with the number of load areas. It is unacceptable in terms of calculation time to independently perform full-time domain power superposition and power flow verification for each candidate combination.

[0063] The above-mentioned technical problems arise because the time span of the cold load start-up power curve in the peak and decay phases varies depending on the duration of the power outage in each load area. The superposition relationship of the curves of different load areas on the time axis cannot be eliminated by simple parallelization. Furthermore, after each new load area is added, the entire time domain needs to be re-integrated and the power flow is checked. This causes the computational burden of the batch construction process to accumulate linearly with the number of candidate load areas, while the combination space expands exponentially with the number of candidate load areas. Both of these factors together cause the computation time to deteriorate sharply as the scale increases.

[0064] To address the aforementioned technical issues, the usual solution is to employ greedy algorithms or heuristic pruning to eliminate some candidate combinations in advance, thereby reducing computational load. However, greedy strategies rely on single-step optimal judgment, which can easily miss globally feasible batch combinations when there is time overlap in the cold load start-up curves. Heuristic pruning, on the other hand, relies on empirical thresholds, and its accuracy is difficult to guarantee when there are significant differences in power characteristics across load areas. Ultimately, it may still miss constrained feasible batch combinations or generate power transmission schemes that exceed constraints.

[0065] This invention effectively solves this technical problem. The spatiotemporal high-order tensor product decoupling algorithm constructs a third-order tensor from the set of multi-zone recovery power curves. The three dimensions correspond to the load zone index, the time discrete point index, and the cold load start-up characteristic component index, respectively. By extracting the time basis matrix and the spatial basis matrix through high-order singular value decomposition, the common change pattern of the power curves of each load zone on the time axis and the inter-zone superposition coupling relationship are separated. Under the decoupling basis, the superimposed additional power demand of any candidate batch combination can be directly reconstructed through a linear combination of the basis matrices, without the need to perform full-time domain integration independently for each candidate combination. The computational complexity is compressed from exponential growth with the number of combinations to linear growth with the basis dimension. At the same time, the explicit safe envelope manifold represents the boundary of the power flow feasible region as a quadratic algebraic surface. The constraint verification for each newly added load zone is transformed into a single algebraic projection operation, completely eliminating the computational overhead of multiple iterations of power flow convergence. The two technologies work together in the batch construction process. The former solves the combined explosion of power superposition, while the latter solves the iterative bottleneck of constraint verification. Together, they ensure that the batch construction process can still be completed within a reasonable calculation time when there are many load areas, making the batch recovery control method of this invention feasible in actual large-scale distribution networks.

[0066] Specifically, the principle of this invention is as follows: This invention solves the aforementioned technical problems by starting from the physical essence of the cold load start-up phenomenon and unifying the pre-power-on estimation and post-power-on dynamic tracking into a closed-loop control framework. The peak power multiple and duration of cold load start-up are essentially determined by the thermodynamic relaxation process of the temperature-controlled load, and the amplitude of thermodynamic relaxation and the power outage duration have a monotonically increasing and saturating relationship. This invention utilizes this physical law to establish a piecewise linear saturated mapping between cold load start-up characteristics and power outage duration for different load types. Before the planned power-on time, the peak multiple, peak duration, and attenuation duration of each load area to be restored can be determined based on the expected power outage duration, thereby generating a piecewise linear cold load start-up power curve. This transforms the originally invisible additional power demand into a calculable deterministic input, providing a quantitative basis for comprehensive load factor verification. At the verification level, this invention introduces a spatiotemporal high-order tensor product decoupling algorithm. Through high-order singular value decomposition, the power recovery curves of multiple zones are decoupled in both time and space, avoiding the combination explosion caused by the exponential growth of candidate batch combinations with the number of load zones. Simultaneously, an explicit safety envelope manifold based on second-order Taylor expansion is constructed, transforming multiple iterative power flow verifications into a single algebraic projection operation, ensuring that the batch construction process can still be completed efficiently in large-scale distribution networks. At the dynamic tracking level after power transmission, this invention constructs a dynamic power supply capacity occupancy and release state transition network based on deep learning. Through a multi-scale self-attention mechanism, it simultaneously perceives short-term power fluctuations and long-term attenuation trends. A long-span residual dense jump matrix retains the cumulative energy occupancy state of multiple historical batches, and a state entropy increase dynamic penalty term based on a thermodynamic model is embedded in the loss function. This ensures that the power attenuation prediction output by the network conforms to the physical attenuation law of temperature-controlled loads, effectively removing spurious peak components caused by environmental temperature disturbances, thereby accurately determining the temporary power supply capacity that can be released at each moment. The above three levels form a complete closed loop in logic: feature mapping provides quantitative prediction before power transmission, batch construction algorithm ensures efficient completion of verification, and state transition network ensures dynamic tracking accuracy after power transmission. The three together support the precise management of power supply capacity of the distribution network during the continuous batch restoration process, making the technical solution physically reasonable, algorithmically feasible, and engineering operable.

[0067] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0068] The specific implementation of step S01 is as follows: Collect the line connection relationships, switch status, power supply and energy storage device operating status, line capacity, node voltage limits, and load outage records after the distribution network fault isolation; divide the power-loss load into multiple load zones to be restored according to the load switch control range; and record the current outage duration of each load zone to be restored. With normal operating power , Number the load area to be restored.

[0069] The specific implementation of step S02 is as follows: based on the planned power transmission time for this round... With the current moment Calculate the estimated outage duration for each load area to be restored:

[0070] ;

[0071] In the formula, Area to be restored The estimated duration of the power outage corresponding to the planned power supply time in this round. Area to be restored The duration of the power outage at the current moment. The scheduled power transmission time for this round of operations. For the current moment, all three units are... The relationship between cold load start-up characteristics and outage duration is linearly related to the peak value, with the peak multiple applied. Peak duration Attenuation duration Three types of features are established separately:

[0072] ;

[0073] In the formula, The duration of the power outage is expressed in units of 1. , The linear slope parameter, with units varying depending on the feature type. And so, when hour Dimensionless, when hour and Units are , For the intercept parameter, when hour Dimensionless, when The unit of time is , For the saturation value of the corresponding feature, the dimensions are... Consistently, the above parameters were obtained by fitting historical measurement data or simulation data, and corrected according to factors such as ambient temperature and season. Characteristics of cold load start-up The relationship function between the power outage duration and the duration of the outage. The characteristic values ​​are determined as follows:

[0074] , , ;

[0075] In the formula, This is a dimensionless multiple of the peak value at the start of the cooling load. Peak duration, in units of , The decay duration, in units of , , , These are the relationships between the peak load multiple, peak duration, decay duration, and outage duration, respectively. At that time, the piecewise linear starting power curve for cold load is as follows:

[0076] when hour, ;

[0077] when hour, ;

[0078] when hour, ;

[0079] In the formula, The time elapsed since the planned power transmission time in this round is calculated from the time of power transmission, in units of... The planned power transmission times for this round correspond to , Area to be restored Time The starting power of the cold load at that time, in units of It sequentially goes through a peak phase, a linear decay phase, and a normal operation phase. When At that time, after the peak phase ends, the power directly recovers to .

[0080] The specific implementation of step S03 is to calculate the additional power demand and the remaining additional energy demand based on the power curve:

[0081] ;

[0082] ;

[0083] In the formula, Area to be restored Time Additional power demand exceeding normal operating power, in units of , To experience time Additional electrical energy required from the start of the cooling load until the end of the cooling load startup period, in units of , The time variable is the integral time variable, and the unit is... A graph topology incremental update mechanism is adopted to maintain the network topology adjacency list. After each switching action, only the affected local subgraph is refreshed to identify the local branch set. The sparse matrix local correction operator updates the admittance matrix and its triangular decomposition factor for the local branch set using the matrix rank-1 correction method. Forward and backward substitution is performed only on the affected rows and columns to map the local power change to the node voltage and branch current change, thus compressing the power flow update calculation after each switching action from the whole network scale to the local branch scale.

[0084] The specific implementation of step S04 is to add the load areas to be restored one by one to the current power supply batch according to the preset power supply sequence, and calculate the additional power demand and remaining additional energy demand of the batch:

[0085] ;

[0086] ;

[0087] In the formula, This refers to the set of load areas awaiting restoration that have already been included in this round of power supply. The time elapsed for this batch of power transmissions Total additional power requirement, in units , The time elapsed for this batch of power transmissions The remaining additional energy requirement, in units of The calculation period length is:

[0088] ;

[0089] In the formula, This refers to the calculation period length of this round of power transmission batches, in units of... The spatiotemporal high-order tensor product decoupling algorithm constructs a third-order tensor for the multi-zone recovery power curve. The three dimensions correspond to the load zone index, the time discrete point index, and the cold load start-up characteristic component index, respectively. Through high-order singular value decomposition, it is decomposed into the product of the time basis matrix, the spatial basis matrix, and the characteristic kernel tensor. This allows the superimposed additional power demand of any candidate batch combination to be directly calculated through a linear combination of the basis matrices, compressing the computational complexity from exponential growth with the number of combinations to linear growth with the basis dimension. The explicit safety envelope manifold constructs a feasible region boundary surface in the power injection parameter space that satisfies node voltage constraints and line capacity constraints by performing a second-order Taylor expansion of the power flow equation at the current operating point. This surface is represented by a quadratic algebraic equation, with coefficients calculated offline using the Jacobian and Hessian matrices and stored. During verification, the power injection vector is projected onto this algebraic surface, and a single algebraic projection operation is used to determine whether the power injection point is located inside the feasible region. The distribution network operating load rate is determined as follows:

[0090] ;

[0091] In the formula, This represents the load power adjustment ratio for this round of power transmission batches, dimensionless. Corresponding to the actual expected load power of this round of power transmission batches, To be based on the proportional coefficient Adjusted nodes Voltage amplitude, in units of , and They are nodes The lower and upper limits of the allowable voltage, in units of , For the line In the proportionality coefficient and the time spent The apparent power below, in units of , For the line Permissible transport capacity, in units of , For the passage of time The load factor of the distribution network at that time is dimensionless. The overall load factor is determined by the following formula:

[0092] ;

[0093] In the formula, For the passage of time The power available to meet the additional demand during cold load startup, in units of , The corresponding available energy is expressed in units of 1. , This represents the overall load factor for this round of power transmission, dimensionless, and the maximum of the three values ​​is taken. All three ratios are dimensionless. At that time, the last added load area to be restored is cancelled, and the load areas that have been added are determined as the current power supply batch.

[0094] The specific implementation of step S05 is as follows: at the planned power supply time of this round, a closing command is issued to the load switch. After confirming the restoration of power supply, the normal operating power and additional demand of the restored load in this round are included in the current power supply demand. The input layer of the dynamic power supply capacity occupancy release state transition network receives the actual load power time series, the remaining power supply capacity time series, and the expected attenuation curve residual time series. The three signals are smoothed in the time domain through a streaming sliding window with a length of 16 to 32 sampling points. The three outputs are merged into a unified input feature tensor through orthogonal linear transformation, and the orthogonal linear transformation weight matrix is ​​used. During the training phase, the following constraints are satisfied through Schmitt orthogonalization:

[0095] ;

[0096] In the formula, The weight matrix is ​​for orthogonalizing linear transformation. Indicates matrix transpose. The identity matrix ensures that the three heterogeneous features do not generate redundant coupling at the input layer. The input feature tensor enters a multi-scale self-attention encoding layer, which consists of three parallel self-attention heads with window lengths of 4, 8, and 16 time steps, respectively. The memory blocks of the three attention heads are allocated in a 2:4:8 ratio. The output is concatenated along the channel dimension and then compressed to a unified hidden layer dimension via linear mapping. The hidden layer features enter a nonlinear state evolution layer, which contains four sequentially stacked nonlinear transformation modules. The input for each module is:

[0097] ;

[0098] In the formula, For the first The input feature vector of a nonlinear transformation module Number the nonlinear transformation module and , Number the historical modules. For the first Module 1 to Module 2 The diagonal sparse jump weight matrix of each module is stored in a diagonal sparse form to reduce memory bandwidth usage during data loops between layers. The activation function is nonlinear. The output of the nonlinear state evolution layer enters the control flow branch layer, which embeds a graph theory shortest path operator. The shortest path operator uses the current batch of power-sending nodes and the set of power-supplying capacity-releasing nodes as the vertices of the graph, and the energy transfer paths between nodes as the edges of the graph. It calculates the length of the shortest energy transfer path using the Dijkstra algorithm. and will As an additional feature injected into the gating unit, the gating adjustment function calculates the comprehensive gating exponent:

[0099] ;

[0100] In the formula, The comprehensive gating index is dimensionless. The normalized decay rate is the additional power demand within the current sampling window, dimensionless. The power supply capacity output by the shortest path operator is transferred to the shortest path length, which is dimensionless. The proportion of current additional power demand to available power, dimensionless. , , The weighting coefficients are and satisfy the following conditions: This was determined through 30 to 60 iterative experiments on the validation set, targeting the accuracy of power supply release timing. When an aggressive release strategy is adopted, all temporary reserved power supply capacity that is no longer needed will be released; when A gradual release strategy is adopted at that time, according to Proportional linear release of temporary power supply capacity; when A conservative maintenance strategy is adopted, and the network is not released for the time being. The network loss function embeds a state entropy increase dynamic penalty term. Based on the construction of a temperature-controlled load thermodynamic model, the entropy growth rate of the predicted additional power sequence on the time axis is constrained, effectively removing the spurious peak components caused by environmental temperature disturbances and enhancing the robustness of state prediction.

[0101] The specific implementation of step S06 is to continuously collect actual load power, node voltage, line power flow, power output, and energy storage state of charge, and update the data in real time according to the output of the dynamic power supply capacity occupancy and release state transition network. and When the additional power demand and the remaining additional energy demand decrease to the corresponding preset thresholds, and the power supply and energy storage equipment have sufficient remaining output, the temporary power supply capacity that is no longer needed will be re-included in the subsequent power supply capacity, and the process will enter the next round and repeat steps S02 to S05 until all loads to be restored are energized or the remaining loads no longer have the conditions for power supply.

[0102] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: This example takes a typical distribution network fault isolation batch load restoration test scenario as the object to illustrate the application process of the control method in solving practical problems.

[0103] The test network consists of 10 nodes and 12 lines, including 3 normally open interconnecting lines. Under normal operation, it is a radial distribution network equipped with one main substation (rated capacity 3200 kVA), two distributed generation sources (rated outputs of 600 kW and 400 kW respectively), and one energy storage device (rated capacity 500 kWh, maximum charge / discharge power 200 kW). The rated current of all lines is 400 A, and the allowable range of node voltage is 93%–107% of the rated voltage. After a fault occurs, isolation operations de-energize 7 load nodes (nodes 4 to 10). Based on the load switch control range, these nodes are divided into 5 load zones to be restored, denoted as... ~ The normal operating power and the current power outage duration of each area are shown in Table 1.

[0104] Table 1 Initial Parameters for Each Load Zone to be Restored

[0105]

[0106] After step S01 is completed, the system confirms the current time. This is 52 minutes after fault isolation. Based on the fault isolation completion status and power startup status, the planned power-on time for this round is determined. for The next 8 minutes, i.e., the estimated duration of the power outage for each load area to be restored. All durations are 60 minutes. The pre-established cold load start-up characteristic mapping relationship is obtained by fitting historical power outage recovery measurement data and building thermal simulation data. For a power outage duration of 60 minutes, the cold load start-up characteristics of each load zone are shown in Table 2. The above characteristics are obtained by calculating a piecewise linear relationship function that satisfies saturation constraints, and the parameters are corrected for the current season (winter) and the proportion of temperature-controlled load (approximately 62%).

[0107] Table 2 Characteristic Parameters of Cold Load Start-up for Each Load Zone to be Restored

[0108]

[0109] like Figure 2 The diagram illustrates the relationship between cold load start-up characteristics and power outage duration. In this embodiment, the cold load start-up peak multiple, peak duration, and decay duration all increase linearly with the power outage duration and approach their respective saturation values. The saturation value for the peak multiple is approximately 1.80, the saturation value for the peak duration is approximately 20 minutes, and the saturation value for the decay duration is approximately 60 minutes. In this test scenario, the parameters under a 60-minute power outage duration are still within the linear growth range and have not yet reached saturation. For example, its peak power is The peak power was 480 kW, lasting for 12 minutes, and then linearly decayed to 480 kW within 38 minutes.

[0110] like Figure 3 The diagram shown is a schematic of the segmented linear cold load start-up power curve after power is supplied to the load area to be restored. In this embodiment, The segmented linear cold load start-up power curve after power transmission sequentially includes a peak phase (0–12 min, power maintained at 772.8 kW), a linear decay phase (12–50 min, power linearly decreasing from 772.8 kW to 480 kW), and a normal operation phase (after 50 min, power stabilizes at 480 kW). Figure 3 It can be seen that the slope of the decay phase is approximately kW / min, compared with the results calculated by the linear decay model The kW / min values ​​are consistent, and the curve shape matches the statistical regularity of historical measurement data.

[0111] After generating the power curves for each load area, proceed to step S03. Based on the graph topology incremental update mechanism and the sparse matrix local correction operator, perform power flow differential reconstruction on the local branches affected by this round of power transmission operations. In the test network, there were a total of 4 switching operations in this round, and the affected local branches were the connection branches from node 4 to node 7, involving a total of 6 branches, accounting for 50% of the total number of branches in the entire network. However, since the coupling interface between this subnet and the main network is only 2 nodes, the admittance matrix correction is concentrated in the corresponding 2 rows and 2 columns. After local correction, only forward back substitution needs to be performed on the 6 local branches, without the need to re-decompose the 10×10 admittance matrix of the entire network, which greatly reduces the power flow calculation load.

[0112] Step S04: Power supply sequence is arranged according to a preset order (based on the duration of the power outage and the number of users). → → → → Load regions are added to the current batch one by one, and the additional demand from multiple regions is decoupled using a high-order spatiotemporal tensor product decoupling algorithm. For the cold load start-up power curves of the five load regions, a third-order tensor with 500 time discrete points and 3 feature components is constructed. After high-order singular value decomposition, the three common patterns extracted from the time basis matrix can cover more than 96% of the total variance. The calculation of the superimposed additional power of candidate combinations in subsequent batch verification can be directly completed under the decoupled basis, without having to perform full-time domain power superposition for each candidate combination.

[0113] After adding load areas sequentially, an explicit safety envelope manifold for the distribution network boundary voltage and power limits is constructed, and algebraic projection verification is performed on each newly added load area. to After adding the load factor, the overall load factors were 0.64, 0.73, and 0.94, respectively, all not exceeding 1, indicating that the network voltage and line capacity both met the constraints. Afterwards, the overall load factor was calculated to be 1.07, which exceeds 1, therefore the operation was cancelled. The addition of [a certain element] determines the current power transmission batch as { , , }. See Table 3 for details of the various comprehensive load factor indicators.

[0114] Table 3. Overall Load Rate Verification Results After Gradually Adding Load Zones

[0115]

[0116] The overall load factor is the maximum of the following three factors: the ratio of additional power demand to available power, the ratio of surplus additional energy demand to available energy, and the operating load factor of the distribution network. During this phase, the distribution network operating load factor (0.94) became a constraint bottleneck, indicating that the main factor restricting the scale of power transmission in this round was the node voltage margin rather than the power supply capacity.

[0117] Step S05: At the planned power transmission time, send to { , , The corresponding load switch issues a closing command, and simultaneously issues force adjustment commands to the power supply and energy storage equipment. After actual power supply, the dynamic power supply capacity occupancy / release state transition network starts operating. The input layer receives the load power time series, the remaining power supply capacity time series, and the expected attenuation curve residual time series processed by three sliding windows (window length 16 sampling points, sampling interval 1 min). After orthogonalization linear transformation, these are merged into an input feature tensor, which is then encoded in a multi-scale self-attention encoding layer (short, medium, and long scale attention windows of 4, 8, and 16 time steps, respectively). Subsequently, the hidden layer features are output through a nonlinear state evolution layer (4 sequentially stacked nonlinear transformation modules, including a long-span residual dense jump matrix). The control flow branch layer, which embeds the graph theory shortest path operator, uses the set of power supply nodes and the set of power supply capacity release nodes as graph vertices to calculate the shortest energy transfer path length for the current power supply capacity to transition from the occupancy state to the release state, and injects it as an additional feature into the gating unit.

[0118] In step S06, 22 minutes after power is supplied, the actual measurement is... , , The load power has decreased to 631 kW, 492 kW, and 430 kW respectively. The total additional power demand has decreased from the peak of 621 kW to 493 kW, and the remaining additional energy demand has decreased from the initial 298 kWh to 83 kWh. The relative error between the predicted additional power demand output by the dynamic power supply capacity occupancy release state transition network and the actual measured value was 4.3% at the 22nd minute, indicating that the network has good accuracy in tracking the power decay process. The network also identified a spurious peak component caused by a temperature disturbance of approximately 28 kW (due to a brief rise in outdoor temperature) superimposed on the actual load power during this period. This spurious peak component was removed from the decay trend through the state entropy increase dynamic penalty term, avoiding misjudging it as an extension of the cold load start-up process.

[0119] Thirty-eight minutes after power supply, the additional power demand decreased to 34 kW, accounting for 8.1% of the available power, which is lower than the preset threshold (approximately 9.5% corresponding to the 10th percentile of the validation set); the remaining additional energy demand decreased to 5 kWh, also lower than the corresponding preset threshold. The comprehensive gating index was calculated using the power supply capacity release gating adjustment function. If the threshold for aggressive power release is exceeded (0.8), the gating unit immediately issues an energy release signal, adding the temporarily unused power capacity, which has been updated and determined to be no longer needed, to the subsequent power supply capacity and triggering the next round of power supply batch calculations. At this point, the total remaining available power is 387kW, and the energy storage device can release 196kWh of energy, which is sufficient to support... and The basic conditions for restoring power supply.

[0120] The next round of power transmission will be and As a candidate, The power outage is expected to last approximately 98 minutes. Approximately 98 minutes later, the peak power multiples are expected to reach 1.72 and 1.70 respectively (due to the extended outage time approaching the saturation zone), with a total peak power of [missing information]. kW. The corresponding comprehensive load factor verification result is 0.91, which meets the power supply conditions, therefore and Together, they constitute the second batch of power supply, and the power supply operation was completed immediately. All lost loads were restored within about 38 minutes after the first batch of power supply.

[0121] The beneficial effects of the above embodiments are reflected in the following aspects. The cold load start-up power monotonically decreases over time after the peak phase ends ( The attenuation slope (approximately 7.71 kW / min) reflects the physical process of the gradual release of heat capacity in temperature-controlled loads. It is this predictable monotonic attenuation characteristic that allows the additional power demand to be incorporated into the overall load factor calculation as a deterministic constraint within a finite time domain. This ensures that the power and energy margin required for cold load startup is accurately reserved during the batch power transmission phase, avoiding the power over-limit problem after power transmission caused by neglecting the cold load startup effect in the background technology. Among the three components of the overall load factor, the distribution network operating load factor (0.94) is the constraint bottleneck after the addition of the third load area, while the additional power ratio (0.78) and additional energy ratio (0.85) have not yet reached the limit boundary. This indicates that in this test scenario, the node voltage constraint tightens earlier than the power supply capacity constraint. If an explicit safety envelope manifold check including node voltage and line capacity is not introduced, and the judgment is based solely on the rated capacity margin of the power supply, it may lead to a misjudgment that the network has the capacity to continue transmitting power, thereby causing a node voltage exceeding the lower limit fault. The dynamic power supply capacity occupancy release state transition network separates the 28 kW spurious peak caused by temperature disturbance from the actual power at the 22nd minute. This makes the input signal of the gating function closer to the physical law of pure cold load start-up attenuation, preventing the spurious peak from delaying the power supply capacity release time. This allows the second round of power supply to be triggered at the 38th minute (instead of the potential 46th minute), enabling... and The power outage duration was reduced by about 8 minutes, and the peak value was reduced by about 0.03 accordingly. This effectively shortened the duration of additional demand for cold load startup and alleviated the positive feedback deterioration trend that the longer the power outage time, the higher the peak value of cold load startup.

[0122] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for batch recovery control of a distribution network based on cold load start-up characteristic mapping and power supply capacity occupancy release, characterized in that, Includes the following steps: Collect data on line connections, switch status, power supply and energy storage equipment operation status, line capacity, node voltage limits and load outage records after the distribution network fault isolation. Divide the power-loss load into multiple load areas to be restored according to the control range of the load switch, and determine the current outage duration and normal operating power of each load area to be restored. Determine the planned power supply time for this round, calculate the expected power outage duration for each load area to be restored, and determine the peak multiple, peak duration, and attenuation duration of the cold load based on the expected power outage duration and the pre-established mapping relationship between the cold load start characteristics, thereby generating a piecewise linear cold load start power curve. Calculate the additional power demand and remaining additional energy demand of each load area to be restored, and use the graph topology incremental update mechanism and sparse matrix local correction operator to perform power flow differential reconstruction on local branches to update the power flow of the distribution network. According to the preset power supply sequence, the load areas to be restored are added to the current power supply batch one by one. The additional power demand and the remaining additional energy demand are calculated and decoupled by using a high-order spatiotemporal tensor product decoupling algorithm. An explicit safety envelope manifold is constructed to check the node voltage and line capacity constraints. The comprehensive load rate is calculated. When the comprehensive load rate is greater than 1, the last added load area to be restored is canceled, and the current power supply batch is determined. At the planned power transmission time, a closing command is issued to the load switch corresponding to the power transmission batch in this round, and a power adjustment command is issued to the power source and energy storage equipment. The dynamic power supply capacity occupancy and release state transition network is used to track and predict the dynamic changes in the additional power demand for cold load start-up and the remaining additional energy demand, and update the power and energy available for subsequent power transmission. Continuously collect operational data and update the additional power demand and remaining additional energy demand for cold load startup in real time. When the additional power demand and remaining additional energy demand drop to preset thresholds and the distribution network has the conditions to continue power supply, the temporary power supply capacity is re-included in the subsequent power supply availability, and the next round is entered and the above steps are repeated until all loads to be restored are energized or the remaining loads do not have the conditions to be energized.

2. The method for phased restoration control of a power distribution network according to claim 1, characterized in that, The graph topology incremental update mechanism and sparse matrix local correction operator are specifically implemented by maintaining the adjacency list of the network topology graph structure and performing topology refresh only on the affected local subgraphs. The matrix rank-1 correction method is used to locally update the admittance matrix and its triangular decomposition factor. The power flow linearization differential calculation is performed with the corrected local branch power as the boundary condition, and the change in local power is mapped to the change in node voltage and branch current.

3. The method for phased restoration control of a power distribution network according to claim 2, characterized in that, The spatiotemporal higher-order tensor product decoupling algorithm specifically constructs a set of multi-zone recovery power curves into a third-order tensor. The three dimensions correspond to the index of the load zone to be restored, the index of the time discrete point, and the index of the cold load start-up characteristic component, respectively. The third-order tensor is decomposed into the product of the time basis matrix, the spatial basis matrix, and the characteristic kernel tensor through higher-order singular value decomposition. The superimposed additional power demand of any candidate batch combination is directly calculated through the linear combination of the basis matrices.

4. The method for phased restoration control of a power distribution network according to claim 3, characterized in that, The construction of the explicit safety envelope manifold is specifically achieved by performing a second-order Taylor expansion of the power flow equations of the distribution network at the current operating point, constructing a feasible domain boundary surface in the power injection parameter space that satisfies node voltage constraints and line capacity constraints. The boundary surface is represented by a quadratic algebraic equation. When checking newly added load areas to be restored, the power injection vector is projected onto the algebraic surface, and the power injection point is determined to be inside the feasible domain boundary by solving the algebraic equation.

5. The method for phased restoration control of a power distribution network according to claim 4, characterized in that, The comprehensive load factor is specifically the maximum value of the following three factors during the entire cold load start-up process: the proportion of additional power demand to available power, the proportion of remaining additional energy demand to available energy, and the distribution network operating load factor.

6. The method for phased restoration control of a power distribution network according to claim 5, characterized in that, The dynamic power supply capacity occupancy and release state transition network is built based on deep learning. The input layer receives the actual collected load power time series, power supply remaining capacity time series and expected decay curve residual time series. After being smoothed by a streaming sliding window, it enters the multi-scale self-attention coding layer, nonlinear state evolution layer and control flow branch layer with embedded graph theory shortest path operator. The output layer outputs the predicted cold load start-up additional power demand and remaining additional energy demand.

7. The method for phased restoration control of a power distribution network according to claim 6, characterized in that, The multi-scale self-attention coding layer consists of three parallel self-attention heads. The attention window lengths of the three attention heads are set to short-scale window value, medium-scale window value, and long-scale window value, respectively. The outputs of the three attention heads are concatenated in the channel dimension and then compressed to a unified hidden layer dimension by a linear mapping layer.

8. The method for phased restoration control of a power distribution network according to claim 7, characterized in that, The nonlinear state evolution layer comprises sequentially stacked nonlinear transformation modules, with long-span residual dense jump matrices set between each nonlinear transformation module. The input of each nonlinear transformation module is the previous one. The weighted summation of the outputs of each module is stored in the form of a diagonal sparse matrix with weighted coefficients.

9. The method for phased restoration control of a power distribution network according to claim 8, characterized in that, The loss function of the dynamic power supply capacity occupancy release state transition network is embedded with a state entropy increase dynamic penalty term. The penalty term is constructed based on the thermodynamic model of temperature-controlled load and constrains the entropy growth rate of the predicted additional power sequence on the time axis.

10. The method for phased restoration control of a power distribution network according to claim 9, characterized in that, The sliding window length is 16 to 32 sampling points; the attention window lengths of the three attention heads are 4 time steps for the short scale, 8 time steps for the medium scale, and 16 time steps for the long scale, respectively; the memory blocks of the three attention heads are allocated within the layer in a ratio of 2:4:8.

Images