Intelligent Power Load Forecasting Methods and Systems for Industrial and Commercial Users

Abstract

This invention relates to the field of power load forecasting, providing a method and system for intelligent power load forecasting for industrial and commercial users. The method includes: generating aligned time-based load records and weather event fragment records based on raw electricity consumption and meteorological data; determining weather scenario conditions and generating a weather scenario sample set based on these records; determining abnormal weather scenario conditions based on the weather scenario sample set; generating sample gap records and an abnormal weather supplementary sample set; and generating a scenario calibration sample library by combining the weather scenario conditions; training a long short-term memory network based on the scenario calibration sample library to generate a basic forecast result table and a weather deviation sample table; generating a scenario calibration network package based on both; generating calibration trigger records based on the scenario calibration network package; determining the calibrated load forecast sequence; and combining both to generate the final load forecast result. This invention integrates multi-dimensional meteorological early warning perception with scenario-based load patterns to construct an intelligent forecasting and deviation calibration mechanism for industrial and commercial power loads under abnormal weather conditions.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting, and in particular to a method and system for intelligent power load forecasting for industrial and commercial users. Background Technology

[0002] With the increasing application of electricity load forecasting in energy management by industrial and commercial users, load identification and forecast calibration under abnormal weather conditions are becoming increasingly important. Typhoons, heavy rains, and high temperatures can alter the electricity load of air conditioning, drainage, and emergency equipment. Existing conventional forecasting models tend to follow normal daily patterns, leading to increased forecast bias under abnormal weather scenarios. How to combine meteorological warning levels to conduct scenario-based forecasting of abnormal weather loads has become an urgent technical problem to be solved.

[0003] Chinese patent application CN113177355A discloses a method for predicting electricity load. The method includes: obtaining original input features of a model; the original input features include electricity load data and meteorological factor data; inputting the original input features into a multilayer RBM network for training and learning, and obtaining second input features through multiple nonlinear transformations of the RBM network and parameter reconstruction and fine-tuning; obtaining third input features based on the second input features and a genetic algorithm; obtaining second weights based on the initial weights of a BP neural network and the genetic algorithm; inputting the third input features and the second weights into the BP neural network, and using the BP algorithm to perform inverse parameter fine-tuning according to a set error threshold until the error is less than or equal to the preset threshold, thus obtaining a model for predicting electricity load; and inputting the time to be predicted into the model for predicting electricity load for prediction.

[0004] However, current technologies still face many challenges. When industrial and commercial users are under a rainstorm warning, the electricity load for purposes such as drainage and emergency response may change with the intensity of rainfall and the warning status. Ordinary power load forecasting models rely heavily on historical electricity consumption patterns under normal weather conditions, making it difficult to incorporate the impact of weather warning levels and weather changes on electricity consumption into forecasting. When heavy rain causes the load of relevant equipment to deviate significantly from normal consumption patterns, the forecast results may still follow the conventional trend, leading to a discrepancy between the load estimate for the forecast period and the actual electricity demand. If this discrepancy is not identified and corrected during the forecasting phase, subsequent energy management or load scheduling will lack reliable references for abnormal weather conditions, easily resulting in response delays. Summary of the Invention

[0005] To achieve the above objectives, this invention provides an intelligent power load forecasting method for industrial and commercial users, the specific technical solution of which is as follows: Based on the acquired raw electricity consumption and meteorological data, aligned time load records and weather event fragment records are generated. Based on the aligned time load records and weather event fragment records, weather scenario conditions are determined, and a weather scenario sample set is generated. Based on the weather scenario sample set, abnormal weather scenario conditions are determined and sample gap records are generated. Based on the sample gap records, an abnormal weather supplementary sample set is generated, and a scenario calibration sample library is generated by combining the weather scenario conditions. A long short-term memory network is trained based on the scenario calibration sample library to generate a basic prediction result table. A weather deviation sample table is generated based on the basic prediction result table. A scenario calibration network package is generated based on the basic prediction result table and the weather deviation sample table. Based on the scenario, calibration network packets are calibrated to generate calibration trigger records. The calibration trigger records are used to determine the post-calibration load prediction sequence. Finally, the load prediction result is generated based on the post-calibration load prediction sequence and the calibration trigger records.

[0006] The present invention also provides an intelligent power load forecasting system for industrial and commercial users, which is used to implement the above-mentioned intelligent power load forecasting method for industrial and commercial users. The system includes a scenario construction module, a scenario calibration module, a deviation calibration module, and a load forecasting module. The scenario construction module is used to generate aligned time load records and weather event fragment records based on the acquired raw electricity consumption and meteorological data, determine weather scenario conditions based on the aligned time load records and weather event fragment records, and generate a weather scenario sample set. The scene calibration module is used to determine abnormal weather scene conditions based on the weather scene sample set and generate sample gap records, generate an abnormal weather supplementary sample set based on the sample gap records, and generate a scene calibration sample library in combination with the weather scene conditions. The deviation calibration module is used to train a long short-term memory network based on a scenario calibration sample library, generate a basic prediction result table, generate a weather deviation sample table based on the basic prediction result table, and generate a scenario calibration network package based on the basic prediction result table and the weather deviation sample table. The load prediction module is used to generate calibration trigger records based on scenario calibration network packets, determine the post-calibration load prediction sequence based on the calibration trigger records, and generate the final load prediction result based on the post-calibration load prediction sequence and the calibration trigger records.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention generates aligned time load records and weather event fragment records based on raw electricity consumption and meteorological data, and determines weather scene conditions to generate a weather scene sample set. This enables the aligned time load records to correspond to the meteorological warning process according to the scene, reducing the sample classification bias caused by using the load pattern of ordinary time periods when abnormal weather occurs. This invention determines abnormal weather scenario conditions based on a weather scenario sample set and generates sample gap records. Then, it generates an abnormal weather supplementary sample set based on the sample gap records and forms a scenario calibration sample library. This improves the problem that historical load samples are insufficient and it is difficult to form corresponding calibration basis under early warning scenarios such as typhoons, rainstorms, and high temperatures. This invention generates a basic prediction result table by training a long short-term memory network using a scene calibration sample library, and generates a weather deviation sample table and a scene calibration network package based on the basic prediction result table, so that the basic prediction results can establish a correspondence with the abnormal weather deviation samples, and avoids the abnormal weather load offset being directly mixed into the normal load prediction pattern. This invention generates calibration trigger records based on scenario calibration network packets, determines the post-calibration load prediction sequence based on the calibration trigger records, and generates the final load prediction result. This ensures that the corresponding calibration deviation is only superimposed when matching weather scenario conditions, thus avoiding improper calibration of load prediction results in the absence of early warning or matching scenarios. Attached Figure Description

[0008] 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, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating the principle of the intelligent power load forecasting method for industrial and commercial users of the present invention. Figure 2 This is a schematic diagram illustrating the principle of weather event segmentation processing in this invention; Figure 3 This is a schematic diagram of the process for generating and writing the basic predicted load value in this invention; Figure 4 This is a functional block diagram of the intelligent power load forecasting system for industrial and commercial users according to the present invention. Detailed Implementation

[0010] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] Example 1: Please see Figure 1As shown, this embodiment provides a method for intelligent power load forecasting for industrial and commercial users, including: Step S1000: Based on the acquired raw electricity consumption and meteorological data... Generate aligned time load records Weather event snippets Based on the alignment time load record Weather event snippets Determine weather scene conditions Generate a weather scene sample set .

[0012] Specifically, this step aims to process the acquired raw electricity consumption and meteorological data. As input, aligned time load records are generated sequentially. Weather event snippets and record the load according to the alignment time. and the recorded weather event segments Determine weather scene conditions Generate a weather scene sample set This processing method is used to ensure consistent temporal representation between electrical load records and meteorological early warning data during abnormal weather events such as typhoons, heavy rains, and high temperatures. The output is a sample set of weather scenarios. To follow up according to weather conditions It provides a data foundation for sample retrieval and model input construction.

[0013] Further, step S1000 includes: Step S1100: Process the raw electricity consumption and meteorological data. By aligning time points according to 15-minute time intervals, missing sub-load values ​​at individual time points are filled in by using the average of neighboring points, generating aligned time load records. .

[0014] In the specific implementation process, this step first obtains raw electricity consumption and meteorological data through the data acquisition link on the industrial and commercial user side. The original electricity consumption and meteorological data This includes: total load data collected by smart meters at the main distribution cabinet; sub-load data collected by sub-meters from the air conditioning system, drainage pump group, and emergency lighting branch circuits; production shift data from the production planning system; and weather warning data, weather warning level, and weather type from the weather warning interface.

[0015] Classifying the total load data, sub-item load data and production shift data into adjacent fifteen-minute time periods according to collection time, and taking the starting moment of the fifteen-minute time period as a unified time point; then, classifying the release time, upgrade time, downgrade time and release time in the meteorological early warning data into corresponding time points according to the same fifteen-minute time periods, so that the total load data, sub-item load data, production shift data and meteorological early warning data are on the same time axis; then, performing adjacent point mean filling for missing sub-item load values at individual time points to obtain time points corresponding sub-item equipment sub-item load value , and writing the sub-item load value into sub-item load records. Wherein, the sub-item equipment is an index variable for sub-item load categories, which sequentially represents air conditioning equipment, drainage equipment and emergency equipment; when , it represents air conditioning equipment; when , it represents drainage equipment; when , it represents emergency equipment.

[0016] The specific execution logic of adjacent point mean filling is as follows: when the sub-item load value of sub-item equipment at time point at time point is missing, and both the sub-item load value corresponding to the previous fifteen-minute time point and the sub-item load value corresponding to the next fifteen-minute time point exist, the average of the two is calculated to obtain the sub-item load value corresponding to time point corresponding sub-item load value , that is: . If the sub-item load value corresponding to either of the adjacent fifteen-minute time points before and after time point is missing, no filling is performed on the sub-item load value corresponding to time point , and the sub-item load record corresponding to time point is marked as a to-be-filled record. Finally, a data source tag is retained for the sub-item load record , wherein, indicates that the sub-item load value is a directly collected value, indicates that the sub-item load value is a value filled by adjacent point mean, so that subsequent steps can identify the record source when dividing weather event fragments.

[0017] After the above processing is completed, time-aligned load records are formed . The time-aligned load records are constructed with fifteen minutes as a time interval, and the total load value , sub-item load value of air conditioning equipment Drainage equipment sub-load values Emergency equipment sub-load values Production shifts Weather warning level and weather type Aligned to the same point in time. This indicates the time points divided into 15-minute time intervals, specifically the start time of the corresponding 15-minute time segment; for example, the time points corresponding to the time segment from 10:00 to 10:15. The time points corresponding to the period from 10:00 to 10:15 and then to 10:30. It is 10:15.

[0018] Finally, the output alignment time load record It consists of a single load record corresponding to multiple time points. An ordered set arranged in chronological order. For any point in time... The alignment time load record Single load record It can be represented as: .in, This indicates that the data is collected by the smart meter at the main distribution cabinet and aligned to the specified time point. Total load value; Indicates a point in time The load values ​​of the air conditioning equipment; Indicates a point in time The sub-load values ​​of the drainage equipment; Indicates a point in time Emergency equipment load values; Indicates a point in time Production shifts; Indicates a point in time The weather warning level; Indicates a point in time The weather types are typhoon, rainstorm, high temperature and no abnormal weather; Indicates the data source.

[0019] Specifically, this step aligns the total load data, sub-load data, production shift data, and weather warning data of industrial and commercial users to the same 15-minute time axis, and performs neighbor mean supplementation and source marking for short-term missing sub-load records to reduce record mismatch caused by inconsistent collection times from different data sources, while providing a time-consistent basic record for subsequent sample division by weather events.

[0020] For example, a commercial or industrial plant receives an orange rainstorm warning at 10:00 AM, and the load on its drainage equipment begins to increase at 10:15 AM. Through this step, the total load value, drainage equipment load, production shift, and orange rainstorm warning at 10:15 AM will be written into the same single load record. In this way, subsequent steps can determine the time correspondence between the load change and the rainstorm warning.

[0021] Step S1200, based on the alignment time load record Medium-level meteorological warning and weather type Determine the weather warning period and base it on the weather warning level. Changes are processed to segment the weather warning period, generating weather event fragment records. .

[0022] In the specific implementation process, this step first involves obtaining the alignment time load record from step S1100. Reading time points Weather warning level and weather type The aforementioned weather warning level The value can be 0, 1, 2, or 3. When the value is 0, it represents a time point. Not under weather warning status; when the value is 1, it indicates the time point. Under yellow alert; when the value is 2, it indicates a time point. Under orange alert; when the value is 3, it indicates a specific time point. The weather warning level is red. The value is 0, and the weather type is... When there is no abnormal weather, the record corresponding to that time point is recorded as a normal time period and is not formed into a weather event segment.

[0023] Next, based on the issuance time, escalation time, downgrade time, and cancellation time in the meteorological warning data, the meteorological warning periods within the same day are segmented. Here, a meteorological warning period refers to a time point within the same natural day where at least one time point exists. Meets meteorological warning level Continuous time interval; when weather type For typhoons, heavy rain, or high temperatures, and the corresponding weather warning level for the time period. In this case, the continuous time interval is used as the object to be segmented into weather event segments.

[0024] The specific execution logic of the segmented processing is as follows: the release time of each meteorological warning data is used as the starting boundary of the warning process, and the cancellation time is used as the ending boundary of the warning process; when the meteorological warning level changes within the same day, the upgrade or downgrade time in the meteorological warning data is used as the new segmentation boundary, and the time interval corresponding to the previous meteorological warning level and the time interval corresponding to the next meteorological warning level are processed as different segments. Here, "the meteorological warning level changes within the same day" means that within the same natural day, the meteorological warning level... A change from a lower warning level to a higher warning level, or vice versa; for example, a weather warning level. A change from 1 to 2 indicates an escalation of the warning, while a change from 3 to 2 indicates a downgrade. In this scenario, each level change serves as a segmentation point, resulting in adjacent time intervals that are sequential and non-overlapping, with each time interval containing a specific weather warning level. It remains unchanged.

[0025] Then, weather event segments are constructed according to the starting boundary, ending boundary, and segment boundaries. Specifically, the four 15-minute time points before the starting boundary are designated as the pre-warning period, the time points between the starting boundary and the ending boundary are designated as the warning duration period, and the four 15-minute time points after the ending boundary are designated as the recovery period after the warning is lifted. If the weather warning level changes during the warning duration period, the warning duration period is further divided into multiple adjacent segments at the corresponding segment boundaries, and the corresponding warning level and event stage are recorded respectively. Each of the four 15-minute time points corresponds to a 1-hour observation window.

[0026] Finally, the output is a record of weather event fragments. The set used to represent all weather event segments consists of multiple weather event segment numbers. Corresponding single weather record Composition. Among them, the single weather record... Used to represent a set The first in A snippet of a weather event; Number the weather event segments.

[0027] For any weather event segment number The single weather record It can be represented as: .in, Indicates the first The weather type corresponding to each weather event segment, and its value is related to... Consistent, but This represents a fragment-level attribute. This represents a point-in-time attribute; for example, When it is one of the following: typhoon, rainstorm, high temperature, or no abnormal weather, The weather type used uniformly within this segment should be consistent. Indicates the first The meteorological warning level corresponding to each weather event segment, and its value is related to... Consistent, but This represents a fragment-level attribute. This represents a point-in-time attribute; for example, This indicates that the segment is an orange alert segment. Indicates the first The event phase corresponding to each weather event segment is one of the following: the period before the warning, the period during which the warning is in effect, and the recovery period after the warning is lifted. Indicates the first The start and end times covered by a weather event segment are determined by the earliest and latest 15-minute times within that segment; if the period before the warning or the recovery period after the warning is lifted has fewer than four 15-minute times due to the beginning or end of the day or missing data, then the start and end times are determined accordingly. Only the actual time range is recorded, without adding other time points outside the weather event segment. Indicates the relationship with the first The set of segment load records corresponding to each weather event segment, i.e., from the aligned time load records Selected from the list, belonging to the first category Start and end times of a weather event segment All individual load records within ;Right now .

[0028] Specifically, this step divides the same weather warning period into three stages: before the warning, during the warning, and after the warning is lifted. When the warning level is upgraded or downgraded on the same day, it is further divided into new adjacent segments at the level change point, ensuring that the weather type and weather warning level remain consistent within each segment. Through this processing, subsequent sample construction obtains load records before the warning occurs, during the warning, and after the warning is lifted, and distinguishes the load changes corresponding to different warning levels.

[0029] For example, an orange typhoon warning is issued at 8:00 and lifted at 18:00. This step defines the period from 7:00 to 8:00 as the pre-warning period, 8:00 to 18:00 as the warning duration period, and 18:00 to 19:00 as the recovery period after the warning is lifted. If the warning is upgraded to a red warning at 12:00, then the periods from 8:00 to 12:00 and from 12:00 to 18:00 will be recorded as weather event segments for different meteorological warning levels. The weather warning level in the previous segment... The weather warning level in the next segment Both belong to the warning period and are saved as two separate weather event segments.

[0030] Further, please refer to Figure 2 As shown, Figure 2 This is a schematic diagram illustrating the principle of segmented processing of weather event fragments in this invention. For example... Figure 2 As shown, the bold horizontal axis represents the timeline arranged in 15-minute increments within the same calendar day. The dashed vertical lines, from left to right, represent the start of the pre-warning period, the initial boundary of the warning duration, the segmental boundaries of the warning duration, the termination boundary of the warning duration, and the end of the recovery period after the warning is lifted. Adjacent solid vertical lines represent the discrete scale of the four 15-minute increments on the timeline, used to characterize the alignment of time load records. The time points of continuous sampling. The orange horizontal line indicates the weather warning level. The warning duration is 2, and the red horizontal line indicates the weather warning level. The warning duration is 3, and the intersection of the orange and red horizontal lines corresponds to the weather warning level. The time points at which changes occur are used as the boundaries for segmenting the warning duration. Referring to the exemplary diagram, 8:00 is used as the starting boundary of the warning duration, and 18:00 is used as the ending boundary. 7:00 to 8:00 is divided into the pre-warning period, 8:00 to 18:00 into the warning duration, and 18:00 to 19:00 into the recovery period after the warning is lifted. Specifically, 7:00 to 8:00 corresponds to the four 15-minute intervals before the starting boundary, and 18:00 to 19:00 corresponds to the four 15-minute intervals after the ending boundary. When the meteorological warning level is at 12:00... When the warning level changes from 2 to 3, 12:00 will be used as the dividing line to further divide the warning period into meteorological warning levels. Segment number 2 and weather warning level The segment is 3.

[0031] Step S1300: Record weather event fragments and alignment time load record Determine weather scene conditions and in weather scene conditions The total load sequence and the time series of changes in individual loads are constructed to generate a weather scenario sample set. .

[0032] In the specific implementation process, firstly, based on the weather event fragments recorded from step S1200... The fragment load record set Read the alignment time load record in step S1100 , and fragment load record set Corresponding single load record .

[0033] Next, read a single load record. Each time point production shifts Single load records corresponding to the same weather type, the same meteorological warning level, the same event stage, and the same production shift will be combined. Classified into the same weather scene conditions Among them, the weather scene conditions It consists of weather type, meteorological warning level, event stage, and production shift.

[0034] Then, construct the conditions for each weather scenario. The total load sequence, and the time sequence of load changes for air conditioning, drainage, and emergency equipment. The total load sequence is based on the same weather scenario conditions. Total load values ​​arranged in chronological order Composition; the timing of load changes for the three types of equipment—air conditioning, drainage, and emergency systems—is determined by the same weather scenario conditions. The following are the load changes of each equipment item arranged in chronological order. composition.

[0035] For any sub-item equipment At a certain point in time The change in load of individual equipment items relative to the previous time point, i.e., the next 15 minutes. The load change value of the sub-item equipment is calculated based on the difference between adjacent 15-minute time points. Numerically equal to the same weather scenario conditions Next, time point Corresponding sub-item equipment Sub-item load values The sub-item equipment corresponding to the adjacent 15-minute time point. Sub-item load values The difference. Wherein, the load change value of the sub-item equipment. Used to characterize weather scene conditions Under constraints, corresponding sub-items of equipment Changes in response during abnormal weather events. For the same weather scenario conditions. The first time point arranged in chronological order, representing the load change value of this sub-item equipment. It is denoted as 0 to ensure that the sequence length is consistent.

[0036] Finally, the weather scene conditions The total load sequence, the time series of changes in individual load components, and the sample type labels are written into the weather scenario sample set. The weather scene sample set It consists of multiple weather scene samples The sample set consists of samples from each weather scenario. This includes weather type, meteorological warning level, event stage, production shift, total load sequence, time series of changes in sub-items of load, and sample type labeling.

[0037] Wherein, the sample type label Used to identify the first Weather scene samples In the weather scene sample set The formation method and subsequent use of the data, whose values ​​include 0 and 1. When the sample type is labeled... A value of 0 indicates that the weather scene sample... Alignment time load record output from step S1100 And the weather event fragment records output by step S1200 Collaborative construction, namely the construction of historically collected samples, serves as real-world samples for subsequent sample gap identification and bias extraction; when sample type is labeled... A value of 1 indicates that the weather scene sample... It is constructed from the production plan data and meteorological warning data corresponding to the target forecast period, i.e., the record of the scenario to be predicted. It is used as the record of the scenario to be predicted to participate in subsequent scenario matching, but is not used as a historical measured load sample to calculate the prediction deviation.

[0038] The target forecast period refers to the future time interval from which load forecast results need to be output. Its length is determined by the load forecasting task of industrial and commercial users, for example, from 0:00 to 24:00 the next day. The production plan data corresponding to the target forecast period comes from the production plan system accessed in step S1100, including the production shifts corresponding to each 15-minute time point within the target forecast period. The meteorological warning data corresponding to the target forecast period comes from the meteorological warning interface accessed in step S1100, including the weather type, meteorological warning level, issuance time, escalation time, downgrade time, and cancellation time within the target forecast period.

[0039] This sample type is marked This is a sample-level label used to identify the entire weather scene sample. Is it a historically collected sample or a record of the scenario to be predicted? And the data source labeling in step S1100... Used to describe individual sub-item load values The data sources are different, and the two do not have the same meaning.

[0040] For any scene sample The first Weather scene samples The structure can be represented as: .in, This indicates the sample number for the weather scene. Indicates the first The weather scene conditions corresponding to each weather scene sample; Indicates the first The total load sequence corresponding to each weather scenario sample; Indicates the first The time series of load changes corresponding to each weather scenario sample includes the time series of air conditioning load changes, the time series of drainage equipment load changes, and the time series of emergency equipment load changes. Indicates the first The sample type label corresponding to each weather scene sample.

[0041] Specifically, this step is based on weather scenario conditions. For classification purposes, the alignment time load record output from step S1100 will be used. Weather event fragments output by step S1200 The constructed historical data samples, along with the predicted scenario records constructed from production plan data and meteorological warning data for the target prediction period, are written into a unified data structure. This data structure stores the total load sequence and the time series of load changes for air conditioning, drainage, and emergency equipment in the historical data samples, and stores the weather scenario conditions in the predicted scenario records. and sample type tag This enables subsequent steps to perform sample retrieval, sample gap identification, and scene calibration matching based on weather type, meteorological warning level, event stage, and production shift.

[0042] For example, in the case of weather scene conditions Heavy rain, orange alert, alert duration, day shift production For example, if the weather conditions are... Corresponding sample type tag If the value is 0, then the weather scene sample set Save the total load time series, air conditioning load change time series, drainage equipment load change time series, and emergency equipment load change time series generated from historically collected samples under this condition; if this weather scenario condition Corresponding sample type tag If the value is 1, then the weather scene sample set Save the weather scene conditions The following is a record of the scenarios to be predicted, formed from the target time period.

[0043] Step S2000, based on the weather scene sample set Determine abnormal weather scenarios and conditions And generate sample gap records Based on the sample gap record Generate supplementary sample set for abnormal weather Combined with weather conditions Generate a scene calibration sample library .

[0044] Specifically, this step aims to process the weather scene sample set from S1300. As input, determine the conditions of the abnormal weather scenario. Corresponding sample gap record Then, based on the sample gap record... Generate supplementary sample set for abnormal weather and in combination with weather conditions Generate a scene calibration sample library This processing method is used to manage abnormal weather supplementary samples, real collected samples, and predicted scene records under the same scene conditions, providing a data foundation for subsequent prediction deviation analysis under the same scene conditions.

[0045] Further, step S2000 includes: Step S2100, from the weather scene sample set Filtering abnormal weather scenarios Based on abnormal weather scenarios Generate sample gap records .

[0046] In the specific implementation process, this step first uses the weather scene sample set from step S1300. Reading weather scene conditions Next, from the aforementioned weather scenario conditions... The selection criteria include abnormal weather scenarios where the weather type is typhoon, heavy rain, or high temperature, and the meteorological warning level is not 0. Statistical analysis was conducted. Among these, the abnormal weather scenario conditions... Weather scene conditions A subset thereof, which is derived from weather scene conditions The selected scenarios meet the criteria of abnormal weather type and are under meteorological warning status.

[0047] For each abnormal weather scenario condition Statistical weather scene sample set Sample type labeling Historical sample count And statistical analysis of the conditions of this abnormal weather scenario. Below, the longest consecutive number of time points recorded at 15-minute intervals within the same historical data sample. Among them, the historical sample count Used to determine the conditions of corresponding abnormal weather scenarios. Whether the actual historical data collected below reaches the sample base required for subsequent prediction bias extraction; the longest continuous time point number Used to determine the conditions of corresponding abnormal weather scenarios. Does a load change process that meets the requirement of continuity exist?

[0048] Subsequently, the conditions of the abnormal weather scenario were examined. Check if the timing sequence of load changes for each sub-item of equipment exists. If the timing sequence of any sub-item load change is missing, add the sub-item load missing type corresponding to the missing sub-item load timing sequence to the sub-item load missing type set. The set of missing sub-item load types. Used to record abnormal weather conditions The missing sub-item load change time series types include air conditioning load change time series, drainage equipment load change time series, and emergency equipment load change time series.

[0049] Then, according to the abnormal weather sample gap determination rules, a sample gap record is generated. The specific execution logic of the abnormal weather sample gap determination rule is as follows: when the historical sample count... Fewer than 3, or the longest consecutive time point Fewer than 8 consecutive 15-minute time points, or a set of missing component load types. When the abnormal weather scenario contains at least one missing sub-load type, the condition will be changed. The corresponding information is written into the sample gap record. Setting the historical sample counting threshold to three is to ensure that the same abnormal weather scenario conditions are considered within the same range. The system has at least several historically collected samples that can be compared with each other, thus supporting subsequent deviation extraction. The threshold for the longest continuous time point is set to 8 fifteen-minute time points to ensure that the sample covers a continuous load change process of no less than 2 hours, thereby preserving the temporal continuity of the load response of the sub-equipment before and after the occurrence of abnormal weather.

[0050] For writing sample gap records Any abnormal weather scenario conditions Determine the conditions of this abnormal weather scenario. The corresponding number of supplementary samples to be generated The number of supplementary samples to be generated. The specific determination logic is as follows: for abnormal weather scenarios that have been determined to have sample gaps... If its historical sample count If there are fewer than 3 samples, then supplementary samples will be generated to make up the number of samples. Equal to 3 and the corresponding historical sample count The difference is used to ensure that the total number of samples under this abnormal weather scenario reaches 3. If its historical sample count... Three or more have been reached, but due to the longest consecutive time point... Fewer than 8 consecutive 15-minute time points, or due to a set of missing sub-load types. It contains at least one missing item load type and is written into the sample gap record. Then the number of supplementary samples to be generated Take 1.

[0051] Finally, output sample gap records The sample gap record For multiple sample gap sub-records The set composed of. The first... Sample gap sub-record Specifically, it is expressed as follows: .in, Indicates the sample gap number; Indicates the first Sample gap sub-record Corresponding abnormal weather scenario conditions; Indicates the first Sample gap sub-record Corresponding abnormal weather scenario conditions The existing historical sample count; Indicates the first Sample gap sub-record Corresponding abnormal weather scenario conditions The longest consecutive time point in the existing historical data collection samples; Indicates the first Sample gap sub-record Corresponding abnormal weather scenario conditions The set of missing sub-load types below; Indicates the first Sample gap sub-record Corresponding abnormal weather scenario conditions The number of supplementary samples to be generated is required.

[0052] Specifically, this step uses abnormal weather scenarios as an example. To identify the unit, a joint judgment is made on the number of historically collected samples, the length of continuous time points, and the completeness of the time sequence of load changes, so that sample gaps can be mapped to specific weather types, meteorological warning levels, event stages, and production shifts. Compared with simply counting the total number of samples according to the type of abnormal weather, this step can distinguish the differences in electricity consumption corresponding to different production states under the same weather warning, reduce the situation of grouping samples with different production load characteristics into the same category, and provide a basis for identifying abnormal weather scenarios. Generate supplementary samples to provide gap records.

[0053] For example, the company's historical data contains historical samples corresponding to red typhoon warnings, but these samples all correspond to nighttime production shutdowns; however, for the abnormal weather scenario of "typhoon, red warning, warning duration, daytime production"... Weather scene sample set There are insufficient historical data samples, and the time series of drainage equipment load changes during daytime production is lacking. Therefore, this step will consider the abnormal weather scenario conditions. It was determined that a sample gap existed, and the sample gap was recorded. Write the corresponding historical sample count into it. Longest consecutive time points Itemized load missing type set and the number of supplementary samples to be generated .

[0054] Step S2200: Record the sample gap Determine the length of the supplementary sample sequence and scene condition vector and perturbation sequence This will supplement the sample sequence length. and scene condition vector and perturbation sequence Input conditions to generate a generative adversarial network to generate candidate abnormal weather supplementary samples. The abnormal weather supplementary sample set was obtained after scene consistency checks. .

[0055] In the specific implementation process, this step first reads the sample gap record output in step S2100. For the sample gap record The first in Sample gap sub-record Read the sample gap sub-record Corresponding abnormal weather scenario conditions Longest consecutive time points Itemized load missing type set and the number of supplementary samples to be generated .

[0056] Subsequently, the number was determined Sample gap sub-record Corresponding supplementary sample sequence length The supplementary sample sequence length The specific determination logic is as follows: The longest consecutive time point number... The sample sequence length is supplemented by comparing it with the preset minimum number of consecutive time points of 8. The preset minimum number of continuous time points, 8, represents the number of fifteen-minute time points corresponding to a two-hour continuous load process.

[0057] Based on the above determination logic, if there is already a longest consecutive time point of historically collected samples... If there are fewer than 8, the sample sequence length will be supplemented. Eight time points are set to ensure that the generated samples cover a continuous load change process of no less than 2 hours; if there are already historically collected samples, the longest continuous time point is... If the number reaches or exceeds 8, then the longest consecutive time point count will be used. As a supplement to the sample sequence length, the time length of the generated sample is adapted to the continuous record length of the existing historical samples.

[0058] Determining the length of the supplementary sample sequence Next, this step will consider the abnormal weather scenario conditions. Processing to generate scene condition vectors acceptable to conditional generative adversarial networks Specifically, abnormal weather scenarios and conditions The weather type, meteorological warning level, event stage, and production shift are converted into corresponding preset numerical codes and sorted in the order of "weather type, meteorological warning level, event stage, and production shift" to form a scene condition vector. Among them, the numerical code for weather type is used to distinguish abnormal weather types such as typhoons, rainstorms, and high temperatures; the numerical code for meteorological warning level is used to distinguish between yellow warnings, orange warnings, and red warnings; the numerical code for event stage is used to distinguish between the period before the warning, the period of the warning, and the recovery period after the warning is lifted; and the numerical code for production shift is used to distinguish between different production states.

[0059] The conditional generative adversarial network is an existing conditional generative adversarial network algorithm, which includes a generator and a discriminator. The generator is used to determine the condition vector based on the scene conditions. Perturbation sequence And supplementing the sample sequence length Generate corresponding candidate abnormal weather supplementary samples. The discriminator is used to determine the load sequence of the input sample and its corresponding scene condition vector. Determine whether the load sequence of the input sample meets the conditions of an abnormal weather scenario. The load change characteristics are described below. The load sequence of the input sample includes at least one of the total load sequence and the time series of sub-items load changes; the time series of sub-items load changes includes the time series of air conditioning load changes, the time series of drainage equipment load changes, and the time series of emergency equipment load changes.

[0060] In the same abnormal weather scenario conditions Next, multiple candidate samples with different perturbation sequences are generated. This step sets the generation sequence number. And according to the sample gap number and generate serial number Generate perturbation sequence Specifically, based on the sample gap number. and generate serial number Using the combination of [variables] as a random seed, the uniform distribution sampling function in the existing random number generator is called to [sample the data in the interval...]. Generate sequentially inside A random number; the... Arrange the random numbers in the order they were generated to obtain the perturbation sequence. The perturbation sequence Each random number corresponds to a consecutive 15-minute time point in the candidate sample, used to characterize the conditions of the same anomalous weather scenario. Load variation disturbances at different consecutive 15-minute time points.

[0061] For the Sample gap sub-record and the Each generation sequence number will be used to generate the scene condition vector. Perturbation sequence And supplementing the sample sequence length The generator of the input conditional generative adversarial network generates the first... Candidate abnormal weather supplementation samples The candidate abnormal weather supplementary samples The specific generation logic is as follows: First, from the scene condition vector Limiting the abnormal weather scenario conditions corresponding to the candidate samples Then, the sample sequence length is supplemented. The number of consecutive 15-minute time points included in the candidate samples is limited; finally, the perturbation sequence is used. Adjusting candidate samples under abnormal weather conditions The magnitude and pattern of load changes. The candidate abnormal weather supplementary samples. This includes the total load sequence, the time series of load changes for each component of equipment, and the corresponding abnormal weather scenarios. .

[0062] During the training of the conditional generative adversarial network, the weather scene sample set output in step S1300 is used. Sample type labeling The weather scene samples are used as historical data collection samples, and the corresponding weather type, meteorological warning level, event stage, and production shift are used as training scenario conditions. The discriminator supplements the samples based on the historical data collection samples and the candidate abnormal weather output by the generator. Determine the candidate abnormal weather supplement samples Whether the load sequence and sub-load change time series in the data match the training scenario conditions; the generator adjusts the generation process based on the discriminator's output to supplement the sample with candidate abnormal weather. The total load sequence, the time series of changes in sub-loads, and the relationship between changes between adjacent time points are similar to historical samples collected under similar or identical abnormal weather scenarios.

[0063] Then, supplement the candidate anomalous weather samples output by the generator. Perform scene consistency checks. The scene consistency check includes: first, supplementing candidate abnormal weather samples. The corresponding weather type, meteorological warning level, event stage, and production shift are respectively related to the abnormal weather scenario conditions. Consistent. Second, supplementing the sample with candidate abnormal weather. The number of consecutive 15-minute time points included is not less than the length of the supplementary sample sequence. Third, supplementing the sample with candidate abnormal weather. Includes time-series types of sub-item load changes, covering the set of missing sub-item load types. Fourth, complete the candidate abnormal weather sample. The total load value at each time point in the total load sequence is not less than 0; fifth, candidate abnormal weather supplements the sample. The timing of changes in air conditioning load, drainage equipment load, and emergency equipment load are consistent with the corresponding sub-item equipment types.

[0064] If candidate abnormal weather is added to the sample If the above scenario consistency check is passed, it will be added to the abnormal weather supplementary sample set. Among them, the abnormal weather supplementary sample set. Sample source markers in The subsequent step S2300 uses a calibration sample library for the synthesized scene. Time is uniformly set to This is used to distinguish abnormal weather and supplement the sample source.

[0065] If candidate abnormal weather is added to the sample If the above scenario consistency check fails, the candidate abnormal weather supplementary sample will be discarded. And continue to generate the next candidate abnormal weather supplementary sample. For the same sample gap sub-record According to the generation sequence number The samples are generated and filtered in ascending order until they are added to the abnormal weather supplementary sample set. The sub-record corresponding to the sample gap. The number of abnormal weather supplementary samples has reached the number of supplementary samples to be generated. .

[0066] Finally, output the abnormal weather supplementary sample set. During the generation and filtering process, This represents the candidate abnormal weather supplement samples output by the generator; the candidate abnormal weather supplement samples pass the scene consistency check and are written into the abnormal weather supplement sample set. Then, it was recorded as an abnormal weather supplement sample. ,in Indicates abnormal weather to supplement the sample set The sample number in the database. Supplement the sample for any abnormal weather event. All samples retain the following information: First, the abnormal weather conditions that supplement the sample with the corresponding abnormal weather scenario conditions. The abnormal weather supplement sample is characterized by four elements: 1) the weather type, weather warning level, event stage, and production shift; 2) the total load sequence corresponding to the abnormal weather supplement sample, used to characterize the change of total load of industrial and commercial users over time in this scenario; 3) the time sequence of load changes for each sub-item of equipment corresponding to the abnormal weather supplement sample, used to characterize the load response changes of each sub-item of equipment in this scenario, including the time sequence of air conditioning load changes, drainage equipment load changes, and emergency equipment load changes; and 4) the source attribute of the abnormal weather supplement sample, used to indicate the source of the abnormal weather supplement sample. Derived from conditional generative adversarial networks, its specific sample source labeling The value is uniformly assigned by subsequent step S2300. .

[0067] Specifically, this step records the sample gaps. As a constraint, supplementary samples are generated only for abnormal weather scenarios where the sample size is insufficient, continuous time points are insufficient, or the time sequence of load changes is missing. The generation process is also constrained by weather type, meteorological warning level, event stage, and production shift, ensuring that the supplementary samples correspond to the abnormal weather power consumption scenarios of specific industrial and commercial users. At the same time, the responsibility for setting the sample source label is uniformly handled by S2300 to distinguish between historically collected samples, abnormal weather supplementary samples, and records of scenarios to be predicted.

[0068] Step S2300, based on the weather scene sample set Supplementing the sample set with abnormal weather Determine the actual sample set collected. Set of records of scenarios to be predicted Supplementing samples for abnormal weather According to weather conditions The actual collected sample set Set of records of scenarios to be predicted Supplementing samples for abnormal weather Assign them to the corresponding scene groups to generate a scene calibration sample library. .

[0069] In the specific implementation process, this step first reads the weather scene sample set from step S1300. The weather scene sample set Sample type labeling The weather scene samples are denoted as the real collected sample set. The weather scene sample set Sample type labeling The weather scene samples are denoted as the set of scene records to be predicted. Next, read the abnormal weather supplementary sample set from step S2200. Abnormal weather supplementation samples The abnormal weather supplementary samples read are used as the third type of input samples for inclusion, without setting separate categories. Other sample names listed side-by-side. Among them, abnormal weather supplementary samples. Sample source marker In this step, set as This indicates that the sample originates from the abnormal weather supplementary sample set generated in step S2200. Then, this step is performed according to the weather conditions. Real samples were collected respectively Record of the scenario to be predicted And supplementary samples for abnormal weather Samples are grouped into scenario groups to create a scenario calibration sample library, ensuring that samples with the same weather type, weather warning level, event stage, and production shift are grouped together. .

[0070] The scenario calibration sample library It consists of multiple scene calibration sample groups, each scene calibration sample group corresponding to a weather scene condition. For any weather scenario conditions The corresponding scene calibration sample group is denoted as ,Right now It is a scene calibration sample library A subset of [a set]. Any weather scenario condition. Scene calibration sample group Represented as: .in, Indicates weather scene conditions The following scene calibration sample group; Represents a weather scene sample set Meets weather scenario conditions Authentic collected samples; Indicates abnormal weather to supplement the sample set Meets weather scenario conditions Abnormal weather supplementary samples ; Represents a weather scene sample set Meets weather scenario conditions The record of the scenario to be predicted; the symbol ∪ indicates that the three types of samples are grouped into the same scenario group. This grouping process does not change the load sequence and sample source label of each sample.

[0071] This step uniformly adopts weather scene conditions. As a basis for dividing scene groups. Weather scene conditions. This includes weather type, weather warning level, event stage, and production shift; abnormal weather supplementary samples generated in step S2200. Although it comes from sample gap records Abnormal weather scenario conditions However, it still has its own characteristics related to weather conditions. The same condition fields. Therefore, in the synthetic scene calibration sample library. At that time, supplement the sample with abnormal weather. According to their corresponding weather type, meteorological warning level, event stage, and production shift, they are classified into the corresponding weather scenario conditions. Instead of setting abnormal weather scenario conditions separately as scenario group indexes, this approach allows for the inclusion of abnormal weather scenario conditions.

[0072] Then, this step is... Each sample retains a sample source marker. The sample source marker Used to indicate the source of the sample, its value is... , or .in, This indicates that the sample was actually collected. This indicates that abnormal weather conditions have been added to the sample. This represents a recorded scenario to be predicted. Specifically, it comes from... The actual collected samples, and their sample source markings. Set as From Abnormal weather supplementary samples Its sample source marker Set as From The record of the scene to be predicted, and its sample source label. Set as The sample source is marked. Unlike the sample type labeling in step S1300 , Used to differentiate weather scene sample sets Historical samples and records of scenarios to be predicted are included. Used to differentiate scene calibration sample libraries The source of the samples.

[0073] When the same weather conditions When both real-world samples and supplementary samples from abnormal weather exist simultaneously, this step does not average the two types of samples. Instead, it retains the original sequences and source markers of both types of samples, enabling the S3000 to distinguish the source of the bias when extracting prediction biases. Finally, the biases will be processed according to the weather scenario conditions. All scene calibration sample groups after organization Write to the scene calibration sample library It also retains the number of samples, source composition, and number of consecutive time points in each scenario group.

[0074] Specifically, this step manages real-collected samples, abnormal weather-supplemented samples, and records of the scenarios to be predicted under the same scenario conditions, while retaining the sample source markers. This is used to reduce the mixing of load sequences from different weather types, different meteorological warning levels, or different production shifts. Therefore, the subsequent step S3000 can extract prediction bias within the same scenario group and distinguish whether the bias originates from actual collected samples or from samples supplemented by abnormal weather.

[0075] Step S3000: Calibrate the sample library according to the scene. Train the Long Short-Term Memory network to generate a basic prediction result table. Based on the basic forecast results table Generate a weather deviation sample table Based on the basic prediction results table Sample table of weather deviations Generate scene calibration network packets .

[0076] Specifically, this step aims to integrate the scene calibration sample library from the S2300. As input, the Long Short-Term Memory network is trained and a basic prediction result table is generated. According to the basic forecast results table Generate a weather deviation sample table Based on the aforementioned basic prediction result table Sample table of weather deviations Generate scene calibration network packets This processing method is used to establish a correspondence between the basic prediction results output by the Long Short-Term Memory Network and the abnormal weather deviation samples in abnormal weather load prediction scenarios such as typhoons, rainstorms, or high temperatures, providing input for subsequent load prediction calibration.

[0077] Further, step S3000 includes: Step S3100, from the scene calibration sample library In the process, normal weather training samples are identified, and a basic prediction input vector is constructed based on these normal weather training samples. And train the Long Short-Term Memory network to use the scene calibration sample library. Abnormal weather samples and records of scenarios to be predicted are input into a trained Long Short-Term Memory (LSTM) network to obtain the baseline forecast load. And write it into the basic prediction results table. .

[0078] In the specific implementation process, this step first obtains the scene calibration sample library from step S2300. Screening of sample source markers Real-world samples with no abnormal weather conditions and a weather warning level of 0 were used as training samples for normal weather. This screening method was used to enable the Long Short-Term Memory Network to learn the normal electricity consumption patterns of industrial and commercial users under conditions without typhoons, rainstorms, or high-temperature warnings, thereby reducing the impact of load fluctuations in air conditioning, drainage, and emergency equipment caused by abnormal weather on the training of normal patterns.

[0079] Next, for each 15-minute time point in the normal weather training sample... Construct the basic prediction input vector The basic prediction input vector From a point in time Total load values ​​and time points for the previous four consecutive 15-minute intervals Corresponding production shifts and time point number The composition can be represented as: .in, , , and Representing time points The total load values ​​at the 4th, 3rd, 2nd, and 1st 15-minute time points were all derived from the scene calibration sample library. The total load sequence corresponding to the actual collected samples in the data; Indicates a point in time Corresponding production shifts; Indicates a point in time The 15-minute time points within a calendar day are numbered from 1 to 96, corresponding to the 96 15-minute time points within a calendar day. For training samples in normal weather conditions where the preceding four consecutive 15-minute time points are incomplete, these are not used as training samples for the Long Short-Term Memory (LSTM) network model to avoid missing terms in the input vector.

[0080] Then, using the base prediction input vector As input, time point Corresponding total load value The training objective is to train a Long Short-Term Memory (LSTM) network. This LSM network is used as a regression network to output load estimates under normal weather conditions; during the training phase, the LSM network is trained based on the baseline prediction input vector. Output time point Corresponding training prediction load value And based on the training predicted load value With time point Corresponding total load value The difference between the values ​​determines the training error, enabling the trained Long Short-Term Memory (LSTM) network to characterize the load variation patterns under normal weather conditions, which are jointly determined by the preceding total load, production shifts, and intraday time. The training predicted load value... Training errors are only used in the training process of the Long Short-Term Memory network and are not included in the basic prediction results table. .

[0081] Subsequently, the scene calibration sample library was used. Abnormal weather samples and predicted scenario records are used as call samples and input into the trained Long Short-Term Memory network to obtain the basic predicted load value at each time point. The basic predicted load value This represents the load estimation result under normal weather patterns, output by the trained Long Short-Term Memory network, without considering the additional effects of abnormal weather.

[0082] For abnormal weather samples, construct the basic prediction input vector using the same input structure as in the training phase. Wherein, the basic prediction input vector The preceding total load value to The abnormal weather sample is from the scene calibration sample library. The total load sequence is stored in the database. Since this abnormal weather sample is a real sample, the corresponding preceding load value can be directly taken from historical measured data.

[0083] For the scenario records to be predicted, the basic predicted load values ​​for each time point are generated sequentially according to time order. At the point of construction time The corresponding basic prediction input vector At that time, production shifts The time point number is determined by the production plan data recorded in the scenario to be predicted. From a point in time The location is determined within a natural day; for the base prediction input vector If the preceding total load position in the forecast scenario does not contain a corresponding total load value, then the already generated basic forecast load value will be used for recursive filling. Specifically, the time point... to Corresponding basic forecast load value to Write the base prediction input vector The corresponding preceding total load position is used to form a continuous input sequence for forecasting subsequent time periods. This allows the scenario to be predicted to be recorded based on the previously output baseline forecast load value even when historical measured loads are unavailable. Complete the recursive prediction.

[0084] Finally, the time point Weather scene conditions Basic predicted load value Total load value of calibration samples for corresponding scenarios and sample source marking Write into the basic prediction results table For sample source labeling The record of the scenarios to be predicted does not include the actual total load value at the target prediction time point, so the basic prediction result table is not included. The corresponding total load value Record as empty.

[0085] Basic Prediction Results Table Single basic prediction results corresponding to multiple time points The individual basic prediction results are composed in chronological order. It can be represented as: .in, Indicates a time point in the fifteen-minute time range; Indicates a point in time The weather conditions of the scene; This represents the basic forecast load value; Represents the scene calibration sample library The total load value of the corresponding sample; Indicates the sample source marker.

[0086] Specifically, this step uses real-world samples collected without abnormal weather conditions to train the Long Short-Term Memory (LSTM) network. This allows the LTM network to learn the impact of shift changes and intraday variations on industrial and commercial user loads, while excluding additional load fluctuations caused by abnormal weather events such as typhoons, heavy rains, and high-temperature warnings from the basic forecasting process. Therefore, in subsequent steps, when abnormal weather samples are input into the LTM network, its output basic forecast load value... It can serve as a baseline load excluding weather disturbances, used for comparison with the actual total load, thereby extracting the load deviation caused by the operation of air conditioning, drainage, and emergency equipment under abnormal weather conditions, and reducing the amount of this load deviation pre-mixed into the baseline forecast results table. The basic forecast load value recorded in the middle middle.

[0087] Further, please refer to Figure 3 As shown, Figure 3 This is a schematic diagram of the process for generating and writing the basic predicted load value in this invention.

[0088] Step S3200, from the basic prediction results table Determining Abnormal Weather Scenarios The corresponding deviation extraction object is used to calculate the prediction deviation. And according to the sequence number within the stage Generate a weather deviation sample table The sequence number within the aforementioned stage According to weather scene conditions The order of 15-minute time points within the mid-event phase is established.

[0089] In the specific implementation process, this step first starts from the basic prediction result table in step S3100. Read sample source markers for or And weather conditions Records with corresponding weather types of typhoon, rainstorm, or high temperature, and whose meteorological warning level value is not 0, are used as the bias extraction targets. (The sample source marker is mentioned.) This indicates that the record corresponds to an actual collected sample. This indicates that the record corresponds to an abnormal weather supplement sample; sample source marker. The recorded scenarios to be predicted are not included in the prediction bias calculation. The weather scenario conditions... It consists of weather type, meteorological warning level, event stage, and production shift, based on the weather scenario conditions. Scenes selected from the data that are typhoon, rainstorm, or high temperature and whose weather warning level is not set to 0 are recorded as abnormal weather scene conditions. This is used to represent a scenario with abnormal weather types and under weather warning conditions.

[0090] Next, according to the abnormal weather scenario conditions The deviation extraction objects are grouped to obtain the conditions for each abnormal weather scenario. The corresponding deviation extraction group; each deviation extraction group includes the conditions of the abnormal weather scenario. The actual sample records collected below, and the conditions used to supplement this abnormal weather scenario. Abnormal weather conditions at locations where actual samples were missing were used to supplement sample records. Within each deviation extraction group, a sequence number was established based on the 15-minute time points within the event phase. The event phase can be one of the following: the pre-warning period, the warning duration, or the recovery period after the warning is lifted. The phase number is specified within each phase. Starting from 1 and incrementing, it is used to indicate the sequential position of a certain point in time within the corresponding event phase. For example, the phase number of the first fifteen minutes within a certain warning duration. The sequence number within the second 15-minute time point is 1. The value is 2.

[0091] Then, for each time point involved in deviation extraction Read the basic prediction results table Corresponding time point Total load value and basic forecast load values To calculate time points Corresponding prediction bias The prediction deviation Numerically equal to, at a point in time. Total load value Subtract the base forecast load value When the prediction deviation A value greater than 0 indicates the basic predicted load value. Less than the corresponding total load value; when the prediction deviation When it is less than 0, it represents the basic predicted load value. Greater than the corresponding total load value; when the prediction deviation When equal to 0, it represents the base forecast load value. Same as the corresponding total load value.

[0092] Subsequently, based on the weather scene sample in step S1300 In and at time point Corresponding sub-item equipment load change value Identify the main sources of load for each sub-item. This applies to any of the following sub-items: air conditioning equipment, drainage equipment, and emergency equipment. Calculate the equipment for each sub-item At the point of time The proportion of load changes The percentage of load changes. The specific calculation logic is as follows: reading and time point Corresponding sub-equipment Sub-item equipment load change value And read the corresponding sub-item load change time series, that is, at the time point. The corresponding load changes for air conditioning equipment, drainage equipment, and emergency equipment. The load changes for these sub-items... Take the absolute value and divide it by the sum of the absolute values ​​of the load changes for the air conditioning equipment, drainage equipment, and emergency equipment to obtain the load change percentage. When the denominator is 0, it indicates that the three components are at the specified time point. No identifiable load changes were observed at any point in time, so the primary load source for that time point was recorded as empty. When the denominator was not zero, the equipment with the largest load change percentage was recorded as the primary load source for that time point. If the load change percentage of two or more equipment items was not recorded, the primary load source for that time point was recorded as empty. If the values ​​are the same and both are the maximum values, then the parallel sub-items will be written together into the main sub-item load source.

[0093] Then, this step is performed under the same abnormal weather scenario conditions. Below, in the actual collected samples and the supplementary samples for abnormal weather, the serial numbers within the same period are... Prediction bias Grouping and summarizing samples by source to generate serial numbers within the same period. The following represents the deviation value. Specifically, if the sequence number within the stage... The sample source marker exists below. The prediction deviation is then determined by the sequence number within that stage. The median of the prediction bias of the actual collected samples is used as the representative bias value. If the sequence number within the stage... There is no prediction bias based on actual collected samples, but there is sample source labeling. The prediction deviation is then determined by the sequence number within that stage. The median of the predicted bias for abnormal weather conditions is used as the representative bias value. The median refers to the median value calculated using the sequence number within the same period. Prediction bias The value in the middle position after sorting by numerical value; when the prediction bias is involved in the sorting. When the quantity is even, the average of the two middle values ​​is taken. By prioritizing the prediction bias of actual collected samples, when actual collected samples are missing, the prediction bias of abnormal weather-supplemented samples is used to fill in the gaps. Actual collected samples are prioritized for forming the bias sequence, and abnormal weather-supplemented samples are used to fill in the missing sequence numbers within the stages from the actual collected samples. .

[0094] Finally, the conditions for each abnormal weather scenario will be... Next, press the sequence number within the stage. The representative deviation value of the arrangement is written into the abnormal weather scenario conditions. Corresponding prediction bias sequence The prediction bias sequence is mentioned above. By the sequence number within the stage The representative deviation values ​​are arranged from smallest to largest.

[0095] At the same time, a sequence of deviation directions is generated sequentially based on the sign of each representative deviation value. Specifically, when the representative deviation value is greater than 0, the corresponding deviation direction is recorded as positive deviation; when the representative deviation value is less than 0, the corresponding deviation direction is recorded as negative deviation; and when the representative deviation value is equal to 0, the corresponding deviation direction is recorded as zero deviation.

[0096] The duration of deviation is calculated based on the length of time during which the representative deviation value continuously meets the effective deviation condition. Specifically, the effective deviation condition refers to a deviation value whose absolute value is greater than or equal to the effective deviation threshold. The number of consecutive 15-minute time points that satisfy the effective deviation condition is multiplied by 15 minutes to obtain the corresponding deviation duration.

[0097] At the same time, the corresponding sequence number within the stage The main load sources identified below are listed in order of their phase number. Write the main component load source sequence .

[0098] The effective deviation threshold is used to exclude minor differences caused by fluctuations in meter sampling, and its value is set to 3% of the average total load value in normal weather training samples. When the absolute value of the representative deviation value is greater than or equal to the effective deviation threshold, the sequence number within this stage... The corresponding time point is recorded as the effective deviation time point; the number of 15-minute time points with consecutive effective deviation time points multiplied by 15 minutes is the corresponding deviation duration.

[0099] Weather Deviation Sample Table A single weather deviation sample in the data, wherein the single weather deviation sample is generated by abnormal weather scenario conditions. Corresponding sequence number within the stage Prediction bias sequence arranged from smallest to largest The predicted deviation sequence Corresponding deviation direction sequence The predicted deviation sequence Corresponding deviation duration The prediction deviation sequence Serial number within each stage Corresponding dominant load source and for recording the prediction bias sequence Serial number within each stage The corresponding source of bias is indicated by whether it is from actual sample collection or from bias sources added due to abnormal weather. composition.

[0100] Specifically, this step applies to the same abnormal weather scenario conditions. The difference between the basic predicted load value and the corresponding total load value is organized into a sequence number within each stage. The system arranges the predicted deviation sequence and simultaneously records the deviation direction, duration, main sub-load sources, and deviation source markers. This allows for the retention of priority for actual collected samples while supplementing missing phases with abnormal weather data, thus enriching the weather deviation sample table. It can be used for subsequent load forecast calibration for abnormal weather scenarios such as typhoons, rainstorms, or high temperatures.

[0101] Step S3300, from the basic prediction results table Read the record of the scene to be predicted from the weather deviation sample table. Searching for time points Weather conditions A matching single weather deviation sample is used to generate a single scene calibration rule, which is then written into the scene calibration network packet. .

[0102] In the specific implementation process, this step first reads the basic prediction result table in step S3100. Medium sample source marker The record of the scene to be predicted is used to obtain each 15-minute time point in the target time period to be predicted. Corresponding basic forecast load value and weather conditions The time point mentioned above. Weather scene conditions This includes weather type, weather warning level, event stage, and production shift.

[0103] Next, the weather deviation sample table output in step S3200 In the search, find the time point mentioned above. Corresponding weather scene conditions A single matching weather deviation sample. Matching means that the weather type, weather warning level, event stage, and production shift are all the same; if any field differs, the corresponding weather deviation sample will not be recorded in the table. As of that point in time The calibration basis.

[0104] Then, for the weather deviation sample table that can be matched The recorded time point Determine the sequence number of this point in time within the event phase. and from the corresponding prediction deviation sequence Read the sequence number within the aforementioned stage. The corresponding representative deviation value. When predicting the deviation sequence There is no sequence number within the stage. When the corresponding representative deviation value is used, this step does not refer to that time point. Bind the representative deviation value, and in this weather scenario condition The corresponding single-scene calibration rule will include the sequence number within this stage. Marked as a location missing marker.

[0105] Then, the time point Basic forecast load value Weather and scene conditions and the weather conditions Matching prediction bias sequence Deviation source marking and the sequence of major component load sources To bind and form a connection with the weather scene conditions. The corresponding single-scene calibration rule is set, and a scene calibration rule number is assigned to each scene calibration rule. and the range of location numbers The scenario calibration rules are a set of rules constructed according to weather scenario conditions. Each individual scenario calibration rule is a calibration rule established within this set for a specific weather scenario condition and its corresponding position within a given stage. When multiple time points exist under the same weather scenario condition, multiple corresponding individual scenario calibration rules can be formed. The position number range... Represents the prediction bias sequence The range of covered location numbers is used to determine whether the corresponding time point within the predicted target time period is within the calibrable range.

[0106] Finally, this step will generate the Long Short-Term Memory network and the basic prediction results table that were trained in step S3100. Weather Deviation Sample Table The set of individual scene calibration rules corresponding to each weather scenario condition is written into the scene calibration network packet. .

[0107] Step S4000: Calibrate network packets according to the scenario. Generate calibration trigger record According to the calibration trigger record Determine the post-calibration load prediction sequence And based on the calibrated load prediction sequence and calibration trigger record Generate final load forecast results .

[0108] Specifically, this step aims to process the scene calibration network packets from the S3300. As input, a calibration trigger record is generated. According to the calibration trigger record Determine the post-calibration load prediction sequence And based on the calibrated load prediction sequence With calibration trigger record Generate the final load forecast results This processing method is used to determine 15-minute time points in abnormal weather scenarios such as typhoons, heavy rain, and high temperatures. Whether to trigger abnormal weather calibration and adjust the calibrated load forecast value The calibration trigger criteria are written into the same prediction result. The generated final load prediction result... This is used to represent the 15-minute load forecast output for industrial and commercial users within the target time period to be predicted.

[0109] Further, step S4000 includes: Step S4100: Read the scene calibration network packet Mid-15 minute time point Corresponding weather scene conditions Based on weather scene conditions Retrieve a matching single-scene calibration rule for the search key, and generate a calibration trigger flag according to the calibration trigger determination rule. Write to calibration trigger record .

[0110] In the specific implementation process, this step first takes the scene calibration network packet output from step S3300. Read each 15-minute time point within the target time period to be predicted. Corresponding weather scene conditions The aforementioned weather scene conditions This includes weather type, weather warning level, event stage, and production shift.

[0111] Next, at the aforementioned time point Corresponding weather scene conditions As a search key, in the scene calibration network packet Searching for the weather scene conditions A matching single-scene calibration rule is found. "Matching" means that all four fields—weather type, weather warning level, event stage, and production shift—are identical. If any field differs, no usable scene calibration rule is considered matched. If a matching single-scene calibration rule is found, its scene calibration rule number is read. and location number range If no matching rule is found, the scene calibration rule number will be changed. Record as empty.

[0112] Then, determine the time point. The sequence number within the phase of its event phase The sequence number within the aforementioned stage. Indicates a point in time The position numbering is arranged in 15-minute time intervals within the same weather event phase, with values ​​increasing from 1. For example, within a certain warning duration, the position number of the first 15-minute time interval. 1, the sequence number within the second 15-minute time point. The value is 2. For time points within the same event phase, this step determines their phase sequence number sequentially according to chronological order. and the sequence number within that stage The range of location numbers in the matched single scene calibration rule Perform an inclusion relationship comparison to determine the sequence number within the stated stage. Does it fall within the range of the stated location number? Inside.

[0113] The specific execution logic for the inclusion relationship comparison is as follows: if the sequence number within the stage... Falling into the range of the specified location numbers If the position corresponding to that time point is within the calibrable range of the rule, then it indicates that the position is within the calibrable range of the rule; if the sequence number within the stage... Not falling within the range of the specified location numbers If the result is negative, it means that the location corresponding to that time point is outside the calibrable range of the rule.

[0114] Subsequently, a calibration trigger flag is generated according to the calibration trigger determination rules. The calibration trigger determination rule refers to the rule used to determine the time point. The specific execution logic for determining whether to apply calibration under abnormal weather conditions is as follows: A calibration trigger flag will be set when the following conditions are met simultaneously. Recorded as 1, representing a time point. Trigger abnormal weather calibration. The conditions are as follows: First, the weather type is typhoon, heavy rain, or high temperature; second, the meteorological warning level is not 0; third, the scene calibration network packet... There are weather scene conditions in it. Completely consistent single-scene calibration rules; fourth, the sequence number within the phase. The position number range that falls within this rule Within the specified area, and the location is not marked as missing. If any of the above conditions are not met, the calibration trigger flag will be activated. Recorded as 0, representing a time point. Do not trigger abnormal weather calibration.

[0115] Generate a non-triggered cause flag when abnormal weather calibration is not triggered. The non-triggered reason flag. The values ​​for include 0, 1, 2, and 3. Among them, This indicates that abnormal weather calibration has been triggered. This indicates that the weather type is normal or the weather warning level is 0. This indicates that the network packet was not calibrated in the scene. Matched with Completely consistent single-scene calibration rules. Indicates a rule was matched, but the sequence number within the stage was not specified. Not within the range of location numbers Or the location may be marked as missing.

[0116] Finally, each time point Corresponding weather scene conditions Calibration trigger flag Scene calibration rule number Serial number within the stage and the flag for not triggering the reason Write calibration trigger record For any given time point The calibration trigger record Single trigger record It can be represented as: .

[0117] Specifically, this step uses weather type, meteorological warning level, event stage, production shift, and stage number as trigger conditions for abnormal weather calibration at each 15-minute time point within the target time period to be predicted. Calibration is only triggered when the corresponding scenario calibration rule exists and the stage number falls within the location number range. Calibration is triggered within the time frame. This avoids using deviations from nighttime shutdown scenarios for daytime production scenarios, avoids using deviations from the duration of an alert for the recovery period after the alert is lifted, and avoids overlaying calibration deviations in areas not covered by the deviation sequence.

[0118] Step S4200, based on the calibration trigger record Mid-calibration trigger flag Determine the time point Corresponding calibration deviation value , will the calibration deviation value With scene calibration network packet Basic forecast load value Add them together to obtain the calibrated load prediction value. And generate a calibrated load prediction sequence in chronological order. .

[0119] In the specific implementation process, this step first takes the scene calibration network packet output from step S3300. Read each 15-minute time point within the target time period to be predicted. Corresponding basic forecast load value The basic predicted load value Derived from the Long Short-Term Memory network trained in step S3100, it represents the 15-minute load forecast value without superimposed abnormal weather calibration bias.

[0120] Next, read the calibration trigger record output in step S4100. Mid-calibration trigger flag When the calibration trigger flag When the value is 0, the scene calibration network packet output in step S3300 is not invoked. The corresponding predicted bias sequence , time point Corresponding calibration deviation value Record it as 0, and set the calibrated load prediction value to 0. Equal to the basic predicted load value The calibration deviation value is mentioned above. This indicates that under the same weather conditions, the scene calibration network packets... Read from the time point Corresponding stage sequence number The corresponding representative deviation value is used to match the base forecast load value. Perform calibration.

[0121] When the calibration trigger flag When the value is 1, it is determined according to the scene calibration rule number. In scene calibration network packet Reading weather scene conditions A completely consistent single-scene calibration rule, and the prediction bias sequence bound to that rule. Reading and the sequence number within the stage The corresponding representative deviation value is used as the calibration deviation value. .

[0122] Then, the calibration deviation value Superimposed on the basic forecast load value The calibrated load prediction value was obtained. When the superimposed calibrated load prediction value If the value is less than 0, then the calibrated load prediction value will be... A value of 0 is used to ensure the predicted load value after calibration. Not less than 0, which conforms to the physical meaning of non-negative electrical load.

[0123] Subsequently, the calibration deviation value was read. Serial number within the same stage Corresponding main component load sources and deviation sources markings When the calibration trigger flag When the value is 1, the main sub-load sources are related to weather scenario conditions. Completely consistent single-scene calibration rules, bound to the main component load source sequences. The value is obtained from [the data source], and can be categorized as air conditioning equipment, drainage equipment, emergency equipment, or multiple parallel sub-items. The deviation source marker is [the reference to the source of the deviation]. From the perspective of weather scene conditions The identical single-scene calibration rule is obtained from the bound deviation source marker, and its value is... or ,in This indicates the calibration deviation value. Derived from real collected samples, This indicates the calibration deviation value. The sample was supplemented based on abnormal weather conditions. When the calibration trigger flag... When the value is 0, the main sub-item load sources and deviation sources are marked. All are empty.

[0124] Finally, the basic forecast load values ​​for each time point are listed in chronological order. Calibration deviation value , post-calibration load prediction Scene calibration rule number Serial number within the stage Marking of main load sources and deviation sources Write the calibrated load prediction sequence .

[0125] For example, under the weather scenario of "heavy rain, orange alert, alert duration, day shift production", the scenario calibration network packet A single scene calibration rule already exists that perfectly matches the weather scenario conditions. If the sequence number within the stage corresponding to the 2nd to 8th 15-minute time points during the forecast period... The location number range that falls within this single scene calibration rule And calibration trigger flag Then, the sequence number within the stage will be read sequentially. The corresponding representative deviation value is used as the calibration deviation value. And superimposed on the basic predicted load value. This yields calibrated load forecast values ​​that reflect the increasing load trend of drainage equipment during the continuous rainstorm warning process. For calibration trigger flags At the specified time point, the basic predicted load value is directly retained. As the predicted load value after calibration .

[0126] Step S4300, based on the calibrated load prediction sequence and calibration trigger record According to time points Perform record correspondence verification, and generate the final load forecast result by sequentially analyzing the records corresponding to each time point that pass the verification. .

[0127] In the specific implementation process, this step first reads the calibrated load prediction sequence. Post-calibration load prediction values Calibration deviation value Main load sources and deviation sources markings Next, read the calibration trigger record output in step S4100. At the same time point Corresponding weather scene conditions Calibration trigger flag Scene calibration rule number Serial number within the stage and the flag for not triggering the reason .

[0128] Subsequently, the above fields were processed by time point. The record correspondence verification. The time-point... The record correspondence verification refers to: based on time points As a related field, check the calibration trigger record. With the calibrated load prediction sequence Do all of them have the same point in time? The corresponding record.

[0129] If the calibration trigger record and the calibrated load prediction sequence The same time point exists in both. The records will be merged by field to form the final load forecast result. A single record in the [database name]. If the calibration trigger record... There are time points in it The calibrated load prediction sequence If a corresponding record is missing, then that time point... Not included in the final load forecast results and the time point Record any missing output records. If the calibrated load prediction sequence... There are time points in it The calibration trigger record If a corresponding record is missing, that time point will not be generated. Corresponding final load forecast results Records are kept to ensure the final load forecast results. Each record in the database has a corresponding calibration trigger basis.

[0130] Finally, this step will be done by time point. The records corresponding to each time point in the record correspondence verification are arranged in chronological order to obtain the final load forecast result. The final load forecast results This is used to represent the 15-minute load forecast output results for industrial and commercial users within the target forecast period.

[0131] For any point in time Final load forecast results Each record contains a 15-minute time point. , post-calibration load prediction Weather and scene conditions Whether to trigger calibration for abnormal weather Scene calibration rule number The sequence number within the stage representing the location of the deviation call. Calibration deviation value Main load sources and deviation sources are marked. And a flag indicating why it was not triggered. The post-calibration load prediction value... As of that point in time The corresponding final load forecast; the weather scenario conditions Used to indicate the weather type, meteorological warning level, event stage, and production shift corresponding to that time point; the calibration trigger mark This is used to indicate whether calibration was performed using abnormal weather deviations at that specific time point; the scene calibration rule number. and the sequence number within the stage Used to indicate calibration deviation value The basis for invocation; the main sub-item load source is used to indicate the sub-item equipment type that mainly corresponds to the calibration deviation value; the deviation source marker This is used to indicate whether the calibration deviation value comes from a real sample or a sample supplemented by abnormal weather.

[0132] When the calibration trigger flag When the value is 0, the calibrated load prediction value Equal to the basic predicted load value And the scene calibration rule number Serial number within the stage Marking of main load sources and deviation sources All values ​​are recorded as empty; calibration deviation value. Recorded as 0. When the calibration trigger flag... When the value is 1, the calibrated load prediction value Based on predicted load values Superimposed corresponding calibration deviation values The result afterwards.

[0133] For example, when industrial and commercial users are still in daytime production during the duration of an orange rainstorm warning, the final load forecast result is... It can record abnormal weather calibration deviations caused by increased load on drainage equipment at corresponding time points; when air conditioning equipment continues to run during the duration of a high-temperature orange alert, the final load forecast results... It can record the increase in electricity load caused by air conditioning equipment at corresponding time points; when emergency equipment is gradually taken out of operation during the recovery period after the typhoon warning is lifted, the final load forecast result is obtained. It can record the corresponding load decline deviation at the corresponding time points. Therefore, the final load forecast result... It not only provides 15-minute load forecasts for industrial and commercial users, but also distinguishes whether the forecasts have been calibrated for abnormal weather and identifies the main equipment types corresponding to the calibration deviations. This enables dispatchers or energy management systems to arrange power management strategies for drainage equipment, air conditioning equipment, or emergency equipment based on the sources of load changes under different abnormal weather scenarios.

[0134] Example 2: This embodiment, based on Embodiment 1, provides an intelligent power load forecasting system for industrial and commercial users, such as... Figure 4 As shown, the system includes a scene construction module, a scene calibration module, a deviation calibration module, and a load prediction module; The scenario construction module is used to construct scenarios based on the acquired raw electricity consumption and meteorological data. Generate aligned time load records Weather event snippets Based on the alignment time load record Weather event snippets Determine weather scene conditions Generate a weather scene sample set .

[0135] The scene calibration module is used to perform calibration based on a weather scene sample set. Determine abnormal weather scenarios and conditions And generate sample gap records Based on the sample gap record Generate supplementary sample set for abnormal weather Combined with weather conditions Generate a scene calibration sample library .

[0136] The deviation calibration module is used to calibrate the sample library according to the scenario. Train the Long Short-Term Memory network to generate a basic prediction result table. Based on the basic forecast results table Generate a weather deviation sample table Based on the basic prediction results table Sample table of weather deviations Generate scene calibration network packets .

[0137] The load prediction module is used to calibrate network packets according to the scenario. Generate calibration trigger record According to the calibration trigger record Determine the post-calibration load prediction sequence And based on the calibrated load prediction sequence and calibration trigger record Generate final load forecast results .

[0138] The parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0139] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent power load forecasting for industrial and commercial users, characterized in that, include: Based on the acquired raw electricity consumption and meteorological data, aligned time load records and weather event fragment records are generated. Based on the aligned time load records and weather event fragment records, weather scenario conditions are determined, and a weather scenario sample set is generated. Based on the weather scenario sample set, abnormal weather scenario conditions are determined and sample gap records are generated. Based on the sample gap records, an abnormal weather supplementary sample set is generated, and a scenario calibration sample library is generated by combining the weather scenario conditions. A long short-term memory network is trained based on the scenario calibration sample library to generate a basic prediction result table. A weather deviation sample table is generated based on the basic prediction result table. A scenario calibration network package is generated based on the basic prediction result table and the weather deviation sample table. Based on the scenario calibration network packets, a calibration trigger record is generated. The calibration trigger record is used to determine the post-calibration load prediction sequence. Finally, the load prediction result is generated based on the post-calibration load prediction sequence and the calibration trigger record. The process involves generating aligned time-load records and weather event fragment records based on the acquired raw electricity consumption and meteorological data, determining weather scenario conditions based on the aligned time-load records and weather event fragment records, and generating a weather scenario sample set, including: The original electricity consumption and meteorological data are aligned to time points in 15-minute time intervals. For missing sub-load values ​​at individual time points, the average value of neighboring points is used to fill in the missing values ​​and generate aligned time load records. The meteorological warning period is determined based on the meteorological warning level and weather type in the aligned time load record, and the meteorological warning period is segmented according to the changes in the meteorological warning level to generate weather event segment records; Weather scenario conditions are determined based on weather event fragment records and aligned time load records. Under these weather scenario conditions, a total load sequence and a time series of sub-item load changes are constructed to generate a weather scenario sample set. The total load sequence consists of total load values ​​arranged in chronological order. The time series of sub-item load changes consists of sub-item equipment load change values ​​arranged in chronological order. The alignment of the raw electricity consumption and meteorological data with time points in 15-minute intervals includes: Acquire raw electricity consumption and meteorological data, which includes total load data collected by smart meters at the main distribution cabinet, sub-load data collected by sub-meters on the air conditioning system, drainage pump group, and emergency lighting branch circuit, production shift data from the production planning system, and meteorological warning data, meteorological warning level, and weather type from the meteorological warning interface. The total load data, sub-load data, and production shift data are grouped into adjacent 15-minute time intervals according to the collection time, and the start time of the 15-minute time interval is used as the unified time point. The release time, upgrade time, downgrade time, and cancellation time of the meteorological warning data are grouped into the corresponding time points according to the same 15-minute time interval.

2. The intelligent power load forecasting method for industrial and commercial users according to claim 1, characterized in that, The method of using the neighboring average to fill in the missing component load values ​​at a single time point includes: When the time point If the sub-item load value of the processing equipment is missing, but the sub-item load values ​​corresponding to the previous 15-minute time point and the next 15-minute time point both exist, then the average of the sub-item load values ​​corresponding to the previous 15-minute time point and the next 15-minute time point is calculated to obtain the time point. The corresponding sub-item load value is written into the sub-item load record; If the time point If the corresponding sub-load value is missing on either side of the 15-minute time point before or after the point, then the time point is not considered. The corresponding sub-load values ​​were supplemented, and the time points were... The corresponding sub-item load records are marked as records to be supplemented; A data source marker is retained for the sub-load record. If the data source marker is 0, the sub-load value is the directly collected value; if the data source marker is 1, the sub-load value is the neighboring average value.

3. The intelligent power load forecasting method for industrial and commercial users according to claim 1, characterized in that, The generated weather event fragment records include: The time of issuance of each meteorological warning data is used as the starting boundary of the warning process, and the time of cancellation of the meteorological warning data is used as the ending boundary of the warning process. When the weather warning level changes on the same day, the upgrade or downgrade time in the weather warning data is used as the new segment boundary, and the time interval corresponding to the previous weather warning level and the time interval corresponding to the next weather warning level are respectively regarded as different segments. Based on the starting boundary, ending boundary, and segmentation boundary, weather event segments are constructed; The construction logic of the weather event segment is as follows: the four 15-minute time points before the starting boundary are designated as the pre-warning period, the time points between the starting boundary and the ending boundary are designated as the warning duration period, and the four 15-minute time points after the ending boundary are designated as the recovery period after the warning is lifted; if the meteorological warning level changes during the warning duration period, the warning duration period is further divided into multiple adjacent segments at the corresponding segment boundaries. Write the weather type, meteorological warning level, event stage, and start and end time corresponding to the weather event segment into the weather event segment record; The event phase can be one of the following: the period before the warning, the period during which the warning is in effect, or the recovery period after the warning is lifted.

4. The intelligent power load forecasting method for industrial and commercial users according to claim 1, characterized in that, The process of determining abnormal weather scenario conditions based on a weather scenario sample set and generating sample gap records, generating an abnormal weather supplementary sample set based on the sample gap records, and generating a scenario calibration sample library by combining the weather scenario conditions includes: Screen abnormal weather scenario conditions from the weather scenario sample set, and generate sample gap records based on the abnormal weather scenario conditions; Based on the sample gap record, determine the length of the supplementary sample sequence, the scene condition vector, and the perturbation sequence. Input the length of the supplementary sample sequence, the scene condition vector, and the perturbation sequence into the conditional generative adversarial network to generate candidate abnormal weather supplementary samples. After scene consistency check, obtain the abnormal weather supplementary sample set. The abnormal weather supplementary sample set includes a sample source marker; the sample source marker can be a real collected sample, an abnormal weather supplementary sample, or a record of a scenario to be predicted; Based on the weather scenario sample set and the abnormal weather supplementary sample set, the real sample set, the scenario record set to be predicted, and the abnormal weather supplementary sample set are determined. According to the weather scenario conditions, the real sample set, the scenario record set to be predicted, and the abnormal weather supplementary sample set are classified into the corresponding scenario groups to generate a scenario calibration sample library.

5. The intelligent power load forecasting method for industrial and commercial users according to claim 4, characterized in that, The step of filtering abnormal weather scene conditions from the weather scene sample set and generating sample gap records based on the abnormal weather scene conditions includes: Read weather scene conditions from the weather scene sample set, and filter out abnormal weather scene conditions whose weather type is typhoon, rainstorm or high temperature and whose meteorological warning level is not 0. For each abnormal weather scenario, count the number of historical samples marked as historical samples in the weather scenario sample set, and count the number of the longest consecutive time points recorded at 15-minute intervals in the same historical sample under the abnormal weather scenario. Check whether the timing sequence of load changes for each component of equipment exists under the abnormal weather scenario. If any timing sequence of load changes is missing, write the missing timing sequence of load change into the missing load type set. When the historical sample count is less than 3, or the longest consecutive time point is less than 8 consecutive 15-minute time points, or the set of missing sub-load types contains at least one missing sub-load type, the information corresponding to the abnormal weather scenario conditions will be written into the sample gap record. For abnormal weather scenarios where a sample gap has been identified, if the historical sample count is less than 3, the number of supplementary samples to be generated in the sample gap record is equal to the difference between 3 and the corresponding historical sample count; if the historical sample count has reached or exceeded 3, but is written into the sample gap record because the longest consecutive time point is less than 8 consecutive 15-minute time points, or because the set of missing sub-load types contains at least one missing sub-load type, the number of supplementary samples to be generated in the sample gap record is 1.

6. The intelligent power load forecasting method for industrial and commercial users according to claim 4, characterized in that, The process involves determining the length of the supplementary sample sequence, the scene condition vector, and the perturbation sequence based on the sample gap record. This supplementary sample sequence length, scene condition vector, and perturbation sequence are then input into a conditional generative adversarial network to generate candidate abnormal weather supplementary samples. A scene consistency check is then performed to obtain an abnormal weather supplementary sample set, including: Read the abnormal weather scenario conditions, the longest consecutive time point count, the set of missing sub-load types, and the number of samples to be generated and supplemented from the sample gap record; The longest consecutive time point is compared with the preset minimum consecutive time point of 8, and the larger of the two values ​​is taken as the length of the supplementary sample sequence. The weather type, meteorological warning level, event stage, and production shift in the abnormal weather scenario conditions are converted into corresponding preset numerical codes, and sorted in the order of weather type, meteorological warning level, event stage, and production shift to form a scenario condition vector. Using the combination of the sample gap number and the generated sequence number as the random seed, the uniform distribution sampling function in the random number generator is called to sample within the interval. A number of random numbers corresponding to the length of the supplemented sample sequence are generated sequentially, and the random numbers are arranged in the order of generation to obtain the perturbation sequence; The scene condition vector, disturbance sequence, and supplementary sample sequence length are input into the generator of the conditional generative adversarial network to generate candidate abnormal weather supplementary samples. The candidate abnormal weather supplementary samples include the total load sequence, the time sequence of load changes of each sub-item equipment, and the corresponding abnormal weather scene conditions. The candidate abnormal weather supplement samples output by the generator are subjected to scene consistency checks. If the candidate abnormal weather supplement sample passes the scene consistency check, it is written into the abnormal weather supplement sample set; if the candidate abnormal weather supplement sample fails the scene consistency check, it is discarded and the next candidate abnormal weather supplement sample is generated.

7. The intelligent power load forecasting method for industrial and commercial users according to claim 1, characterized in that, The process involves training a long short-term memory network based on a scenario calibration sample library, generating a basic prediction result table, generating a weather deviation sample table based on the basic prediction result table, and generating a scenario calibration network package based on the basic prediction result table and the weather deviation sample table, including: Normal weather training samples are determined from the scene calibration sample library. Based on the normal weather training samples, a basic prediction input vector is constructed and a long short-term memory network is trained. Abnormal weather samples and the scene to be predicted are recorded in the scene calibration sample library and input into the trained long short-term memory network to obtain the basic prediction load value and write it into the basic prediction result table. The deviation extraction objects corresponding to the abnormal weather scenario conditions are determined from the basic prediction result table. The prediction deviation is calculated based on the deviation extraction objects, and a weather deviation sample table is generated according to the sequence number within the stage. The sequence number within the stage is established according to the chronological order of the 15-minute time points within the event stage in the weather scenario conditions. Read the record of the scene to be predicted from the basic prediction result table, find the single weather deviation sample that matches the weather scene conditions at the time point in the weather deviation sample table, generate a single scene calibration rule based on the single weather deviation sample, and write it into the scene calibration network packet.

8. A smart power load forecasting system for industrial and commercial users, used to implement the smart power load forecasting method for industrial and commercial users as described in any one of claims 1-7, characterized in that, The system includes a scenario construction module, a scenario calibration module, a deviation calibration module, and a load prediction module; The scenario construction module is used to generate aligned time load records and weather event fragment records based on the acquired raw electricity consumption and meteorological data, determine weather scenario conditions based on the aligned time load records and weather event fragment records, and generate a weather scenario sample set. The scene calibration module is used to determine abnormal weather scene conditions based on the weather scene sample set and generate sample gap records, generate an abnormal weather supplementary sample set based on the sample gap records, and generate a scene calibration sample library in combination with the weather scene conditions. The deviation calibration module is used to train a long short-term memory network based on a scenario calibration sample library, generate a basic prediction result table, generate a weather deviation sample table based on the basic prediction result table, and generate a scenario calibration network package based on the basic prediction result table and the weather deviation sample table. The load prediction module is used to generate calibration trigger records based on scenario calibration network packets, determine the post-calibration load prediction sequence based on the calibration trigger records, and generate the final load prediction result based on the post-calibration load prediction sequence and the calibration trigger records.

Images