Active power distribution network situation prediction method and system based on multiple attention mechanism

Abstract

The application discloses a method and system for active power distribution network situation prediction based on a multiple attention mechanism, and relates to the technical field of power systems.The method comprises: collecting multi-source data of bus layers, feeder layers and transformer area layers in a hierarchical manner, and dividing the data into base class data and new class data; obtaining a set of twin multiple attention situation feature extractors; matching historical difficult-to-classify samples based on the base class data and directionally fine-tuning a first extractor to obtain a first optimized extractor; using the first optimized extractor to identify features of the base class data and using a second extractor to identify features of the new class data, and inputting the identification results into a weighted decoupling module to obtain multi-layer situation prediction results.The application solves the technical problems of insufficient identification accuracy and low prediction reliability of existing active power distribution network situation prediction when processing multi-source multi-dimensional data and rare working conditions, and achieves the technical effects of improving situation prediction accuracy and reliability under complex working conditions through twin multiple attention mechanisms and difficult-to-classify sample optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to an active power distribution network situation prediction method and system based on a multiple attention mechanism. BACKGROUND

[0002] With the large-scale access of distributed power sources and the increasing complexity of user load characteristics, the operation state of the active power distribution network presents characteristics such as multi-source, volatility and uncertainty. Traditional situation prediction methods mostly rely on single feature extraction or experience-based model inference, and are difficult to effectively cope with the complex coupling relationship of multi-level and multi-type operation data, especially when dealing with boundary conditions and new rare conditions, the prediction results often have deviations or delays. At the same time, the existing methods lack flexible weight allocation and decoupling mechanisms among multi-dimensional operation indicators, resulting in insufficient accuracy and stability of the situation prediction results. SUMMARY

[0003] The present application provides an active power distribution network situation prediction method and system based on a multiple attention mechanism, which is used to solve the technical problems of insufficient recognition accuracy and low prediction reliability of the active power distribution network situation prediction in the prior art when dealing with multi-source and multi-dimensional data and rare conditions.

[0004] In a first aspect, the present application provides an active power distribution network situation prediction method based on a multiple attention mechanism, which comprises: performing hierarchical multi-source data collection on the bus layer, the feeder layer and the district layer of a target active power distribution network, and dividing the collection results according to a preset base class division rule to obtain a base class multi-source operation data sequence set and a new class multi-source operation data sequence set; obtaining a pre-constructed twin multiple attention situation feature extractor set, wherein the twin multiple attention situation feature extractor set includes a first multiple attention situation feature extractor and a second multiple attention situation feature extractor; matching a historical difficult sample multi-source operation data sequence set based on the base class multi-source operation data sequence set, and directionally fine-tuning the first multiple attention situation feature extractor based on the historical difficult sample multi-source operation data sequence set to obtain a first optimized multiple attention situation feature extractor; using the first optimized multiple attention situation feature extractor to perform situation feature recognition on the base class multi-source operation data sequence set, and using the second multiple attention situation feature extractor to perform situation feature recognition on the new class multi-source operation data sequence set, and transmitting the recognition results to a weighted decoupling module for situation prediction to obtain a multi-layer situation prediction result.

[0005] In a second aspect of the present application, a multi-attention mechanism-based active power distribution network situation prediction system is provided, which comprises: a data acquisition and division module, configured to perform hierarchical multi-source data acquisition on bus layers, feeder layers and transformer area layers of a target active power distribution network, and to divide the acquisition results according to a preset base class division rule to obtain a base class multi-source operation data sequence set and a new class multi-source operation data sequence set; a feature extractor acquisition module, configured to acquire a pre-constructed twin multi-attention situation feature extractor set, wherein the twin multi-attention situation feature extractor set comprises a first multi-attention situation feature extractor and a second multi-attention situation feature extractor; a extractor fine-tuning module, configured to match a historical difficult-to-divide sample multi-source operation data sequence set based on the base class multi-source operation data sequence set, and to perform directional fine-tuning on the first multi-attention situation feature extractor based on the historical difficult-to-divide sample multi-source operation data sequence set to obtain a first optimized multi-attention situation feature extractor; and a situation prediction module, configured to perform situation feature identification on the base class multi-source operation data sequence set using the first optimized multi-attention situation feature extractor, and to perform situation feature identification on the new class multi-source operation data sequence set using the second multi-attention situation feature extractor, and to transmit the identification results to a weighted decoupling module for situation prediction to obtain a multi-layer situation prediction result.

[0006] The one or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0007] The multi-attention mechanism-based active power distribution network situation prediction method and system provided in the present application relate to the technical field of power systems, construct a twin multi-attention situation feature extractor, jointly model feature attention and time attention, fine-tune the base class extractor in a directional manner in combination with historical difficult-to-divide samples, and introduce a weighted decoupling module to fuse the base class and new class identification results, thereby realizing classification processing and unified prediction of multi-level data of an active power distribution network, and solving the technical problems of insufficient identification accuracy and low prediction reliability of existing active power distribution network situation prediction when processing multi-source multi-dimensional data and rare working conditions, and achieving the technical effects of improving situation prediction accuracy and reliability under complex working conditions through twin multi-attention mechanism and difficult-to-divide sample optimization. BRIEF DESCRIPTION OF DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0009] Figure 1A flowchart of a method for active power distribution network situation prediction based on a multiple attention mechanism is provided in the embodiments of the present application.

[0010] Figure 2 A structural diagram of a system for active power distribution network situation prediction based on a multiple attention mechanism is provided in the embodiments of the present application.

[0011] Legend: data acquisition and division module 11, feature extractor acquisition module 12, extractor fine-tuning module 13, situation prediction module 14. DETAILED DESCRIPTION

[0012] The present application provides a method and system for active power distribution network situation prediction based on a multiple attention mechanism, which is used to solve the technical problems of insufficient recognition accuracy and low prediction reliability of the existing active power distribution network situation prediction in processing multi-source multi-dimensional data and rare working conditions.

[0013] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0014] It should be noted that the terms "first", "second", etc. in the specification and the above drawings of the present application are used to distinguish similar objects, and do not necessarily mean a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices.

[0015] Embodiment one, as shown in the present application, a method for active power distribution network situation prediction based on a multiple attention mechanism is provided, which comprises: Figure 1

[0016] P10: Hierarchical multi-source data acquisition is performed on the bus layer, feeder layer and district layer of the target active power distribution network, and the acquisition results are divided according to the preset base class division rule, to obtain a base class multi-source operation data sequence set and a new class multi-source operation data sequence set.

[0017] Further, the step P10 of the embodiments of the present application further comprises: ​

[0018] P11: extract a preset partition index set in the preset base class partition rule, wherein the preset partition index set includes occurrence frequency, duration, running index deviation, and external driving factor; P12: based on the preset partition index set, respectively perform index analysis on the collected result of the multi-source running data sequence set, and perform weighted analysis according to a preset weight, to obtain a multi-source running data base class reliability factor set; P13: extract a preset base class reliability factor threshold in the preset base class partition rule, divide the multi-source running data base class reliability factor set, and map and divide the multi-source running data sequence set according to the division result, to obtain a base class multi-source running data sequence set and a new class multi-source running data sequence set.

[0019] It should be understood that the hierarchical multi-source data collection is first carried out for the bus layer, the feeder layer, and the transformer area layer of the target active power distribution network. Due to significant differences in data characteristics of different levels, for example, the bus layer data mainly reflects global characteristics, the feeder layer data more reflects power flow distribution and load transfer, and the transformer area layer data represents power quality and load fluctuation on the user side, therefore, multiple different monitoring collection points need to be arranged at each level to ensure that the collected data can cover multi-dimensional information such as voltage, current, power, frequency, load curve, and distributed power output. Through this top-down layer-by-layer refinement collection method, the comprehensiveness and divisibility of the collected data can be ensured at the source, laying a solid foundation for subsequent data division.

[0020] Specifically, first, the index set required for partitioning needs to be extracted from the preset base class partition rule. The index set includes parameters such as occurrence frequency, duration, running index deviation, and external driving factor. The occurrence frequency is used to measure the probability and coverage of a certain class of working conditions in historical data, which can reflect whether the working condition belongs to the typical running state of the power distribution network; the duration is used to describe the stability of the working condition in the time dimension, for example, the peak-valley load state or the daytime mode of photovoltaic output, which often represents a stable working condition with a long duration, while load surge or short-time reverse power flow belongs to a working condition with a short duration; the running index deviation is used to measure the deviation degree of voltage, current, power factor, and other running parameters compared with the normal running interval, through the measurement of these deviation amounts, it can be judged whether the working condition has potential risks or abnormalities; the external driving factor mainly considers the influence of external conditions such as weather, season, and special events on the working condition, for example, seasonal load change, running state under extreme weather conditions, or fluctuation of electricity consumption behavior during holidays.

[0021] After the above division index set is extracted, it is necessary to carry out index analysis based on the collected multi-source running data sequence, and to carry out weighted processing on each index in combination with the preset weight system, so as to obtain a multi-source running data base class reliability factor set which can comprehensively reflect the typicality, stability and deviation degree of the working condition. The so-called base class reliability factor is a quantitative description of the possibility of a certain running data sequence belonging to a historical typical working condition. The numerical value can directly reflect the closeness between the data sequence and the typical working condition. Through the method of weighted analysis, not only the influence of the non-uniform dimension of different indexes can be eliminated, but also the role of key indicators can be highlighted according to the actual operation experience, and the rationality and accuracy of the division result can be improved.

[0022] Finally, after obtaining the base class reliability factor set, the base class reliability factor threshold set in the preset base class division rule needs to be further called. The system compares the reliability factor of each running data sequence with the threshold value. When the reliability factor is higher than the threshold value, it indicates that the sequence has high typicality and stability, and should be classified as base class data; when the reliability factor is lower than the threshold value, it indicates that the sequence belongs to a very rare or never appeared working condition, and should be classified as new class data. Through this mapping division method based on reliability factor, the base class multi-source running data sequence set and the new class multi-source running data sequence set can be accurately and quickly distinguished in a large and complex multi-level running data, providing high-quality input data support for subsequent multi-attention feature extraction and situation prediction.

[0023] P20: Obtain a pre-constructed twin multi-attention situation feature extractor set, wherein the twin multi-attention situation feature extractor set includes a first multi-attention situation feature extractor and a second multi-attention situation feature extractor.

[0024] Further, the step P20 of the embodiments of the present application further includes:

[0025] P21: Obtain a set of historical multi-source running data sequences and a set of historical multi-layer situation features as a training data set; P22: train an LSTM time series model with fusion feature attention layer and time series attention layer using the training data set to obtain an initial multi-attention situation feature extractor; P23: duplicate the initial multi-attention situation feature extractor to obtain a first multi-attention situation feature extractor and a second multi-attention situation feature extractor, and obtain the twin multi-attention situation feature extractor set.

[0026] Optionally, a pre-constructed twin multiple attention situation feature extractor set is obtained, which consists of two identical multiple attention situation feature extractors, named the first multiple attention situation feature extractor and the second multiple attention situation feature extractor. Both are based on historical situation features and belong to situation feature recognition models with feature attention mechanism and time attention mechanism.

[0027] First, a historical multi-source operation data sequence set and a historical multi-layer situation feature set are obtained as the training data set required for model training. The historical multi-source operation data sequence set covers long-term operation data at the busbar layer, feeder layer and transformer area layer, while the historical multi-layer situation feature set includes multi-dimensional situation information such as voltage level, load state, power factor, distributed power output feature and power quality index. This combined training set can provide sufficient feature support for model learning.

[0028] Next, the long short-term memory network (LSTM) time series model with fusion feature attention layer and time sequence attention layer is trained using the training data set. In this process, the model will learn how to extract key features from multi-source data and identify the pattern of these features over time. The feature attention layer is used to assign different weights to different feature dimensions when multi-source data is input, highlighting the contribution of key features in situation recognition. The time sequence attention layer is used to distinguish the importance of different historical time points in the time dimension, so that the model can effectively capture the dynamic characteristics of the working condition evolution. Through the introduction of this double attention mechanism, the model can not only identify the relevance of different dimensional data, but also grasp the key nodes in the time series, thereby achieving more accurate situation feature extraction.

[0029] After obtaining the initial multiple attention situation feature extractor, the extractor is duplicated to obtain the first multiple attention situation feature extractor and the second multiple attention situation feature extractor. These two extractors are the same in structure and parameters, both of which are situation feature recognition models constructed based on historical situation features. By combining them, a twin multiple attention situation feature extractor set is finally formed. This set can ensure that the processing of base class and new class data follows the same feature extraction logic, while also being able to carry out different data prediction tasks, thereby achieving efficient modeling and recognition of multi-level situations in active distribution networks.

[0030] P30: Based on the base class multi-source operation data sequence set, match the historical difficult sample multi-source operation data sequence set, based on the historical difficult sample multi-source operation data sequence set, directionally fine-tune the first multiple attention situation feature extractor, and obtain the first optimized multiple attention situation feature extractor.

[0031] Further, based on matching the base class multi-source operation data sequence set with the historical difficult-to-distinguish sample multi-source operation data sequence set, the embodiment of the present application step P30 further comprises:

[0032] P31: performing same-class aggregation on the base class multi-source operation data sequence set to obtain a plurality of same-class aggregation results; P32: performing mean shift screening on the plurality of same-class aggregation results to determine a plurality of centralized base class multi-source operation data sequences; and P33: performing historical difficult-to-distinguish sample matching based on the plurality of centralized base class multi-source operation data sequences to obtain a historical difficult-to-distinguish sample multi-source operation data sequence set.

[0033] Specifically, the base class multi-source operation data sequence set is used to match and extract historical difficult-to-distinguish samples, and on this basis, the directionality of the first multi-attention situation feature extractor is fine-tuned, so as to obtain the first optimized multi-attention situation feature extractor. The difficult-to-distinguish sample refers to a sample data whose feature performance is highly similar to other categories and is easily confused in the situation feature recognition process, for example, a boundary working condition between light voltage over-limit and normal voltage fluctuation, or a critical working condition between slight line overload and normal load peak. By introducing the difficult-to-distinguish sample for targeted optimization, the discrimination ability and generalization performance of the model under the boundary situation can be significantly improved.

[0034] First, the base class multi-source operation data sequence set is aggregated in the same class. Same-class aggregation refers to aggregating data sequences with similar features into a class. Specifically, according to the typical operation features defined in the base class division rule, data samples of the same class are aggregated into the corresponding category, thereby forming a plurality of same-class aggregation results. This process can reduce the discreteness and redundancy of data, so that the samples within different categories are more compact in structure, which is beneficial to subsequent sample screening.

[0035] Next, the plurality of same-class aggregation results are screened using the mean shift method to obtain a plurality of centralized base class multi-source operation data sequences. Mean shift is a density-based clustering algorithm that identifies high-density regions in data by calculating the local mean and variance of data points. By iteratively calculating the density distribution center of the sample, the samples close to the center are gradually converged to the high-density region. Application of this method can effectively eliminate discrete samples and noise points, thereby extracting more representative centralized sample sequences in each class of working conditions.

[0036] Next, based on these centralized base class multi-source operation data sequences, historical difficult-to-classify sample matching is performed. The historical difficult-to-classify sample refers to a sample data that is easily confused in the situation feature extraction result. By calculating the feature distance or similarity index between samples, samples that are located in the boundary region between classes and have high similarity with multiple classes can be identified and divided into a historical difficult-to-classify sample multi-source operation data sequence set. This set is input into the first multi-attention situation feature extractor as an important optimization data set in subsequent training.

[0037] Finally, the first multi-attention situation feature extractor is directionally fine-tuned using the historical difficult-to-classify sample multi-source operation data sequence set described above. Directional fine-tuning refers to a small range update of the weight parameters of the model under specific class boundary conditions based on the feature distribution of the difficult-to-classify sample, rather than complete retraining of the entire model. In this way, the recognition accuracy of the model for critical operating conditions can be significantly enhanced, avoiding misjudgment or omission in complex and variable power grid operating scenarios, thereby obtaining an optimized first multi-attention situation feature extractor that provides more reliable feature extraction capability for subsequent situation prediction.

[0038] Further, the step P33 of the embodiment of the present application further includes:

[0039] P33-1: indexing the plurality of centralized base class multi-source operation data sequences, performing historical situation feature error recognition result retrieval, and obtaining a plurality of historical situation feature error recognition result sets; P33-2: extracting multi-source operation data sequences in the plurality of historical situation feature error recognition result sets to obtain the historical difficult-to-classify sample multi-source operation data sequence set.

[0040] Optionally, the process of historical difficult-to-classify sample matching is further refined to improve the efficiency and pertinence of the acquisition of historical difficult-to-classify samples.

[0041] Specifically, first, the aforementioned plurality of centralized base class multi-source operation data sequences are used as indexes to perform retrieval in the error recognition records formed in the historical situation feature recognition process. The historical situation feature error recognition result refers to a sample record whose model prediction result is inconsistent with the actual operating condition class when the previous model recognizes the multi-source operation data. For example, when a certain operation data sequence belongs to a voltage mild excursion operating condition in the actual label, but the model identifies it as normal voltage fluctuation, this data sample constitutes an error recognition result. By using the centralized base class data as a retrieval index, the search range can be narrowed down, and only the error samples highly related to the current base class are extracted, avoiding the introduction of irrelevant data noise.

[0042] Next, the multi-source running data sequences in the retrieved historical situation feature error recognition result set are extracted one by one and summarized to form a historical difficult-to-distinguish sample multi-source running data sequence set. Since these data samples themselves are derived from the historical recognition process in the situation that is easy to confuse and misjudge, their features naturally have high similarity and boundary fuzziness, and can provide the most representative training materials for the optimization of the recognition performance of the model under the critical situation condition. By inputting the set into the first multi-attention situation feature extractor for directional fine-tuning, the robustness and sensitivity of the model under complex operating conditions can be effectively enhanced, thereby improving the reliability and accuracy of the overall situation recognition and prediction.

[0043] Specifically, the fine-tuning process can be: inputting the historical difficult-to-distinguish sample multi-source running data sequence set as additional training data into the first multi-attention situation feature extractor. Adjust the parameters of the model, especially those related to error recognition, to reduce classification errors on these samples. Evaluate the performance of the fine-tuned model through cross-validation and other methods to ensure that the fine-tuning process is effective.

[0044] Further, based on the historical difficult-to-distinguish sample multi-source running data sequence set, the first multi-attention situation feature extractor is directionally fine-tuned to obtain a first optimized multi-attention situation feature extractor. The embodiment P30 of the present application further comprises:

[0045] P34: Obtain the initial hyperparameters of the first multi-attention situation feature extractor, and randomly adjust the initial hyperparameters according to a preset adjustment method to obtain an adjusted hyperparameter solution space; P35: Obtain a historical difficult-to-distinguish sample multi-layer situation feature set corresponding to the historical difficult-to-distinguish sample multi-source running data sequence set, and identify the historical difficult-to-distinguish sample multi-layer situation feature set to construct a directional fine-tuning validation set; P36: Based on the directional fine-tuning validation set and the adjusted hyperparameter solution space, the first multi-attention situation feature extractor is directionally fine-tuned to obtain the first optimized multi-attention situation feature extractor.

[0046] In a possible embodiment of the present application, for the optimization process of the first multi-attention situation feature extractor, a hyperparameter adjustment and validation set construction mechanism can be further introduced to ensure that the directional fine-tuning can achieve higher recognition accuracy and generalization performance under difficult-to-distinguish sample conditions.

[0047] First, the initial hyperparameters of the first multi-attention situation feature extractor are obtained. Hyperparameters refer to parameters that are not learned through backpropagation during model training, but are set externally, such as learning rate, regularization coefficient, attention weight dimension, time window length, and LSTM hidden layer unit number. To avoid the model falling into a local optimal solution under fixed hyperparameters, the initial hyperparameters need to be randomly disturbed or changed according to the preset adjustment method, thereby generating a wider adjustment hyperparameter solution space. This solution space contains a variety of possible hyperparameter combinations, which can provide rich choices for subsequent fine-tuning. In this way, a variety of candidate configurations can be provided for the subsequent fine-tuning process, improving the adaptability of the model under complex working conditions.

[0048] Subsequently, based on the historical difficult-to-distinguish sample multi-source operating data sequence set, the corresponding historical difficult-to-distinguish sample multi-layer situation feature set is further obtained. This set mainly includes node voltage distribution, branch power flow state, voltage fluctuation amplitude, line overload coefficient and other feature parameters directly related to system situation. On this basis, the feature set is labeled to generate a directional fine-tuning verification set. Each data in the verification set has a clear working condition label and can be used as a standard sample to evaluate the recognition ability of the model, ensuring that the fine-tuning process is measurable and comparable.

[0049] Next, based on the directional fine-tuning verification set and the adjustment hyperparameter solution space, the first multi-attention situation feature extractor is directionally fine-tuned. Specifically, different parameter combinations in the adjustment hyperparameter solution space are used to evaluate the recognition performance of the model on the verification set, and through iterative optimization and selection, the optimal hyperparameter configuration is gradually approached, and the weight parameters of the model under difficult-to-distinguish sample conditions are continuously corrected in the process.

[0050] For example, during fine-tuning, a set of hyperparameters in the adjustment hyperparameter solution space is randomly selected, the first multi-attention situation feature extractor is configured with this set of hyperparameters, and training and testing are performed on the directional fine-tuning verification set. The performance of the model on the verification set is evaluated, and performance indicators such as accuracy and recall rate are recorded. Repeat this process to traverse all possible combinations in the adjustment hyperparameter solution space. Select the best hyperparameter combination and apply it to the first multi-attention situation feature extractor to obtain the first optimized multi-attention situation feature extractor. This first optimized multi-attention situation feature extractor has higher accuracy and stability when processing critical working conditions and confusing samples, thereby providing more reliable feature extraction capability for subsequent multi-layer situation prediction.

[0051] Further, the step P36 of the embodiment of the present application further comprises:

[0052] P36-1: randomly select a plurality of adjustment hyperparameters from the adjustment hyperparameter solution space, use the plurality of adjustment hyperparameters to fine-tune the parameters of the first multi-attention situation feature extractor, and obtain a plurality of fine-tuned first multi-attention situation feature extractors; P36-2: based on the directional fine-tuning verification set, verify the adjustment hyperparameter reliability of the plurality of fine-tuned first multi-attention situation feature extractors, and obtain a plurality of adjustment hyperparameter reliability coefficients; P36-3: take the adjustment hyperparameter corresponding to the maximum value in the plurality of adjustment hyperparameter reliability coefficients as the fine-tuning direction, and directionally fine-tune the plurality of adjustment hyperparameters in the adjustment hyperparameter solution space to obtain a plurality of updated adjustment hyperparameters; P36-4: fine-tune the parameters of the first multi-attention situation feature extractor based on the plurality of updated adjustment hyperparameters again, and call the directional fine-tuning verification set to verify the updated adjustment hyperparameter reliability of the plurality of fine-tuned first multi-attention situation feature extractors, and obtain a plurality of updated adjustment hyperparameter reliability coefficients; P36-5: determine whether there is an updated adjustment hyperparameter reliability coefficient greater than or equal to the adjustment hyperparameter reliability coefficient corresponding to the fine-tuning direction in the plurality of updated adjustment hyperparameter reliability coefficients, if so, update the fine-tuning direction based on the updated adjustment hyperparameter corresponding to the maximum value in the plurality of updated adjustment hyperparameter reliability coefficients, and continue to directionally fine-tune the plurality of updated adjustment hyperparameters, and count the number of updates; P36-6: when the number of updates is greater than or equal to a preset number of update thresholds, determine whether the maximum value of the updated hyperparameter reliability coefficient in the update process is greater than or equal to a preset hyperparameter reliability coefficient, if so, take the updated adjustment hyperparameter corresponding to the maximum value of the updated hyperparameter reliability coefficient in the update process as the target adjustment hyperparameter, and adjust the parameters of the first multi-attention situation feature extractor based on the target adjustment hyperparameter to obtain a first optimized multi-attention situation feature extractor; P36-7: if the maximum value of the updated hyperparameter reliability coefficient in the update process is less than the preset hyperparameter reliability coefficient, obtain a re-selection instruction, reselect a plurality of adjustment hyperparameters in the adjustment hyperparameter solution space according to the re-selection instruction, and directionally fine-tune the first multi-attention situation feature extractor again to obtain the first optimized multi-attention situation feature extractor.

[0053] Specifically, the fine-tuning process of the first multi-attention situation feature extractor can be further refined to ensure the optimal model performance.

[0054] First, a plurality of candidate hyperparameter combinations are randomly selected from the adjustment hyperparameter solution space, and these combinations are used to fine-tune the parameters of the first multi-attention situation feature extractor, respectively, to obtain a plurality of fine-tuned first multi-attention situation feature extractors. This step can ensure that the model can be trained under different hyperparameter conditions, increasing the diversity of solution space exploration.

[0055] Next, the multiple fine-tuned models are verified one by one based on the directional fine-tuning verification set, and corresponding multiple adjusted hyperparameter reliability coefficients are obtained. The reliability coefficient is a comprehensive evaluation index for measuring the model recognition accuracy, stability and sensitivity to difficult samples under the hyperparameter combination on the verification set. The higher the value, the stronger the effectiveness of the hyperparameter combination in directional fine-tuning.

[0056] Next, the adjusted hyperparameters corresponding to the maximum value in the multiple reliability coefficients are taken as the fine-tuning direction, and based on this, further directional fine-tuning is performed on the multiple adjusted hyperparameters in the adjusted hyperparameter solution space, and multiple updated adjusted hyperparameters are obtained. Through this directional convergence strategy, a more optimal hyperparameter region can be gradually locked, avoiding ineffective search.

[0057] Then, the first multiple attention situation feature extractor is fine-tuned again based on the multiple updated adjusted hyperparameters, and the directional fine-tuning verification set is called for verification, and multiple updated adjusted hyperparameter reliability coefficients are obtained. Through this process, fine-tuning verification can be performed on the basis of preliminary screening results, gradually improving the performance of the model on the verification set.

[0058] Further, it is determined whether there is a case where the multiple updated adjusted hyperparameter reliability coefficients are greater than or equal to the reliability coefficient corresponding to the fine-tuning direction. If so, the updated adjusted hyperparameters corresponding to the maximum value of the reliability coefficient are taken as the new fine-tuning direction, and iterative directional fine-tuning is continued in the solution space, while recording the number of updates. Through this way, the fine-tuning direction can be constantly adjusted, so that the model gradually converges to the global optimal solution.

[0059] When the number of updates reaches or exceeds the preset number of updates threshold, it is further determined whether the maximum value of the reliability coefficient in the updating process is greater than or equal to the preset hyperparameter reliability coefficient. If the condition is met, the updated adjusted hyperparameters corresponding to the maximum value are determined as the target adjusted hyperparameters, and the first multiple attention situation feature extractor is finally adjusted based on the target hyperparameters to obtain an optimized first multiple attention situation feature extractor.

[0060] If all the maximum values of the reliability coefficients in the updating process are lower than the preset hyperparameter reliability coefficient, it means that the current solution space exploration has failed to converge to an effective hyperparameter combination. In this case, the re-selection instruction is received, and multiple hyperparameter combinations are randomly selected in the adjusted hyperparameter solution space according to the instruction, and the fine-tuning and verification process is repeated until the first optimized multiple attention situation feature extractor is obtained.

[0061] Through the above-mentioned cycle process of step-by-step screening, directional fine-tuning and reliability verification, the efficiency and accuracy of hyperparameter optimization can be significantly improved, and the first optimized multi-attention situation feature extractor finally obtained can have better robustness and generalization ability when processing historical difficult samples.

[0062] Further, the step P36-2 of the embodiment of the present application further comprises:

[0063] P36-21: using the multiple fine-tuned first multi-attention situation feature extractors to perform situation feature extraction on the historical difficult sample multi-source operation data sequence set in the directional fine-tuning verification set respectively, and obtaining multiple fine-tuned output multi-layer situation feature sets; P36-22: comparing and verifying the multiple fine-tuned output multi-layer situation feature sets with the historical difficult sample multi-layer situation feature sets identified in the directional fine-tuning verification set respectively, and obtaining multiple adjusted hyperparameter reliability coefficients.

[0064] Optionally, the directional fine-tuning process can be further refined to ensure that the obtained optimized model can provide reliable performance in actual application.

[0065] First, using the multiple fine-tuned first multi-attention situation feature extractors obtained in the previous step, the historical difficult sample multi-source operation data sequence set contained in the directional fine-tuning verification set is processed for feature extraction. Through the joint action of the feature attention mechanism and the time attention mechanism, the model can extract multi-dimensional and time-sequenced situation feature information from the input multi-source operation data. Finally, for each fine-tuned extractor, a corresponding fine-tuned output multi-layer situation feature set can be obtained, which reflects the feature recognition ability of the model for difficult samples under different hyperparameter combinations.

[0066] Next, the multiple fine-tuned output multi-layer situation feature sets are compared with the historical difficult sample multi-layer situation feature sets identified in the directional fine-tuning verification set one by one. The comparison process can be based on feature similarity measurement, classification accuracy, bias sum of squares, etc., to ensure that the verification result is quantitative and reproducible. Through the comparison, the identification reliability and accuracy performance of the model for difficult samples under certain hyperparameters can be effectively reflected. After the comparison is completed, the system assigns an adjusted hyperparameter reliability coefficient to each group of hyperparameters, which quantifies the performance of the current hyperparameters on the verification set and provides an objective decision basis for subsequent directional fine-tuning. Through this refined fine-tuning and verification process, the obtained optimized model can not only perform well in theory, but also provide stable and reliable performance in actual application.

[0067] P40: The first optimized multi-attention situation feature extractor is used to identify the situation features of the base class multi-source operation data sequence set, and the second multi-attention situation feature extractor is used to identify the situation features of the new class multi-source operation data sequence set. The identification results are transmitted to the weighted decoupling module for situation prediction to obtain multi-layer situation prediction results.

[0068] The weighted decoupling module is embedded with active power distribution network situation factors, including node voltage out-of-limit margin, branch load severity, voltage fluctuation coefficient, and current fluctuation coefficient.

[0069] It should be understood that the process of using the optimized multi-attention situation feature extractor to identify the situation features of different types of data sequence sets and using the identification results for situation prediction. First, the first optimized multi-attention situation feature extractor with directional fine-tuning and hyperparameter optimization is used to identify the situation features of the base class multi-source operation data sequence set. Since the extractor has been optimized for historical typical working conditions and difficult samples, it has high stability and accuracy when identifying base class data, and can fully extract key situation features under base class working conditions.

[0070] At the same time, the second multi-attention situation feature extractor with the same structure is used to identify the situation features of the new class multi-source operation data sequence set. Since the new class data appears less frequently or not at all in historical records, the twin model without base class optimization is directly used for identification, avoiding overfitting or bias of the model on new class data. Through this dual model parallel mode, common working conditions and rare working conditions can be covered, ensuring the comprehensiveness of the feature identification results.

[0071] After the above feature extraction, the identification results from the base class and the new class are transmitted to the weighted decoupling module at the same time. The core function of this module is to fuse and decompose the feature results from different sources to realize cross-level and cross-category situation prediction. In the process of weighted decoupling, the module is internally pre-embedded with multi-dimensional situation factors closely related to the operation safety of the active power distribution network, including node voltage out-of-limit margin, branch load severity, voltage fluctuation coefficient, and current fluctuation coefficient. Among them, the node voltage out-of-limit margin is used to represent the degree of voltage level deviation from the allowed interval, the branch load severity is used to describe the overload risk of lines or devices, and the voltage fluctuation coefficient and the current fluctuation coefficient are used to measure the change amplitude of short-term dynamic stability in system operation.

[0072] By introducing the above-mentioned situation factors into the weighted decoupling module, the indicators most directly related to the safety and stability of power grid operation can be extracted from the multi-level and multi-category recognition results, and the contribution of different situation characteristics can be balanced through weighted calculation. At the same time, the decoupling mechanism can effectively distinguish the feature redundancy and coupling effect between different levels, so that the prediction results have both global consistency and level difference. Finally, the multi-level situation prediction results output by the system can comprehensively reflect the operation state of the active distribution network at the bus layer, feeder layer and transformer area layer, and provide a scientific basis for operation dispatching and risk warning.

[0073] In summary, the embodiments of the present application have at least the following technical effects:

[0074] The present application realizes comprehensive coverage of data sources by collecting and dividing multi-source operation data at the bus layer, feeder layer and transformer area layer; realizes differentiated processing of different working conditions by using the twin multiple attention situation feature extractor to identify features of base class and new class data respectively; improves the recognition accuracy of the model in critical working conditions by introducing historical difficult-to-divide samples to fine-tune the direction of the first multiple attention situation feature extractor; embeds situation factors such as voltage out-of-limit margin, load severity and fluctuation coefficient to make the prediction method adaptable to different types and levels of distribution network data, and improve the generality and adaptability of the model; and by predicting and identifying potential power grid faults in advance, it is helpful to take preventive measures, thereby improving the stability and reliability of the power grid.

[0075] The technical effects of improving the situation prediction accuracy and reliability under complex working conditions through the twin multiple attention mechanism and difficult-to-divide sample optimization are achieved.

[0076] Embodiment two, based on the same inventive concept as the multiple attention mechanism-based active distribution network situation prediction method in the preceding embodiments, as shown in Figure 2 The present application provides a multiple attention mechanism-based active distribution network situation prediction system, and the system and method embodiments in the present application are based on the same inventive concept. The system comprises:

[0077] The data collection and division module 11 is used for collecting multi-source data at the bus layer, feeder layer and transformer area layer of the target active distribution network in a hierarchical manner, and dividing the collected results according to a preset base class division rule to obtain a base class multi-source operation data sequence set and a new class multi-source operation data sequence set.

[0078] The feature extractor acquisition module 12 is used for acquiring a set of pre-constructed twin multiple attention situation feature extractors, wherein the set of twin multiple attention situation feature extractors includes a first multiple attention situation feature extractor and a second multiple attention situation feature extractor.

[0079] The extractor fine-tuning module 13 is configured to match the historical difficult-to-classify sample multi-source operation data sequence set based on the base class multi-source operation data sequence set, fine-tune the first multi-attention situation feature extractor based on the historical difficult-to-classify sample multi-source operation data sequence set, and obtain a first optimized multi-attention situation feature extractor.

[0080] The situation prediction module 14 is configured to perform situation feature identification on the base class multi-source operation data sequence set by using the first optimized multi-attention situation feature extractor, perform situation feature identification on the new class multi-source operation data sequence set by using the second multi-attention situation feature extractor, and transmit the identification results to the weighted decoupling module for situation prediction to obtain a multi-layer situation prediction result.

[0081] Further, the data collection and division module 11 is further configured to perform the following steps:

[0082] extracting a preset division index set in the preset base class division rule, wherein the preset division index set includes occurrence frequency, duration, operation index deviation, and external driving factor; performing index analysis on the multi-source operation data sequence set in the collection result based on the preset division index set, and performing weighted analysis according to a preset weight to obtain a multi-source operation data base class reliability factor set; extracting a preset base class reliability factor threshold in the preset base class division rule, dividing the multi-source operation data base class reliability factor set, and mapping and dividing the multi-source operation data sequence set according to the division result to obtain a base class multi-source operation data sequence set and a new class multi-source operation data sequence set.

[0083] Further, the feature extractor acquisition module 12 is further configured to perform the following steps:

[0084] acquiring a historical multi-source operation data sequence set and a historical multi-layer situation feature set as a training data set; training an LSTM time series model with a fusion feature attention layer and a time series attention layer by using the training data set to obtain an initial multi-attention situation feature extractor; performing twin replication on the initial multi-attention situation feature extractor to obtain a first multi-attention situation feature extractor and a second multi-attention situation feature extractor, and collecting the twin multi-attention situation feature extractor set.

[0085] Further, the extractor fine-tuning module 13 is further configured to perform the following steps:

[0086] The same kind of aggregation is performed on the base class multi-source operation data sequence set, a plurality of same kind of aggregation results are obtained, mean shift screening is performed on the plurality of same kind of aggregation results, a plurality of centralized base class multi-source operation data sequences are determined, and history difficult-to-distinguish sample multi-source operation data sequence set is obtained based on the plurality of centralized base class multi-source operation data sequences.

[0087] Further, the extractor fine-tuning module 13 is further used to perform the following steps:

[0088] With the plurality of centralized base class multi-source operation data sequences as indexes, history situation feature error recognition result retrieval is performed, a plurality of history situation feature error recognition result sets are obtained, and the multi-source operation data sequences in the plurality of history situation feature error recognition result sets are extracted to obtain the history difficult-to-distinguish sample multi-source operation data sequence set.

[0089] Further, the extractor fine-tuning module 13 is further used to perform the following steps:

[0090] An initial hyperparameter of the first multi-attention situation feature extractor is obtained, and the initial hyperparameter is randomly adjusted according to a preset adjustment mode to obtain an adjustment hyperparameter solution space; a history difficult-to-distinguish sample multi-layer situation feature set corresponding to the history difficult-to-distinguish sample multi-source operation data sequence set is obtained, and the history difficult-to-distinguish sample multi-layer situation feature set is labeled to construct a directional fine-tuning verification set, and based on the directional fine-tuning verification set and the adjustment hyperparameter solution space, the first multi-attention situation feature extractor is directionally fine-tuned to obtain the first optimized multi-attention situation feature extractor.

[0091] Further, the extractor fine-tuning module 13 is further used to perform the following steps:

[0092] randomly select a plurality of adjustment hyperparameters from the adjustment hyperparameter solution space, use the plurality of adjustment hyperparameters to fine-tune the parameters of the first multiple attention situation feature extractor, and obtain a plurality of fine-tuned first multiple attention situation feature extractors; based on the directional fine-tuning verification set, verify the reliability of the adjustment hyperparameters of the plurality of fine-tuned first multiple attention situation feature extractors, and obtain a plurality of adjustment hyperparameter reliability coefficients; take the adjustment hyperparameter corresponding to the maximum value in the plurality of adjustment hyperparameter reliability coefficients as the fine-tuning direction, and directionally fine-tune the plurality of adjustment hyperparameters in the adjustment hyperparameter solution space to obtain a plurality of updated adjustment hyperparameters; again fine-tune the parameters of the first multiple attention situation feature extractor based on the plurality of updated adjustment hyperparameters, and call the directional fine-tuning verification set to verify the reliability of the updated adjustment hyperparameters of the plurality of fine-tuned first multiple attention situation feature extractors, and obtain a plurality of updated adjustment hyperparameter reliability coefficients; determine whether there is an updated adjustment hyperparameter reliability coefficient greater than or equal to the adjustment hyperparameter reliability coefficient corresponding to the fine-tuning direction in the plurality of updated adjustment hyperparameter reliability coefficients, if yes, update the fine-tuning direction based on the updated adjustment hyperparameter corresponding to the maximum value in the plurality of updated adjustment hyperparameter reliability coefficients, and continue to directionally fine-tune the plurality of updated adjustment hyperparameters, and count the number of updates; when the number of updates is greater than or equal to a preset update number threshold, determine whether the maximum value of the updated hyperparameter reliability coefficient in the update process is greater than or equal to a preset hyperparameter reliability coefficient, if yes, take the updated adjustment hyperparameter corresponding to the maximum value of the updated hyperparameter reliability coefficient in the update process as the target adjustment hyperparameter, and fine-tune the parameters of the first multiple attention situation feature extractor based on the target adjustment hyperparameter to obtain a first optimized multiple attention situation feature extractor; if the maximum value of the updated hyperparameter reliability coefficient in the update process is less than the preset hyperparameter reliability coefficient, obtain a re-selection instruction, randomly select a plurality of adjustment hyperparameters in the adjustment hyperparameter solution space according to the re-selection instruction, and directionally fine-tune the first multiple attention situation feature extractor again to obtain the first optimized multiple attention situation feature extractor.

[0093] Further, the extractor fine-tuning module 13 is further configured to perform the following steps:

[0094] The plurality of fine-tuned first multiple attention situation feature extractors are used to extract situation features from the historical difficult-to-distinguish sample multi-source operation data sequence set in the directional fine-tuning verification set, and a plurality of fine-tuned output multi-layer situation feature sets are obtained; the plurality of fine-tuned output multi-layer situation feature sets are compared and verified with the historical difficult-to-distinguish sample multi-layer situation feature set identified in the directional fine-tuning verification set, and a plurality of adjustment hyperparameter reliability coefficients are obtained.

[0095] Further, in the situation prediction module 14:

[0096] The weighted decoupling module is embedded with an active power distribution network situation factor, wherein the active power distribution network situation factor comprises a node voltage out-of-limit margin, a branch load severity, a voltage fluctuation coefficient and a current fluctuation coefficient.

[0097] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the above describes a specific embodiment of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0098] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0099] The present application and the drawings are only exemplary descriptions of the present application, and are considered to cover any and all modifications, changes, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the scope of the present application. Thus, if these modifications and changes of the present application belong to the scope of the present application and its equivalent technology, the present application intends to include these modifications and changes.

Claims

1. A method for active power distribution network situation prediction based on multiple attention mechanisms, characterized in that, The method comprises: The bus layer, feeder layer and district layer of the target active power distribution network are subjected to hierarchical multi-source data collection, and the collection results are divided according to a preset base class division rule, to obtain a base class multi-source operation data sequence set and a new class multi-source operation data sequence set; Obtain a pre-constructed twin multi-attention situation feature extractor set, wherein the twin multi-attention situation feature extractor set comprises a first multi-attention situation feature extractor and a second multi-attention situation feature extractor; Match the base class multi-source operation data sequence set with a historical difficult-to-divide sample multi-source operation data sequence set, and directionally fine-tune the first multi-attention situation feature extractor based on the historical difficult-to-divide sample multi-source operation data sequence set, to obtain a first optimized multi-attention situation feature extractor; Perform situation feature recognition on the base class multi-source operation data sequence set by using the first optimized multi-attention situation feature extractor, and perform situation feature recognition on the new class multi-source operation data sequence set by using the second multi-attention situation feature extractor, and transmit the recognition results to a weighted decoupling module for situation prediction, to obtain a multi-layer situation prediction result.

2. The active power distribution grid situation prediction method based on multiple attention mechanisms according to claim 1, characterized in that, The bus layer, feeder layer and district layer of the target active power distribution network are subjected to hierarchical multi-source data collection, and the collection results are divided according to a preset base class division rule, to obtain a base class multi-source operation data sequence set and a new class multi-source operation data sequence set, comprising: Extract a preset division index set in the preset base class division rule, wherein the preset division index set comprises occurrence frequency, duration, operation index deviation and external driving factor; Perform index analysis on the multi-source operation data sequence set in the collection results based on the preset division index set, and perform weighted analysis according to a preset weight, to obtain a multi-source operation data base class reliability factor set; Extract a preset base class reliability factor threshold in the preset base class division rule, divide the multi-source operation data base class reliability factor set, and map and divide the multi-source operation data sequence set according to the division result, to obtain the base class multi-source operation data sequence set and the new class multi-source operation data sequence set. 3.The active power distribution network situation prediction method based on multiple attention mechanisms of claim 1, wherein, Obtain a pre-constructed twin multi-attention situation feature extractor set, wherein the twin multi-attention situation feature extractor set comprises a first multi-attention situation feature extractor and a second multi-attention situation feature extractor, comprising: Obtain a historical multi-source operation data sequence set and a historical multi-layer situation feature set as a training data set; Train an LSTM time series model with a fusion feature attention layer and a time series attention layer by using the training data set, to obtain an initial multi-attention situation feature extractor; Duplicate the initial multi-attention situation feature extractor to obtain the first multi-attention situation feature extractor and the second multi-attention situation feature extractor, and obtain the twin multi-attention situation feature extractor set by summarizing.

4. The active power distribution grid situation prediction method based on multiple attention mechanisms according to claim 2, characterized in that, Match the base class multi-source operation data sequence set with a historical difficult-to-divide sample multi-source operation data sequence set, comprising: The same kind of aggregation is performed on the base class multi-source operation data sequence set, and a plurality of same kind of aggregation results are obtained; The plurality of same kind of aggregation results are subjected to mean shift screening to determine a plurality of centralized base class multi-source operation data sequences; Based on the plurality of centralized base class multi-source operation data sequences, historical difficult-to-distinguish sample matching is performed to obtain a historical difficult-to-distinguish sample multi-source operation data sequence set.

5. The active power distribution grid situation prediction method based on multiple attention mechanisms according to claim 4, characterized in that, Based on the plurality of centralized base class multi-source operation data sequences, historical difficult-to-distinguish sample matching is performed to obtain a historical difficult-to-distinguish sample multi-source operation data sequence set. The plurality of historical situation feature error recognition result sets are subjected to historical situation feature error recognition result retrieval to obtain a plurality of historical situation feature error recognition result sets; The plurality of historical situation feature error recognition result sets are subjected to historical situation feature error recognition result retrieval to obtain a plurality of historical situation feature error recognition result sets; 6. The active power distribution grid situation prediction method based on multiple attention mechanisms according to claim 2, characterized in that, Based on the historical difficult-to-distinguish sample multi-source operation data sequence set, the first multi-attention situation feature extractor is directionally fine-tuned to obtain a first optimized multi-attention situation feature extractor, including: An initial hyperparameter of the first multi-attention situation feature extractor is obtained, and the initial hyperparameter is randomly adjusted according to a preset adjustment mode to obtain an adjustment hyperparameter solution space; A historical difficult-to-distinguish sample multi-layer situation feature set corresponding to the historical difficult-to-distinguish sample multi-source operation data sequence set is obtained, and the historical difficult-to-distinguish sample multi-layer situation feature set is labeled to construct a direction fine-tuning verification set, Based on the direction fine-tuning verification set and the adjustment hyperparameter solution space, the first multi-attention situation feature extractor is directionally fine-tuned to obtain the first optimized multi-attention situation feature extractor.

7. The active power distribution grid situation prediction method based on multiple attention mechanisms according to claim 6, characterized in that, Based on the direction fine-tuning verification set and the adjustment hyperparameter solution space, the first multi-attention situation feature extractor is directionally fine-tuned to obtain the first optimized multi-attention situation feature extractor, including: A plurality of adjustment hyperparameters are randomly selected from the adjustment hyperparameter solution space, and the plurality of adjustment hyperparameters are used to fine-tune the parameters of the first multi-attention situation feature extractor to obtain a plurality of fine-tuned first multi-attention situation feature extractors; Based on the direction fine-tuning verification set, the plurality of fine-tuned first multi-attention situation feature extractors are subjected to adjustment hyperparameter reliability verification to obtain a plurality of adjustment hyperparameter reliability coefficients; The adjustment hyperparameter corresponding to the maximum value in the plurality of adjustment hyperparameter reliability coefficients is taken as a fine-tuning direction, and the plurality of adjustment hyperparameters are directionally fine-tuned in the adjustment hyperparameter solution space to obtain a plurality of updated adjustment hyperparameters; The first multi-attention situation feature extractor is again subjected to parameter fine-tuning based on the plurality of updated adjustment hyperparameters, and the plurality of updated fine-tuned first multi-attention situation feature extractors are subjected to updated adjustment hyperparameter reliability verification by calling the direction fine-tuning verification set to obtain a plurality of updated adjustment hyperparameter reliability coefficients; determining whether there is an update adjustment hyperparameter reliability coefficient greater than or equal to the adjustment hyperparameter reliability coefficient corresponding to the fine-tuning direction in the plurality of update adjustment hyperparameter reliability coefficients, if yes, updating the fine-tuning direction based on the update adjustment hyperparameter corresponding to the maximum value in the plurality of update adjustment hyperparameter reliability coefficients, and continuing to fine-tune the plurality of update adjustment hyperparameters in direction, and counting the number of updates; when the number of updates is greater than or equal to a preset update number threshold, determining whether the maximum value of the update hyperparameter reliability coefficient in the update process is greater than or equal to a preset hyperparameter reliability coefficient, if yes, taking the update adjustment hyperparameter corresponding to the maximum value of the update hyperparameter reliability coefficient in the update process as a target adjustment hyperparameter, and adjusting the parameters of the first multi-attention situation feature extractor based on the target adjustment hyperparameter to obtain a first optimized multi-attention situation feature extractor; if the maximum value of the update hyperparameter reliability coefficient in the update process is less than the preset hyperparameter reliability coefficient, an instruction for reselecting is obtained, a plurality of adjustment hyperparameters are randomly selected in the adjustment hyperparameter solution space according to the reselection instruction, and the first multi-attention situation feature extractor is fine-tuned in direction to obtain the first optimized multi-attention situation feature extractor.

8. The active power distribution grid situation prediction method based on multiple attention mechanisms according to claim 7, characterized in that, Based on the directional fine-tuning verification set, the plurality of fine-tuned first multi-attention situation feature extractors are adjusted hyperparameter reliability verified to obtain a plurality of adjustment hyperparameter reliability coefficients, including: using the plurality of fine-tuned first multi-attention situation feature extractors to extract situation features from the historical difficult sample multi-source running data sequence set in the directional fine-tuning verification set respectively, and obtaining a plurality of fine-tuned output multi-layer situation feature sets; comparing and verifying the plurality of fine-tuned output multi-layer situation feature sets with the historical difficult sample multi-layer situation feature set identified in the directional fine-tuning verification set respectively, and obtaining a plurality of adjustment hyperparameter reliability coefficients. 9.The active power distribution grid situation prediction method based on multiple attention mechanisms of claim 1, wherein, The weighting decoupling module is embedded with an active power distribution network situation factor, wherein the active power distribution network situation factor includes node voltage out-of-limit margin, branch load severity, voltage fluctuation coefficient and current fluctuation coefficient.

10. An active power distribution grid situation prediction system based on multiple attention mechanism, characterized in that, The system comprises: a data acquisition and division module for hierarchical multi-source data acquisition of bus layers, feeder layers and district layers of a target active power distribution network, and dividing the acquisition results according to a preset base class division rule to obtain base class multi-source running data sequence sets and new class multi-source running data sequence sets; a feature extractor acquisition module for acquiring a pre-constructed twin multi-attention situation feature extractor set, wherein the twin multi-attention situation feature extractor set includes a first multi-attention situation feature extractor and a second multi-attention situation feature extractor; an extractor fine-tuning module for matching a historical difficult sample multi-source running data sequence set based on the base class multi-source running data sequence set, fine-tuning the first multi-attention situation feature extractor in direction based on the historical difficult sample multi-source running data sequence set, and obtaining a first optimized multi-attention situation feature extractor; The situation prediction module is configured to perform situation feature identification on the base-class multi-source operation data sequence set by using the first optimized multi-attention situation feature extractor, perform situation feature identification on the new-class multi-source operation data sequence set by using the second multi-attention situation feature extractor, and transmit the identification results to the weighted decoupling module to perform situation prediction and obtain multi-layer situation prediction results.

Images