An anti-overfitting power prediction method for digital twin power grid

Abstract

The application discloses an anti-overfitting power prediction method for a digital twin power grid, comprising the following steps: inputting power network scheduling acceptance data of the digital twin power grid into a variational autoencoder for conversion to obtain mean and variance data, constructing a simulated Gaussian distribution, randomly sampling from the Gaussian distribution to obtain a feature, dividing the feature into two parts, denoted as z b Part and z p Part, using the feature to fit the power network scheduling acceptance data, using z b Part data to fit the bias input data, training through a generative adversarial network, thereby improving the independence of z b Part and z p Part, and using the finally output z p Part as an encoded feature of the removed bias input to input into a regression network, and then obtaining a power demand prediction value. The anti-overfitting power prediction method provided by the application robustly represents the training data, reduces the influence of historical data bias on the model, and improves the environmental adaptability of prediction.

Description

Technical Field

[0001] This invention relates to the field of automated management technology for network resource scheduling in digital twin power grids, and more particularly to a method for preventing overfitting in power forecasting for digital twin power grids. Background Technology

[0002] Digital twin power grids are a new technology proposed in recent years. By collecting comprehensive information data on the entire life cycle status of the power system, and adopting a multi-scale, multi-physical quantity, multi-space, and multi-disciplinary cross-integration method, a digital twin model of the power system is constructed. Using a digital twin computing platform and service platform, real-time mapping and interaction between the physical and virtual power systems are achieved, thereby enabling real-time improvement and updating of the real and virtual systems. This allows for real-time capture of the power system's operating status, real-time monitoring and prediction of potential risks, and other functions, ensuring that the power system operates in the best possible way.

[0003] The construction and management of power system twin data is an important component of digital twin power grids. Based on existing data, power companies utilize digital twin technology, integrating big data analytics, artificial intelligence, machine learning, and other techniques to predict regional electricity demand. Within the digital twin power network, they can rationally schedule and allocate power resources, optimizing the processes of power generation, transmission, transformation, distribution, and consumption. This results in efficient power resource scheduling and allocation schemes, which are then applied to the actual power system, thereby improving the overall operation and management effectiveness of the power system, reducing power losses, and increasing effective consumption.

[0004] However, current common forecasting methods are based on existing power grid data, designing and learning intelligent models, and then using these models to predict electricity consumption to assist in intelligent power dispatch. However, these methods suffer from inherent shortcomings of traditional machine learning methods, primarily the problem of overfitting in the prediction model. That is, the accuracy of the prediction model depends heavily on the collected data, leading to significant biases in the resulting predictions. For example, if a user's electricity consumption was high during a certain period based on historical data, the model might predict that the user will have high electricity demand in the future. While this data-driven forecasting method performs well when the data distribution remains unchanged, over-reliance on historical data can lead to overfitting. When historical data becomes biased or the data distribution changes, the prediction results become unreliable. For instance, due to technological upgrades (such as coal / oil-to-electricity conversion) or the promotion of new energy equipment, the user might require more electricity, or temporary shutdowns due to site maintenance might cause electricity consumption to be significantly lower than in previous years. Using the electricity consumption data from this period as model training input could easily result in inaccurate predictions or overfitting. If these biased models, or those overfitting to historical data, are used for power resource scheduling and allocation optimization in digital twin power networks, the resulting scheduling schemes will inevitably be unreasonable. Therefore, there is an urgent need to design an overfitting power forecasting method for digital twin power grids.

[0005] The above background information is provided only to assist in understanding the inventive concept and technical solution of this invention. It does not necessarily belong to the prior art of this patent application, nor does it necessarily provide technical teaching. In the absence of clear evidence that the above information was disclosed before the filing date of this patent application, the above background information should not be used to evaluate the novelty and inventiveness of this application. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides an overfitting-resistant power prediction method for digital twin power grids, the specific technical solution of which is as follows:

[0007] On the one hand, an overfitting-resistant power prediction method for digital twin power grids is provided, comprising the following steps:

[0008] A deep learning model is built using machine learning to process the power network dispatch data received by the digital twin power grid. The corresponding model input data (c, s, y) and model output data are provided. By definition, the power network dispatch data received by the digital twin power grid is divided into non-deviation input data and deviation input data. The power network dispatch data received by the digital twin power grid is denoted as x, c is non-deviation input data, s is deviation input data, c and s are distinguished by a predefined information stability, and y is the actual value of electricity demand. This is a forecast of electricity demand.

[0009] In the deep learning model, the power network dispatch data received by the digital twin power grid is input into a variational autoencoder for transformation to obtain mean and variance data. A simulated Gaussian distribution is constructed using the obtained mean and variance. Random sampling is performed from the Gaussian distribution to obtain a representational feature, denoted as z. This representational feature is then divided into two parts, denoted as zi and zj respectively. b Part and z p Part of the data involves using z to fit the power network dispatching data of the digital twin power grid, and using z... b Partial data is used to fit the biased input data, and then trained using a generative adversarial network to improve z-axis performance. b Part and z p Partial independence will affect the output z after training. p A portion of the biased data is used as a feature to input into the regression network to obtain the predicted electricity demand. The predicted electricity demand is then examined to determine whether it meets the preset requirements. If it does not meet the preset requirements, the deep learning model needs to be further optimized until the preset requirements are met.

[0010] Furthermore, the evaluation includes an evaluation of prediction accuracy and an evaluation of the degree of bias correction, wherein the evaluation of prediction accuracy is performed using the following formula:

[0011]

[0012] In the formula, A is the prediction accuracy, and y i It is the actual power demand value for the corresponding region i given by the dispatch automation system in the digital twin power grid. This is the predicted electricity demand value for region i given by the deep learning model. is the average of the actual electricity demand, and n is the total number of regions;

[0013] The degree of deviation correction is assessed using the following formula.

[0014]

[0015] In the formula, ΔEO represents the degree of deviation correction, and TP... s＝i The number of data points predicted as biased and actually representing biased data on type i data; FN s＝iIt refers to the number of data points predicted as biased but actually being unbiased on type i data; TP s＝j The number of data points predicted as biased and actually represented as biased data on class j data; FN s＝j It represents the number of data points predicted as biased but actually being unbiased on class j data.

[0016] Furthermore, the variational autoencoder includes a mean encoder and a variance encoder. Both the mean encoder and the variance encoder define a two-layer neural network, both use the ELU function as the inter-layer activation function, and both use the Tanh function as the tail layer activation function.

[0017] The conversion formula corresponding to the mean encoder is as follows:

[0018] u = Tanh(W u2 ×ELU(W u1 ×x))

[0019] The conversion formula corresponding to the variance encoder is as follows:

[0020] v = Tanh(W v2 ×ELU(W v1 ×x))

[0021] In the formula, u and v represent the mean and variance of the data, respectively, and W *i This represents the weight of the i-th network layer in the processing layer *, where b *i This represents the offset of the i-th network layer being processed, where * is either u or v.

[0022] Furthermore, z is used to fit the power network dispatching and receiving data of the digital twin power grid, and the corresponding fitting formula is as follows:

[0023]

[0024] Using z b To fit the biased input data to partial data, the corresponding fitting formula is as follows:

[0025]

[0026] In the formula, To receive data for power network dispatching in the fitted digital twin power grid, Input data for the fitted bias, W *i This represents the weight of the i-th network layer being processed, where * is either x or a.

[0027] Furthermore, z is used to fit the power network dispatching received data of the digital twin power grid to a neural network that reconstructs the unbiased input information at output. The loss function of the neural network that reconstructs the unbiased input information at output is calculated as follows:

[0028]

[0029] Using z b Partial data is used to fit the biased input data to a neural network that reconstructs the output biased input information. The loss function of the neural network for reconstructing the output biased input information is calculated as follows:

[0030]

[0031] Make z b Part and z p The loss functions that are independent of each other are calculated as follows:

[0032]

[0033] In the formula, p(x|z) is the decoding distribution, q(z|x) is the encoding distribution, p(z) is the prior representation of features, and p(s|z) is the encoding distribution. b ) is the distribution of the deviation information decoding, z bk It is the k-th prediction bias input representation dimension;

[0034] Based on the above three loss functions, the overall loss function for training representations can be derived as follows:

[0035] Loss = Loss VAE +Loss b +α×Loss z

[0036] In the formula, α is a parameter that controls the degree to which bias information is eliminated.

[0037] Furthermore, if the prediction accuracy meets the requirements but the bias correction does not, the value of α needs to be lowered and the learning optimization needs to be performed again; if the bias correction meets the requirements but the prediction accuracy does not, the value of α needs to be increased and the learning optimization needs to be performed again. The value of α is in the range of [0,1].

[0038] Furthermore, the regression network is denoted as f, and its corresponding loss function is as follows:

[0039]

[0040] in, It is the regression threshold.

[0041] Furthermore, the non-biased input data includes regional GDP, regional resident population, regional user count, regional power equipment count, regional transmission line length, regional transmission line type, and historical electricity demand; the biased input data is potential historical biased input information, which includes defined explicit biased data, such as specific information related to electricity consumption during special periods, such as the construction of new industrial parks, the demolition of factories, and new energy subsidies.

[0042] In another aspect, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the overfitting power prediction method for digital twin power grids.

[0043] In another aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the overfitting power prediction method for digital twin power grids as described in any one of the claims.

[0044] Compared with the prior art, the present invention has the following advantages: it uses variational autoencoders and generative adversarial networks to robustly represent training data, reduces model overfitting, and reduces the impact of historical data bias on the model, thereby improving the environmental adaptability of prediction. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the variational autoencoding data processing process in the anti-overfitting power prediction method for digital twin power grids provided in this embodiment of the invention;

[0046] Figure 2 This is an overall training flowchart of the anti-overfitting power prediction method for digital twin power grids provided in this embodiment of the invention;

[0047] Figure 3 This is a flowchart of the method for removing bias information in the anti-overfitting power prediction method for digital twin power grids provided in an embodiment of the present invention. Detailed Implementation

[0048] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0049] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0050] To address the problems in the background technology, this invention provides robust representation of training data to reduce model overfitting and mitigate the impact of historical data bias on the model. Deep learning is a machine learning method that uses multi-layer neural networks for data fitting, prediction, and analysis. Representation learning is a data processing method in deep learning that encodes features, extracting potential information from the data by transforming its mapping space. Current intelligent methods used in the dispatch automation system of digital twin power grids ignore the overfitting effect caused by unstable factors in historical data. Variational autoencoders, on the other hand, are unsupervised data generation models that encode input data into a representation space and then reconstruct the data from that space. The encoding process not only encrypts the data but also obtains more independent representation features between columns. These representation features essentially contain all the information of the original data, including information required for power network system dispatching in digital twin power grids and information identifying certain regions. To reduce the impact of overfitting and efficiently utilize it for power resource dispatching in digital twin power grids, it is essential to remove bias information from the model training data.

[0051] In one embodiment of the present invention, an overfitting-prediction method for digital twin power grids is provided, comprising the following steps:

[0052] Step 1: Mathematical Modeling

[0053] A deep learning model is built using machine learning to process the power network dispatch data received by the digital twin power grid. The corresponding model input data (c, s, y) and model output data are provided. By definition, the power network dispatch data received by the digital twin power grid is divided into non-deviation input data and deviation input data. The power network dispatch data received by the digital twin power grid is denoted as x, c is non-deviation input data, s is deviation input data, c and s are distinguished by a predefined information stability, and y is the actual value of electricity demand. This is a forecast of electricity demand.

[0054] For example, the formulation and execution of annual contracted electricity volume can be used to provide the corresponding model input data;

[0055] Unbiased input information c: Input data used for electricity demand forecasting. It mainly includes some correct historical regional electricity-related information, such as regional GDP, regional resident population, regional user number, regional power equipment number, regional transmission line length, regional transmission line type, historical electricity demand, etc.

[0056] Deviation input information s: Input data used for electricity demand forecasting, which belongs to potential historical deviation input information; it includes clearly defined deviation data, such as specific information related to electricity consumption during special periods, such as the construction of new industrial parks, the demolition of factories, new energy subsidies, etc.

[0057] Actual electricity demand y: The actual electricity demand given by the dispatch automation system in the digital twin power grid, that is, the actual electricity consumption of the region in reality;

[0058] Electricity demand forecast The predicted regression value represents the resource value that the power grid dispatch automation system in the digital twin power grid should allocate to the region based on the input information.

[0059] Step 2: Model Network Construction

[0060] In deep learning models, see Figure 1 and Figure 2 The z is input into a variational autoencoder for transformation to obtain mean and variance data. A simulated Gaussian distribution is constructed using the obtained mean and variance. Random sampling is performed from the Gaussian distribution to obtain the characterizing feature, denoted as z. The characterizing feature is divided into two parts, denoted as zi and zv respectively. b Part and z p Part of the data involves using z to fit the power network dispatching data of the digital twin power grid, and using z... b Partial data is used to fit the biased input data, and then trained using a generative adversarial network to improve z-axis performance. b Part and z p Partial independence.

[0061] Specifically, variational autoencoders and generative adversarial networks are used to perform bias-free information processing on the data feature representation. First, an autoencoder model is built. Equations (1)-(3) describe the transformation process of data through the variational autoencoder. The variational autoencoder consists of a mean encoder and a variance encoder. Both encoders define a two-layer neural network, using the ELU function as the inter-layer activation function and the Tanh function as the tail layer activation function. The predicted mean and variance are used to construct a simulated distribution, and the final representation features are randomly sampled from the distribution.

[0062] The obtained representation z is divided into z b Part and z p Partial, using [z b +zp Fit the original data, using z b To fit the biased input data, equations (4)-(6) describe this adversarial learning process. Finally, z is improved. b and z p The independence of z. The final output z p As encoded features that have had potential biases removed.

[0063] u = Tanh(W u2 ×ELU(W u1 ×x)) (1)

[0064] v = Tanh(W v2 ×ELU(W v1 ×x)) (2)

[0065]

[0066] z b ,z p =split(z) (4)

[0067]

[0068]

[0069] In the formula, u and v represent the mean and variance of the data, respectively. To receive data for power network dispatching in the fitted digital twin power grid, Input data for the fitted bias, W *i This represents the weights of the i-th network layer, where * is u, v, x, a. This indicates sampling from a normal distribution with parameters u and v. The above operations are for bias representation. The variational autoencoder and the adversarial network are used together as the representation model. Therefore, the overall representation model includes the following sub-networks: a neural network that outputs the mean of the data, a neural network that outputs the variance of the data, a neural network that outputs the unbiased input information reconstruction, and a neural network that outputs the biased input information reconstruction.

[0070] Step 3: Model Optimization

[0071] The final output z p A portion of the coded features, which have been used as bias-free inputs, are input into the regression network to obtain the predicted electricity demand. The predicted electricity demand is then examined to determine whether it meets the preset requirements. If it does not meet the preset requirements, the deep learning model needs to be further optimized until the preset requirements are met.

[0072] Specifically, model optimization includes the following:

[0073] In variational autoencoders, each feature column is sampled from a Gaussian distribution generated by its own unique predicted mean and variance. By constraining the relationship between the generated Gaussian distribution and the standard Gaussian distribution, the representation features become more independent.

[0074] The power network dispatching data received by the digital twin power grid is mapped to a neural network that reconstructs the unbiased input information at output by using z as a fit. A simple mean square error loss function is used to constrain the information carried in the representation. The loss function of the neural network that reconstructs the unbiased input information at output is calculated as follows:

[0075]

[0076] Using z p Partial data fitting of the biased input data corresponds to a neural network that reconstructs the input information from the output bias. A simple mean squared error loss function is used to constrain the representation of z. p Partial information is carried; the neural network loss function for restoring the output deviation input information is calculated as follows:

[0077]

[0078] Make z b Part and z p The loss functions that are independent of each other are calculated as follows:

[0079]

[0080] In the formula, p(x|z) is the decoding distribution, q(z|x) is the encoding distribution, p(z) is the prior representation of the features, and p(s|z) is the encoding distribution. b ) is the distribution of the deviation information decoding, z bk The k-th prediction bias input representation dimension, z,z p ,z b The relationship is z = [z p ,z b ];

[0081] Based on the above three loss functions, the overall loss function for training representations can be derived as follows:

[0082] Loss = Loss VAE +Loss b +α×Loss z

[0083] In the formula, α is a parameter that controls the degree to which the characterization eliminates bias information, and the value of α ranges from [0,1]. The larger the α is, the deeper the optimization against overfitting. If the characterization improves its resistance to overfitting, then the complete historical information it contains will be reduced. Therefore, choosing an appropriate α as the optimization balance parameter is also very important.

[0084] Through the final obtained characterization z p We use a naive regression network f to output efficient resource allocation values, i.e., electricity demand forecasts. The loss function corresponding to the regression network f is as follows:

[0085]

[0086] in, It is the regression threshold.

[0087] The obtained electricity demand forecast is examined to determine whether it meets the preset requirements. This examination includes an assessment of forecast accuracy and a determination of the degree of bias correction. The forecast accuracy assessment is performed using the following formula.

[0088]

[0089] In the formula, A is the prediction accuracy, and y i It is the actual power demand value for the corresponding region i given by the dispatch automation system in the digital twin power grid. This is the predicted electricity demand value for region i given by the deep learning model. The average value of actual electricity demand is given, and n is the total number of regions. The prediction accuracy A is obtained by normalizing the regression values, providing a reasonable model accuracy evaluation data. Its value is between 0 and 1, with a larger value indicating better accuracy.

[0090] The degree of deviation correction is assessed using the following formula.

[0091]

[0092] A naive classifier is used to classify the data to be tested, determining whether the data is biased. The data is then evenly distributed across k batches, where i is the i-th batch. In the formula, ΔEO represents the degree of bias removal; a larger ΔEO value indicates better bias removal. TP s＝i The number of data points predicted as biased and actually representing biased data on type i data; FN s＝i It refers to the number of data points predicted as biased but actually being unbiased on type i data; TP s＝j The number of data points predicted as biased and actually represented as biased data on class j data; FN s＝j It represents the number of data points predicted as biased but actually being unbiased on class j data.

[0093] If the prediction accuracy meets the requirements but the bias correction does not, the α value needs to be lowered and the learning optimization needs to be performed again. If the bias correction meets the requirements but the prediction accuracy does not, the α value needs to be increased and the learning optimization needs to be performed again. Through continuous adjustment and optimization, the requirements for both prediction accuracy and bias correction are met simultaneously.

[0094] In this embodiment of the invention, step one mainly involves mathematically modeling the overfitting problem of power prediction in a digital twin power grid; step two mainly involves building the various network frameworks of the model, where the output of the representation is achieved by a variational autoencoder, information separation of the representation is achieved by strengthening the independence between different parts, and information preservation of the representation is achieved by a decoding network; step three mainly involves iteratively optimizing different neural networks using their respective optimization objectives. In a preferred embodiment of the invention, see [link to specific embodiment]. Figure 3 After mathematical modeling, the network module is initialized, and then the representation model is trained, namely, the variational autoencoder and the adversarial network are trained. By adjusting the setting α, the expected information shielding ability is achieved. Then, the regression model is further trained. By adjusting the setting α and training the representation model, the regression model is made to meet the expected evaluation error. Finally, the representation model and the regression model are output. The input data of the model are passed through the representation model and the regression model in sequence to obtain the predicted value of electricity demand.

[0095] The overfitting power forecasting method for digital twin power grids provided in this embodiment uses data representation methods in deep learning to process the data features of the region to be loaded. It efficiently represents the data based on variational autoencoders and generative adversarial networks, and uses the strategy of deep learning models to schedule power resources, replacing the previous dispatch automation system that focused too much on distribution efficiency, thus preventing overfitting of the power grid dispatch model of the digital twin power grid.

[0096] The overfitting power forecasting method for digital twin power grids provided by this invention extracts the difference in input features, removes biased data that hinders forecasting ability, reduces the correlation between predicted power consumption and biased data using variational autoencoders and generative adversarial networks, and uses representation learning to remove biased data from power grid dispatching data to achieve reasonable forecasting of contracted power consumption and promote the efficient operation of dispatch automation systems.

[0097] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for preventing overfitting in power prediction for digital twin power grids. The concept of this device embodiment is the same as the working process of the detection method in the above embodiments. Therefore, all contents of the above detection method embodiments are incorporated into this device embodiment by full reference, and will not be repeated here.

[0098] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for preventing overfitting in power prediction for digital twin power grids. The concept of this storage medium embodiment is the same as the working process of the detection method in the above embodiments. Therefore, the entire content of the above detection method embodiments is incorporated into this storage medium embodiment by means of full reference, and will not be repeated here.

[0099] The overfitting power forecasting method for digital twin power grids provided by this invention is based on variational autoencoders and generative adversarial networks to robustly represent data, and uses deep learning model strategies to schedule power resources, replacing the previous automated scheduling system that focused too much on historical data.

[0100] The above description is merely a preferred embodiment of the present invention and does not limit its patent scope. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, whether directly or indirectly applied to other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for preventing overfitting in power forecasting for digital twin power grids, characterized in that, Includes the following steps: A deep learning model is built using machine learning to process the power network dispatch data received by the digital twin power grid, and the corresponding model input data (c, s, ...) is provided. and model output data By definition, the power network dispatching received data of the digital twin power grid is divided into non-deviation input data and deviation input data. The power network dispatching received data of the digital twin power grid is denoted as... c represents unbiased input data, and s represents biased input data. c and s are distinguished by a predefined information stability. This represents the actual electricity demand. This is a forecast of electricity demand. In the deep learning model, the power network dispatch data received by the digital twin power grid is input into a variational autoencoder for transformation to obtain mean and variance data. A simulated Gaussian distribution is constructed using the obtained mean and variance. Random sampling is performed from the Gaussian distribution to obtain a representational feature, denoted as z. This representational feature is divided into two parts, denoted as zi and zv respectively. Parts and Partially, z is used to fit the power network dispatching received data of the digital twin power grid, utilizing... Partial data is used to fit the biased input data, and then trained using a generative adversarial network to improve performance. Parts and Partial independence will result in the output after training. A portion of the biased data is used as a feature to input into the regression network to obtain the predicted electricity demand. The predicted electricity demand is then examined to determine whether it meets the preset requirements. If it does not meet the preset requirements, the deep learning model needs to be further optimized until the preset requirements are met. The power network dispatching data received by the digital twin power grid is mapped to a neural network that reconstructs the unbiased input information at output by using z as a fit. The loss function of the neural network that reconstructs the unbiased input information at output is calculated as follows: ； use Partial data is used to fit the biased input data to a neural network that reconstructs the output biased input information. The loss function of the neural network for reconstructing the output biased input information is calculated as follows: ； Make Parts and The loss functions that are independent of each other are calculated as follows: ； In the formula, It is a decoding distribution. It is a coding distribution. It represents the prior characteristics. It is the distribution of deviation information decoding. It is the k-th prediction bias input representation dimension; Based on the above three loss functions, the overall loss function for training representations can be derived as follows: ； In the formula, Parameters used to control the degree to which bias information is eliminated.

2. The method for preventing overfitting in power prediction for digital twin power grids according to claim 1, characterized in that, The evaluation includes an evaluation of prediction accuracy and an evaluation of the degree of bias correction. The prediction accuracy evaluation is performed using the following formula. ； In the formula, To improve prediction accuracy, It is the corresponding area given by the dispatch automation system in the digital twin power grid. The actual value of electricity demand, It is the corresponding region given by the deep learning model. Electricity demand forecasts This is the average of the actual electricity demand. Total number of regions; The degree of deviation correction is assessed using the following formula. ； In the formula, To reduce the degree of bias, Is i The prediction on the data type is the number of biased data; Is i The number of data points where the predicted data is biased but the actual data is unbiased; Is j The prediction on the data type is the number of biased data; Is j The number of data points where the predicted data is biased but the actual data is unbiased.

3. The method for preventing overfitting in power prediction for digital twin power grids according to claim 1, characterized in that, The variational autoencoder includes a mean encoder and a variance encoder, both of which define a two-layer neural network and use... The function is used as an inter-layer activation function. The function is used as the tail activation function. The conversion formula corresponding to the mean encoder is as follows: ； The conversion formula corresponding to the variance encoder is as follows: ； In the formula, These represent the mean and variance of the data, respectively. This represents the weight of the i-th network layer, where * is the weight of the i-th layer. .

4. The method for preventing overfitting in power prediction for digital twin power grids according to claim 2, characterized in that, The following is the fitting formula for using z to fit the power network dispatching data of the digital twin power grid: ； use To fit the biased input data to partial data, the corresponding fitting formula is as follows: ； In the formula, To receive data for power network dispatching in the fitted digital twin power grid, Input data for the fitted deviation. This represents the weight of the i-th network layer, where * is... .

5. The method for preventing overfitting in power prediction for digital twin power grids according to claim 1, characterized in that, If the prediction accuracy meets the requirements but the bias correction does not, then the setting needs to be lowered. The numerical values ​​are then further optimized through learning. If the bias correction meets the requirements but the prediction accuracy does not, then the setting needs to be increased. Further learning and optimization were conducted, including... The value range is [0, 1].

6. The method for preventing overfitting in power prediction for digital twin power grids according to claim 1, characterized in that, The regression network is denoted as The corresponding loss function is as follows: ； in, It is the regression threshold.

7. The method for preventing overfitting in power prediction for digital twin power grids according to claim 1, characterized in that, The non-biased input data includes regional GDP, regional resident population, regional user count, regional power equipment count, regional transmission line length, regional transmission line type, and historical electricity demand. The biased input data is potential historical biased input information, which includes defined explicit biased data.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the overfitting power prediction method for digital twin power grids as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the overfitting power prediction method for digital twin power grids as described in any one of claims 1 to 7.

Images