Overload risk assessment methods, apparatus, computer equipment and storage media

Abstract

This application relates to an overload risk assessment method, apparatus, computer equipment, and storage medium. The method includes: collecting historical load data from multiple users and performing decoupling processing to decompose the total load sequence into multiple independent load subsequences; inputting the total load sequence and the multiple independent load subsequences into corresponding prediction models of decomposed linear probabilities to obtain the distribution data of load conditional probabilities within the prediction window; calculating a first risk index and a second risk index based on the load conditional probability distribution data; constructing a two-dimensional risk decision function based on the first risk index, the second risk index, and a preset thermal inertia time threshold to map the continuous probability space into discrete operating state levels; and determining the core load components causing distribution network overload risk from the multiple independent load subsequences in response to triggering a preset operating state level. This method can improve the accuracy of overload risk assessment.

Description

Technical Field

[0001] This application relates to the field of power grid risk assessment technology, and in particular to an overload risk assessment method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] Short-term load forecasting is a crucial aspect of ensuring the safe operation and economical dispatch of distribution networks. Traditionally, deep learning-based forecasting methods such as LSTM and CNN have been widely studied. However, these models often employ iterative strategies when handling multi-step forecasting tasks, leading to error accumulation, and suffer from difficulties in hyperparameter tuning and low computational efficiency. To overcome these shortcomings, Transformer and its variants have been introduced, achieving breakthroughs in long-term time-series forecasting, but their complex structures are prone to overfitting. Recent research indicates that simplified decomposed linear models actually outperform traditional models in both forecasting performance and efficiency.

[0003] However, the aforementioned methods are mostly point-based forecasts, failing to quantify the uncertainty of the forecast results. When facing highly uncertain distribution network loads, point forecasts have significant safety blind spots: when the load is in a critical state, even if the forecast average does not exceed the threshold, the actual load may still exceed safety limits due to random fluctuations, leading to missed overload risks. Furthermore, current assessment systems lack consideration for the physical characteristics of power equipment, such as thermal inertia, and using a single threshold for judgment is prone to frequent false alarms due to short-term fluctuations. Simultaneously, most studies focus only on total system load analysis, making it impossible to trace the source of risk when an alarm is triggered, thus hindering the provision of decision-making basis for accurate demand-side response.

[0004] Therefore, there is an urgent need for an overload risk assessment method, device, computer equipment, computer-readable storage medium, and computer program product that can improve the accuracy of overload risk assessment. Summary of the Invention

[0005] Therefore, it is necessary to provide an overload risk assessment method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the accuracy of overload risk assessment in response to the above-mentioned technical problems.

[0006] Firstly, this application provides an overload risk assessment method, including:

[0007] Collect historical load data from multiple users within the transformer area of ​​the distribution network;

[0008] The multi-user historical load data is decoupled to decompose the total load sequence into multiple independent load subsequences;

[0009] The total load sequence and the multiple independent load subsequences are respectively input into the corresponding prediction model of decomposed linear probability to obtain the distribution data of load conditional probability within the prediction window;

[0010] Based on the load conditional probability distribution data corresponding to the total load sequence, a first risk index and a second risk index are calculated; the first risk index is the extreme probability of the load exceeding the preset safety threshold within the prediction window, and the second risk index is the maximum continuous duration for which the overload probability continuously exceeds the preset warning threshold.

[0011] Based on the first risk indicator, the second risk indicator, and the preset thermal inertia time threshold, a two-dimensional risk decision function is constructed to map the continuous probability space into discrete operating state levels.

[0012] In response to triggering a preset operating state level, the core load components that cause the risk of overload in the distribution network are determined from multiple independent load subsequences based on the load condition probability distribution data corresponding to each independent load subsequence.

[0013] In one embodiment, the decoupling process of the multi-user historical load data, which decomposes the total load sequence into multiple independent load subsequences, includes:

[0014] Based on the multi-user historical load data, users whose peak load is lower than the preset physical stratification threshold are classified into the first user set, and users whose peak load is higher than the preset physical stratification threshold are classified into the second user set.

[0015] The original load data of the first user set is combined as an independent load subsequence, and the historical load curves of each user in the second user set are normalized to obtain a shape vector.

[0016] The K-Means clustering algorithm is used to cluster the morphological vectors, grouping users with similar load curve shapes into the same subclass;

[0017] Based on the clustering labels, the original load data of the second user set is recombined to obtain several independent load subsequences;

[0018] The independent load subsequences are composed of the independent load subsequences corresponding to the first user set and the several independent load subsequences corresponding to the second user set.

[0019] In one embodiment, the prediction model that decomposes linear probabilities is configured as follows:

[0020] The input total load sequence and the multiple independent load subsequences are decoupled into trend components and periodic components using a moving average operation;

[0021] The trend component and the periodic component are respectively input into independent linear layers for feature mapping;

[0022] The distribution data of load condition probabilities within the prediction window are output through a dual-channel probability header. The dual-channel probability header includes a mean channel and a variance channel. The mean channel is used to output the expected value of the predicted distribution data, and the variance channel outputs the standard deviation of the predicted distribution data after constraining the linear layer output through the Softplus activation function.

[0023] In one embodiment, the training process of the prediction model that decomposes linear probabilities includes:

[0024] Obtain a training sample set, which includes historical load data and corresponding real load observation data;

[0025] The training sample set is input into the initial model, and the initial model outputs the mean and standard deviation of the predicted distribution data.

[0026] The loss value is calculated based on the negative log-likelihood loss function, which is used to simultaneously optimize the accuracy of the prediction mean and the reasonableness of the prediction standard deviation.

[0027] The parameters of the initial model are adjusted based on the loss value until the model converges, resulting in a prediction model that decomposes linear probabilities.

[0028] In one embodiment, the step of outputting the distribution data of load condition probabilities within the prediction window via a dual-channel probability head includes:

[0029] The original value of the linear layer output corresponding to the variance channel is input into the Softplus activation function;

[0030] The original values ​​are mapped to positive real numbers using the Softplus activation function, which are then output as the standard deviation of the predicted distribution data; wherein, the Softplus activation function is used to ensure that the standard deviation of the output is always greater than zero.

[0031] In one embodiment, the step of constructing a two-dimensional risk decision function based on the first risk indicator, the second risk indicator, and a preset thermal inertia time threshold, mapping a continuous probability space to discrete operating state levels, includes:

[0032] The first risk indicator is compared with the preset warning threshold and the preset deterministic threshold, respectively, and the second risk indicator is compared with the preset thermal inertia time threshold; the preset deterministic threshold is greater than the preset warning threshold.

[0033] If the first risk indicator is less than the preset warning threshold, the operating status is determined to be a safe state;

[0034] If the first risk indicator is greater than or equal to the preset warning threshold and less than the preset deterministic threshold, the operating status is determined to be a warning status.

[0035] If the first risk indicator is greater than or equal to the preset deterministic threshold and the second risk indicator is less than the preset thermal inertia time threshold, the operating state is determined to be a warning state.

[0036] If the first risk indicator is greater than or equal to the preset deterministic threshold and the second risk indicator is greater than or equal to the preset thermal inertia time threshold, the operating state is determined to be an overload state.

[0037] Secondly, this application also provides an overload risk assessment device, comprising:

[0038] The data acquisition module is used to collect historical load data from multiple users within the transformer area of ​​the distribution network.

[0039] The decomposition module is used to decouple the multi-user historical load data and decompose the total load sequence into multiple independent load subsequences.

[0040] The processing module is used to input the total load sequence and the multiple independent load subsequences into the corresponding prediction model of decomposed linear probability, respectively, to obtain the distribution data of load conditional probability within the prediction window;

[0041] The calculation module is used to calculate a first risk indicator and a second risk indicator based on the load conditional probability distribution data corresponding to the total load sequence; the first risk indicator is the extreme probability of the load exceeding the preset safety threshold within the prediction window, and the second risk indicator is the maximum continuous duration for which the overload probability continuously exceeds the preset warning threshold.

[0042] The construction module is used to construct a two-dimensional risk decision function based on the first risk indicator, the second risk indicator and the preset thermal inertia time threshold, so as to map the continuous probability space into discrete operating state levels.

[0043] The evaluation module is used to determine the core load components that cause the risk of overload in the distribution network from multiple independent load subsequences based on the load condition probability distribution data corresponding to each independent load subsequence, in response to the triggering of a preset operating state level.

[0044] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0045] Collect historical load data from multiple users within the transformer area of ​​the distribution network;

[0046] The multi-user historical load data is decoupled to decompose the total load sequence into multiple independent load subsequences;

[0047] The total load sequence and the multiple independent load subsequences are respectively input into the corresponding prediction model of decomposed linear probability to obtain the distribution data of load conditional probability within the prediction window;

[0048] Based on the load conditional probability distribution data corresponding to the total load sequence, a first risk index and a second risk index are calculated; the first risk index is the extreme probability of the load exceeding the preset safety threshold within the prediction window, and the second risk index is the maximum continuous duration for which the overload probability continuously exceeds the preset warning threshold.

[0049] Based on the first risk indicator, the second risk indicator, and the preset thermal inertia time threshold, a two-dimensional risk decision function is constructed to map the continuous probability space into discrete operating state levels.

[0050] In response to triggering a preset operating state level, the core load components that cause the risk of overload in the distribution network are determined from multiple independent load subsequences based on the load condition probability distribution data corresponding to each independent load subsequence.

[0051] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0052] Collect historical load data from multiple users within the transformer area of ​​the distribution network;

[0053] The multi-user historical load data is decoupled to decompose the total load sequence into multiple independent load subsequences;

[0054] The total load sequence and the multiple independent load subsequences are respectively input into the corresponding prediction model of decomposed linear probability to obtain the distribution data of load conditional probability within the prediction window;

[0055] Based on the load conditional probability distribution data corresponding to the total load sequence, a first risk index and a second risk index are calculated; the first risk index is the extreme probability of the load exceeding the preset safety threshold within the prediction window, and the second risk index is the maximum continuous duration for which the overload probability continuously exceeds the preset warning threshold.

[0056] Based on the first risk indicator, the second risk indicator, and the preset thermal inertia time threshold, a two-dimensional risk decision function is constructed to map the continuous probability space into discrete operating state levels.

[0057] In response to triggering a preset operating state level, the core load components that cause the risk of overload in the distribution network are determined from multiple independent load subsequences based on the load condition probability distribution data corresponding to each independent load subsequence.

[0058] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0059] Collect historical load data from multiple users within the transformer area of ​​the distribution network;

[0060] The multi-user historical load data is decoupled to decompose the total load sequence into multiple independent load subsequences;

[0061] The total load sequence and the multiple independent load subsequences are respectively input into the corresponding prediction model of decomposed linear probability to obtain the distribution data of load conditional probability within the prediction window;

[0062] Based on the load conditional probability distribution data corresponding to the total load sequence, a first risk index and a second risk index are calculated; the first risk index is the extreme probability of the load exceeding the preset safety threshold within the prediction window, and the second risk index is the maximum continuous duration for which the overload probability continuously exceeds the preset warning threshold.

[0063] Based on the first risk indicator, the second risk indicator, and the preset thermal inertia time threshold, a two-dimensional risk decision function is constructed to map the continuous probability space into discrete operating state levels.

[0064] In response to triggering a preset operating state level, the core load components that cause the risk of overload in the distribution network are determined from multiple independent load subsequences based on the load condition probability distribution data corresponding to each independent load subsequence.

[0065] The aforementioned overload risk assessment methods, devices, computer equipment, computer-readable storage media, and computer program products, by collecting and decoupling historical load data from multiple users, decompose the complex and heterogeneous total load into several load subsequences with independent morphological characteristics, laying a data foundation for subsequent accurate risk source location. By inputting the total load sequence and each independent load subsequence into the decomposed linear probability prediction model, load conditional probability distribution data within the future prediction window is obtained, realizing a paradigm shift from deterministic point prediction to probabilistic interval prediction. This effectively quantifies the uncertainty brought about by load fluctuations and compensates for the safety blind spots in traditional point prediction methods under critical operating conditions. Furthermore, based on the total load probability distribution, two core risk indicators—extreme probability and maximum continuous duration—are calculated, and a two-dimensional risk decision function is constructed by integrating the thermal inertia characteristics of power equipment. This maps the continuous probability space into three discrete states: safety, early warning, and overload. This avoids false alarms caused by short-term disturbances and provides dispatchers with a clear and intuitive operational situation awareness. Finally, in response to triggering an early warning or overload state, based on the probability distribution data of each independent load subsequence, the core load component causing overload risk is accurately located from multiple subsequences. This achieves a technological breakthrough from macro-system alarm to micro-level component tracing, comprehensively improving the accuracy, robustness, and interpretability of overload risk assessment. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a diagram illustrating the application environment of an overload risk assessment method in one embodiment.

[0068] Figure 2 This is a general framework diagram of an overload risk assessment method in one embodiment;

[0069] Figure 3 This is a flowchart illustrating an overload risk assessment method in one embodiment;

[0070] Figure 4 This is a schematic diagram of the load prediction results of a point prediction model in one embodiment;

[0071] Figure 5 This is a schematic diagram of the load overload probability prediction result of a probability prediction model in one embodiment;

[0072] Figure 6This is a schematic diagram illustrating the overload situation of the actual load value under pressure test conditions in one embodiment;

[0073] Figure 7 This is a diagram showing the clustering effect of the load data of a probabilistic prediction model in one embodiment.

[0074] Figure 8 This is a summary graph of the overload probability for various user types in one embodiment;

[0075] Figure 9 This is a structural block diagram of an overload risk assessment device in one embodiment;

[0076] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0078] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0079] The overload risk assessment method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on another network server.

[0080] The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle systems, and projection devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0081] In one exemplary embodiment, such as Figure 2 As shown, an overload risk assessment method is provided, which is then applied to... Figure 1 Taking server 104 as an example, the explanation is as follows: Figure 3 As shown, the procedure includes steps S302 to S312. Wherein:

[0082] Step S302: Collect historical load data of multiple users within the transformer area of ​​the distribution network.

[0083] Specifically, intelligent measurement terminals or data acquisition and monitoring systems deployed on the transformer side of the distribution network acquire the time-series records of the active power load of each independent user within the transformer's power supply area at a fixed sampling frequency (e.g., 15 minutes / time). The collected data covers complete daily, weekly, and quarterly periodic characteristics, forming a two-dimensional data structure containing multiple users and multiple continuous time points. Each row represents a sampling time, and each column represents the historical load curve of an independent user. This raw data is the source of multi-user load aggregation and the basic input for subsequent decoupling analysis and probabilistic prediction.

[0084] Step S304: Decouple the historical load data of multiple users and decompose the total load sequence into multiple independent load subsequences.

[0085] Specifically, the collected historical load data from multiple users is decoupled, decomposing the original total load sequence into multiple load subsequences with independent morphological characteristics. The total load within the transformer area of ​​the distribution network is composed of numerous user loads with vastly different electricity consumption behaviors, exhibiting significant heterogeneity: different types of users (such as industrial and commercial users and residential users) have fundamental differences in electricity consumption levels, work and rest patterns, and fluctuation patterns. If the total load sequence is directly predicted, the mixing and interference between different electricity consumption patterns will obscure the inherent patterns of each component, increasing the difficulty of prediction and reducing the accuracy of risk tracing. Based on this, by decoupling the historical load data from multiple users, and according to the inherent statistical characteristics and physical properties of the load data, the mixed and superimposed total load sequence is reorganized and separated, generating several independent load subsequences with strong internal homogeneity and significant external differences.

[0086] Step S306: Input the total load sequence and multiple independent load subsequences into the corresponding prediction models of decomposed linear probabilities to obtain the distribution data of load conditional probabilities within the prediction window.

[0087] Specifically, a dual-track prediction architecture of macro and micro levels was constructed: the macro track corresponds to the total load sequence, used to monitor the overall operational status of the system; the micro track corresponds to multiple independent load sub-sequences, used to capture the personalized fluctuation patterns of different user categories. Each decomposed linear probability prediction model adopts the same core architecture, possessing the ability to map the input historical sequence to a future probability distribution, but the parameters of each model are trained independently, each adapted to the statistical characteristics of its corresponding load sequence. The load conditional probability distribution data output by the model is not a single point prediction value, but rather distribution information presented in the form of a probability density function, typically characterized by statistical parameters such as mean and standard deviation, used to characterize the probability distribution of load values ​​at future times. This distribution data can quantify the inherent uncertainty of the prediction results, providing rich information beyond point estimation for subsequent risk assessment, enabling the system not only to predict the most likely load level, but also to perceive the potential range of load fluctuations, especially the tail risk of the load upper bound exceeding the safety threshold.

[0088] Step S308: Calculate the first risk index and the second risk index based on the load conditional probability distribution data corresponding to the total load sequence; the first risk index is the extreme probability of the load exceeding the preset safety threshold within the prediction window, and the second risk index is the maximum continuous duration for which the overload probability continuously exceeds the preset warning threshold.

[0089] Specifically, in the overload probability time series, all consecutive time periods where the overload probability exceeds the warning threshold are identified, and the duration of each period is calculated. The maximum value among these periods is taken as the second risk indicator. This indicator, by introducing a time dimension, reflects the duration of the overload risk, thus distinguishing between transient disturbances and persistent overloads. These two indicators characterize overload risk from the dimensions of intensity and persistence, respectively. The former captures the peak intensity of the risk, while the latter captures the cumulative effect of the risk over time. Together, they provide a quantitative basis for subsequent decision-making based on the thermal inertia characteristics of power equipment, overcoming the limitation of traditional single-threshold judgments in distinguishing between transient spikes and persistent overloads.

[0090] Step S310: Based on the first risk indicator, the second risk indicator, and the preset thermal inertia time threshold, a two-dimensional risk decision function is constructed to map the continuous probability space into discrete operating state levels.

[0091] Specifically, due to the significant thermal inertia characteristics of power system equipment (such as distribution transformers), their physical temperature does not change instantaneously with load fluctuations, but rather exhibits a thermal accumulation effect: short-term load overruns usually do not cause a sudden rise in winding temperature and thus equipment damage; only overloads lasting for a certain period of time will cause substantial heat accumulation and insulation aging. Traditional safety judgment methods based on a single threshold ignore this physical characteristic and are prone to misjudging instantaneous disturbances or brief fluctuations as overload events, leading to frequent alarms and dispatch interference. To address this issue, this step introduces a two-dimensional risk decision function, using a first risk indicator (overload intensity) and a second risk indicator (overload duration) as input dimensions, and combining them with a preset thermal inertia time threshold (characterizing the allowable short-term overload tolerance duration of the equipment) to construct a joint judgment logic. This function first compares the first risk indicator with preset warning thresholds and deterministic thresholds to determine the severity of the overload intensity; simultaneously, it compares the second risk indicator with the thermal inertia time threshold to determine whether the duration of the overload risk is sufficient to cause thermal damage to the equipment. Through the cross-judgment of the above two dimensions, the two-dimensional risk decision function maps the originally continuous and complex probability space (i.e., the continuous interval of overload probability from low to high and the continuous interval of duration from short to long) into a finite and discrete operating state level.

[0092] Step S312: In response to triggering a preset operating state level, the core load components that cause the risk of overload in the distribution network are determined from multiple independent load subsequences based on the load condition probability distribution data corresponding to each independent load subsequence.

[0093] Specifically, when the constructed two-dimensional risk decision function determines that the current operating state has entered a warning state or an overload state, the system triggers a risk tracing mechanism to initiate the location analysis of the risk source. Multiple independent load subsequences obtained through decoupling are used as the analysis objects, with each independent load subsequence corresponding to a user group with similar electricity consumption behavior patterns. During the risk triggering period, the system retrieves the load conditional probability distribution data obtained from parallel predictions of each independent load subsequence and calculates the overload risk index for each subsequence within the same time period based on this data. By conducting a horizontal comparative analysis of the risk indices of each subsequence, the load subsequence with the highest overload probability amplitude or the longest duration of overload probability exceeding the warning threshold during the risk triggering period is identified. This identification result is the core load component causing the overall overload risk of the system.

[0094] In the aforementioned overload risk assessment method, historical load data from multiple users is collected and decoupled to decompose the complex and heterogeneous total load into several load subsequences with independent morphological characteristics, laying a data foundation for subsequent accurate risk source location. By inputting the total load sequence and each independent load subsequence into the decomposed linear probability prediction model, load conditional probability distribution data within the future prediction window is obtained, realizing a paradigm shift from deterministic point prediction to probabilistic interval prediction. This effectively quantifies the uncertainty brought about by load fluctuations and compensates for the safety blind spots in traditional point prediction methods under critical operating conditions. Furthermore, based on the total load probability distribution, two core risk indicators—extreme probability and maximum continuous duration—are calculated, and a two-dimensional risk decision function is constructed by integrating the thermal inertia characteristics of power equipment. This maps the continuous probability space into three discrete states: safety, early warning, and overload. This avoids false alarms caused by short-term disturbances and provides dispatchers with a clear and intuitive operational situation awareness. Finally, in response to triggering an early warning or overload state, based on the probability distribution data of each independent load subsequence, the core load component causing overload risk is accurately located from multiple subsequences. This achieves a technological breakthrough from macro-system alarm to micro-level component tracing, comprehensively improving the accuracy, robustness, and interpretability of overload risk assessment.

[0095] In one exemplary embodiment, decoupling processing is performed on multi-user historical load data, decomposing the total load sequence into multiple independent load subsequences, including:

[0096] Based on historical load data of multiple users, users whose peak load is lower than the preset physical stratification threshold are classified into the first user set, and users whose peak load is higher than the preset physical stratification threshold are classified into the second user set.

[0097] The original load data of the first user set are combined as independent load subsequences, and the historical load curves of each user in the second user set are normalized to obtain a shape vector.

[0098] The K-Means clustering algorithm is used to cluster morphological vectors, grouping users with similar load curve shapes into the same subclass;

[0099] Based on the clustering labels, the original load data of the second user set is recombined to obtain several independent load subsequences;

[0100] Multiple independent load subsequences are formed by the independent load subsequences corresponding to the first user set and several independent load subsequences corresponding to the second user set.

[0101] Specifically, firstly, coarse-grained stratification is performed based on the physical magnitude characteristics of the load. By setting a physical stratification threshold, the historical peak load of each user is compared with this threshold: users with peak loads below the threshold are classified into the first user group, typically corresponding to residential users or micro-enterprises, whose load curves are more volatile but smaller in magnitude; users with peak loads above the threshold are classified into the second user group, typically corresponding to industrial and commercial users or large user groups, whose load magnitudes are larger and whose work and rest patterns are obvious. This stratification effectively eliminates the interference of random noise on the low-voltage side on the regular load on the high-voltage side.

[0102] Secondly, a morphological decoupling was performed on the second user set. Since industrial and commercial users still exhibit different production schedules, such as continuous production, single-shift production, and double-shift production, their load curve shapes show significant differences. To finely distinguish these electricity consumption patterns, the historical load curves of each user were normalized to eliminate the influence of absolute amplitude, retaining only the morphological characteristics of load fluctuations, resulting in morphological vectors. Subsequently, the K-Means clustering algorithm was used to perform unsupervised clustering of all morphological vectors, using Euclidean distance as a similarity measure. Users with similar daily load curve shapes were grouped into the same subclass, ensuring high homogeneity of electricity consumption patterns within the same subclass and significant differences in electricity consumption patterns between different subclasses.

[0103] Finally, based on the labels obtained from clustering, the original load data of the second user set is reorganized, generating an independent load subsequence for each subclass. Simultaneously, the original load data of the first user set is directly combined into an independent load subsequence. Thus, the load subsequence corresponding to the first user set and several load subsequences corresponding to the second user set together constitute multiple independent load subsequences.

[0104] To further address the complex differences in work and rest schedules within the industrial and commercial user group, an unsupervised learning method is employed for morphological segmentation.

[0105] First, residential users and industrial / commercial users are distinguished based on data volume. Then, normalization is performed: the historical load curves of each industrial / commercial user are normalized using Min-Max to obtain a shape vector. :

[0106] ;

[0107] This operation can eliminate the influence of absolute amplitude and retain only the morphological characteristics of load fluctuation.

[0108] After normalization, the K-Means algorithm is applied to cluster the normalized morphological vectors. This is done by calculating the Euclidean distance.

[0109] ;

[0110] in, The number of clusters, For the first Subclasses This serves as the shape center for this subclass. Users with similar daily load curve shapes are grouped into the same subclass.

[0111] Finally, based on the clustering labels, the original load data is recombine into several independent load subsequences, which serve as independent input channels for subsequent probabilistic prediction models. Through the above two-step hierarchical process, the complex total load is deconstructed into several distinctive and highly regular sub-load components.

[0112] In this embodiment, by performing two-layer decoupling processing on multi-user load data, the complex and heterogeneous total load is reconstructed into homogeneous independent subsequences. The load probability distribution is output by decomposing the linear probability prediction model, which solves the safety blind spot problem in point prediction. By constructing a two-dimensional risk index that integrates thermal inertia characteristics, the continuous probability space is mapped to discrete operating states, effectively filtering out false alarms caused by instantaneous disturbances.

[0113] In an exemplary embodiment, the prediction model that decomposes linear probabilities is configured as follows:

[0114] The moving average operation is used to decouple the input total load sequence and multiple independent load subsequences into trend components and periodic components.

[0115] The trend component and the periodic component are respectively input into independent linear layers for feature mapping;

[0116] The dual-channel probability header outputs the distribution data of the load condition probability within the prediction window. The dual-channel probability header includes a mean channel and a variance channel. The mean channel is used to output the expected value of the predicted distribution data, and the variance channel outputs the standard deviation of the predicted distribution data after constraining the output of the linear layer through the Softplus activation function.

[0117] Specifically, in the configuration of the decomposed linear probabilistic prediction model, the model first uses a moving average operation to decouple the input historical load sequence into trend and periodic components. The trend component reflects the long-term trend of load changes, while the periodic component captures the periodic fluctuations of load. This decomposition operation reduces the prediction difficulty of non-stationary time series. Subsequently, the decomposed trend and periodic components are input into independent linear layers for feature mapping, enabling the model to learn independent mapping parameters for time series components of different properties, enhancing the targeting of feature extraction. Finally, the model outputs the prediction distribution through a dual-channel probability head: the mean channel outputs the expected value of the prediction, and the variance channel constrains the output of the linear layer through the Softplus activation function, ensuring that the standard deviation of the output is always a positive real number, thus constructing a complete normal distribution. This configuration realizes a paradigm shift from deterministic point prediction to probabilistic interval prediction.

[0118] To quantify uncertainty, standards The linear probability prediction model (i.e., decomposed linear probability prediction model) has been probabilistically modified to enable it to output conditional probability distributions.

[0119] To reduce the difficulty of predicting non-stationary load data, the model first introduces a sequence decomposition module. For the input historical sequence... The moving average operation is used to decouple it into trend components. and periodic components :

[0120] In the sequence decomposition module, for the input historical sequence We use moving average operations to decouple it into trend components. and periodic components Among them, the trend component The calculation formula is defined as follows:

[0121] ;

[0122] in, Indicates at time step The trend component value is calculated at that time. Represents the original sequence exist Load observations at any given time. The window length represents the moving average (usually an odd number to ensure symmetry). and These represent rounding down and rounding up, respectively, to ensure that the moving average window circles... Symmetrical at all times.

[0123] ;

[0124] in, The periodic component is obtained by subtracting the trend component value from the load observation value. This decomposition process enables the model to capture the long-term growth trend and periodic fluctuations of the load, thereby enhancing the targeting of feature extraction.

[0125] In the feature mapping stage, independent linear layers are applied to the trend component and the periodic component for regression prediction to obtain their respective future predicted values, and these values ​​are superimposed to obtain the final prediction result.

[0126] To achieve probabilistic prediction, the model output was improved by designing a dual-channel probability header:

[0127] Mean channel: Outputs the expected value of the predicted distribution. This represents the most likely load level at a future moment.

[0128] Variance channel: Outputs the standard deviation of the predicted distribution. To ensure the non-negativity of the standard deviation, we use... The activation function constrains the output of the linear layer:

[0129] ;

[0130] in, It represents the standard deviation of the Gaussian distribution of the predicted load, used to quantify the uncertainty of the prediction results, and its value is always greater than zero. This represents the input data from the feature extraction layer inside the model. The independent fully connected linear transform layer representing the variance channel, its output value Let be any real number. Right now Activation function. This function maps the output of the linear layer to... Within the scope, ensured It always uses the effective standard deviation, thus avoiding mathematical errors during the training process.

[0131] In this embodiment, a model configuration combining sequence decomposition and dual-channel probability output achieves probabilistic prediction capabilities while maintaining the high efficiency of a linear model. Sequence decomposition reduces the difficulty of predicting non-stationary loads, while independent linear layers enhance the specificity of feature extraction for different time-series components. The dual-channel probability head outputs the effective standard deviation through Softplus constraints, enabling the model to learn the heteroscedasticity characteristics of the load data. This allows the model to output narrow confidence intervals during stable periods and wide confidence intervals during periods of sharp fluctuations, thus truly quantifying prediction uncertainty.

[0132] In an exemplary embodiment, the training process of a prediction model that decomposes linear probabilities includes:

[0133] Obtain a training sample set, which includes historical load data and corresponding real load observation data;

[0134] The training sample set is input into the initial model, and the initial model outputs the mean and standard deviation of the predicted distribution data.

[0135] The loss value is calculated based on the negative log-likelihood loss function, which is used to simultaneously optimize the accuracy of the prediction mean and the reasonableness of the prediction standard deviation.

[0136] Adjust the parameters of the initial model based on the loss value until the model converges, thus obtaining a prediction model that decomposes linear probabilities.

[0137] Specifically, in the training process of the decomposed linear probability prediction model, the core lies in using supervised learning to enable the model not only to accurately predict the expected value of future loads but also to reasonably estimate the uncertainty of the prediction results, thereby providing complete probability distribution information for subsequent risk assessment. This training process specifically includes the following steps:

[0138] First, the training sample set is derived from multi-user load data collected over historical periods. Each sample consists of two parts: one part is the input historical sequence used for prediction, i.e., load observations at multiple consecutive time points; the other part is the corresponding actual load observations, i.e., the actual load data at each time point within the future prediction window after the input sequence. This sample set forms the basis for the model to learn the load evolution pattern.

[0139] Secondly, the input historical sequences from the training sample set are fed into the initialized decomposed linear probability prediction model, which performs forward computation and outputs the prediction results. Unlike traditional point prediction models that only output a single value, this model's output is designed with a dual-channel structure, outputting the mean and standard deviation of the load prediction distribution at each time point within the future prediction window. The mean represents the most likely load level, and the standard deviation represents the degree of uncertainty in the prediction results.

[0140] Next, based on the mean and standard deviation of the model output and the corresponding actual load observations in the sample, the negative log-likelihood loss function is calculated. The core idea of ​​the negative log-likelihood loss function is: assuming the actual load follows a normal distribution defined by the mean and standard deviation of the model output, the likelihood probability of observing the actual load value under this distribution is calculated, and its negative logarithm is taken as the loss. The unique feature of this loss function is that it not only penalizes the bias in mean prediction (through the mean term in the likelihood function) but also automatically adjusts the reasonableness of the standard deviation: when the prediction error is large, if the model outputs a small standard deviation (indicating overconfidence), the loss will increase sharply; conversely, if the model outputs a large standard deviation (indicating conservative estimation), the loss will be relatively moderate. This mechanism forces the model to adaptively learn the heteroscedasticity characteristics of the load data during training, that is, outputting a narrower confidence interval when the load fluctuates smoothly, and outputting a wider confidence interval when the load changes drastically or has weak regularity, thus truly reflecting the inherent uncertainty of the prediction results.

[0141] Finally, based on the calculated loss value, the parameters of the initial model are adjusted using the backpropagation algorithm, and the above forward calculation and loss optimization process is iteratively executed until the model's performance converges on the validation set. The model after training convergence is a usable decomposed linear probability prediction model, which has the ability to map any input historical sequence to a load conditional probability distribution within the future prediction window, and this distribution can accurately characterize the uncertainty of the prediction.

[0142] To train the aforementioned probability model, this embodiment abandons the traditional mean squared error loss and instead employs negative log-likelihood (NLR). Loss function. Assume the target variable follows a Gaussian distribution. The training objective is to minimize the NLL value of the observed data:

[0143] ;

[0144] in, Indicates the length of the entire predicted sequence The average negative log-likelihood loss value within. This represents the total number of time steps within the prediction window. Indicates at time step The actual load observation value at the location. Indicates the model at time step The mean of the predicted distribution. Indicates the model at time step The standard deviation of the predicted distribution.

[0145] This joint optimization mechanism enables the model to automatically learn the heteroscedasticity of the data, that is, to output a narrow confidence interval when the load is stable and a wide confidence interval when the load fluctuates sharply, thereby truly reflecting the potential risks of the system.

[0146] In this embodiment, by using a negative log-likelihood loss function to train the decomposed linear probability prediction model, a paradigm shift from point prediction to probability prediction is achieved. This not only optimizes the accuracy of the prediction mean but also jointly optimizes the rationality of the standard deviation, enabling the model to adaptively learn the heteroscedasticity characteristics of the load data. Specifically, it outputs a narrow confidence interval when the load is stable and a wide confidence interval when there are sharp fluctuations, thus truly quantifying the inherent uncertainty of the prediction results.

[0147] In one exemplary embodiment, the distribution data of load condition probabilities within the prediction window is output via a dual-channel probability header, including:

[0148] The original values ​​of the linear layer output corresponding to the variance channel are input into the Softplus activation function;

[0149] The original values ​​are mapped to positive real numbers using the Softplus activation function, which is then output as the standard deviation of the predicted distribution data. The Softplus activation function ensures that the standard deviation of the output is always greater than zero.

[0150] Specifically, the linear layer corresponding to the variance channel first performs a linear transformation on the input features, outputting an original numerical value that spans the entire real number domain and may be negative. However, the standard deviation, as a statistic describing the dispersion of a probability distribution, requires by definition to be a non-negative real number. Directly using the original output of the linear layer as the standard deviation would lead to mathematical errors and training instability. Therefore, the original output of the linear layer is input into the Softplus activation function, which maps any real number to a positive real number and is continuously differentiable across the entire domain. This ensures that the standard deviation is always greater than zero and maintains gradient smoothness, which is beneficial for backpropagation training. The output value after the Softplus mapping serves as the final standard deviation, which, together with the expected value of the mean channel output, constitutes the complete normal distribution parameters, used to characterize the probability distribution of the load values ​​at future times.

[0151] In this embodiment, the output of the variance channel is constrained by the Softplus activation function, ensuring that the standard deviation of the model output is always a valid positive real number. This avoids mathematical errors and training divergence problems caused by negative outputs from linear layers. The continuous differentiability of the Softplus function allows the variance channel to participate in end-to-end gradient backpropagation. The model can simultaneously optimize the mean and standard deviation through the negative log-likelihood loss function, thereby adaptively learning the heteroscedasticity of the learning load data.

[0152] In an exemplary embodiment, a two-dimensional risk decision function is constructed based on a first risk indicator, a second risk indicator, and a preset thermal inertia time threshold, mapping a continuous probability space to discrete operating state levels, including:

[0153] The first risk indicator is compared with the preset warning threshold and the preset deterministic threshold respectively, and the second risk indicator is compared with the preset thermal inertia time threshold; the preset deterministic threshold is greater than the preset warning threshold.

[0154] If the first risk indicator is less than the preset warning threshold, the operating status is determined to be safe.

[0155] If the first risk indicator is greater than or equal to the preset warning threshold and less than the preset deterministic threshold, the operating status is determined to be a warning status.

[0156] If the first risk indicator is greater than or equal to the preset deterministic threshold and the second risk indicator is less than the preset thermal inertia time threshold, the operating state is determined to be a warning state.

[0157] If the first risk indicator is greater than or equal to a preset deterministic threshold and the second risk indicator is greater than or equal to a preset thermal inertia time threshold, the operating state is determined to be overloaded.

[0158] Specifically, to comprehensively characterize the intensity and persistence of overload risk, this embodiment defines two core evaluation metrics: the maximum overload probability aims to quantify the maximum likelihood of the load exceeding the physical safety threshold within the prediction window. For the prediction window... any time step within Define the overload probability sequence Extreme value probability The maximum value of the sequence within the window:

[0159] ;

[0160] This metric reflects the upper bound of confidence that the system will experience overload under the worst operating conditions.

[0161] Considering the significant thermal inertia of power equipment, short-term overload exceeding limits typically does not lead to a sudden rise in winding temperature; only sustained overload will cause substantial thermal damage. Therefore, the maximum continuous duration is defined as follows: when the overload probability in the current time window exceeds the warning threshold. If the probability of overload continues to exceed the warning threshold Longest time span:

[0162] ;

[0163] This indicator effectively filters out false alarm signals caused by data noise or instantaneous disturbances by introducing a time dimension.

[0164] Based on the above two indicators, this embodiment establishes a three-level state decision-making mechanism of safety-early warning-overload, mapping continuous risk indicators into discrete signals:

[0165] This embodiment defines an early warning threshold. =0.35, deterministic threshold =0.4 and thermal inertia time threshold The current operating status of the system is determined as follows:

[0166] ;

[0167] The physical meaning of this decision logic is as follows:

[0168] Red status: Indicates effective overload. That is, the average load has exceeded the physical limit and the duration is sufficient to cause thermal damage to the equipment.

[0169] Yellow status: Indicates a risk warning. It covers two critical scenarios: high-intensity transient disturbances and medium-intensity sustained high pressure. The system prompts close monitoring.

[0170] Green state: Indicates safe operation. That is, the probability of extreme values ​​is low, and the system is operating within the safety boundary.

[0171] In this embodiment, a two-dimensional risk decision function integrating overload intensity and duration is constructed to map the continuous probability space into discrete three-level operating states: safe, warning, and overload. The decision logic first compares a first risk indicator with a preset warning threshold and a deterministic threshold, then compares a second risk indicator with a thermal inertia time threshold, achieving state division through four conditional branches. This fully considers the thermal inertia physical characteristics of power equipment, effectively distinguishing between three different scenarios: high-intensity instantaneous disturbances, medium-intensity continuous high voltage, and substantial overload. This avoids frequent false alarms caused by short-term fluctuations and provides dispatchers with a warning buffer window, enabling the identification of potential risks before overload occurs.

[0172] The most detailed embodiment of this application is as follows:

[0173] The algorithm was validated using a publicly available electricity load dataset. This dataset, collected from a real power distribution network area, contains long-term electricity load records from 321 independent users, with a sampling interval of 15 minutes, which can accurately reflect the short-term fluctuation characteristics and intraday rhythm of the load. The dataset covers complete periodic characteristics and exhibits significant non-stationarity and heterogeneity.

[0174] To verify the advantages of the proposed probabilistic prediction architecture in risk perception, a comparative experiment was designed with the traditional deterministic point prediction model. The point prediction model used was the unmodified point prediction model.

[0175] To ensure fairness and consistency in the comparison, both the proposed probabilistic model and the benchmark point prediction model adhere to the following strict settings: all models are trained and tested on the same dataset. All models are trained on the same hardware environment to ensure a fair comparison of computational cost and efficiency. All models employ a consistent sliding window mechanism and use the same optimization and learning rates. The parameter settings are as follows:

[0176] The learning rate is continuously optimized by the same Adam optimizer. The historical lookback window length is 366 time steps, the future prediction window length is 96 time steps, the number of experimental iterations is 1, and the training rounds are 20. If a better solution is not obtained after 20 rounds, the training is terminated early.

[0177] Through the rigorous experimental controls described above, interference caused by differences in data or training can be eliminated, thereby allowing for a focused verification of the unique advantages of the probabilistic prediction model proposed in this paper in capturing long-tail risks and reducing the false negative rate.

[0178] Experiment 1: Comparison of point prediction and probability prediction:

[0179] To verify the advantages of this embodiment in risk perception capability, this embodiment uses the traditional deterministic point prediction model as a benchmark and compares it with the probabilistic prediction model of this embodiment on the same test set.

[0180] (1) Point prediction results:

[0181] Figure 4 The graph shows the performance of the point prediction model on the full test set: the horizontal axis represents the time window index, covering the entire test period; the vertical axis represents the load power value, representing the peak load within each prediction window; the dashed line at the top of the graph represents the set physical overload threshold, and the solid black line represents the peak load curve of the window output by the point prediction model.

[0182] like Figure 4 As shown, the output curve of the point prediction model remained below the threshold line throughout the entire test period. Even during peak load periods near the 2500-hour window, the predicted mean only approached the threshold and never substantially exceeded it. If scheduling is based on this result, the system will determine that the current power grid is in a safe state in the vast majority of cases.

[0183] (2) Probability prediction results:

[0184] To verify the effectiveness of the tiered early warning mechanism in actual time series data, this paper conducted a risk assessment experiment on 5,000 consecutive time windows in the test set. The experimental results are as follows: Figure 5 As shown.

[0185] In the figure, the gray curve represents the time window with no overload risk, while the black crosses and gray dots correspond to the time windows with high overload risk and some overload risk, respectively. Observing the alarm distribution in the figure reveals that this mechanism can accurately classify risks based on probability amplitude and duration. When the overload probability curve exceeds 0.4 and meets the high-risk condition of lasting for more than four time windows, the system densely marks the cross alarm points, accurately capturing the sustained high overload risk segment near time steps 1500 to 2500 in the figure. Simultaneously, for potential risk intervals with probabilities between 0.35 and 0.4, and for some sudden spikes with probabilities briefly exceeding 0.4 but not meeting the duration requirement, the system correctly marks them as gray dot alarms. This warning result based on dual-judgment logic not only intuitively demonstrates the evolution of risk but also effectively avoids false alarms and frequent switching caused by a single threshold judgment, proving the robustness of this method in capturing true overload risks.

[0186] To verify the value of the proposed three-level state determination mechanism (safety-early warning-overload) in actual operation, the model successfully captured 13 instances of high overload risk, which will be analyzed in detail later. Furthermore, the yellow light status plays a crucial buffering and filtering role in risk assessment, specifically in the following two aspects:

[0187] Statistical data shows that 86 transient disturbance events were correctly identified by the system as a yellow alert rather than a red alert. This verifies the effectiveness of the thermal inertia criterion introduced in this paper. If a traditional single threshold judgment were used, these 86 brief probability spikes would be misjudged as severe overload, leading to serious alarm fatigue for dispatchers. This model successfully intercepted these invalid alarms before they reached the red light state, ensuring a high confidence level for red light alarms.

[0188] The system identified 911 high-load warning states, corresponding to periods when the load was in a critical region. This intermediate state provides dispatchers with a valuable buffer window. In practical engineering, dispatchers can use this warning signal to take flexible load reduction or energy storage dispatch measures before a red overload occurs, thereby eliminating risks at the nascent stage and avoiding the passive situation of the system directly entering a hard load shedding phase.

[0189] (3) Comparison of overload warning effects:

[0190] Because real power grid operation data has extremely high stability, overload faults are typical long-tailed sparse samples, making it difficult to effectively verify the triggering mechanism of the risk warning model using the original design threshold directly on a limited test set. To comprehensively evaluate the model's ability to capture faults under critical and overload conditions, this embodiment employs a stress testing strategy, compressing the safety margin of the original rated load threshold (i.e., setting the experimental threshold to a fraction of the original threshold).

[0191] like Figure 6 As shown, the curve illustrates the actual load fluctuations in the test set. Since the dataset used in this embodiment has a high-frequency sampling characteristic at the 15-minute level, and the distribution network load exhibits significant daily periodicity and a huge peak-valley difference, under continuous observation for more than 5,000 time windows, the high-density intraday load oscillations show significant high-frequency filling characteristics on the time axis.

[0192] The dashed line represents the adjusted pressure test threshold. It can be seen that under normal operating conditions, the load curves are all below the safety line; however, around time step 2100, the load significantly increases and reaches a peak of 1.602MW, successfully exceeding the set pressure threshold. This artificially constructed quasi-overload scenario compensates for the lack of real positive samples.

[0193] Because it can only output red or green lights based on whether the predicted value absolutely exceeds the threshold, lacking a measure of uncertainty, it exhibits an extremely high false negative rate when dealing with critical overload samples. At these moments, even though the actual load has already experienced a positive overload, the point prediction model still determines it to be safe, which poses a serious safety hazard in actual power grid operation.

[0194] In contrast, the probabilistic model demonstrated superior risk sensitivity and tiered early warning capabilities. Firstly, it successfully captured all overload events missed by the point prediction system, correctly marking them as red. Furthermore, throughout the entire experimental project, the probabilistic prediction model detected a total of 13 red alarms, far exceeding the single red alarm detected by the point prediction system, significantly improving the system's safety margin. Secondly, within the time window of 2010, although the actual overload state was negative but close to the critical point, the probabilistic model, based on distribution prediction, sensitively issued a yellow warning. This proves that the model possesses the ability to identify potential risks, providing a buffer period before a disaster occurs, thus overcoming the shortcomings of traditional models that only consider black and white scenarios.

[0195] (4) Comparison of prediction accuracy:

[0196] To comprehensively evaluate the prediction accuracy of the probabilistic prediction model proposed in this paper, this embodiment uses the traditional deterministic point prediction model as a benchmark and conducts a comparative analysis of mean squared error and mean absolute error under the same test set and training conditions.

[0197] In terms of traditional evaluation metrics, the probabilistic prediction model proposed in this embodiment performs slightly better than the deterministic point prediction model. This phenomenon is expected and has significant engineering implications. The point prediction model, with minimizing the mean squared error (MSE) as its sole objective, tends to overfit the mean of historical data. While achieving lower numerical error, it loses the data's volatility characteristics. In contrast, the probabilistic model in this embodiment aims to minimize the mean range (NLL), simultaneously optimizing mean accuracy and variance reasonableness to cover potential random disturbances. Although the probabilistic model shows a slight increase in MSE, this minor concession in accuracy is compensated by a keen ability to detect long-tail risks. In power system security assessment, avoiding underreporting is far more important than pursuing the ultimate numerical fitting accuracy. Combined with the aforementioned risk warning experiments, it is precisely this preservation of distribution characteristics that enabled this model to successfully identify the 12 high overload risks that the point prediction model missed.

[0198] Experiment 2: Risk Source Tracing

[0199] To verify the effectiveness of the physical stratification and morphological clustering strategy, this embodiment performed cluster analysis based on the load data of the first 30 days in the original dataset. The results are as follows: Figure 7 As shown.

[0200] The curve for residents and micro-enterprises is D-shaped, exhibiting a typical bimodal characteristic with a significant evening peak and large overall fluctuations, which is consistent with residents' daily routines.

[0201] The load data of Class A users exhibits a regular "square wave" shape, with huge differences between day and night loads, and extremely stable daytime loads.

[0202] Category B users experience drastic load fluctuations and lack a clear diurnal rhythm, exhibiting strong randomness.

[0203] Class C users exhibit a single-peak characteristic in their electricity consumption throughout the day, with a significant evening peak.

[0204] The above results demonstrate that this embodiment successfully deconstructs a mixed load with significant heterogeneity into several homogeneous subsets. This data preprocessing based on clustering effectively eliminates mutual interference between different modalities, laying the foundation for subsequent independent prediction and risk attribution.

[0205] In order to track the main sources of overload, this embodiment performs probability prediction on the four types of data obtained from cluster analysis, in order to obtain the electricity consumption of different types of users and find the main influencing factors that cause overload.

[0206] Figure 8The risk evolution curves for four load components are presented. The experimental results reveal striking differences in risk distribution: First, user type B exhibits fluctuations highly consistent with the overall load risk. Within the time window of 1200 to 2500, its overload probability consistently and significantly exceeds the warning threshold of 0.4, with peak values ​​even approaching 1.0. This indicates that this type of load is the absolute dominant factor causing overall system overload. In contrast, while user type A shows some fluctuations, it consistently remains within a relatively safe range below 0.4; while user types C and D exhibit stable performance and have minimal impact on system overload risk.

[0207] This source analysis demonstrates that the system-level overload was not caused by all users, but rather by specific high-energy-consuming or highly volatile components. This conclusion provides a basis for developing precise demand response strategies, namely, that peak shaving and valley filling only for type B users can effectively eliminate the overall overload risk while minimizing intervention costs.

[0208] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0209] Based on the same inventive concept, this application also provides an overload risk assessment device for implementing the overload risk assessment method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more overload risk assessment device embodiments provided below can be found in the limitations of the overload risk assessment method described above, and will not be repeated here.

[0210] In one exemplary embodiment, such as Figure 9 As shown, an overload risk assessment device is provided, comprising:

[0211] The data acquisition module 902 is used to acquire historical load data of multiple users within the transformer area of ​​the distribution network;

[0212] The decomposition module 904 is used to decouple the historical load data of multiple users and decompose the total load sequence into multiple independent load subsequences.

[0213] The processing module 906 is used to input the total load sequence and multiple independent load subsequences into the corresponding prediction model of decomposed linear probability, respectively, to obtain the distribution data of load conditional probability within the prediction window;

[0214] The calculation module 908 is used to calculate the first risk index and the second risk index based on the load condition probability distribution data corresponding to the total load sequence. The first risk index is the extreme probability of the load exceeding the preset safety threshold within the prediction window, and the second risk index is the maximum continuous duration for which the overload probability continuously exceeds the preset warning threshold.

[0215] The construction module 910 is used to construct a two-dimensional risk decision function based on the first risk indicator, the second risk indicator and the preset thermal inertia time threshold, so as to map the continuous probability space into discrete operating state levels.

[0216] The evaluation module 912 is used to determine the core load components that cause the risk of overload in the distribution network from multiple independent load subsequences based on the load condition probability distribution data corresponding to each independent load subsequence, in response to the triggering of a preset operating state level.

[0217] In an exemplary embodiment, the decomposition module 904 is specifically configured to, based on multi-user historical load data, classify users whose peak load is below a preset physical stratification threshold into a first user set, and users whose peak load is above the preset physical stratification threshold into a second user set; combine the original load data of the first user set as independent load subsequences; normalize the historical load curves of each user in the second user set to obtain a shape vector; cluster the shape vectors using the K-Means clustering algorithm, aggregating users with similar load curve shapes into the same subclass; recombine the original load data of the second user set according to the clustering labels to obtain several independent load subsequences; the independent load subsequences corresponding to the first user set and the several independent load subsequences corresponding to the second user set together constitute multiple independent load subsequences.

[0218] In an exemplary embodiment, the prediction model that decomposes linear probabilities is configured as follows:

[0219] The input total load sequence and multiple independent load subsequences are decoupled into trend components and periodic components using a moving average operation. The trend components and periodic components are then input into independent linear layers for feature mapping. The distribution data of the load conditional probability within the prediction window is output through a dual-channel probability head. The dual-channel probability head includes a mean channel and a variance channel. The mean channel is used to output the expected value of the predicted distribution data, and the variance channel outputs the standard deviation of the predicted distribution data after constraining the output of the linear layer through the Softplus activation function.

[0220] In an exemplary embodiment, the training process of a prediction model that decomposes linear probabilities includes:

[0221] Obtain a training sample set, which includes historical load data and corresponding actual load observation data; input the training sample set into the initial model, which outputs the mean and standard deviation of the predicted distribution data; calculate the loss value based on the negative log-likelihood loss function, which is used to simultaneously optimize the accuracy of the predicted mean and the reasonableness of the predicted standard deviation; adjust the parameters of the initial model according to the loss value until the model converges, thus obtaining a prediction model that decomposes linear probabilities.

[0222] In an exemplary embodiment, the processing module 906 is specifically used to input the original value of the linear layer output corresponding to the variance channel into the Softplus activation function; the Softplus activation function maps the original value to a positive real number, which is output as the standard deviation of the predicted distribution data; wherein, the Softplus activation function is used to ensure that the standard deviation of the output is always greater than zero.

[0223] In an exemplary embodiment, the construction module 910 is specifically configured to compare a first risk indicator with a preset warning threshold and a preset deterministic threshold, and compare a second risk indicator with a preset thermal inertia time threshold; the preset deterministic threshold is greater than the preset warning threshold; if the first risk indicator is less than the preset warning threshold, the operating state is determined to be a safe state; if the first risk indicator is greater than or equal to the preset warning threshold and less than the preset deterministic threshold, the operating state is determined to be a warning state; if the first risk indicator is greater than or equal to the preset deterministic threshold and the second risk indicator is less than the preset thermal inertia time threshold, the operating state is determined to be a warning state; if the first risk indicator is greater than or equal to the preset deterministic threshold and the second risk indicator is greater than or equal to the preset thermal inertia time threshold, the operating state is determined to be an overload state.

[0224] Each module in the aforementioned overload risk assessment device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0225] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores multi-user historical load data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements an overload risk assessment method.

[0226] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0227] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0228] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0229] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0230] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0231] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0232] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0233] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An overload risk assessment method, characterized in that, The method includes: Collect historical load data from multiple users within the transformer area of ​​the distribution network; The multi-user historical load data is decoupled to decompose the total load sequence into multiple independent load subsequences; The total load sequence and the multiple independent load subsequences are respectively input into the corresponding prediction model of decomposed linear probability to obtain the distribution data of load conditional probability within the prediction window; Based on the load conditional probability distribution data corresponding to the total load sequence, a first risk index and a second risk index are calculated; the first risk index is the extreme probability of the load exceeding the preset safety threshold within the prediction window, and the second risk index is the maximum continuous duration for which the overload probability continuously exceeds the preset warning threshold. Based on the first risk indicator, the second risk indicator, and the preset thermal inertia time threshold, a two-dimensional risk decision function is constructed to map the continuous probability space into discrete operating state levels. In response to triggering a preset operating state level, the core load components that cause the risk of overload in the distribution network are determined from multiple independent load subsequences based on the load condition probability distribution data corresponding to each independent load subsequence.

2. The method according to claim 1, characterized in that, The decoupling process for the multi-user historical load data, which decomposes the total load sequence into multiple independent load sub-sequences, includes: Based on the multi-user historical load data, users whose peak load is lower than the preset physical stratification threshold are classified into the first user set, and users whose peak load is higher than the preset physical stratification threshold are classified into the second user set. The original load data of the first user set is combined as an independent load subsequence, and the historical load curves of each user in the second user set are normalized to obtain a shape vector. The K-Means clustering algorithm is used to cluster the morphological vectors, grouping users with similar load curve shapes into the same subclass; Based on the clustering labels, the original load data of the second user set is recombined to obtain several independent load subsequences; The independent load subsequences are composed of the independent load subsequences corresponding to the first user set and the several independent load subsequences corresponding to the second user set.

3. The method according to claim 2, characterized in that, The prediction model that decomposes linear probabilities is configured as follows: The input total load sequence and the multiple independent load subsequences are decoupled into trend components and periodic components using a moving average operation; The trend component and the periodic component are respectively input into independent linear layers for feature mapping; The distribution data of load condition probabilities within the prediction window are output through a dual-channel probability header. The dual-channel probability header includes a mean channel and a variance channel. The mean channel is used to output the expected value of the predicted distribution data, and the variance channel outputs the standard deviation of the predicted distribution data after constraining the linear layer output through the Softplus activation function.

4. The method according to claim 3, characterized in that, The training process of the prediction model that decomposes linear probabilities includes: Obtain a training sample set, which includes historical load data and corresponding real load observation data; The training sample set is input into the initial model, and the initial model outputs the mean and standard deviation of the predicted distribution data. The loss value is calculated based on the negative log-likelihood loss function, which is used to simultaneously optimize the accuracy of the prediction mean and the reasonableness of the prediction standard deviation. The parameters of the initial model are adjusted based on the loss value until the model converges, resulting in a prediction model that decomposes linear probabilities.

5. The method according to claim 3, characterized in that, The distribution data of load condition probabilities within the prediction window output through the dual-channel probability head includes: The original value of the linear layer output corresponding to the variance channel is input into the Softplus activation function; The original values ​​are mapped to positive real numbers using the Softplus activation function, which are then output as the standard deviation of the predicted distribution data; wherein, the Softplus activation function is used to ensure that the standard deviation of the output is always greater than zero.

6. The method according to claim 1, characterized in that, The step of constructing a two-dimensional risk decision function based on the first risk indicator, the second risk indicator, and a preset thermal inertia time threshold, mapping a continuous probability space to discrete operating state levels, includes: The first risk indicator is compared with the preset warning threshold and the preset deterministic threshold, respectively, and the second risk indicator is compared with the preset thermal inertia time threshold; the preset deterministic threshold is greater than the preset warning threshold. If the first risk indicator is less than the preset warning threshold, the operating status is determined to be a safe state; If the first risk indicator is greater than or equal to the preset warning threshold and less than the preset deterministic threshold, the operating status is determined to be a warning status. If the first risk indicator is greater than or equal to the preset deterministic threshold and the second risk indicator is less than the preset thermal inertia time threshold, the operating state is determined to be a warning state. If the first risk indicator is greater than or equal to the preset deterministic threshold and the second risk indicator is greater than or equal to the preset thermal inertia time threshold, the operating state is determined to be an overload state.

7. An overload risk assessment device, characterized in that, The device includes: The data acquisition module is used to collect historical load data from multiple users within the transformer area of ​​the distribution network. The decomposition module is used to decouple the multi-user historical load data and decompose the total load sequence into multiple independent load subsequences. The processing module is used to input the total load sequence and the multiple independent load subsequences into the corresponding prediction model of decomposed linear probability, respectively, to obtain the distribution data of load conditional probability within the prediction window; The calculation module is used to calculate a first risk indicator and a second risk indicator based on the load conditional probability distribution data corresponding to the total load sequence; the first risk indicator is the extreme probability of the load exceeding the preset safety threshold within the prediction window, and the second risk indicator is the maximum continuous duration for which the overload probability continuously exceeds the preset warning threshold. The construction module is used to construct a two-dimensional risk decision function based on the first risk indicator, the second risk indicator and the preset thermal inertia time threshold, so as to map the continuous probability space into discrete operating state levels. The evaluation module is used to determine the core load components that cause the risk of overload in the distribution network from multiple independent load subsequences based on the load condition probability distribution data corresponding to each independent load subsequence, in response to the triggering of a preset operating state level.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images