A transformer area net load prediction method and system

Abstract

This application discloses a method and system for predicting net load in a power distribution area, relating to the field of power load prediction technology. The method includes: acquiring current core data for a target power distribution area, including current net load (NL) data and current temperature; constructing a dynamic NL map of the target power distribution area based on the current temperature; and using a preset extreme high-temperature weather NL prediction model, making a prediction based on the current NL data and the dynamic NL map to obtain the predicted NL value for the target power distribution area. The method of this application can improve the accuracy of power distribution area NL prediction.

Description

Technical Field

[0001] This application relates to the field of power load forecasting technology, and in particular to a method and system for forecasting net load in a transformer substation. Background Technology

[0002] Net load (NL) refers to the total load of the power system after deducting the output of renewable energy sources, representing the actual power demand that traditional power sources must bear. Against the backdrop of global climate change, extreme heatwaves are becoming increasingly frequent in China. Such weather leads to the concentrated use of cooling equipment such as air conditioners and industrial evaporative coolers, causing a surge in electricity demand in distribution areas. However, the power generation capacity of traditional power sources and photovoltaic (PV) systems is limited under extreme heat, exacerbating the contradiction between peak NL and system power supply capacity. Significant deviations in NL forecasting under extreme heat can easily lead to power supply and demand imbalances. Therefore, accurate NL forecasting under extreme heat is crucial for ensuring a stable power supply in distribution areas.

[0003] Currently, extreme high-temperature dynamic range (NL) predictions are generally based on the patterns of variation contained in historical NL data. However, the historical variation pattern prediction method is not very adaptable to nonlinear sudden NL fluctuations caused by extreme high temperatures, resulting in poor accuracy in NL predictions for power station areas. Summary of the Invention

[0004] To overcome the problem that existing historical variation pattern prediction methods are not very adaptable to nonlinear extreme high temperature sudden NL fluctuations, resulting in poor NL prediction accuracy for transformer areas, this application provides a transformer area net load prediction method and system.

[0005] Firstly, in order to solve the above-mentioned technical problems, this application provides a method for predicting the net load of a transformer substation, including: Obtain the current core data for the target area, which includes the current NL data and the current temperature; A dynamic NL map of the target transformer area is constructed based on the current temperature. Using a pre-defined extreme high-temperature weather NL prediction model, predictions are made based on current NL data and the current NL dynamic map of the current weather station area to obtain the predicted NL value of the target weather station area.

[0006] Secondly, this application also provides a net load forecasting system for transformer substations, comprising: The acquisition module is used to acquire the current core data of the target area, which includes the current NL data and the current temperature. The graph construction module is used to construct a dynamic NL graph of the target area based on the current temperature. The model prediction module is used to make predictions based on the current NL data and the current NL dynamic map of the current high temperature weather area using a preset extreme high temperature weather area NL prediction model, and to obtain the NL prediction value of the target area.

[0007] Thirdly, this application also provides a computing device, including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the above-described net load forecasting method for transformer areas.

[0008] Fourthly, this application also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the steps of the transformer area net load forecasting method.

[0009] The beneficial effects of this application are as follows: A dynamic NL map of the target substation is constructed based on the current temperature included in the current core data of the target substation. Then, using a pre-defined extreme high-temperature weather substation NL prediction model, predictions are made based on the current NL data included in the current core data and the current substation NL dynamic map to obtain the predicted NL value for the target substation at future times. In this way, by introducing the substation NL dynamic map of the target substation, the NL model can consider more comprehensive NL influencing factors during prediction, flexibly responding to the dynamic changes of the target substation, thereby improving the model's adaptability to nonlinear extreme high-temperature sudden NL fluctuations, and ultimately improving the accuracy of substation NL prediction. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an exemplary embodiment of a transformer area net load forecasting method. Figure 2 This is a flowchart illustrating the net load forecasting method for transformer substations provided in an exemplary embodiment of this application. Figure 3 This is a schematic diagram illustrating the structure of a transformer area net load prediction system, as shown in an exemplary embodiment of this application. Figure 4 This is a schematic diagram of the structure of a computer system of a computing device, illustrating an exemplary embodiment of this application.

[0011] The labels for each figure are as follows: 300-Net load forecasting system for transformer area, 301-Acquisition module, 302-Graph construction module, 303-Model forecasting module, 400-Computer system, 401-CPU, 402-ROM, 403-RAM, 404-Bus, 405-I / O interface, 406-Input section, 407-Output section, 408-Storage section, 409-Communication section, 410-Driver, 411-Removable media. Detailed Implementation

[0012] The following embodiments are further explanations and supplements to this application and do not constitute any limitation on this application.

[0013] Existing extreme high-temperature (NL) prediction methods fall into two categories: statistical prediction methods and machine learning prediction methods. Statistical prediction methods build models based on the statistical regularities of historical NL data, with the core assumption that the patterns of change inherent in historical NL will continue into the future. Based on this assumption, this method predicts future NL by analyzing the historical correlation between influencing factors and NL changes. For example, one NL prediction method based on multiple linear regression separates meteorologically sensitive NLs through trend analysis and then uses multiple linear regression to predict them. Another example is an NL prediction method based on quadratic exponential smoothing, which processes NL data using quadratic exponential smoothing to obtain the fluctuation characteristics of NL residuals, thereby improving NL prediction accuracy. Yet another example is an NL prediction method based on autoregressive integral moving average, which constructs multiple NL subsequence models using the autoregressive integral moving average algorithm to complete NL prediction. While these methods offer advantages such as strong interpretability and low computational cost, they are less adaptable to nonlinear, sudden fluctuations in extreme high-temperature NL, resulting in poor prediction accuracy.

[0014] Machine learning prediction methods automatically learn the mapping relationship between features and nonlinear loads (NLs) through data-driven approaches. They can uncover the nonlinear patterns of NLs without explicit mathematical assumptions, thus enabling NL prediction. For example, one NL prediction method based on deep random forests integrates meteorological and historical load data, using deep random forests to improve prediction accuracy. Another example is an NL prediction method based on Transformers, which uses the self-attention mechanism in the model to capture the time-dependent features of NL sequences and then completes the prediction through a feedforward neural network. Yet another example is an NL prediction method based on Long Short-Term Memory (LSTM) networks, which uses LSTM to analyze the impact of multiple energy sources on NLs at different time scales to achieve prediction. These methods automatically extract high-dimensional features through deep networks, capturing the long-term dependencies of NL sequences and are suitable for NL prediction in complex scenarios. However, the limited amount of extreme high-temperature NL data makes these methods prone to overfitting, and the model decision-making process is difficult to interpret, resulting in low prediction accuracy.

[0015] To address the aforementioned problems, embodiments of this application provide a method and system for predicting net load in a transformer substation, which will be described in detail below.

[0016] The net load forecasting method for a transformer substation provided in this application can be executed by a server. It should be noted that the server can be a standalone server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. No limitation is imposed here.

[0017] Please see Figure 1 , Figure 1 A method for predicting net load in a transformer substation is illustrated in an exemplary embodiment of this application, such as... Figure 1 As shown, this application provides a method for predicting net load in a transformer substation, including: S101, Obtain the current core data of the target area, which includes the current NL data and the current temperature; S102, Based on the current temperature, construct the current NL dynamic map of the target area; S103 uses a preset extreme high-temperature weather NL prediction model to make predictions based on current NL data and the current NL dynamic map of the current weather station area, and obtains the NL prediction value of the target weather station area.

[0018] The transformer substation net load forecasting method provided in this application constructs a dynamic NL map of the target substation based on the current temperature, which is included in the current core data of the target substation. Then, using a preset extreme high-temperature weather substation NL forecasting model, it predicts the NL value of the target substation at future times based on the current NL data and the current NL dynamic map included in the current core data. By introducing the dynamic NL map of the target substation, the NL model can consider more comprehensive NL influencing factors during forecasting, flexibly responding to the dynamic changes of the target substation, thereby improving the model's adaptability to nonlinear extreme high-temperature sudden NL fluctuations and ultimately improving the accuracy of substation NL forecasting. The current NL data includes real-time load rate, current voltage, or current current.

[0019] Optionally, a dynamic NL map of the target area is constructed based on the current temperature, including: Obtain the current topology of the target distribution area, which includes multiple distribution transformer nodes and multiple user meter box nodes; The static edge weights between nodes in the transformer area topology are calculated. The formula for calculating static edge weights is as follows:

[0020] in, Indicates the node v of the distribution transformerj With user table box node v k Static edge weights between them This represents the balance coefficient, used to adjust the contribution of the line length from the distribution transformer to the user's meter box to the correlation between line length and NL (Neutral Level). This represents the natural exponential function. Indicates the node v of the distribution transformer j With user table box node v k The length of the lines between them reflects the fundamental constraints of spatial physical connections. Indicates the attenuation coefficient of the line length. Indicates the node v of the distribution transformer j NL data and user table box node v k The Pearson correlation coefficient between the NL data, This represents the maximum Pearson correlation coefficient of all nodes in the target area, used to ensure the consistency of the dimensions of line attenuation; Based on the current transformer area topology and static edge weights, a static NL graph of the target transformer area is generated. The static NL map of the current transformer area is dynamically updated based on the current temperature to obtain the dynamic NL map of the current transformer area.

[0021] In the embodiment provided in this application, the static edge weights between nodes in the current topology of the target transformer area are calculated to form a static NL graph of the current transformer area containing the current topology and static edge weights. This static NL graph is then dynamically updated based on the current temperature to obtain a dynamic NL graph of the current transformer area. This introduces a time-varying temperature difference factor, ensuring that the dynamic NL graph of the current transformer area matches the real-time temperature dynamics of the target transformer area. This facilitates the subsequent incorporation of the dynamic NL graph into the NL model prediction, improving the model's adaptability to nonlinear, extreme high-temperature sudden NL fluctuations, thereby enhancing the accuracy of transformer area NL prediction.

[0022] Optionally, the static NL map of the current transformer area is dynamically updated based on the current temperature to obtain a dynamic NL map of the current transformer area, including: The static edge weights are dynamically updated based on the current temperature to obtain the corresponding dynamic edge weights; where the current temperature includes the first air temperature at the location of each distribution transformer node and the second air temperature at the location of each user meter box node. The formula for updating dynamic edge weights is as follows:

[0023] in, Indicates the node v of the distribution transformer j With user table box node v k Dynamic edge weights between them Indicates the node v of the distribution transformer j With user table box node v k Static edge weights between them This represents the natural exponential function. Indicates the node v of the distribution transformer j The first temperature at time t Represents the user's table box node v k The second temperature at time t This represents the temperature sensitivity coefficient of the transformer area, which is used to adjust the degree of influence of temperature differences in the transformer area on the correlation strength. The static graph of the current transformer area NL is updated by applying multiple dynamic edge weights to obtain the dynamic graph of the current transformer area NL.

[0024] In the embodiment provided in this application, the static edge weights are dynamically updated based on the current temperature to obtain the corresponding dynamic edge weights. The static NL map of the current transformer area is then updated based on multiple dynamic edge weights to obtain the dynamic NL map of the current transformer area. This allows the time-varying temperature difference factor to be introduced into the dynamic NL map of the current transformer area, making the dynamic NL map of the current transformer area match the real-time temperature dynamics of the target transformer area. This facilitates the subsequent incorporation of the dynamic NL map of the current transformer area into the NL model prediction, improving the model's adaptability to nonlinear extreme high-temperature sudden NL fluctuations, thereby improving the accuracy of transformer area NL prediction.

[0025] As can be seen, this embodiment accurately depicts complex correlation characteristics through a static and dynamic combination of transformer area map structures. The static transformer area NL map, constructed based on the line length attenuation effect and historical NL correlation, establishes the basic spatial correlation between transformer area nodes; and a dynamic map that dynamically updates edge weights using a real-time temperature difference factor is introduced to achieve adaptive adaptation to the temporal changes in node correlation strength under extreme high temperatures.

[0026] In an exemplary embodiment provided in this application, during extreme high-temperature weather NL analysis in a substation area, the graph structure can characterize the spatial and dynamic relationships of the substation area. Specifically, the static NL map of the substation area is a basic correlation structure constructed based on the substation area's topology and its historical NL correlation; the dynamic NL map of the substation area, based on the static NL map, introduces a time-varying temperature difference factor to obtain the dynamic impact of extreme high-temperature meteorological conditions on the substation area's NL.

[0027] Specifically, the static NL diagram of a distribution area uses the distribution transformers and user meter boxes of the area as nodes, with static edge weights. Used to characterize the distribution transformer V j To user box node v kThe static correlation strength between nodes can be determined by integrating the distance attenuation effect of the 0.4kV line with the historical NL correlation. Under extreme high-temperature conditions, the NL of a transformer substation is significantly affected by temperature, and the electrical correlation strength between nodes dynamically adjusts with real-time temperature changes. Therefore, the static edge weights of the transformer substation need to be corrected in real-time to obtain dynamic edge weights that are corrected over time. .

[0028] In this way, when the real-time temperature difference between the distribution transformer node and the user meter box node increases, the exponential term decays, resulting in a decrease in dynamic weight. This can reflect the operating condition characteristics of the larger the temperature difference and the weaker the load correlation under extreme high temperature weather, thereby realizing the time-series dynamic adaptation of topology weight.

[0029] Optionally, the extreme high temperature weather NL prediction model includes an input layer, a dynamic graph convolutional layer, a temporal feature extraction layer, and an output layer; Using a pre-defined extreme high-temperature weather NL prediction model, based on current NL data and the current NL dynamic map of the current weather station area, predictions are made to obtain the NL prediction value for the target weather station area, including: Using the input layer, the current NL data and the current NL dynamic map of the current transformer area are input into the NL prediction model for the extreme high temperature weather transformer area; By using a dynamic graph convolutional layer, spatial convolution operations are performed on the current NL data based on the current NL dynamic graph of the current transformer area to obtain the current spatial features of the target transformer area. The temporal feature extraction layer is used to extract features from the current spatial features to obtain the current temporal features of the target station area; The output layer is used to map and activate the current time-series features to obtain the NL prediction value of the target station area.

[0030] In the embodiment provided in this application, firstly, the input layer of the extreme high-temperature weather station area NL prediction model is used to input the current NL data and the current station area NL dynamic map into the model. Then, a dynamic map convolution layer is used to perform spatial convolution operations on the current NL data based on the current station area NL dynamic map to capture the dynamic correlation and temperature dynamic impact of the target station area in the spatial dimension, adapting to the dynamic changes in the topological spatial correlation between stations and obtaining the current spatial characteristics of the target station area. Secondly, a temporal feature extraction layer is used to extract features from the current spatial features to further capture the temporal dependence characteristics of the target station area, such as the continuous high-temperature accumulation effect in the time dimension, obtaining the current temporal features of the target station area in both the spatial and temporal dimensions. Finally, the output layer is used to map and activate the current temporal features to obtain the predicted NL value of the target station area. This allows for dynamic fusion prediction by combining the correlation features of the target station area in both the spatial and temporal dimensions, flexibly responding to the dynamic changes of the target station area, thereby improving the model's adaptability to nonlinear extreme high-temperature sudden NL fluctuations and ultimately improving the accuracy of station area NL prediction.

[0031] Optionally, the temporal feature extraction layer is provided with a gated loop unit, which includes an update gate, a reset gate, a candidate hidden state unit, and a hidden state unit; The temporal feature extraction layer is used to extract features from the current spatial features to obtain the current temporal features of the target station area, including: In the temporal feature extraction layer, the current spatial features are transformed using an update gate to obtain updated features; The updated features are transformed using a reset gate to obtain the reset features; The candidate hidden state features are obtained by transforming the reset features using candidate hidden state units. The candidate hidden state features are gated and fused using hidden state units to obtain the current temporal features of the target station area.

[0032] In the embodiment provided in this application, the update gate is used to determine the amount of old information to retain and the amount of new information to add; the reset gate is used to control the degree of integration of historical information; the candidate hidden state unit is calculated through the input and historical state, containing a mixed representation of new input and historical information; the hidden state unit is the output of the current time step and is passed to the next time step. Therefore, this embodiment utilizes the update gate, reset gate, candidate hidden state unit and hidden state unit included in the gated recurrent unit (GRU) in the temporal feature extraction layer to process the current spatial features in sequence, which can retain a certain amount of old information and add a certain amount of new information, and control the degree of integration of historical information, to obtain the current temporal features of the target area that integrate the mixed representation of new input and historical information, thereby capturing the temporal dependence characteristics of the target area such as the continuous high temperature accumulation effect in the time dimension, so as to adapt to the dynamic temporal changes of the target area, thereby improving the model's adaptability to nonlinear extreme high temperature sudden NL fluctuations, and thus improving the NL prediction accuracy of the area.

[0033] In this embodiment, the transformation processes performed in the update gate, reset gate, and candidate hidden state unit of the gated loop unit (GRU) can all include linear transformation and sigmoid nonlinear transformation (S-shaped nonlinear transformation).

[0034] In addition, the gating fusion process performed in the candidate hidden state unit of the gated recurrent unit (GRU) can be replaced by element-wise weighted summation or linear interpolation.

[0035] Optionally, the output layer includes a fully connected layer and a linear activation function; By mapping and activating the current temporal features using the output layer, the NL prediction value of the target station area is obtained, including: In the output layer, a fully connected layer is used to perform a linear transformation on the current temporal features to obtain the current linear features. The NL prediction value of the target area is obtained by activating the current linear features using a linear activation function.

[0036] In the embodiment provided in this application, a fully connected layer is used to linearly transform the current time-series features in the output layer to obtain the current linear features. Then, a linear activation function is used to activate the current linear features to obtain the NL prediction value of the target area. In this way, the processing of time-series features by the fully connected layer and the linear activation function can realize the linear regression mapping from features to target values ​​by combining the time-series features of the target area in both spatial and temporal dimensions. The output is a continuous and unbounded real number, which improves the interpretability of the model output parameters. At the same time, it can flexibly respond to the dynamic changes of the target area, thereby improving the model's adaptability to nonlinear extreme high-temperature sudden NL fluctuations, and thus improving the NL prediction accuracy of the area.

[0037] In one exemplary embodiment provided in this application, the extreme high-temperature weather NL prediction model is a deep learning model that integrates graph structure data and dynamic feature capture capabilities. It is used to process non-Euclidean structure data with dynamic correlation characteristics and can be represented as DGCNN (dynamic graph convolutional neural network). DGCNN introduces a dynamic mechanism on the basis of traditional graph neural network (GCN), and accurately captures the dynamic evolution of the correlation between nodes in the graph by adjusting the topology of the graph or the convolution kernel parameters in real time, thereby achieving efficient modeling of complex dynamic systems.

[0038] Specifically, the Extreme High Temperature Weather Zone NL Prediction Model (DGCNN) consists of four layers: an input layer, a dynamic graph convolutional layer, a temporal feature extraction layer, and an output layer. During model application, the input layer receives the current NL data; during model training, it receives extreme high temperature sample data containing multi-source features, including historical core data, historical auxiliary data, and historical spatial correlation data. The dynamic graph convolutional layer performs spatial convolution on the input data at each time step based on the zone's NL dynamic graph to capture the dynamic correlation between zones and obtain the spatial features of the zone. The temporal feature extraction layer uses a gated recurrent unit (GRU) based on the temporal dependency of the modeling load to obtain the current temporal features of the zone. The output layer outputs the predicted NL values ​​for future time periods through a fully connected layer.

[0039] The dynamic graph convolutional layer is based on the dynamic graph of the station area NL at each time t. Perform convolution operations to extract spatial features of the transformer area; dynamic NLR map of the transformer area. The structure can be represented as: ,in, Indicates the first A set of nodes in each transformer area; Indicates the first The set of nodes and edges in each platform area Indicates the first The combination of the topology of each transformer area and dynamic edge weights.

[0040] For example, a dynamic graph convolutional layer fuses the spatial correlation and dynamic temperature influence of a platform area through graph convolution operations. The features obtained from each layer processed by the dynamic graph convolutional layer... for:

[0041] in, Represents the first convolutional layer of a dynamic graph. Layer at time eigenmatrix Represents the first convolutional layer of a dynamic graph. Layer at time The feature matrix is ​​used to record the features of each node in the transformer area; the 0th layer feature... This represents the multi-source feature matrix input to the input layer, which contains the original observation data such as station NL and temperature at the current time t. Represents the symmetric normalization degree matrix. Normalization is used for graph convolution to ensure that the features of nodes in the platform area with different degrees are aggregated fairly. Represents the k-th dynamic adjacency matrix, derived from the dynamic graph. topology generate , This represents the dynamic correlation weight matrix between platforms. (as an identity matrix), used to integrate spatial correlation and dynamic temperature effects between stations; Indicates the first The first layer Each learnable convolutional kernel can generate dynamic weights through attention mechanisms or temporal modeling. To adapt to the topology update of the dynamic NL map of the transformer area; This indicates the number of convolutional kernels in DGCNN, used to control the multi-scale extraction capability of spatial features; This represents the activation function, used to introduce nonlinearity and enhance the expressive power of the model.

[0042] In this way, by extracting the time sequence and features from different levels through the dynamic graph convolutional layers of DGCNN, the outputs from different layers or at different times are concatenated along the feature dimension to form a NL spatial feature sequence, thereby aggregating the features of the transformer area nodes and ultimately outputting the spatial features of the transformer area. Simultaneously, by utilizing the dynamic adjacency matrix and dynamic convolutional kernel in the dynamic graph convolutional layers, it is possible to accurately capture the topological relationships, spatial connections, and dynamic temperature influences between transformer areas, providing a high-quality NL spatial feature sequence (the final output spatial features) for subsequent temporal feature extraction layers.

[0043] The temporal feature extraction layer inputs the NL spatial feature sequence (the final output spatial features) output from the dynamic graph convolutional layer into the gated recurrent unit (GRU) built into the gate, which is used to analyze the temporal dependency characteristics of NL under extreme high temperature weather. The GRU unit structure includes an update gate, a reset gate, candidate hidden state units, and hidden state units. The update gate is used to determine the amount of old information to retain and the amount of new information to add; the reset gate is used to control the degree of integration of historical information; the candidate hidden state unit is calculated from the input and historical states, containing a mixed representation of new input and historical information; the hidden state unit is the output of the current time step and is passed to the next time step.

[0044] GRU update gate output for: ,in, This represents the activation function. This indicates that the dynamic convolution kernel of the updated gate is in... The attention weight of a moment depends on time. This weight changes over time; that is, the dynamic convolutional layer adaptively adjusts the fusion method of the convolutional kernels according to the input features at different times. Represents the first dynamic convolutional layer Layer at time The feature matrix; GRU's reset gate output for: ; Output of GRU candidate hidden state units for: ,in, Indicates the dynamic convolution kernel of the candidate hidden state. Attention weight at any moment This represents element-wise multiplication; Indicates the activation function; Output of GRU hidden state unit for: in, express The output of the hidden state unit from the previous moment.

[0045] In the output layer, the temporal features output by the GRU are mapped to NL predictions through a fully connected layer and a linear activation function. for: , where linear is the linear activation function.

[0046] As shown above, under extreme high-temperature weather conditions, the load balancing (NL) of a transformer substation is not only affected by spatial correlations caused by topological relationships, but also by temporal dependence characteristics such as the cumulative effect of sustained high temperatures. To accurately capture these spatial and temporal correlation features and improve prediction accuracy, this application uses DGCNN for analysis and NL prediction. DGCNN effectively captures the spatial correlations and dynamic temperature effects between substations through dynamic graph convolutional layers, thereby adapting to the dynamic changes in the topological spatial correlations between substations. At the same time, it uses gated recurrent units (GRUs) to capture the temporal dependence characteristics of the substation load. Finally, the spatial and temporal features are fused to output the substation load prediction results. The organic combination of the two achieves deep fusion of spatial and temporal features, providing effective support for accurate prediction of substation NL under extreme high-temperature weather conditions, thereby improving the prediction accuracy of substation NL.

[0047] Optionally, the training process for the NL prediction model for extreme high-temperature weather zones is as follows: Acquire multiple extreme high temperature sample data of the target area, and the static map of the NL sample of the area corresponding to each extreme high temperature sample data; Based on the extreme high temperature sample data, the corresponding static map of NL sample in the transformer area is dynamically updated to obtain the corresponding dynamic map of NL sample in the transformer area. Using the initial prediction model, the training NL prediction value corresponding to each extreme high temperature sample data is obtained by processing the extreme high temperature sample data and the dynamic map of the NL sample of the transformer area; wherein, the initial prediction model is a prediction model obtained by pre-training on non-extreme high temperature sample data. Using a preset loss function, loss is calculated based on each training NL prediction value to obtain the corresponding adversarial training loss value. The initial prediction model is iteratively converged based on multiple adversarial training loss values ​​to obtain the NL prediction model for extreme high temperature weather zones.

[0048] In the embodiment provided in this application, an initial prediction model is used to process multiple extreme high-temperature sample data of the target area and the corresponding dynamic map of the area's NL samples. This yields a training NL prediction value corresponding to each extreme high-temperature sample data point. The initial prediction model is then iteratively converged using the adversarial training loss value corresponding to each training NL prediction value, calculated based on a preset loss function. This ensures that the trained extreme high-temperature weather NL prediction model meets the required prediction accuracy, thereby improving the NL prediction accuracy of the area. Simultaneously, because the dynamic map of the target area's NL samples is incorporated into the model training, the NL model can learn more comprehensive NL influencing factors during training, enabling it to flexibly respond to the dynamic changes of the target area. This improves the model's adaptability to nonlinear extreme high-temperature sudden NL fluctuations, further enhancing the NL prediction accuracy of the area.

[0049] Among them, the non-extreme high temperature sample data is sample data obtained by feature splicing based on the core features, auxiliary features and spatial correlation features of non-extreme high temperature dates. Non-extreme high temperature dates are characterized as dates in which the total duration of temperature greater than or equal to the preset temperature does not exceed the preset duration, or dates in which the temperature is always less than the preset temperature. The pre-training steps of the initial prediction model are the same as the training steps of the extreme high temperature weather station NL prediction model.

[0050] In this implementation, a domain discriminator D is used for adversarial training of the model to achieve convergence, and the corresponding loss function is... for:

[0051] in, This indicates the predicted NL loss; Represents the NL neighborhood classification loss; Indicates domain label; This represents the target domain NL concatenation feature values; Indicates the trade-off parameters; The domain probability output by the domain discriminator aims to distinguish whether features originate from the NL source domain (including multiple ordinary high temperature candidate data) or the NL target domain (including multiple extreme high temperature weather data). This is an NL feature extractor whose goal is to allow D to distinguish NL source features; it is achieved by minimizing... The feature extractor extracts features that are insensitive to the domain but useful for NL prediction tasks.

[0052] Specifically, and The calculation formula is as follows:

[0053]

[0054] in, Indicates the number of samples. This represents the total number of categories in the NL prediction task. Represented as the true NL value, This represents the training NL prediction value, where N is the number of samples. Indicates domain label, This represents the neighborhood probability output by the neighborhood discriminator. This represents the temporal features output by the feature extractor (temporal feature extraction layer).

[0055] In this embodiment, the initial prediction model is iteratively converged based on multiple adversarial training loss values ​​to obtain an extreme high-temperature weather NL prediction model. The specific steps are as follows: when multiple adversarial training loss values ​​do not meet the preset convergence conditions, the network parameters of the initial prediction model are adjusted to obtain an intermediate prediction model; extreme high-temperature sample data and dynamic graphs of NL samples in the NL region are input into the intermediate prediction model for processing to obtain a new adversarial training loss value corresponding to each extreme high-temperature sample data; when multiple new adversarial training loss values ​​meet the preset convergence conditions, the corresponding intermediate prediction model is determined as the extreme high-temperature weather NL prediction model.

[0056] The preset convergence condition is that the model has converged when the difference between the current round of adversarial training loss value and the previous round of adversarial training loss value for each extreme high temperature sample data is less than a set range.

[0057] Optionally, acquire multiple extreme high-temperature sample data of the target area, and a static map of the area's NL sample corresponding to each extreme high-temperature sample data, including: Acquire historical core data, historical auxiliary data, and historical spatial correlation data for multiple historical high-temperature dates in the target area; For each historical high-temperature date, feature concatenation is performed based on the historical core features, historical auxiliary features, and historical spatial correlation features of the historical high-temperature date to obtain the corresponding initial historical data. Multiple ordinary high-temperature candidate data that meet the preset requirements for the similarity of extreme high-temperature weather day characteristics are selected from multiple initial historical data, as well as multiple extreme high-temperature weather data; Data augmentation was performed based on multiple ordinary high-temperature candidate data and multiple extreme high-temperature weather data to obtain multiple extreme high-temperature sample data for the target area.

[0058] In this embodiment, multiple ordinary high-temperature candidate data that meet the preset similarity requirement of extreme high-temperature weather days are selected from multiple initial historical data obtained by splicing, and multiple extreme high-temperature weather data are used for data augmentation to obtain multiple extreme high-temperature sample data of the target area. This expands the extreme high-temperature sample set, effectively solving the problem of data scarcity. This addresses the issue of decreased prediction accuracy caused by the scarcity of NL samples in the area under extreme high temperatures, ensuring that the trained model meets the accuracy requirements and thus improving the prediction accuracy of NL in the area. The similarity requirement for extreme high-temperature weather days is that the similarity is less than a preset value.

[0059] The core historical data includes historical load factor (NL) data and historical temperature. Historical NL data includes historical load rate, historical voltage, or historical current. Historical temperature includes the temperature of the distribution area at various times on the corresponding historical dates, specifically the first temperature at the location of each distribution transformer node within the target distribution area, and the second temperature at the location of each user meter box node. Historical auxiliary data includes historical load type and historical date type. Historical spatial correlation data consists of spatial correlation features extracted from the distribution area topology map structure constructed through geographical distances and topological relationships between distribution areas.

[0060] In addition, historical high-temperature dates are defined as dates where the highest temperature is greater than or equal to a preset temperature. Historical core data includes historical NL data and historical temperatures. Historical high-temperature dates are divided into ordinary high-temperature dates and extreme high-temperature dates. Ordinary high-temperature dates are defined as dates where the total duration of temperatures greater than or equal to the preset temperature does not exceed the preset duration. Extreme high-temperature dates are defined as dates where the total duration of temperatures greater than or equal to the preset temperature exceeds the preset duration. Ordinary high-temperature candidate data consists of sample data from ordinary high-temperature dates. Extreme high-temperature weather data consists of sample data from extreme high-temperature dates.

[0061] In an exemplary embodiment provided in this application, data preprocessing is a crucial step in improving the accuracy of the prediction model when forecasting NL (Noise Level) in a substation during extreme high-temperature weather. Daily NL data for substations during extreme high-temperature weather are characterized by scarce samples, high volatility, and the influence of multiple coupled factors, making traditional processing methods insufficient for prediction requirements. Therefore, it is necessary to construct a comprehensive NL feature set through multi-source information fusion and expand the effective training data using a small-sample NL enhancement strategy to provide high-quality data input for the subsequent NL prediction model.

[0062] The core of NL prediction lies in accurately capturing the key factors that affect NL changes. By integrating multiple types and dimensions of feature inputs, it comprehensively depicts the NL change patterns under extreme high-temperature weather scenarios. Therefore, the multiple types and dimensions of features in this implementation include three categories: core features, auxiliary features, and spatial correlation features.

[0063] Specifically, the core features of NL forecasting are crucial in determining the underlying trend of NL changes and are also the core inputs for NL forecasting. These features mainly include historical NL data and meteorological data (such as the temperature parameters in this application). Historical NL data covers NL sequences under both ordinary and extreme high-temperature weather conditions, providing fundamental data support for analyzing key characteristics such as the periodicity and trend of NL changes over time, and is the core foundational data for NL forecasting. Meteorological data, with daily temperature changes as its core, indirectly affects the intensity of use of cooling equipment such as air conditioners and industrial evaporative coolers, ultimately influencing NL changes.

[0064] The auxiliary features of NL forecasting can further refine the forecast dimensions. The proportion of load type determines the weight of electricity consumption for each type of load, which can characterize the differences in electricity consumption characteristics of different transformer areas. Date types such as weekdays and weekends correspond to different electricity consumption behavior patterns. Generally, industrial electricity consumption accounts for a higher proportion on weekdays, while residential electricity consumption accounts for a higher proportion on weekends.

[0065] The spatial correlation features of NL prediction are obtained by constructing a topological map of the substations through the geographical distance and topological relationships between them. This enables the model to learn the NL transmission rules between substations and provides structural support for the NL prediction model based on DGCNN.

[0066] For example, for the NL prediction of the i-th transformer area at time t, the above features can be concatenated into initial historical data in the form of a feature vector. :

[0067] in, This represents the historical NL data feature vector of the i-th station area; This represents the historical temperature feature vector at time t; This is the historical load type feature vector of the i-th transformer area; Let i be the historical date type feature vector of the i-th station area; The spatial correlation features are extracted from the topology of the i-th transformer area.

[0068] In another exemplary embodiment provided in this application, the feature dimensions can be further expanded by introducing more meteorological data such as humidity and wind speed, as well as user electricity consumption behavior characteristics, to optimize the dynamic adjustment mechanism of graph structure construction and optimize the training of the NL prediction model for transformer substations in extreme high-temperature weather. This allows the NL model to learn more comprehensive NL influencing factors during training, thereby further improving the model's generalization ability and prediction accuracy under extreme weather conditions, flexibly responding to the dynamic changes in the transformer substation, and thus improving the model's adaptability to nonlinear extreme high-temperature sudden NL fluctuations, thereby improving the NL prediction accuracy of the transformer substation.

[0069] Optionally, data augmentation is performed based on multiple ordinary high-temperature candidate data and multiple extreme high-temperature weather data to obtain multiple extreme high-temperature sample data for the target area, including: Based on the similarity feature transfer of multiple extreme high temperature weather data, pseudo extreme high temperature weather data corresponding to each ordinary high temperature candidate data is obtained; For each extreme high-temperature weather data, data perturbation processing is performed on the extreme high-temperature weather data to obtain the corresponding perturbation sample data; Multiple extreme high-temperature weather data, multiple pseudo-extreme high-temperature weather data, and multiple disturbance sample data are used to form multiple extreme high-temperature sample data.

[0070] In the embodiment provided in this application, firstly, similar feature transfer is performed based on multiple extreme high-temperature weather data to obtain pseudo-extreme high-temperature weather data corresponding to each ordinary high-temperature candidate data. Then, for each extreme high-temperature weather data, data perturbation processing is performed based on the extreme high-temperature weather data to obtain corresponding perturbation sample data, thereby forming multiple extreme high-temperature sample data containing multiple extreme high-temperature weather data, multiple pseudo-extreme high-temperature weather data, and multiple perturbation sample data. In this way, through similar feature transfer and data perturbation processing, the extreme high-temperature sample set can be expanded, effectively solving the data scarcity problem. This addresses the problem of decreased prediction accuracy caused by the scarcity of NL samples in the substation area under extreme high temperatures, enabling the trained model to meet accuracy requirements and thus improving the prediction accuracy of NL in the substation area.

[0071] Among them, the static map of the NL sample corresponding to the sample data that has been augmented and deformed in multiple extreme high temperature sample data is the static map of the NL sample corresponding to the sample data before deformation.

[0072] In one exemplary embodiment provided in this application, extreme high-temperature days are rare, resulting in a scarcity of extreme high-temperature nocturnal (NL) samples, which can easily lead to overfitting when directly training the model. Although there are few NL samples for extreme high-temperature days, there are many NL samples for ordinary summer high-temperature days. Therefore, it is possible to find NL samples similar to extreme high-temperature days from ordinary high-temperature days and adjust their distribution to be similar to that of extreme days, thereby achieving similar feature transfer and generating pseudo-extreme high-temperature weather data.

[0073] Specifically, firstly, multiple candidate data sets of ordinary high-temperature days that meet the preset similarity requirements for extreme high-temperature weather days are selected from multiple initial historical data sets: In the historical high-temperature sample selection, based on the temperature feature similarity index, a sample set similar to the characteristics of extreme high-temperature weather days is selected from the samples of ordinary high-temperature weather days. The corresponding selection steps can be as follows: First, the NL source domain is defined as abundant ordinary high-temperature candidate data. The target domain for NL is extreme high-temperature weather data where data is scarce. .

[0074] Second, during the similarity sample screening process, the Euclidean distance between each sample in the NL source domain and the sample in the NL target domain on key temperature features is calculated. Then, based on the Euclidean distance, the DTW similarity of the temperature sequences between each ordinary high-temperature candidate data and each extreme high-temperature weather data is calculated as follows: ,in, Indicates length is Temperature time series data corresponding to ordinary high temperature candidate data. Indicates length is Temperature time series data corresponding to extreme high temperature weather data, The distance function is the Euclidean distance between two temperature time series data. In this embodiment, DTW (Dynamic Time Warping) is an algorithm for measuring the similarity between two time series. DTW constructs a distance equation between the two time series curves, uses slope range constraints, and calculates the time series path with the optimal cumulative distance. DTW can compare the similarity of time series of different lengths and has the advantage of low computational cost.

[0075] Third, the m samples with the smallest distance in the NL source domain are selected as ordinary high temperature candidate data to form a candidate set of extreme high temperature weather target domain.

[0076] Secondly, domain adaptation (DA) is used to transfer the NL source domain based on the NL target domain to achieve similar feature transfer from multiple extreme high-temperature weather data, resulting in pseudo-extreme high-temperature weather samples for each ordinary high-temperature candidate data set, which are then used to train the NL prediction model. The goal of DA is to minimize the squared estimate between the NL target domain and the NL source domain. The maximum mean discrepancy (MMD) of DA is used to measure the distribution difference between the source domain of the ordinary high-temperature day sample set and the target domain of the extreme high-temperature day sample set. In this embodiment, DA is a key technology in machine learning and pattern recognition. Its core objective is to solve the problem of mismatch between the NL source domain and the NL test data distribution. Through specific methods, the knowledge learned by the model in the source domain can be effectively transferred to the target domain, ultimately improving the model's generalization performance in the target domain.

[0077] Squared estimates between the NL target domain and the NL source domain The corresponding calculation formula is:

[0078] in, This represents the reproducing kernel Hilbert space (RKHS). Represents a mapping function. and These represent the number of samples in the NL source domain and the NL target domain, respectively. Indicates the first Source domain splicing feature values ​​of each station area; For the first The target domain splicing feature values ​​of each station area.

[0079] In another exemplary embodiment provided in this application, data perturbation generates new samples by applying small, physically constrained random perturbations to existing daily samples of extreme high-temperature weather. The data perturbation processing in this application may include temperature perturbation and NL data perturbation of the historical core data (historical temperature and historical NL data) included in the extreme high-temperature weather data. The steps of the data perturbation processing are as follows: Temperature disturbance: Temperature after disturbance for ; NL data perturbation: NL data after perturbation for .

[0080] The physical constraints for temperature perturbation and NL data perturbation are: , in, This represents the temperature disturbance factor. Indicates the first Historical NL data feature vectors for each station area For NL perturbation factor, For the first The highest historical temperature in each of the three districts; For the first The maximum load capacity (NL) for high-temperature scenarios in each transformer substation area.

[0081] Based on the disturbed temperature corresponding to this extreme high temperature weather data NL data Together with undisturbed historical auxiliary data and historical spatial correlation data, corresponding perturbed sample data are formed.

[0082] As shown above, a comprehensive NL feature set was constructed by integrating historical NL data, temperature and meteorological data, topological relationships, and other features. By combining DTW for similar sample selection, DA for generating pseudo-extreme high-temperature samples, and data perturbation techniques under physical constraints, the extreme high-temperature sample set was expanded, effectively solving the data scarcity problem and addressing the decline in prediction accuracy caused by the scarcity of NL samples in the substation area under extreme high temperatures.

[0083] Please see Figure 2 , Figure 2 This is a schematic flowchart illustrating the net load forecasting method for transformer substations provided in an exemplary embodiment of this application, as shown below. Figure 2 As shown, the extreme high-temperature weather net load (NL) prediction process based on dynamic graph convolution in this application consists of three parts: NL data preprocessing, NL graph structure construction, and NL prediction. The application steps of the NL prediction method are as follows: (1) Preprocessing of NL data in the transformer area.

[0084] First, historical core data such as NL (Neural Time) data and temperature for historical high-temperature days are obtained through systems such as electricity consumption information collection and numerical weather forecasting. Historical auxiliary data such as load type of users in the distribution area on historical high-temperature days are obtained through the marketing system. Historical spatial correlation data such as the topology of the distribution area on historical high-temperature days are also collected through the marketing system to obtain initial historical data, forming the NL input dataset. Second, multiple ordinary high-temperature candidate data that meet the preset similarity requirements of extreme high-temperature weather day features are selected from multiple initial historical data. NL samples with similar characteristics to extreme high-temperature days are selected from multiple ordinary high-temperature candidate data. Similarity feature transfer is performed using domain adaptation (DA) to transform them into pseudo-extreme high-temperature weather data to expand the training set. Finally, the temperature and NL data of the existing extreme high-temperature weather data selected from multiple initial historical data are slightly perturbed to generate new samples, resulting in perturbed sample data. Ultimately, multiple extreme high-temperature sample data are formed, consisting of multiple extreme high-temperature weather data, multiple pseudo-extreme high-temperature weather data, and multiple perturbed sample data, to enhance the small extreme high-temperature samples.

[0085] The initial prediction model was trained based on multiple extreme high temperature sample data to obtain the extreme high temperature weather NL prediction model (DGCNN).

[0086] Thus, firstly, static and dynamic graphs representing the spatial relationships of the substation area are constructed based on the substation topology and historical load curves. Secondly, using DGCNN, the graph structure at the current moment is dynamically learned and generated based on features such as real-time temperature and historical load, and an NL prediction model (DGCNN) for extreme high-temperature weather is established. Finally, a few-shot learning strategy is adopted, pre-training the model on a large amount of normal weather data and then fine-tuning it using a small number of extreme high-temperature samples, thereby improving the NL prediction accuracy under extreme high-temperature weather.

[0087] (2) Construction of the structure of the area map.

[0088] The distribution transformer area's topology is constructed using multiple distribution transformers and multiple user meter boxes as nodes. Static edge weights between nodes are calculated based on the spatial correlation strength between the distribution transformers and user meter boxes, forming a static NL graph for the distribution transformer area. On this basis, the static edge weights are dynamically updated by incorporating temperature-time correlation under extreme high temperatures, resulting in dynamic edge weights. These dynamic edge weights are then used to update the static NL graph of the distribution transformer area, yielding a dynamic NL graph.

[0089] (3) NL prediction for the Taiwan area.

[0090] The trained Extreme High Temperature Weather NL Prediction Model (DGCNN) includes an input layer, a dynamic graph convolutional layer, a temporal feature extraction layer, and an output layer.

[0091] First, based on the current core data of the transformer area input from the input layer, a dynamic graph convolutional layer is used to perform spatial convolution operations on the current core data based on the current transformer area's dynamic range (NL) graph corresponding to the current core data. This captures the spatial correlation of NL and the dynamic influence of temperature, thus obtaining the current spatial features of the transformer area. Second, the current spatial features output from the dynamic graph convolutional layer are input into the gated recurrent unit (GRU) built into the temporal feature extraction layer for feature extraction. This captures the temporal dependence of NL under extreme high temperatures, thus obtaining the current temporal features of the transformer area. Finally, the current temporal features are processed sequentially through the fully connected layer and linear activation function of the output layer to output the predicted NL value of the transformer area.

[0092] In summary, this application constructs a comprehensive NL feature set by integrating historical NL data, temperature and meteorological data, and topological relationships. By combining DTW for similar sample selection, DA for generating pseudo-extreme high-temperature samples, and data perturbation techniques under physical constraints, the extreme high-temperature sample set is expanded, effectively solving the data scarcity problem and addressing the decline in prediction accuracy caused by the scarcity of NL samples in the substation area under extreme high temperatures.

[0093] Meanwhile, the complex correlation characteristics are accurately depicted through a combination of static and dynamic distribution area diagrams. The static distribution area NL diagram, constructed based on the line length attenuation effect and historical NL correlation, establishes the basic spatial correlation between distribution area nodes; and a dynamic diagram that dynamically updates edge weights using real-time temperature difference factors is introduced to achieve adaptive adaptation to the temporal changes in node correlation strength under extreme high temperatures.

[0094] Furthermore, the DGCNN in this application effectively captures the spatial correlation and temperature dynamics between stations through dynamic graph convolutional layers, while GRU accurately mines the temporal dependency characteristics of NL. The organic combination of the two achieves deep fusion of spatial and temporal features, improving the prediction accuracy of NL.

[0095] Please see Figure 3 , Figure 3 An exemplary embodiment of this application illustrates a net load forecasting system for transformer substations, such as... Figure 3 As shown, this application provides a net load forecasting system 300 for transformer substations, comprising: The acquisition module 301 is used to acquire the current core data of the target area, which includes the current NL data and the current temperature. The graph construction module 302 is used to construct a dynamic NL graph of the target area based on the current temperature. The model prediction module 303 is used to make predictions based on the current NL data and the current NL dynamic map of the current high temperature weather area using a preset extreme high temperature weather area NL prediction model, and to obtain the NL prediction value of the target area.

[0096] The transformer substation net load forecasting system 300 provided in this application utilizes a graph construction module 302 to construct a dynamic NL map of the target substation based on the current temperature included in the current core data of the target substation acquired by the acquisition module 301. Then, a model prediction module 303 uses a preset extreme high-temperature weather substation NL prediction model to predict the NL value of the target substation at future times based on the current NL data included in the current core data and the current NL dynamic map. By introducing the dynamic NL map of the target substation, the NL model prediction can consider more comprehensive NL influencing factors, flexibly responding to the dynamic changes of the target substation, thereby improving the model's adaptability to nonlinear extreme high-temperature sudden NL fluctuations and ultimately improving the accuracy of substation NL prediction.

[0097] Optionally, the graph construction module 302 is specifically used for: Obtain the current topology of the target distribution area, which includes multiple distribution transformer nodes and multiple user meter box nodes; The static edge weights between nodes in the transformer area topology are calculated. Based on the current transformer area topology and static edge weights, a static NL graph of the target transformer area is generated. The static NL map of the current transformer area is dynamically updated based on the current temperature to obtain the dynamic NL map of the current transformer area.

[0098] Optionally, the graph construction module 302 is specifically used for: The static edge weights are dynamically updated based on the current temperature to obtain the corresponding dynamic edge weights. The static graph of the current transformer area NL is updated by applying multiple dynamic edge weights to obtain the dynamic graph of the current transformer area NL.

[0099] Optionally, the extreme high temperature weather NL prediction model includes an input layer, a dynamic graph convolutional layer, a temporal feature extraction layer, and an output layer; Model prediction module 303 is specifically used for: Using the input layer, the current NL data and the current NL dynamic map of the current transformer area are input into the NL prediction model for the extreme high temperature weather transformer area; By using a dynamic graph convolutional layer, spatial convolution operations are performed on the current NL data based on the current NL dynamic graph of the current transformer area to obtain the current spatial features of the target transformer area. The temporal feature extraction layer is used to extract features from the current spatial features to obtain the current temporal features of the target station area; The output layer is used to map and activate the current time-series features to obtain the NL prediction value of the target station area.

[0100] Optionally, the temporal feature extraction layer is provided with a gated loop unit, which includes an update gate, a reset gate, a candidate hidden state unit, and a hidden state unit; Model prediction module 303 is specifically used for: In the temporal feature extraction layer, the current spatial features are transformed using an update gate to obtain updated features; The updated features are transformed using a reset gate to obtain the reset features; The candidate hidden state features are obtained by transforming the reset features using candidate hidden state units. The candidate hidden state features are gated and fused using hidden state units to obtain the current temporal features of the target station area.

[0101] Optionally, the output layer includes a fully connected layer and a linear activation function; Model prediction module 303 is specifically used for: In the output layer, a fully connected layer is used to perform a linear transformation on the current temporal features to obtain the current linear features. The NL prediction value of the target area is obtained by activating the current linear features using a linear activation function.

[0102] Optionally, the transformer area net load prediction system 300 also includes a model training module, which is used to train a transformer area NL prediction model for extreme high temperature weather. The model training module is specifically used for: Acquire multiple extreme high temperature sample data of the target area, and the static map of the NL sample of the area corresponding to each extreme high temperature sample data; Based on the extreme high temperature sample data, the corresponding static map of NL sample in the transformer area is dynamically updated to obtain the corresponding dynamic map of NL sample in the transformer area. Using the initial prediction model, the training NL prediction value corresponding to each extreme high temperature sample data is obtained by processing the extreme high temperature sample data and the dynamic map of the NL sample of the transformer area; wherein, the initial prediction model is a prediction model obtained by pre-training on non-extreme high temperature sample data. Using a preset loss function, loss is calculated based on each training NL prediction value to obtain the corresponding adversarial training loss value. The initial prediction model is iteratively converged based on multiple adversarial training loss values ​​to obtain the NL prediction model for extreme high temperature weather zones.

[0103] Optionally, the model training module is specifically used for: Acquire historical core data, historical auxiliary data, and historical spatial correlation data for multiple historical high-temperature dates in the target area; For each historical high-temperature date, feature concatenation is performed based on the historical core features, historical auxiliary features, and historical spatial correlation features of the historical high-temperature date to obtain the corresponding initial historical data. Multiple ordinary high-temperature candidate data that meet the preset requirements for the similarity of extreme high-temperature weather day characteristics are selected from multiple initial historical data, as well as multiple extreme high-temperature weather data; Data augmentation was performed based on multiple ordinary high-temperature candidate data and multiple extreme high-temperature weather data to obtain multiple extreme high-temperature sample data for the target area.

[0104] Optionally, the model training module is specifically used for: Based on the similarity feature transfer of multiple extreme high temperature weather data, pseudo extreme high temperature weather data corresponding to each ordinary high temperature candidate data is obtained; For each extreme high-temperature weather data, data perturbation processing is performed on the extreme high-temperature weather data to obtain the corresponding perturbation sample data; Multiple extreme high-temperature weather data, multiple pseudo-extreme high-temperature weather data, and multiple disturbance sample data are used to form multiple extreme high-temperature sample data.

[0105] It should be noted that the transformer area net load forecasting system and the transformer area net load forecasting method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the transformer area net load forecasting system provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0106] A computing device according to an embodiment of this application includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements some or all of the steps of the above-described net load forecasting method for transformer areas.

[0107] The computing device can be a computer, and the corresponding program is computer software. The parameters and steps in the computing device described above can be referred to the parameters and steps in the embodiment of the net load forecasting method for transformer areas mentioned above, and will not be repeated here.

[0108] Figure 4 A schematic diagram of a computer system suitable for implementing the embodiments of this application is shown. It should be noted that... Figure 4 The computer system 400 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0109] like Figure 4 As shown, the computer system 400 includes a CPU 401 (Central Processing Unit), which can perform various appropriate actions and processes based on programs stored in ROM 402 (Read-Only Memory) or programs loaded from storage portion 408 into RAM 403 (Random Access Memory), such as performing the methods described in the above embodiments. The RAM 403 also stores various programs and data required for system operation. The CPU 401, ROM 402, and RAM 403 are interconnected via a bus 404. An I / O interface 405 (Input / Output interface) is also connected to the bus 404.

[0110] The following components are connected to I / O interface 405: an input section 406 including a keyboard, mouse, etc.; an output section 407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 409 performs communication processing via a network such as the Internet. A drive 410 is also connected to I / O interface 405 as needed. A removable medium 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 410 as needed so that computer programs read from it can be installed into storage section 408 as needed.

[0111] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by CPU 401, it performs various functions defined in the system of this application.

[0112] This application embodiment provides a computer-readable storage medium storing instructions that, when executed, perform the steps of the above-described transformer area net load forecasting method. The computer-readable storage medium can be either transient or non-transient.

[0113] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of this disclosure. The aforementioned computer-readable storage medium can be a non-transitory computer-readable storage medium, including: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, and other media capable of storing program code; it can also be a transient computer-readable storage medium.

[0114] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0115] Those skilled in the art will recognize that this application can be implemented as a system, method, or computer program product. Therefore, this disclosure can be implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "module" or "system." Furthermore, in some embodiments, this application can also be implemented as a computer program product contained in one or more computer-readable media, which contains computer-readable program code. Computer-readable storage media can be, for example, but not limited to—electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof.

[0116] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0117] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for predicting net load in a transformer substation, characterized in that, include: Obtain the current core data of the target area, including the current NL data and the current temperature; Based on the current temperature, a dynamic NL map of the target area is constructed. Using a preset extreme high-temperature weather NL prediction model, predictions are made based on the current NL data and the current NL dynamic map of the current weather station area to obtain the predicted NL value of the target weather station area.

2. The method according to claim 1, characterized in that, The process of constructing the current NL dynamic map of the target area based on the current temperature includes: Obtain the current topology of the target distribution area, which includes multiple distribution transformer nodes and multiple user meter box nodes; The static edge weights between nodes in the topology of the transformer area are calculated. Based on the current transformer area topology and the static edge weights, a static NL graph of the target transformer area is formed. The static NL map of the current transformer area is dynamically updated based on the current temperature to obtain the dynamic NL map of the current transformer area.

3. The method according to claim 2, characterized in that, The step of dynamically updating the static NL map of the current transformer area based on the current temperature to obtain the dynamic NL map of the current transformer area includes: Based on the current temperature, the static edge weight is dynamically updated to obtain the corresponding dynamic edge weight. The current transformer area NL static graph is updated based on multiple dynamic edge weights to obtain the current transformer area NL dynamic graph.

4. The method according to claim 1, characterized in that, The extreme high temperature weather NL prediction model for the weather zone includes an input layer, a dynamic graph convolutional layer, a temporal feature extraction layer, and an output layer. The method of using a preset extreme high-temperature weather NL prediction model to predict the NL value of the target weather station based on the current NL data and the current NL dynamic map of the weather station includes: Using the input layer, the current NL data and the current NL dynamic map of the current transformer area are input into the extreme high temperature weather transformer area NL prediction model; Using the dynamic graph convolutional layer, spatial convolution operation is performed on the current NL data based on the current station area NL dynamic graph to obtain the current spatial features of the target station area; The temporal feature extraction layer is used to extract features from the current spatial features to obtain the current temporal features of the target station area; The output layer is used to map and activate the current time-series features to obtain the NL prediction value of the target station area.

5. The method according to claim 4, characterized in that, The temporal feature extraction layer is provided with a gated loop unit, which includes an update gate, a reset gate, a candidate hidden state unit, and a hidden state unit. The step of extracting features from the current spatial features using the temporal feature extraction layer to obtain the current temporal features of the target station area includes: In the temporal feature extraction layer, the update gate is used to transform the current spatial features to obtain updated features; The reset gate is used to transform the updated feature to obtain the reset feature; The candidate hidden state unit is used to transform the reset feature to obtain the candidate hidden state feature; The candidate hidden state features are subjected to gated fusion processing using the hidden state unit to obtain the current temporal features of the target station area.

6. The method according to claim 4, characterized in that, The output layer includes a fully connected layer and a linear activation function; The step of mapping and activating the current temporal features using the output layer to obtain the NL prediction value of the target station area includes: In the output layer, the fully connected layer is used to perform a linear transformation on the current temporal features to obtain the current linear features; The current linear feature is activated using the linear activation function to obtain the NL prediction value of the target station area.

7. The method according to any one of claims 1 to 6, characterized in that, The training process of the extreme high-temperature weather NL prediction model for the area is as follows: Acquire multiple extreme high temperature sample data of the target area, and a static map of the NL sample of the area corresponding to each extreme high temperature sample data; Based on the extreme high temperature sample data, the corresponding static map of NL sample in the transformer area is dynamically updated to obtain the corresponding dynamic map of NL sample in the transformer area. Using the initial prediction model, the extreme high temperature sample data and the dynamic map of the NL sample of the transformer area are processed to obtain the training NL prediction value corresponding to each extreme high temperature sample data; wherein, the initial prediction model is a prediction model obtained by pre-training on non-extreme high temperature sample data. Using a preset loss function, loss is calculated based on each training NL prediction value to obtain the corresponding adversarial training loss value. The initial prediction model is iteratively converged based on multiple adversarial training loss values ​​to obtain the NL prediction model for extreme high temperature weather zones.

8. The method according to claim 7, characterized in that, The acquisition of multiple extreme high-temperature sample data of the target area, and the corresponding static map of the area's NL sample for each extreme high-temperature sample data, includes: Acquire historical core data, historical auxiliary data, and historical spatial correlation data for multiple historical high-temperature dates in the target area; For each historical high-temperature date, feature concatenation is performed based on the historical core features, historical auxiliary features, and historical spatial correlation features of the historical high-temperature date to obtain the corresponding initial historical data. Multiple ordinary high-temperature candidate data that meet the preset requirements for the similarity of extreme high-temperature weather day characteristics are selected from multiple initial historical data, as well as multiple extreme high-temperature weather data; Data augmentation is performed based on multiple ordinary high-temperature candidate data and multiple extreme high-temperature weather data to obtain multiple extreme high-temperature sample data for the target area.

9. The method according to claim 8, characterized in that, The data augmentation based on multiple ordinary high-temperature candidate data and multiple extreme high-temperature weather data yields multiple extreme high-temperature sample data for the target area, including: Based on the similarity feature transfer of multiple extreme high temperature weather data, pseudo extreme high temperature weather data corresponding to each ordinary high temperature candidate data is obtained; For each extreme high-temperature weather data, data perturbation processing is performed on the extreme high-temperature weather data to obtain the corresponding perturbation sample data; Multiple extreme high-temperature weather data, multiple pseudo-extreme high-temperature weather data, and multiple disturbance sample data are used to form multiple extreme high-temperature sample data.

10. A net load forecasting system for transformer substations, characterized in that, include: The acquisition module is used to acquire the current core data of the target station area, which includes the current NL data and the current temperature; The graph construction module is used to construct a dynamic NL graph of the target area based on the current temperature. The model prediction module is used to make predictions based on the current NL data and the current NL dynamic map of the current high temperature weather area using a preset extreme high temperature weather area NL prediction model, so as to obtain the NL prediction value of the target area.

Images