Power distribution network load pseudo measurement generation method in combination with weather and scene labels

Abstract

The invention discloses a power distribution network load pseudo measurement generation method in combination with meteorology and scene labels. The method comprises the steps of collecting historical data of multiple groups of different types of installed nodes in a power distribution network; constructing meteorological feature vectors corresponding to historical moments according to the meteorological data, and mapping time data of the historical moments into time period scene label vectors; mapping the user behavior pattern into a user behavior pattern label vector; constructing a multi-dimensional feature vector, and taking the historical load data actually measured by the corresponding node at the corresponding historical moment as the label of the corresponding multi-dimensional feature vector so as to obtain a training set; training the multi-task deep learning model by using samples in the training set; and obtaining an active power prediction value and a reactive power prediction value of the unloaded node at the prediction moment by using the multi-task deep learning model. By adopting the method provided by the invention, the accuracy and the robustness of the load pseudo measurement generation of the unloaded nodes can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of power distribution network data processing technology, specifically a method for generating pseudo-measurements of power distribution network load by combining meteorological and scene tags. Background Technology

[0002] As the end point of the power system, the accurate perception of the operating status of the distribution network is crucial for ensuring power supply reliability. Due to the numerous and widely distributed nodes in the distribution network, limitations in construction costs and communication conditions make it difficult to achieve real-time measurement equipment coverage for all nodes. Therefore, utilizing existing measurement data to generate high-precision pseudo-load measurement data for nodes without meters is a key step in achieving comprehensive observability and accurate state estimation of the distribution network.

[0003] Distribution network loads are influenced by a variety of factors, including weather conditions, temporal patterns, user behavior, and holidays, exhibiting strong nonlinear, volatile, and random characteristics. Current technologies typically employ data-driven deep learning algorithms to address this issue. For example, models such as Long Short-Term Memory (LSTM) networks or temporal convolutional networks are used to mine temporal dependencies in historical load data to establish load forecasting models and generate pseudo-measurement data.

[0004] However, the aforementioned existing technologies still have certain limitations in practical applications. First, most traditional deep learning models adopt a single-task learning mode, typically focusing only on the prediction of active power, often ignoring the inherent physical connection and strong coupling between active and reactive power, resulting in low accuracy of the generated pseudo-measured active power data. Second, existing methods are relatively crude in feature construction, usually directly inputting raw time data or meteorological data into the model, failing to fully explore the specific scene semantics behind the time period (such as morning and evening peak hours, valleys, etc.) and the differentiated impact of the combination of meteorological factors and specific scene labels on load changes, thus limiting the model's generalization ability and prediction accuracy in complex and variable scenarios. Summary of the Invention

[0005] To address the technical problem of low accuracy in existing methods for generating pseudo-measurement data for distribution networks, this invention provides a method for generating pseudo-measurement data of distribution network load that combines meteorological and scene tags, thereby effectively improving the accuracy and robustness of generating pseudo-measurement data for loads at nodes without meters installed.

[0006] This invention adopts the following technical solution: a method for generating pseudo-measurements of power distribution network load by combining meteorological and scene tags, comprising: Historical data of multiple sets of metered nodes of different types in the distribution network are collected. The historical data includes meteorological data, time data, user behavior patterns, holiday markers and historical load data of each metered node at multiple historical moments. Meteorological feature vectors corresponding to historical moments are constructed based on meteorological data, and the time data of historical moments are mapped to time period scene label vectors; user behavior patterns are mapped to user behavior pattern label vectors. A multidimensional feature vector is constructed based on the meteorological feature vector of historical time, the time period scene label vector, the user behavior pattern label vector, and the holiday marker. The historical load data measured at the corresponding node at the corresponding historical time is used as the label of the corresponding multidimensional feature vector to obtain the training set. The multi-task deep learning model is trained using samples from the training set. The input of the multi-task deep learning model is the multi-dimensional feature vector of the node at the corresponding time, and the output is the predicted active power and reactive power of the node. For nodes without meters installed, a multi-dimensional feature vector is constructed based on the meteorological data, time data, user behavior patterns, and holiday markers corresponding to the prediction time, and then input into the trained multi-task deep learning model to obtain the predicted active power and reactive power values ​​of the node at the prediction time.

[0007] Preferably, the types of nodes with installed tables include residential nodes, commercial / office nodes, and industrial production nodes.

[0008] Preferably, for node n, the expression for its meteorological feature vector at time t is: ; In the formula, , , as well as Let represent the normalized values ​​of temperature, relative humidity, solar irradiance, and wind speed at node n at time t, respectively. Let be the meteorological feature vector of node n at time t.

[0009] Preferably, the shared layer of the multi-task deep learning model adopts a long short-term memory network or a temporal convolutional network.

[0010] Preferably, the expression for the loss function L used to train the multi-task deep learning model is: ; In the formula, and , respectively, are the weights of the first and second tasks, MSE is the mean square error, and P and Q are the measured values ​​of node active power and node reactive power, respectively. and These represent the predicted active power and reactive power of the nodes, respectively.

[0011] Preferably, mapping historical time data to time period scene label vectors includes: Extract the hour from the timestamp; Define the scene meanings corresponding to different time intervals: 00:00-06:00 corresponds to the scene of late night / deep valley; 06:00-10:00 corresponds to the scene of morning rush hour; 10:00-16:00 corresponds to the scene of heavy traffic; 16:00-21:00 corresponds to the scene of evening rush hour; and 21:00-24:00 corresponds to the scene of night transition. Based on the hour of the current time, obtain the corresponding time period scene and map it to an integer index.

[0012] Preferably, the holiday flag is obtained by: obtaining the date information corresponding to the historical moment, determining whether it is a holiday, and setting the holiday flag to 1 if it is a holiday, otherwise setting it to 0.

[0013] Preferably, the multidimensional feature vector expression of node n at time t is: ; in: Let n be the meteorological feature vector of node n at time t; Let n be the scene label vector for node n at time t. Let n be the user behavior pattern label vector. It serves as a sign for holidays.

[0014] Preferably, it further includes: performing noise-perturbation-based data augmentation on the training set before training the multi-task deep learning model using samples from the training set, including: A normally distributed random noise is superimposed on the meteorological feature vector to form a new meteorological feature vector. ; The newly added meteorological feature vectors are used to obtain the newly added multidimensional feature vectors of the training samples. The set of the newly added multidimensional feature vectors and the original multidimensional feature vectors of the training samples is used as the enhanced training samples. The expression for the newly added multidimensional feature vector of node n at time t is: .

[0015] Preferably, data augmentation based on noise perturbation of the training set further includes: setting one of the terms in the multidimensional feature vector to zero with a preset probability, thereby obtaining a new multidimensional feature vector of the training samples.

[0016] The beneficial effects of this invention are as follows: by comprehensively considering multi-source heterogeneous data such as meteorological data, time scene labels, user behavior patterns, and holiday markers to construct multi-dimensional feature vectors, it can more comprehensively capture the complex factors affecting the load changes of the distribution network; at the same time, by using a multi-task deep learning model to predict active power and reactive power simultaneously, it effectively utilizes the strong coupling relationship between active and reactive power. Compared with traditional single-task prediction or methods based solely on historical load inference, it significantly improves the accuracy and robustness of generating pseudo-measurements for loads at nodes without meters, and effectively solves the problem of unobservable loads caused by insufficient distribution network measurement configuration. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart illustrating a method for generating pseudo-measurements of power distribution network load by combining meteorological and scene tags according to an embodiment of the present invention. Detailed Implementation

[0018] 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, not all, of the embodiments of the present invention. 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.

[0019] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] Example of a method for generating pseudo-measurements of power distribution network load by combining meteorological and scene tags: like Figure 1 As shown, the method for generating pseudo-measurements of power distribution network load combining meteorological and scene tags of the present invention includes: S101. Collect historical data of metered nodes in the distribution network, specifically: collect historical data of multiple sets of metered nodes of different types in the distribution network. The historical data includes meteorological data, time data, user behavior patterns, holiday markers and historical load data of each metered node at multiple historical moments.

[0021] In this embodiment, in order for the model to learn the load characteristics under different scenarios, the installed nodes should cover a variety of typical user behavior patterns. The types of installed nodes include, but are not limited to: Residential node data: Their load data exhibits a bimodal distribution, occurring in both morning and evening. Commercial / office node data: its load data reflects the peak characteristics during weekdays; Industrial production node data: Its load data reflects the characteristics of continuous or periodic production load.

[0022] By subdividing the metered nodes in the distribution network into different types such as residential, commercial / office, and industrial production, more refined feature extraction and model training can be performed for the specific electricity consumption behavior characteristics of different types of users. Compared with the general model that treats all nodes indiscriminately, this further improves the model's adaptability and prediction accuracy to different load characteristics (such as the difference between morning and evening peak hours and base load level).

[0023] S102. Obtain meteorological feature vectors, time period scene label vectors, and user behavior pattern label vectors. Specifically, construct meteorological feature vectors corresponding to historical moments based on meteorological data, and map the time data of historical moments to time period scene label vectors; map user behavior patterns to user behavior pattern label vectors. In this embodiment, meteorological data includes: temperature, relative humidity, solar irradiance, and wind speed at the location of the node. For node n, its location... The expression for the meteorological feature vector at time t is: ; In the formula, , , as well as Let represent the normalized values ​​of temperature, relative humidity, solar irradiance, and wind speed at node n at time t, respectively. Let be the meteorological feature vector of node n at time t.

[0024] By selecting four key meteorological factors—temperature, relative humidity, solar irradiance, and wind speed—and normalizing them to construct meteorological feature vectors, the influence of numerical differences between data of different dimensions is eliminated. Compared with schemes that use only single meteorological data or unprocessed data, this approach can more accurately reflect the comprehensive physical impact of meteorological conditions on power grid load, and improve the convergence speed and numerical stability of deep learning model training.

[0025] S103. Obtain the training set, specifically: construct a multi-dimensional feature vector based on the meteorological feature vector, time period scene label vector, user behavior pattern label vector, and holiday markers at historical moments, and use the historical load data measured at the corresponding historical moment of the corresponding node as the label of the corresponding multi-dimensional feature vector to obtain the training set. In this embodiment, the multidimensional feature vector expression of node n at time t is: ; in: Let n be the meteorological feature vector of node n at time t; Let n be the scene label vector for node n at time t. Let n be the user behavior pattern label vector. It serves as a sign for holidays.

[0026] In this embodiment, the composition structure of the multidimensional feature vector is clearly defined through standardized mathematical expressions. Meteorological features, time period scene labels, user behavior pattern labels, and holiday markers are organically integrated, providing the deep learning model with input data that is structurally clear, information-rich, and formatted in a unified manner. This ensures that the model can efficiently process multi-source features in parallel and accurately establish the mapping relationship between input features and output load.

[0027] S104. Train the multi-task deep learning model, specifically by using samples in the training set to train the multi-task deep learning model; the input of the multi-task deep learning model is the multi-dimensional feature vector of the node at the corresponding time, and the output is the predicted active power and reactive power of the node.

[0028] In this embodiment, the multi-task deep learning model includes an input layer, a shared layer, and a multi-task output layer. The input layer receives multi-dimensional feature vectors; the shared layer extracts the time-series dependencies between meteorological and behavioral features and outputs a high-dimensional hidden state vector; the multi-task output layer includes a first prediction head and a second prediction head. The first prediction head is used to regress the active power of the prediction node, and the second prediction head is used to regress the reactive power of the prediction node. Both the first and second prediction heads are fully connected layers.

[0029] The shared layer can be a Long Short-Term Memory (LSTM) network or a Temporal Convolutional Network (TCN).

[0030] Long Short-Term Memory (LSTM) networks are an improved type of recurrent neural network (RNN) used for processing and predicting sequential data. Their key feature is their ability to capture dependencies over long time spans, addressing the vanishing and exploding gradient problems common in traditional RNNs. LSTMs utilize a "gating mechanism" to control the input, retention, and forgetting of information, enabling the network to remember dependencies spanning tens or even hundreds of time steps.

[0031] Temporal convolutional networks (TCNNs) are deep convolutional network structures used for sequence modeling. They replace the recurrent mechanisms of RNNs / LSTMs with convolutional operations, enabling the memorization and feature extraction of long sequences. TCNNs construct feature extraction modules with long temporal dependencies through causal convolution and dilated convolution structures, enabling deep feature modeling of multi-dimensional time series inputs such as weather, time periods, and behaviors, thereby improving the fitting accuracy of pseudo-measurements from untabled nodes.

[0032] Using Long Short-Term Memory (LSTM) or Temporal Convolutional Network (TCN) in the shared layers of multi-task deep learning models can more effectively capture long-term temporal dependencies and complex temporal features in load data compared to ordinary fully connected networks or traditional RNNs. This results in better feature extraction capabilities when processing power load data with strong temporal and periodic characteristics, preventing gradient vanishing or exploding problems.

[0033] The expression for the loss function L used to train a multi-task deep learning model is as follows: ; In the formula, and , respectively, are the weights of the first and second tasks, MSE is the mean square error, and P and Q are the measured values ​​of node active power and node reactive power, respectively. and These represent the predicted active power and reactive power of the nodes, respectively.

[0034] By employing a mean squared error loss function that includes weighting coefficients, the prediction errors of both active and reactive power are constrained simultaneously. By adjusting the weighting coefficients to balance the learning weights of the two tasks, compared to training alone or unweighted multi-task loss, the optimization process of the two related tasks can be better coordinated, ensuring that the model can achieve better accuracy in the predicted values ​​of active and reactive power at the output node, and avoiding the dominance of a single task in model training.

[0035] S105. The active and reactive power of unmetered nodes are predicted using the trained multi-task deep learning model. Specifically, for unmetered nodes, a multi-dimensional feature vector is constructed based on the meteorological data, time data, user behavior patterns, and holiday markers corresponding to the prediction time, and then input into the trained multi-task deep learning model to obtain the predicted active and reactive power values ​​of the node at the prediction time.

[0036] This embodiment of the method abandons the strong dependence of traditional load forecasting methods on historical load data of target nodes. It can generate load data solely based on meteorological data, time data, user behavior patterns, and holiday markers, filling the blind spots in distribution network observation. Furthermore, through "user behavior pattern quantification" and the construction of a "multi-dimensional feature space," the model can understand the electricity consumption logic of different types of users, rather than simply fitting numerical values. Moreover, by employing a multi-task learning framework and utilizing the implicit physical correlation between active and reactive power (such as power factor characteristics), the physical consistency of the generated pseudo-measurement data is further improved.

[0037] In one embodiment, mapping historical time data to time period scene label vectors includes: S201. Extract time features: Extract the number of hours at the current moment from the timestamp; S202. Define the scene meanings corresponding to different time intervals, where the time interval 00:00-06:00 corresponds to the late night / deep valley scene; 06:00-10:00 corresponds to the morning rush hour scene; 10:00-16:00 corresponds to the midday rush hour scene; 16:00-21:00 corresponds to the evening rush hour scene; and 21:00-24:00 corresponds to the night transition scene.

[0038] S203. Based on the hour of the current time, obtain the corresponding time period scene and map it to an integer index.

[0039] In this embodiment, the mapping method is as follows: the integer index corresponding to late night / deep valley is 0, the integer index corresponding to morning peak is 1, the integer index corresponding to waist load is 2, the integer index corresponding to evening peak is 3, and the integer index corresponding to night transition is 4.

[0040] The day is divided into distinct time periods with clear physical meanings, such as late night / lower hours, morning peak, mid-load, evening peak, and nighttime transition. Time data is mapped to corresponding integer indices. Compared to directly using continuous linear time values, this discretized scene labeling can more intuitively reflect the differences in the operating status and load characteristics of the power grid at different times, helping the model to better understand and learn the load change patterns in the time dimension.

[0041] In one embodiment, the holiday flag is obtained by: obtaining the date information corresponding to the historical moment, determining whether it is a holiday, and setting the holiday flag to 1 if it is a holiday, otherwise setting it to 0.

[0042] By introducing a holiday flag, date information is converted into binary features, clearly distinguishing between weekdays and holidays. Compared with prediction methods that ignore special dates, this method can effectively capture sudden changes in load patterns caused by changes in user schedules and production plans during holidays, significantly reducing prediction errors for special dates and improving the model's generalization ability to atypical days.

[0043] In one embodiment, mapping a user behavior pattern to a user behavior pattern label vector includes: using a mapping rule to map the user behavior pattern to a user behavior pattern label vector, wherein different user behavior patterns in the mapping rule correspond to different behavior pattern label vectors.

[0044] In this embodiment, the behavior pattern label vector can be a One-Hot vector.

[0045] User behavior patterns include residential lifestyle patterns, commercial office patterns, industrial production patterns, new load patterns, public facility patterns, and mixed load patterns.

[0046] In one embodiment, the method further includes: performing noise-perturbation-based data augmentation on the training set before training the multi-task deep learning model using samples from the training set, including: S301. Random noise following a normal distribution is superimposed on the meteorological feature vector to form a new meteorological feature vector. The expression for the new meteorological feature vector of node n at time t is: , ; Represents random noise It follows a normal distribution with a mean of 0 and a variance of 0.05.

[0047] S302. Utilize the newly added meteorological feature vectors to obtain the newly added multidimensional feature vectors of the training samples, and use the set of the newly added multidimensional feature vectors and the original multidimensional feature vectors of the training samples as the enhanced training samples; wherein, the expression for the newly added multidimensional feature vector of node n at time t is: .

[0048] In this embodiment, by introducing a "noise disturbance data enhancement" mechanism, sensor errors and micro-environmental differences in the actual environment are simulated, making the generated pseudo-measurement data more stable when facing meteorological fluctuations and avoiding model overfitting.

[0049] Before training, data augmentation was performed by superimposing normally distributed random noise onto the meteorological feature vectors. This simulated the errors or fluctuations that may exist in meteorological measurements in the actual environment. Compared with training methods that only use the original collected data, this increased the diversity of training samples and significantly improved the model's generalization ability in the face of noise interference, making it more robust in practical applications.

[0050] In one embodiment, performing noise perturbation-based data augmentation on the training set further includes: By setting one of the terms in the multidimensional feature vector to zero with a preset probability, a new multidimensional feature vector of the training sample is obtained.

[0051] By setting one of the items in the multidimensional feature vector to zero with a preset probability for data augmentation (similar to the Dropout mechanism), abnormal situations such as missing data or sensor failure are simulated. This forces the model to make predictions without relying on a single feature. Compared with training with all features, this further prevents overfitting of deep learning models and enhances the model's adaptability to incomplete data in practical applications.

[0052] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for generating pseudo-measurements of distribution network load by combining meteorological and scene tags, characterized in that, include: Historical data of multiple sets of metered nodes of different types in the distribution network are collected. The historical data includes meteorological data, time data, user behavior patterns, holiday markers and historical load data of each metered node at multiple historical moments. Meteorological feature vectors corresponding to historical moments are constructed based on meteorological data, and the time data of historical moments are mapped to time period scene label vectors; user behavior patterns are mapped to user behavior pattern label vectors. A multidimensional feature vector is constructed based on the meteorological feature vector of historical time, the time period scene label vector, the user behavior pattern label vector, and the holiday marker. The historical load data measured at the corresponding node at the corresponding historical time is used as the label of the corresponding multidimensional feature vector to obtain the training set. The multi-task deep learning model is trained using samples from the training set. The input of the multi-task deep learning model is the multi-dimensional feature vector of the node at the corresponding time, and the output is the predicted active power and reactive power of the node. For nodes without meters installed, a multi-dimensional feature vector is constructed based on the meteorological data, time data, user behavior patterns, and holiday markers corresponding to the prediction time, and then input into the trained multi-task deep learning model to obtain the predicted active power and reactive power values ​​of the node at the prediction time.

2. The method for generating pseudo-measurements of distribution network load combining meteorological and scene tags as described in claim 1, characterized in that, The types of nodes that have been installed include residential nodes, commercial / office nodes, and industrial production nodes.

3. The method for generating pseudo-measurements of distribution network load combining meteorological and scene tags as described in claim 1, characterized in that, For node n, it is in The expression for the meteorological feature vector at time t is: ； In the formula, , , as well as Let represent the normalized values ​​of temperature, relative humidity, solar irradiance, and wind speed at node n at time t, respectively. Let be the meteorological feature vector of node n at time t.

4. The method for generating pseudo-measurements of distribution network load combining meteorological and scene tags as described in claim 1, characterized in that, The shared layers of multi-task deep learning models employ long short-term memory networks or temporal convolutional networks.

5. The method for generating pseudo-measurements of distribution network load combining meteorological and scene tags as described in claim 1, characterized in that, The expression for the loss function L used to train a multi-task deep learning model is as follows: ； In the formula, and , respectively, are the weights of the first and second tasks, MSE is the mean square error, and P and Q are the measured values ​​of node active power and node reactive power, respectively. and These represent the predicted active power and reactive power of the nodes, respectively.

6. The method for generating pseudo-measurements of distribution network load combining meteorological and scene tags as described in claim 1, characterized in that, Mapping historical time data into time period scene label vectors includes: Extract the hour from the timestamp; Define the scene meanings corresponding to different time intervals: 00:00-06:00 corresponds to the scene of late night / deep valley; 06:00-10:00 corresponds to the scene of morning rush hour; 10:00-16:00 corresponds to the scene of heavy traffic; 16:00-21:00 corresponds to the scene of evening rush hour; and 21:00-24:00 corresponds to the scene of night transition. Based on the hour of the current time, obtain the corresponding time period scene and map it to an integer index.

7. The method for generating pseudo-measurements of distribution network load combining meteorological and scene tags as described in claim 1, characterized in that, The holiday flag is obtained by: obtaining the date information corresponding to the historical moment, determining whether it is a holiday, and setting the holiday flag to 1 if it is a holiday, otherwise setting it to 0.

8. The method for generating pseudo-measurements of distribution network load combining meteorological and scene tags as described in any one of claims 1 to 7, characterized in that, The multidimensional eigenvector expression of node n at time t is: ； in: Let n be the meteorological feature vector of node n at time t; Let n be the scene label vector for node n at time t. Let n be the user behavior pattern label vector. It serves as a sign for holidays.

9. The method for generating pseudo-measurements of distribution network load combining meteorological and scene tags as described in claim 8, characterized in that, Also includes: Before training a multi-task deep learning model using samples from the training set, data augmentation based on noise perturbation is performed on the training set, including: A normally distributed random noise is superimposed on the meteorological feature vector to form a new meteorological feature vector. ; The newly added meteorological feature vectors are used to obtain the newly added multidimensional feature vectors of the training samples. The set of the newly added multidimensional feature vectors and the original multidimensional feature vectors of the training samples is used as the enhanced training samples. The expression for the newly added multidimensional feature vector of node n at time t is: 。 10. The method for generating pseudo-measurements of distribution network load combining meteorological and scene tags as described in claim 9, characterized in that, Data augmentation based on noise perturbation for the training set also includes setting a certain term in the multidimensional feature vector to zero with a preset probability, thereby obtaining a new multidimensional feature vector for the training samples.

Images