Distribution transformer load prediction method, system, equipment and medium

Abstract

The invention belongs to the technical field of power distribution, and discloses a distribution transformer load prediction method, system and device and a medium, and the method comprises the steps: collecting historical load data and original load data of a target distribution transformer, and carrying out the normalization processing of the historical load data and the original load data, and obtaining the load data of the distribution transformer; carrying out decomposition processing on the load data of the distribution transformer by utilizing a variational mode decomposition algorithm to obtain a stable intrinsic mode function sub-sequence group; carrying out diversified processing on the stable intrinsic mode function sub-sequence group by utilizing an improved deep learning model, and predicting a load sequence of the target distribution transformer; and based on the load sequence of the target distribution transformer, performing parameter optimization on the improved deep learning model through an adaptive moment estimation optimization algorithm, and evaluating the performance of the improved deep learning model by using the evaluation index. According to the method, high-precision prediction of the distribution transformer load is realized, and the increasing demand of a power system on load prediction is met.

Description

Technical Field

[0001] This invention relates to the field of power distribution technology, and in particular to a method, system, equipment and medium for predicting the load of a distribution transformer. Background Technology

[0002] As a crucial link connecting the power transmission network and users, the accurate forecasting of distribution transformer loads is essential for the rational allocation of power resources, the stable operation of the power grid, and the economic dispatching of power companies. With rapid socio-economic development and the widespread application of various electrical devices, the power load carried by distribution transformers is exhibiting increasingly complex characteristics. From the demand side, improved living standards have led to a significant increase in the types and frequency of household appliances. Prolonged operation of air conditioners during hot summers and frequent use of various heating devices in winter result in significant seasonal and diurnal variations in residential electricity load. Simultaneously, the diversity of commercial activities also impacts distribution transformer loads; the operating hours and electricity consumption patterns of shopping malls, office buildings, and other venues differ, and there are significant variations in electricity load between weekdays and weekends. In industrial production, the production processes and electricity consumption patterns of different industries vary considerably, and the production schedules of some high-energy-consuming enterprises can cause significant fluctuations in distribution transformer loads.

[0003] Currently, distribution transformer load forecasting methods mainly include traditional methods and machine learning-based methods. Traditional methods, such as time series analysis, assume that load data has stationary and linear characteristics. However, actual distribution transformer load data is affected by a variety of complex factors, exhibiting significant nonlinearity and volatility, resulting in low prediction accuracy. Time series analysis methods struggle to accurately capture sudden load changes during unexpected weather changes or special holidays. While regression analysis can consider the impact of multiple factors on the load, its ability to handle complex nonlinear relationships is limited.

[0004] Machine learning-based methods, such as artificial neural networks (ANNs), while possessing some nonlinear fitting capabilities, are prone to overfitting, especially with limited data, resulting in poor model generalization. Support vector machines (SVMs) offer advantages in handling small sample data, but their computational complexity is high for large-scale distribution transformer load data, and parameter selection is difficult, with different parameter settings potentially leading to significant variations in prediction results. Furthermore, existing methods are ineffective at handling noise in distribution transformer load data; electromagnetic interference and measurement errors in the power system can interfere with the model's learning of the true load patterns, further reducing prediction accuracy. In summary, existing distribution transformer load forecasting methods struggle to meet the demands for high-precision prediction when facing complex and variable actual load conditions, and urgently require improvement.

[0005] Therefore, how to provide a method, system, equipment, and medium for predicting the load of distribution transformers is an urgent problem to be solved. Summary of the Invention

[0006] This invention provides a method, system, device, and medium for predicting the load of a distribution transformer, in order to solve the problems mentioned above in the prior art.

[0007] According to a first aspect of the present invention, a method for predicting the load of a distribution transformer is provided.

[0008] In one embodiment, the distribution transformer load forecasting method, system, equipment, and medium include:

[0009] Collect historical load data and raw load data of the target distribution transformer, and normalize the historical load data and raw load data to obtain the distribution transformer load data;

[0010] The load data of the distribution transformer is decomposed using the variational mode decomposition algorithm to obtain a set of stationary intrinsic mode function subsequences;

[0011] An improved deep learning model is used to diversify the processing of stationary intrinsic mode function subsequences to predict the load sequence of the target distribution transformer.

[0012] Based on the load sequence of the target distribution transformer, the parameters of the improved deep learning model are optimized using an adaptive moment estimation optimization algorithm, and the performance of the improved deep learning model is evaluated using evaluation metrics.

[0013] In one embodiment, the process of collecting historical load data and raw load data of the target distribution transformer, and normalizing the historical load data and raw load data to obtain distribution transformer load data includes:

[0014] Retrieve the minimum and maximum values ​​from historical load data;

[0015] Calculate the difference between the original load data and the minimum value;

[0016] Calculate the difference between the minimum and maximum values;

[0017] The load data of the distribution transformer is obtained by calculating the ratio of the difference between the original load data and the minimum value to the difference between the minimum and the maximum value, thus eliminating the dimensional differences.

[0018] In one embodiment, the decomposition of the distribution transformer load data using the variational mode decomposition algorithm yields a set of stationary intrinsic mode function subsequences, including:

[0019] The variational mode decomposition algorithm is used to decompose the load data of the distribution transformer into subsequence groups with different center frequencies, and to remove noise and interference from the original data.

[0020] The number of mode decompositions is determined based on the characteristics of the subsequence set. Based on the constraints, the optimal center frequency and bandwidth in the subsequence set are searched to separate the intrinsic mode functions and obtain a stationary intrinsic mode function subsequence set.

[0021] In one embodiment, the formula for the constraint is:

[0022]

[0023] In the formula, K represents the number of mode decompositions; f indicates that the original signal has k components; This represents the time partial differential operator; t represents the time variable; j represents the imaginary unit; The square of the L2 norm is represented; st represents the constraints of the explicit optimization problem; u k ω k Let be the k-th component and its center frequency, respectively; δ(t) represents the Dirac function; * represents the convolution operator. Introducing the Lagrange multiplication operator, the constrained variational problem is transformed into an unconstrained variational problem, yielding the augmented Lagrange expression:

[0024]

[0025] In the formula, α represents the quadratic penalty factor; L represents the augmented Lagrange function; and λ represents the Lagrange multiplier.

[0026] In one embodiment, the method of using an improved deep learning model to diversify the set of stationary intrinsic mode function subsequences to predict the load sequence of the target distribution transformer includes:

[0027] By using a time attention mechanism to weight the stationary intrinsic mode function subsequence group, the key time steps of the distribution transformer load data change are automatically focused, and the feature vector after time attention processing is obtained.

[0028] By using convolutional neural networks to perform convolution operations on feature vectors, local spatial features in the stable intrinsic mode function subsequence group are extracted. Furthermore, by using convolution and pooling operations, the data dimensionality is reduced and local key features are extracted.

[0029] Long Short-Term Memory (LSTM) networks are used to capture long-term time dependencies in stable intrinsic mode function subsequences through a gating structure, in order to adapt to the dynamic changes in distribution transformer load data over time steps.

[0030] By concatenating local key features and long-term time dependencies with feature vectors through residuals, the load sequence of the target distribution transformer is predicted to avoid the gradient vanishing problem.

[0031] In one embodiment, the weighting of the stationary intrinsic mode function subsequence group using a time attention mechanism to automatically focus on the key time steps of the distribution transformer load data changes, resulting in a feature vector after time attention processing, includes:

[0032] Based on a set of stationary intrinsic mode function subsequences, attention weights are calculated at each time step using a time attention mechanism.

[0033] The attention score is obtained by calculating the correlation between each time step and other time steps in a set of subsequences based on stationary intrinsic mode functions using a time attention mechanism.

[0034] Attention scores are converted into attention weights using a normalized exponential function;

[0035] The feature vector after temporal attention processing is obtained by multiplying the data at each time step with the corresponding attention weight and summing the results.

[0036] In one embodiment, the step of optimizing the parameters of the improved deep learning model using an adaptive moment estimation optimization algorithm based on the load sequence of the target distribution transformer, and evaluating the performance of the improved deep learning model using evaluation metrics, includes:

[0037] The optimal parameters of the Long Short-Term Memory network are found by using an adaptive moment estimation optimization algorithm, and the optimal parameters are set at the maximum number of iterations, the learning rate decay period, the initial learning rate, and the learning rate decay factor during the training process.

[0038] The performance of the improved deep learning model was evaluated using root mean square error and mean absolute percentage error as evaluation metrics. The smaller the value, the higher the prediction accuracy of the improved deep learning model.

[0039] According to a second aspect of the present invention, a load prediction system for distribution transformers is provided.

[0040] In one embodiment, a distribution transformer load forecasting system includes:

[0041] The data collection and processing module is used to collect historical load data and raw load data of the target distribution transformer, and to normalize the historical load data and raw load data to obtain the distribution transformer load data.

[0042] The data decomposition module is used to decompose the load data of the distribution transformer using the variational mode decomposition algorithm to obtain a set of stationary intrinsic mode function subsequences.

[0043] The load forecasting module is used to perform multivariate processing on the stationary intrinsic mode function subsequence group using an improved deep learning model to predict the load sequence of the target distribution transformer.

[0044] The model optimization and evaluation module is used to optimize the parameters of the improved deep learning model based on the load sequence of the target distribution transformer using an adaptive moment estimation optimization algorithm, and to evaluate the performance of the improved deep learning model using evaluation metrics.

[0045] In one embodiment, the process of collecting historical load data and raw load data of the target distribution transformer, and normalizing the historical load data and raw load data to obtain distribution transformer load data includes:

[0046] Retrieve the minimum and maximum values ​​from historical load data;

[0047] Calculate the difference between the original load data and the minimum value;

[0048] Calculate the difference between the minimum and maximum values;

[0049] The load data of the distribution transformer is obtained by calculating the ratio of the difference between the original load data and the minimum value to the difference between the minimum and the maximum value, thus eliminating the dimensional differences.

[0050] In one embodiment, the decomposition of the distribution transformer load data using the variational mode decomposition algorithm yields a set of stationary intrinsic mode function subsequences, including:

[0051] The variational mode decomposition algorithm is used to decompose the load data of the distribution transformer into subsequence groups with different center frequencies, and to remove noise and interference from the original data.

[0052] The number of mode decompositions is determined based on the characteristics of the subsequence set. Based on the constraints, the optimal center frequency and bandwidth in the subsequence set are searched to separate the intrinsic mode functions and obtain a stationary intrinsic mode function subsequence set.

[0053] In one embodiment, the formula for the constraint is:

[0054]

[0055]

[0056] In the formula, K represents the number of mode decompositions; f indicates that the original signal has k components; This represents the time partial differential operator; t represents the time variable; j represents the imaginary unit; The square of the L2 norm is represented; st represents the constraints of the explicit optimization problem; u k ω kLet be the k-th component and its center frequency, respectively; δ(t) represents the Dirac function; * represents the convolution operator. Introducing the Lagrange multiplication operator, the constrained variational problem is transformed into an unconstrained variational problem, yielding the augmented Lagrange expression:

[0057]

[0058] In the formula, α represents the quadratic penalty factor; L represents the augmented Lagrange function; and λ represents the Lagrange multiplier.

[0059] In one embodiment, the method of using an improved deep learning model to diversify the set of stationary intrinsic mode function subsequences to predict the load sequence of the target distribution transformer includes:

[0060] By using a time attention mechanism to weight the stationary intrinsic mode function subsequence group, the key time steps of the distribution transformer load data change are automatically focused, and the feature vector after time attention processing is obtained.

[0061] By using convolutional neural networks to perform convolution operations on feature vectors, local spatial features in the stable intrinsic mode function subsequence group are extracted. Furthermore, by using convolution and pooling operations, the data dimensionality is reduced and local key features are extracted.

[0062] Long Short-Term Memory (LSTM) networks are used to capture long-term time dependencies in stable intrinsic mode function subsequences through a gating structure, in order to adapt to the dynamic changes in distribution transformer load data over time steps.

[0063] By concatenating local key features and long-term time dependencies with feature vectors through residuals, the load sequence of the target distribution transformer is predicted to avoid the gradient vanishing problem.

[0064] In one embodiment, the weighting of the stationary intrinsic mode function subsequence group using a time attention mechanism to automatically focus on the key time steps of the distribution transformer load data changes, resulting in a feature vector after time attention processing, includes:

[0065] Based on a set of stationary intrinsic mode function subsequences, attention weights are calculated at each time step using a time attention mechanism.

[0066] The attention score is obtained by calculating the correlation between each time step and other time steps in a set of subsequences based on stationary intrinsic mode functions using a time attention mechanism.

[0067] Attention scores are converted into attention weights using a normalized exponential function;

[0068] The feature vector after temporal attention processing is obtained by multiplying the data at each time step with the corresponding attention weight and summing the results.

[0069] In one embodiment, the step of optimizing the parameters of the improved deep learning model using an adaptive moment estimation optimization algorithm based on the load sequence of the target distribution transformer, and evaluating the performance of the improved deep learning model using evaluation metrics, includes:

[0070] The optimal parameters of the Long Short-Term Memory network are found by using an adaptive moment estimation optimization algorithm, and the optimal parameters are set at the maximum number of iterations, the learning rate decay period, the initial learning rate, and the learning rate decay factor during the training process.

[0071] The performance of the improved deep learning model was evaluated using root mean square error and mean absolute percentage error as evaluation metrics. The smaller the value, the higher the prediction accuracy of the improved deep learning model.

[0072] According to a third aspect of the present invention, a computer device is provided.

[0073] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0074] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0075] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0076] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0077] (1) This invention combines VMD decomposition with the improved deep learning model described above, giving full play to the advantages of VMD in data preprocessing and the powerful nonlinear fitting capability of the deep learning model, thereby improving the accuracy and robustness of distribution transformer load forecasting. This improved method aims to solve the problems of existing technologies being unable to effectively process complex distribution transformer load data, difficult to accurately capture load change patterns, and sensitive to noise, so as to achieve high-precision forecasting of distribution transformer load and meet the growing demand of power systems for load forecasting.

[0078] (2) This invention utilizes the VMD algorithm to decompose the original distribution transformer load data, effectively suppressing noise and interference, resulting in cleaner data input to the subsequent deep learning model and clearer feature representation. The temporal attention mechanism can focus on key time step information, weakening the interference of irrelevant features, thereby improving the accuracy of feature extraction. CNN and LSTM perform deep feature mining on the load sequence from the spatial and temporal dimensions, respectively, which can fully learn the complex evolution law of distribution transformer load data. At the same time, the introduction of residual connections effectively alleviates the gradient vanishing problem in the deep training process, ensuring the stability and convergence of the model, enabling it to capture richer and more discriminative features.

[0079] (3) In actual power systems, distribution transformer load data is inevitably affected by noise such as electromagnetic interference and measurement errors. The introduction of VMD significantly enhances the model's robustness to noise, preventing large deviations in prediction results due to noise. The improved deep learning structure, especially the design of residual connections, further enhances the model's generalization ability, enabling it to adapt to distribution transformer load data with different characteristics. Even when the data distribution changes or new influencing factors appear, the model can still maintain high prediction accuracy and stability.

[0080] (4) The advance load forecasting method of this invention can provide strong support for power grid operation and maintenance. Maintenance personnel can use this method to identify potential overload or fault risks, rationally allocate inspection and maintenance resources, thereby reducing equipment failure rates and extending their service life. When the improved deep learning model predicts that a distribution transformer has an overload risk during a specific period, maintenance personnel can conduct inspections and maintenance in advance to avoid equipment failures and power outages caused by overload. This method not only improves the power supply reliability of the distribution system but also provides technical support for ensuring continuous and stable power supply for users.

[0081] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0082] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0083] Figure 1 This is one of the flowcharts illustrating a load forecasting method for a distribution transformer according to an exemplary embodiment;

[0084] Figure 2 This is a schematic block diagram of a distribution transformer load prediction system according to an exemplary embodiment;

[0085] Figure 3This is a schematic diagram of the structure of a computer device according to an exemplary embodiment;

[0086] Figure 4 This is a structural diagram of an LSTM gated unit according to an exemplary embodiment;

[0087] Figure 5 This is an LSTM timing unfolding diagram illustrated according to an exemplary embodiment;

[0088] Figure 6 This is a second flowchart illustrating a load forecasting method for a distribution transformer according to an exemplary embodiment;

[0089] Figure 7 This is a VMD-based distribution transformer load decomposition diagram illustrated according to an exemplary embodiment;

[0090] Figure 8 This is a comparison chart of load forecasting analysis according to an exemplary embodiment;

[0091] Figure 9 This is a comparison chart showing the results of an exemplary embodiment. Detailed Implementation

[0092] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0093] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0094] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0095] Figure 1 An embodiment of a distribution transformer load prediction method according to the present invention is shown.

[0096] In this optional embodiment, the method for predicting the load of a distribution transformer includes:

[0097] Step 101: Collect historical load data and raw load data of the target distribution transformer, and normalize the historical load data and raw load data to obtain the distribution transformer load data;

[0098] Step 102: Use the variational mode decomposition (VMD) algorithm to decompose the distribution transformer load data to obtain a set of stationary intrinsic mode functions (IMF) subsequences;

[0099] Step 103: Use an improved deep learning model to perform multivariate processing on the stationary intrinsic mode function subsequence group to predict the load sequence of the target distribution transformer.

[0100] Step 104: Based on the load sequence of the target distribution transformer, the parameters of the improved deep learning model are optimized using the adaptive moment estimation optimization algorithm (Adam optimization algorithm), and the performance of the improved deep learning model is evaluated using evaluation metrics.

[0101] Specifically, such as Figure 6 As shown, the VMD algorithm is used to decompose the original distribution transformer load data. The algorithm adaptively decomposes complex distribution transformer load signals into multiple subsequences with different center frequencies. Compared to traditional methods such as empirical mode decomposition, VMD has a solid theoretical foundation and exhibits stronger robustness in handling noise and non-stationary signals. Through VMD decomposition, noise and interference in the original data can be effectively removed, and the distribution transformer load data can be decomposed into subsequences with different frequency characteristics, making the data characteristics clearer and providing higher-quality data for subsequent prediction models.

[0102] After decomposing the subsequences, a deep learning model integrating temporal attention, convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and residual connections is constructed. The temporal attention mechanism helps the model automatically focus on key time steps closely related to changes in the transformer load when processing subsequences, highlighting important information and improving the targeting of feature extraction. The CNN part is responsible for mining local spatial features in the subsequences, extracting effective features from the data and reducing data dimensionality through convolution and pooling operations. LSTM, with its special gating structure, can effectively capture long-term temporal dependencies in the subsequences, adapting to the dynamic changes in transformer load over time. The introduction of residual connections provides a shortcut for feature transfer in the model, avoiding the gradient vanishing problem during deep network training, enabling the model to learn deeper features and enhancing its expressive power.

[0103] In this optional embodiment, the process of collecting historical load data and raw load data of the target distribution transformer, and normalizing the historical load data and raw load data to obtain distribution transformer load data includes:

[0104] Retrieve the minimum and maximum values ​​from historical load data;

[0105] Calculate the difference between the original load data and the minimum value;

[0106] Calculate the difference between the minimum and maximum values;

[0107] The load data of the distribution transformer is obtained by calculating the ratio of the difference between the original load data and the minimum value to the difference between the minimum and the maximum value, thus eliminating the dimensional differences.

[0108] Specifically, historical load data of the target distribution transformers should be collected. The data time span should be long enough to cover various situations such as different seasons, weekdays, and special dates to ensure data representativeness. The collected raw distribution transformer load data should be normalized using the min-max normalization method, mapping the data of each variable to the interval [0, 1]. The calculation formula for the min-max normalization method is as follows:

[0109]

[0110] In the formula, x represents the original load data; x min and x max These represent the minimum and maximum values ​​of historical load data, respectively; x norm This represents the normalized value.

[0111] In this optional embodiment, the decomposition of the distribution transformer load data using the variational mode decomposition algorithm yields a set of stationary intrinsic mode function subsequences, including:

[0112] The variational mode decomposition algorithm is used to decompose the load data of the distribution transformer into subsequence groups with different center frequencies, and to remove noise and interference from the original data.

[0113] The number of mode decompositions is determined based on the characteristics of the subsequence set. Based on the constraints, the optimal center frequency and bandwidth in the subsequence set are searched to separate the intrinsic mode functions and obtain a stationary intrinsic mode function subsequence set.

[0114] Specifically, VMD is an adaptive signal decomposition method that determines the number of mode decompositions based on the characteristics of the sequence itself. This method has a more robust mathematical theoretical foundation. During signal decomposition, the optimal center frequency and bandwidth of each mode can be searched, achieving effective separation of IMFs. Each IMF is a finite bandwidth signal, and its stationarity is higher than that of the original signal. The dissolved gas concentration sequence in oil is a highly nonlinear and non-stationary time series, which is not conducive to accurate prediction by the model. Using VMD to decompose the original sequence into a series of more stationary signals helps improve its prediction accuracy.

[0115] The core idea of ​​VMD is to construct and solve variational problems. Assume the original signal f is decomposed into k components. To ensure that each mode has a center frequency and the sum of its bandwidths is minimized, the constraints are as follows.

[0116] In this optional embodiment, the formula for the constraint condition is:

[0117]

[0118] In the formula, K represents the number of mode decompositions; f indicates that the original signal has k components; This represents the time partial differential operator, used for differentiating signals along their time dimension; t represents the time variable; j represents the imaginary unit. The square of the L2 norm is represented; st represents the constraints of the explicit optimization problem; u k ω k Let be the k-th component and its center frequency, respectively; δ(t) represents the Dirac function; * represents the convolution operator. Introducing the Lagrange multiplication operator, the constrained variational problem is transformed into an unconstrained variational problem, yielding the augmented Lagrange expression:

[0119]

[0120] In the formula, α represents the quadratic penalty factor, which aims to reduce Gaussian noise interference. The modal components and their optimal center frequencies can be obtained by using the alternating direction multiplier iterative algorithm; L represents the augmented Lagrangian function; and λ represents the Lagrange multiplier.

[0121] In this optional embodiment, the step of using an improved deep learning model to diversify the stationary intrinsic mode function subsequence set to predict the load sequence of the target distribution transformer includes:

[0122] By using a time attention mechanism to weight the stationary intrinsic mode function subsequence group, the key time steps of the distribution transformer load data change are automatically focused, and the feature vector after time attention processing is obtained.

[0123] By using convolutional neural networks (CNNs) to perform convolution operations on feature vectors, local spatial features in the stationary intrinsic mode function subsequence group are extracted. Furthermore, by using convolution and pooling operations, the data dimensionality is reduced and key local features are extracted.

[0124] Long Short-Term Memory (LSTM) networks are used to capture long-term time dependencies in a set of stationary intrinsic mode function subsequences through a gating structure (gating unit) to adapt to the dynamic changes in distribution transformer load data over time steps;

[0125] By concatenating local key features and long-term time dependencies with feature vectors through residuals, the load sequence of the target distribution transformer is predicted to avoid the gradient vanishing problem.

[0126] Specifically, each subsequence obtained from VMD decomposition is taken as input, and the temporal attention mechanism calculates an attention weight for each time step.

[0127] The feature vectors processed by time attention are fed into the CNN layer. The CNN layer consists of multiple convolutional kernels; for example, a kernel size of 3×1, a stride of 1, and 16 kernels can be set. The convolutional kernels slide across the feature vectors to perform convolution operations, extracting local spatial features from the subsequence. The result after the convolution operation undergoes a non-linear transformation using the ReLU activation function, as shown in the formula:

[0128] y = max(0,x);

[0129] In the formula, x represents the output of the convolution operation; y represents the output after ReLU activation.

[0130] Max pooling is employed with a 2×1 pooling window and a stride of 2 to reduce the dimensionality of the convolutional features, highlighting key features. The pooled features are then input into an LSTM layer. LSTM is a variant of RNN that adds gating units to measure the importance of historical information to the current time step. It is a recurrent neural network model capable of preserving and transmitting long-term temporal information, solving the gradient vanishing problem during traditional RNN training. The structure of the gating unit and the temporal unfolding diagram of LSTM are shown below. Figure 4 and Figure 5 As shown.

[0131] The gating unit is the core component of LSTM, and its states at time t are C t and S t C t It contains long-term historical information about the sequence, S t It contains short-term historical information of the sequence. The input of the LSTM at time t is x. t The output value S at the previous time step t-1 and C t-1 The output at time t is the long-term memory state C. t and short-term memory state S t In the gating unit, the forget gate, input gate, and output gate control C respectively. t-1 For C t x t For C t and C t For S t The degree of influence. The formulas for the forget gate, input gate, and output gate are as follows:

[0132] f t =σ(W f S t-1 +W f x t +b f );

[0133] i t =σ(W i S t-1 +W i x t +b i );

[0134] o t =σ(W o S t-1 +W o x t +b o );

[0135] In the formula, f t i t and o t These represent the outputs of the forget gate, input gate, and output gate, respectively; W f W i and W o b are the weight matrices for the forget gate, input gate, and output gate, respectively; f b i and b o These are the bias terms for the forget gate, input gate, and output gate, respectively; S t-1 σ represents the output value at the previous time step; σ represents the sigmoid activation function. The output of the LSTM is determined by the following formula:

[0136]

[0137] In the formula, W represents the state of the input unit at time t; c b represents the weight matrix of the input unit; c The output unit bias is represented by σ; the sigmoid activation function is represented by σ; the hyperbolic tangent activation function is represented by tanh; and C represents the output unit bias. t Represents long-term historical information of the sequence; This indicates that the matrix is ​​multiplied by corresponding elements.

[0138] The output of the CNN layer is directly added to the input of the LSTM layer, using the following formula:

[0139] x LSTM =x CNN +x input ;

[0140] In the formula, xLSTM x represents the input to the LSTM layer. CNN x represents the output of the CNN layer. input This represents the original input to the CNN layer. During model training, gradients can be backpropagated more easily, avoiding the vanishing gradient problem and enabling the model to learn deeper features.

[0141] In this optional embodiment, the weighting of the stationary intrinsic mode function subsequence group using a time attention mechanism, automatically focusing on the key time steps of the distribution transformer load data changes, and obtaining the feature vector after time attention processing includes:

[0142] Based on a set of stationary intrinsic mode function subsequences, attention weights are calculated at each time step using a time attention mechanism.

[0143] The attention score is obtained by calculating the correlation between each time step and other time steps in a set of subsequences based on stationary intrinsic mode functions using a time attention mechanism.

[0144] Attention scores are converted into attention weights using a normalized exponential function (Softmax function);

[0145] The feature vector after temporal attention processing is obtained by multiplying the data at each time step with the corresponding attention weight and summing the results.

[0146] Specifically, each subsequence obtained from VMD decomposition is used as input, and the temporal attention mechanism calculates an attention weight for each time step. By calculating the correlation between each time step in the subsequence and other time steps, an attention score is obtained, and then the attention score is converted into attention weights through the Softmax function.

[0147]

[0148] In the formula, α t e represents the attention weight at time step t; t The attention score represents time step t; T represents the total number of time steps in the subsequence; exp represents the exponential function. Represents all exp(e) from the 1st category to the Tth category. t Then, sum the data at each time step. Next, multiply the data at each time step by its corresponding attention weight and sum them to obtain the feature vector after time attention processing.

[0149] In this optional embodiment, the step of optimizing the parameters of the improved deep learning model using an adaptive moment estimation optimization algorithm based on the load sequence of the target distribution transformer, and evaluating the performance of the improved deep learning model using evaluation metrics, includes:

[0150] The optimal parameters of the Long Short-Term Memory network are found by using an adaptive moment estimation optimization algorithm, and the optimal parameters are set at the maximum number of iterations, the learning rate decay period, the initial learning rate, and the learning rate decay factor during the training process.

[0151] The performance of the improved deep learning model was evaluated using root mean square error and mean absolute percentage error as evaluation metrics. The smaller the value, the higher the prediction accuracy of the improved deep learning model.

[0152] Specifically, the weights and biases between layers have a significant impact on the final prediction results. The Adam optimization algorithm is a commonly used deep learning model training method, widely adopted due to its excellent performance. Therefore, the Adam optimization algorithm is used to find the optimal parameters for the LSTM. During training, the maximum number of iterations, the learning rate decay period, the initial learning rate, and the learning rate decay factor are set to 250, 125, 0.005, and 0.2, respectively.

[0153] The following two evaluation metrics are selected to assess the performance of the prediction model: Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAS).

[0154] Error (MAPE). The calculation formulas are shown below:

[0155]

[0156] In the formula: y i , These represent the actual value and the predicted value, respectively; n represents the number of data points used for verification; y RMSE y represents the root mean square error; MAPE It represents the mean absolute percentage error.

[0157] Figure 2 An embodiment of a distribution transformer load prediction system according to the present invention is shown.

[0158] In this optional embodiment, a distribution transformer load forecasting system includes:

[0159] The data collection and processing module 201 is used to collect historical load data and raw load data of the target distribution transformer, and to normalize the historical load data and raw load data to obtain the distribution transformer load data.

[0160] Data decomposition module 202 is used to decompose the load data of the distribution transformer using the variational mode decomposition algorithm to obtain a set of stationary intrinsic mode function subsequences.

[0161] The load forecasting module 203 is used to perform multivariate processing on the stationary intrinsic mode function subsequence group using an improved deep learning model to predict the load sequence of the target distribution transformer.

[0162] The model optimization and evaluation module 204 is used to optimize the parameters of the improved deep learning model based on the load sequence of the target distribution transformer through an adaptive moment estimation optimization algorithm, and to evaluate the performance of the improved deep learning model using evaluation metrics.

[0163] In this optional embodiment, the step of collecting historical load data and original load data of the target distribution transformer, and normalizing the historical load data and original load data to obtain distribution transformer load data includes: obtaining the minimum and maximum values ​​in the historical load data; calculating the difference between the original load data and the minimum value; calculating the difference between the minimum and the maximum value; and calculating the ratio of the difference between the original load data and the minimum value to the difference between the minimum and the maximum value to obtain distribution transformer load data, thereby eliminating dimensional differences.

[0164] In this optional embodiment, the decomposition of the distribution transformer load data using the variational mode decomposition algorithm yields a set of stationary intrinsic mode function subsequences, including:

[0165] The variational mode decomposition algorithm is used to decompose the load data of the distribution transformer into subsequence groups with different center frequencies, and to remove noise and interference from the original data.

[0166] The number of mode decompositions is determined based on the characteristics of the subsequence set. Based on the constraints, the optimal center frequency and bandwidth in the subsequence set are searched to separate the intrinsic mode functions and obtain a stationary intrinsic mode function subsequence set.

[0167] In this optional embodiment, the formula for the constraint condition is:

[0168]

[0169] In the formula, K represents the number of mode decompositions; f indicates that the original signal has k components; This represents the time partial differential operator, used for differentiating signals along their time dimension; t represents the time variable; j represents the imaginary unit. The square of the L2 norm is represented; st represents the constraints of the explicit optimization problem; u k ω k Let be the k-th component and its center frequency, respectively; δ(t) represents the Dirac function; * represents the convolution operator. Introducing the Lagrange multiplication operator, the constrained variational problem is transformed into an unconstrained variational problem, yielding the augmented Lagrange expression:

[0170]

[0171] In the formula, α represents the quadratic penalty factor, which aims to reduce Gaussian noise interference. The modal components and their optimal center frequencies can be obtained by using the alternating direction multiplier iterative algorithm; L represents the augmented Lagrangian function; and λ represents the Lagrange multiplier.

[0172] In this optional embodiment, the step of using an improved deep learning model to diversify the stationary intrinsic mode function subsequence set to predict the load sequence of the target distribution transformer includes:

[0173] By using a time attention mechanism to weight the stationary intrinsic mode function subsequence group, the key time steps of the distribution transformer load data change are automatically focused, and the feature vector after time attention processing is obtained.

[0174] By using convolutional neural networks to perform convolution operations on feature vectors, local spatial features in the stable intrinsic mode function subsequence group are extracted. Furthermore, by using convolution and pooling operations, the data dimensionality is reduced and local key features are extracted.

[0175] Long Short-Term Memory (LSTM) networks are used to capture long-term time dependencies in stable intrinsic mode function subsequences through a gating structure, in order to adapt to the dynamic changes in distribution transformer load data over time steps.

[0176] By concatenating local key features and long-term time dependencies with feature vectors through residuals, the load sequence of the target distribution transformer is predicted to avoid the gradient vanishing problem.

[0177] In this optional embodiment, the weighting of the stationary intrinsic mode function subsequence group using a time attention mechanism, automatically focusing on the key time steps of the distribution transformer load data changes, and obtaining the feature vector after time attention processing includes:

[0178] Based on a set of stationary intrinsic mode function subsequences, attention weights are calculated at each time step using a time attention mechanism.

[0179] The attention score is obtained by calculating the correlation between each time step and other time steps in a set of subsequences based on stationary intrinsic mode functions using a time attention mechanism.

[0180] Attention scores are converted into attention weights using a normalized exponential function;

[0181] The feature vector after temporal attention processing is obtained by multiplying the data at each time step with the corresponding attention weight and summing the results.

[0182] In this optional embodiment, the step of optimizing the parameters of the improved deep learning model using an adaptive moment estimation optimization algorithm based on the load sequence of the target distribution transformer, and evaluating the performance of the improved deep learning model using evaluation metrics, includes:

[0183] The optimal parameters of the Long Short-Term Memory network are found by using an adaptive moment estimation optimization algorithm, and the optimal parameters are set at the maximum number of iterations, the learning rate decay period, the initial learning rate, and the learning rate decay factor during the training process.

[0184] The performance of the improved deep learning model was evaluated using root mean square error and mean absolute percentage error as evaluation metrics. The smaller the value, the higher the prediction accuracy of the improved deep learning model.

[0185] To facilitate understanding of the above technical solutions of the present invention, the following further describes the above technical solutions of the present invention from the perspectives of architecture and principle, as follows:

[0186] Taking a 10kV distribution transformer in a residential community of a certain city as the research object, hourly historical load data from January 1 to December 31, 2023 were collected, and the date type and seasonal information of each data point were labeled. Minimum-maximum normalization was used to map the original load to the [0, 1] interval to eliminate magnitude differences. Subsequently, six stationary subsequences were obtained through adaptive decomposition using a variational mode decomposition algorithm (e.g., ...). Figure 7 As shown in the figure, subsequences 1-2 correspond to the high-frequency components of intraday short-term fluctuations, subsequences 3-4 correspond to the medium-frequency components of weekly patterns, and subsequences 5-6 correspond to the low-frequency components of seasonal trends. Figure 7 The process of VMD effectively separating the multi-scale frequency characteristics of the load is presented in a visual way, while reducing noise interference.

[0187] Six subsequences were divided into training and testing sets in a 7:3 ratio. An improved deep learning model was constructed, incorporating a temporal attention mechanism, a convolutional neural network (CNN), a long short-term memory network (LSTM), and residual connections. The CNN layer used 3×1 convolutional kernels, 16 kernels, and a 2×1 max-pooling window to extract local spatial features of the subsequences. The LSTM layer used 64 hidden units to capture long-term temporal dependencies. The temporal attention mechanism focused on moments of load abrupt change, and residual connections prevented gradient vanishing during model training. After training with the Adam optimization algorithm (maximum number of iterations 250, learning rate decay period 125, initial learning rate 0.005, learning rate decay factor 0.2), the model's prediction performance on the test set was as follows: Figure 8As shown, its root mean square error (RMSE) is 0.032 and mean absolute percentage error (MAPE) is 2.1%, significantly better than the traditional ARIMA method (RMSE = 0.085, MAPE = 6.8%) and the ordinary LSTM model (RMSE = 0.051, MAPE = 3.7%). The performance metrics of each model are compared as follows: Figure 9 The results are presented intuitively in the form of a bar chart. Furthermore, the model maintains stable predictive performance even under non-stationary scenarios such as sudden weather events and special holidays, fully demonstrating the effectiveness and robustness of the proposed method in distribution transformer load forecasting.

[0188] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0189] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0190] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0191] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0192] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0193] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A method for predicting the load of a distribution transformer, characterized in that, include: Collect historical load data and raw load data of the target distribution transformer, and normalize the historical load data and raw load data to obtain the distribution transformer load data; The load data of the distribution transformer is decomposed using the variational mode decomposition algorithm to obtain a set of stationary intrinsic mode function subsequences; An improved deep learning model is used to diversify the processing of stationary intrinsic mode function subsequences to predict the load sequence of the target distribution transformer. Based on the load sequence of the target distribution transformer, the parameters of the improved deep learning model are optimized using an adaptive moment estimation optimization algorithm, and the performance of the improved deep learning model is evaluated using evaluation metrics.

2. The method for predicting the load of a distribution transformer according to claim 1, characterized in that, The process involves collecting historical load data and raw load data of the target distribution transformer, and then normalizing the historical load data and raw load data to obtain the distribution transformer load data, including: Retrieve the minimum and maximum values ​​from historical load data; Calculate the difference between the original load data and the minimum value; Calculate the difference between the minimum and maximum values; The load data of the distribution transformer is obtained by calculating the ratio of the difference between the original load data and the minimum value to the difference between the minimum and the maximum value, thus eliminating the dimensional differences.

3. The method for predicting the load of a distribution transformer according to claim 1, characterized in that, The variational mode decomposition algorithm is used to decompose the load data of the distribution transformer, resulting in a set of stationary intrinsic mode function subsequences, including: The variational mode decomposition algorithm is used to decompose the load data of the distribution transformer into subsequence groups with different center frequencies, and to remove noise and interference from the original data. The number of mode decompositions is determined based on the characteristics of the subsequence set. Based on the constraints, the optimal center frequency and bandwidth in the subsequence set are searched to separate the intrinsic mode functions and obtain a stationary intrinsic mode function subsequence set.

4. The method for predicting the load of a distribution transformer according to claim 3, characterized in that, The formula for the constraint condition is: In the formula, K represents the number of mode decompositions; f indicates that the original signal has k components; This represents the time partial differential operator; t represents the time variable; j represents the imaginary unit; The square of the L2 norm is represented; st represents the constraints of the explicit optimization problem; u k ω k Let be the k-th component and its center frequency, respectively; δ(t) represents the Dirac function; * represents the convolution operator. Introducing the Lagrange multiplication operator, the constrained variational problem is transformed into an unconstrained variational problem, yielding the augmented Lagrange expression: In the formula, α represents the quadratic penalty factor; L represents the augmented Lagrange function; and λ represents the Lagrange multiplier.

5. The method for predicting the load of a distribution transformer according to claim 1, characterized in that, The method of using an improved deep learning model to perform multivariate processing on the stationary intrinsic mode function subsequence set to predict the load sequence of the target distribution transformer includes: By using a time attention mechanism to weight the stationary intrinsic mode function subsequence group, the key time steps of the distribution transformer load data change are automatically focused, and the feature vector after time attention processing is obtained. By using convolutional neural networks to perform convolution operations on feature vectors, local spatial features in the stable intrinsic mode function subsequence group are extracted. Furthermore, by using convolution and pooling operations, the data dimensionality is reduced and local key features are extracted. Long Short-Term Memory (LSTM) networks are used to capture long-term time dependencies in stable intrinsic mode function subsequences through a gating structure, in order to adapt to the dynamic changes in distribution transformer load data over time steps. By concatenating local key features and long-term time dependencies with feature vectors through residuals, the load sequence of the target distribution transformer is predicted to avoid the gradient vanishing problem.

6. The method for predicting the load of a distribution transformer according to claim 5, characterized in that, The time attention mechanism is used to weight the stationary intrinsic mode function subsequence group, automatically focusing on the key time steps of the distribution transformer load data changes, and obtaining the feature vector after time attention processing, including: Based on a set of stationary intrinsic mode function subsequences, attention weights are calculated at each time step using a time attention mechanism. The attention score is obtained by calculating the correlation between each time step and other time steps in a set of subsequences based on stationary intrinsic mode functions using a time attention mechanism. Attention scores are converted into attention weights using a normalized exponential function; The feature vector after temporal attention processing is obtained by multiplying the data at each time step with the corresponding attention weight and summing the results.

7. The method for predicting the load of a distribution transformer according to claim 5, characterized in that, The load sequence based on the target distribution transformer is used to optimize the parameters of the improved deep learning model through an adaptive moment estimation optimization algorithm, and the performance of the improved deep learning model is evaluated using evaluation metrics, including: The optimal parameters of the Long Short-Term Memory network are found by using an adaptive moment estimation optimization algorithm, and the optimal parameters are set at the maximum number of iterations, the learning rate decay period, the initial learning rate, and the learning rate decay factor during the training process. The performance of the improved deep learning model was evaluated using root mean square error and mean absolute percentage error as evaluation metrics. The smaller the value, the higher the prediction accuracy of the improved deep learning model.

8. A load forecasting system for distribution transformers, characterized in that, The system includes: The data collection and processing module is used to collect historical load data and raw load data of the target distribution transformer, and to normalize the historical load data and raw load data to obtain the distribution transformer load data. The data decomposition module is used to decompose the load data of the distribution transformer using the variational mode decomposition algorithm to obtain a set of stationary intrinsic mode function subsequences. The load forecasting module is used to perform multivariate processing on the stationary intrinsic mode function subsequence group using an improved deep learning model to predict the load sequence of the target distribution transformer. The model optimization and evaluation module is used to optimize the parameters of the improved deep learning model based on the load sequence of the target distribution transformer using an adaptive moment estimation optimization algorithm, and to evaluate the performance of the improved deep learning model using evaluation metrics.

9. A distribution transformer load prediction system according to claim 8, characterized in that, The process involves collecting historical load data and raw load data of the target distribution transformer, and then normalizing the historical load data and raw load data to obtain the distribution transformer load data, including: Retrieve the minimum and maximum values ​​from historical load data; Calculate the difference between the original load data and the minimum value; Calculate the difference between the minimum and maximum values; The load data of the distribution transformer is obtained by calculating the ratio of the difference between the original load data and the minimum value to the difference between the minimum and the maximum value, thus eliminating the dimensional differences.

10. A distribution transformer load prediction system according to claim 8, characterized in that, The variational mode decomposition algorithm is used to decompose the load data of the distribution transformer, resulting in a set of stationary intrinsic mode function subsequences, including: The variational mode decomposition algorithm is used to decompose the load data of the distribution transformer into subsequence groups with different center frequencies, and to remove noise and interference from the original data. The number of mode decompositions is determined based on the characteristics of the subsequence set. Based on the constraints, the optimal center frequency and bandwidth in the subsequence set are searched to separate the intrinsic mode functions and obtain a stationary intrinsic mode function subsequence set.

11. A distribution transformer load prediction system according to claim 10, characterized in that, The formula for the constraint condition is: In the formula, K represents the number of mode decompositions; f indicates that the original signal has k components; This represents the time partial differential operator; t represents the time variable; j represents the imaginary unit; The square of the L2 norm is represented; st represents the constraints of the explicit optimization problem; u k ω k Let be the k-th component and its center frequency, respectively; δ(t) represents the Dirac function; * represents the convolution operator. Introducing the Lagrange multiplication operator, the constrained variational problem is transformed into an unconstrained variational problem, yielding the augmented Lagrange expression: In the formula, α represents the quadratic penalty factor; L represents the augmented Lagrange function; and λ represents the Lagrange multiplier.

12. A distribution transformer load prediction system according to claim 8, characterized in that, The method of using an improved deep learning model to perform multivariate processing on the stationary intrinsic mode function subsequence set to predict the load sequence of the target distribution transformer includes: By using a time attention mechanism to weight the stationary intrinsic mode function subsequence group, the key time steps of the distribution transformer load data change are automatically focused, and the feature vector after time attention processing is obtained. By using convolutional neural networks to perform convolution operations on feature vectors, local spatial features in the stable intrinsic mode function subsequence group are extracted. Furthermore, by using convolution and pooling operations, the data dimensionality is reduced and local key features are extracted. Long Short-Term Memory (LSTM) networks are used to capture long-term time dependencies in stable intrinsic mode function subsequences through a gating structure, in order to adapt to the dynamic changes in distribution transformer load data over time steps. By concatenating local key features and long-term time dependencies with feature vectors through residuals, the load sequence of the target distribution transformer is predicted to avoid the gradient vanishing problem.

13. A distribution transformer load prediction system according to claim 12, characterized in that, The time attention mechanism is used to weight the stationary intrinsic mode function subsequence group, automatically focusing on the key time steps of the distribution transformer load data changes, and obtaining the feature vector after time attention processing, including: Based on a set of stationary intrinsic mode function subsequences, attention weights are calculated at each time step using a time attention mechanism. The attention score is obtained by calculating the correlation between each time step and other time steps in a set of subsequences based on stationary intrinsic mode functions using a time attention mechanism. Attention scores are converted into attention weights using a normalized exponential function; The feature vector after temporal attention processing is obtained by multiplying the data at each time step with the corresponding attention weight and summing the results.

14. A distribution transformer load prediction system according to claim 12, characterized in that, The load sequence based on the target distribution transformer is used to optimize the parameters of the improved deep learning model through an adaptive moment estimation optimization algorithm, and the performance of the improved deep learning model is evaluated using evaluation metrics, including: The optimal parameters of the Long Short-Term Memory network are found by using an adaptive moment estimation optimization algorithm, and the optimal parameters are set at the maximum number of iterations, the learning rate decay period, the initial learning rate, and the learning rate decay factor during the training process. The performance of the improved deep learning model was evaluated using root mean square error and mean absolute percentage error as evaluation metrics. The smaller the value, the higher the prediction accuracy of the improved deep learning model.

15. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images