Method and system for predicting dynamic simultaneous rate intervals

Abstract

This invention provides an artificial intelligence-based method and system for predicting the dynamic simultaneity rate interval of distribution transformers. The system first quantifies and grades the acquired historical data into three levels: beginner, intermediate, and advanced. An innovative time-weighted loss function is then used to progressively train the CVAE model, generating a high-precision initial prediction model. In the online prediction and update phase, the model receives real-time distribution transformer characteristics as input and generates core probability interval prediction results for the application layer. Simultaneously, the invention's feedback update mechanism is activated. A deviation calculation module compares the prediction results with actual observation data in real time. If the deviation exceeds a threshold, incremental updates are triggered to fine-tune the prediction model's parameters, thus achieving online adaptive modeling and forming a complete technical closed loop.

Description

Technical Field

[0001] This invention relates to the field of power system management and power grid analysis technology, specifically to a probabilistic interval prediction and risk quantification method and system for random loads in distribution networks based on a conditional generative artificial intelligence model. Background Technology

[0002] The rapid proliferation of electric vehicles (EVs) has brought significant and novel challenges to power distribution network management. Unlike traditional residential or industrial loads, EV charging behavior is highly stochastic, time-varying, and exhibits strong temporal correlations (e.g., a large number of users charging within similar timeframes after get off work), making it difficult to predict peak loads on distribution transformers. Existing technologies have significant limitations in addressing this challenge. First, traditional load forecasting methods typically rely on static simultaneity factor coefficients or simple time-of-use averages. These methods cannot adapt to the vastly different scenarios of different distribution transformers (e.g., distribution transformers serving residential areas, commercial areas, or highway service areas have drastically different load characteristics) or capture the dynamic changes in charging behavior. Second, although some studies have employed more advanced artificial intelligence models for load forecasting, these models typically provide a single-point, deterministic prediction. This output fails to effectively quantify the inherent uncertainties in the forecasting process and all the seemingly realistic future scenarios. In particular, these deterministic prediction models aim to minimize average errors (such as mean squared error), which inherently underestimates the probability and magnitude of low-probability, high-impact load peaks caused by highly correlated user behavior. These extreme events are precisely the main causes of distribution transformer overload and damage. Therefore, such single-point prediction results cannot adequately support risk-based decision-making and may even give grid operators a false sense of security.

[0003] In summary, the industry urgently needs a new technical solution that can not only predict load, but more importantly, provide a probabilistic prediction range, perform customized predictions based on the specific attributes of power grid assets, and evolve adaptively online, thereby truly achieving proactive and reliable power grid operation and planning. Summary of the Invention

[0004] This invention solves the problem of how to perform customized load forecasting based on the specific physical and environmental characteristics of distribution transformers through the following technical means.

[0005] The method for predicting the dynamic simultaneity rate interval of distribution transformers includes the following steps:

[0006] Data acquisition and preprocessing: Acquire historical data of distribution transformers; historical data includes time series data and distribution transformer characteristics; quantify and classify the historical data into three levels: beginner, intermediate, and advanced.

[0007] Model training: Based on the beginner, intermediate and advanced courses, the CVAE model is trained using a time-weighted loss function to obtain the initial prediction model;

[0008] Online prediction: The initial prediction model receives the characteristics of the distribution transformer as input in real time and generates the core probability interval prediction results to be delivered to the application layer;

[0009] Online adaptive model adjustment: The deviation calculation module compares the prediction results with the actual observation data in real time. Once the deviation exceeds the threshold, it will trigger an incremental update to fine-tune the parameters of the prediction model.

[0010] Furthermore, the method for dividing the primary, intermediate, and advanced courses is as follows: Obtain the simultaneity rate curves for different distribution transformers at set durations; and calculate the sample difficulty coefficient based on the maximum and minimum values ​​of the curves, as well as the changes between continuous data points, according to the following formula. :

[0011]

[0012] in For the first i Simultaneity rate curve; α These are weighting coefficients used to balance the two dimensions of volatility and peak-to-valley difference; A normalized volatility indicator, The normalized peak-to-valley difference index; based on the calculated difficulty coefficient By setting two thresholds and All training samples are automatically assigned to three courses:

[0013] Beginner Course: Includes all requirements < ;

[0014] Intermediate course: Includes all requirements ≤ < ;

[0015] Advanced courses: include all those who meet the requirements ≥ The sample.

[0016] Furthermore, based on the three levels of courses (beginner, intermediate, and advanced), the method of training the CVAE model using the time-weighted loss function is as follows: First, the CVAE model is initially trained using data from the beginner course. After the model converges, data from the intermediate and advanced courses are gradually introduced to perform incremental fine-tuning of the model, thereby enhancing the model's final fitting and generalization ability to complex working conditions.

[0017] Furthermore, the expression for the time-weighted loss function is:

[0018] Where T represents the set of times t, For the true value, To generate values, w ( t ) is the time weighting function. w ( t The value of ) is positively correlated with the historical average electricity price over different periods.

[0019] Furthermore, the specific process of online prediction is as follows:

[0020] 1) Provide conditions: Determine the target power distribution characteristic condition vector c;

[0021] 2) Latent space sampling: From the prior distribution of the model, i.e., the standard normal distribution N(0,I), independent random sampling is performed N times to obtain a set of latent vectors { z 1, z 2,..., z N}; each z i This represents a possible pattern or style of load variation that has not been precisely observed in historical data.

[0022] 3) Conditional generation: For each sampled latent vector... z i This, along with a fixed target feature condition vector c, is input into the trained decoder; the decoder will then generate a corresponding, true future simultaneous rate curve. ;

[0023] 4) Construct the prediction set: Repeat step 3) to obtain a set containing N prediction curves. x 1′, x 2′,..., x N This set together constitutes a probabilistic description of the future simultaneity rate.

[0024] 5) Define the prediction interval: Perform statistical analysis on this prediction set at each time point; calculate the percentile at each time point to determine...t Lower confidence limit of the simultaneity rate for time prediction B L ( t ) and confidence upper limit B U ( t ); lower confidence limit B L ( t ) and confidence upper limit B U ( t Together, they constitute a dynamic simultaneous rate prediction interval with a certain confidence level. B L ( t ), B U ( t )];

[0025] The specific process of online adaptive model adjustment is as follows: First, the trained CVAE model is run online, continuously outputting prediction intervals; after model deployment, the system will monitor the simultaneity curves observed in real time. x ( t ) and the confidence interval of the model prediction [ B L ( t ), B U ( t A comparison is performed; and an incremental update process is triggered using a sliding window-based quantization triggering mechanism; the sliding window-based quantization triggering mechanism is as follows: First, a single-point deviation indicator function is defined:

[0026]

[0027] This function determines that... t At time 1, whether the actual value exceeds the prediction interval; then, within a length of 1... W Calculate the cumulative deviation within the sliding time window. :

[0028]

[0029] Representative in the past W Within a given timeframe, the frequency with which the actual value exceeds the predicted range; finally, setting an update trigger threshold. x When the conditions are met When the system detects a significant shift in operating conditions, it automatically triggers an incremental update process: storing the new data samples into the supplementary dataset.

[0030] This invention also provides a dynamic simultaneity rate interval prediction system for distribution transformers, comprising:

[0031] Data acquisition and preprocessing module: acquires historical data of distribution transformers; historical data includes time series data and distribution transformer characteristics; the historical data is quantified and graded into three levels: beginner, intermediate, and advanced.

[0032] Model training module: Based on the beginner, intermediate and advanced courses, the CVAE model is trained using a time-weighted loss function to obtain the initial prediction model;

[0033] Online prediction module: The initial prediction model receives the characteristics of the distribution transformer as input in real time, generates the core probability interval prediction results, and delivers them to the application layer.

[0034] Online adaptive adjustment module: By comparing the prediction results with the actual observation data in real time through the deviation calculation module, once the deviation exceeds the threshold, incremental updates will be triggered to fine-tune the parameters of the prediction model.

[0035] Furthermore, the method for dividing the primary, intermediate, and advanced courses is as follows: Obtain the simultaneity rate curves for different distribution transformers at set durations; and calculate the sample difficulty coefficient based on the maximum and minimum values ​​of the curves, as well as the changes between continuous data points, according to the following formula. :

[0036]

[0037] in This is the simultaneity rate curve for the i-th line; α These are weighting coefficients used to balance the two dimensions of volatility and peak-to-valley difference; A normalized volatility indicator, The normalized peak-to-valley difference index; based on the calculated difficulty coefficient By setting two thresholds and All training samples are automatically assigned to three courses:

[0038] Beginner Course: Includes all requirements < ;

[0039] Intermediate course: Includes all requirements ≤ < ;

[0040] Advanced courses: include all those who meet the requirements ≥ The sample.

[0041] Furthermore, based on the three levels of courses (beginner, intermediate, and advanced), the method of training the CVAE model using the time-weighted loss function is as follows: First, the CVAE model is initially trained using data from the beginner course. After the model converges, data from the intermediate and advanced courses are gradually introduced to perform incremental fine-tuning of the model, thereby enhancing the model's final fitting and generalization ability to complex working conditions.

[0042] Furthermore, the expression for the time-weighted loss function is:

[0043] Where T represents the set of times t, For the true value, To generate values, w ( t ) is the time weighting function. w ( t The value of ) is positively correlated with the historical average electricity price over different periods.

[0044] Furthermore, the specific process of online prediction is as follows:

[0045] 1) Provide conditions: Determine the target power distribution characteristic condition vector c;

[0046] 2) Latent space sampling: From the prior distribution of the model, i.e., the standard normal distribution N(0,I), independent random sampling is performed N times to obtain a set of latent vectors { z 1, z 2,..., z N}; each z i This represents a possible pattern or style of load variation that has not been precisely observed in historical data.

[0047] 3) Conditional generation: For each sampled latent vector... z i This, along with a fixed target feature condition vector c, is input into the trained decoder; the decoder will then generate a corresponding, true future simultaneous rate curve. ;

[0048] 4) Construct the prediction set: Repeat step 3) to obtain a set containing N prediction curves. x 1′, x 2′,..., x N This set together constitutes a probabilistic description of the future simultaneity rate.

[0049] 5) Define the prediction interval: Perform statistical analysis on this prediction set at each time point; calculate the percentile at each time point to determine...t Lower confidence limit of the simultaneity rate for time prediction B L ( t ) and confidence upper limit B U ( t ); lower confidence limit B L ( t ) and confidence upper limit B U ( t Together, they constitute a dynamic simultaneous rate prediction interval with a certain confidence level. B L ( t ), B U ( t )];

[0050] The specific process of online adaptive model adjustment is as follows: First, the trained CVAE model is run online, continuously outputting prediction intervals; after model deployment, the system will monitor the simultaneity curves observed in real time. x ( t ) and the confidence interval of the model prediction [ B L ( t ), B U ( t A comparison is performed; and an incremental update process is triggered using a sliding window-based quantization triggering mechanism; the sliding window-based quantization triggering mechanism is as follows: First, a single-point deviation indicator function is defined:

[0051]

[0052] This function determines that... t At time 1, whether the actual value exceeds the prediction interval; then, within a length of 1... W Calculate the cumulative deviation within the sliding time window. :

[0053]

[0054] Represents the past W Within a given timeframe, the frequency with which the actual value exceeds the predicted range; finally, setting an update trigger threshold. x When the conditions are met When the system detects a significant shift in operating conditions, it automatically triggers an incremental update process: storing the new data samples into the supplementary dataset.

[0055] The advantages of this invention are:

[0056] 1. This invention proposes a method for generating charging simultaneity rate intervals based on specific physical feature constraints. It abandons the fuzzy selection of input features in traditional load forecasting and uniquely identifies and structures a set of physical and environmental features strongly correlated with the shape of the charging simultaneity rate curve of distribution transformers. These features include: the year the distribution transformer was put into operation, its geographical location attributes (such as residential areas or commercial areas), the total number of connected charging facilities, and their total power. This feature set is used as a strong conditional vector for a conditional variational autoencoder (CVAE), fundamentally solving the problem of insufficient generalization ability of the prediction model due to the lack of key information.

[0057] 2. Unlike existing technologies that only provide a single prediction curve, this invention utilizes the ability of a conditional variational autoencoder (CVAE) to sample multiple times from the latent space, generating a diverse cluster of future simultaneity rate curves that conform to the characteristics of a specific distribution transformer. This allows for the construction of a probability prediction interval with clear statistical significance. This probability interval can accurately quantify the uncertainty of future loads, providing unprecedented decision-making basis for power grid operation risk assessment and the development of robust dispatch strategies.

[0058] 3. To ensure the model's sophistication and practicality, this invention does not employ the conventional offline training mode. Instead, it develops a comprehensive advanced model construction methodology that includes customized objectives, adaptive capabilities, and robust training. During the training phase, a progressively challenging learning strategy is used, effectively enhancing the customization requirements for specific scenarios. Meanwhile, the feedback and update mechanism during application ensures the model's online adaptability and robustness.

[0059] 4. A complete "data-model-application" system architecture has been constructed. It deeply couples the unique probability interval prediction output with downstream modules such as risk assessment, orderly charging, and capacity expansion planning, ensuring the integrity of the technical solution and its direct usability in industrial scenarios, and solving the common problem in existing technologies where prediction models are disconnected from actual decision-making applications. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the overall workflow of the dynamic simultaneity rate interval prediction method for distribution transformers in this embodiment of the invention. The diagram illustrates the complete workflow from data input to strategy output. Historical data and distribution transformer characteristics serve as model inputs, which are processed by the core CVAE-based prediction model. After processing, the model generates a dynamic simultaneity rate interval as the key output, which is ultimately used to support applications such as risk assessment, orderly charging, and capacity expansion planning.

[0061] Figure 2This document illustrates the entire flowchart of the CVAE model, incorporating an innovative training strategy, from initial construction to online adaptive optimization, as described in this invention. In the offline training phase on the left, the system first quantifies and grades the acquired historical data into three levels: beginner, intermediate, and advanced. An innovative time-weighted loss function is then used to progressively train the CVAE model, generating a high-precision initial prediction model. Moving to the online prediction and update phase on the right, the model receives real-time distribution transformer features as input, generating core probability interval prediction results for the application layer. Simultaneously, the feedback update mechanism of this invention is activated. The deviation calculation module compares the prediction results with the actual observed data in real time. Once the deviation exceeds a threshold, incremental updates are triggered to fine-tune the prediction model's parameters, thus achieving online adaptation and forming a complete technical closed loop.

[0062] Figure 3 This is a set of typical dynamic simultaneity rate interval prediction results for distribution transformers obtained using the method shown in the embodiments of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] In this embodiment, a method for predicting the dynamic simultaneity rate of charging load based on a conditional variational autoencoder is described, such as... Figure 1 As shown, the specific steps are as follows:

[0065] Step 1: System Architecture and Data Acquisition. Establish the infrastructure and data foundation required for implementing this invention.

[0066] Step 1.1, System Architecture

[0067] The implementation system of this invention can be deployed on a central server or cloud platform, which possesses high-performance computing capabilities (such as GPU acceleration) to support the training of deep learning models. Logically, the system includes:

[0068] 1) Data Interface Module: Responsible for secure and reliable data exchange with external data sources (such as the power grid's SCADA system, Advanced Metering Architecture (AMI), charging pile operation platform, etc.).

[0069] 2) Data storage module: Employs a time-series database or distributed file system for efficient storage and retrieval of large amounts of historical load data and distribution transformer attribute data.

[0070] 3) Model training and inference engine: The core computing unit, responsible for executing all algorithms from step two to step four.

[0071] 4) Application Interface Module: Push the prediction results and alarm signals generated in step 5 to the power grid dispatching system, asset management system or demand response platform in a standardized format (such as API call).

[0072] Step 1.2, Data Acquisition. Two highly correlated data types are collected. In this embodiment, the input data required by the model is collectively referred to as historical data. This historical data consists of two main categories of information: the first is time-series data, which describes the dynamic changes in the load of the distribution transformer over time; the second is distribution transformer characteristics, which describes the inherent, relatively static physical and environmental attributes of the distribution transformer itself. The former serves as samples for model learning, while the latter provides the conditions for the model to make personalized predictions.

[0073] Step 1.2.1, Time series data ( X This is a sample for model learning. Historical load data is collected from smart metering terminals (such as smart meters or concentrators) on the low-voltage side of the distribution transformer. The collection frequency is selected to be 15 minutes or higher to capture the dynamic changes in charging load. The raw load data (unit: kW) needs to be processed and converted into a dimensionless simultaneity rate. Essentially, these data depict the load change curve of the distribution transformer over a period of time, specifically the load value at each moment. The calculation formula is:

[0074]

[0075] in, For the simultaneous rate, For real-time load, The rated capacity of the distribution transformer is used. This normalization operation eliminates the dimensional differences between distribution transformers of different capacities, allowing the model to learn across distribution transformers of different sizes.

[0076] Step 1.2.2, Distribution Transformer Conditional Characteristic Data ( CThis refers to the static conditions for the model to make personalized predictions. These data describe the unique physical and socioeconomic attributes of each distribution transformer that affect its load characteristics. These features help the model understand the fundamental reasons for the different load patterns of different distribution transformers. In this invention, it mainly includes information such as the number of charging facilities, user composition, the year the distribution transformer was put into operation, and its rated capacity. Essentially, these data are used to describe the objects analyzed and predicted in this invention. These features are formally defined as shown in Table 1.

[0077] Table 1: Definition of Conditional Feature Variables

[0078]

[0079] Step 2: Data Preprocessing and Feature Encoding. This step transforms the raw, heterogeneous data into a standardized numerical format suitable for deep learning models.

[0080] Step 2.1: Normalization of Time Series Data

[0081] Although already calculated as the synchronization rate, to further improve the stability and convergence speed of model training, the synchronization rate curve can be... x The numerical range is further scaled up to a standard interval, for example [ [1,1]. To prevent gradient vanishing or gradient exploding problems in gradient calculation of neural networks.

[0082] Step 2.2, Conditional Feature Encoding. Neural networks can only process numerical inputs. Therefore, categorical features must be encoded.

[0083] First, one-hot encoding is performed on the categorical feature c2 (user composition). This method creates an independent binary feature for each category. This encoding method avoids introducing spurious ordinal relationships into the categories (e.g., incorrectly leading the model to believe "commercial area = 2" is greater than "residential area = 1"). Then, the one-hot encoded vector is concatenated with other numerical features (such as c1, c3, c4) to form the final conditional vector. c All numerical features should also be normalized.

[0084] Step 3: Adaptive Conditional Variational Autoencoder (CVAE) Model Architecture. This step is the core technology of this invention, detailing the internal structure and working principle of the CVAE model. For example... Figure 2 As shown, CVAE is a generative model whose goal is not to simply copy the input, but to learn a latent representation of data and generate new, seemingly realistic data from that representation.

[0085] Step 3.1, ELBO objective function

[0086] CVAE training does not directly maximize the log-likelihood of the data. This is because the term is difficult to compute. Instead, the model maximizes its Evidence Lower Bound (ELBO). This is used for training. ELBO is a variational lower bound of the true log-likelihood; maximizing ELBO is equivalent to simultaneously maximizing the log-likelihood and minimizing the difference between the true posterior and the approximate posterior. Its mathematical expression is:

[0087]

[0088] This objective function consists of two key parts: reconstruction loss (the first term). This measure is derived from the input. x Encoding to latent vectors z , and then from z Decoding x During the process of ', the original input x and reconstructed output x The similarity between ''. Its goal is to enable the decoder to accurately reconstruct the input data. For continuous data (such as simultaneous rate curves), this term is typically implemented as a negative mean squared error. KL divergence (second term): This term is a regularization term that measures the approximate posterior distribution generated by the encoder. With a pre-defined simple prior distribution The difference lies in the fact that the goal is to force the latent space learned by the encoder to have a good structure (smooth and continuous), avoiding the model from assigning an isolated and unique code to each input data point, thereby ensuring the model's generalization and generative capabilities.

[0089] Step 3.2, Encoder Network

[0090] Encoder Network Its purpose is to learn a mapping from the data space to the latent space. It receives a historical concurrency rate curve. x and its corresponding condition vector c The encoder takes the input as input and outputs parameters describing the Gaussian distribution corresponding to that input in the latent space. It employs a recurrent neural network to process the temporal data. x This is to effectively capture its time dependency. Conditional vector c The hidden vectors are concatenated with the final hidden state of the recurrent neural network and then fed into two separate fully connected layers. The fully connected layers do not directly output the hidden vectors. z Instead, it outputs the mean vector of the Gaussian distribution. m And the logarithmic variance vector log( s 2Learn the logarithmic variance rather than directly learning the variance. s 2 This is because the numerical domain is the entire real number axis, and the variance must be positive, which makes network optimization easier.

[0091] Step 3.3, Reparameterization

[0092] During training, it is necessary to obtain the distribution from the encoder output. Sampling a latent vector z The data is then fed into the decoder. However, the sampling operation itself is random and non-differentiable, which prevents gradients from propagating back from the decoder and loss function to the encoder, thus hindering encoder training. Reparameterization solves this problem by separating randomness from the network path. This is achieved by not directly inputting the data from the decoder and loss function. N ( m , s 2 Instead of sampling from a single, fixed, standard normal distribution that is independent of the model parameters, the model samples from the standard normal distribution. N (0, I Sample a noise vector in ) Then, the latent vector z It is obtained through a deterministic, differentiable transformation:

[0093]

[0094] in s =exp(0.5 log( s 2 )), ⊙ represents element-wise multiplication. Thus, randomness originates from external factors. The gradient can flow through a deterministic path to... m and s This updates the encoder's weight parameters. .

[0095] Step 3.4, Decoder Network

[0096] The decoder's role is to learn a mapping from the latent space to the data space. It receives a vector sampled from the latent space. z and a condition vector c And generate a new simultaneous rate curve. x The decoder's architecture is typically symmetrical to the encoder's; it will... z and c After concatenation, the data is used as input, and then passed through a series of fully connected layers and upsampling layers to gradually restore the low-dimensional latent vector into a high-dimensional time series output. x ′. Conditions cThe input decoder is also crucial for conditional VAEs, ensuring that the curves generated by the decoder meet the specified conditions.

[0097] Step 3.5, Loss Function

[0098] A time-weighted loss function considering power grid safety and economy is proposed: Standard CVAE training typically uses mean squared error as the reconstruction loss, aiming to make each point of the generated curve as close as possible to the original curve. In power grid applications, the economic losses or safety risks caused by prediction errors at different time points vary greatly. For example, a prediction deviation at 3 a.m. is insignificant, but a prediction deviation at 7 p.m. (peak electricity consumption) could lead to transformer overload.

[0099]

[0100] Where T represents the set of times t, For the true value, To generate values, and w ( t This is a time-weighted function. It is dynamically generated based on historical electricity prices, typical load curves, or defined risk periods. w ( t The value of ) is positively correlated with the historical average electricity price for that period, during known peak hours (such as 18:00-21:00). w ( t The value of ) is significantly higher than in other periods. This invention is the first to introduce the economic and safety risks of power grid operation into the design of the loss function of the generative model. By imposing a greater penalty on the prediction error during peak periods, this training mechanism can force the model to prioritize learning and accurately fit the curve shape of the key periods that have the greatest impact on power grid safety. This solves the technical problem of the disconnect between traditional loss functions and actual industrial application goals, and significantly improves the practicality and risk avoidance ability of the model.

[0101] When the posterior distribution is Gaussian and the prior is standard normal distribution N (0, I When the KL divergence has an analytical solution, it does not require estimation through sampling, making the computation efficient and accurate. The divergence loss is expressed as:

[0102]

[0103] The final loss function minimized during training is the weighted sum of the two terms mentioned above. (Weight factors) β The hyperparameter used to balance the trade-off between reconstruction quality and latent space regularity is an important one. The total loss is expressed as:

[0104]

[0105] Step 4: Continuous Learning Based on Closed-Loop Feedback. This step describes how to use the defined model and prepared data to complete the model's learning and application.

[0106] Step 4.1: Robust Training Strategies Based on Course Learning

[0107] Standard training mixes all historical data (whether simple, complex, clear, or noisy) together and randomly shuffles it as input to the model. In the data environment of a power system, this learning method is inefficient because the model struggles to identify truly influential key parameters. This invention introduces the concept of "curriculum learning," simulating the human learning process, allowing the model to learn from easy to difficult.

[0108] To automate the design of training courses, the training data is first categorized based on its complexity and signal-to-noise ratio. This invention proposes a method for calculating the difficulty coefficient of data samples:

[0109]

[0110] in α It is a weighting coefficient used to balance the two dimensions of volatility and peak-to-valley difference. It is a normalized volatility indicator used to measure the smoothness of a curve. The calculation formula is:

[0111]

[0112] Using the above formula V The value is normalized to the range [0, 1]. The more volatile the fluctuation, the larger the value, and the higher the learning difficulty. It is a normalized peak-to-trough difference index used to measure the dynamic range of a curve. The calculation formula is:

[0113]

[0114] The above formula normalizes the value of P to the interval [0, 1]. The larger the peak-to-valley difference, the larger this value, and the higher the learning difficulty. Based on the calculated difficulty coefficient... By setting two thresholds and All training samples are automatically assigned to three courses:

[0115] Beginner Course: Includes all requirements < The sample.

[0116] Intermediate course: Includes all requirements ≤ < The sample.

[0117] Advanced courses: include all those who meet the requirements ≥ The sample.

[0118] Finally, phased training is conducted: First, the model is initially trained using beginner-level course data. After the model converges, intermediate and advanced course data are gradually introduced for incremental fine-tuning, thereby enhancing the model's final fitting and generalization ability to complex working conditions. The goal of the training phase is to find a set of optimal model parameters (encoder weights) through iteration on the training dataset. and decoder weights i ), making the total loss function L t Minimal. The training process employs the Adam optimization algorithm based on mini-batch stochastic gradient descent. This course difficulty quantification and grading method, for the first time, provides a mathematical definition of the learning difficulty of the charging load curve, transforming the abstract teaching idea of ​​"from easy to difficult" into concrete, automatically executable algorithmic steps. It frees the course learning strategy from reliance on human experience, improves the objectivity and reproducibility of the training process, and is a key invention in the methodology for achieving robust model training.

[0119] The following example illustrates the data hierarchy in course learning:

[0120] Suppose we collected four 24-hour simultaneity rate curves from four different distribution transformers. For simplicity, we will focus on the key features of these curves: maximum values, minimum values, and variations between consecutive data points.

[0121] Curve A (representing a stable, predictable load) is very smooth, with minimal variation between hours. The minimum value is 0.10, the maximum value is 0.30, and the sum of the absolute differences is 0.8.

[0122] Curve B (representing a typical residential load) shows a distinct nighttime peak and moderate fluctuations. The minimum value is 0.05, the maximum value is 0.65, and the sum of the absolute differences is 2.5.

[0123] Curve C (representing an area with irregular commercial charging) has a large peak-to-trough difference and moderate to high volatility. The minimum value is 0.15, the maximum value is 0.85, and the sum of the absolute differences is 4.5.

[0124] Curve D (representing highly volatile loads, such as highway fast charging stations) is extremely unstable, with sharp and unpredictable peaks. The minimum value is 0.02, the maximum value is 0.92, and the sum of the absolute differences is 7.0.

[0125] Then use the normalized volatility indicator and normalized peak-to-valley difference index Perform normalization calculations. Assume the maximum volatility is 8.0 and the maximum peak-to-trough difference is 1.0 across the entire dataset. The calculations yield:

[0126] Curve A has a peak-to-trough difference of 0.20 and a volatility of 0.80. The normalized peak-to-trough difference is 0.20 and the normalized volatility is 0.10.

[0127] Curve B has a peak-to-trough difference of 0.60 and a volatility of 2.5. The normalized peak-to-trough difference is 0.60 and the normalized volatility is 0.31.

[0128] Curve C has a peak-to-trough difference of 0.70 and a volatility of 4.5. The normalized peak-to-trough difference is 0.70 and the normalized volatility is 0.56.

[0129] Curve D has a peak-to-trough difference of 0.90 and a volatility of 7.0. The normalized peak-to-trough difference is 0.90 and the normalized volatility is 0.88.

[0130] In this embodiment, the weighting factor α is set to 0.5 when calculating the sample difficulty coefficient, indicating that volatility and peak-to-valley difference are equally important. The calculated difficulties of the four curves are 0.15, 0.46, 0.63, and 0.89, respectively. Finally, these samples are divided into three courses according to the difficulty coefficient. The beginner course (difficulty < 0.40) has low volatility and small peak-to-valley difference, which is very suitable for the initial training of the model; the intermediate course (0.40 ≤ difficulty < 0.70) represents a more typical and moderately complex load pattern, which is introduced after the model has completed the initial data learning and converged; the advanced course (difficulty ≥ 0.70) represents difficult samples, characterized by high volatility and extreme peaks. These are introduced in the final stage of training to improve the robustness of the model and its ability to predict rare high-risk events.

[0131] Step 4.2, Reasoning (Prediction) Stage

[0132] The goal is to generate a probabilistic prediction interval for a given distribution transformer with specific conditions, based on its dynamic simultaneity rate. This process eliminates the need for an encoder, utilizing only a pre-trained decoder. The specific workflow is as follows:

[0133] 1) Provide conditions: Determine the condition vector of the target distribution transformer. c (For example, the number of charging piles = 23, the user composition = commercial area, the year of commissioning = 2023, etc., and preprocessing is performed).

[0134] 2) Latent space sampling: from the model's prior distribution, i.e., the standard normal distribution. N (0, I In independent random sampling N Next, a set of implicit vectors is obtained { z 1, z 2,...,z N Each one. z i This can be understood as representing a possible pattern or style of load change that has not been precisely observed in historical data.

[0135] 3) Conditional generation: For each sampled latent vector... z i Compare it with the fixed target condition vector c These are input together into the trained decoder. The decoder will then generate a corresponding, seemingly realistic future concurrency rate curve: .

[0136] 4) Construct the prediction set: Repeat step 3) to obtain a set containing N prediction curves. x 1′, x 2′,..., x N This set together constitutes a probabilistic description of the future simultaneity rate, reflecting its inherent uncertainty.

[0137] 5) Define the prediction interval: Perform statistical analysis on this prediction set at every time point (e.g., every 15 minutes of the day). Calculate the percentile for each time point; for example, calculate the 5th percentile as the lower bound of the prediction interval and the 95th percentile as the upper bound. These two boundary curves together constitute the dynamic concurrency prediction interval with a 90% confidence level.

[0138] Step 4.3, Closed-Loop Continuous Learning Framework

[0139] Models are typically trained offline once on a fixed historical dataset, and their parameters remain unchanged after deployment. This approach cannot adapt to dynamic changes such as the construction of new charging stations or changes in user behavior patterns, leading to performance degradation over time. This invention designs a closed-loop continuous learning framework of "prediction-validation-feedback-update," enabling the model to self-evolve. The specific implementation is as follows: First, the trained CVAE model is run online, continuously outputting prediction intervals. After model deployment, the system will display the real-time observed simultaneity curves... x ( t ) and the confidence interval of the model prediction [ B L ( t ), B U ( t A comparison was made. Among them, B L ( t )and BU ( t They are respectively t The lower and upper confidence limits for the simultaneity rate at any given time.

[0140] To avoid frequent model updates due to single-point noise or transient disturbances, this invention designs a quantization triggering mechanism based on a sliding window. First, a single-point deviation indicator function is defined:

[0141]

[0142] This function determines that... t At time [time], whether the actual value exceeds the predicted range. Then, within a length of [length]... W Calculate the cumulative deviation within the sliding time window. (Cumulative Deviation Score, SCD):

[0143]

[0144] Represents the past W Within a given timeframe, the frequency with which the actual value exceeds the predicted range. Finally, an update trigger threshold is set. x When the conditions are met When a significant shift in operating conditions occurs, the system automatically triggers an incremental update process: new data samples (containing input features and the true curve) are stored in the supplementary dataset. Instead of retraining the model, the system uses this small supplementary dataset to fine-tune the existing model several times.

[0145] This invention proposes an incremental update method for CVAE models based on operating condition drift detection. This method establishes a data closed-loop feedback system, enabling online continuous learning and adaptive evolution of the model, effectively addressing the industry pain point of performance degradation of static offline models in dynamically changing environments. Compared to periodic full retraining, the incremental update mechanism of this invention has extremely low computational cost, fast response speed, and can ensure the accuracy and reliability of the model's long-term operation.

[0146] Step 5: Results and Reasoning. This step transforms the model's predictive output into decision-making criteria that can guide the actual operation and planning of the power grid, forming a closed-loop intelligent management system. Combined with... Figure 3The dynamic simultaneity rate curve shown in the figure illustrates this. In the figure, each light gray thin line represents a possible future load generated by the CVAE model based on the characteristics of the current distribution transformer (such as location and number of charging piles), visually demonstrating the inherent randomness and volatility of future charging load. The blue shaded area represents the core of this invention—the probabilistic prediction range of the dynamic simultaneity rate. The actual future load curve will fall within this blue safety channel, transforming uncertainty into a clearly defined range. The red dashed line represents the set operating safety threshold for the distribution transformer, such as 80% of its rated capacity. At any time, the actual load should not continuously exceed this line; otherwise, there is a risk of overload, shortened equipment lifespan, or even damage.

[0147] Step 5.1: Short-term operational risk early warning and proactive control

[0148] When the upper edge of the blue shaded area crosses the red dotted line upwards, it's a strong high-risk warning signal. Mathematically, this means there's a non-negligible high probability of overload occurring within this timeframe. This intersection not only tells us there's a risk, but also how significant it is. The duration and height of the intersection provide precise dispatch instructions. For example, we can clearly see that at least approximately 40kW of peak load needs to be reduced between 18:00 and 22:00 to ensure safety. This visualized warning signal directly triggers the backend active control system. The system sends instructions to charging stations within the area, precisely reducing the load through price incentives or direct power adjustment, thus actively pushing the sharp corner of the blue area back below the red line.

[0149] Step 5.2, Medium- and Long-Term Planning and Asset Management

[0150] By monitoring the forecast range of a distribution transformer over a long period (such as weeks or months), the severity and frequency of its overload risk can be quantified. If a distribution transformer's chart frequently shows blue areas crossing above red lines, and the area of ​​this crossing is large, then even if it has never experienced an actual power outage, we can determine from a data perspective that this distribution transformer is in a sub-healthy or high-risk state. Long-term probabilistic assessments provide strong quantitative evidence for power grid companies' investment decisions. This shifts investment decisions from traditional reactive (replace after damage) to proactive (preventative) approaches, thereby optimizing capital expenditures and improving the overall reliability of the power grid.

[0151] The main objective of this embodiment is to provide a probabilistic interval prediction method and system for the dynamic simultaneity rate of distribution transformer charging. Its core is a complete Conditional Variational Autoencoder (CVAE) model management system incorporating an innovative training methodology, systematically overcoming the fundamental shortcomings of existing static, deterministic, and non-adaptive prediction methods. This invention does not pursue single-point prediction, but rather generates a dynamic probabilistic prediction interval by imposing conditional constraints on the specific physical and environmental characteristics of the distribution transformer and utilizing CVAE, thereby transforming invisible uncertainty risks into quantifiable decision-making basis. Furthermore, this invention proposes an adaptive training process that includes not only a progressive training strategy based on course learning and a time-weighted loss function designed to prioritize accuracy during peak periods, but also online incremental updates of the model through a closed-loop feedback mechanism, addressing the performance degradation shortcomings of static models. Ultimately, the system provides accurate probabilistic prediction results to support downstream risk warning, proactive control, and asset planning applications.

[0152] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the dynamic simultaneity rate interval of distribution transformers, characterized in that, Includes the following steps: Data acquisition and preprocessing: Acquire historical data of distribution transformers; historical data includes time series data and distribution transformer characteristics; The historical data was quantified and graded into three levels of courses: beginner, intermediate, and advanced. Model Training: Based on the beginner, intermediate, and advanced courses, the CVAE model is trained using a time-weighted loss function to obtain an initial prediction model; the expression for the time-weighted loss function is: Where T represents the set of times t, For the true value, To generate values, w ( t ) is the time weighting function. w ( t The value of ) is positively correlated with the historical average electricity price over different periods; Online prediction: The initial prediction model receives the characteristics of the distribution transformer as input in real time, generates a set containing N prediction simultaneity rate curves, calculates the percentile for the set of prediction simultaneity rate curves at each time point, and determines... t The lower and upper confidence limits of the simultaneous rate prediction are calculated at any given time. The lower and upper confidence limits together constitute a dynamic simultaneous rate prediction interval with a certain confidence level, which is delivered to the application layer. Online adaptive model adjustment: The deviation calculation module compares the prediction results with the actual observation data in real time. Once the deviation exceeds the threshold, it will trigger an incremental update to fine-tune the parameters of the prediction model.

2. The method for predicting the dynamic simultaneity rate interval of distribution transformers according to claim 1, characterized in that, The method for dividing the courses into elementary, intermediate, and advanced levels is as follows: Obtain the simultaneity rate curves for different distribution transformers at set durations; and calculate the sample difficulty coefficient based on the maximum and minimum values ​​of the curves, as well as the changes between continuous data points, according to the following formula. : in This is the simultaneity rate curve for the i-th line; α These are weighting coefficients used to balance the two dimensions of volatility and peak-to-valley difference; A normalized volatility indicator, The normalized peak-to-valley difference index; based on the calculated difficulty coefficient By setting two thresholds and All training samples are automatically assigned to three courses: Beginner Course: Includes all requirements < ; Intermediate course: Includes all requirements ≤ < ; Advanced courses: include all those who meet the requirements ≥ The sample.

3. The method for predicting the dynamic simultaneity rate interval of distribution transformers according to claim 1 or 2, characterized in that, The method for training the CVAE model based on the three levels of courses (beginner, intermediate, and advanced) using the time-weighted loss function is as follows: First, the CVAE model is initially trained using data from the beginner course. After the model converges, data from the intermediate and advanced courses are gradually introduced to perform incremental fine-tuning of the model, thereby enhancing the model's final fitting and generalization ability to complex working conditions.

4. The method for predicting the dynamic simultaneity rate interval of distribution transformers according to claim 1 or 2, characterized in that, The online prediction process is as follows: 1) Provide conditions: Determine the target power distribution characteristic condition vector c; 2) Latent space sampling: From the prior distribution of the model, i.e., the standard normal distribution N(0,I), N independent random samples are taken to obtain a set of latent vectors { z 1, z 2,..., z N }; each z i This represents a possible pattern or style of load variation that has not been precisely observed in historical data. 3) Conditional generation: For each sampled latent vector... z i It is then input together with the fixed target feature condition vector c into the trained decoder; The decoder will generate a corresponding, real future simultaneous rate curve: ; 4) Construct the prediction set: Repeat step 3) to obtain a set containing N prediction curves. x 1′, x 2′,..., x N This set together constitutes a probabilistic description of the future simultaneity rate. 5) Define the prediction interval: Perform statistical analysis on this prediction set at each time point; Calculate the percentile at each time point to determine t Lower confidence limit of the simultaneity rate for time prediction B L ( t ) and confidence upper limit B U ( t ); lower confidence limit B L ( t ) and confidence upper limit B U ( t Together, they constitute a dynamic simultaneous rate prediction interval with a certain confidence level. B L ( t ), B U ( t )).

5. The method for predicting the dynamic simultaneity rate interval of distribution transformers according to claim 4, characterized in that, The specific process of online adaptive model adjustment is as follows: First, the trained CVAE model is run online, continuously outputting prediction intervals; after model deployment, the system will monitor the simultaneity curves observed in real time. x ( t ) and the confidence interval of the model prediction [ B L ( t ), B U ( t Compare them; The incremental update process is triggered using a sliding window-based quantization triggering mechanism. The sliding window-based quantization triggering mechanism is as follows: First, a single-point deviation indicator function is defined: This function determines that... t At time 1, whether the actual value exceeds the prediction interval; then, within a length of 1... W Calculate the cumulative deviation within the sliding time window. : Represents the past W Within a given timeframe, the frequency with which the actual value exceeds the predicted range; finally, setting an update trigger threshold. χ When the conditions are met When a condition drift occurs, the system determines that an incremental update process has been triggered: the new data sample is stored in the supplementary dataset.

6. A distribution transformer dynamic simultaneity rate interval prediction system, characterized in that, include: Data acquisition and preprocessing module: acquires historical data of distribution transformers; historical data includes time series data and distribution transformer characteristics; The historical data was quantified and graded into three levels of courses: beginner, intermediate, and advanced. Model training module: Based on the beginner, intermediate, and advanced courses, the CVAE model is trained using a time-weighted loss function to obtain an initial prediction model; the expression of the time-weighted loss function is: ,in, It is the actual value. It is a generated value. w ( t ) is a time weighting function. w ( t The value of ) is positively correlated with the historical average electricity price over different periods; Online prediction module: The initial prediction model receives the characteristics of the distribution transformer as input in real time, generates a set containing N prediction simultaneity rate curves, calculates the percentile for the set of prediction simultaneity rate curves at each time point, and determines... t The lower and upper confidence limits of the simultaneous rate prediction are calculated at any given time. The lower and upper confidence limits together constitute a dynamic simultaneous rate prediction interval with a certain confidence level, which is delivered to the application layer. Online adaptive adjustment module: By comparing the prediction results with the actual observation data in real time through the deviation calculation module, once the deviation exceeds the threshold, incremental updates will be triggered to fine-tune the parameters of the prediction model.

7. The distribution transformer dynamic simultaneity rate interval prediction system according to claim 6, characterized in that, The method for dividing the courses into elementary, intermediate, and advanced levels is as follows: Obtain the simultaneity rate curves for different distribution transformers at set durations; and calculate the sample difficulty coefficient based on the maximum and minimum values ​​of the curves, as well as the changes between continuous data points, according to the following formula. : in For the first i Simultaneity rate curve; α These are weighting coefficients used to balance the two dimensions of volatility and peak-to-valley difference; A normalized volatility indicator, The normalized peak-to-valley difference index; based on the calculated difficulty coefficient By setting two thresholds and All training samples are automatically assigned to three courses: Beginner Course: Includes all requirements < ; Intermediate course: Includes all requirements ≤ < ; Advanced courses: include all those who meet the requirements ≥ The sample.

8. The distribution transformer dynamic simultaneity rate interval prediction system according to claim 6 or 7, characterized in that, The method for training the CVAE model based on the three levels of courses (beginner, intermediate, and advanced) using the time-weighted loss function is as follows: First, the CVAE model is initially trained using data from the beginner course. After the model converges, data from the intermediate and advanced courses are gradually introduced to perform incremental fine-tuning of the model, thereby enhancing the model's final fitting and generalization ability to complex working conditions.

9. The distribution transformer dynamic simultaneity rate interval prediction system according to claim 6 or 7, characterized in that, The online prediction process is as follows: 1) Provide conditions: Determine the target power distribution characteristic condition vector c; 2) Latent space sampling: From the prior distribution of the model, i.e., the standard normal distribution N(0,I), N independent random samples are taken to obtain a set of latent vectors { z 1, z 2,..., z N }; each z i This represents a possible pattern or style of load variation that has not been precisely observed in historical data. 3) Conditional generation: For each sampled latent vector... z i It is then input together with the fixed target feature condition vector c into the trained decoder; The decoder will generate a corresponding, real future simultaneous rate curve: ; 4) Construct the prediction set: Repeat step 3) to obtain a set containing N prediction curves. x 1′, x 2′,..., x N This set together constitutes a probabilistic description of the future simultaneity rate. 5) Define the prediction interval: Perform statistical analysis on this prediction set at each time point; Calculate the percentile at each time point to determine t Lower confidence limit of the simultaneity rate for time prediction B L ( t ) and confidence upper limit B U ( t ); lower confidence limit B L ( t ) and confidence upper limit B U ( t Together, they constitute a dynamic simultaneous rate prediction interval with a certain confidence level. B L ( t ), B U ( t )).

10. The distribution transformer dynamic simultaneity rate interval prediction system according to claim 9, characterized in that, The specific process of online adaptive model adjustment is as follows: First, the trained CVAE model is run online, continuously outputting prediction intervals; after model deployment, the system will monitor the simultaneity curves observed in real time. x ( t ) and the confidence interval of the model prediction [ B L ( t ), B U ( t Compare them; The incremental update process is triggered using a sliding window-based quantization triggering mechanism. The sliding window-based quantization triggering mechanism is as follows: First, a single-point deviation indicator function is defined: This function determines that... t At time 1, whether the actual value exceeds the prediction interval; then, within a length of 1... W Calculate the cumulative deviation within the sliding time window. : Represents the past W Within a given timeframe, the frequency with which the actual value exceeds the predicted range; finally, setting an update trigger threshold. χ When the conditions are met When the system detects a significant shift in operating conditions, it automatically triggers an incremental update process: storing the new data samples into the supplementary dataset.

Images