A method and apparatus for load forecasting

Abstract

The application discloses a kind of distribution transformer load prediction method and device, based on self-encoder to distribution transformer historical load data carries out time series feature extraction, obtains load feature vector, carries out clustering to load feature vector, obtains several clustering groups, respectively to the clustering group constructs load time series sample library, by combining the time series convolution network based on attention mechanism and convolution neural network trains the target load prediction model of each clustering group, can make full use of the structural similarity between clustering internal distribution transformer load time series data, effectively capture distribution transformer load multidimension, nonlinear and long short-term dependence, significantly improve the expression ability and prediction accuracy to complex load mode, solve the existing distribution transformer load prediction method difficult to effectively capture the deep-seated characteristics of distribution transformer load, and ignore the time series similarity between different distribution transformers, leading to prediction model difficult to adapt to diversified load mode, the technical problem of low reliability.

Description

Technical Field

[0001] This invention relates to the field of distribution transformer load prediction technology, and in particular to a method and apparatus for distribution transformer load prediction. Background Technology

[0002] In the context of building a new power system, the importance of distribution transformer load forecasting is becoming increasingly prominent. With the large-scale integration of new energy sources, the operating environment of the distribution network is becoming more complex, exhibiting a trend of frequent load fluctuations and increased forecasting difficulty. Simultaneously, the rapid development of new loads such as electric vehicles and distributed photovoltaic power further exacerbates the uncertainty and dynamism of distribution loads. As a key link connecting the distribution system and end users, the load characteristics of distribution transformers directly affect the stability of power grid operation and the formulation of control strategies. Accurate distribution transformer load forecasting not only provides basic data for the planning and construction of the distribution network but also plays a supporting role in operation scheduling, load transfer, and flexible control, serving as a prerequisite for achieving coordinated interaction among power sources, grid, load, and energy storage.

[0003] Existing methods for forecasting distribution transformer loads are unable to effectively capture the deep-seated characteristics of distribution transformer loads and ignore the temporal similarities between different distribution transformers, resulting in forecasting models that are difficult to adapt to diverse load patterns and have low reliability. Summary of the Invention

[0004] This invention provides a method and apparatus for predicting distribution transformer load, which solves the technical problems of existing distribution transformer load prediction methods that are difficult to effectively capture the deep characteristics of distribution transformer load and ignore the temporal similarity between different distribution transformers, resulting in prediction models that are difficult to adapt to diverse load patterns and have low reliability.

[0005] In view of this, the first aspect of the present invention provides a method for predicting distribution transformer load, comprising: Obtain historical load data of distribution transformers; The time series of historical load data of distribution transformers is characterized by pre-trained autoencoders to obtain load feature vectors; The load feature vectors are clustered to obtain several cluster groups; Standardize the load feature vectors in each cluster group to construct a load time series sample library; Based on temporal convolutional networks and convolutional neural networks, a load prediction model for each cluster group is constructed. Based on the load time series sample library, the load prediction model for each cluster group is trained to obtain the target load prediction model for each cluster group; Based on the target load prediction model for each cluster group, load prediction is performed on the distribution transformer to obtain the load prediction results.

[0006] Optionally, the pre-trained autoencoders include LTSTM-ATT-AE and CNN-ATT-AE.

[0007] Optionally, it also includes: Reconstruct the load forecast results.

[0008] Optionally, the historical load data of the distribution transformer includes the historical active power, historical reactive power, historical apparent power, historical power factor, historical voltage amplitude, and historical current amplitude of the distribution transformer.

[0009] Optionally, the load feature vectors are clustered to obtain several cluster groups, including: The load feature vectors are clustered using the K-Means algorithm to obtain several cluster groups. A second aspect of the present invention provides a distribution transformer load prediction device, comprising: The data acquisition module is used to acquire historical load data of the distribution transformer; The feature extraction module is used to perform feature processing on the time series of historical load data of distribution transformers based on a pre-trained autoencoder to obtain load feature vectors; The clustering module is used to cluster the load feature vectors to obtain several cluster groups; The sample construction module is used to standardize the load feature vectors in each cluster group and build a load time series sample library; The prediction model building module is used to build a load prediction model for each cluster group based on temporal convolutional networks and convolutional neural networks; The training module is used to train the load prediction model for each cluster group based on the load time series sample library, so as to obtain the target load prediction model for each cluster group. The prediction module is used to predict the load of distribution transformers based on the target load prediction model of each cluster group, and obtain the load prediction results.

[0010] Optionally, the pre-trained autoencoders include LSTM-ATT-AE and CNN-ATT-AE.

[0011] Optionally, it also includes: The load prediction restoration module is used to restore the load prediction results.

[0012] Optionally, the historical load data of the distribution transformer includes the historical active power, historical reactive power, historical apparent power, historical power factor, historical voltage amplitude, and historical current amplitude of the distribution transformer.

[0013] Optionally, the clustering module is specifically used for: The load feature vectors are clustered using the K-Means algorithm to obtain several cluster groups.

[0014] As can be seen from the above technical solutions, the distribution transformer load prediction method provided by the present invention has the following advantages: The distribution transformer load prediction method provided by this invention extracts time-series features from historical distribution transformer load data using an autoencoder to obtain load feature vectors. These load feature vectors are then clustered to obtain several cluster groups. A load time-series sample library is constructed for each cluster group. By combining a time-series convolutional network based on an attention mechanism and a convolutional neural network to train the target load prediction model for each cluster group, this method can fully utilize the structural similarity between distribution transformer load time-series data within each cluster. It effectively captures the multi-dimensional, nonlinear, and long- and short-term dependencies of distribution transformer loads, significantly improving the expressive power and prediction accuracy for complex load patterns. This method solves the technical problems of existing distribution transformer load prediction methods, which struggle to effectively capture the deep-seated features of distribution transformer loads and ignore the time-series similarities between different distribution transformers, leading to prediction models that are difficult to adapt to diverse load patterns and have low reliability.

[0015] Meanwhile, the distribution transformer load prediction method provided by this invention extracts the potential characteristics of the historical load of distribution transformers through an autoencoder and groups the distribution transformers by combining a clustering method, thereby constructing a more targeted and clearer prediction model system, realizing high-precision and scalable load prediction for multiple types of distribution transformers, and effectively improving the intelligence level of distribution network operation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a distribution transformer load prediction method provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the LSTM-ATT-AE provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the CNN-ATT-AE structure provided in an embodiment of the present invention; Figure 4 The load prediction model structure based on TCN-CNN provided in this embodiment of the invention; Figure 5 This is a schematic diagram of the structure of a distribution transformer load prediction device provided in an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] For easier understanding, please refer to Figure 1 This invention provides an embodiment of a distribution transformer load prediction method, comprising: Step 101: Obtain historical load data of distribution transformers.

[0020] It should be noted that, in this embodiment of the invention, historical load data of the distribution transformer is acquired first. The historical load data of the distribution transformer includes historical active power, historical reactive power, historical apparent power, historical power factor, historical voltage amplitude, and historical current amplitude.

[0021] Step 102: Based on the pre-trained autoencoder, perform feature processing on the time series of historical load data of distribution transformers to obtain load feature vectors.

[0022] It should be noted that after training the autoencoder to convergence using historical load data of distribution transformers, the model parameters of the autoencoder are fixed, and only the encoder part of the autoencoder is retained as the feature extraction network. The new time series of historical load data of distribution transformers is then input into the feature extraction network to extract features and obtain the load feature vector.

[0023] In one embodiment, the pre-trained autoencoders include LSTM-ATT-AE (Long Short-Term Memory Network-Attention Mechanism-Autoencoder) and CNN-ATT-AE (Convolutional Neural Network-Attention Mechanism-Autoencoder). LSTM-ATT-AE combines the long-term dependency modeling capability of LSTM with the key-period focusing capability of the attention mechanism, making it suitable for mining load characteristics in industries with obvious periodicity and peak-and-trough patterns, such as manufacturing and commercial areas. The LSTM model handles the problem of inconsistent data lengths for different distribution transformer loads through sequence padding and masking mechanisms, then uses LSTM to extract long-term dependencies, and the attention module dynamically focuses on key time periods such as peak electricity consumption. Finally, the compressed feature vectors are used for industry classification, similarity analysis, or cross-industry generalization prediction. CNN-ATT-AE is suitable for modeling industries with high-frequency fluctuations and obvious abrupt changes. Its one-dimensional convolutional layers extract local features of the load sequence layer by layer, and the attention mechanism enhances the representation of key load fluctuation segments. Through low-dimensional encoding of the latent space, not only can various key components such as abrupt change patterns and periodic patterns be extracted, but their intrinsic relationship with external variables can also be revealed.

[0024] The structure of LSTM-ATT-AE is as follows Figure 2 As shown, it includes the original time series input layer, sequence padding and masking layer, LSTM encoder layer, single-head self-attention layer, bottleneck layer, LSTM decoder layer, and reconstructed output layer. Its working principle is: Raw time series input layer: Extracts the load sequence arranged in chronological order from the historical load monitoring data of distribution transformers. The load sequence is normalized, outlier processing is performed, and missing values ​​are imputed. Training samples are constructed using a sliding window method, so that each sample corresponds to a continuous distribution transformer load curve at one end, which serves as the input time series for the network.

[0025] Sequence padding and masking layer: After the original time series input layer is constructed, a uniform maximum sequence length is set. For the load sequence whose length is less than the maximum sequence length, zero padding is performed at the end to make the length of all samples consistent in the time dimension. At the same time, a binary mask matrix is ​​generated according to the padding position to mask these padding time steps in subsequent LSTM calculation and gradient backpropagation, thereby avoiding the padding value from interfering with the real load information while maintaining batch training.

[0026] LSTM encoder layer: The load sequence after padding and masking is input into one or more LSTM encoders. The LSTM encoder calculates the hidden state time by time. Under the action of the gating mechanism, the encoder explicitly models the short-term fluctuation and long-term periodic relationship of the distribution transformer load during the state transition process, forming a comprehensive representation of the time series patterns such as work and holidays, day and night peaks and valleys, and seasonal changes.

[0027] Single-head self-attention layer: After obtaining the hidden state sequence of the LSTM encoder, a single-head self-attention mechanism is introduced to score the hidden state at each time step and obtain the attention weight through softmax. Then, a weighted sum (i.e., the sum of the product of the hidden state sequence and the attention weight) is calculated based on the hidden state sequence and the attention weight as the context vector. This allows the model to focus on key moments when the distribution transformer load pattern has a significant impact (such as peak load, abrupt change points, and morning and evening peak periods) during feature extraction, while suppressing information redundancy or time periods with small contributions.

[0028] Bottleneck layer: The context vector output by the single-head self-attention layer is input to the fully connected transformation or linear projection layer and mapped to a latent feature vector of a preset dimension. This vector serves as the bottleneck representation of the autoencoder. Under the constraint of dimensionality compression, it focuses on characterizing the overall shape, peak and valley distribution and fluctuation level of the distribution transformer load curve within the time window, providing a compact and information-intensive feature description for subsequent reconstruction and prediction tasks.

[0029] The LSTM decoder layer: In the decoding stage, the latent feature vector is used as the initial hidden state or additional input of the decoder LSTM. The output sequence is generated step by step by the LSTM decoding network, which is roughly symmetrical with the encoder structure. If necessary, the attention mechanism is combined at the decoding end to enable the decoding process to adaptively utilize the bottleneck features at different time steps, so as to achieve fine reconstruction of the time structure of the original distribution transformer load curve.

[0030] Reconstructed Output Layer: The decoder's output at each time step is restored to the load scalar value through a linear mapping layer, forming a complete reconstructed load sequence. and with mean squared error loss function As a training objective, the encoder, attention module, bottleneck layer and decoder are optimized end-to-end to ensure that the latent feature vectors fully retain the key statistical features and temporal structure of the distribution transformer load sequence under reconstruction constraints.

[0031] After LSTM-ATT-AE is pre-trained and converged on large-scale historical load data of distribution transformers, its parameters are fixed, and only the encoder part (including LSTM, attention layer and bottleneck layer) is retained as the feature extraction network. New historical load data of distribution transformers is input into the encoder to obtain the corresponding latent feature vector (i.e. load feature vector).

[0032] The structure of CNN-ATT-AE is as follows: Figure 3 As shown, it includes the original time series input layer, sequence padding and masking layer, one-dimensional convolutional encoder layer, single-head self-attention layer, bottleneck layer, one-dimensional deconvolutional decoder layer, and reconstructed output layer. Its working principle is: Raw time series input layer: Extracts the load sequence arranged in chronological order from the historical load monitoring data of distribution transformers. The load sequence is normalized, outlier removal is performed, and missing value imputation is performed. Training samples are constructed through a sliding time window so that each sample corresponds to a continuous distribution transformer load curve (such as several hours or days in the past) as the input time series of CNN-ATT-AE.

[0033] Sequence padding and masking layer: After the sample construction is completed, a uniform maximum sequence length is set. For load sequences with a length less than the maximum sequence length, zero padding is performed at the end to align the lengths of all samples in the time dimension. At the same time, a binary mask matrix is ​​generated based on the padding position. These padding time steps are masked in subsequent convolution operations and loss calculations, thereby eliminating the influence of padding values ​​on the real load information of the distribution transformer while achieving batch training.

[0034] One-dimensional convolutional encoder layer: The load sequence after padding and masking is input into an encoder composed of multiple layers of one-dimensional convolutions. The convolution kernel slides on the time axis to perform convolution operations on the load changes within the local time window and superimpose nonlinear activation and downsampling operations. The number of feature maps is increased layer by layer and the length of the time dimension is reduced, mapping the original high-dimensional time series into a multi-scale hierarchical representation, which fully characterizes the short-term fluctuations, peak-valley structure and periodic patterns of the distribution transformer load.

[0035] Single-head self-attention layer: After obtaining the time series feature map at the end of the encoder, a single-head self-attention mechanism is introduced. The correlation score of the feature vector at each time step or position is calculated and the attention weight is obtained through softmax. The feature sequence is weighted and summed to form a context vector. This allows the network to focus on key moments and local morphology with higher distinguishability of distribution transformer load mode during the feature extraction stage, suppress noise segments, and thus improve the ability to distinguish bottleneck features in the later stage.

[0036] Bottleneck layer: The context vector output by the single-head self-attention layer is input into the fully connected transformation or linear projection layer to obtain a latent feature vector of a preset dimension. This vector serves as the bottleneck layer representation of the autoencoder. Under the premise of explicit compression of time and feature dimension, it centrally encodes key information such as the overall level, peak and valley distribution, and fluctuation intensity of the distribution transformer load curve within the current time window, providing a compact feature description for decoding reconstruction and subsequent load prediction.

[0037] One-dimensional deconvolutional decoder layer: In the decoding stage, the bottleneck feature vector is fed into the decoder, which consists of multiple layers of one-dimensional deconvolution. Through continuous upsampling and convolution operations, the time dimension length is increased layer by layer and the number of feature maps is reduced. The time resolution and local morphology are gradually restored according to the structure symmetrical with the encoder, so that the network generates an intermediate reconstructed sequence with the same length as the original distribution transformer load sequence.

[0038] Reconstructed Output Layer: The last layer of the decoder maps intermediate features into scalar payloads through one-dimensional convolution or linear mapping, outputting a complete reconstructed payload sequence. And the mean square error function is used. As a training objective, the convolutional encoder, attention module, bottleneck layer, and deconvolutional decoder are optimized end-to-end to ensure that the latent feature vectors fully retain the statistical characteristics and temporal structure of the distribution transformer load time series under reconstruction constraints.

[0039] After CNN-ATT-AE completes pre-training and converges on large-scale distribution transformer historical load data, its parameters are fixed, and only the encoder part (including a one-dimensional convolutional encoder layer, a single-head self-attention layer, and a bottleneck layer) is retained as a feature extraction network. For new distribution transformer historical load data, the time series is fed into the encoder to obtain the corresponding latent feature vector (i.e., load feature vector).

[0040] Step 103: Cluster the load feature vectors to obtain several cluster groups.

[0041] It should be noted that, based on a pre-set clustering algorithm, the load feature vectors are clustered, and the time series of historical load data of distribution transformers is divided into several clusters to identify the similarity between time series, thereby grouping the distribution transformers according to users' electricity consumption habits and assigning cluster labels. In one embodiment, the pre-set clustering algorithm is the K-Means algorithm.

[0042] Step 104: Standardize the load feature vectors in each cluster group to construct a load time series sample library.

[0043] It should be noted that the load feature vectors in each cluster are standardized to generate load time series samples of distribution transformers, thereby constructing a load time series sample library.

[0044] Step 105: Construct a load prediction model for each cluster group based on temporal convolutional networks and convolutional neural networks.

[0045] It should be noted that a Temporal Convolutional Network (TCN) combined with a Convolutional Neural Network (CNN) is used to construct the load forecasting model. Each cluster group corresponds to a shared load forecasting model, used to learn the future trend of the load time series for that category. To address the seasonal and cyclical differences between different industries, the original series are first decomposed using STL (Seasonal-Trend-Residual) to unify the time frequency and trend patterns of each industry. Subsequently, a TCN-CNN hybrid model is constructed for each category: the Temporal Convolutional Network (TCN) identifies medium- to long-term trend dependencies through dilated convolutions; the CNN branches enhance the ability to identify sudden fluctuation patterns; and the covariate inputs integrate multi-source information such as temperature, holidays, and industry policies to improve the model's robustness and interpretability to external disturbances. The structure of the TCN-CNN-based load forecasting model is as follows: Figure 4 As shown, it includes an input layer, a fully connected layer, a flattened layer, a TCN module, a CNN module, and an output layer. Its working principle is: Input layer: As the input to the entire load forecasting model, each sample corresponds to the load characteristics of a time window, which is used to predict the load at future times.

[0046] Fully connected layer: Performs a linear transformation on the input followed by non-linear activation, recombines the original features and maps them to a new feature space. On the one hand, it adjusts the feature dimensions to make them more suitable for subsequent structures; on the other hand, it mixes between channels to improve the expressive power of the features.

[0047] Flattening layer: The output of the fully connected layer enters the flattening layer. The flattening layer reorganizes the current features according to the time dimension and the channel dimension, and reorganizes a long vector into a two-dimensional structure of time step multiplied by the feature channel, so that the subsequent one-dimensional convolution kernel can slide along the time axis to perform convolution operation on the sequence, thereby explicitly treating the representation as a time series.

[0048] The TCN module, through structures such as one-dimensional convolution, causal convolution, and dilated convolution, expands the receptive field while maintaining the advantages of parallel convolutional computation, characterizing the dynamic changes of the load over a longer time range, and stabilizing the training process through residual connections, is the core part for capturing long-term dependencies.

[0049] CNN module: One-dimensional convolutional kernels slide along the time axis, performing nonlinear transformations on local windows at different time scales. The first few layers focus more on local fluctuations and peaks and valleys in a short period of time. Based on the large receptive field provided by the TCN module, information from a longer time period is integrated to extract distribution transformer load time-series features that contain both local details and global trends. Finally, it provides a high-quality deep representation for the output layer to generate future load predictions.

[0050] Output layer: Outputs predicted load time series.

[0051] Step 106: Based on the load time series sample library, train the load prediction model for each cluster group to obtain the target load prediction model for each cluster group.

[0052] It should be noted that the load prediction model for each cluster group is trained using sample data from the load time series sample library. After the load prediction model converges, the load prediction model parameters are fixed to obtain the target load prediction model for each cluster group.

[0053] Step 107: Based on the target load prediction model of each cluster group, perform load prediction on the distribution transformer to obtain the load prediction results.

[0054] It should be noted that, for each user of the distribution transformer, after collecting the user's load data, the target load prediction model corresponding to the cluster group to which the user belongs is used to obtain the predicted load of the user, thereby obtaining the predicted load of each user under the distribution transformer and obtaining the load prediction result of the distribution transformer.

[0055] The distribution transformer load prediction method provided by this invention extracts time-series features from historical distribution transformer load data using an autoencoder to obtain load feature vectors. These load feature vectors are then clustered to obtain several cluster groups. A load time-series sample library is constructed for each cluster group. By combining a time-series convolutional network based on an attention mechanism and a convolutional neural network to train the target load prediction model for each cluster group, this method can fully utilize the structural similarity between distribution transformer load time-series data within each cluster. It effectively captures the multi-dimensional, nonlinear, and long- and short-term dependencies of distribution transformer loads, significantly improving the expressive power and prediction accuracy for complex load patterns. This method solves the technical problems of existing distribution transformer load prediction methods, which struggle to effectively capture the deep-seated features of distribution transformer loads and ignore the time-series similarities between different distribution transformers, leading to prediction models that are difficult to adapt to diverse load patterns and have low reliability.

[0056] Meanwhile, the distribution transformer load prediction method provided by this invention extracts the potential characteristics of the historical load of distribution transformers through an autoencoder and groups the distribution transformers by combining a clustering method, thereby constructing a more targeted and clearer prediction model system, realizing high-precision and scalable load prediction for multiple types of distribution transformers, and effectively improving the intelligence level of distribution network operation.

[0057] In one embodiment, after step 107, the method further includes: Step 108: Restore the load forecast results.

[0058] It should be noted that the sample constructed in step 104 has been standardized. Therefore, the prediction result obtained in step 107 is also represented in a standardized form. In order to improve the observability of the load prediction result, the load prediction result obtained in step 107 can be restored so that the representation of the load prediction result is consistent with the representation of the actual collected load data.

[0059] For easier understanding, please refer to Figure 5 This invention provides an embodiment of a distribution transformer load prediction device, comprising: The data acquisition module is used to acquire historical load data of the distribution transformer; The feature extraction module is used to perform feature processing on the time series of historical load data of distribution transformers based on a pre-trained autoencoder to obtain load feature vectors; The clustering module is used to cluster the load feature vectors to obtain several cluster groups; The sample construction module is used to standardize the load feature vectors in each cluster group and build a load time series sample library; The prediction model building module is used to build a load prediction model for each cluster group based on temporal convolutional networks and convolutional neural networks; The training module is used to train the load prediction model for each cluster group based on the load time series sample library, so as to obtain the target load prediction model for each cluster group. The prediction module is used to predict the load of distribution transformers based on the target load prediction model of each cluster group, and obtain the load prediction results.

[0060] In one embodiment, the pre-trained autoencoders include LSTM-ATT-AE and CNN-ATT-AE.

[0061] In one embodiment, it also includes: The load prediction restoration module is used to restore the load prediction results.

[0062] In one embodiment, the historical load data of the distribution transformer includes the historical active power, historical reactive power, historical apparent power, historical power factor, historical voltage amplitude, and historical current amplitude of the distribution transformer.

[0063] In one embodiment, the clustering module is specifically used for: The load feature vectors are clustered using the K-Means algorithm to obtain several cluster groups.

[0064] The distribution transformer load prediction device provided in this invention is used to execute the distribution transformer load prediction method provided in this invention. Its principle and the technical effects achieved are the same as those of the distribution transformer load prediction method provided in this invention, and will not be repeated here.

[0065] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0066] The above-described 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 distribution transformer load, characterized in that, include: Obtain historical load data of distribution transformers; The time series of historical load data of distribution transformers is characterized by pre-trained autoencoders to obtain load feature vectors; The load feature vectors are clustered to obtain several cluster groups; Standardize the load feature vectors in each cluster group to construct a load time series sample library; Based on temporal convolutional networks and convolutional neural networks, a load prediction model for each cluster group is constructed. Based on the load time series sample library, the load prediction model for each cluster group is trained to obtain the target load prediction model for each cluster group; Based on the target load prediction model for each cluster group, load prediction is performed on the distribution transformer to obtain the load prediction results.

2. The distribution transformer load prediction method according to claim 1, characterized in that, Pre-trained autoencoders include LSTM-ATT-AE and CNN-ATT-AE.

3. The distribution transformer load prediction method according to claim 1, characterized in that, Also includes: Reconstruct the load forecast results.

4. The distribution transformer load prediction method according to claim 1, characterized in that, The historical load data of the distribution transformer includes the historical active power, historical reactive power, historical apparent power, historical power factor, historical voltage amplitude, and historical current amplitude of the distribution transformer.

5. The distribution transformer load prediction method according to claim 1, characterized in that, The load feature vectors are clustered to obtain several cluster groups, including: The load feature vectors are clustered using the K-Means algorithm to obtain several cluster groups.

6. A distribution transformer load prediction device, characterized in that, include: The data acquisition module is used to acquire historical load data of the distribution transformer; The feature extraction module is used to perform feature processing on the time series of historical load data of distribution transformers based on a pre-trained autoencoder to obtain load feature vectors; The clustering module is used to cluster the load feature vectors to obtain several cluster groups; The sample construction module is used to standardize the load feature vectors in each cluster group and build a load time series sample library; The prediction model building module is used to build a load prediction model for each cluster group based on temporal convolutional networks and convolutional neural networks; The training module is used to train the load prediction model for each cluster group based on the load time series sample library, so as to obtain the target load prediction model for each cluster group. The prediction module is used to predict the load of distribution transformers based on the target load prediction model of each cluster group, and obtain the load prediction results.

7. The distribution transformer load prediction device according to claim 6, characterized in that, Pre-trained autoencoders include LSTM-ATT-AE and CNN-ATT-AE.

8. The distribution transformer load prediction device according to claim 6, characterized in that, Also includes: The load prediction restoration module is used to restore the load prediction results.

9. The distribution transformer load prediction device according to claim 6, characterized in that, The historical load data of the distribution transformer includes the historical active power, historical reactive power, historical apparent power, historical power factor, historical voltage amplitude, and historical current amplitude of the distribution transformer.

10. The distribution transformer load prediction device according to claim 6, characterized in that, The clustering module is specifically used for: The load feature vectors are clustered using the K-Means algorithm to obtain several cluster groups.

Images