A reinforcement learning-based dynamic optimization prediction method for power load

Abstract

This invention relates to the field of power load technology, specifically a power load dynamic optimization prediction method based on reinforcement learning. The method collects real-time voltage, frequency, and power data from nodes, tracks the directional changes in frequency differences between adjacent nodes, marks reverse disturbances to generate a distribution map, identifies frequent disturbance areas, clusters nodes with consistent fluctuation trends to establish target areas, monitors power fluctuation reversals to lock energy-shifting nodes, constructs a prediction input set, compares the prediction with the actual trend to extract deviation intervals, adjusts the step size and rate of the reinforcement learning model, updates the decision and smooths the curve, and outputs a dynamic power load prediction curve. This invention achieves spatiotemporal dynamic identification of disturbances through multi-layer correlation analysis, improves judgment accuracy through joint frequency and power calibration, reflects load path consistency and regional extension through node clustering, strengthens power transfer tracking through energy node identification, constructs a feedback loop through offset monitoring to improve curve continuity and response rate, and achieves stable convergence and trend consistency in prediction under multiple disturbances.

Description

Technical Field

[0001] This invention relates to the field of power load technology, and in particular to a method for dynamic optimization and prediction of power load based on reinforcement learning. Background Technology

[0002] The field of power load technology encompasses the analysis and prediction of load changes in power systems. Its core lies in studying the electricity consumption patterns of users at different times and under different operating environments through calculation, modeling, and data acquisition. This field typically relies on multi-source information such as historical load data, meteorological information, holiday characteristics, and socio-economic activities to construct predictive models, thereby describing the dynamic distribution characteristics of power grid load in time and space. The field also involves fundamental research directions such as power dispatching, demand response, and energy management. By systematically characterizing load features, it provides a basis for power grid operation safety, supply-demand balance, and resource allocation, thus forming a technical system centered on power data processing, load modeling, and result inference.

[0003] Among them, the power load dynamic optimization prediction method based on reinforcement learning refers to the technical solution of using reinforcement learning models to intelligently make decisions and model the time series data of power load, and to realize load prediction optimization in a dynamic environment. Its technical aspects cover the entire process from load data feature extraction, state space construction, action strategy generation to prediction result update. By determining the state variables to represent the system characteristics of power load at a given time, the prediction error or prediction accuracy is quantitatively described according to the reward function, and the model parameters are adjusted by strategy iteration or value function update, so that the prediction process can achieve adaptive optimization for load changes in multiple time periods and multiple scenarios during continuous training.

[0004] Existing load forecasting methods are mostly based on static feature extraction and linear modeling. When local disturbances or unsteady energy fluctuations occur in power operation, they are difficult to reflect the directional differences between frequency and voltage in real time. The models are not sensitive enough to the dynamic impact between spatial nodes, which weakens the temporal correlation of load changes. Predictions often show lag or over-smoothing in areas with concentrated local fluctuations. The lack of a continuous identification mechanism for disturbance propagation paths leads to the fragmentation of energy transfer between different time periods. When load offset continues to accumulate, the prediction results deviate from the actual trend, causing errors in load allocation judgment and delays in control feedback. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a method for dynamic optimization and prediction of power load based on reinforcement learning. The technical solution is as follows:

[0006] A power load dynamic optimization prediction method based on reinforcement learning includes the following steps:

[0007] S1: Acquire voltage, frequency and power data of outgoing terminals, switches and branch monitoring points, identify frequency direction changes of adjacent monitoring nodes, mark changes from the same direction to the opposite direction as disturbance events, classify and sort them to form a node event sequence, establish a correspondence table, and generate a load disturbance distribution map;

[0008] S2: Based on the number and density of disturbance events in the load disturbance distribution map, extract the endpoints, power consumption side and segment nodes as analysis nodes, analyze the consistency of frequency and voltage fluctuation direction, form a cluster of converging nodes and integrate them into an identifiable target area, determine the priority according to the fluctuation concentration, and generate a disturbance focus node table.

[0009] S3: Based on the identification information of the converging nodes in the disturbance focusing node table, call the regional input and output power data, compare the fluctuation direction to determine the energy alternation section, determine the node where the power turns and the frequency voltage reverses accordingly as the energy turning node, and generate the load prediction input set;

[0010] S4: Call the turning node features and time-series power in the load prediction input set to form the target area load prediction sequence. Compare it with the actual load change trend. If the deviation exceeds the set threshold and continues to occur, it is marked as a prediction offset interval, and a prediction offset interval sequence is generated.

[0011] As a further embodiment of the present invention, the load disturbance distribution map includes node disturbance density distribution, time interval distribution characteristics and disturbance intensity classification information; the disturbance focusing node table includes target area priority information, node clustering relationship and fluctuation trend characteristic parameters; the load prediction input set includes power characteristic parameter set, energy shift characteristic information and time series change data; and the prediction offset interval sequence includes offset persistence characteristics, offset amplitude distribution and offset occurrence frequency data.

[0012] As a further aspect of the present invention, the step of obtaining the load disturbance distribution map is as follows:

[0013] S101: Obtain real-time voltage, frequency and active power change information of distribution transformer outgoing terminal nodes, distribution switch nodes and branch line monitoring points, arrange the frequency values ​​of each monitoring point in a time series at continuous sampling time, calculate the frequency difference change rate between adjacent sampling times, determine the frequency change direction of each sampling time based on the change rate, and establish a direction mapping matrix with node identifier and sampling time as index to generate a frequency direction mapping table.

[0014] S102: Based on the power distribution line topology, call the frequency direction mapping table to clarify the correspondence between the previous and next monitoring points of each monitoring node in the line. Compare the frequency change direction of adjacent monitoring nodes at the same sampling time to determine whether the direction difference is consistent. Track the continuous change of the difference direction in adjacent sampling times. When the direction state changes from the same direction to opposite direction, record the disturbance identifier. Calculate the number of disturbances of each node in the continuous time period to obtain the disturbance direction mapping table.

[0015] S103: Based on the disturbance direction mapping table, call the corresponding distribution transformer outgoing terminal node, distribution switch node and branch line monitoring point identification information, classify the disturbance event records into nodes, arrange them in chronological order to form a node event sequence, count the number of disturbance events in each time interval according to the time interval, calculate the correspondence between the time interval and the number of events, and generate a load disturbance distribution map.

[0016] As a further aspect of the present invention, the step of obtaining the disturbance focusing node table is as follows:

[0017] S201: Obtain the number of disturbance events and spatial distribution density of each monitoring node in the load disturbance distribution map in different time intervals, calculate the disturbance density value in the line area based on the node's geographical coordinates and the number of events, compare the disturbance density differences between adjacent areas and filter areas with a density exceeding a set threshold, record the spatial range and index number of the area, and generate a disturbance concentration interval.

[0018] S202: Based on the disturbance concentration interval, extract the distribution line endpoint nodes, power consumption side monitoring points and segmented acquisition unit nodes in the corresponding area, record the real-time frequency and voltage change data of each analysis node, calculate the sign difference between the frequency change direction and the voltage change direction, determine the node group with the same sign, summarize the node set with the same direction, and obtain the node group with the same direction.

[0019] S203: Call the group of nodes with consistent direction, and merge the set of nodes with consistent continuous direction to form an identifiable target area based on the spatial distance and frequency direction continuity between nodes. Calculate the concentration of frequency and voltage fluctuation amplitude in each target area, sort them according to the concentration to determine the identification priority, and generate a disturbance focusing node table.

[0020] As a further aspect of the present invention, the step of obtaining the load forecasting input set is as follows:

[0021] S301: Based on the identification information of each identifiable target area and convergent node in the disturbance focusing node table, call the input power and output power data of the corresponding power distribution line area segment, compare the input power change direction and output power change direction at the same sampling time in the same area, record the direction difference status and count the number of direction changes in a continuous period, locate the period when mutual changes occur continuously, and generate energy alternation segments.

[0022] S302: Based on the energy alternation section, detect the points where the input and output power fluctuation directions change in each time period, extract the nodes where the frequency and voltage directions change in opposite directions at the same time, determine the position and state of the nodes during the input and output power switching, and establish a corresponding relationship with the node number and time index to obtain the energy turning node set.

[0023] S303: Call the energy steering node set, sort and reorganize the node set according to the time sequence and power change amplitude of each node, merge the nodes with related characteristics into the input end recording area according to the time sequence, filter out duplicate or invalid items, establish the input parameter sequence based on the power direction change information, and generate the load prediction input set.

[0024] As a further aspect of the present invention, the step of obtaining the predicted offset interval sequence is as follows:

[0025] S401: Call the energy shift node characteristic parameters and corresponding time-series power data in the load forecast input set, calculate the time-series power difference based on the power change direction of each node in the same area, merge the node data with the same fluctuation direction in the same time period, obtain the load power forecast change sequence of different target areas, and arrange the sequence intervals according to the time index to generate the load forecast change sequence.

[0026] S402: Based on the load prediction change sequence, call the real-time load change information of the substation monitoring terminal and the low-voltage bus acquisition equipment, calculate the power deviation between the predicted value and the actual value in the same time interval, and judge the deviation status based on the power deviation threshold, filter out the time intervals in which the deviation direction is consistent and the magnitude exceeds the threshold, and obtain the set of predicted deviation intervals.

[0027] S403: Based on the predicted deviation interval set, rearrange the records of each interval in chronological order, organize and calculate the deviation magnitude and duration within a continuous time period to form a time-series result, and organize and output it according to the deviation time index number to generate a predicted offset interval sequence.

[0028] As a further aspect of the present invention, the method further includes:

[0029] S5: Based on the duration, offset amplitude and frequency of occurrence of each predicted offset interval in the predicted offset interval sequence, adjust the parameters of the reinforcement learning model, import the current load stage state and energy shift node behavior records into the reinforcement learning process, update the model decision to generate a new load trend curve, and straighten out the trend change pattern in continuous time to obtain the dynamic prediction curve of power load.

[0030] The power load dynamic prediction curve includes load change trend characteristics, prediction accuracy evaluation information, and dynamic load time sequence records.

[0031] As a further aspect of the present invention, the step of obtaining the dynamic prediction curve of power load is as follows:

[0032] S501: Based on the duration, offset amplitude, and frequency of occurrence of each predicted offset interval in the predicted offset interval sequence, calculate the weighted ratio of the three indicators for each interval, determine the fluctuation intensity state based on the offset amplitude benchmark value, and perform segment normalization processing on the duration and frequency of occurrence. Based on the calculation results of the three indicators, adjust the action update step size and learning rate parameters to generate a parameter adjustment matrix.

[0033] S502: Based on the parameter adjustment matrix, call the current load stage status information and the behavior records of the analysis nodes and energy shifting nodes in the identifiable target area related to the prediction offset interval, classify and match the node power, frequency and voltage change sequences, adjust the corresponding weights according to the parameter matrix, update the node behavior weight ratio, and obtain the behavior feature weight set.

[0034] S503: Based on the aforementioned behavioral feature weight set, perform trend straightening processing on the power change sequence over a continuous time period, calculate the directional extension degree of power fluctuation and the span of the stable interval for each time period, reconnect the trend segments in chronological order to form a complete curve, record the trend data of the continuous change pattern, and generate a dynamic prediction curve for power load.

[0035] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0036] In this invention, through multi-layer correlation analysis of the relationship between frequency, voltage and power changes in the power distribution network, dynamic identification of disturbance distribution in the time and space dimensions is achieved. The joint calibration of frequency difference direction switching and power fluctuation reverse characteristics improves the accuracy of disturbance event judgment in continuous time periods. The trend cluster formed by node aggregation can reflect the local consistency of load change path and the extension of regional influence. The identification of energy alternation sections and energy turning nodes enhances the traceability of power transfer direction. Predictive offset monitoring constructs a feedback closed loop of load trend. The model parameters are self-adjusted to correct long-term load offset. The continuity and response rate of dynamic curve output are improved, enabling load forecasting to maintain stable convergence and trend consistency under multiple disturbances and uncertain operating conditions. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This is a flowchart illustrating the process of obtaining the load disturbance distribution map according to the present invention.

[0039] Figure 3 This is a flowchart illustrating the process of obtaining the perturbation focusing node table in this invention.

[0040] Figure 4 This is a flowchart illustrating the process of obtaining the load prediction input set for this invention.

[0041] Figure 5 This is a flowchart illustrating the process of obtaining the predicted offset interval sequence according to the present invention.

[0042] Figure 6 This is a flowchart illustrating the process of obtaining the dynamic prediction curve of power load in this invention. Detailed Implementation

[0043] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0044] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0045] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0046] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0047] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0048] Please see Figure 1 This invention provides a technical solution: a method for dynamic optimization and prediction of power load based on reinforcement learning, comprising the following steps:

[0049] S1: Acquire real-time voltage, frequency, and active power change information of distribution transformer outgoing terminal nodes, distribution switch nodes, and branch line monitoring points. Record the frequency change trend of each monitoring point over time during continuous sampling. Based on the distribution line topology, identify the preceding and following monitoring points on the line as adjacent monitoring nodes for each monitoring point. Identify the frequency difference direction of each adjacent monitoring node at the same sampling time and track the continuous change of the frequency difference direction in adjacent sampling times. When the frequency change direction of adjacent monitoring nodes changes from the same direction to opposite directions during a continuous period, it is marked as a disturbance event. Record all disturbance events according to the corresponding distribution transformer outgoing terminal nodes, distribution switch nodes, and branch line monitoring points, and arrange them in chronological order to form a node event sequence. Establish a time interval and event quantity correspondence table and generate a load disturbance distribution map.

[0050] S2: Based on the number of disturbance events and the density of spatial distribution of each monitoring node in different time intervals in the load disturbance distribution map, identify the target area where disturbances occur frequently, extract the distribution line end nodes, power consumption side monitoring points and segmented acquisition unit nodes located in the target area as analysis nodes, record the real-time frequency and voltage change data of each analysis node, identify whether the frequency fluctuation direction of each analysis node is consistent with the corresponding voltage fluctuation direction, determine the sub-segment where the frequency change direction among the analysis nodes in the target area tends to be consistent, mark the set of analysis nodes with the same change trend in each sub-segment as the convergent node group, integrate multiple convergent node groups to form an identifiable target area based on the spatial continuity and trend consistency among the analysis nodes, determine the identification priority based on the concentration of frequency and voltage fluctuation amplitude in each identifiable target area, and generate a disturbance focusing node table;

[0051] S3: Based on the identification information of each identifiable target area and converging node in the disturbance focusing node table, the input power and output power data of the corresponding distribution line area are called, and the input and output fluctuation directions are obtained by comparing them in different time periods. When the fluctuation directions of input power and output power repeatedly change each other in a continuous interval, the continuous interval is determined as the energy alternation section. In the energy alternation section, the nodes where the fluctuation directions of input and output power change and the corresponding frequency and voltage change directions reverse are determined as energy turning nodes. The identified energy turning nodes are marked and reorganized into a set of input items according to the time sequence and power change characteristics to generate the load forecast input set.

[0052] S4: Call the energy shift node characteristic parameters and corresponding time-series power data in the load forecast input set to form a load forecast result sequence for each identifiable target area, and analyze the load power forecast change trend. Compare the obtained load forecast change trend with the actual load change trend obtained by the substation monitoring terminal and low-voltage bus acquisition equipment to determine whether there is a continuous deviation in the change path. When the deviation exceeds the set threshold and continues to occur in multiple consecutive time periods, record the corresponding time period and mark it as the prediction offset interval. Arrange all prediction offset intervals in chronological order to generate a prediction offset interval sequence.

[0053] S5: Based on the duration, offset amplitude, and frequency of occurrence of each predicted offset interval in the predicted offset interval sequence, adjust the action update step size and learning rate parameters in the reinforcement learning model. Import the current load stage state information, as well as the behavior records of analysis nodes and energy shifting nodes in the identifiable target area related to the predicted offset interval into the reinforcement learning process to advance the decision behavior update of the reinforcement learning model, and regenerate the load trend curve. Smooth the trend change pattern over continuous time and output a complete dynamic record to obtain the dynamic prediction curve of power load.

[0054] The load disturbance distribution map includes node disturbance density distribution, time interval distribution characteristics, and disturbance intensity classification information. The disturbance focus node table includes target area priority information, node clustering relationships, and fluctuation trend characteristic parameters. The load forecast input set includes power characteristic parameter set, energy shift characteristic information, and time series change data. The forecast offset interval sequence includes offset persistence characteristics, offset amplitude distribution, and offset occurrence frequency data. The power load dynamic forecast curve includes load change trend characteristics, forecast accuracy evaluation information, and dynamic load time series records.

[0055] Please see Figure 2 The steps for obtaining the load disturbance distribution map are as follows:

[0056] S101: Obtain real-time voltage, frequency and active power change information of distribution transformer outgoing terminal nodes, distribution switch nodes and branch line monitoring points, arrange the frequency values ​​of each monitoring point in a time series at continuous sampling time, calculate the frequency difference change rate between adjacent sampling times, determine the frequency change direction of each sampling time based on the change rate, and establish a direction mapping matrix with node identifier and sampling time as index to generate a frequency direction mapping table.

[0057] Real-time voltage, frequency, and active power changes are acquired at distribution transformer outgoing terminals, distribution switch nodes, and branch line monitoring points. High-precision synchronous phasor measurement units (PMUs) or intelligent circuit breakers with high-frequency sampling capabilities are deployed as data acquisition devices at key monitoring locations at the transformer outgoing terminals, switches controlling current flow, and branch lines at the end of the distribution network. The sampling frequency is set to 50 times per second, i.e., the sampling interval... The acquisition device reads the instantaneous voltage values ​​of each node in real time, with a time interval of 0.02 seconds. Instantaneous frequency value and active power The collected frequency data is stored in a first-in-first-out queue according to the chronological order, forming a queue of length [length missing]. Time series data For example, in the first At a certain sampling time, the frequency value of a monitoring point on a certain branch line was 50.02Hz, while... The frequency at time 50.01Hz is at which point a subtraction operation is performed. The frequency difference was found to be 0.01 Hz. Dividing this difference by the sampling interval of 0.02 seconds yielded the rate of change of the frequency difference. Set a frequency fluctuation judgment threshold This threshold is based on the standard deviation of frequency fluctuations during normal grid operation. Settings, take Assuming historical data shows a normal fluctuation standard deviation of 0.005 Hz / s, then set... , calculate and and Perform numerical comparisons, if If the frequency is determined to be increasing at that moment, it is recorded as state flag "1". If the frequency of judgment shows a decreasing trend, it is recorded as a status flag "-1". If the absolute value If the frequency remains stable, it is marked as state identifier "0". Repeat the above comparison and assignment operation for all sampling times of each monitoring node to construct a two-dimensional matrix with the number of rows equal to the total number of monitoring nodes and the number of columns equal to the total number of sampling points. Fill the corresponding cells of the matrix with the state values ​​of "1", "-1" or "0" of each node at each time. The row index is associated with the unique ID of each node, and the column index is associated with the corresponding timestamp to generate a frequency direction mapping table.

[0058] S102: Based on the power distribution line topology, call the frequency direction mapping table to clarify the correspondence between the previous and next monitoring points of each monitoring node in the line. Compare the frequency change direction of adjacent monitoring nodes at the same sampling time to determine whether the direction difference is consistent. Track the continuous change of the difference direction in adjacent sampling times. When the direction state changes from the same direction to opposite direction, record the disturbance identifier. Calculate the number of disturbances of each node in the continuous time period to obtain the disturbance direction mapping table.

[0059] Based on the power distribution line topology, the frequency direction mapping table is invoked, and the adjacency table or database storing line connection relationships is read. Then, based on the node ID connection pairs recorded in the adjacency table, the node to be analyzed is located. Upstream directly connected nodes As the previous monitoring point and the directly connected downstream node As the next monitoring point, extract from the frequency direction mapping table and At the same sampling time Direction indicator value and Perform multiplication operations ,like A positive result indicates that the frequencies change in the same direction, while a negative result indicates that they change in opposite directions. For example... The value is 1 (increase). If it is -1 (decreasing), then A value of -1 indicates a discrepancy in direction; in this case, the state at that moment is temporarily stored as "reverse," and the next sampling moment is retrieved. The direction indicator is used to repeat the above comparison process for comparison. Time and The state of the node at time t, if The time node pair is in the same direction (i.e.) ),and The time node transitions to the "reverse" state (i.e., If a directional change has occurred, a counter variable is set. The initial value is 0. Each time a transition from the same direction to the opposite direction is detected, [the value will be set to 0]. Increment the value by 1 and record the moment the mutation occurred. and the node IDs involved, for and The above difference comparison and jump tracking steps are repeated in combination. The total number of directional changes that occur between a node and all its neighboring nodes within a continuous monitoring period is accumulated. For example, if a switch node changes from the same direction to the opposite direction 5 times with the upstream node and from the same direction to the opposite direction 3 times with the downstream node within 1 minute, then the cumulative disturbance count value of the node in this period is 8. The cumulative disturbance value calculated for each node is associated with the corresponding node ID and time period and filled into a new data table to obtain the disturbance direction mapping table.

[0060] S103: Based on the disturbance direction mapping table, call the corresponding distribution transformer outgoing terminal node, distribution switch node and branch line monitoring point identification information, classify the disturbance event records into nodes, arrange them in chronological order to form a node event sequence, count the number of disturbance events in each time interval according to the time interval, calculate the correspondence between the time interval and the number of events, and generate a load disturbance distribution map.

[0061] Based on the disturbance direction mapping table, the system retrieves the corresponding identification information of distribution transformer outgoing node, distribution switch node, and branch line monitoring point. It iterates through each record in the disturbance direction mapping table, reading the unique node identification code contained in the record. This identification code is then matched against a pre-set equipment attribute database. If the attribute corresponding to the identification code is "Transformer_Out", it is classified as a transformer outgoing node; if it is "Switch_Node", it is classified as a distribution switch; and if it is "Branch_Monitor", it is classified as a branch line. Three independent list containers are created to store the categorized disturbance event records. Records in each list are then sorted by timestamp. Sort the data in ascending order of size to ensure events are arranged chronologically, and set the time analysis window length. For example, take The entire monitoring period is divided into seconds (5 minutes). a continuous time interval For each time interval The system iterates through the lists of each category and counts the number of disturbance event records falling within the time interval. For example, if the transformer outgoing terminal list contains 12 records marked as disturbances in the interval from 10:00 to 10:05, then the number of disturbances at the transformer node in that interval is 12. After counting the number of disturbances for all nodes in all time intervals, a dataset containing four fields is constructed: "start point of time interval", "node ID", "node type" and "number of disturbances". The time interval is used as the horizontal axis variable, the node's geographical location or logical topological location is used as the vertical axis variable, and the count of disturbances is used as the numerical intensity. A three-dimensional data matrix or heat map data source is established, and the numerical magnitude reflects the degree of disturbance density, generating a load disturbance distribution map.

[0062] Please see Figure 3 The steps to obtain the perturbation focus node table are as follows:

[0063] S201: Obtain the number of disturbance events and spatial distribution density of each monitoring node in the load disturbance distribution map in different time intervals, calculate the disturbance density value in the line area based on the node's geographical coordinates and the number of events, compare the difference in disturbance density between adjacent areas and filter areas with a density exceeding the set threshold, record the spatial range and index number of the area, and generate a disturbance concentration interval.

[0064] Obtain the number of disturbance events and spatial density of each monitoring node in the load disturbance distribution map within different time intervals, and extract the cumulative disturbance count value of each node in the distribution map within a specific time window. and the geographic location coordinates of the node (These coordinates are obtained through a GIS geographic information system or GPS positioning module), with a radius set centered on each monitoring node. The circular scan area, for example, taking The disturbance count is calculated by iterating through all neighboring monitoring nodes within the circular area, and summing the disturbance counts of the central node and all neighboring nodes to obtain the total disturbance count for the area. Using the formula Calculate the disturbance density value of this area. Assuming there are 5 nodes within a radius of 1.5 kilometers in a certain area, and the total number of disturbances is 150, then... Times / square kilometer, setting a density screening threshold. This threshold is taken as the historical average density of the entire distribution network. 1.5 times that, if the historical average is 10 times per square kilometer, then times per square kilometer, the calculated results for each region Value and Compare them one by one and filter out all of them. For high-density areas, calculate the distance between the center points of any two spatially adjacent high-density areas. If the distance is less than... If the two regions are found to have spatial overlap or connection, the spatial boundary coordinates of the two regions are combined into a continuous geographical range. The maximum and minimum latitude and longitude coordinates and the corresponding time interval index number of the combined range are recorded to generate the disturbance concentration interval.

[0065] S202: Based on the disturbance concentration area, extract the distribution line endpoint nodes, power consumption side monitoring points and segmented acquisition unit nodes in the corresponding area, record the real-time frequency and voltage change data of each analysis node, calculate the sign difference between the frequency change direction and the voltage change direction, determine the node group with the same sign, summarize the node set with the same direction, and obtain the node group with the same direction.

[0066] Based on the concentrated disturbance range, the distribution line endpoint nodes, power consumption side monitoring points, and segmented acquisition unit nodes within the corresponding area are extracted. Using the recorded latitude and longitude coordinates, a search is performed in the power grid asset database to identify all key node IDs falling within this range, defining them as an analysis node set. For each analysis node in the set, its real-time frequency sequence within the corresponding time interval is retrieved from the historical database of the Distribution Automation System (DAS). and voltage sequence Calculate the difference between adjacent time points. and Extract the sign attribute of the difference, if The frequency direction sign is denoted as "+" and the opposite as "-". The voltage direction sign is handled similarly, and the frequency direction sign and voltage direction sign of a node at the same moment are combined to form a feature vector. For example, if the frequency and voltage of a node decrease at time t, its eigenvector is (-,-). Then, when analyzing any two nodes in the node set... and Compare the feature vectors, if and If they are completely identical, then the two nodes are determined to have the same oscillation characteristics at the current time. All nodes that satisfy the consistency of feature vectors are added to the same temporary list. For example, all nodes with the feature (-,-) are grouped together. The empty set that does not contain any nodes in the list is removed. Each of the remaining lists is defined as an independent set unit, resulting in a group of nodes with the same direction.

[0067] S203: Call the node group with consistent direction, determine the spatial distance and frequency direction continuity between nodes, merge the node set with consistent continuous direction to form an identifiable target area, calculate the concentration of frequency and voltage fluctuation amplitude in each target area, sort them according to the concentration to determine the identification priority, and generate a disturbance focusing node table.

[0068] The system calls a node group with consistent calling direction. Based on the spatial distance and frequency direction continuity between nodes, it reads the topological connection relationship and line length data of each member node in the node group and calculates the electrical distance between any two nodes in the group. Set a threshold for spatial continuity determination If the electrical distance between a node in a group and any other node in the group is less than 500 meters and there is a direct or indirect connection path on the topology map, then spatial continuity is confirmed. Multiple groups of nodes that satisfy spatial continuity and are aligned in the same direction are merged. For example, if node group A is located on the main line and node group B is located on a branch extending from the main line, and the two are connected by a switch at a distance of only 200 meters, then A and B are merged. The merged area is defined as the identifiable target area. For each identifiable target area, the mean absolute value of the frequency fluctuation amplitude of all nodes within the area during the observation period is calculated. The mean of the absolute values ​​of voltage fluctuation amplitude ;

[0069] Constructing Concentration Assessment Indicators The weighting coefficient Set it to 0.6 (unit: 1 / Hz). Set to 0.4 (unit: 1 / V) to eliminate the dimensional difference between frequency and voltage and achieve normalized weighting. Assume a target region... Hz, V;

[0070] but The calculated values ​​for all identifiable target areas The values ​​are sorted in descending order; larger values ​​represent more concentrated fluctuation energy, and these are placed at the top of the list as high-priority targets. The sorted target area ID, the list of contained nodes, and... Values ​​are stored in a structured manner, generating a table of perturbation-focused nodes.

[0071] Please see Figure 4 The steps for obtaining the load forecast input set are as follows:

[0072] S301: Based on the identification information of each identifiable target area and convergent node in the disturbance focusing node table, call the input power and output power data of the corresponding power distribution line area segment, compare the input power change direction and output power change direction at the same sampling time in the same area, record the direction difference status and count the number of direction changes in a continuous period, locate the period when the mutual changes occur continuously, and generate energy alternation segments.

[0073] Based on the identification information of each identifiable target area and converging node in the disturbance focusing node table, the input and output power data of the corresponding power distribution line segment are retrieved. The active power sequence of the target area's power supply side (low-voltage side of the transformer or the beginning of the line) is obtained from the SCADA system or bidirectional smart meters installed on the low-voltage side of the transformer and at the end of the line, aligned with the timestamp. As input power, and active power sequence on the load side (regional end or branch aggregation). As the output power, it is obtained by performing a difference operation on the values ​​of adjacent time points in the sequence. and Through symbolic functions and Determine the direction of change;

[0074] like Then the direction indicator is 1, if The direction indicator is then set to -1, and comparisons are performed time-by-time. and ,like Record this moment as the "direction divergence point". Recorded as "direction coordination point", within the set sliding window time length. Within a 30-second period, count the number of times the "direction divergence point" occurs. Set a deviation frequency threshold Take 30% of the total number of sampling points in the window (assuming a 30-second window contains 1500 sampling points). If a certain window contains This indicates that the input-output power fluctuation relationship is extremely unstable during this period, with energy oscillations or reverse power flow present. The time range corresponding to this window is marked. And merge all consecutive or overlapping marked windows on the timeline, that is, merge the intervals. and merged into Record the start and end times of the merged time period to generate energy alternation segments.

[0075] S302: Based on the energy alternation section, detect the points where the direction of input and output power fluctuation changes in each time period, extract the nodes where the frequency and voltage direction change in opposite directions at the same time, determine the position and state of the nodes during the switching of input and output power, and establish a corresponding relationship with the node number and time index to obtain the set of energy turning nodes.

[0076] Based on the energy alternation intervals, the points of change in the direction of input and output power fluctuations within each time period are detected to lock the time range of the energy alternation intervals. Within this range, the direction of input power is scanned point by point. The moment of sign flip (i.e., from 1 to -1 or from -1 to 1), and the direction of output power. The moments when the sign flips are observed are defined as the "power inflection point set". ,against Every moment in Backtrack and query each analysis node in this area Indicator of frequency change direction at any moment and voltage change direction indicator and the sign of the previous moment. and Check for the presence of reverse change characteristics, i.e., determine whether the conditions are met: and If this condition is met, it indicates that the voltage frequency characteristics of the node also underwent a synchronous reversal at the moment of power switching. This node is thus identified as a critical node sensitive to energy fluctuations. The physical location identifier of the node (e.g., line branch point A, user access point B) is recorded, and combined with the current power flow direction (input to output or output feedback to input), a location label "Source_Side" (source-side response) or "Load_Side" (load-side response) is assigned to the node. A structure containing "node ID" and "occurrence time" is then constructed. "Power change amplitude" The record entries with the four attributes of "position label" are summarized to obtain the energy turning node set.

[0077] S303: Call the energy steering node set, sort and reorganize the node set according to the time sequence and power change amplitude of each node, merge the nodes with related characteristics into the input end recording area according to the time sequence, filter out duplicate or invalid items, establish the input parameter sequence based on the power direction change information, and generate the load forecast input set.

[0078] The energy redirection node set is retrieved, and the set is sorted and reorganized according to the time sequence and power change amplitude of each node, based on the recorded occurrence times. Sort all entries in the set in ascending order. For multiple records with identical timestamps, sort them according to the absolute value of the power change amplitude. Perform a two-level descending sort, prioritizing the retention of node records with large power surges, execute data cleaning operations, and set a time-based deduplication threshold. If the time difference between two adjacent records is less than 0.5 seconds and they have the same node ID, the record with the larger amplitude is retained, and the record with the smaller amplitude is discarded to eliminate redundant data caused by sampling jitter. Node feature parameters, including node ID and time, are extracted from the cleaned records. Power amplitude, voltage frequency fluctuation state values, these parameters are assembled into a feature vector sequence according to the time series. Each of them A complete feature description representing an energy shift event is generated, and the corresponding power direction change identifier (1 or -1) is used as the sequence guide label. A standard data row in the format of "timestamp-feature vector-direction label" is constructed. All standard data rows are stored in a structured data table as the direct input source for subsequent prediction models, generating the load prediction input set.

[0079] Please see Figure 5 The steps for obtaining the predicted offset interval sequence are as follows:

[0080] S401: Call the energy shift node characteristic parameters and corresponding time series power data in the load forecast input set, calculate the time series power difference based on the power change direction of each node in the same area, merge the node data with the same fluctuation direction in the same time period, obtain the load power forecast change sequence of different target areas, and arrange the sequence intervals according to the time index to generate the load forecast change sequence.

[0081] The system retrieves the energy shifting node characteristic parameters and corresponding time-series power data from the load forecast input set. It then groups all node data in the input set according to their target region ID; for example, nodes belonging to region A are grouped together, and those belonging to region B are grouped together. Within each group, the system extracts the data for each node. Power change value at time ,in Represents the node index, calculates the sign of the average power change direction for all nodes within that region. Iterate through each node in the group and check the direction of its individual changes. Whether or not Consistent, if Then retain the node. The values ​​are included in the aggregation calculation. If they are inconsistent, they are considered local noise and removed. The remaining node power change values ​​are then summed to obtain the value for that region. Total predicted power variation at time point With time step Calculate the future continuously at intervals of 5 minutes (e.g.). Each time step (e.g., the next hour, i.e., 12 steps) The value constitutes a length of Predicted change sequence:

[0082] ;

[0083] The above operations are performed on all target regions to obtain multiple sets of parallel predicted change sequences. Finally, these sequences are sorted according to their start timestamps. The data is sorted in ascending order. If there is time overlap, each data point retains its independent time index label to generate a load forecast change sequence.

[0084] S402: Based on the load forecast change sequence, call the real-time load change information of the substation monitoring terminal and the low-voltage bus acquisition equipment, calculate the power deviation between the predicted value and the actual value in the same time interval, and judge the deviation status based on the power deviation threshold. Filter out time periods in which the deviation direction is consistent and the magnitude exceeds the threshold in multiple consecutive intervals to obtain the set of predicted deviation intervals.

[0085] Based on the load forecast change sequence, real-time load change information from the substation monitoring terminal and low-voltage busbar acquisition equipment is retrieved to obtain the measured active power value from the substation RTU (Remote Terminal Unit) or the low-voltage busbar side gate meter within the corresponding time period. ;

[0086] And calculate the measured power change. This corresponds to each prediction point in the load forecast change sequence. Calculate the power deviation value Set power deviation threshold The threshold is taken as the rated capacity of the area. 2%, assuming the rated capacity of the area is 1000kW, then kW, check the absolute value of the deviation at each time step ,like If a significant deviation is detected at a given moment, the deviation status is recorded as 1; otherwise, it is recorded as 0. The process continues by scanning consecutive time intervals, searching for instances where all deviation status indicators are 1 and the deviation direction is (…). (the sign of the sign) remains consistent for a period of time, for example from arrive If the deviation of five consecutive sampling points is greater than 20kW and is positive (the predicted value is consistently high), then the interval will be... Mark it as a candidate deviation event and calculate the duration of this continuous interval. Set a minimum duration threshold For 3 sampling periods (i.e., 15 minutes), if If the interval is confirmed as a valid deviation, the start time, end time, average deviation magnitude, and deviation direction sign of the interval are recorded. All confirmed valid deviation intervals are stored in a list to obtain the set of predicted deviation intervals.

[0087] S403: Based on the set of predicted deviation intervals, rearrange the records of each interval in chronological order, organize and calculate the deviation magnitude and duration within a continuous time period, form a time-series result, and organize and output it according to the deviation time index number to generate a predicted offset interval sequence.

[0088] Based on the predicted deviation interval set, rearrange the records of each interval in chronological order, and read the start timestamp of each interval in the set. Using either bubble sort or quick sort algorithm, based on Reorder all interval records in ascending order to ensure the correct chronological order on the timeline. For the sorted list, check whether adjacent intervals have temporal continuity or overlap. If intervals... End time and interval start time The time difference between them is less than one sampling period. If the two intervals have the same sign of deviation direction, then merge these two intervals into a larger interval and update the starting time of the merged interval. The end time is And recalculate the average deviation magnitude of the merged interval:

[0089] ;

[0090] After the merging operation is completed, a unique sequence number is assigned to each of the finally determined independent intervals. (e.g., SEQ_001, SEQ_002…), integrate the attributes such as number, start and end time, duration, average deviation magnitude, and deviation direction into a standard output format, write them line by line into the result file in numerical order, and generate a predicted offset interval sequence.

[0091] Please see Figure 6 The steps for obtaining the dynamic forecast curve of power load are as follows:

[0092] S501: Based on the duration, offset amplitude, and frequency of occurrence of each predicted offset interval in the predicted offset interval sequence, calculate the weighted ratio of the three indicators for each interval, determine the fluctuation intensity state based on the offset amplitude benchmark value, and perform segment normalization on the duration and frequency of occurrence. Based on the calculation results of the three indicators, adjust the action update step size and learning rate parameters to generate a parameter adjustment matrix.

[0093] Based on the duration, offset amplitude, and frequency of occurrence of each predicted offset interval in the predicted offset interval sequence, each predicted offset interval in the sequence is read. Extract duration from attribute data. (Unit: minutes) Average offset (Unit: kilowatts) and the number of times this type of offset has occurred in the past 24 hours. Set the offset range reference value This benchmark value is taken as 5% of the rated capacity of the distribution transformer. For example, if the transformer capacity is 2000kVA, then... kW, calculate wave intensity coefficient ,like kW, then At the same time, a reference value for the maximum duration is set. Reference values ​​for minutes and maximum frequency Next, using the formula and The duration and frequency of occurrence were normalized separately to obtain normalized values. and Construct comprehensive impact indicators The weighting coefficients are set as follows:

[0094] ;

[0095] Suppose a certain interval ,but ;

[0096] Calculated For reinforcement learning models (such as Q-learning or DQN), retrieve the current learning rate parameter. (Initially set to 0.01) and action update stride (Initially set to 0.1), based on comprehensive impact indicators Dynamically adjust the parameters when If the error is identified as a high-impact bias, the parameters need to be increased to speed up the model's adaptation. New parameters need to be calculated. and Set adjustment factor Substituting the numerical values, we get , It will be targeted at different time periods. Calculated and Numerical values ​​are stored in a two-dimensional array according to time index. The rows of the array correspond to the time step, and the columns correspond to the parameter types, generating a parameter adjustment matrix.

[0097] S502: Based on the parameter adjustment matrix, call the current load stage status information and the behavior records of the analysis nodes and energy shifting nodes in the identifiable target area related to the prediction offset interval, classify and match the node power, frequency and voltage change sequences, adjust the corresponding weights according to the parameter matrix, update the node behavior weight ratio, and obtain the behavior feature weight set.

[0098] Based on the parameter adjustment matrix, the system retrieves the current load phase status information, as well as the behavior records of identifiable target area analysis nodes and energy shifting nodes related to the prediction offset interval, and loads the current forecast period from the database. Power grid load basic state vector (Including total load level, ambient temperature, and date type), and retrieve data within the predicted offset range. All active analysis nodes and energy shifting node IDs during the occurrence period were analyzed, and the power sequences of these nodes within the corresponding period were extracted. Frequency sequence and voltage sequence Calculate the variance of each sequence. To quantify node activity, the variance value exceeding a set threshold (such as power variance) is considered. The nodes are marked as "strongly correlated nodes", and the correction parameters for the corresponding time period are read from the parameter adjustment matrix. and ;

[0099] Constructing weight adjustment coefficients using parameter values ;

[0100] For example For each strongly correlated node, retrieve its original feature weights in the reinforcement learning state space. Perform multiplication operations Weight amplification is applied; if the original weight of an energy redirection node is 0.3, the updated weight will be... For nodes not marked as strongly correlated, keep the weights unchanged or decrease them proportionally to maintain the normalization of the total weights, and update the weight values ​​of all nodes. The node's corresponding feature identifiers (such as "Node A - Voltage Fluctuation" and "Node B - Power Reversal") are bound to key-value pairs to form a new feature importance list. This list accurately reflects the adjustment of each node's contribution to load forecasting under the current disturbance environment, thus obtaining a set of behavioral feature weights.

[0101] S503: Based on the behavioral feature weight set, the power change sequence over a continuous time period is processed for trend straightening, the directional extension of power fluctuation and the span of stable intervals in each time period are calculated, the trend segments are reconnected in chronological order to form a complete curve, the trend data of continuous change patterns are recorded, and a dynamic prediction curve of power load is generated.

[0102] Based on the behavioral feature weight set, the power change sequence over a continuous time period is trend-normalized. Using the updated behavioral feature weight set as input weights, the load forecasting algorithm is re-run to calculate the preliminary power points for future time periods. This forms the original prediction sequence, and for each time point in the sequence... , cut off A local window is formed by five points before and after the center, and the slope of the data within the window is calculated. ,judge The sign continuity is determined by calculating the start and end time difference of a consecutive segment if eight consecutive points within the window have the same slope sign (both positive or both negative). This difference is used as the "degree of directional extension". ,like If the fluctuation is less than the minimum extension threshold (e.g., 10 minutes), it is determined to be noise, and the moving average method is used to smooth and correct the data segment. Simultaneously identify the absolute value of the slope in the sequence. For continuous periods of kW / min, the duration is recorded as the "stable interval span". When connecting two paragraphs with different trends, check the connection point. The slope changes abruptly at a certain point, if the difference between the slopes before and after is significant. To address this, a Bézier curve interpolation algorithm is introduced, adding three transition points before and after the connection point. The power values ​​at these transition points are recalculated to achieve a smooth transition. All processed piecewise curves are then stitched together sequentially along the time axis. For any potential numerical jumps at the stitching points, a linear weighted fusion method is used for fine-tuning to ensure... The power value at any given time is unique and continuous. The final output is a continuous data stream containing timestamps and smoothed power values, generating a dynamic power load prediction curve.

[0103] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A power load dynamic optimization prediction method based on reinforcement learning, characterized in that, The method comprises the following steps: S1: acquiring voltage, frequency and power data of the outlet end, switch and branch monitoring point, identifying frequency direction change of adjacent monitoring nodes, marking disturbance events in the same direction as reverse, classifying and sorting to form a node event sequence, establishing a corresponding table, and generating a load disturbance distribution map; S2: according to the number and density of disturbance events in the load disturbance distribution map, extracting the end point, power consumption side and segmented node as an analysis node, analyzing the consistency of frequency and voltage fluctuation direction, forming a convergent node group and integrating it into an identifiable target area, determining the priority according to the fluctuation concentration, and generating a disturbance focusing node table; S3: based on the identification information of the convergent nodes in the disturbance focusing node table, calling the regional input and output power data, comparing the fluctuation direction to determine the energy exchange section, determining the nodes with power turning and frequency and voltage reversing as energy turning nodes, and generating a load prediction input set; S4: calling the turning node characteristics and time sequence power in the load prediction input set, forming a target area load prediction sequence, comparing the actual load change trend, and if the deviation exceeds the set threshold and continues to appear, marking it as a prediction deviation interval, and generating a prediction deviation interval sequence; S5: according to the duration, deviation amplitude and occurrence frequency characteristics of each prediction deviation interval in the prediction deviation interval sequence, adjusting the reinforcement learning model parameters, importing the current load stage state and energy turning node behavior record into the reinforcement learning process, updating the model decision to generate a new load trend curve, and performing smoothing processing on the trend change form in continuous time to obtain a power load dynamic prediction curve; The acquisition step of the power load dynamic prediction curve is: S501: according to the duration, deviation amplitude and occurrence frequency characteristics of each prediction deviation interval in the prediction deviation interval sequence, calculating the weighted ratio of the three indexes of each interval, judging the fluctuation intensity state according to the deviation amplitude reference value, and performing section normalization processing on the duration and occurrence frequency, adjusting the action update step and learning rate parameters according to the operation results of the three indexes, and generating a parameter adjustment matrix; S502: based on the parameter adjustment matrix, calling the current load stage state information, and the behavior record of the analysis node and the energy turning node in the identifiable target area related to the prediction deviation interval, classifying and matching the node power, frequency and voltage change sequence, and updating the node behavior weight proportion according to the parameter matrix adjustment corresponding weight, to obtain a behavior characteristic weight set; S503: according to the behavior characteristic weight set, performing trend smoothing processing on the power change sequence in continuous time, calculating the direction extension degree and stable interval span of power fluctuation in each time period, reconnecting the trend paragraphs in time sequence to form a complete curve, recording the trend data of continuous change form, and generating a power load dynamic prediction curve; The load disturbance distribution map includes node disturbance density distribution, time interval distribution characteristics and disturbance intensity classification information, the disturbance focusing node table includes target area priority information, node clustering relationship and fluctuation trend characteristic parameters, the load prediction input set includes power characteristic parameter set, energy steering characteristic information and time sequence change data, the prediction offset interval sequence includes offset duration characteristics, offset amplitude distribution and offset occurrence frequency data, and the power load dynamic prediction curve includes load change trend characteristics, prediction accuracy evaluation information and dynamic load time sequence record; The specific method of identifying the frequency direction change of adjacent monitoring nodes and marking the same direction change as a reverse direction as a disturbance event comprises the following steps: obtaining a frequency difference change rate, setting a frequency fluctuation determination threshold, comparing the frequency difference change rate with the frequency fluctuation determination threshold, and assigning a state identifier accordingly, determining that the frequency is rising if the frequency difference change rate is greater than the frequency fluctuation determination threshold, assigning a positive number as the state identifier, determining that the frequency is falling if the frequency difference change rate is less than the negative frequency fluctuation determination threshold, assigning a negative number as the state identifier, and determining that the frequency remains stable if the absolute value of the frequency difference change rate is less than or equal to the frequency fluctuation determination threshold, assigning zero as the state identifier, repeating the process for all sampling time points of each monitoring node, and generating a frequency direction mapping table; According to the power distribution line topology relationship, the data in the frequency direction mapping table is read, the state identifier of one monitoring node at each time point is multiplied by the state identifier of the connected monitoring node to obtain the frequency change direction, the frequency change directions of two consecutive time points are compared, if the frequency change directions are different, it means that the frequency change directions of the two monitoring nodes are inconsistent, i.e. the same direction changes to the reverse direction, otherwise, it is the reverse direction, at this time, the time point is marked as a reverse state, if the frequency change directions of the two monitoring nodes change from the same direction to the reverse direction, it is determined that the direction suddenly changes, and the time point and the involved monitoring node ID are recorded, and marked as a disturbance event; The specific method of identifying the frequency direction change of adjacent monitoring nodes and marking the same direction change as a reverse direction as a disturbance event comprises the following steps: obtaining a frequency difference change rate, setting a frequency fluctuation determination threshold, comparing the frequency difference change rate with the frequency fluctuation determination threshold, and assigning a state identifier accordingly, determining that the frequency is rising if the frequency difference change rate is greater than the frequency fluctuation determination threshold, assigning a positive number as the state identifier, determining that the frequency is falling if the frequency difference change rate is less than the negative frequency fluctuation determination threshold, assigning a negative number as the state identifier, and determining that the frequency remains stable if the absolute value of the frequency difference change rate is less than or equal to the frequency fluctuation determination threshold, assigning zero as the state identifier, repeating the process for all sampling time points of each monitoring node, and generating a frequency direction mapping table; The specific method of identifying the frequency direction change of adjacent monitoring nodes and marking the same direction change as a reverse direction as a disturbance event comprises the following steps: obtaining a frequency difference change rate, setting a frequency fluctuation determination threshold, comparing the frequency difference change rate with the frequency fluctuation determination threshold, and assigning a state identifier accordingly, determining that the frequency is rising if the frequency difference change rate is greater than the frequency fluctuation determination threshold, assigning a positive number as the state identifier, determining that the frequency is falling if the frequency difference change rate is less than the negative frequency fluctuation determination threshold, assigning a negative number as the state identifier, and determining that the frequency remains stable if the absolute value of the frequency difference change rate is less than or equal to the frequency fluctuation determination threshold, assigning zero as the state identifier, repeating the process for all sampling time points of each monitoring node, and generating a frequency direction mapping table; The specific method of identifying the frequency direction change of adjacent monitoring nodes and marking the same direction change as a reverse direction as a disturbance event comprises the following steps: obtaining a frequency difference change rate, setting a frequency fluctuation determination threshold, comparing the frequency difference change rate with the frequency fluctuation determination threshold, and assigning a state identifier accordingly, determining that the frequency is rising if the frequency difference change rate is greater than the frequency fluctuation determination threshold, assigning a positive number as the state identifier, determining that the frequency is falling if the frequency difference change rate is less than the negative frequency fluctuation determination threshold, assigning a negative number as the state identifier, and determining that the frequency remains stable if the absolute value of the frequency difference change rate is less than or equal to the frequency fluctuation determination threshold, assigning zero as the state identifier, repeating the process for all sampling time points of each monitoring node, and generating a frequency direction mapping table; ； wherein, is a fluctuation intensity coefficient, is a duration normalization index, is a frequency of occurrence normalization index, is a comprehensive influence index, , and are weight coefficients; For the reinforcement learning model, the current learning rate parameter is called and the action update step size , according to the comprehensive influence index to dynamically correct the parameters; The specific method of identifying the frequency direction change of adjacent monitoring nodes and marking the same direction change as a reverse direction as a disturbance event comprises the following steps: obtaining a frequency difference change rate, setting a frequency fluctuation determination threshold, comparing the frequency difference change rate with the frequency fluctuation determination threshold, and assigning a state identifier accordingly, determining that the frequency is rising if the frequency difference change rate is greater than the frequency fluctuation determination threshold, assigning a positive number as the state identifier, determining that the frequency is falling if the frequency difference change rate is less than the negative frequency fluctuation determination threshold, assigning a negative number as the state identifier, and determining that the frequency remains stable if the absolute value of the frequency difference change rate is less than or equal to the frequency fluctuation determination threshold, assigning zero as the state identifier, repeating the process for all sampling time points of each monitoring node, and generating a frequency direction mapping table; 2. The method of claim 1, wherein the method is based on reinforcement learning. ​ S101: Real-time voltage values, frequency values and active power change information of the outgoing terminal node of the distribution transformer, the distribution switch node and the branch line monitoring point are acquired, the frequency values of each monitoring point at continuous sampling time are arranged in time sequence, the frequency difference change rate of adjacent sampling time is calculated, the frequency change direction of each sampling time is determined according to the change rate, and a direction mapping matrix is established with the node identifier and the sampling time as the index to generate a frequency direction mapping table; S102: According to the topological relationship of the distribution line, the frequency direction mapping table is called to determine the corresponding relationship between the previous monitoring point and the next monitoring point of each monitoring node in the line, the frequency change direction of adjacent monitoring nodes at the same sampling time is compared, whether the direction difference is consistent is judged, and the continuation change of the difference direction in adjacent sampling time is tracked, the disturbance identifier is recorded when the direction state changes from the same direction to the reverse direction, the disturbance times of each node in the continuous period are calculated, and a disturbance direction mapping table is obtained; S103: Based on the disturbance direction mapping table, the corresponding identifier information of the outgoing terminal node of the distribution transformer, the distribution switch node and the branch line monitoring point is called to classify the disturbance event records, the node event sequence is arranged in time sequence, the number of disturbance events in each interval is counted, and the corresponding relationship between the time interval and the number of events is calculated to generate a load disturbance distribution diagram.

3. The method of claim 1, wherein: The acquisition step of the disturbance focusing node table is: S201: The number of disturbance events and the spatial distribution density of each monitoring node in different time intervals in the load disturbance distribution diagram are acquired, the disturbance density value in the line area is calculated according to the node geographic coordinates and the number of events, the disturbance density difference of adjacent areas is compared, and the area with the disturbance density exceeding the set threshold is selected, the spatial range and index number of the area are recorded, and a disturbance concentrated interval is generated; S202: According to the disturbance concentrated interval, the distribution line endpoint node, the monitoring point on the power consumption side and the segmented collection unit node in the corresponding area are extracted, the real-time frequency and voltage change data of each analysis node are recorded, the sign difference of the frequency change direction and the voltage change direction is calculated, and the node group with the same sign is judged, the node set with the same direction is summarized, and the direction consistent node group is obtained; S203: The direction consistent node group is called, the spatial distance between nodes and the frequency direction continuity are determined, the node set with continuous consistent direction is merged to form an identifiable target area, the concentration degree of the frequency and voltage fluctuation amplitude in each target area is calculated, the recognition priority is determined according to the concentration degree, and a disturbance focusing node table is generated.

4. The method of claim 1, wherein: The acquisition step of the load prediction input set is: S301: Based on the identifier information of each identifiable target area and the converging node in the disturbance focusing node table, the input power and output power data of the corresponding distribution line area segment are called, the input power change direction and the output power change direction at the same sampling time in the same area are compared, the direction difference state is recorded, the number of direction changes in the continuous period is counted, the period with continuous mutual change is positioned, and an energy alternation section is generated; S302: According to the energy alternating section, the change point of input and output power fluctuation direction in each period is detected, the node with simultaneous frequency and voltage direction reverse change is extracted, the position state of the node in the input and output power switching is judged, and the corresponding relationship between the node number and the time index is established to obtain an energy turning node set; S303: The energy turning node set is called, the node set is sorted and reorganized according to the time sequence of each node and the power change amplitude, the nodes with associated characteristics are integrated into the input end record area according to the time sequence, the repeated or invalid items are screened out, the input parameter sequence is established according to the power direction change information, and the load prediction input set is generated.

5. The method of claim 1, wherein: The acquisition step of the prediction deviation interval sequence is: S401: The energy turning node characteristic parameters and the corresponding time sequence power data in the load prediction input set are called, the time sequence power difference is calculated according to the power change direction of each node in the same area, the node data with consistent fluctuation direction in the same time period is merged, the load power prediction change sequence of different target areas is obtained, the sequence interval is arranged according to the time index, the load prediction change sequence is generated; S402: According to the load prediction change sequence, the real-time load change information of the substation monitoring terminal and the low-voltage bus collection equipment is called, the power deviation between the predicted value and the actual value in the same time interval is calculated, the deviation state is judged according to the power deviation threshold, the time period with consistent deviation direction and amplitude exceeding the threshold in continuous multiple intervals is screened, and the prediction deviation interval set is obtained; S403: According to the prediction deviation interval set, the interval records are rearranged in time sequence, the deviation amplitude and the continuous time length in continuous time are calculated, the time sequence result is formed, and the output is arranged according to the deviation time index number to generate the prediction deviation interval sequence.

Images