A transformer heavy overload risk grade prediction method, device, equipment and medium

Abstract

The application discloses a transformer heavy overload risk grade prediction method, device and equipment and medium, relates to the transformer heavy overload prediction technical field, and the method comprises the following steps: according to the holiday type and the festival category of target prediction period, holiday behavior semantic vector and festival activity semantic vector are obtained respectively;According to the holiday behavior semantic vector and the festival activity semantic vector, a periodic event combination effect vector is generated;According to the periodic event combination effect vector, the static attribute feature vector and the dynamic time sequence feature vector, the predicted load value of the target transformer is obtained by load prediction;According to the predicted load value and the transformer rated capacity, the heavy overload risk grade of the target transformer in the target prediction period is predicted.The application can solve the technical problems that the prior art encodes holidays and festivals as isolated independent labels respectively, and cannot capture the synergistic superposition effect generated when both occur simultaneously, resulting in inaccurate prediction.

Description

Technical Field

[0001] This invention relates to the field of transformer overload prediction technology, and in particular to a method, device, equipment and medium for predicting the risk level of transformer overload. Background Technology

[0002] As a core piece of equipment in the power grid, transformers are a major cause of accelerated aging, insulation damage, and even burnout accidents due to severe overload. With the rapid growth of seasonal electricity demand, such as air conditioning loads and electric vehicle charging loads, the overload risk faced by transformers during specific periods is becoming increasingly prominent. Accurately predicting the risk of transformer overload is of significant engineering practical importance for power grid dispatching departments to formulate load transfer plans in advance, for equipment operation and maintenance units to arrange timely inspections and troubleshooting, and for ensuring the reliability of power supply to residential and industrial users.

[0003] Currently, the main methods used in the power industry for predicting transformer overload include: statistical models based on time series analysis (such as ARIMA and exponential smoothing), which use extrapolated trends from historical load data for prediction; prediction methods based on traditional machine learning (such as support vector machines, backpropagation neural networks, and random forests), which use a limited number of features such as historical load, temperature, and holiday codes as input for load prediction; and time series prediction methods based on deep learning (such as LSTM and Transformer), which use historical load sequences for end-to-end time series modeling. In these methods, holiday information and solar term information are usually used as independent numerical labels or classification codes as input to the model. However, these coding methods share a common technical flaw: the model treats holidays and solar terms as independent variables that are not related, ignoring the potential synergistic effect when both occur simultaneously. For example, load increases when summer vacation occurs alone, and load also increases when the Great Heat (a solar term) occurs alone. However, when summer vacation and the Great Heat occur simultaneously, multiple factors such as student residential electricity consumption, high-temperature air conditioning load, and agricultural irrigation electricity consumption are superimposed, resulting in a load increase that is not a simple linear sum of their individual effects, but exhibits significant nonlinear abrupt changes. Because the model cannot establish a correlation between holidays and solar terms, its prediction results in the above combined scenarios often deviate significantly from the actual load, leading to false alarms or missed alarms of heavy overload risk. Summary of the Invention

[0004] This invention provides a method, device, equipment, and medium for predicting the risk level of transformer overload. It can solve the technical problem that existing technologies encode holidays and solar terms as isolated independent labels, which cannot capture the synergistic superposition effect when the two occur simultaneously, resulting in inaccurate prediction of load change phenomena in combined holiday and solar term scenarios.

[0005] In a first aspect, embodiments of the present invention provide a method for predicting the risk level of transformer overload, including: Obtain the holiday type and solar term category for the target prediction period, and based on the holiday type and solar term category, obtain the holiday behavior semantic vector and solar term activity semantic vector of the target transformer for the target prediction period, respectively. Based on the holiday behavior semantic vector and the solar term activity semantic vector, a periodic event input sequence is generated, and the periodic event input sequence is input into a pre-constructed interactive encoder so that the interactive encoder performs feature interaction on the periodic event input sequence to output a periodic event combination effect vector. Obtain the static attribute feature vector and dynamic time series feature vector of the target transformer, and perform load prediction based on the periodic event combination effect vector, static attribute feature vector and dynamic time series feature vector to obtain the predicted load value of the target transformer in the target prediction period. Obtain the rated capacity of the target transformer, and based on the predicted load value and rated capacity, predict the heavy overload risk level of the target transformer during the target prediction period.

[0006] This invention maps discrete holiday types and solar term categories into feature vectors rich in behavioral / activity semantics (personnel flow, energy consumption patterns, agricultural patterns, and climate characteristics), providing a semantic carrier for subsequent relationship modeling. By generating input sequences, the two are organized into a sequence structure that the model can process, enabling the subsequent interactive encoder to perform joint calculations on both. Then, through feature interaction via the interactive encoder, a dynamic relationship between holidays and solar terms is established, extracting the combined effect generated when the two act together, thereby capturing the synergistic superposition pattern when the two occur simultaneously, rather than treating them in isolation. By performing load forecasting based on three types of heterogeneous features of different properties, the synergistic effect of holidays and solar terms, along with equipment attributes and temporal patterns, is incorporated into the forecasting decision, correcting the prediction bias of traditional models in combined scenarios. By transforming the predicted load values ​​into executable risk assessment results, the output results have engineering applicability.

[0007] This invention utilizes holiday behavior semantic vectors and solar term activity semantic vectors to map holidays and solar terms into high-dimensional spaces carrying deep semantic information such as population movement and agricultural activities. This transforms them from isolated numbers into semantically rich entities, providing a foundation for subsequent models to understand their essential behavioral patterns. Based on this, the two are organized into a periodic event input sequence and processed by an interactive encoder. This establishes a dedicated channel for the model to calculate the correlation between the two, transforming them from independent input variables into interconnected collaborative objects. This allows the model to understand holidays and solar terms as combined events for the first time, learning the synergistic superposition patterns beyond their independent influences. Furthermore, the combined effect vector, static attribute features, and dynamic time-series features are subsequently incorporated into load forecasting. This allows the forecasting model to comprehensively consider all three information simultaneously when making load predictions, avoiding ignoring sudden superposition effects based solely on historical trends or blindly predicting sudden load changes based solely on combined effects. Finally, the risk assessment result is output based on the corrected predicted load value. Therefore, the embodiments of the present invention can solve the technical problem that the prior art encodes holidays and solar terms as isolated independent labels, which fails to capture the synergistic superposition effect when the two occur simultaneously, resulting in the model's inaccurate prediction of load change phenomena in the combined scenario of holidays and solar terms.

[0008] In some preferred embodiments of the first aspect, a periodic event input sequence is generated based on the holiday behavior semantic vector and the solar term activity semantic vector, specifically as follows: The event token sequence is obtained by combining the holiday behavior semantic vector and the solar term activity semantic vector in a preset order. Different positional codes are inserted into the holiday behavior semantic vector and the solar term activity semantic vector in the event token sequence to generate a periodic event input sequence.

[0009] This invention organizes two semantic vectors into a sequence structure according to a predetermined order, forming an event token sequence. This ensures that each event has a fixed position within the sequence. By inserting different position codes, different position identifiers are added to each semantic vector, enabling the model to clearly distinguish between holiday semantic vectors and solar term semantic vectors when processing the sequence. This allows the subsequent self-attention mechanism to accurately identify which event the query vector and key vector originate from when calculating association weights. Consequently, it calculates the association weights in both directions: from holiday to solar term and from solar term to holiday, providing an accurate computational basis for extracting the combined effect. Without this feature, the model would be unable to distinguish the origin of the two events, and the calculated association weights would lose their physical meaning.

[0010] In some preferred embodiments of the first aspect, feature interaction is performed on the input sequence of periodic events to output a vector of combined effects of periodic events, specifically as follows: The holiday behavior semantic vector and the solar term activity semantic vector in the periodic event input sequence are linearly transformed to generate their respective query vector, key vector and value vector; Calculate the first similarity between the query vector of the holiday behavior semantic vector and the key vector of the solar term activity semantic vector; calculate the second similarity between the query vector of the solar term activity semantic vector and the key vector of the holiday behavior semantic vector; and calculate the first association weight matrix between the holiday behavior semantic vector and the solar term activity semantic vector based on the first and second similarities. Based on the first association weight matrix, the value vectors of the holiday behavior semantic vector and the value vectors of the solar term activity semantic vector are weighted and aggregated to obtain the basic interaction feature vector; wherein, the basic interaction feature vector includes the holiday behavior semantic vector after incorporating solar term information and the solar term activity semantic vector after incorporating holiday information. The basic interaction feature vector is subjected to a nonlinear transformation to obtain a nonlinear interaction feature vector, and the nonlinear interaction feature vector is aggregated to generate and output a periodic event combination effect vector.

[0011] This invention projects two semantic vectors into a computationally calculable similarity space through linear transformation. Then, by calculating the first similarity (the similarity between a holiday query and a solar term key) and the second similarity (the similarity between a solar term query and a holiday key), the influence of holidays on solar terms and the influence of solar terms on holidays are quantified, respectively. An association weight matrix is ​​constructed, transforming the synergistic relationship from a fuzzy concept into a precisely calculable numerical matrix. Next, the value vectors are weighted and aggregated according to this matrix, so that the value vectors of holidays carry solar term information, and the value vectors of solar terms carry holiday information, completing a two-way information exchange between the two events. This transforms them from isolated semantic entities into composite entities carrying each other's information. Furthermore, a nonlinear transformation is used to perform nonlinear activation mapping on the above-mentioned basic interaction features, elevating them from independent representations carrying each other's information to a fused representation of their synergistic effect. Finally, the vectors are aggregated into a unified combined effect vector. These features work together to upgrade the model from simply viewing two events independently to quantifying the strength of their two-way influence and extracting their synergistic nonlinear combined effect, achieving a representation of the superimposed explosive pattern of 1+1>2.

[0012] In some preferred embodiments of the first aspect, the basic interaction feature vector is subjected to a nonlinear transformation to obtain a nonlinear interaction feature vector, specifically as follows: By using a preset self-attention mechanism, the correlation weights between each element in the basic interaction feature vector are calculated to obtain a second correlation weight matrix. Based on the second correlation weight matrix, the basic interaction feature vector is weighted and aggregated to obtain an intermediate interaction feature vector. The intermediate interactive feature vector is nonlinearly mapped to obtain a nonlinear interactive feature vector.

[0013] This invention employs a self-attention mechanism to calculate the correlation weights between elements within basic interactive features. This allows the model to automatically identify which combinations of components, already carrying information about each other, contribute most to load surges. Weighted aggregation then focuses the model on the interactions between these key components. Finally, nonlinear mapping performs cross-dimensional multiplicative recombination of intermediate interactive features, enabling the model to generate entirely new feature components that represent the simultaneous existence of both independent components, such as the triple-overlapping surge pattern of student residential electricity consumption, high-temperature air conditioning load, and agricultural irrigation electricity consumption. These technical features enable the model to express a synergistic effect where 1+1>2, thus completing the full extraction from information exchange to synergistic superposition representation.

[0014] In some preferred embodiments of the first aspect, load forecasting is performed based on the periodic event combination effect vector, static attribute feature vector, and dynamic time series feature vector to obtain the predicted load value of the target transformer during the target forecast period, specifically as follows: The combined effect vector of the periodic events, the static attribute feature vector, and the dynamic time series feature vector are fused to obtain the fused features; The fused features are input into a pre-trained machine learning model to enable the machine learning model to perform load prediction and obtain the predicted load value of the target transformer during the target prediction period.

[0015] This invention integrates three elements: a periodic event combination effect vector representing the combined effect, a static attribute feature vector representing the inherent properties of the equipment, and a dynamic time-series feature vector representing the real-time operating status. This allows the prediction model to simultaneously obtain information on the degree of synergistic impact, the equipment's capacity, and the current operating level when making load predictions. Furthermore, a pre-trained machine learning model jointly maps these three types of information into specific load prediction values, enabling them to complement and synergistically correct each other. This avoids blindly predicting sudden load changes based solely on the combination effect vector and avoids ignoring sudden superposition effects based solely on historical trends, ultimately improving the accuracy of load prediction values ​​in combined scenarios.

[0016] In some preferred embodiments of the first aspect, the heavy overload risk level of the target transformer during the target prediction period is predicted based on the predicted load value and rated capacity, specifically as follows: Based on the predicted load value and rated capacity, calculate the predicted load rate of the target transformer during the target prediction period; The predicted load rate is compared with preset heavy load thresholds and overload thresholds, and the heavy overload risk level of the target transformer during the target prediction period is determined based on the comparison results.

[0017] The embodiments of the present invention obtain risk warning results by accurately identifying the combined effects of holidays and solar terms. Compared with the risk false alarms caused by inaccurate predictions in traditional solutions, the risk warning accuracy is higher, and the complete transformation from synergistic effect identification to engineering executable risk signals is achieved.

[0018] In some preferred embodiments of the first aspect, based on the holiday type and solar term category, the holiday behavior semantic vector and solar term activity semantic vector of the target transformer during the target prediction period are obtained, respectively, including: The holiday types are mapped to holiday behavior semantic vectors by using a pre-trained holiday behavior knowledge embedding matrix; By using a pre-trained knowledge embedding matrix of solar term activities, the solar term categories are mapped to semantic vectors of solar term activities.

[0019] This invention, through pre-training the embedding parameters, ensures that the holiday behavior knowledge embedding matrix fully captures the patterns of population movement and residents' energy consumption habits corresponding to holidays during the learning process. Similarly, the solar term activity knowledge embedding matrix fully captures the patterns of agricultural production activities and climate characteristics corresponding to solar terms. By inputting discrete type labels into these embedding matrices, the output vector already condenses the behavioral and activity semantics behind the event. This means that even if subsequent combined scenario samples are sparse, the model can establish reasonable associations based on the existing semantic knowledge in the embedding vectors, significantly reducing the dependence on the number of combined scenario samples and improving training efficiency and prediction generalization ability.

[0020] Secondly, embodiments of the present invention provide a transformer overload risk level prediction device, including a semantic vector acquisition module, a combined effect acquisition module, a load prediction module, and a risk level acquisition module, wherein... The semantic vector acquisition module is used to acquire the holiday type and solar term category of the target prediction period, and obtain the holiday behavior semantic vector and solar term activity semantic vector of the target transformer in the target prediction period according to the holiday type and solar term category, respectively. The combined effect acquisition module is used to generate a periodic event input sequence based on the holiday behavior semantic vector and the solar term activity semantic vector, and input the periodic event input sequence into a pre-built interactive encoder so that the interactive encoder performs feature interaction on the periodic event input sequence to output a periodic event combined effect vector. The load forecasting module is used to obtain the static attribute feature vector and dynamic time series feature vector of the target transformer, and perform load forecasting based on the periodic event combination effect vector, static attribute feature vector and dynamic time series feature vector to obtain the predicted load value of the target transformer in the target forecasting period. The risk level acquisition module is used to acquire the rated capacity of the target transformer and, based on the predicted load value and rated capacity, predict the heavy overload risk level of the target transformer during the target prediction period.

[0021] This invention employs a semantic vector acquisition module to map discrete holiday types and solar term categories into feature vectors rich in behavioral / activity semantics (personnel flow, energy consumption patterns, agricultural patterns, and climate characteristics), providing a semantic carrier for subsequent relationship modeling. A combined effect acquisition module generates an input sequence, organizing both into a model-processable sequence structure, enabling the subsequent interactive encoder to perform joint calculations. The interactive encoder then interacts with features to establish a dynamic relationship between holidays and solar terms, extracting the combined effect generated when they act together, thus capturing the synergistic superposition pattern when they occur simultaneously, rather than treating them in isolation. A load prediction module performs load prediction based on three different types of heterogeneous features, incorporating the holiday-solar term synergistic effect with equipment attributes and temporal patterns into the prediction decision, correcting the prediction bias of traditional models in combined scenarios. Finally, a risk level acquisition module transforms the predicted load value into an executable risk assessment result, making the output practical for engineering applications.

[0022] Thirdly, embodiments of the present invention provide a terminal device, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction that causes the processor to perform the operation of the transformer overload risk level prediction method as described in any of the above.

[0023] Fourthly, embodiments of the present invention provide a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device or apparatus containing the computer-readable storage medium to perform the transformer overload risk level prediction method as described in any of the preceding claims.

[0024] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of a method for predicting the risk level of transformer overload provided in an embodiment of the present invention; Figure 2 This is a structural diagram of a transformer overload risk level prediction device provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1: like Figure 1 As shown, an embodiment of the present invention provides a method for predicting the risk level of transformer overload, comprising: S101, obtain the holiday type and solar term category of the target prediction period, and obtain the holiday behavior semantic vector and solar term activity semantic vector of the target transformer in the target prediction period according to the holiday type and solar term category respectively; In one specific embodiment, before obtaining the holiday type and solar term category for the target prediction period, it is necessary to establish a holiday behavior knowledge embedding matrix and a solar term activity knowledge embedding matrix. The embedding parameters are jointly trained by historical load data, meteorological data and periodic event information so that the embedding vector can characterize the population flow pattern, residential energy consumption pattern, agricultural production activity pattern and historical load change characteristics corresponding to the periodic events.

[0028] In this embodiment, based on the holiday type and the solar term category, the holiday behavior semantic vector and the solar term activity semantic vector of the target transformer in the target prediction period are obtained, respectively, including: mapping the holiday type to a holiday behavior semantic vector through a pre-trained holiday behavior knowledge embedding matrix; and mapping the solar term category to a solar term activity semantic vector through a pre-trained solar term activity knowledge embedding matrix.

[0029] In one specific embodiment, the current holiday category is mapped to a holiday behavior semantic vector and the current solar term category is mapped to a solar term activity semantic vector using the embedded matrix after training, providing input for subsequent Transformer combined effect modeling.

[0030] Specifically, holiday information is collected through the State Council's annual announcements, and dates are mapped to five types of codes: no holidays, weekends, short holidays, long holidays, and special days. After embedding the codes, the output is a dense vector of holidays. , as a component of the feature vector of periodic events.

[0031] Specifically, information on the 24 solar terms is collected through an astronomical calendar, and dates are mapped to solar term IDs (0-24). After embedding and encoding, the output is a dense vector of solar terms. It is used to capture load changes related to farming and traditional customs as a component of the periodic event feature vector.

[0032] S102, Based on the holiday behavior semantic vector and the solar term activity semantic vector, a periodic event input sequence is generated, and the periodic event input sequence is input into a pre-constructed interactive encoder so that the interactive encoder performs feature interaction on the periodic event input sequence to output a periodic event combination effect vector; In this embodiment, a periodic event input sequence is generated based on the holiday behavior semantic vector and the solar term activity semantic vector. Specifically, the holiday behavior semantic vector and the solar term activity semantic vector are combined in a preset order to obtain an event token sequence; different position codes are inserted into the holiday behavior semantic vector and the solar term activity semantic vector in the event token sequence to generate a periodic event input sequence.

[0033] In one specific embodiment, the acquisition of the event token sequence is specifically as follows: using the holiday dense vector (i.e., holiday behavior semantic vector) and the solar term dense vector (i.e., solar term activity semantic vector) as independent event tokens, and assembling them into an event token sequence according to a preset order. .

[0034] In one specific embodiment, the process of generating the input sequence of periodic events is as follows: Since the Transformer model itself does not contain event sequence information, in order to enable the model to distinguish the position attributes of different periodic events in the sequence, position encoding is introduced for holiday tokens and solar term tokens respectively.

[0035] Specifically, let the first position be encoded as The second position is encoded as Then, construct the enhanced event token representation: Further, the input sequence of periodic events is obtained: By assigning fixed location identifiers to different periodic events through location encoding, the Transformer can identify the source order of holiday events and solar term events. In the subsequent self-attention calculation process, it establishes the correlation between holiday behavioral features and solar term activity features, providing an input basis for extracting the combined effect of periodic events.

[0036] In this embodiment, feature interaction is performed on the input sequence of periodic events to output a combined effect vector of periodic events. Specifically, the following steps are taken: Linear transformations are performed on the semantic vectors of holiday behavior and solar term activities in the input sequence of periodic events to generate corresponding query vectors, key vectors, and value vectors; a first similarity is calculated between the query vector of the semantic vector of holiday behavior and the key vector of the semantic vector of solar term activities; a second similarity is calculated between the query vector of the semantic vector of solar term activities and the key vector of the semantic vector of holiday behavior; based on the first and second similarities, a first association weight matrix is ​​calculated between the semantic vector of holiday behavior and the semantic vector of solar term activities; based on the first association weight matrix, the value vectors of the semantic vectors of holiday behavior and solar term activities are weighted and aggregated to obtain a basic interaction feature vector; wherein, the basic interaction feature vector includes a semantic vector of holiday behavior after incorporating solar term information and a semantic vector of solar term activities after incorporating holiday information; a nonlinear transformation is performed on the basic interaction feature vector to obtain a nonlinear interaction feature vector, and vector aggregation is performed on the nonlinear interaction feature vector to generate and output a combined effect vector of periodic events.

[0037] In a preferred embodiment, the above-mentioned basic interaction feature vectors are obtained through a cross-attention mechanism.

[0038] In one specific embodiment, the process of obtaining the basic interaction feature vector is as follows: Let the input sequence be represented as: in, Represents holiday behavior tokens (i.e. holiday behavior semantic vectors). Represents the token of solar term activities (i.e., the semantic vector of solar term activities).

[0039] First, the query matrix Q, key matrix K, and value matrix V are generated through linear mapping: in, , and These represent the trainable weight matrices.

[0040] Subsequently, the scaling dot product attention mechanism is used to calculate the correlation weights between events: Here, A represents the correlation matrix of periodic events, and the elements in matrix A are used to characterize the degree of influence between different periodic events. Let represent the transpose of the key matrix K, where d is the dimension of the query vector and the key vector. This is a scaling factor used to control the size of the attention score to prevent gradient vanishing.

[0041] Further utilize the correlation matrix to perform weighted aggregation of event features: Obtain the event representation after the interaction (i.e., the basic interaction feature vector): in, This indicates the interactive features of holidays after integrating solar term information. This indicates the interactive features of the solar terms after integrating holiday information.

[0042] Preferably, the process of obtaining the above basic interaction feature vectors is performed in a pre-trained Transformer interaction encoder.

[0043] It should be noted that, unlike traditional feature splicing methods, the above interactive process enables each event feature to contain not only its own information, but also information about another periodic event, thereby establishing a dynamic relationship between holiday events and solar term events.

[0044] In this embodiment, the basic interaction feature vector is subjected to a nonlinear transformation to obtain a nonlinear interaction feature vector. Specifically, the association weights between each element in the basic interaction feature vector are calculated through a preset self-attention mechanism to obtain a second association weight matrix. Based on the second association weight matrix, the basic interaction feature vector is weighted and aggregated to obtain an intermediate interaction feature vector. The intermediate interaction feature vector is then nonlinearly mapped to obtain a nonlinear interaction feature vector.

[0045] In one specific embodiment, the basic interaction feature vector is subjected to a nonlinear transformation to obtain a nonlinear interaction feature vector. Specifically, the basic interaction feature vector is input into a second-layer transformer coding layer. The second-layer transformer coding layer includes a self-attention module and a feedforward network module, and performs the following processing on the input basic interaction feature vector in sequence.

[0046] The second-layer transformer encoding layer first inputs the basic interaction feature vector into the self-attention module. In the self-attention module, according to the self-attention mechanism, the basic interaction feature vector is mapped into a query matrix, a key matrix, and a value matrix, respectively. The similarity between the transpose of the query matrix and the key matrix is ​​calculated, and the calculation result is normalized to obtain a second association weight matrix. Each element in the second association weight matrix represents the degree of association between elements in the basic interaction feature vector, including the degree of association within the holiday behavior semantic vector after incorporating solar term information, the degree of association within the solar term activity semantic vector after incorporating holiday information, and the degree of mutual association between the two. Based on the second association weight matrix, the value matrix is ​​weighted and aggregated to output an intermediate interaction feature vector.

[0047] The intermediate interaction feature vector is then input into the feedforward network module of the second-layer transformer encoding layer. In the feedforward network module, the intermediate interaction feature vector is nonlinearly mapped by passing it through a fully connected layer and an activation function, thereby reorganizing the feature space of the intermediate interaction feature vector to extract the collaborative superposition pattern that transcends the independent information in the basic interaction features, and outputting a nonlinear interaction feature vector.

[0048] It should be noted that the first-layer converter coding layer and the second-layer converter coding layer cooperate with each other and progress step by step in terms of function, and together complete the complete extraction process from the original periodic event input to the combined effect vector. However, the objects of their operation and the data processing objectives are fundamentally different.

[0049] The first-layer transformer encoding layer receives a sequence of periodic events as initial input. The two elements in this sequence are a holiday behavior semantic vector and a solar term activity semantic vector, respectively, which do not yet carry information about each other. The first-layer transformer encoding layer calculates the bidirectional association weights between the two vectors through a cross-attention mechanism, enabling each vector to carry information about the other after weighted aggregation. The output basic interaction feature vector includes both the holiday behavior semantic vector and the solar term activity semantic vector, which incorporate solar term information. Essentially, this layer facilitates information exchange, establishing a direct relationship between two previously isolated semantic entities, thus enabling information sharing between them.

[0050] The second-layer transformer encoding layer receives the basic interactive feature vector output from the first layer. Each element of this vector carries partial information about the other. The second-layer transformer encoding layer first calculates the multidimensional correlation weights between elements within the sequence using a self-attention mechanism, performs weighted aggregation, and outputs an intermediate interactive feature vector. Then, a feedforward network module performs a nonlinear mapping on this intermediate interactive feature vector, outputting a nonlinear interactive feature vector. Essentially, this layer performs feature recombination; that is, based on the information exchange already completed in the first layer, it reorganizes and nonlinearly combines the two features that already carry information about each other, generating entirely new feature representations that transcend the independent information of each feature. This allows the extraction of the superimposed burst pattern generated by the synergistic effect of holidays and solar terms.

[0051] It should be noted that without the information exchange in the first layer, the self-attention mechanism in the second layer would deal with two unrelated independent vectors, making it impossible to establish a meaningful internal connection. Similarly, without the nonlinear recombination in the second layer, the information exchange in the first layer would only remain at the level of linear weighted superposition, failing to express the nonlinear synergistic effect where 1+1>2. The two layers are sequentially linked, from establishing connections to extracting synergies, jointly completing the generation of the combined effect vector of periodic events.

[0052] In one specific embodiment, the output of the second-layer Transformer is finally subjected to feature aggregation to obtain a periodic event combination effect vector: in, Let f(·) represent the vector of combined effects of periodic events, and f(·) represent the feature aggregation function.

[0053] S103, obtain the static attribute feature vector and dynamic time series feature vector of the target transformer, and perform load prediction based on the periodic event combination effect vector, static attribute feature vector and dynamic time series feature vector to obtain the predicted load value of the target transformer in the target prediction period. In one specific embodiment, it is necessary to collect static data and construct static feature vectors in advance, wherein the static data includes: Altitude: The altitude value is extracted from the DEM (Digital Elevation Model) using the latitude and longitude coordinates of the transformer installation location. After normalization, the output is a scalar value, which serves as the first component of the static feature vector.

[0054] Locational differences: By using the administrative division information and city level classification of the transformer installation location, the economic development level and load characteristic category of the region where the transformer is located are obtained. After embedding encoding, the output is an 8-dimensional dense vector, which serves as a component of the static feature vector.

[0055] User attributes: Type labels for the users supplied by the transformer, including residential users, commercial users, industrial users, mixed users, agricultural users, etc. After embedding encoding, the output is a 6-dimensional dense vector, which serves as a component of the static feature vector.

[0056] Intrinsic attributes: The inherent technical parameters of the transformer, including its years of operation, rated capacity, number of power supply users, number of heavy overload cycles, and power supply radius. After normalization, the output is a 5-dimensional vector, which serves as a component of the static feature vector.

[0057] After processing, the above four types of data are concatenated in sequence to form a complete static feature vector. The vector includes one component for altitude, eight components for location differences, six components for user attributes, and five components for the transformer's own attributes, totaling 20 dimensions, used to characterize the transformer's inherent properties.

[0058] In one specific embodiment, dynamic data collection and dynamic feature vector construction need to be performed in advance, wherein the dynamic data includes: Current operating data: The SCADA system collects three-phase current and voltage data of the transformer every 15 minutes and calculates the load rate. The four sampling points per hour are aggregated into an hourly average value, which is output as a scalar value as the first component of the dynamic time-series vector.

[0059] Historical benchmark: The median load rate at the same time in the past 4 weeks is taken as the historical benchmark. The difference between the current load rate and the benchmark is calculated and output as a scalar value to reflect the abnormal load fluctuations caused by external factors such as personnel mobility and special events. It serves as the second component of the dynamic time series vector.

[0060] Weather data: Hourly temperature, humidity, and rainfall data for the area where the transformer is located are collected via the meteorological bureau's API, and the perceived temperature is calculated. The output is a 3D vector (temperature, humidity, perceived temperature), which serves as a component of the dynamic time-series vector.

[0061] The dynamic data is organized chronologically, with each time step outputting a 5-dimensional vector (load rate, load rate offset, temperature, humidity, and perceived temperature). The historical window spans the past 7 days (168 hours), outputting a tensor of 168 rows x 5 columns. The future window spans 24 hours, outputting a tensor of 24 rows x 3 columns (temperature, humidity, and rainfall), used to characterize the temporal variation of the transformer.

[0062] It should be noted that the input sequence of the periodic events outputs 8 components for holiday data and 16 components for solar term data, totaling 24 dimensions. This vector is replicated and extended to the next 24 time steps, outputting a tensor of shape 24 rows × 24 columns, used to characterize the impact of periodic events on the load.

[0063] Preferably, the static feature vector needs to be pre-encoded using a pre-constructed static attribute encoder to obtain a static attribute feature vector. The process of obtaining the static attribute feature vector specifically involves the static attribute encoder receiving the aforementioned static feature vector as input, which is time-independent and remains unchanged throughout the prediction period. Therefore, a fully connected network is used for feature extraction and dimensionality transformation. The first layer has a 20-dimensional input and a 64-dimensional output, using the ReLU activation function. The second layer has a 64-dimensional input and a 128-dimensional output, also using the ReLU activation function. The third layer has a 128-dimensional input and a 128-dimensional output. The static attribute encoder outputs a dense 128-dimensional vector representing the inherent attribute features of the transformer.

[0064] in, , This represents the aforementioned static feature vector. This is the first layer weight matrix. This is the first layer bias vector.

[0065] in, , For the real number field, This is the weight matrix for the second layer. This is the bias vector for the second layer.

[0066] in, , , This is the third layer weight matrix. This is the bias vector for the second layer.

[0067] To facilitate subsequent fusion with dynamic data, the vector copy was extended to 24 time steps. (Static attribute feature vector), the output shape becomes [batch 24, 128].

[0068] Preferably, the dynamic feature vector needs to be encoded in advance using a pre-built dynamic temporal encoder to obtain the dynamic temporal feature vector. The process of obtaining the dynamic temporal feature vector is as follows: the input of the dynamic temporal encoder is divided into two parts: historical data (shaped [batch, 168, 5], including 7 days of load rate, load rate offset, temperature, humidity, and perceived temperature) and future weather (shaped [batch, 24, 3], including temperature, humidity, and rainfall in the next 24 hours).

[0069] Dynamic time-series data exhibits time dependence and periodicity. Therefore, a multi-scale temporal convolutional network (TCN) is employed to extract feature patterns along the time dimension. The short-time branch has a kernel size of 3, an expansion rate of 1, and a receptive field of approximately 7 hours, capturing short-term spikes and rapid changes; the medium-time branch has a kernel size of 5, an expansion rate of [1, 2, 4], and a receptive field of approximately 43 hours, capturing daily cycle patterns; and the long-time branch has a kernel size of 7, an expansion rate of [1, 4, 16], and a receptive field of approximately 337 hours, capturing weekly trends. The outputs of the three branches are concatenated along the feature dimension and then processed through a time attention mechanism, allowing the model to automatically focus on key moments within the historical window. The role of time attention: For example, when forecasting high temperatures in the future, the model will pay more attention to load patterns during historical high-temperature periods.

[0070] The input for each branch, Where B is the batch size, and short-time branching is used. China Times Branch: Long-term branches: Multi-scale splicing: After the time attention mechanism, the last time step is taken as the overall representation: The dynamic time-series encoder outputs two parts: historical data is transformed into a 256-dimensional vector, representing the overall load pattern of the past 7 days; future weather features are kept in their original shape [batch, 24, 3] and directly passed to the fusion module.

[0071] In this embodiment, load prediction is performed based on the periodic event combination effect vector, static attribute feature vector, and dynamic time series feature vector to obtain the predicted load value of the target transformer during the target prediction period. Specifically, the periodic event combination effect vector, static attribute feature vector, and dynamic time series feature vector are fused to obtain fused features. The fused features are then input into a pre-trained machine learning model to enable the machine learning model to perform load prediction and obtain the predicted load value of the target transformer during the target prediction period.

[0072] In one specific embodiment, the acquisition of the predicted load value involves concatenating the periodic event combination effect vector, static attribute feature vector, and dynamic time-series feature vector along the feature dimension to obtain a fused feature. The fused feature simultaneously includes information on the synergistic effect of holidays and solar terms, the inherent attribute information of the transformer, and historical time-series change information.

[0073] The fused features are input into a pre-trained load prediction model. The load prediction model is a machine learning model pre-trained using historical sample data. The load prediction model performs a non-linear mapping on the input fused features, mapping the fused features from a high-dimensional feature space to load prediction values, and outputs the predicted load value of the target transformer during the target prediction period.

[0074] Preferably, the specific number of dimensions and encoding method of static attribute feature vectors and dynamic time-series feature vectors can be flexibly set according to the actual application scenario.

[0075] Optionally, the load prediction model can adopt a fully connected network structure or other machine learning model structures that can achieve feature-to-prediction value mapping, without affecting the implementation of the present invention.

[0076] S104, obtain the rated capacity of the target transformer, and predict the heavy overload risk level of the target transformer during the target prediction period based on the predicted load value and rated capacity.

[0077] In this embodiment, the heavy overload risk level of the target transformer during the target prediction period is predicted based on the predicted load value and rated capacity. Specifically, the predicted load rate of the target transformer during the target prediction period is calculated based on the predicted load value and rated capacity; the predicted load rate is compared with preset heavy load thresholds and overload thresholds; and the heavy overload risk level of the target transformer during the target prediction period is determined based on the comparison results.

[0078] In one specific embodiment, based on the predicted load value and rated capacity, the heavy overload risk level of the target transformer during the target prediction period is predicted. Specifically, the rated capacity of the distribution transformer is assumed to be... The predicted load is Then calculate the predicted load factor: Furthermore, the heavy overload risk level is determined based on the predicted load rate: when When, it is determined to be an overload warning; when When, it is judged as a high-risk heavy load; when At that time, it was determined to be a risk of severe overload.

[0079] Further output of heavy overload prediction results: in, Indicates the predicted load. Indicates the predicted load factor. Indicates the risk level.

[0080] This invention maps discrete holiday types and solar term categories into feature vectors rich in behavioral / activity semantics (personnel flow, energy consumption patterns, agricultural patterns, and climate characteristics), providing a semantic carrier for subsequent relationship modeling. By generating input sequences, the two are organized into a sequence structure that the model can process, enabling the subsequent interactive encoder to perform joint calculations on both. Then, through feature interaction via the interactive encoder, a dynamic relationship between holidays and solar terms is established, extracting the combined effect generated when the two act together, thereby capturing the synergistic superposition pattern when the two occur simultaneously, rather than treating them in isolation. By performing load forecasting based on three types of heterogeneous features of different properties, the synergistic effect of holidays and solar terms, along with equipment attributes and temporal patterns, is incorporated into the forecasting decision, correcting the prediction bias of traditional models in combined scenarios. By transforming the predicted load values ​​into executable risk assessment results, the output results have engineering applicability.

[0081] This invention utilizes holiday behavior semantic vectors and solar term activity semantic vectors to map holidays and solar terms into high-dimensional spaces carrying deep semantic information such as population movement and agricultural activities. This transforms them from isolated numbers into semantically rich entities, providing a foundation for subsequent models to understand their essential behavioral patterns. Based on this, the two are organized into a periodic event input sequence and processed by an interactive encoder. This establishes a dedicated channel for the model to calculate the correlation between the two, transforming them from independent input variables into interconnected collaborative objects. This allows the model to understand holidays and solar terms as combined events for the first time, learning the synergistic superposition patterns beyond their independent influences. Furthermore, the combined effect vector, static attribute features, and dynamic time-series features are subsequently incorporated into load forecasting. This allows the forecasting model to comprehensively consider all three information simultaneously when making load predictions, avoiding ignoring sudden superposition effects based solely on historical trends or blindly predicting sudden load changes based solely on combined effects. Finally, the risk assessment result is output based on the corrected predicted load value. Therefore, the embodiments of the present invention can solve the technical problem that the prior art encodes holidays and solar terms as isolated independent labels, which fails to capture the synergistic superposition effect when the two occur simultaneously, resulting in the model's inaccurate prediction of load change phenomena in the combined scenario of holidays and solar terms.

[0082] Example 2: like Figure 2 As shown, this embodiment provides a transformer overload risk level prediction device, including a semantic vector acquisition module 201, a combined effect acquisition module 202, a load prediction module 203, and a risk level acquisition module 204, wherein... The semantic vector acquisition module 201 is used to acquire the holiday type and solar term category of the target prediction period, and obtain the holiday behavior semantic vector and solar term activity semantic vector of the target transformer in the target prediction period according to the holiday type and solar term category. In this embodiment, the semantic vector acquisition module 201 obtains the holiday behavior semantic vector and the solar term activity semantic vector of the target transformer during the target prediction period according to the holiday type and the solar term category, respectively. This includes: the semantic vector acquisition module 201 maps the holiday type to the holiday behavior semantic vector through a pre-trained holiday behavior knowledge embedding matrix; and maps the solar term category to the solar term activity semantic vector through a pre-trained solar term activity knowledge embedding matrix.

[0083] The combined effect acquisition module 202 is used to generate a periodic event input sequence based on the holiday behavior semantic vector and the solar term activity semantic vector, and input the periodic event input sequence into a pre-built interactive encoder so that the interactive encoder performs feature interaction on the periodic event input sequence to output a periodic event combined effect vector. In this embodiment, the combined effect acquisition module 202 generates a periodic event input sequence based on the holiday behavior semantic vector and the solar term activity semantic vector. Specifically, the combined effect acquisition module 202 combines the holiday behavior semantic vector and the solar term activity semantic vector in a preset order to obtain an event token sequence; and inserts different position codes into the holiday behavior semantic vector and the solar term activity semantic vector in the event token sequence to generate a periodic event input sequence.

[0084] In this embodiment, the combined effect acquisition module 202 performs feature interaction on the periodic event input sequence to output a periodic event combined effect vector. Specifically, the combined effect acquisition module 202 performs linear transformation on the holiday behavior semantic vector and the solar term activity semantic vector in the periodic event input sequence to generate their respective query vectors, key vectors, and value vectors; calculates the first similarity between the query vector of the holiday behavior semantic vector and the key vector of the solar term activity semantic vector; calculates the second similarity between the query vector of the solar term activity semantic vector and the key vector of the holiday behavior semantic vector; and calculates the second similarity between the query vector of the solar term activity semantic vector and the key vector of the holiday behavior semantic vector; based on the first similarity and the second similarity, the combined effect acquisition module 202 performs feature interaction on the periodic event input sequence to output a periodic event combined effect vector. The second similarity is calculated by obtaining a first association weight matrix between the holiday behavior semantic vector and the solar term activity semantic vector. Based on the first association weight matrix, the value vectors of the holiday behavior semantic vector and the value vectors of the solar term activity semantic vector are weighted and aggregated to obtain a basic interaction feature vector. The basic interaction feature vector includes the holiday behavior semantic vector after incorporating solar term information and the solar term activity semantic vector after incorporating holiday information. The basic interaction feature vector is subjected to a nonlinear transformation to obtain a nonlinear interaction feature vector, and the nonlinear interaction feature vector is aggregated to generate and output a periodic event combination effect vector.

[0085] In this embodiment, the combination effect acquisition module 202 performs a nonlinear transformation on the basic interaction feature vector to obtain a nonlinear interaction feature vector. Specifically, the combination effect acquisition module 202 calculates the correlation weights between each element in the basic interaction feature vector through a preset self-attention mechanism to obtain a second correlation weight matrix. Based on the second correlation weight matrix, the basic interaction feature vector is weighted and aggregated to obtain an intermediate interaction feature vector. The intermediate interaction feature vector is then nonlinearly mapped to obtain a nonlinear interaction feature vector.

[0086] The load prediction module 203 is used to obtain the static attribute feature vector and dynamic time series feature vector of the target transformer, and perform load prediction based on the periodic event combination effect vector, static attribute feature vector and dynamic time series feature vector to obtain the predicted load value of the target transformer in the target prediction period. In this embodiment, the load forecasting module 203 performs load forecasting based on the periodic event combination effect vector, static attribute feature vector, and dynamic time series feature vector to obtain the predicted load value of the target transformer during the target forecast period. Specifically, the load forecasting module 203 performs feature fusion on the periodic event combination effect vector, static attribute feature vector, and dynamic time series feature vector to obtain fused features; the fused features are input into a pre-trained machine learning model so that the machine learning model performs load forecasting to obtain the predicted load value of the target transformer during the target forecast period.

[0087] The risk level acquisition module 204 is used to acquire the rated capacity of the target transformer and, based on the predicted load value and rated capacity, predict the heavy overload risk level of the target transformer during the target prediction period.

[0088] In this embodiment, the risk level acquisition module 204 predicts the heavy overload risk level of the target transformer during the target prediction period based on the predicted load value and rated capacity. Specifically, the risk level acquisition module 204 calculates the predicted load rate of the target transformer during the target prediction period based on the predicted load value and rated capacity; compares the predicted load rate with preset heavy load thresholds and overload thresholds; and determines the heavy overload risk level of the target transformer during the target prediction period based on the comparison results.

[0089] For a more detailed explanation of the working principle and procedures of this embodiment, please refer to the relevant description in Embodiment 1.

[0090] In this embodiment of the invention, the semantic vector acquisition module 201 maps discrete holiday types and solar term categories into feature vectors rich in behavioral / activity semantics (personnel flow, energy consumption patterns, agricultural patterns, and climate characteristics), providing a semantic carrier for subsequent relationship modeling. The combined effect acquisition module 202 generates an input sequence, organizing the two into a sequence structure that the model can process, enabling the subsequent interactive encoder to perform joint calculations on both. Then, through feature interaction via the interactive encoder, a dynamic relationship between holidays and solar terms is established, extracting the combined effect generated when they act together, thereby capturing the synergistic superposition pattern when they occur simultaneously, rather than treating them in isolation. The load prediction module 203 performs load prediction based on three different types of heterogeneous features, incorporating the holiday-solar term synergistic effect with equipment attributes and temporal patterns into the prediction decision, correcting the prediction bias of traditional models in combined scenarios. Finally, the risk level acquisition module 204 transforms the predicted load value into an executable risk assessment result, making the output practical for engineering applications.

[0091] Example 3: This embodiment provides a terminal device, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction that causes the processor to perform the operation of the transformer overload risk level prediction method as described in any of the above.

[0092] Example 4: This invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device or apparatus containing the computer-readable storage medium to perform the transformer overload risk level prediction method as described in any of the above embodiments.

[0093] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0094] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting the risk level of transformer overload, characterized in that, include: Obtain the holiday type and solar term category for the target prediction period, and based on the holiday type and solar term category, obtain the holiday behavior semantic vector and solar term activity semantic vector of the target transformer for the target prediction period, respectively. Based on the holiday behavior semantic vector and the solar term activity semantic vector, a periodic event input sequence is generated, and the periodic event input sequence is input into a pre-constructed interactive encoder so that the interactive encoder performs feature interaction on the periodic event input sequence to output a periodic event combination effect vector. Obtain the static attribute feature vector and dynamic time series feature vector of the target transformer, and perform load prediction based on the periodic event combination effect vector, static attribute feature vector and dynamic time series feature vector to obtain the predicted load value of the target transformer in the target prediction period. Obtain the rated capacity of the target transformer, and based on the predicted load value and rated capacity, predict the heavy overload risk level of the target transformer during the target prediction period.

2. The method for predicting the risk level of transformer overload as described in claim 1, characterized in that, Based on the holiday behavior semantic vector and the solar term activity semantic vector, a periodic event input sequence is generated, specifically as follows: The event token sequence is obtained by combining the holiday behavior semantic vector and the solar term activity semantic vector in a preset order. Different positional codes are inserted into the holiday behavior semantic vector and the solar term activity semantic vector in the event token sequence to generate a periodic event input sequence.

3. The method for predicting the risk level of transformer overload as described in claim 2, characterized in that, The input sequence of periodic events is subjected to feature interaction to output a vector of combined effects of periodic events, specifically as follows: The holiday behavior semantic vector and the solar term activity semantic vector in the periodic event input sequence are linearly transformed to generate their respective query vector, key vector and value vector; Calculate the first similarity between the query vector of the holiday behavior semantic vector and the key vector of the solar term activity semantic vector; Calculate the second similarity between the query vector of the semantic vector of the solar term activities and the key vector of the semantic vector of the holiday behavior; and calculate the first association weight matrix between the semantic vector of the holiday behavior and the semantic vector of the solar term activities based on the first similarity and the second similarity. Based on the first association weight matrix, the value vectors of the holiday behavior semantic vector and the value vectors of the solar term activity semantic vector are weighted and aggregated to obtain the basic interaction feature vector; wherein, the basic interaction feature vector includes the holiday behavior semantic vector after incorporating solar term information and the solar term activity semantic vector after incorporating holiday information. The basic interaction feature vector is subjected to a nonlinear transformation to obtain a nonlinear interaction feature vector, and the nonlinear interaction feature vector is aggregated to generate and output a periodic event combination effect vector.

4. The method for predicting the risk level of transformer overload as described in claim 3, characterized in that, The basic interaction feature vector is subjected to a nonlinear transformation to obtain a nonlinear interaction feature vector, specifically: By using a preset self-attention mechanism, the correlation weights between each element in the basic interaction feature vector are calculated to obtain a second correlation weight matrix. Based on the second correlation weight matrix, the basic interaction feature vector is weighted and aggregated to obtain an intermediate interaction feature vector. The intermediate interactive feature vector is nonlinearly mapped to obtain a nonlinear interactive feature vector.

5. A method for predicting the risk level of transformer overload as described in any one of claims 1-4, characterized in that, Load forecasting is performed based on the combined effect vector of the periodic events, the static attribute feature vector, and the dynamic time-series feature vector to obtain the predicted load value of the target transformer during the target forecast period, specifically: The combined effect vector of the periodic events, the static attribute feature vector, and the dynamic time series feature vector are fused to obtain the fused features; The fused features are input into a pre-trained machine learning model to enable the machine learning model to perform load prediction and obtain the predicted load value of the target transformer during the target prediction period.

6. The method for predicting the risk level of transformer overload as described in claim 1, characterized in that, Based on the predicted load value and rated capacity, the heavy overload risk level of the target transformer during the target prediction period is predicted as follows: Based on the predicted load value and rated capacity, calculate the predicted load rate of the target transformer during the target prediction period; The predicted load rate is compared with preset heavy load thresholds and overload thresholds, and the heavy overload risk level of the target transformer during the target prediction period is determined based on the comparison results.

7. The method for predicting the risk level of transformer overload as described in claim 1, characterized in that, Based on the holiday type and solar term category, the holiday behavior semantic vector and solar term activity semantic vector of the target transformer during the target prediction period are obtained, including: The holiday types are mapped to holiday behavior semantic vectors by using a pre-trained holiday behavior knowledge embedding matrix; By using a pre-trained knowledge embedding matrix of solar term activities, the solar term categories are mapped to semantic vectors of solar term activities.

8. A device for predicting the risk level of transformer overload, characterized in that, It includes a semantic vector acquisition module, a combined effect acquisition module, a load prediction module, and a risk level acquisition module, among which, The semantic vector acquisition module is used to acquire the holiday type and solar term category of the target prediction period, and obtain the holiday behavior semantic vector and solar term activity semantic vector of the target transformer in the target prediction period according to the holiday type and solar term category, respectively. The combined effect acquisition module is used to generate a periodic event input sequence based on the holiday behavior semantic vector and the solar term activity semantic vector, and input the periodic event input sequence into a pre-built interactive encoder so that the interactive encoder performs feature interaction on the periodic event input sequence to output a periodic event combined effect vector. The load forecasting module is used to obtain the static attribute feature vector and dynamic time series feature vector of the target transformer, and perform load forecasting based on the periodic event combination effect vector, static attribute feature vector and dynamic time series feature vector to obtain the predicted load value of the target transformer in the target forecasting period. The risk level acquisition module is used to acquire the rated capacity of the target transformer and, based on the predicted load value and rated capacity, predict the heavy overload risk level of the target transformer during the target prediction period.

9. A terminal device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation of the transformer overload risk level prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device or apparatus containing the computer-readable storage medium to perform the transformer overload risk level prediction method as described in any one of claims 1 to 7.

Images