A property park power load prediction system and method

Abstract

The application discloses a kind of property park electric power load prediction system and method, belong to electric power prediction technical field, the topological membership relation index table of park total entrance layer-distribution feeder layer-tenant branch layer is constructed in the present application, after obtaining the historical load data and tenant information data of each hierarchical monitoring node, baseline prediction model is constructed for each monitoring node;Real-time acquisition of the load data of each monitoring node in tenant branch layer, identify the mutation node of which load baseline mutation event occurs;Combined with tenant information data, the mutation source node caused by tenant change is screened out;Based on each mutation source node, analyze the mutation influence degree of each distribution feeder loop of distribution feeder layer and park total entrance layer;According to the mutation influence degree, determine the adjustment strategy of the baseline prediction model of corresponding monitoring node, realize the model updating mode of each layer monitoring node is selected according to the mutation influence degree, in guaranteeing prediction accuracy and efficiency.

Description

Technical Field

[0001] This invention relates to the field of power forecasting technology, and in particular to a power load forecasting system and method for residential parks. Background Technology

[0002] Electricity load forecasting is a crucial foundational task for power system planning, operation, and dispatch. Its accuracy directly impacts the safe and stable operation of the power system, the formulation of generation plans, electricity market trading strategies, and the optimization decisions of energy management systems. In residential communities (including commercial complexes, office buildings, and industrial parks), accurate electricity load forecasting not only affects the safe operation of the distribution system and the rational allocation of transformer capacity, but also directly influences the property operator's demand management, the optimized dispatch of distributed energy resources, and the accuracy of online carbon emission monitoring.

[0003] There are various existing power load forecasting methods. Traditional forecasting methods include per capita electricity consumption method, elasticity coefficient method, output value unit consumption method, load density method, linear regression method, and time series analysis method. With the development of artificial intelligence technology, load forecasting methods based on deep learning have been widely used.

[0004] However, existing technologies still have the following technical problems: The power distribution system of a property park usually presents a three-level topology structure of "main entrance of the park - power distribution feeder - tenant branch circuit". The load forecast results of different levels serve completely different business scenarios. In a property park, events such as tenant move-in, lease termination or business type change will cause significant and lasting changes in the power load of the corresponding tenant branch circuit. This mutation will propagate upward along the power distribution topology, affecting the load pattern of the power distribution feeder circuit and even the main entrance of the park. However, existing technologies fail to systematically identify mutation events, trace mutation source nodes and quantify their impact on nodes at different levels. This makes it impossible to perceive the real changes in the power consumption structure of the park in a timely manner. Consequently, there is a lack of differentiated model adjustment strategies for power distribution feeder circuits and to the main entrance of the park based on the degree of mutation impact. That is, it is impossible to adaptively select different model update methods such as complete retraining or parameter fine-tuning according to mutations, which will result in either over-updating or inaccurate prediction accuracy. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a method for predicting the power load of a residential park, comprising the following steps: S1. Construct a topological membership index table of the park's main entrance layer - power distribution feeder layer - tenant branch layer, and obtain historical load data and tenant information data of monitoring nodes at each level in the park; S2. Based on the historical load data, construct a baseline prediction model for each monitoring node; S3. Collect load data of each monitoring node in the tenant branch layer in real time and identify whether it is a node that has experienced a load baseline change event. S4. Combine tenant information data to filter out mutation nodes caused by tenant changes and mark them as mutation source nodes; S5. Based on each mutation source node, analyze the degree of mutation impact of each power distribution feeder circuit in the power distribution feeder layer and the main entrance layer of the park; S6. Determine the adjustment strategy for the baseline prediction model of the monitoring nodes of the corresponding power distribution feeder circuit and the monitoring nodes of the park's main entrance layer based on the degree of impact of the sudden change.

[0006] Furthermore, the construction of the topological membership index table specifically includes the following steps: Based on the power distribution topology of the park, the load monitoring system is divided into three levels: the park's main entrance layer, the power distribution feeder layer, and the tenant branch layer. Establish a first topology affiliation index table between each tenant branch and the power distribution feeder, and a second topology affiliation index table between each power distribution feeder and the main entrance of the park. Combine the first topology affiliation index table and the second topology affiliation index table into a topology affiliation index table. The first topology affiliation index table records the power distribution feeder number to which each tenant branch belongs. The second topology affiliation index table records each power distribution feeder under the main entrance of the park.

[0007] Furthermore, the construction of a baseline prediction model for each monitoring node specifically includes the following steps: For the monitoring nodes at the main entrance of the park, the monitoring nodes of each power distribution feeder circuit, and the monitoring nodes of each tenant branch circuit, their respective historical load time series data are extracted and preprocessed. Periodic features are extracted from the preprocessed historical load time-series data; For each monitoring node, a baseline prediction model based on a long short-term memory network is constructed. The input of the baseline prediction model is the historical load time series data and periodic characteristics of the corresponding monitoring node, and the output is the load prediction value within a preset future time window.

[0008] Furthermore, the load data of each monitoring node in the tenant branch layer is collected in real time, and it is identified whether it is a node that has experienced a load baseline change event. This includes the following steps: With a preset sliding window length and sliding step size, load data of each monitoring node in the tenant branch layer is collected in real time to construct a real-time load window sequence for each monitoring node. For each monitoring node, the sliding window corresponding to the real-time load window sequence is used to calculate its load statistical characteristics, including window mean, window standard deviation, window maximum, window minimum, and window coefficient of variation; the window coefficient of variation is the ratio of the window standard deviation to the window mean. The current load statistics are compared with the historical load statistics of the historical window corresponding to the baseline prediction model of the corresponding monitoring node. The deviation of the statistical characteristics is calculated. If the absolute value of the deviation of the statistical characteristics is greater than the first preset deviation threshold, the corresponding dynamic window is marked as a suspected mutation window. If a preset number of consecutive sliding windows in a monitoring node are marked as suspected mutation windows, then the monitoring node is determined to have experienced a load baseline mutation event, is marked as a mutation node, and the start timestamp, end timestamp, average load level before mutation, and average load level after mutation of the load baseline mutation event are recorded to generate a mutation event record.

[0009] Further, nodes that undergo mutations due to tenant changes are identified, specifically through the following steps: Obtain the contract status change records of the tenant branch corresponding to the mutation node from the property management system; If the start time of a load baseline mutation event of a mutation node is within a preset time window after the change time of any change record in the contract status of the corresponding tenant branch, then the mutation node is marked as a candidate mutation source node. Calculate the load change of the candidate mutation source node before and after the load baseline mutation event. If the change is greater than the preset change threshold, the candidate mutation source node is determined to be caused by the tenant change and is marked as the mutation source node.

[0010] Furthermore, the load change of the candidate mutation source node before and after the load baseline mutation event is the absolute value of the difference between the steady-state mean load of the candidate mutation source node after the load baseline mutation event and the steady-state mean load before the load baseline mutation event. The steady-state mean load before the load baseline abrupt event is the average load data for the same period within a preset number of days before the abrupt start time of the candidate abrupt source node; the steady-state mean load after the load baseline abrupt event is the average load data for the same period within a preset number of days after the abrupt end time of the candidate abrupt source node.

[0011] Furthermore, the analysis of the impact of abrupt changes on each power distribution feeder circuit at the power distribution feeder layer and the overall entrance layer of the park includes the following steps: S51. Based on the topology membership index table, obtain the distribution feeder circuit to which the mutation source node belongs; S52. The load change rate of the mutation source node is the ratio of the absolute value of the load change of the mutation source node to the absolute value of the steady-state mean load of the mutation source node before the occurrence of the load baseline mutation event. S53. The time corresponding to the second preset number of days before the load baseline change event occurs at the mutation source node is recorded as the first time period. The average load P1 of the mutation source node in the first time period, the average load P2 of the distribution feeder monitoring node of the distribution feeder circuit to which the mutation source node belongs, and the average load P3 of the total inlet monitoring node in the first time period are obtained. S54. The ratio of average load P1 to average load P2 is used as the distribution feeder influence coefficient of the sudden change source node, and the ratio of average load P1 to average load P3 is used as the total inlet influence coefficient of the sudden change source node. S55. The first comprehensive impact value of the sudden change source node on its corresponding distribution feeder circuit is obtained by weighting the load change rate of the sudden change source node and the influence coefficient of the distribution feeder through preset weights. The second comprehensive impact value of the sudden change source node on the total entrance layer of the park is obtained by weighting the load change rate of the sudden change source node and the influence coefficient of the total entrance through preset weights. S56. According to steps S51 to S55, obtain the first comprehensive impact value and the second comprehensive impact value of each mutation source node; take the sum of the first comprehensive impact values ​​of all mutation source nodes under the power distribution feeder circuit as its mutation impact degree, and take the sum of the second comprehensive impact values ​​of all mutation source nodes under the park's main entrance layer as its mutation impact degree.

[0012] Furthermore, based on the degree of impact of the sudden change, the adjustment strategy for the baseline prediction model of the monitoring nodes of the corresponding distribution feeder circuits and the monitoring nodes at the main entrance layer of the park is determined, specifically as follows: If the impact of the mutation is greater than or equal to the first preset impact level, then the baseline prediction model of the corresponding monitoring node will be completely retrained. If the impact of the mutation is less than the first preset impact level but greater than or equal to the second preset impact level, then the model parameter update strategy will be executed on the baseline prediction model of the corresponding monitoring node. For mutation source nodes, the same strategy of completely retraining their baseline prediction models is employed.

[0013] Furthermore, for the baseline prediction model that has undergone a complete retraining strategy, the following steps are performed: Starting from the latest mutation end time among the mutation source nodes, load data of the monitoring node is collected within the third preset number of days to complete the construction of new training samples. The corresponding baseline prediction model is then retrained using the new training samples. For baseline prediction models that employ model parameter update strategies, perform the following steps: Starting from the latest mutation end time among the mutation source nodes, load data of the monitoring node is collected within the fourth preset number of days to complete the construction of the incremental training dataset. Using the incremental learning algorithm, the model parameters are fine-tuned online using the incremental training dataset while keeping the network structure of the corresponding baseline prediction model unchanged.

[0014] The present invention also provides a power load forecasting system for residential parks, comprising: The index table acquisition module constructs a topological membership index table of the park's main entrance layer, power distribution feeder layer, and tenant branch layer. The data acquisition module acquires historical load data and tenant information data from monitoring nodes at various levels within the park. The model acquisition module constructs a baseline prediction model for each monitoring node based on the historical load data. The first identification module collects load data of each monitoring node in the tenant branch layer in real time and identifies whether it is a node that has experienced a load baseline change event. The second identification module combines tenant information data to filter out mutation nodes caused by tenant changes and marks them as mutation source nodes. The analysis module analyzes the impact of each power distribution feeder circuit in the power distribution feeder layer and the total entrance layer of the park on each mutation source node. The model update module determines the adjustment strategy for the baseline prediction model of the monitoring nodes of the corresponding power distribution feeder circuit and the monitoring nodes of the main entrance layer of the park, based on the degree of impact of the sudden change.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention collects load data from tenant branch layers in real time and identifies abrupt change nodes where load baseline abrupt changes occur. Combined with tenant information data, it filters out abrupt change source nodes caused by tenant changes. This allows for the automatic and accurate identification and tracing of load abrupt changes in property parks caused by tenant occupancy, lease termination, or changes in business type, overcoming the shortcomings of existing technologies that cannot promptly perceive real changes in the power structure. By quantifying the impact of abrupt change source nodes on the feeder layer and the main inlet layer, it achieves a quantitative assessment of the upward propagation of abrupt change impact along the distribution topology. By sequentially completing the entire process of topology modeling, baseline model construction, abrupt change identification, source node tracing, impact quantification, and model adaptive adjustment, it constructs a complete closed-loop technical path from abrupt change identification to model adaptive updates. This enables the prediction system to continuously track changes in the park's power structure and respond promptly, implementing differentiated model adjustment strategies based on the degree of abrupt change impact. Therefore, in the specific application scenario of property parks, it significantly improves the accuracy and efficiency of load prediction at each layer. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a property park power load prediction method according to the present invention; Figure 2 This is a flowchart illustrating the construction process of the baseline prediction model in the method of this invention; Figure 3 This is a flowchart of the identification process for mutation nodes in the method of this invention; Figure 4 This is a flowchart of the identification process for mutation source nodes in the method of this invention; Figure 5 This is a flowchart illustrating the analysis of the degree of influence of mutations in the method of this invention; Figure 6 This is a structural block diagram of a property park power load prediction system according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0021] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" and "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0022] Example 1 See Figure 1 As shown, the present invention provides a method for predicting the power load of a property park, which specifically includes the following steps: S1. Construct a topological membership index table of the park's main entrance layer - power distribution feeder layer - tenant branch layer, and obtain historical load data and tenant information data of monitoring nodes at each level in the park; S2. Based on the historical load data, construct a baseline prediction model for each monitoring node; S3. Real-time acquisition of load data sequences of each monitoring node in the tenant branch layer, and identification of whether it is a node that has experienced a load baseline change event; S4. Combine tenant information data to filter out mutation nodes caused by tenant changes and mark them as mutation source nodes; S5. Based on each mutation source node, analyze the degree of mutation impact of each power distribution feeder circuit in the power distribution feeder layer and the main entrance layer of the park; S6. Determine the adjustment strategy for the baseline prediction model of the monitoring nodes of the corresponding power distribution feeder circuit and the monitoring nodes of the park's main entrance layer based on the degree of impact of the sudden change.

[0023] This invention collects tenant branch load data in real time and identifies the mutation nodes where load baseline mutation events occur. Then, by combining tenant information data, it filters out mutation source nodes caused by tenant changes. This enables load mutations caused by tenant move-in, move-out, or business changes in the property park to be automatically and accurately identified and traced, thus enabling timely understanding of the real changes in the park's power consumption structure.

[0024] Simultaneously, based on each mutation source node, the impact of mutations on each distribution feeder circuit at the distribution feeder layer and the main entrance layer of the park is analyzed. Since the property park's power distribution system has a three-level topology of "main entrance of the park - distribution feeder - tenant branch circuit", load mutations in tenant branch circuits will propagate upwards along the topology and affect the load patterns of feeder circuits and even the main entrance. This step quantifies the impact level by level with the mutation source node as the starting point, solving the technical problem in the existing technology that failed to quantify the impact of load mutations in tenant branch circuits on nodes at different levels.

[0025] This invention determines the adjustment strategy of the baseline prediction model for the corresponding distribution feeder circuit monitoring node and the park's main entrance monitoring node based on the degree of impact of sudden changes. Since the degree of impact of sudden changes varies among nodes at different levels, a uniform model update strategy either wastes computational resources due to "over-updating" or leads to a decrease in prediction accuracy due to "under-updating". This invention formulates an adjustment strategy based on the degree of impact, so that the model update intensity matches the actual magnitude of load pattern changes. This solves the technical problems in the prior art of "lacking differentiated model adjustment strategies for distribution feeder circuits and the park's main entrance based on the degree of impact of sudden changes" and "being unable to adaptively select different intensity model update methods such as complete retraining and parameter fine-tuning according to sudden changes", thereby achieving a balance between prediction accuracy and computational efficiency.

[0026] In step S1, constructing the topological membership index table of the park's main entrance layer - power distribution feeder layer - tenant branch layer specifically includes the following steps: S11. According to the power distribution topology of the park, the load monitoring system is divided into three levels: the park's main entrance layer, the power distribution feeder layer, and the tenant branch layer. The park's main entrance layer includes one monitoring node, the power distribution feeder layer includes monitoring nodes for multiple power distribution feeder circuits, and the tenant branch layer includes monitoring nodes for multiple tenant branches. For example, based on the primary wiring diagram of the power distribution system of the property park, the power distribution topology of the park is identified, and the load monitoring system is divided into three levels.

[0027] The main entrance layer of the park is the connection point between the park and the external power grid. The total load data of the park is collected by the smart meters in the main distribution room of the park. For example, a smart meter is installed in the mains power incoming cabinet of the main distribution room of the park. The smart meter collects load data such as the total active power, total reactive power, three-phase current, three-phase voltage and power factor of the park through current transformers and voltage transformers.

[0028] The power distribution feeder layer consists of various power distribution feeder circuits in the park. The load data of each feeder is collected by smart meters in each feeder circuit. For example, smart meters are installed on the outgoing feeder circuits of the power distribution rooms on each floor or the power distribution cabinets in each area. Each feeder circuit corresponds to one smart meter, which is used to collect the total load data of the area supplied by the feeder.

[0029] The tenant branch layer consists of independent power branches for each tenant. The load data of each tenant is collected through the smart meters of each tenant branch. For example, a smart meter is installed in the independent meter box of each tenant, with one smart meter corresponding to each tenant branch, which is used to collect the independent power load data of that tenant.

[0030] In some embodiments, the hierarchy can be defined by reading the configuration file of the park's power distribution system. The configuration file records the incoming circuit number of the park's main power distribution room, the feeder circuit number of each floor's power distribution room, and the branch circuit number of each tenant's meter box, as well as the hierarchical relationship between the three, in a structured data format (such as JSON or XML).

[0031] S12. Establish a first topology membership index table between each tenant branch and the power distribution feeder, and a second topology membership index table between each power distribution feeder and the main entrance of the park. Combine the first topology membership index table and the second topology membership index table into a topology membership index table. For example, the first topology membership index table records the power distribution feeder number to which each tenant branch belongs; the second topology membership index table records each power distribution feeder under the main entrance of the park. Through these two index tables, it is possible to trace the higher-level nodes affected by any tenant branch from the smallest to the largest level, or to drill down to the smallest level and find all the lower-level nodes contained in any high-level node.

[0032] In the power load forecasting of the industrial park, the forecast results at different levels serve completely different business scenarios and cannot be substituted for one another: Tenant branch layer: Prediction results are typically used for tenant energy consumption allocation, internal cost accounting, or demand monitoring for individual tenants; Power distribution feeder layer: Corresponding to specific areas of the park (such as a building or a commercial area on a certain floor), the prediction results are used to support the air conditioning / lighting group control strategy and local fault backup plan for that area. The main entrance layer of the park corresponds to the main high-voltage incoming line table. The forecast results are directly used for basic electricity fee (demand fee) declaration, transformer economic operation strategy, and participation in power demand response.

[0033] The acquisition of historical load data and tenant information data of monitoring nodes at each level within the park specifically includes: Load data of smart meters at each monitoring node in each level are collected at a preset sampling period; The load data includes active power, reactive power, current, voltage, and power factor. The collected load data is timestamped and then stored in a time-series database. For example, the smart meters at each monitoring node are polled via fieldbus or industrial Ethernet at a preset sampling period (e.g., once every 15 minutes) to read the load data in the meter registers. Each data point is labeled with a timestamp of the collection time and then stored in a time-series database.

[0034] Tenant information data is obtained from the property management system. The tenant information data includes each tenant's branch number, tenant name, tenant business type, leased area, move-in date, move-out date, and contract status. For example, the prediction server periodically synchronizes tenant information data through the API interface or direct database connection provided by the property management system. Among them, the branch number is used to uniquely identify the electricity branch corresponding to the tenant; the tenant business type is used to characterize the tenant's business category (such as office, catering, retail, fitness, cinema, etc.); the move-in date and move-out date are used to record the start and end time of the tenant's electricity consumption on that branch; and the contract status is used to indicate whether the tenant is currently in a valid lease status.

[0035] See Figure 2 As shown, in step S2, a baseline prediction model is constructed for each monitoring node based on the historical load data, specifically including the following steps: S21. For the monitoring nodes at the main entrance of the park, the monitoring nodes of each power distribution feeder circuit, and the monitoring nodes of each tenant branch, extract their respective historical load time series data and perform preprocessing. For example, the prediction server extracts active power time-series data for a past period (such as the past 365 days) from the time-series database for each monitoring node, forming an independent historical load sequence for each node.

[0036] The preprocessing includes missing value imputation, outlier removal, and normalization.

[0037] The missing value imputation adopts the linear interpolation method. For time periods with consecutive missing values ​​exceeding a preset threshold, the data segments are marked as invalid and removed. For example, when individual sampling points are missing in the historical load sequence, linear interpolation between the two adjacent valid data points before and after the missing point is used to fill in the missing data. If the number of consecutively missing data points exceeds a preset threshold (e.g., more than 6 consecutively missing sampling points), the entire period of consecutive missing data is marked as an invalid data segment and removed to avoid introducing unreliable interpolation results due to a long period without data.

[0038] The outlier removal method is based on the interquartile range, which marks data points that exceed 1.5 times the interquartile range above and below the quartile as outliers and removes them. For example, for each level of historical load sequence, its first quartile Q1 and third quartile Q3 are calculated, and then the interquartile range IQR = Q3 - Q1 is calculated. Data points less than Q1 - 1.5 × IQR or greater than Q3 + 1.5 × IQR are marked as outliers and removed from the sequence. The resulting gaps are filled by linear interpolation of the preceding and following valid values.

[0039] The normalization process uses the minimum-maximum normalization method to map the load data to the [0,1] interval.

[0040] S22. Extract periodic features from the preprocessed historical load time series data, wherein the periodic features include daily periodic features and weekly periodic features; The daily cycle characteristics include the load mean, load standard deviation, and load peak-to-valley difference at each time point; For example, for each day, the average load, standard deviation of load, and peak-to-valley load difference (the difference between the daily maximum load and the daily minimum load) are calculated for each of the 96 sampling points (with a 15-minute sampling interval) to form a daily periodic feature vector.

[0041] The weekly cycle characteristics include the load difference characteristics between weekdays and non-weekdays; For example, based on the date type (Monday to Friday are working days, Saturday, Sunday and public holidays are non-working days), the average load curves for working days and non-working days are statistically analyzed, and the load difference between the two at the same time is calculated to form a weekly cycle feature vector.

[0042] S23. For each monitoring node, a baseline prediction model based on a long short-term memory network is constructed. The input of the baseline prediction model is the historical load time series data and periodic characteristics of the corresponding monitoring node, and the output is the load prediction value within a future preset time window. The Long Short-Term Memory (LSTM) network model comprises an input layer, an LSTM hidden layer, and an output layer. The LSTM model structure includes: an input layer receiving an input tensor with a dimension of (time step × feature dimension); an LSTM hidden layer containing two LSTM units used to capture long-term dependencies in time-series data; and an output layer that is a fully connected layer that maps the output of the LSTM hidden layer to a sequence of predicted load values ​​for a prediction time window.

[0043] The baseline prediction model is trained using corresponding historical load data as training samples, with mean squared error as the loss function and the Adam optimization algorithm used to optimize the model parameters.

[0044] See Figure 3 As shown, in step S3, load data of each monitoring node in the tenant branch layer is collected in real time, and it is identified whether it is a node that has experienced a load baseline change event. Specifically, this includes the following steps: S31. With a preset sliding window length and sliding step size, collect the load data of each monitoring node in the tenant branch layer in real time, and construct the real-time load window sequence of each monitoring node. For example, the prediction server sets the sliding window length to W sampling points (e.g., W=96, corresponding to 24 hours) and the sliding step size to S sampling points (e.g., S=1, corresponding to 15 minutes). Each time a new sampling data point is received, the point is added to the current window and the oldest point in the window is removed, forming a real-time sliding window sequence.

[0045] S32. For each monitoring node, calculate the load statistical characteristics of the sliding window corresponding to the real-time load window sequence, including the window mean, window standard deviation, window maximum value, window minimum value and window coefficient of variation. The window coefficient of variation is the ratio of the window standard deviation to the window mean.

[0046] S33. Compare the current load statistics with the historical load statistics of the historical window corresponding to the baseline prediction model of the corresponding monitoring node, calculate the deviation of the statistical characteristics, and if the absolute value of the deviation of the statistical characteristics is greater than the first preset deviation threshold, mark the corresponding dynamic window as a suspected mutation window. For example, historical load data for the same time period as the current window (e.g., Monday morning 9:00-next day 9:00) is extracted from the historical load sequence, its statistical characteristics are calculated, and then the load statistical characteristics of the current window are compared with the historical load statistical characteristics of the historical window.

[0047] The statistical feature deviation is calculated as follows: the difference between the current window mean and the window mean of the corresponding historical window is divided by the window standard deviation of the corresponding historical window.

[0048] S34. If a preset number of consecutive sliding windows in a monitoring node are all marked as suspected mutation windows, then the monitoring node is determined to have experienced a load baseline mutation event, is marked as a mutation node, and the start timestamp, end timestamp, average load level before mutation, and average load level after mutation of the load baseline mutation event are recorded to generate a mutation event record.

[0049] Mutation initiation time: The start time of the first window marked as a suspected mutation; Mutation end time: The end time of the last window marked as a suspected mutation; Average load water before mutation: The average load during the same period 7 days before the onset of mutation; Average load water after mutation: The average load during the same period 7 days after the end of the mutation.

[0050] See Figure 4 As shown, in step S4, nodes that have undergone mutations due to tenant changes are identified by combining tenant information data and marked as mutation source nodes. This specifically includes the following steps: S41. Obtain the contract status change records of the tenant branch corresponding to the mutation node from the property management system, including tenant move-in records, tenant move-out records, and tenant business type change records. For example, query the property management system for contract status change records within a period of time before and after the sudden event (such as 30 days before and after), including: new tenant move-in (contract takes effect), existing tenant move-out (contract terminates), and existing tenant business type change (business type changes when contract is renewed).

[0051] S42. If the start time of the load baseline mutation event of the mutation node is within a preset time window after the change time of any change record in the contract status of the corresponding tenant branch, then the mutation node is marked as a candidate mutation source node. S43. Calculate the load change of the candidate mutation source node before and after the load baseline mutation event. If it is greater than the preset change threshold, the candidate mutation source node is determined to be caused by the tenant change, and the candidate mutation source node is marked as the mutation source node.

[0052] In step S43, the load change of the candidate mutation source node before and after the load baseline mutation event is the absolute value of the difference between the steady-state mean load of the candidate mutation source node after the load baseline mutation event and the steady-state mean load before the load baseline mutation event.

[0053] The steady-state mean load before the occurrence of a load baseline mutation event is the mean load data for the same period within a preset number of days before the mutation start time of the candidate mutation source node.

[0054] The steady-state mean load after a load baseline mutation event is the mean load data for the same period within a preset number of days after the mutation end time of the candidate mutation source node.

[0055] For example, if the mutation start time of a candidate mutation source node is 6:00 and the mutation end time is 9:00, and the preset number of days is 7 days, then the steady-state load mean before the load baseline mutation event is the average of the load data between 6:00 and 9:00 every day for the 7 days before the mutation start time, and the steady-state load mean after the load baseline mutation event is the average of the load data between 6:00 and 9:00 every day for the 7 days after the mutation end time.

[0056] See Figure 5 As shown, in step S5, the degree of abrupt change impact on each power distribution feeder circuit of the power distribution feeder layer and the total entrance layer of the park is analyzed, specifically including the following steps: S51. Based on the topology membership index table, obtain the distribution feeder circuit to which the mutation source node belongs; S52. The load change rate of the mutation source node is the ratio of the absolute value of the load change of the mutation source node to the absolute value of the steady-state mean load of the mutation source node before the occurrence of the load baseline mutation event. S53. The time corresponding to the second preset number of days before the load baseline change event occurs at the mutation source node is recorded as the first time period. The average load P1 of the mutation source node in the first time period, the average load P2 of the distribution feeder monitoring node of the distribution feeder circuit to which the mutation source node belongs, and the average load P3 of the total inlet monitoring node in the first time period are obtained. S54. The ratio of average load P1 to average load P2 is used as the distribution feeder influence coefficient of the sudden change source node, and the ratio of average load P1 to average load P3 is used as the total inlet influence coefficient of the sudden change source node. S55. The first comprehensive impact value of the sudden change source node on its corresponding distribution feeder circuit is obtained by weighting the load change rate of the sudden change source node and the influence coefficient of the distribution feeder through preset weights. The second comprehensive impact value of the sudden change source node on the total entrance layer of the park is obtained by weighting the load change rate of the sudden change source node and the influence coefficient of the total entrance through preset weights. S56. According to steps S51 to S55, obtain the first comprehensive impact value and the second comprehensive impact value of each mutation source node; take the sum of the first comprehensive impact values ​​of all mutation source nodes under the power distribution feeder circuit as its mutation impact degree, and take the sum of the second comprehensive impact values ​​of all mutation source nodes under the park's main entrance layer as its mutation impact degree.

[0057] In some embodiments, the preset weights in step S55 can be determined in the following way: using a combined weighting method that combines the analytic hierarchy process (AHP) and the entropy weighting method; first, inviting power system dispatching experts to compare the relative importance of load change rate and influence coefficient pairwise, constructing a judgment matrix and calculating subjective weights; at the same time, collecting historical sudden event data of the park, constructing an evaluation matrix, calculating objective weights using the entropy weighting method, and then using a linear weighting method to synthesize the subjective weights and objective weights to obtain the final weight w1 of load change rate and the final weight w2 of influence coefficient.

[0058] In step S6, the adjustment strategy for the baseline prediction model of the monitoring nodes of the corresponding distribution feeder circuit and the monitoring nodes of the park's main entrance layer is determined according to the degree of impact of the sudden change. Specifically: If the impact of the mutation is greater than or equal to the first preset impact level, then the baseline prediction model of the corresponding monitoring node will be completely retrained. If the impact of the mutation is less than the first preset impact level but greater than or equal to the second preset impact level, then the model parameter update strategy will be executed on the baseline prediction model of the corresponding monitoring node. For mutation source nodes, the same strategy of completely retraining their baseline prediction models is employed.

[0059] When the impact of a mutation is greater than or equal to the first preset impact level, it indicates that the load pattern of the monitoring node has undergone a fundamental change, and the original baseline prediction model is no longer applicable, requiring a complete retraining of the model. When the impact of a mutation is less than the first preset impact level but greater than or equal to the second preset impact level, it indicates that the load pattern of the monitoring node has changed to a certain extent, but the magnitude of the change is not enough to completely negate the original model, and the model parameters can be fine-tuned through incremental learning.

[0060] For the baseline prediction model that has undergone a complete retraining strategy, perform the following steps: Starting from the latest mutation end time among the mutation source nodes, load data of the monitoring node is collected within the third preset number of days to complete the construction of new training samples. The corresponding baseline prediction model is then retrained using the new training samples.

[0061] The third preset number of days is set based on the preset minimum number of training samples. The minimum number of training samples is set according to the minimum data requirements for training the LSTM model, for example, 30 days × 96 sampling points / day = 2880 samples.

[0062] For baseline prediction models that employ model parameter update strategies, perform the following steps: Starting from the latest mutation end time among the mutation source nodes, load data of the monitoring node is collected within the fourth preset number of days to complete the construction of the incremental training dataset. An incremental learning algorithm is used to fine-tune the model parameters online using the incremental training dataset while keeping the network structure of the corresponding baseline prediction model unchanged.

[0063] During the process of fine-tuning the model parameters online using the incremental training dataset, the incremental learning algorithm adopts a stochastic gradient descent algorithm with learning rate decay, wherein the learning rate is set to one-tenth of the initial learning rate of the original model training. For example, the initial learning rate of the original model training is 0.001, and the learning rate during incremental fine-tuning is set to 0.0001. The lower learning rate ensures that the model parameters do not change too much during the fine-tuning process, thereby retaining the effective knowledge already learned by the original model while adapting to new data patterns and preventing catastrophic forgetting. The termination condition for the online fine-tuning is: the loss function value on the incremental training dataset no longer decreases for a preset number of consecutive rounds, or reaches the preset maximum number of fine-tuning rounds.

[0064] Example 2 See Figure 6As shown, the present invention also provides a power load forecasting system for residential parks, specifically including: The system includes an index table acquisition module, a data acquisition module, a model acquisition module, a first identification module, a second identification module, an analysis module, a model update module, and a central processing unit that communicates with the above modules.

[0065] The index table acquisition module constructs a topological membership index table of the park's main entrance layer, power distribution feeder layer, and tenant branch layer. The data acquisition module acquires historical load data and tenant information data from monitoring nodes at various levels within the park. The model acquisition module constructs a baseline prediction model for each monitoring node based on the historical load data. The first identification module collects load data of each monitoring node in the tenant branch layer in real time and identifies whether it is a node that has experienced a load baseline change event. The second identification module combines tenant information data to filter out mutation nodes caused by tenant changes and marks them as mutation source nodes. The analysis module analyzes the impact of each power distribution feeder circuit in the power distribution feeder layer and the total entrance layer of the park on each mutation source node. The model update module determines the adjustment strategy for the baseline prediction model of the monitoring nodes of the corresponding power distribution feeder circuit and the monitoring nodes of the main entrance layer of the park, based on the degree of impact of the sudden change.

[0066] For example, during the initial system configuration phase: Maintenance personnel import the primary wiring diagram (CAD or VISIO format) of the park's power distribution system into the system, or manually enter the power distribution hierarchy through the topology configuration interface provided by the system. The index table acquisition module automatically generates a topology hierarchy index table of the park's main entrance layer—power distribution feeder layer—tenant branch layer based on the entered information, and stores it as structured data (such as JSON format).

[0067] The data acquisition module polls all smart meters in the park (including the main inlet meter, feeder meters, and tenant branch meters) via industrial Ethernet or RS-485 bus, reads historical load data, and stores it in a local time-series database after missing value imputation, outlier removal, and normalization preprocessing. Simultaneously, the data acquisition module retrieves tenant information data such as contract status, business type (restaurant / office / retail, etc.), and move-in / move-out dates through the property management system's API interface.

[0068] The model acquisition module extracts historical load data for each monitoring node within the park, and extracts daily and weekly periodic features. Then, a baseline prediction model based on a long short-term memory network is built for each node, trained using mean squared error as the loss function and Adam as the optimizer. After training, the model parameters are persistently stored in the model library.

[0069] Example 3 The present invention also provides an electronic device, including: a processor, a transmitting device, an input device, an output device, and a memory. The processor may be implemented using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory may be implemented using a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM), and is used to store computer program code. The computer program code includes computer instructions. When the processor executes the computer instructions, the electronic device executes a method as described in any of the above possible implementation methods.

[0070] Example 4 The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor of an electronic device, cause the processor to perform a method as described in any of the above possible implementations.

[0071] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0072] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0073] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for predicting the power load of a residential park, characterized in that, Includes the following steps: S1. Construct a topological membership index table of the park's main entrance layer - power distribution feeder layer - tenant branch layer, and obtain historical load data and tenant information data of monitoring nodes at each level in the park; S2. Based on the historical load data, construct a baseline prediction model for each monitoring node; S3. Collect load data of each monitoring node in the tenant branch layer in real time and identify whether it is a node that has experienced a load baseline change event. S4. Combine tenant information data to filter out mutation nodes caused by tenant changes and mark them as mutation source nodes; S5. Based on each mutation source node, analyze the degree of mutation impact of each power distribution feeder circuit in the power distribution feeder layer and the main entrance layer of the park; S6. Determine the adjustment strategy for the baseline prediction model of the monitoring nodes of the corresponding power distribution feeder circuit and the monitoring nodes of the park's main entrance layer based on the degree of impact of the sudden change.

2. The method for predicting power load in residential parks according to claim 1, characterized in that, The construction of the topological membership index table specifically includes the following steps: Based on the power distribution topology of the park, the load monitoring system is divided into three levels: the park's main entrance layer, the power distribution feeder layer, and the tenant branch layer. Establish a first topology affiliation index table between each tenant branch and the power distribution feeder, and a second topology affiliation index table between each power distribution feeder and the main entrance of the park. Combine the first topology affiliation index table and the second topology affiliation index table into a topology affiliation index table. The first topology affiliation index table records the power distribution feeder number to which each tenant branch belongs. The second topology affiliation index table records each power distribution feeder under the main entrance of the park.

3. The method for predicting power load in residential parks according to claim 1, characterized in that, The process of constructing a baseline prediction model for each monitoring node specifically includes the following steps: For the monitoring nodes at the main entrance of the park, the monitoring nodes of each power distribution feeder circuit, and the monitoring nodes of each tenant branch circuit, their respective historical load time series data are extracted and preprocessed. Periodic features are extracted from the preprocessed historical load time-series data; For each monitoring node, a baseline prediction model based on a long short-term memory network is constructed. The input of the baseline prediction model is the historical load time series data and periodic characteristics of the corresponding monitoring node, and the output is the load prediction value within a preset future time window.

4. The method for predicting power load in residential parks according to claim 1, characterized in that, The system collects load data from each monitoring node in the tenant branch layer in real time and identifies whether it is a node that has experienced a load baseline abrupt change event. This includes the following steps: With a preset sliding window length and sliding step size, load data of each monitoring node in the tenant branch layer is collected in real time to construct a real-time load window sequence for each monitoring node. For each monitoring node, the sliding window corresponding to the real-time load window sequence is used to calculate its load statistical characteristics, including window mean, window standard deviation, window maximum, window minimum, and window coefficient of variation; the window coefficient of variation is the ratio of the window standard deviation to the window mean. The current load statistics are compared with the historical load statistics of the historical window corresponding to the baseline prediction model of the corresponding monitoring node. The deviation of the statistical characteristics is calculated. If the absolute value of the deviation of the statistical characteristics is greater than the first preset deviation threshold, the corresponding dynamic window is marked as a suspected mutation window. If a preset number of consecutive sliding windows in a monitoring node are marked as suspected mutation windows, then the monitoring node is determined to have experienced a load baseline mutation event, is marked as a mutation node, and the start timestamp, end timestamp, average load level before mutation, and average load level after mutation of the load baseline mutation event are recorded to generate a mutation event record.

5. The method for predicting power load in residential parks according to claim 4, characterized in that, The process of identifying mutation nodes caused by tenant changes includes the following steps: Obtain the contract status change records of the tenant branch corresponding to the mutation node from the property management system; If the start time of a load baseline mutation event of a mutation node is within a preset time window after the change time of any change record in the contract status of the corresponding tenant branch, then the mutation node is marked as a candidate mutation source node. Calculate the load change of the candidate mutation source node before and after the load baseline mutation event. If the change is greater than the preset change threshold, the candidate mutation source node is determined to be caused by the tenant change and is marked as the mutation source node.

6. The method for predicting power load in a property park according to claim 5, characterized in that, The load change of the candidate mutation source node before and after the load baseline mutation event is the absolute value of the difference between the steady-state mean load of the candidate mutation source node after the load baseline mutation event and the steady-state mean load before the load baseline mutation event. The steady-state mean load before the load baseline abrupt event is the average load data for the same period within a preset number of days before the abrupt start time of the candidate abrupt source node; the steady-state mean load after the load baseline abrupt event is the average load data for the same period within a preset number of days after the abrupt end time of the candidate abrupt source node.

7. The method for predicting power load in residential parks according to claim 5, characterized in that, The analysis of the impact of sudden changes in each power distribution feeder circuit at the power distribution feeder layer and the main entrance layer of the park includes the following steps: S51. Based on the topology membership index table, obtain the distribution feeder circuit to which the mutation source node belongs; S52. The load change rate of the mutation source node is the ratio of the absolute value of the load change of the mutation source node to the absolute value of the steady-state mean load of the mutation source node before the occurrence of the load baseline mutation event. S53. The time corresponding to the second preset number of days before the load baseline change event occurs at the mutation source node is recorded as the first time period. The average load P1 of the mutation source node in the first time period, the average load P2 of the distribution feeder monitoring node of the distribution feeder circuit to which the mutation source node belongs, and the average load P3 of the total inlet monitoring node in the first time period are obtained. S54. The ratio of average load P1 to average load P2 is used as the distribution feeder influence coefficient of the sudden change source node, and the ratio of average load P1 to average load P3 is used as the total inlet influence coefficient of the sudden change source node. S55. The first comprehensive impact value of the sudden change source node on its corresponding distribution feeder circuit is obtained by weighting the load change rate of the sudden change source node and the influence coefficient of the distribution feeder through preset weights. The second comprehensive impact value of the sudden change source node on the total entrance layer of the park is obtained by weighting the load change rate of the sudden change source node and the influence coefficient of the total entrance through preset weights. S56. According to steps S51 to S55, obtain the first comprehensive impact value and the second comprehensive impact value of each mutation source node; take the sum of the first comprehensive impact values ​​of all mutation source nodes under the power distribution feeder circuit as its mutation impact degree, and take the sum of the second comprehensive impact values ​​of all mutation source nodes under the park's main entrance layer as its mutation impact degree.

8. The method for predicting power load in a property park according to claim 1, characterized in that, The adjustment strategy for the baseline prediction model of the monitoring nodes of the corresponding distribution feeder circuits and the monitoring nodes at the main entrance layer of the park is determined based on the degree of impact of the sudden change. Specifically: If the impact of the mutation is greater than or equal to the first preset impact level, then the baseline prediction model of the corresponding monitoring node will be completely retrained. If the impact of the mutation is less than the first preset impact level but greater than or equal to the second preset impact level, then the model parameter update strategy will be executed on the baseline prediction model of the corresponding monitoring node. For mutation source nodes, the same strategy of completely retraining their baseline prediction models is employed.

9. The method for predicting power load in a property park according to claim 8, characterized in that, For the baseline prediction model that has undergone a complete retraining strategy, perform the following steps: Starting from the latest mutation end time among the mutation source nodes, load data of the monitoring node is collected within the third preset number of days to complete the construction of new training samples. The corresponding baseline prediction model is then retrained using the new training samples. For baseline prediction models that employ model parameter update strategies, perform the following steps: Starting from the latest mutation end time among the mutation source nodes, load data of the monitoring node is collected within the fourth preset number of days to complete the construction of the incremental training dataset. Using the incremental learning algorithm, the model parameters are fine-tuned online using the incremental training dataset while keeping the network structure of the corresponding baseline prediction model unchanged.

10. A property park power load forecasting system, employing the property park power load forecasting method as described in any one of claims 1 to 9, characterized in that, include: The index table acquisition module constructs a topological membership index table of the park's main entrance layer, power distribution feeder layer, and tenant branch layer. The data acquisition module acquires historical load data and tenant information data from monitoring nodes at various levels within the park. The model acquisition module constructs a baseline prediction model for each monitoring node based on the historical load data. The first identification module collects load data of each monitoring node in the tenant branch layer in real time and identifies whether it is a node that has experienced a load baseline change event. The second identification module combines tenant information data to filter out mutation nodes caused by tenant changes and marks them as mutation source nodes. The analysis module analyzes the impact of each power distribution feeder circuit in the power distribution feeder layer and the total entrance layer of the park on each mutation source node. The model update module determines the adjustment strategy for the baseline prediction model of the monitoring nodes of the corresponding power distribution feeder circuit and the monitoring nodes of the main entrance layer of the park, based on the degree of impact of the sudden change.

Images