Short-term load forecasting method for thermal power unit based on feature decoupling domain generalization network

Abstract

The application discloses a thermal power unit short-term load prediction method based on a feature decoupling domain generalization network, belongs to the technical field of thermal power unit short-term load prediction, obtains operation data and generator active power load labels of multiple source domain thermal power units, generates a data set through pretreatment and feature screening, independently trains a domain private feature extractor for each source domain thermal power unit, freezes parameters after training convergence, and obtains a domain private feature extractor with fixed parameters; a domain generalization network model is constructed, the fixed domain private feature extractor is used as a decoupling reference, an adversarial training mechanism is combined, a joint loss function composed of a prediction loss, a domain classification loss and a decoupling loss is used, and an alternating optimization strategy is used to complete model optimization; target domain data is input into the optimal model, and a short-term load prediction result is output. The application can effectively solve the problems of insufficient cross-unit generalization ability and feature confusion of the existing thermal power unit short-term load prediction method.

Description

Technical Field

[0001] This invention belongs to the field of short-term load forecasting technology for thermal power units, specifically relating to a short-term load forecasting method for thermal power units based on a feature decoupling domain generalization network. Background Technology

[0002] Short-term load forecasting for thermal power units is a key technology for optimized power system dispatch and safe operation, and its accuracy directly affects the operating efficiency and economy of power plants. However, existing mainstream forecasting methods generally suffer from insufficient generalization ability across units. Traditional methods typically rely on historical data of specific units for modeling. When these models are directly applied to different units with varying equipment parameters and operating conditions, the forecasting accuracy decreases significantly due to domain shifts in data distribution.

[0003] Some existing methods based on domain adaptive or domain adversarial neural networks attempt to extract domain-invariant general features across units through adversarial training. However, they still face a core challenge in complex industrial scenarios such as thermal power plants: their models fail to effectively distinguish and decouple unit-specific features from true general features during the learning process, leading to feature confusion. The learned invariant features are still mixed with private information of specific domains, which limits the prediction accuracy and generalization performance of the model on new, unseen units (i.e., the target domain).

[0004] Therefore, there is an urgent need for a new method that can separate private features from general features at the mechanism level, thereby fundamentally improving the model's adaptability across units. Summary of the Invention

[0005] The purpose of this invention is to solve the problem of poor cross-unit adaptability of short-term load forecasting models for thermal power units, and to propose a short-term load forecasting method for thermal power units based on feature decoupling domain generalization networks.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a short-term load forecasting method for thermal power units based on a feature decoupling domain generalization network, comprising the following steps: The system acquires operating data and generator active power load labels from multiple source-domain thermal power units, performs preprocessing and feature filtering on the operating data, and generates a dataset. Based on the dataset and load labels, an independent domain-private feature extractor is trained for each source domain thermal power unit. After training converges, the network parameters are frozen to obtain domain-private feature extractors for each source domain thermal power unit with fixed parameters. The domain-private feature extractors for each source domain thermal power unit with fixed parameters are used to perform forward feature extraction with constant parameters in the domain generalization network model to output the domain-private features of each source domain thermal power unit. A domain generalization network model is constructed. An adversarial training mechanism is used to learn the initial mixed feature representation across power units. The domain-private features output by the domain-private feature extractors of each source domain thermal power unit with fixed parameters are used as the decoupling benchmark. Through decoupling loss constraints, the domain-private information of each source domain thermal power unit is extracted from the initial mixed feature representation. This purification of the domain-private information yields domain-invariant features, completing the training of the domain generalization network model and obtaining the trained model. Based on the joint loss function, the dataset, and load labels, an alternating optimization strategy is used to iteratively optimize the feature extraction, adversarial training, and feature decoupling processes to obtain the optimal domain generalization network model. The joint loss function consists of prediction loss, domain classification loss, and decoupling loss. The operating data of thermal power units in the target domain are input into the optimal domain generalization network model, and the short-term load forecast results of thermal power units in the target domain are output.

[0007] Furthermore, the domain-private feature extractor adopts an Informer network structure, which includes a multi-head sparse self-attention module and a one-dimensional convolutional downsampling module.

[0008] Furthermore, the domain generalization network model includes a domain-invariant feature extractor, a label predictor, a domain classifier, and a gradient inversion layer; The domain-invariant feature extractor uses the same Informer network structure as the domain-private feature extractor, with the gradient inversion layer placed between the domain-invariant feature extractor and the domain classifier.

[0009] Furthermore, the decoupling loss is obtained by calculating the L1 distance between the domain-invariant features output by the domain-invariant feature extractor and the domain-private features output by the source domain thermal power unit domain-private feature extractors with fixed parameters. The overall decoupling loss is the sum of the single-domain decoupling losses of all source domains.

[0010] Furthermore, the gradient inversion layer directly transmits features during forward propagation and multiplies the gradient by a negative coefficient during backward propagation.

[0011] Furthermore, the joint loss function is shown in the following equation:

[0012]

[0013]

[0014]

[0015]

[0016] in, For the joint loss function, For the prediction loss function, As the first weighted parameter, For domain classification loss, β The second weighting parameter, For decoupling loss, The loss is the single-domain decoupling loss, where N is the total number of samples in the preprocessed running data. For the i-th sample of running parameters in the preprocessing running data, To input the i-th sample of the running parameters in the preprocessed running data into the output of the predictor in the domain adversarial neural network framework, where n represents the total number of source domains, The total number of categories in the domain. It is a one-hot vector representing the true domain label of the sample. E[·] represents the probability predicted by the domain classifier that sample k belongs to the m-th domain; E[·] represents the mean operation of samples in the batch, and ||·||1 represents the L1 distance. This indicates that the domain-invariant features generated by the domain-invariant feature extractor are... This represents the k-th frozen domain-private feature output by the domain-private feature extractor.

[0017] Furthermore, preprocessing includes missing value imputation, outlier removal, and normalization; feature selection uses Pearson correlation coefficient to screen operating parameters whose correlation with generator active power load is higher than a threshold.

[0018] Secondly, the present invention provides a short-term load forecasting system for thermal power units, comprising: The dataset acquisition module is used to acquire the operating data of multiple source domain thermal power units and the active power load labels of generators, preprocess the operating data and perform feature filtering to generate a dataset. The domain-private feature extractor training module is used to train an independent domain-private feature extractor for each source domain thermal power unit based on the dataset and load labels. After the training converges, the network parameters are frozen to obtain domain-private feature extractors for each source domain thermal power unit with fixed parameters. The domain-private feature extractors for each source domain thermal power unit with fixed parameters perform forward feature extraction with constant parameters in the domain generalization network model to output the domain-private features of each source domain thermal power unit. The domain generalization network model optimization module is used to construct the domain generalization network model. It learns the initial mixed feature representation across power units through an adversarial training mechanism. Using the domain-private features output by the domain-private feature extractors of each source domain thermal power unit with fixed parameters as the decoupling benchmark, it extracts the domain-private information of each source domain thermal power unit from the initial mixed feature representation through decoupling loss constraints. This purifies the domain-private information to obtain domain-invariant features, completing the training of the domain generalization network model and obtaining the trained domain generalization network model. Based on the joint loss function, the dataset, and load labels, an alternating optimization strategy is used to iteratively optimize the decoupling and feature extraction processes, as well as the adversarial training and feature decoupling processes, to obtain the optimal domain generalization network model. The joint loss function consists of prediction loss, domain classification loss, and decoupling loss. The short-term load forecasting output module is used to input the operating data of thermal power units in the target domain into the optimal domain generalization network model and output the short-term load forecasting results of thermal power units in the target domain.

[0019] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a short-term load forecasting method for thermal power units based on a feature decoupling domain generalization network.

[0020] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a short-term load forecasting method for thermal power units based on a feature decoupling domain generalization network.

[0021] Compared with the prior art, the present invention has the following beneficial technical effects: This invention proposes a short-term load forecasting method for thermal power units based on a feature-decoupled domain generalization network. Through a two-stage training architecture, it first independently trains and freezes domain-specific feature extractors for each source domain thermal power unit, explicitly modeling the unique operating characteristics of each unit and providing a clear, fixed benchmark for feature decoupling. The domain generalization network model built upon this foundation, combined with an adversarial training mechanism and decoupling loss constraints, can extract domain-specific private information from mixed features, refining cross-unit common domain-invariant features. This overcomes the defect of mixed common and private features in traditional domain adversarial methods. Employing a joint loss function consisting of prediction loss, domain classification loss, and decoupling loss, along with an alternating optimization strategy, it balances the model's prediction accuracy with cross-domain transferability. The training process does not require target domain unit data; the trained model can be directly applied to new, unseen units, significantly reducing the cost of repetitive modeling across different units and significantly improving the accuracy and robustness of cross-unit load forecasting. This effectively solves the problems of insufficient cross-unit generalization ability and feature confusion in existing short-term load forecasting methods for thermal power units. Attached Figure Description

[0022] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of the invention in any way. Furthermore, the shapes and proportions of the components in the drawings are merely schematic to aid in understanding the invention and do not specifically limit the shapes and proportions of the components. In the drawings: Figure 1 This is a flowchart of the short-term load forecasting method for thermal power units based on feature decoupling domain generalization network of the present invention.

[0023] Figure 2 This is a simplified structural diagram of the short-term load forecasting system for thermal power units of the present invention.

[0024] Figure 3 This is an electronic device diagram of the short-term load forecasting method for thermal power units based on feature decoupling domain generalization network of the present invention.

[0025] Figure 4 This is a diagram of the Informer (a neural network architecture) network structure in an embodiment of the present invention.

[0026] Figure 5 This is a block diagram of the DANN network in an embodiment of the present invention.

[0027] Figure 6 This is an overall block diagram of the short-term load forecasting method for thermal power units based on feature decoupling domain generalization network in an embodiment of the present invention.

[0028] Figure 7(a) is a comparison of short-term load forecast results under different models of unit 1, 2 → unit 3 in the embodiment of the present invention.

[0029] Figure 7(b) is a comparison of short-term load forecast results under different models for units 1, 2 → 4 in the embodiments of the present invention.

[0030] Figure 8(a) is a comparison of the dimensionality reduction of the original feature distribution in the cross-unit (unit 1, 2 → unit 4) migration prediction in the embodiment of the present invention using t-SNE (t-distributed stochastic neighbor embedding).

[0031] Figure 8(b) is a comparison of t-SNE dimensionality reduction of the feature distribution after DANN alignment in the cross-unit (unit 1, 2 → unit 4) migration prediction in the embodiment of the present invention.

[0032] Figure 8(c) is a comparison of the t-SNE dimensionality reduction of the feature distribution after DDGN alignment in the cross-unit (unit 1, 2 → unit 4) migration prediction in the embodiment of the present invention. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0034] Feature Decoupling for Generalization Network (DDGN); Domain-Adversarial Neural Network (DANN); Mean Squared Error (MSE); Multi-Layer Perceptron (MLP); Gradient Reversal Layer (GRL). Example 1 See Figure 1 A method for short-term load forecasting of thermal power units based on feature decoupling domain generalization networks includes the following steps: The system acquires operating data and generator active power load labels from multiple source-domain thermal power units, performs preprocessing and feature filtering on the operating data, and generates a dataset. Based on the dataset and load labels, an independent domain-private feature extractor is trained for each source domain thermal power unit. After training convergence, the network parameters are frozen to obtain domain-private feature extractors for each source domain thermal power unit with fixed parameters. The domain-private feature extractors for each source domain thermal power unit with fixed parameters perform forward feature extraction with constant parameters in the domain generalization network model, and output the domain-private features of each source domain thermal power unit. During the training process of the domain generalization network model, the domain-private feature extractors for each source domain thermal power unit with fixed parameters are only used to output fixed domain-private features and do not participate in gradient updates. A domain generalization network model is constructed. An adversarial training mechanism is used to learn initial mixed feature representations across power units. The domain-private features output by the domain-private feature extractors of each source domain thermal power unit with fixed parameters are used as the decoupling benchmark. Decoupling loss constraints are applied to extract domain-private information from the initial mixed feature representations, purifying the domain-private information to obtain domain-invariant features. This completes the training of the domain generalization network model, resulting in the trained domain generalization network model. Based on the joint loss function, dataset, and load labels, an alternating optimization strategy is used to iteratively optimize the feature extraction, adversarial training, and feature decoupling processes to obtain the optimal domain generalization network model. The joint loss function consists of prediction loss, domain classification loss, and decoupling loss. The operating data of thermal power units in the target domain are input into the optimal domain generalization network model, and the short-term load forecast results of thermal power units in the target domain are output.

[0035] Specifically, the method for short-term load forecasting of thermal power units based on feature decoupling domain generalization networks, see [link to relevant documentation]. Figure 6 This includes the following steps: S1: This invention acquires full-process operation monitoring data of thermal power units through multi-unit synchronous acquisition. The data categories are mainly divided into three categories: time-series chemical condition characteristic data, equipment operating status data, and generator active power load label data. In step S1, the time resolution of the collected data is set to 15 minutes, and all data are real-time monitoring data of the unit. This mainly includes key monitoring signals such as generator active power output, total coal feed rate, main steam and reheat steam operating temperatures, flue gas oxygen content required for furnace combustion optimization, NOx concentration at the inlet of the SCR denitrification system for environmental emission monitoring, and real-time coal feed rate of each coal mill for regulating combustion efficiency.

[0036] S2: Perform data preprocessing on the original data and reduce the data dimensionality using the Pearson coefficient; Step S2 preprocesses and reduces the dimensionality of the raw data collected from the SIS (Supervisory Information System) to improve the robustness and generalization ability of the parameter prediction model. The specific steps of S2 are as follows: S2.1: Relying on the isnan (is Not a Number) detection function built into the MATLAB (Matrix Laboratory) numerical computing platform, the original dataset is traversed row by row and column by column to accurately identify missing samples such as null values ​​and invalid values ​​in the dataset.

[0037] S2.2: Use the 3σ (standard deviation) criterion to screen outlier data. Remove abrupt outliers and steady-state noise data during unit operation to ensure the overall stability and accuracy of the dataset distribution.

[0038] S2.3: Perform normalization transformation on the cleaned standard dataset to map all feature parameters to the same numerical range, adapting to the training requirements of the neural network model. The normalization formula is as follows:

[0039] in: This represents the original monitoring data corresponding to operating parameter A for the a-th sample group. 'a' represents the standardized data after normalization of the sample; 'b' is the feature dimension identifier, used to distinguish different types of unit operating parameters; 'min(A)' and 'max(A)' correspond to the minimum and maximum values ​​of all sample data for parameter A, respectively. This linear transformation method can uniformly map all sample data to the numerical range [0,1] to eliminate modeling bias caused by differences in units and magnitudes.

[0040] S2.4: Quantitatively analyze the linear correlation between each parameter and unit load changes. By selecting the core feature variables with the best correlation, dimensionality reduction of the dataset features is completed, and the optimal input feature set is selected for model training. The specific calculation formula is as follows:

[0041] Where: In the formula Pearson correlation coefficient between operating parameters and unit load parameters; , These are the monitored values ​​of the operating parameter and load parameter to be measured in the i-th sample, respectively; The values ​​are the mean values ​​of the measured parameter and the load parameter in all c groups of valid samples, respectively. Here, c represents the total number of valid samples. The closer the absolute value of the correlation coefficient is to 1, the more significant the linkage between the operating parameter and the unit load condition, and the higher the characteristic reference value.

[0042] S3: First stage training: Train the domain-private feature extractor for each unit in the source domain separately to extract the unique operating features of a single unit. After training, freeze the parameters of the domain-private feature extractor as a reference for domain decoupling. Step S3 uses the domain-private feature extractor, an Informer structure, to extract the unique load variation patterns of individual units, specifically including: S3.1: Informer as a feature extractor. This module employs a one-dimensional convolutional kernel design and performs downsampling with a stride of 2 in the time dimension to achieve efficient processing of the input data. Let the original input data be... After convolution, the output features are obtained. It can be represented as:

[0043] in, Represents a one-dimensional convolution operation, stride The value is 2. It is the feature sequence output after the convolution operation. For the sample size, The length of the original input sequence. The feature dimension of the convolution output is given. This design addresses two limitations of the Transformer: the original sequence length L is reduced to... This effectively reduces the memory footprint of the feature matrix. Furthermore, it reduces the computational complexity from... Reduce to This greatly improves the computational efficiency of long sequences. It represents the set of real numbers.

[0044] The sparse attention mechanism in Informer aims to improve the computational efficiency of self-attention for long input sequences. Its core idea is to avoid indiscriminate computation on all key-value pairs by filtering them out. This mechanism is based on introducing the following sparsity metric. formula:

[0045] The formula evaluates each key-value pair in the key set K relative to the current query. The strength of the dependency. A higher sparsity metric M value indicates a higher number of key-value pairs. The more significant the contribution. Indicates the first query vectors set of keys Sparsity measure between them For the first A query vector, For key set, It is the total number of key vectors in the key set. The first key in the key set A key vector, Represents the query vector With key vector dot product, The feature dimensions for querying and key vectors.

[0046] Based on this, the most representative set of keys can be filtered out. Given the filtered sparse key set... The attention calculation process can be represented as:

[0047] This represents the result of sparse attention calculation. This is the filtered key vector matrix. This is the corresponding key value vector matrix. T denotes transpose. The dimension of the key vector is represented. In this process, the attention mechanism no longer relies on the complete key matrix, but only performs sparse aggregation operations on the key keys. This reduces computational complexity while maintaining the ability to model important dependencies. Combined with the aforementioned convolutional downsampling methods, these methods collectively enhance the scalability of the Informer model in processing long-term sequence data, enabling it to efficiently extract long-term dependencies and key local features from thermal power unit operation data. Its specific structure is as follows: Figure 4 As shown, the Informer network structure includes a multi-head sparse self-attention module and a one-dimensional convolutional downsampling module.

[0048] S3.2: The domain-private feature extractor is connected to the regression head at the end, consisting of a 16-dimensional fully connected layer and a 1-dimensional output layer, used to output the source domain load prediction value; S3.3: The domain-private feature extractor is trained using the MSE (Mean Squared Error Loss) loss function and the Adam (Adaptive Moment Estimation) optimizer. The training epochs are 1000, the learning rate is set to 0.001, and the batch size is 64. S3.4: Train independent domain-private feature extractors for the data from machine 1 and machine 2 respectively. After training, freeze their parameters as the benchmark for subsequent decoupling loss calculation.

[0049] S4: Second-stage training: Constructing an integrated domain generalization model with decoupled private features. This model includes a domain-invariant feature extractor, a label predictor, a domain classifier, and a gradient inversion layer. It learns cross-unit domain-invariant features through adversarial training. The network structure of the DANN model is as follows: Figure 5 As shown, its construction and training process includes: S4.1: Domain-invariant feature extractor, also known as domain hybrid extractor: It adopts the same Informer network structure as the domain-private feature extractor to learn general feature representations across units; S4.2: Label Predictor: Consistent with the regression head structure of the domain-private feature extractor, it outputs the load prediction value of the target domain unit based on the domain-invariant features of the source domain thermal power unit; S4.3: Domain Classifier: Composed of a 16-dimensional fully connected layer, a batch normalization layer, a ReLU (Rectified Linear Unit) activation function, a 1-dimensional output layer, and a Sigmoid (non-linear) activation function, used to determine the domain to which a feature belongs; S4.4: Gradient Reversal Layer (GRL): A gradient reversal layer is inserted between the feature extractor and the domain classifier. During forward propagation, features are directly passed, and during backward propagation, the gradient is multiplied by a negative coefficient. This enables adversarial training between the feature extractor and the domain classifier. Adjust dynamically as training progresses:

[0050] Where h = p / 1000, p represents the number of training iterations, and the value ranges from 0 to 1000.

[0051] S5: Define the joint loss function, including prediction loss. Domain classification loss Decoupling loss L dec Through weighted parameters and β To balance the optimization objectives of the three factors, an alternating optimization strategy is adopted during the training process.

[0052] S5.1: The prediction loss function is:

[0053] in, For the prediction loss function, For the i-th sample of running parameters in the preprocessing running data, To input the i-th sample of the running parameters in the preprocessed running data into the output of the predictor in the domain adversarial neural network framework, S is the total number of samples in the preprocessed running data; S5.2: Decoupling Loss Function: By calculating the L1 distance between domain-invariant features and domain-private features, inter-domain differences can be effectively quantified. Therefore, the decoupling loss of the k-th source domain is... It can be defined as:

[0054] In the formula, E[·] represents the mean operation of samples within the batch, and ||·||1 represents the L1 distance, which is used to characterize the absolute difference between the elements of the feature vector. This indicates that the domain-invariant features generated by the domain-invariant feature extractor are... This represents the k-th frozen domain-private feature output by the domain-private feature extractor. The overall multi-domain decoupling loss is the weighted sum of all single-domain decoupling losses:

[0055] Where n represents the total number of source domains. By maximizing Domain-invariant features are forcibly separated from the domain-private features of all source domains. This process strips domain-private information from the features, retaining only the domain-invariant features common to all source domains.

[0056] S5.3: In the domain generalization task, the domain classifier performs two key functions: identifying the domain origin of the data and guiding feature learning through adversarial training. This module processes high-dimensional features from the feature extractor, transforms them into a high-dimensional space using a multilayer perceptron (MLP), and then generates domain probabilities using the softmax (normalization exponent) function. During adversarial training, the gradient inversion layer (GRL) inverts the gradient of the domain classification loss (scaling factor 0.5) during backpropagation. This mechanism forces the feature extractor to produce domain-invariant features—while maintaining discriminative power, aligning the distribution between source domains, and ultimately improving the generalization ability to unseen target domains.

[0057] Its loss function for:

[0058] in, It is a one-hot vector representing the true domain label of the sample. It is the probability predicted by the domain classifier that sample k belongs to the m-th domain; This represents the total number of categories for the domain.

[0059] S5.4: Joint Loss Function: The total loss consists of regression loss, neighborhood loss, and decoupling loss.

[0060] The total loss function consists of three key parts: predicted loss, loss prediction, and loss prediction. Ensure the accuracy of the main prediction task; adversarial domain loss. The domain classifier and gradient inversion layer (GRL) promote feature distribution alignment between source domains; while decoupling loss Used to measure the degree of feature privatization. , and Minimize the total loss jointly L Update:

[0061] in, , and These are the learnable parameters for the domain hybrid feature extractor, predictor, and domain classifier, respectively. The learning rate represents the training rate of the model. It is the gradient operator. Figure 5 In represent ; represent Adversarial training is achieved through a gradient inversion layer (GRL): during backpropagation, the GRL inverts the gradient from the previous layer to the next. The gradient sign to the domain fusion feature extractor. This drives a minimax game to align the feature distribution, ultimately enabling the model to learn domain-invariant features with high generalization ability. After training, the domain-invariant features can be directly used to predict the target domain, without the target domain participating in any training stage.

[0062] S6: Input the target domain unit operation data into the trained model and output the short-term load forecast results of the target domain units; S7: Visualize and verify the domain feature alignment effect through t-SNE features, and evaluate the prediction performance using metrics such as RMSE (root mean square error), MAPE (mean absolute percentage error), and R² (coefficient of determination).

[0063] Experimental Analysis: This invention uses actual unit load data from a thermal power plant from July 10th to July 24th, 2024 as experimental data for verification. Units 1 and 2 have a rated capacity of 600 MW, and units 3 and 4 have a rated capacity of 660 MW. Data was sampled at 15-minute intervals, resulting in 96 equidistant sampling points per day. To evaluate the model's domain-invariant feature learning ability and cross-unit generalization performance, units 1 and 2 were selected as the source domain for each experiment, and the other two units as the target domain. The power generation load of thermal power units is affected by various factors, such as fuel supply, operating status, and grid dispatch requirements. To identify the most relevant variables, correlation analysis was performed on the data obtained from the monitoring information system. Features with high correlation to power generation load were selected as model inputs, as summarized in Table 1.

[0064] Table 1. Selected Influence Characteristics and Their Correlation Coefficients

[0065] To verify the effectiveness of the proposed method, the following methods were selected as comparison methods: (1) No Domain Generalization (No DG): The network structure contains only a feature extractor and a predictor. Experiments use a baseline model (Informer) trained on the source domain and tested on the target domain to verify the necessity of transfer learning for feature alignment in cross-domain prediction problems.

[0066] (2) Generative Adversarial Network (GAN): learns the feature distribution of source domain data through minimax game between feature generator and discriminator.

[0067] (3) Domain-Adversarial Neural Networks (DANN).

[0068] The specific prediction comparison results and quantitative indicators are shown in Figures 7(a), 7(b), and Table 2, respectively. In the figures, DANN stands for Domain-Adversarial Neural Network; GAN stands for Generative Adversarial Network, a generative domain alignment method used for comparison; NO DG stands for Informer model without domain generalization (i.e., using only the basic Informer without any domain generalization module; this comparison term is used to verify the necessity of the domain generalization module in cross-unit load forecasting tasks); and DDGN stands for the feature-decoupled domain generalization network proposed in this invention.

[0069] Table 2. Comparison of Quantitative Indicators of Predictive Performance Results of Different Models

[0070] In the two migration prediction scenarios shown in Figures 7(a) and 7(b), the prediction curve of DDGN shows the highest degree of agreement with the actual load curve, accurately tracking the actual fluctuation trend even during the peak and trough phases of rapid load changes. In contrast, models such as DANN exhibit systematic fluctuations at certain periodic nodes, especially showing overfitting to source domain thermal power units under stable operating conditions. The No-DG baseline model deviates significantly from the actual values ​​throughout the prediction range, even exhibiting directional errors at load inflection points. Its MAPE ranges from 7.52% to 8.11%, MAE from 26.27 to 36.45, and the lowest R² value is only 0.573, clearly reflecting the fundamental defects of traditional methods in dealing with distributional shifts. The quantitative data in Table 2 further validate the superiority of DDGN: its MAPE remains stable between 1.74% and 1.89%, its MAE between 6.29 and 7.25, and its R² remains above 0.97 throughout. Compared with various benchmark models, DDGN reduces MAPE by an average of 36.1% and MAE by an average of 43.6% compared with DANN, confirming the effectiveness of the explicit feature decoupling mechanism. At the same time, MAE is improved by 40.9%-46.3% compared with DANN and 40.8%-46.3% compared with GAN, and by 55.4%-58.9% compared with GAN, which fully demonstrates that the explicit feature decoupling mechanism it adopts can effectively extract more accurate domain-invariant features and can cope with the distribution offset problem between different units.

[0071] This invention uses t-SNE to visually compare the performance of DDGN and DANN in domain feature alignment. As shown in Figures 8(a), 8(b), and 8(c), where Dimension is the dimension, the feature distribution after DANN processing exhibits a certain clustering trend, but obvious domain-private features still remain in the shared representation. Feature points from different source domains are not completely fused, and a scattered distribution divided by domain can still be observed in some regions. This indicates that DANN only achieves coarse-grained domain alignment and fails to completely remove domain-private information. In contrast, DDGN, through an explicit feature decoupling mechanism and loss function optimization strategy, enables the feature distribution to exhibit a better global alignment effect. Feature points from different source domains are more tightly clustered in the domain-invariant feature space, and the domain-private features remaining in DANN are effectively stripped away. DDGN's explicit feature decoupling mechanism actively separates domain-private information from domain-invariant information, achieving accurate cross-domain feature alignment.

[0072] This invention constructs a generalization framework based on domain adversarial neural networks, employing an adversarial training mechanism to force the feature extractor to learn feature representations capable of confusing the domain classifier. This design aims to align the feature distributions among thermal power units in different source domains, thereby extracting general operating rules insensitive to domain changes, which is the theoretical basis for cross-unit prediction. During the training phase, this framework only requires data from the source domain thermal power units, without needing data from the target domain units, and can be directly applied to unseen target units during the prediction phase. This invention constructs a two-stage training architecture of decoupling followed by generalization, solving the feature confusion problem of traditional domain adaptive methods in complex industrial scenarios from the perspective of representation learning. Unlike domain adversarial networks that directly align features, this method trains independently for each source domain thermal power unit in the first stage and freezes its domain-specific feature extractor, thus explicitly modeling the unit-specific operating modes and equipment differences. In the second stage, the proposed integrated dual-private feature decoupled domain generalization model introduces an L1-norm-based decoupling loss function to enforce constraints on the distance between the learned domain-invariant feature representation and all frozen private features in the vector space. This achieves the goal of actively stripping domain-specific information from the feature level while preserving cross-domain shared general rules. Regarding the model structure and optimization mechanism, this invention designs a joint optimization objective consisting of prediction loss, adversarial domain classification loss, and decoupling loss, and employs an alternating optimization strategy for training, systematically coordinating the three objectives of model accuracy, domain invariance, and feature purity. The introduction of decoupling loss is a key mechanism, ensuring that the domain-invariant feature extractor not only deceives the domain classifier to achieve distribution alignment but also distances itself from each private feature to guarantee its generality. This multi-objective constraint optimization framework ensures that the feature representation ultimately extracted by the model has both discriminative power for downstream prediction tasks and good transferability to unseen target groups.

[0073] Example 2 See Figure 2 The short-term load forecasting system for thermal power units includes: The dataset acquisition module is used to acquire the operating data of multiple source domain thermal power units and the active power load labels of generators, preprocess the operating data and perform feature filtering to generate a dataset. The domain-private feature extractor training module is used to train an independent domain-private feature extractor for each source domain thermal power unit based on the dataset and load labels, and freeze the network parameters after training convergence to obtain domain-private feature extractors for each source domain thermal power unit with fixed parameters. The domain generalization network model optimization module is used to construct the domain generalization network model. It learns the domain-invariant features across units through an adversarial training mechanism. Based on the domain-private feature extractors of each source domain thermal power unit with fixed parameters, it decouples the domain-private features of each source domain thermal power unit from the domain-invariant features to obtain the trained domain generalization network model. Based on the joint loss function, dataset, and load labels, an alternating optimization strategy is used to iteratively optimize the decoupling and adversarial training mechanism to obtain the optimal domain generalization network model. The joint loss function consists of prediction loss, domain classification loss, and decoupling loss. The short-term load forecasting output module is used to input the operating data of thermal power units in the target domain into the optimal domain generalization network model and output the short-term load forecasting results of thermal power units in the target domain.

[0074] Example 3 See Figure 3 An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a short-term load forecasting method for thermal power units based on a feature decoupling domain generalization network.

[0075] Example 4 A computer-readable storage medium storing a computer program that, when executed by a processor, implements a short-term load forecasting method for thermal power units based on a feature decoupling domain generalization network.

[0076] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, read-only optical discs, optical storage, etc.) containing computer-usable program code.

[0077] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for short-term load forecasting of thermal power units based on feature decoupling domain generalization networks, characterized in that, Includes the following steps: The system acquires operating data and generator active power load labels from multiple source-domain thermal power units, preprocesses and filters the operating data to generate a dataset. Based on the dataset and the load labels, an independent domain-private feature extractor is trained for each of the source domain thermal power units. After the training converges, the network parameters are frozen to obtain domain-private feature extractors for each source domain thermal power unit with fixed parameters. The domain-private feature extractors for each source domain thermal power unit with fixed parameters perform forward feature extraction with constant parameters in the domain generalization network model to output the domain-private features of each source domain thermal power unit. A domain generalization network model is constructed. An initial mixed feature representation across power units is learned through an adversarial training mechanism. Using the domain-private features output by the domain-private feature extractors of each source domain thermal power unit with fixed parameters as the decoupling benchmark, domain-private information of each source domain thermal power unit is extracted from the initial mixed feature representation through decoupling loss constraints. This domain-private information is purified to obtain domain-invariant features, completing the training of the domain generalization network model and obtaining the trained domain generalization network model. Based on the joint loss function, the dataset, and the load labels, an alternating optimization strategy is used to iteratively optimize the feature extraction, adversarial training, and feature decoupling processes to obtain the optimal domain generalization network model. The joint loss function consists of prediction loss, domain classification loss, and decoupling loss. The operating data of the target domain thermal power units are input into the optimal domain generalization network model, and the short-term load forecast results of the target domain thermal power units are output.

2. The method for short-term load forecasting of thermal power units based on feature decoupling domain generalization network according to claim 1, characterized in that, The domain-private feature extractor adopts an Informer network structure, which includes a multi-head sparse self-attention module and a one-dimensional convolutional downsampling module.

3. The method for short-term load forecasting of thermal power units based on feature decoupling domain generalization network according to claim 2, characterized in that, The domain generalization network model includes a domain-invariant feature extractor, a label predictor, a domain classifier, and a gradient inversion layer. The domain-invariant feature extractor adopts the same Informer network structure as the domain-private feature extractor, and the gradient inversion layer is placed between the domain-invariant feature extractor and the domain classifier.

4. The method for short-term load forecasting of thermal power units based on feature decoupling domain generalization network according to claim 3, characterized in that, The decoupling loss is obtained by calculating the L1 distance between the domain-invariant features output by the domain-invariant feature extractor and the domain-private features output by the source domain thermal power unit domain-private feature extractors with fixed parameters. The overall decoupling loss is the sum of the decoupling losses of all source domains.

5. The method for short-term load forecasting of thermal power units based on feature decoupling domain generalization network according to claim 3, characterized in that, The gradient inversion layer directly transmits features during forward propagation and multiplies the gradient by a negative coefficient during backward propagation.

6. The method for short-term load forecasting of thermal power units based on feature decoupling domain generalization network according to claim 1, characterized in that, The joint loss function is shown in the following equation: in, For the joint loss function, For the prediction loss function, As the first weighted parameter, For domain classification loss, β The second weighting parameter, For decoupling loss, The loss is the single-domain decoupling loss, where N is the total number of samples in the preprocessed running data. For the i-th sample of running parameters in the preprocessing running data, To input the i-th sample of the running parameters in the preprocessed running data into the output of the predictor in the domain adversarial neural network framework, where n represents the total number of source domains, The total number of categories in the domain. It is a one-hot vector representing the true domain label of the sample. E[·] represents the probability predicted by the domain classifier that sample k belongs to the m-th domain; E[·] represents the mean operation of samples in the batch, and ||·||1 represents the L1 distance. This indicates that the domain-invariant features generated by the domain-invariant feature extractor are... This represents the kth frozen domain-private feature output by the domain-private feature extractor.

7. The method for short-term load forecasting of thermal power units based on feature decoupling domain generalization network according to claim 1, characterized in that, The preprocessing includes missing value imputation, outlier removal, and normalization; the feature selection uses Pearson correlation coefficient to select operating parameters whose correlation with generator active power load is higher than a threshold.

8. A short-term load forecasting system for thermal power units, characterized in that, include: The dataset acquisition module is used to acquire the operating data of multiple source domain thermal power units and the active power load labels of generators, preprocess the operating data and perform feature filtering to generate a dataset; The private feature extractor training module is used to train independent domain private feature extractors for each of the source domain thermal power units based on the dataset and the load labels, and freeze the network parameters after training convergence to obtain domain private feature extractors for each source domain thermal power unit with fixed parameters. The domain private feature extractors for each source domain thermal power unit with fixed parameters perform forward feature extraction with constant parameters in the domain generalization network model, outputting the domain private features of each source domain thermal power unit. The domain generalization network model optimization module is used to construct the domain generalization network model, learn the initial mixed feature representation across units through an adversarial training mechanism, and optimize the model with the specified parameters. The domain-private features output by the fixed source domain thermal power unit domain-private feature extractor serve as the decoupling benchmark. Domain-private information of each source domain thermal power unit is extracted from the initial mixed feature representation through decoupling loss constraints. This domain-private information is purified to obtain domain-invariant features, completing the training of the domain generalization network model and yielding the trained domain generalization network model. Based on the joint loss function, the dataset, and the load labels, an alternating optimization strategy is used to iteratively optimize the feature extraction, adversarial training, and feature decoupling processes to obtain the optimal domain generalization network model. The joint loss function consists of prediction loss, domain classification loss, and decoupling loss. The short-term load forecasting output module is used to input the operating data of the target domain thermal power units into the optimal domain generalization network model and output the short-term load forecasting results of the target domain thermal power units.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for short-term load forecasting of thermal power units based on feature decoupling domain generalization networks as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the short-term load forecasting method for thermal power units based on feature decoupling domain generalization networks as described in any one of claims 1-7.

Images