Intelligent fusion terminal-based supply-demand matching scheduling method and system
By constructing a multivariate supply and demand state sequence through intelligent fusion terminals and introducing an improved TimeMixer model, the problems of inaccurate supply and demand deviation identification and resource matching in existing scheduling methods are solved. This achieves accuracy in supply and demand deviation prediction and precision in resource scheduling, thereby improving the operational stability and resource utilization of the distribution area.
Patent Information
- Application Number
- CN202611122160.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-08-25
AI Technical Summary
Existing supply and demand scheduling methods are unable to capture in a timely manner rapid changes in supply and demand security margins, fluctuations in photovoltaic output, and sudden increases in charging loads, resulting in delayed response of scheduling strategies, inaccurate resource matching, and insufficient closed-loop updates.
By collecting multi-source operational data through intelligent fusion terminals, constructing a multivariate supply and demand state sequence, introducing risk boundary identification and improving the TimeMixer model, generating a supply and demand deviation prediction sequence, calculating the impact of resource scheduling, and forming a closed-loop scheduling strategy.
It has achieved accuracy in identifying supply and demand discrepancies and precision in resource scheduling, improved edge response speed and resource utilization, and enhanced the operational stability of distribution radio areas.
Smart Images

Figure CN122639166A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart distribution network operation and control technology, and in particular to a supply and demand matching scheduling method and system based on a smart fusion terminal. Background Technology
[0002] As the proportion of distributed photovoltaic systems, energy storage devices, charging piles, and adjustable loads in distribution transformer areas continues to increase, the operating status of these areas is gradually shifting from single-load operation to a complex state involving coordinated changes in the power supply side, power consumption side, and resource side. Existing intelligent converged terminals are now able to collect operational data such as voltage, current, power, load, energy metering, and communication status, and upload the relevant data to the main station or cloud for analysis and processing.
[0003] Existing supply and demand scheduling methods mostly rely on fixed time windows and centralized scheduling logic. They typically determine supply and demand based on real-time load, distributed power generation output, and energy storage status, and then generate scheduling instructions according to preset rules. While these methods can complete routine peak shaving and valley filling and resource allocation, they lack the ability to identify risk phase boundaries in situations such as rapid changes in supply and demand safety margins, local overloads, fluctuations in photovoltaic output, and sudden increases in charging load. They also struggle to capture key moments before and after the formation of supply and demand deviations in a timely manner.
[0004] Existing methods focus on the available capacity or responsiveness of individual resources during resource scheduling, lacking analysis of the propagation of influence between the power supply side, the power consumption side, and adjustable resources. It is difficult to determine the chain reaction of a resource adjustment on subsequent time windows, line power flow, energy storage capacity, and load changes, which leads to scheduling strategies that are prone to problems such as delayed response, inaccurate resource matching, and insufficient closed-loop updates.
[0005] Therefore, how to provide a supply and demand matching and scheduling method and system based on intelligent converged terminals is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] One objective of this invention is to propose a supply and demand matching scheduling method and system based on intelligent fusion terminals. This invention collects multi-source operation data from the power supply side, power consumption side, and adjustable resource side of the distribution transformer area through intelligent fusion terminals. It introduces an improved TimeMixer model with a supply and demand risk phase boundary identification mechanism and a risk modulation topology hybrid mechanism, as well as a method for constructing the supply and demand scheduling influence domain. This enables the prediction of supply and demand deviations, analysis of resource impact propagation, and generation of collaborative scheduling strategies. It has the advantages of accurate supply and demand deviation identification, precise resource scheduling, fast edge response speed, high resource utilization, and good operation stability of the distribution transformer area.
[0007] The supply and demand matching and scheduling method based on intelligent converged terminals according to embodiments of the present invention includes:
[0008] The intelligent fusion terminal collects multi-source operation data from the power supply side, power consumption side and adjustable resource side within the distribution transformer area, performs preprocessing on the multi-source operation data, and generates a supply and demand status set for the transformer area.
[0009] Based on the supply and demand status set of the transformer area, the characteristics of power supply capacity, load demand and resource adjustability are extracted, and a multivariate supply and demand status sequence is constructed according to the time series.
[0010] Calculate the supply and demand security margin and the rate of change of the margin for each time slice, identify the risk phase boundary points of supply and demand, perform risk phase boundary anchoring segment recombination on the multivariate supply and demand state sequence, and generate multi-scale supply and demand state subsequences;
[0011] An improved TimeMixer model is introduced by incorporating multi-scale supply and demand state subsequences into a risk modulation topology mixing mechanism. Within each TimeMixer block, a time-series token-mixing channel, a resource topology attention channel, and a risk gating channel are set in parallel to generate a supply and demand deviation prediction sequence.
[0012] Based on the supply and demand deviation prediction sequence, the propagation relationship of supply and demand influence among the power supply side, power consumption side and adjustable resources is calculated. The supply and demand scheduling influence domain corresponding to the future time window is constructed, the scheduling contribution value of each adjustable resource is calculated, and the resource scheduling influence set is generated.
[0013] Based on the resource scheduling impact set and the supply-demand deviation prediction sequence, a scheduling strategy set is generated for the allocation and adjustment tasks of energy storage devices, adjustable loads, charging piles and distributed power sources. This set is then distributed to the corresponding execution devices at the edge via intelligent fusion terminals. The supply and demand status after execution is collected and compared in real time, and the supply and demand status data set and resource scheduling impact set are updated to form a closed-loop scheduling processing result.
[0014] Optionally, the multi-source operation data includes distribution substation transformer operation data, distributed power output data, energy storage device operation data, user load data, charging pile load data, adjustable load status data, line power flow data, switch status data, electricity metering data, terminal communication status data, and corresponding timestamps.
[0015] Optionally, the step of preprocessing the multi-source operating data to generate a supply and demand status set for the transformer area includes performing time alignment, anomaly removal, and status normalization processing on the multi-source operating data to generate the supply and demand status set for the transformer area.
[0016] Optionally, constructing a multivariate supply and demand state sequence according to the time series includes:
[0017] The supply and demand status set of the transformer area is divided into continuous time slices according to the unified sampling period. Based on the transformer area number, smart fusion terminal number, equipment number and sampling timestamp, the power supply side, power consumption side and adjustable resource side data in the same time slice are collected to generate time slice status records.
[0018] Read the transformer's carrying capacity, line power flow margin, external grid input power, real-time output of distributed power sources, and energy storage discharge capacity from the time slice status record, integrate the fields according to the power source and power carrying object, and generate power supply capacity characteristics.
[0019] Read real-time user load, charging pile load, rigid load, interruptible load and historical load of the same period from the time slice status record, integrate the fields according to load source, load type and load change direction to generate load demand characteristics;
[0020] Read the remaining energy storage capacity, energy storage charging and discharging limits, adjustable load response capacity, charging pile delay power and controllable output range of distributed power sources from the time slice status record, integrate the fields according to resource type, adjustment direction and response status, and generate resource adjustable capability characteristics.
[0021] The power supply capacity characteristics, load demand characteristics, and resource adjustability characteristics within the same time slice are written into the same supply and demand status record in a fixed field order, and multiple supply and demand status records are continuously arranged according to the sampling time order to generate a multivariate supply and demand status sequence.
[0022] Optionally, the adaptive time window reconstruction of the multivariate supply and demand state sequence to obtain a multi-scale supply and demand state sub-sequence includes:
[0023] Read the supply and demand status records in the multivariate supply and demand status sequence according to the sampling time sequence, determine the available power supply capacity from the power supply capacity characteristics corresponding to the same time slice, determine the actual load demand from the load demand characteristics, and determine the supply and demand safety margin of the corresponding time slice by subtracting the actual load demand from the available power supply capacity.
[0024] The supply and demand safety margins of the current time slice and the previous time slice are read according to the sampling order of adjacent time slices. The supply and demand safety margins of the current time slice and the previous time slice are processed by difference. The rate of change of the margin is determined by combining the sampling interval between adjacent time slices.
[0025] Based on the characteristics of supply and demand security margin, margin change rate and resource adjustability in each time slice, the supply and demand risk boundary point is identified. When the supply and demand security margin moves from the security margin range to the critical margin range, from the critical margin range to the imbalance margin range, the margin change rate reaches the preset rate threshold, or the resource adjustability margin is less than the preset margin threshold, the corresponding time slice is marked as the supply and demand risk boundary point.
[0026] Using the supply and demand risk boundary point as the anchor point, extract time slices in front of the anchor point where the supply and demand safety margin is continuously shrinking and the load demand is continuously increasing to form a contraction segment; extract time slices in the neighborhood of the anchor point where the supply and demand safety margin is in the critical margin range and the resource adjustability margin is continuously decreasing to form a critical segment; extract time slices in behind the anchor point where the supply and demand safety margin returns from the imbalance margin range to the safety margin range to form a recovery segment.
[0027] The contracted segments, critical segments, and recovery segments are correlated and recombined according to the sampling time order, phase boundary point number, segment type, and resource adjustability margin. Time slices that have not entered the neighborhood of the supply and demand risk phase boundary point and whose supply and demand safety margin is continuous and stable are retained as stable segments to generate multi-scale supply and demand state subsequences.
[0028] Optionally, generating the supply-demand deviation prediction sequence includes:
[0029] Read stable segments, contraction segments, critical segments, and recovery segments from the multi-scale supply and demand state subsequence, write the segment type, sampling time, resource number, and risk phase boundary number, and add prediction window control flags to generate the risk phase boundary TimeMixer input sequence;
[0030] The TimeMixer input sequence is fed into the temporal token-mixing channel. The temporal token-mixing channel uses a hierarchical expansion mixing module. An expansion step size is used in stable segments to capture long-term dependencies, and a compact step size is used in contracted segments and critical segments to capture short-term details. Reusable caches are inserted between layers to reduce edge inference overhead and output temporal evolution mixing features.
[0031] Based on the connection relationship, power flow relationship and regulation response relationship between the power supply side, power consumption side and adjustable resources, a two-way resource topology relationship table is constructed. The topology relationship table is input into the resource topology attention channel. Inside the channel, a dual-flow direction attention mechanism is used to calculate the correlation strength of supply-demand flow and demand-supply flow respectively. The correlation strength weight is updated in real time using the scheduling result of the previous moment, and the resource topology hybrid feature is output.
[0032] Input the supply and demand security margin, margin change rate, segment type and resource adjustability margin into the risk gating channel. The channel generates modulation weights according to the risk level, performs risk weight weighted fusion on the time evolution hybrid features and resource topology hybrid features, injects the risk level code into the fusion result in the form of residuals, and outputs the risk modulation hybrid features.
[0033] The risk modulation hybrid feature is input into the TimeMixer output mapping layer, and the supply and demand deviation prediction values for the corresponding time slices are generated according to the time slice order of the future time window. The supply and demand deviation prediction values of each time slice are arranged continuously to form a supply and demand deviation prediction sequence.
[0034] Optionally, the calculation of the scheduling contribution value of each adjustable resource and the generation of a resource scheduling impact set includes:
[0035] Read the supply and demand deviation prediction sequence according to the time slice order of the future time window, identify the supply and demand deviation direction, duration of supply and demand deviation and trend of supply and demand deviation corresponding to each time slice, and write the supply and demand deviation direction, duration of supply and demand deviation and trend of supply and demand deviation into the prediction deviation status record.
[0036] Based on the predicted deviation status record, the resource status of the power supply side, power consumption side and adjustable resource side is read, and the adjustment direction, response capacity, response duration and execution constraints of energy storage equipment, adjustable load, charging pile and distributed power source in the corresponding time slice are determined, and a resource responsive status record is generated.
[0037] By matching the prediction deviation status record with the resource responsiveness status record in time slices, the direction, object and duration of the impact of each adjustable resource adjustment on changes in load demand, power supply capacity, line power flow and energy storage remaining capacity are identified, and the supply and demand impact propagation relationship is generated.
[0038] Using the time slice where the supply-demand deviation occurs as the starting point of the impact, and the power supply side nodes, power consumption side nodes, and adjustable resource nodes affected by the adjustment as the impact nodes, the corresponding nodes are connected according to the direction of the impact, the object of the impact, and the duration of the impact to construct the supply-demand scheduling impact domain within the future time window;
[0039] Based on the supply and demand scheduling influence domain, the scheduling contribution value is calculated according to the reduction of supply and demand deviation, response timeliness, continuous support capacity, resource occupancy level and secondary deviation suppression capacity of each adjustable resource. The resource number, affected node, affected direction, affected duration, scheduling contribution value and executable adjustment task are written into the same record to generate a resource scheduling influence set.
[0040] Optionally, the construction of the supply and demand scheduling influence domain within the future time window includes:
[0041] The power deviation, reactive power deviation, frequency deviation, line power flow deviation, energy storage remaining capacity deviation, and voltage deviation of adjacent time slots in the supply and demand deviation prediction sequence are read in time slot order. The six deviations are arranged in sequence to generate a six-dimensional supply and demand deviation vector. All supply and demand deviation vectors are arranged in time slot order to generate a deviation matrix.
[0042] For the corresponding elements in adjacent rows of the deviation matrix, calculate the sum of squares of the differences in power, reactive power, frequency, power flow, capacity and voltage deviations, and then perform a square root operation on the sum of squares to obtain the deviation connection distance.
[0043] Read the available adjustability, expected response time and execution constraints of each adjustable resource within the future time window, arrange the three resource parameters in sequence to generate a three-dimensional resource capability vector, and arrange all resource capability vectors in order of resource number to generate a resource capability matrix.
[0044] For the corresponding elements in adjacent rows of the resource capability matrix, calculate the sum of squares of the differences between capacity, duration and constraint, and then perform a square root operation on the sum of squares to obtain the resource connection distance.
[0045] When both the deviation connection distance and the resource connection distance do not exceed the preset deviation threshold, an influence connection relationship is established between the corresponding time slice and the resource node, and written into the influence adjacency matrix. The supply and demand scheduling influence domain is generated with the influence adjacency matrix as the edge set and the time slice node and the resource node as the vertices.
[0046] Optionally, the set of generated scheduling policies includes:
[0047] Read the scheduling contribution value, response capacity, response time and execution constraints of each resource node in the resource scheduling impact set, write the resource nodes with scheduling contribution values not lower than the first threshold into the first priority queue, write the resource nodes with scheduling contribution values lower than the first threshold but not lower than the second threshold into the second priority queue, and sort the two queues respectively according to the scheduling contribution value from high to low.
[0048] Within the rolling forecast window, the active power deviation corresponding to each time slot in the supply and demand deviation forecast sequence is read. When the active power deviation of any time slot is greater than the first trigger threshold, the time slot is marked as the main adjustment time slot. Resource nodes are selected from the first priority queue in descending order of scheduling contribution value to generate the main adjustment instruction for the main adjustment time slot.
[0049] After the main adjustment instruction is executed, the remaining active power deviation of the corresponding time slice is recalculated. When the remaining active power deviation is still greater than the second trigger threshold, resource nodes are selected from the second priority queue in descending order of scheduling contribution value to generate compensation adjustment instructions for the time slice.
[0050] The main adjustment instructions and the compensation adjustment instructions are written into the same record according to the time slice order, resource type and execution constraints to form a scheduling strategy set.
[0051] According to an embodiment of the present invention, a supply and demand matching and scheduling system based on an intelligent converged terminal includes:
[0052] The data preprocessing module is used to collect multi-source operation data from the power supply side, power consumption side and adjustable resource side, and generate a supply and demand status set for the transformer area after preprocessing.
[0053] The sequence construction module is used to extract power supply capacity, load demand and resource adjustability features from the supply and demand status of the transformer area and construct a multivariable supply and demand status sequence.
[0054] The phase boundary recombination module is used to calculate the supply and demand security margin and the rate of change of the margin, identify the phase boundary points of supply and demand risks, and recombine to generate multi-scale supply and demand state subsequences.
[0055] The deviation prediction module is used to input multi-scale supply and demand state sub-sequences into the improved TimeMixer model and generate a supply and demand deviation prediction sequence through three-channel mixing.
[0056] The influence domain construction module is used to construct the supply and demand scheduling influence domain based on the supply and demand deviation prediction sequence, calculate the scheduling contribution value, and generate the resource scheduling influence set.
[0057] The scheduling and execution module is used to allocate adjustment tasks based on the resource scheduling impact set, generate and issue a set of scheduling strategies, and update the supply and demand status and impact results in a closed loop.
[0058] The beneficial effects of this invention are:
[0059] This invention uses an intelligent fusion terminal to uniformly collect and process multi-source operational data from the power supply side, power consumption side, and adjustable resource side of a distribution transformer area, constructing a multivariable supply and demand state sequence. It then identifies supply and demand risk boundary points by combining supply and demand safety margins and margin change rates, and performs risk boundary anchoring segment recombination on the supply and demand state sequence. This allows the model input to focus on the key stages of supply and demand state changes, avoiding the problem that traditional fixed time windows cannot accurately represent the dynamic evolution of supply and demand, and improving the completeness of supply and demand state modeling and the accuracy of supply and demand deviation identification.
[0060] This invention introduces an improved TimeMixer model with a risk modulation topology mixing mechanism in the prediction stage. Within each TimeMixer block, a time-series token-mixing channel, a resource topology attention channel, and a risk gating channel are coordinated to jointly model time evolution characteristics, resource association characteristics, and risk states. This enables the model to simultaneously focus on supply and demand change patterns, resource topology relationships, and the evolution process of risk states, effectively improving the accuracy and stability of future supply and demand deviation predictions, enhancing prediction capabilities under complex operating conditions, and providing a more reliable decision-making basis for subsequent scheduling.
[0061] This invention constructs a supply and demand scheduling influence domain, analyzes the propagation relationship of supply and demand influence among the power supply side, power consumption side, and adjustable resources, calculates the scheduling contribution value of each adjustable resource, and generates a scheduling strategy based on the resource scheduling influence set. This realizes the transformation from traditional independent resource scheduling to collaborative resource scheduling, which can give full play to the collaborative adjustment capabilities of energy storage equipment, adjustable loads, charging piles, and distributed power sources, improve resource utilization efficiency, reduce the risk of supply and demand imbalance, and enhance the supply and demand matching accuracy of distribution substations, edge scheduling response speed, and system operation stability. Attached Figure Description
[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0063] Figure 1 The flowchart shows the supply and demand matching and scheduling method based on intelligent fusion terminals proposed in this invention.
[0064] Figure 2 This is a schematic diagram of the supply and demand matching and scheduling system based on intelligent fusion terminals proposed in this invention. Detailed Implementation
[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0066] refer to Figure 1 A supply and demand matching and scheduling method based on intelligent converged terminals includes:
[0067] The intelligent fusion terminal collects multi-source operation data from the power supply side, power consumption side and adjustable resource side within the distribution transformer area, performs preprocessing on the multi-source operation data, and generates a supply and demand status set for the transformer area.
[0068] Based on the supply and demand status set of the transformer area, the characteristics of power supply capacity, load demand and resource adjustability are extracted, and a multivariate supply and demand status sequence is constructed according to the time series.
[0069] Calculate the supply and demand security margin and the rate of change of the margin for each time slice, identify the risk phase boundary points of supply and demand, perform risk phase boundary anchoring segment recombination on the multivariate supply and demand state sequence, and generate multi-scale supply and demand state subsequences;
[0070] An improved TimeMixer model is introduced by incorporating multi-scale supply and demand state subsequences into a risk modulation topology mixing mechanism. Within each TimeMixer block, a time-series token-mixing channel, a resource topology attention channel, and a risk gating channel are set in parallel to generate a supply and demand deviation prediction sequence.
[0071] Based on the supply and demand deviation prediction sequence, the propagation relationship of supply and demand influence among the power supply side, power consumption side and adjustable resources is calculated. The supply and demand scheduling influence domain corresponding to the future time window is constructed, the scheduling contribution value of each adjustable resource is calculated, and the resource scheduling influence set is generated.
[0072] Based on the resource scheduling impact set and the supply-demand deviation prediction sequence, a scheduling strategy set is generated for the allocation and adjustment tasks of energy storage devices, adjustable loads, charging piles and distributed power sources. This set is then distributed to the corresponding execution devices at the edge via intelligent fusion terminals. The supply and demand status after execution is collected and compared in real time, and the supply and demand status data set and resource scheduling impact set are updated to form a closed-loop scheduling processing result.
[0073] In this embodiment, the multi-source operation data includes distribution substation transformer operation data, distributed power output data, energy storage device operation data, user load data, charging pile load data, adjustable load status data, line power flow data, switch status data, power metering data, terminal communication status data, and corresponding timestamps.
[0074] In this embodiment, the step of performing preprocessing on multi-source operational data to generate a supply and demand status set for the transformer area includes performing time alignment, anomaly removal, and status normalization on the multi-source operational data to generate the supply and demand status set for the transformer area.
[0075] In this embodiment, constructing a multivariate supply and demand state sequence according to the time series includes:
[0076] The supply and demand status set of the transformer area is divided into continuous time slices according to the unified sampling period. Based on the transformer area number, smart fusion terminal number, equipment number and sampling timestamp, the power supply side, power consumption side and adjustable resource side data in the same time slice are collected to generate time slice status records.
[0077] The transformer's capacity, line power flow margin, external grid input power, distributed generation real-time output, and energy storage discharge capacity are read from the time-slice status record. These fields are then integrated according to the power source and power-carrying object to generate power supply capacity characteristics. Specifically, the generated power supply capacity characteristics are as follows:
[0078] The system reads the transformer's carrying capacity, line power flow margin, external grid input power, distributed generation real-time output, and energy storage discharge power within the same time slice. These five power values are then written into fixed field positions in the order of transformer, line, external feeder, distributed generation, and energy storage, and concatenated into a power supply vector of length 5. All power supply vectors are arranged according to feeder number and transformer substation number in time slice order to generate a power supply capacity matrix. The power values of corresponding fields in adjacent time slices in the power supply capacity matrix are subtracted, and the squared differences are summed and the square root is taken to obtain the power supply capacity change distance. When the power supply capacity change distance does not exceed a set threshold and both the line power flow margin and the energy storage discharge power meet safety margin requirements, a stable power supply connection is established between time slices. To generate a power supply capacity adjacency matrix, positions satisfying the connection relationship are written with 1, and the rest with 0. Using the power supply capacity matrix as the vertex set and the power supply capacity adjacency matrix as the edge set, a power supply capacity association graph is constructed. The transformer upper limit, line margin, external feed, distributed output, and energy storage discharge capacity are kept in a fixed order in the graph to obtain a five-dimensional vector representing the power supply capacity characteristics of a time slice. A stable power supply connection relationship between a time slice and the previous time slice is determined only when the following three conditions are met simultaneously: power supply capacity change distance ≤ 25kW, line power flow margin ≥ 10% of rated and ≥ 20kW, and energy storage discharge power ≥ 15% of rated and remaining energy storage capacity ≥ 10kWh. Otherwise, 0 is written.
[0079] The system reads real-time user load, charging pile load, rigid load, interruptible load, and historical load within the same period from the time-slice status record. It then integrates these data according to load source, load type, and load change direction to generate load demand characteristics. Specifically, the generated load demand characteristics are as follows:
[0080] Read the real-time load of users, the load of charging piles, the rigid load, the interruptible load and the historical load of the same period within the same time slice, write the five load values into a fixed field in the order of user, charging pile, rigid, interruptible and historical, and concatenate them into an initial load vector with a length of 5.
[0081] A load change direction field is appended to the end of the initial load vector. If the current user's real-time load is more than 20kW higher than the previous time slice, the load is written as rising; if it is more than 20kW lower than the previous time slice, the load is written as falling; otherwise, the load is written as stable.
[0082] Arrange the 6-field vector in time slice order to generate a load demand matrix. Take the absolute value of the difference between each field in adjacent rows of the matrix and sum them up to obtain the load change distance.
[0083] When the load change distance does not exceed 15kW and the rigid load accounts for more than 30%, it is determined that the current time slice and the previous time slice are in the same demand stability range. Write 1 in the corresponding position of the load stability adjacency matrix and write 0 in the rest.
[0084] Finally, a load demand association graph is constructed using the load demand matrix as the vertex set and the load stable adjacency matrix as the edge set. The real-time load of users, the load of charging piles, rigid loads, interruptible loads, historical loads of the same period and the direction of load change are kept in a fixed order in the graph, thus forming the load demand feature vector of the time slice.
[0085] The remaining energy storage capacity, energy storage charge / discharge limits, adjustable load response capacity, delayed charging pile power, and controllable output range of distributed power sources are read from the time-slice status record. These fields are then integrated according to resource type, adjustment direction, and response status to generate resource adjustable capability characteristics. Specifically, the generated resource adjustable capability characteristics are as follows:
[0086] Read the remaining energy storage capacity, energy storage charging and discharging limits, adjustable load response capacity, charging pile delay power and distributed power controllable output range within the same time slice, and write the five values into a fixed field in the order of energy storage, limit, adjustable, delay and controllable, and concatenate them into an initial resource vector with a length of 5.
[0087] An adjustment direction field is appended to the end of the initial resource vector. If the energy storage can be discharged and the adjustable load can be reduced, then write "load reduction". If the energy storage can be charged and the charging pile can be delayed, then write "load increase". Otherwise, write "hold".
[0088] Add a response status field. When the energy storage charge / discharge limit is less than 10% of the rated power or the adjustable load response capacity is less than 5kW, writing is restricted. Otherwise, writing is available, resulting in a resource capability vector of length 7.
[0089] Arrange the 7-field vector in time slice order to generate a resource capability matrix. Take the absolute value of the difference between corresponding fields in adjacent rows of the matrix and sum them to obtain the distance of resource capability change.
[0090] When the distance of resource capacity change does not exceed 10kW and the remaining energy storage capacity is higher than 30%, it is determined that the current time slice and the previous time slice are in the same resource adjustable and stable range. Write 1 in the corresponding position of the resource stable adjacency matrix and write 0 in the rest.
[0091] A resource adjustable capability association graph is constructed using the resource capability matrix as the vertex set and the resource stable adjacency matrix as the edge set. The remaining energy storage capacity, energy storage charging and discharging limits, adjustable load response capacity, charging pile delay power, controllable output range of distributed power sources, adjustment direction and response status are kept in a fixed order in the graph, thus forming the resource adjustable capability feature vector for that time slice.
[0092] The power supply capacity characteristics, load demand characteristics, and resource adjustability characteristics within the same time slice are written into the same supply and demand status record in a fixed field order. Multiple supply and demand status records are then sequentially arranged according to their sampling time to generate a multivariate supply and demand status sequence. Specifically, the generation of the multivariate supply and demand status sequence is as follows:
[0093] Write the five power fields from the power supply capacity feature vector within the same time slice into record fields 1-5 in sequence;
[0094] Write the six load fields from the load demand feature vector into record fields 6-11 in sequence;
[0095] Write the 7 resource fields from the resource adjustability feature vector into record fields 12-18 in sequence;
[0096] Write the sampling timestamp in record field 19 and the time slice sequence number in record field 20 to form a supply and demand status record with a length of 20;
[0097] All supply and demand status records are arranged continuously in order of timestamp from early to late, and an increasing sequence index is assigned to each record to obtain a multivariate supply and demand status sequence ordered by sampling time.
[0098] In this embodiment, the step of performing adaptive time window reconstruction on the multivariate supply and demand state sequence to obtain a multi-scale supply and demand state sub-sequence includes:
[0099] The supply and demand status records in the multivariate supply and demand status sequence are read in chronological order of sampling time. Available power supply capacity is determined from the power supply capacity characteristics corresponding to the same time slot, and actual load demand is determined from the load demand characteristics. The result of subtracting the actual load demand from the available power supply capacity is determined as the supply and demand safety margin for the corresponding time slot. The supply and demand safety margin is the difference between the available power supply capacity and the actual load demand for the same time slot, expressed in kilowatts.
[0100] If the difference is ≥1kW, it is recorded as a positive margin;
[0101] If the difference is between -1kW and +1kW, it is considered balanced;
[0102] If the difference is ≤-1kW, it is recorded as a negative margin, and the absolute value is used to represent the size of the supply-demand gap;
[0103] The supply and demand safety margins of the current time slice and the previous time slice are read according to the sampling order of adjacent time slices. The difference between the supply and demand safety margins of the current time slice and the previous time slice is processed. The rate of change of the margin is determined by combining the sampling interval between adjacent time slices. Specifically, the rate of change of the margin is determined by combining the sampling interval between adjacent time slices as follows:
[0104] Read the sampling timestamps of the current time slice and the previous time slice, calculate the sampling interval by subtracting the previous timestamp from the current timestamp, and the result is in seconds;
[0105] Calculate the current supply and demand safety margin minus the previous supply and demand safety margin to obtain the margin increment, in kilowatts;
[0106] Divide the margin increment by the sampling interval, and then convert the result to a uniform unit of kilowatts per minute or kilowatts per second to obtain the margin change rate of the time slice.
[0107] Based on the supply and demand security margin, margin change rate, and resource adjustability characteristics of each time slice, the supply and demand risk boundary points are identified. When the supply and demand security margin moves from the security margin range into the critical margin range, from the critical margin range into the imbalance margin range, the margin change rate reaches a preset rate threshold, or the resource adjustability margin is less than a preset margin threshold, the corresponding time slice is marked as a supply and demand risk boundary point. The division of the security margin range, critical margin range, and imbalance margin range is as follows:
[0108] When the supply and demand safety margin of the same time slot is greater than or equal to 25kW, it is determined to be within the safety margin range;
[0109] When the supply and demand safety margin in the same time slot is less than 25kW and greater than or equal to 0kW, it is determined to be in the critical margin range.
[0110] When the supply and demand safety margin in the same time slot is less than 0kW, it is determined to be in the imbalance margin range;
[0111] The preset rate threshold is set to 15 kW / min;
[0112] The preset margin threshold is set to 10kW;
[0113] Using the supply and demand risk boundary point as the anchor point, extract time slices in front of the anchor point where the supply and demand safety margin is continuously shrinking and the load demand is continuously increasing to form a contraction segment; extract time slices in the neighborhood of the anchor point where the supply and demand safety margin is in the critical margin range and the resource adjustability margin is continuously decreasing to form a critical segment; extract time slices in behind the anchor point where the supply and demand safety margin returns from the imbalance margin range to the safety margin range to form a recovery segment.
[0114] The contracted segments, critical segments, and recovery segments are correlated and recombined according to sampling time order, phase boundary point number, segment type, and resource adjustability margin. Time slices that have not entered the neighborhood of the supply and demand risk phase boundary point and whose supply and demand safety margin is continuously stable are retained as stable segments, generating a multi-scale supply and demand state subsequence. Specifically, the generation of the multi-scale supply and demand state subsequence is as follows:
[0115] The contraction fragments, critical fragments, and recovery fragments are sorted from earliest to latest according to the sampling time, and each fragment is written with a continuously increasing phase boundary point number;
[0116] The status records within each contraction segment are written into four index fields in sequence according to segment type 1, phase boundary point number, sampling time order, and resource adjustability margin, and the segment step size is uniformly set to 2 time slices.
[0117] The state records within each critical segment are written into four index fields according to segment type 2, phase boundary point number, sampling time sequence, and resource adjustability margin, and the segment step size is uniformly set to 1 time slice.
[0118] The status records within each recovery segment are written into four index fields in sequence according to segment type 3, phase boundary point number, sampling time order, and resource adjustability margin, and the segment step size is uniformly set to 3 time slices.
[0119] Time slices that have not entered the neighborhood of the supply and demand risk phase boundary point and whose supply and demand safety margin is continuous and stable are written into 4 index fields in sequence according to the segment type 0, phase boundary point number 0, sampling time order, and resource adjustable capacity margin, and the segment step size is uniformly set to 5 time slices.
[0120] All records are merged and sorted using three keywords: segment type encoding, phase boundary point number, and sampling time order, ultimately yielding a multi-scale supply and demand state subsequence containing four segment step sizes.
[0121] In this embodiment, generating the supply-demand deviation prediction sequence includes:
[0122] Read stable segments, contraction segments, critical segments, and recovery segments from the multi-scale supply and demand state subsequence, write the segment type, sampling time, resource number, and risk phase boundary number, and add prediction window control flags to generate the risk phase boundary TimeMixer input sequence;
[0123] The TimeMixer input sequence is fed into the temporal token-mixing channel. The temporal token-mixing channel employs a hierarchical expansion mixing module. An expansion stride is used in stable segments to capture long-term dependencies, while a compact stride is used in contracted and critical segments to capture short-term details. Reusable buffers are inserted between layers to reduce edge inference overhead. The output is a temporal evolution mixing feature, specifically:
[0124] In the first layer of the temporal token-mixing channel, a stride of 4 and a kernel width of 64 are set for the sampling window of the stable segment, and 64 consecutive time slices are mapped to 1 mixed token to capture long-term dependencies of 256s. A stride of 1 and a kernel width of 8 are set for the sampling window of the contracted segment and the critical segment, and 8 consecutive time slices are mapped to 1 mixed token to capture short-term details of 8s.
[0125] In the second layer, the stable fragment tokens output from the first layer are further expanded and mixed using a stride of 2 and a kernel width of 32, compressing the cross-fragment dependencies within 32 seconds into a single higher-order token; the shrunken fragment and critical fragment tokens are compactly mixed using a stride of 1 and a kernel width of 4, preserving the original granularity of short-term fluctuation details within 4 seconds.
[0126] In the third layer, a fine-tuning and mixing module with a stride of 1 and a kernel width of 2 is used uniformly for all segments to perform a final local feature compensation on the upper-layer tokens, ensuring the continuity of information across segment boundaries;
[0127] A circular cache with a capacity of 128 tokens is inserted between each layer. The tokens processed in the previous time slice are stored in the cache and reused during inference in the next time slice, reducing the amount of redundant calculations by more than 40%.
[0128] The stable fragments, contracted fragments, and critical fragment tokens output from the third layer are concatenated in chronological order and written with fragment type labels and time indices to finally obtain the temporal evolution mixing features of the temporal token-mixing channel.
[0129] A bidirectional resource topology table is constructed based on the connection relationships, power flow relationships, and regulation response relationships among the power supply side, power consumption side, and adjustable resources. This table is input into a resource topology attention channel. Within the channel, a dual-flow attention mechanism is used to calculate the correlation strength of the supply-demand flow and the demand-supply flow, respectively. The correlation strength weights are updated in real-time using the scheduling results from the previous time step, and the resource topology hybrid features are output. Specifically, the bidirectional resource topology table is constructed based on the connection relationships, power flow relationships, and regulation response relationships among the power supply side, power consumption side, and adjustable resources.
[0130] Read the node list of the current distribution area and number them in the following order: power supply side nodes include transformers, external grid feed points, distributed power generation output ports, and energy storage discharge ports; power consumption side nodes include user master meters, charging pile input ports, rigid loads, and interruptible loads; adjustable resource nodes include energy storage charging ports, charging pile delay ports, adjustable load response ports, and distributed power generation output reduction ports, and write the node numbers into the node index table.
[0131] Read the wiring diagram and real-time power flow data once, and write the connection relationship field for node pairs with physical lines or communication links; when the measured active power of the line from node A to node B is greater than 1kW, write the power flow direction field as positive flow in the direction from A to B, and write negative flow in the opposite direction; if the power is less than 1kW, write zero flow.
[0132] Read the scheduling execution result from the previous moment. If resource node R has already performed a discharge or load reduction task and is still within the continuous period, write the adjustment response relationship field as continuous in the direction from R to the target node; otherwise, write it as idle.
[0133] Generate a six-field record in a row according to the node pair sequence: starting node number, target node number, connection relationship, power flow direction, adjustment response status, and remaining line capacity. After writing all node pairs, a unidirectional topology table is formed.
[0134] Copy the unidirectional topology table and swap the starting and target node numbers to generate a reverse record. At the same time, in the power flow direction field, rewrite positive flow as reverse flow, negative flow as positive flow, and leave zero flow unchanged. Merge the two tables to obtain a bidirectional resource topology table.
[0135] Valid edges are filtered based on the conditions that the connection exists and the power flow is non-zero or the regulation response state is continuous. The bidirectional records retained after removing invalid edges are used as inputs for the resource topology attention channel.
[0136] Within the channel, a dual-flow attention mechanism is used to calculate the correlation strength of the supply-demand flow and the demand-supply flow, respectively, as follows:
[0137] For edges in the topology table where the connection relationship field is present and the power flow direction field is positive, generate a supply-to-demand edge set; for edges where the power flow direction field is negative, generate a demand-to-supply edge set.
[0138] For each starting node in the supply-to-demand direction edge set, extract two attributes: real-time output power and sustainable output duration. Concatenate the two attributes into a key vector. For the target node, extract two attributes: real-time load power and remaining adjustable capacity. Concatenate the two attributes into a query vector. Then calculate the absolute value of the inner product of the key vector and the query vector as the original association score for that direction.
[0139] For each starting node in the demand-to-supply direction edge set, extract two attributes: real-time load power and remaining adjustable capacity, and concatenate them into a key vector. For the target node, extract two attributes: real-time output power and sustainable output duration, and concatenate them into a query vector. Then, calculate the absolute value of the inner product of the key vector and the query vector as the original association score for that direction.
[0140] The normalized correlation strength is obtained by dividing the original correlation scores of the supply-to-demand side set and the demand-to-supply side set by the sum of their respective direction scores, and then writing the normalized strength value into the attention weight matrix to complete the correlation strength calculation of the dual-flow directional attention mechanism.
[0141] The output resource topology hybrid features are as follows:
[0142] The attention weight matrix from supply to demand is traversed row by row. The weights from the same starting node to all target nodes are multiplied by the sum of the real-time load power and remaining adjustable capacity of the corresponding target node, and then accumulated according to the target node number to obtain the aggregate vector from supply to demand.
[0143] The attention weight matrix from demand to supply is traversed row by row. The weights from the same starting node to all target nodes are multiplied by the sum of the real-time output power and the sustainable output duration of the corresponding target node. Then, the weights are accumulated according to the target node number to obtain the aggregate vector from demand to supply.
[0144] The supply-to-demand aggregation vector and the demand-to-supply aggregation vector are concatenated one-to-one at the node level, and the real-time status flag of the current node is added to form a node-level fusion vector with a length of three.
[0145] All node-level fusion vectors are written into the same record in the order of node number to obtain a resource topology hybrid feature matrix containing three columns: supply-to-demand aggregation value, demand-to-supply aggregation value, and node status flag. A time slice index is written into the beginning of the record as the resource topology hybrid feature output for the time slice.
[0146] The supply and demand security margin, margin change rate, segment type, and resource adjustability margin are input into the risk gating channel. Modulation weights are generated internally according to risk level. Risk-weighted fusion is performed on the temporal evolution hybrid features and resource topology hybrid features. The risk level code is injected into the fusion result as a residual, outputting the risk-modulated hybrid features. Specifically, the modulation weights generated internally according to risk level are as follows:
[0147] Read the supply and demand safety margin, margin change rate, fragment type, and resource adjustability margin of the current time slice:
[0148] A supply and demand safety margin ≥ 25kW and a margin change rate > -10kW / min corresponds to a risk level of 0.
[0149] 0kW≦supply and demand safety margin<25kW and -10kW / min≦margin change rate≦10kW / min, corresponding to risk level 1;
[0150] If the supply and demand safety margin is less than 0 kW and the rate of change of the margin is less than -10 kW / min, or the resource adjustability margin is less than 10 kW, the corresponding risk level is 2.
[0151] If the supply and demand safety margin is less than -20kW or the margin change rate is less than -20kW / min, and the resource adjustability capacity margin is less than 5kW, the corresponding risk level is 3.
[0152] Establish a modulation weight table between time-series token-mixing weights and resource topology weights:
[0153] Risk level 0, time evolution weight 0.3, topological weight 0.7;
[0154] Risk level 1: Time evolution weight 0.5, topological weight 0.5;
[0155] Risk level 2: Time evolution weight 0.7, topological weight 0.3;
[0156] Risk level 3: Time evolution weight 0.85, topological weight 0.15;
[0157] The corresponding weights are obtained by looking up the table, the time evolution hybrid features are multiplied by the time evolution weights, the resource topology hybrid features are multiplied by the topology weights, and then the elements are added together to complete the weighted fusion.
[0158] The risk level value of 0-3 is linearly scaled to the range of 0-0.3 as the residual offset and added to each dimension of the fusion result to output the risk modulation hybrid feature;
[0159] The risk modulation mixing feature is input into the TimeMixer output mapping layer. Supply and demand deviation prediction values for the corresponding time slots are generated according to the time slot order of the future time window. The supply and demand deviation prediction values for each time slot are then continuously arranged into a supply and demand deviation prediction sequence. Specifically, generating the supply and demand deviation prediction values for the corresponding time slots according to the time slot order of the future time window involves:
[0160] Read the time dimension length of the risk modulation hybrid feature and determine the future prediction window length to be 12 time slices, with each time slice interval of 5 seconds;
[0161] In the TimeMixer output mapping layer, a 128-dimensional to 1-dimensional fully connected mapping weight is configured for each time slice, which maps the 128-dimensional mixed features of the corresponding time slice to a single supply and demand deviation value, in kW;
[0162] The mapping weights are called sequentially according to the time slice order to generate the supply and demand deviation prediction values for the 1st to 12th time slices, and written to the prediction buffer in real time.
[0163] The 12 supply and demand deviation prediction values in the prediction buffer are concatenated in the sampling order to form a supply and demand deviation prediction sequence of length 12, and the prediction window identifier and timestamp are written at the beginning of the sequence.
[0164] The improved TimeMixer model in this invention uses historical distribution substation operation data for supervised training during the offline phase. The training data consists of a multivariate supply and demand state sequence composed of power supply capacity characteristics, load demand characteristics, and resource adjustability characteristics as input, and the corresponding actual collected supply and demand deviation values as labels. During training, the time series is sampled using a sliding window of length 64, and the prediction window is set to 12 time slices. The model contains 3 layers of TimeMixer blocks, with each layer having a channel dimension of 128. The time series token-mixing channel expansion step size is set to 4, and the compaction step size is set to 1. The resource topology attention channel node number is set according to the actual number of nodes in the substation area. The risk gating channel adopts a 4-level risk level mapping. The model training uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 50 training rounds. The loss function adopts a combination of mean squared error loss and weighted absolute error loss to minimize the squared error between the predicted and actual supply and demand deviation values. The weights are increased for time slices with large deviation changes to improve the model's prediction accuracy for sudden supply and demand imbalances.
[0165] This invention structurally improves the traditional TimeMixer model, proposing a multi-channel risk modulation time-series modeling method for distribution area supply and demand matching scheduling scenarios. In terms of time-series modeling, a risk phase boundary anchoring segment input mechanism is introduced. Based on the supply and demand safety margin and the margin change rate, the original supply and demand state sequence is reorganized. Stable segments, contraction segments, critical segments, and recovery segments are organized into a multi-scale input sequence in chronological order, enabling the model to focus on the key stages of supply and demand deviation formation and evolution. A combination of hierarchical expansion step size and compact step size is set in the time-series token-mixing channel to model long-term dependencies and short-term fluctuations separately, improving the ability to characterize complex supply and demand changes.
[0166] To address the interrelationships among various resource types in supply and demand scheduling, this invention introduces a resource topology attention channel. By constructing a bidirectional resource topology relationship table, the connection relationships, power flow relationships, and adjustment response relationships between the power supply side, power consumption side, and adjustable resource nodes are uniformly encoded. Within the channel, a dual-flow direction attention mechanism is used to calculate the correlation strength from supply to demand and from demand to supply, respectively. The correlation weights are updated in real time based on the scheduling execution results of the previous moment, enabling the model to dynamically reflect the propagation path and intensity of influence between resources, thereby enhancing its ability to express the influence relationships of supply and demand scheduling.
[0167] This invention proposes a risk-gated channel that maps supply and demand security margins, margin change rates, segment types, and resource adjustability margins into discrete risk levels. Differentiated modulation weights are generated based on the risk levels to perform weighted fusion of temporal hybrid features and topological hybrid features. At the same time, the risk level is injected as residual information into the fusion result, enabling the model to automatically adjust the focus of features under different risk states, thus achieving continuous modeling from stable operation to imbalanced state.
[0168] The improved TimeMixer model not only achieves collaborative modeling of the time evolution characteristics of supply and demand status, resource topology characteristics, and risk status characteristics, but also significantly improves the accuracy of supply and demand deviation prediction and accurately captures key moments in scheduling through risk phase boundary anchoring and multi-channel fusion mechanism. This enhances the real-time performance and reliability of supply and demand matching scheduling, and strengthens the operational stability and resource utilization efficiency of distribution substations under complex operating conditions.
[0169] In this embodiment, calculating the scheduling contribution value of each adjustable resource and generating a resource scheduling impact set includes:
[0170] Read the supply and demand deviation prediction sequence according to the time slice order of the future time window, identify the supply and demand deviation direction, duration of supply and demand deviation and trend of supply and demand deviation corresponding to each time slice, and write the supply and demand deviation direction, duration of supply and demand deviation and trend of supply and demand deviation into the prediction deviation status record.
[0171] Based on the predicted deviation state record, the resource status of the power supply side, power consumption side, and adjustable resource side is read. The adjustment direction, response capacity, response duration, and execution constraints of energy storage devices, adjustable loads, charging piles, and distributed power sources within the corresponding time slice are determined, and a resource responsiveness state record is generated. Specifically, the generation of the resource responsiveness state record is as follows:
[0172] Read the current charging and discharging status of the energy storage device. If the remaining energy storage capacity is ≥30kWh and the converter is in the discharge allowable state, then write the discharge direction, write the dischargeable power for the response capacity, write the allowable discharge duration in minutes for the response duration, and write the temperature-limited constraint for the execution constraint. If the remaining energy storage capacity is ≤80kWh and the system needs to absorb surplus energy, then write the charging direction for the adjustment, and write the other fields according to the same rules.
[0173] Read the real-time operating power and adjustable power of the adjustable load. If the system needs to reduce the load and the adjustable power is ≥5kW, then write "downward" for the adjustment direction, "adjustable power" for the response capacity, "allowed downward adjustment duration" for the response duration, and "process allowed" for the execution constraint. If the system needs to increase the load and the adjustable power is ≥5kW, then write "upward" for the adjustment direction.
[0174] Read the charging pile queue status. If the total power of the vehicles in the queue is ≥20kW and the system needs to release the power supply capacity, adjust the direction and write the delay, write the delayable power for the response capacity, write the maximum delayable minutes for the response duration, and write the user's consent for the execution constraints.
[0175] Read the real-time output and adjustable output range of the distributed power inverter. If the system needs to increase its power supply capacity and the upward adjustment margin is ≥10kW, then write the adjustment direction as "increase power generation", write the response capacity as "upward adjustment margin", write the response duration as the number of minutes allowed for increased power generation, and write the execution constraint as the inverter temperature rise limit. If it needs to reduce the feed-in power and the downward adjustment margin is ≥10kW, then write the adjustment direction as "decrease power generation".
[0176] Write the above data in the order of fixed fields such as resource type, adjustment direction, response capacity, response duration, and execution constraints to generate a resource responsiveness status record for the time slice. Write the unique resource number and time slice index in the first field of the record and store it in the resource responsiveness status table.
[0177] The predicted deviation status records are time-sliced and matched with the resource responsiveness status records to identify the direction, objects, and duration of the impact of each adjustable resource adjustment on changes in load demand, power supply capacity, line power flow, and remaining energy storage capacity. This generates the supply and demand impact propagation relationship, specifically as follows:
[0178] Using the time slice index as the primary key, the prediction deviation status records are matched one by one with the resource responsiveness status records;
[0179] For each time slice, read the supply-demand deviation direction field. If the deviation direction is supply less than demand, set the influence direction flag to +1; if the deviation direction is supply greater than demand, set it to -1.
[0180] Read the matched resource responsive status records and write them into the affected object field according to the resource type: write "energy storage" for energy storage devices, "load" for adjustable loads, "charging pile" for charging piles, and "distributed power source" for distributed power sources.
[0181] The minimum value of the deviation between the resource response capacity and the prediction is used as the impact magnitude. Records with an impact magnitude greater than 5kW are retained and recorded in the impact duration field: energy storage response is written for 15 minutes, adjustable load and charging pile delay is written for 10 minutes, and distributed power output adjustment is written for 5 minutes.
[0182] Write the time slice index, direction of influence, object of influence, magnitude of influence, and duration of influence into a single row record and save it to the supply and demand influence propagation relationship table.
[0183] Using the time slice where the supply-demand deviation occurs as the starting point of the impact, and the power supply-side nodes, power consumption-side nodes, and adjustable resource nodes affected by the adjustment as the impact nodes, the corresponding nodes are connected according to the direction of impact, the object of impact, and the duration of impact to construct the supply-demand scheduling impact domain within the future time window. Specifically, the construction of the supply-demand scheduling impact domain within the future time window is as follows:
[0184] The six-dimensional deviations of power, reactive power, frequency, line power flow, energy storage capacity and voltage are written into the deviation matrix in the order of time slices. Each row corresponds to one time slice. First, each column is normalized to the interval of -1 to 1 according to the maximum absolute value. Then, the difference between adjacent rows is performed column by column. The squares of the differences are summed and the square root is taken to obtain the deviation connection distance sequence.
[0185] Compare the deviation connection distance with the threshold 0.35. If it is less than or equal to 0.35, it is recorded as connected and 1 is written at the position of time slice i and time slice i+1. Otherwise, 0 is written to form a time slice interconnection submatrix.
[0186] Read the three parameters of each adjustable resource in the future window: available adjustable capacity, expected response time, and execution constraint level. Write them into the resource capability matrix in the order of capacity, time, and constraint. Normalize the capacity column to 0-1 and the time column to 0-1. Map the constraint level to 0, 0.5, and 1 and then normalize it.
[0187] Perform column-by-column difference on adjacent rows of the resource capability matrix, square the difference, sum and take the square root to obtain the resource connection distance sequence. Compare the resource connection distance with the threshold 0.25. If it is less than or equal to 0.25, write 1 at the position of resource node j and j+1, otherwise write 0 to form a resource interconnection submatrix.
[0188] Construct an empty adjacency matrix, write the time slice interconnection submatrix into the upper left block, write the resource interconnection submatrix into the lower right block, and then write the influence connection between the time slice node and the resource node into the upper right block and write 1 according to the influence direction, and write the feedback connection between the resource node and the time slice node into the lower left block and write 1. Finally, the influence adjacency matrix containing bidirectional connections is obtained. Use this adjacency matrix as the edge set and the time slice node and the resource node as vertices to generate the supply and demand scheduling influence domain.
[0189] Based on the supply and demand scheduling influence domain, the scheduling contribution value is calculated according to the reduction magnitude of supply and demand deviation, response timeliness, continuous support capacity, resource utilization, and secondary deviation suppression capacity of each adjustable resource. The resource number, affected node, influence direction, influence duration, scheduling contribution value, and executable adjustment task are written into the same record to generate a resource scheduling influence set. The specific calculation of the scheduling contribution value is as follows:
[0190] Read five indicators of the target resource in the corresponding time slice: the reduction of supply and demand deviation, the timeliness of response, the continuous support capacity, the resource utilization and the ability to suppress secondary deviation. Scale the five indicators to the range of 0-100 respectively.
[0191] Weighting coefficients were assigned to the five indicators: reduction magnitude weight 0.35, timely response weight 0.25, continuous support weight 0.2, resource consumption reverse weight 0.1, and second-order deviation suppression weight 0.1.
[0192] The weighted score is obtained by multiplying the four indicators (reduction magnitude, timely response, continuous support, and second-order deviation suppression) directly according to their respective weights, then summing them up, subtracting the resource occupancy value from 100, multiplying it by its corresponding weight, and then adding it to the summation result.
[0193] Divide the weighted score by the sum of the weighting coefficients and round it to 1kW precision to obtain the final scheduling contribution value;
[0194] When the contribution value is less than 10kW, it is marked as low contribution; when it is between 10-30kW, it is marked as medium contribution; and when it is greater than 30kW, it is marked as high contribution. The resource number, affected node, affected direction, affected duration, scheduling contribution value level, and executable adjustment task are written into the resource scheduling affected set.
[0195] In this embodiment, constructing the supply and demand scheduling influence domain within the future time window includes:
[0196] The power deviation, reactive power deviation, frequency deviation, line power flow deviation, energy storage remaining capacity deviation, and voltage deviation of adjacent time slots in the supply and demand deviation prediction sequence are read in time slot order. The six deviations are arranged in sequence to generate a six-dimensional supply and demand deviation vector. All supply and demand deviation vectors are arranged in time slot order to generate a deviation matrix.
[0197] For the corresponding elements in adjacent rows of the deviation matrix, calculate the sum of squares of the differences in power, reactive power, frequency, power flow, capacity and voltage deviations, and then perform a square root operation on the sum of squares to obtain the deviation connection distance.
[0198] Read the available adjustability, expected response time and execution constraints of each adjustable resource within the future time window, arrange the three resource parameters in sequence to generate a three-dimensional resource capability vector, and arrange all resource capability vectors in order of resource number to generate a resource capability matrix.
[0199] For the corresponding elements in adjacent rows of the resource capability matrix, calculate the sum of squares of the differences between capacity, duration and constraint, and then perform a square root operation on the sum of squares to obtain the resource connection distance.
[0200] When both the deviation connection distance and the resource connection distance do not exceed the preset deviation threshold, an influence connection relationship is established between the corresponding time slice and the resource node, and written into the influence adjacency matrix. The supply and demand scheduling influence domain is generated with the influence adjacency matrix as the edge set and the time slice node and the resource node as the vertices.
[0201] In this embodiment, the generation of the scheduling policy set includes:
[0202] Read the scheduling contribution value, response capacity, response time and execution constraints of each resource node in the resource scheduling impact set. Write the resource nodes with scheduling contribution values not lower than the first threshold into the first priority queue. Write the resource nodes with scheduling contribution values lower than the first threshold but not lower than the second threshold into the second priority queue. Sort the two queues according to the scheduling contribution value from high to low. The first threshold is set to 30kW and the second threshold is set to 10kW.
[0203] Within the rolling forecast window, the active power deviation corresponding to each time slot in the supply and demand deviation forecast sequence is read. When the active power deviation of any time slot is greater than the first trigger threshold, the time slot is marked as the main adjustment time slot. Resource nodes are selected from the first priority queue in descending order of scheduling contribution value, and a main adjustment instruction is generated for the main adjustment time slot. The first trigger threshold is set to 15kW.
[0204] After the main control command is executed, the remaining active power deviation of the corresponding time slice is recalculated. When the remaining active power deviation is still greater than the second trigger threshold, resource nodes are selected from the second priority queue in descending order of scheduling contribution value to generate compensation control commands for the time slice. The second trigger threshold is set to 5kW.
[0205] The main regulation command and the compensation regulation command are written into the same record according to the time slice order, resource category and execution constraints to form a scheduling strategy set, which includes energy storage charging and discharging command, adjustable load power adjustment command, charging pile load migration command and distributed power output correction command.
[0206] refer to Figure 2 A supply and demand matching and scheduling system based on intelligent converged terminals includes:
[0207] The data preprocessing module is used to collect multi-source operation data from the power supply side, power consumption side and adjustable resource side, and generate a supply and demand status set for the transformer area after preprocessing.
[0208] The sequence construction module is used to extract power supply capacity, load demand and resource adjustability features from the supply and demand status of the transformer area and construct a multivariable supply and demand status sequence.
[0209] The phase boundary recombination module is used to calculate the supply and demand security margin and the rate of change of the margin, identify the phase boundary points of supply and demand risks, and recombine to generate multi-scale supply and demand state subsequences.
[0210] The deviation prediction module is used to input multi-scale supply and demand state sub-sequences into the improved TimeMixer model and generate a supply and demand deviation prediction sequence through three-channel mixing.
[0211] The influence domain construction module is used to construct the supply and demand scheduling influence domain based on the supply and demand deviation prediction sequence, calculate the scheduling contribution value, and generate the resource scheduling influence set.
[0212] The scheduling and execution module is used to allocate adjustment tasks based on the resource scheduling impact set, generate and issue a set of scheduling strategies, and update the supply and demand status and impact results in a closed loop.
[0213] Example 1: In a continuous distribution transformer area supply and demand scheduling simulation cycle, the system connects to one intelligent fusion terminal. The rated capacity of the transformer in the area is 630kVA, which is equivalent to an available active power carrying capacity of 504kW. The installed capacity of distributed photovoltaic power generation is 180kW, the rated capacity of energy storage is 200kWh, the rated power of the energy storage converter is 100kW, the total connected capacity of charging piles is 120kW, and the interruptible load capacity is 60kW. The system continuously collects 12 hours of operating data with a sampling period of 5 seconds, resulting in 8640 time slices. Each time slice includes the transformer's carrying capacity, line power flow margin, external grid input power, real-time output of distributed power sources, energy storage discharge power, real-time user load, charging pile load, rigid load, interruptible load, historical load in the same period, remaining energy storage capacity, energy storage charging and discharging limits, adjustable load response capacity, delayed power of charging piles, and controllable output range of distributed power sources.
[0214] After the data enters the processing flow, the intelligent fusion terminal performs field alignment and anomaly removal for each time slice. In a certain time slice, the transformer's carrying capacity is 504kW, the line power flow margin is 46kW, the external grid input power is 312kW, the real-time photovoltaic output is 74kW, and the energy storage's discharge power is 62kW. The system writes five fields in the order of transformer, line, external feed, distributed power source, and energy storage to generate power supply capacity characteristics. The square root of the sum of the squares of the differences between the five power supply fields of two adjacent time slices yields a power supply capacity change distance of 18.6kW, which is less than 25kW. Furthermore, the line power flow margin is greater than 20kW, and the remaining energy storage capacity is 118kWh, which is greater than 30kWh. Therefore, a 1 is written at the corresponding position in the power supply stability adjacency matrix.
[0215] Within the same time slice, the real-time user load is 398kW, the charging pile load is 84kW, the rigid load is 286kW, the interruptible load is 42kW, and the historical load for the same period is 436kW. The system generates an initial load vector with 5 fields in the order of user, charging pile, rigid, interruptible, and historical load. The real-time user load in the previous time slice was 376kW, and it increased by 22kW in this time slice, so the load change direction is written as upward. The system arranges the 6-field vector in the order of time slices to form a load demand matrix. After summing the absolute values of the differences between adjacent rows, the load change distance is 14.2kW, which is less than 15kW. The rigid load accounts for 71.9%, which is greater than 30%, and the corresponding position in the load stability adjacency matrix is written as 1. The resource side reads the remaining energy storage capacity of 118kWh, the energy storage charge and discharge limit of 82kW, the adjustable load response capacity of 38kW, the charging pile delay power of 31kW, and the controllable output range of distributed power sources of 22kW, and adds the adjustment direction of load reduction and the response status of availability, forming a 7-field resource adjustable capability characteristic.
[0216] The system writes 5 power supply fields, 6 load fields, 7 resource fields, sampling timestamps, and time slice sequence numbers into a single 20-byte supply and demand status record. 64 consecutive supply and demand status records form a training input window, with the supply and demand deviation values for the next 12 time slices used as labels, and each prediction time slice spaced 5 seconds apart. The training phase uses 120,000 samples, including 42,000 normal supply and demand samples, 26,000 samples of rapid photovoltaic decline, 22,000 samples of concentrated charging pile access, 16,000 samples of limited energy storage, and 14,000 samples of local overload. The improved TimeMixer model uses 3 TimeMixer blocks with 128 channel dimensions, a stable segment expansion step size of 4, a shrinking segment and critical segment compaction step size of 1, an inter-layer circular buffer capacity of 128 tokens, and a prediction window length of 12. The Adam optimizer was used for training with a learning rate of 0.001 and a batch size of 32. After 50 training rounds, the mean squared error of the training set decreased from 31.6 to 3.8, and the mean absolute error of the validation set decreased from 18.4kW to 4.9kW.
[0217] In the simulated disturbance segment, photovoltaic output decreased from 92kW to 61kW, charging pile load increased from 46kW to 91kW, and user real-time load increased from 351kW to 407kW. The supply and demand safety margins for eight consecutive time slots were 36kW, 31kW, 24kW, 16kW, 7kW, -4kW, -13kW, and -9kW, respectively. The system subtracts the actual load demand from the available power supply capacity to obtain the supply and demand safety margin, and calculates the margin change rate based on the difference between adjacent time slots combined with a 5-second sampling interval. In the third time slot, the supply and demand safety margin decreased from 31kW to 24kW, entering the critical margin range; in the sixth time slot, the supply and demand safety margin decreased from 7kW to -4kW, entering the imbalance margin range; among them, the margin change rate in the sixth time slot reached -18kW / min, exceeding 15kW / min, and the system marked both the third and sixth time slots as the supply and demand risk boundary points. Subsequently, using the risk phase boundary point as the anchor point, 16 time slices in which the supply and demand safety margin continuously shrinks are extracted forward as contraction segments, 12 time slices in the neighborhood of the phase boundary point are extracted as critical segments, 10 time slices in which the supply and demand safety margin recovers from -13kW to 6kW are extracted backward as recovery segments, and the remaining time slices in which the supply and demand safety margin is continuously stable are taken as stable segments, thus generating a multi-scale supply and demand state subsequence.
[0218] After inputting multi-scale supply and demand state subsequences into the improved TimeMixer model, the time-series token-mixing channel performs long-term mixing with a step size of 4 for stable segments and fine-grained mixing with a step size of 1 for contracting and critical segments. The resource topology attention channel reads 12 resource nodes to form a bidirectional resource topology relationship table, in which the edges from the energy storage discharge port to the user master table, from the energy storage discharge port to the charging pile entrance, and from the interruptible load response port to the line power flow node are all written into the valid edges. The risk gating channel reads the current supply and demand safety margin, the margin change rate, and the resource adjustability margin. In critical segments, the time evolution weight is set to 0.70 and the topology weight is set to 0.30, and in unbalanced segments, the time evolution weight is set to 0.85 and the topology weight is set to 0.15. The model outputs active power deviations for the next 12 time slots as 9kW, 13kW, 18kW, 26kW, 31kW, 36kW, 34kW, 29kW, 21kW, 12kW, 6kW, and 2kW. Among them, the deviations for the 3rd to 9th predicted time slots exceed 15kW, and the system marks these as the main control time slots.
[0219] The system constructs a supply and demand scheduling influence domain. Within the future window, the six-dimensional deviation is written into the deviation matrix in the order of power, reactive power, frequency, line power flow, energy storage capacity, and voltage. The deviation connection distance is calculated by normalizing the difference between adjacent rows. The deviation connection distances for the 4th to 8th time slices are 0.21, 0.28, 0.33, 0.31, and 0.24, respectively, all less than 0.35. The resource connection distances for energy storage, interruptible load, charging pile delay, and distributed power source correction are 0.18, 0.22, 0.19, and 0.24, respectively, all less than 0.25, thus influencing the writing of bidirectional edges in the adjacency matrix. The energy storage device's supply and demand deviation reduction score is 88, its timely response score is 91, its continuous support score is 76, its resource utilization score is 38, and its secondary deviation suppression score is 72. The weighted scheduling contribution is 78.6kW; interruptible load is 34.2kW; charging pile delay is 27.5kW; and distributed power source output correction is 18.3kW. Based on the first threshold of 30kW and the second threshold of 10kW, energy storage devices and interruptible loads enter the first priority queue, while charging pile delays and distributed power output corrections enter the second priority queue.
[0220] During the scheduling execution phase, the system first generates a main regulation command for energy storage discharge, setting the discharge power to 45kW and the execution duration to 10 minutes, with the constraint that the remaining energy storage capacity is not less than 30kWh. Simultaneously, it generates an interruptible load reduction command, setting the reduction amount to 22kW and the execution duration to 10 minutes. After the commands are issued, the intelligent fusion terminal continues to collect execution results at 5-second intervals. The original maximum active power deviation of 36kW decreased to 14.6kW after 3 time slices and to 4.8kW after 8 time slices, below the second trigger threshold of 5kW, thus not triggering compensation regulation. In another set of disturbance samples, the remaining deviation after the main regulation was 7.3kW. The system invoked the second priority queue, delaying the 18kW charging power by 10 minutes, and the remaining deviation eventually decreased to 2.6kW.
[0221] In the comparative experiment, the method of this invention and the traditional fixed-window rule scheduling method used the same 120,000 training samples and 200 test disturbance samples. The traditional method does not identify risk boundary points, does not construct the supply-demand scheduling influence domain, and only calls energy storage according to a fixed priority. Test results show:
[0222] The average supply-demand deviation prediction error of the traditional method is 13.8kW, while that of this invention is 4.9kW;
[0223] The average scheduling response time of the traditional method is 42 seconds, while that of this invention is 13 seconds;
[0224] The average residual active power deviation after implementation of the traditional method is 11.6kW, while that of the present invention is 3.7kW;
[0225] Traditional energy storage methods have an average power consumption of 58kW, while this invention has a power consumption of 45kW.
[0226] The traditional method resulted in 27 instances of secondary deviations exceeding 5kW, while the present invention resulted in 7 such instances.
[0227] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A supply and demand matching and scheduling method based on intelligent converged terminals, characterized in that, include: The intelligent fusion terminal collects multi-source operation data from the power supply side, power consumption side and adjustable resource side within the distribution transformer area, performs preprocessing on the multi-source operation data, and generates a supply and demand status set for the transformer area. Based on the supply and demand status set of the transformer area, the characteristics of power supply capacity, load demand and resource adjustability are extracted, and a multivariate supply and demand status sequence is constructed according to the time series. Calculate the supply and demand security margin and the rate of change of the margin for each time slice, identify the risk phase boundary points of supply and demand, perform risk phase boundary anchoring segment recombination on the multivariate supply and demand state sequence, and generate multi-scale supply and demand state subsequences; An improved TimeMixer model is introduced by incorporating multi-scale supply and demand state subsequences into a risk modulation topology mixing mechanism. Within each TimeMixer block, a time-series token-mixing channel, a resource topology attention channel, and a risk gating channel are set in parallel to generate a supply and demand deviation prediction sequence. Based on the supply and demand deviation prediction sequence, the propagation relationship of supply and demand influence among the power supply side, power consumption side and adjustable resources is calculated. The supply and demand scheduling influence domain corresponding to the future time window is constructed, the scheduling contribution value of each adjustable resource is calculated, and the resource scheduling influence set is generated. Based on the resource scheduling impact set and the supply-demand deviation prediction sequence, a scheduling strategy set is generated for the allocation and adjustment tasks of energy storage devices, adjustable loads, charging piles and distributed power sources. This set is then distributed to the corresponding execution devices at the edge via intelligent fusion terminals. The supply and demand status after execution is collected and compared in real time, and the supply and demand status data set and resource scheduling impact set are updated to form a closed-loop scheduling processing result.
2. The supply and demand matching and scheduling method based on intelligent converged terminals according to claim 1, characterized in that, The multi-source operation data includes distribution area transformer operation data, distributed power output data, energy storage equipment operation data, user load data, charging pile load data, adjustable load status data, line power flow data, switch status data, electricity metering data, terminal communication status data, and corresponding timestamps.
3. The supply and demand matching and scheduling method based on intelligent converged terminals according to claim 1, characterized in that, The step of performing preprocessing on multi-source operational data to generate a supply and demand status set for the transformer area includes performing time alignment, anomaly removal, and status normalization on the multi-source operational data to generate the supply and demand status set for the transformer area.
4. The supply and demand matching and scheduling method based on intelligent converged terminals according to claim 1, characterized in that, The construction of a multivariate supply and demand state sequence according to the time series includes: The supply and demand status set of the transformer area is divided into continuous time slices according to the unified sampling period. Based on the transformer area number, smart fusion terminal number, equipment number and sampling timestamp, the power supply side, power consumption side and adjustable resource side data in the same time slice are collected to generate time slice status records. Read the transformer's carrying capacity, line power flow margin, external grid input power, real-time output of distributed power sources, and energy storage discharge capacity from the time slice status record, integrate the fields according to the power source and power carrying object, and generate power supply capacity characteristics. Read real-time user load, charging pile load, rigid load, interruptible load and historical load of the same period from the time slice status record, integrate the fields according to load source, load type and load change direction to generate load demand characteristics; Read the remaining energy storage capacity, energy storage charging and discharging limits, adjustable load response capacity, charging pile delay power and controllable output range of distributed power sources from the time slice status record, integrate the fields according to resource type, adjustment direction and response status, and generate resource adjustable capability characteristics. The power supply capacity characteristics, load demand characteristics, and resource adjustability characteristics within the same time slice are written into the same supply and demand status record in a fixed field order, and multiple supply and demand status records are continuously arranged according to the sampling time order to generate a multivariate supply and demand status sequence.
5. The supply and demand matching and scheduling method based on intelligent converged terminals according to claim 1, characterized in that, The adaptive time window reconstruction of the multivariate supply and demand state sequence yields a multi-scale supply and demand state sub-sequence, including: Read the supply and demand status records in the multivariate supply and demand status sequence according to the sampling time sequence, determine the available power supply capacity from the power supply capacity characteristics corresponding to the same time slice, determine the actual load demand from the load demand characteristics, and determine the supply and demand safety margin of the corresponding time slice by subtracting the actual load demand from the available power supply capacity. The supply and demand safety margins of the current time slice and the previous time slice are read according to the sampling order of adjacent time slices. The supply and demand safety margins of the current time slice and the previous time slice are processed by difference. The rate of change of the margin is determined by combining the sampling interval between adjacent time slices. Based on the characteristics of supply and demand security margin, margin change rate and resource adjustability in each time slice, the supply and demand risk boundary point is identified. When the supply and demand security margin moves from the security margin range to the critical margin range, from the critical margin range to the imbalance margin range, the margin change rate reaches the preset rate threshold, or the resource adjustability margin is less than the preset margin threshold, the corresponding time slice is marked as the supply and demand risk boundary point. Using the supply and demand risk boundary point as the anchor point, extract time slices in front of the anchor point where the supply and demand safety margin is continuously shrinking and the load demand is continuously increasing to form a contraction segment; extract time slices in the neighborhood of the anchor point where the supply and demand safety margin is in the critical margin range and the resource adjustability margin is continuously decreasing to form a critical segment; extract time slices in behind the anchor point where the supply and demand safety margin returns from the imbalance margin range to the safety margin range to form a recovery segment. The contracted segments, critical segments, and recovery segments are correlated and recombined according to the sampling time order, phase boundary point number, segment type, and resource adjustability margin. Time slices that have not entered the neighborhood of the supply and demand risk phase boundary point and whose supply and demand safety margin is continuous and stable are retained as stable segments to generate multi-scale supply and demand state subsequences.
6. The supply and demand matching and scheduling method based on intelligent converged terminals according to claim 1, characterized in that, The generation of the supply-demand deviation prediction sequence includes: Read stable segments, contraction segments, critical segments, and recovery segments from the multi-scale supply and demand state subsequence, write the segment type, sampling time, resource number, and risk phase boundary number, and add prediction window control flags to generate the risk phase boundary TimeMixer input sequence; The TimeMixer input sequence is fed into the temporal token-mixing channel. The temporal token-mixing channel uses a hierarchical expansion mixing module. An expansion step size is used in stable segments to capture long-term dependencies, and a compact step size is used in contracted segments and critical segments to capture short-term details. Reusable caches are inserted between layers to reduce edge inference overhead and output temporal evolution mixing features. Based on the connection relationship, power flow relationship and regulation response relationship between the power supply side, power consumption side and adjustable resources, a two-way resource topology relationship table is constructed. The topology relationship table is input into the resource topology attention channel. Inside the channel, a dual-flow direction attention mechanism is used to calculate the correlation strength of supply-demand flow and demand-supply flow respectively. The correlation strength weight is updated in real time using the scheduling result of the previous moment, and the resource topology hybrid feature is output. Input the supply and demand security margin, margin change rate, segment type and resource adjustability margin into the risk gating channel. The channel generates modulation weights according to the risk level, performs risk weight weighted fusion on the time evolution hybrid features and resource topology hybrid features, injects the risk level code into the fusion result in the form of residuals, and outputs the risk modulation hybrid features. The risk modulation hybrid feature is input into the TimeMixer output mapping layer, and the supply and demand deviation prediction values for the corresponding time slices are generated according to the time slice order of the future time window. The supply and demand deviation prediction values of each time slice are arranged continuously to form a supply and demand deviation prediction sequence.
7. The supply and demand matching and scheduling method based on intelligent converged terminals according to claim 1, characterized in that, The calculation of the scheduling contribution value of each adjustable resource generates a resource scheduling impact set, including: Read the supply and demand deviation prediction sequence according to the time slice order of the future time window, identify the supply and demand deviation direction, duration of supply and demand deviation and trend of supply and demand deviation corresponding to each time slice, and write the supply and demand deviation direction, duration of supply and demand deviation and trend of supply and demand deviation into the prediction deviation status record. Based on the predicted deviation status record, the resource status of the power supply side, power consumption side and adjustable resource side is read, and the adjustment direction, response capacity, response duration and execution constraints of energy storage equipment, adjustable load, charging pile and distributed power source in the corresponding time slice are determined, and a resource responsive status record is generated. By matching the prediction deviation status record with the resource responsiveness status record in time slices, the direction, object and duration of the impact of each adjustable resource adjustment on changes in load demand, power supply capacity, line power flow and energy storage remaining capacity are identified, and the supply and demand impact propagation relationship is generated. Using the time slice where the supply-demand deviation occurs as the starting point of the impact, and the power supply side nodes, power consumption side nodes, and adjustable resource nodes affected by the adjustment as the impact nodes, the corresponding nodes are connected according to the direction of the impact, the object of the impact, and the duration of the impact to construct the supply-demand scheduling impact domain within the future time window; Based on the supply and demand scheduling influence domain, the scheduling contribution value is calculated according to the reduction of supply and demand deviation, response timeliness, continuous support capacity, resource occupancy level and secondary deviation suppression capacity of each adjustable resource. The resource number, affected node, affected direction, affected duration, scheduling contribution value and executable adjustment task are written into the same record to generate a resource scheduling influence set.
8. The supply and demand matching and scheduling method based on intelligent converged terminals according to claim 7, characterized in that, The constructed supply and demand scheduling influence domain within the future time window includes: The power deviation, reactive power deviation, frequency deviation, line power flow deviation, energy storage remaining capacity deviation, and voltage deviation of adjacent time slots in the supply and demand deviation prediction sequence are read in time slot order. The six deviations are arranged in sequence to generate a six-dimensional supply and demand deviation vector. All supply and demand deviation vectors are arranged in time slot order to generate a deviation matrix. For the corresponding elements in adjacent rows of the deviation matrix, calculate the sum of squares of the differences in power, reactive power, frequency, power flow, capacity and voltage deviations, and then perform a square root operation on the sum of squares to obtain the deviation connection distance. Read the available adjustability, expected response time and execution constraints of each adjustable resource within the future time window, arrange the three resource parameters in sequence to generate a three-dimensional resource capability vector, and arrange all resource capability vectors in order of resource number to generate a resource capability matrix. For the corresponding elements in adjacent rows of the resource capability matrix, calculate the sum of squares of the differences between capacity, duration and constraint, and then perform a square root operation on the sum of squares to obtain the resource connection distance. When both the deviation connection distance and the resource connection distance do not exceed the preset deviation threshold, an influence connection relationship is established between the corresponding time slice and the resource node, and written into the influence adjacency matrix. The supply and demand scheduling influence domain is generated with the influence adjacency matrix as the edge set and the time slice node and the resource node as the vertices.
9. The supply and demand matching and scheduling method based on intelligent converged terminals according to claim 1, characterized in that, The set of generated scheduling policies includes: Read the scheduling contribution value, response capacity, response time and execution constraints of each resource node in the resource scheduling impact set, write the resource nodes with scheduling contribution values not lower than the first threshold into the first priority queue, write the resource nodes with scheduling contribution values lower than the first threshold but not lower than the second threshold into the second priority queue, and sort the two queues respectively according to the scheduling contribution value from high to low. Within the rolling forecast window, the active power deviation corresponding to each time slot in the supply and demand deviation forecast sequence is read. When the active power deviation of any time slot is greater than the first trigger threshold, the time slot is marked as the main adjustment time slot. Resource nodes are selected from the first priority queue in descending order of scheduling contribution value to generate the main adjustment instruction for the main adjustment time slot. After the main adjustment instruction is executed, the remaining active power deviation of the corresponding time slice is recalculated. When the remaining active power deviation is still greater than the second trigger threshold, resource nodes are selected from the second priority queue in descending order of scheduling contribution value to generate compensation adjustment instructions for the time slice. The main adjustment instructions and the compensation adjustment instructions are written into the same record according to the time slice order, resource type and execution constraints to form a scheduling strategy set.
10. A supply and demand matching and scheduling system based on intelligent converged terminals, executing the supply and demand matching and scheduling method based on intelligent converged terminals as described in any one of claims 1 to 9, characterized in that, include: The data preprocessing module is used to collect multi-source operation data from the power supply side, power consumption side and adjustable resource side, and generate a supply and demand status set for the transformer area after preprocessing. The sequence construction module is used to extract power supply capacity, load demand and resource adjustability features from the supply and demand status of the transformer area and construct a multivariable supply and demand status sequence. The phase boundary recombination module is used to calculate the supply and demand security margin and the rate of change of the margin, identify the phase boundary points of supply and demand risks, and recombine to generate multi-scale supply and demand state subsequences. The deviation prediction module is used to input multi-scale supply and demand state sub-sequences into the improved TimeMixer model and generate a supply and demand deviation prediction sequence through three-channel mixing. The influence domain construction module is used to construct the supply and demand scheduling influence domain based on the supply and demand deviation prediction sequence, calculate the scheduling contribution value, and generate the resource scheduling influence set. The scheduling and execution module is used to allocate adjustment tasks based on the resource scheduling impact set, generate and issue a set of scheduling strategies, and update the supply and demand status and impact results in a closed loop.