A thermal power unit short-term load prediction method, system, device and medium based on a Gaussian weighted domain generalization network

Abstract

The application discloses a kind of short-term load prediction method, system, equipment and medium of thermal power unit based on Gauss weighting area generalization network, belong to thermal power unit operation optimization and intelligent power generation technical field, including: the historical operation data of thermal power unit is collected and pretreated;The probability distribution of first operation data is fitted to obtain probability density function, and the first operation data is divided into Q time sequence sub-domain;Domain adversarial neural network framework is constructed and soft label vector is constructed;Calculate the joint loss function and determine the parameters of domain adversarial neural network framework;According to the real-time operation data of target unit, output the short-term load prediction result of target unit.The application solves the problem that the heterogeneity of load data distribution and the time sequence distribution offset cannot be captured by traditional load prediction method, resulting in the generalization performance of traditional prediction method suddenly drops when migrating across units or long-term running.

Description

Technical Field

[0001] This invention relates to the field of thermal power unit operation optimization and intelligent power generation technology, specifically to a method, system, equipment and medium for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization network. Background Technology

[0002] Short-term load forecasting for thermal power units is a crucial technical aspect of grid dispatching and unit optimization. However, actual operation of thermal power units presents two types of distribution offsets: heterogeneity in load data distribution among units and temporal distribution offset within a single unit. Specifically, different units exhibit significant heterogeneity in load data distribution due to differences in installed capacity or boiler type; a single unit's load data also shows temporal distribution offsets at different times due to operating condition switching and fuel fluctuations. Traditional load forecasting methods, which rely on the assumption of independent and identically distributed loads based on training and testing data, struggle to capture these offsets. This leads to a sharp decline in the generalization performance of traditional forecasting methods during cross-unit migration or long-term operation, failing to meet the flexible peak-shaving requirements of modern power systems.

[0003] If we can analyze these data using a data-driven approach and construct an adaptive distribution offset prediction method, we can significantly improve the prediction performance of load forecasting in complex operating conditions and break through the generalization bottleneck of traditional prediction methods.

[0004] Therefore, conducting research on short-term load forecasting of thermal power units based on domain generalization and adversarial learning, and solving the model generalization problem under distribution offset, has important engineering and theoretical value for flexible grid dispatching and efficient unit operation. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is that traditional load forecasting methods, which adopt the assumption of independent and identical distribution based on training and test data, are difficult to capture the heterogeneity of load data distribution and temporal distribution offset. This leads to a sharp drop in the generalization performance of traditional forecasting methods during cross-unit migration or long-term operation, and they cannot meet the needs of flexible peak shaving in new power systems.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network, comprising the following steps, The first operating data is divided into several time-domain subdomains by performing a time-domain partitioning operation. The first operating data is the preprocessed historical operating data of the target thermal power unit. A domain adversarial neural network framework is constructed, which is used to obtain soft label vectors of several temporal subdomains; The parameters of the domain adversarial neural network framework are determined based on a joint loss function, which is calculated based on the soft label vector and the output of the domain adversarial neural network framework. The short-term load forecast results of the target unit are output. These results are generated by inputting the real-time operating data of the target unit into a trained domain adversarial neural network framework.

[0008] As a preferred embodiment of the short-term load forecasting method for thermal power units based on Gaussian weighted domain generalization network described in this invention, the step of obtaining the preprocessed historical operating data of the target thermal power unit includes: Historical operating data related to load forecasting are collected from the operating database of thermal power units; Outliers in historical operational data are removed and normalized to obtain normalized operational data; The normalized operating data is preprocessed, the correlation coefficient between each operating data and load data in the normalized operating data is calculated, and the normalized operating data is filtered according to the correlation coefficient to obtain the first operating data.

[0009] As a preferred embodiment of the short-term load forecasting method for thermal power units based on Gaussian weighted domain generalization network described in this invention, the step of performing time-domain partitioning on the first operating data includes: The probabilistic graphical model uses a Gaussian mixture model, which is used to fit the probability distribution of each data point in the first running data to construct the probability density function. Calculate the posterior probability of a data point in the first running data belonging to the k-th Gaussian component based on the probability density function; Update the parameters of the probability density function based on the posterior probability of the k-th Gaussian component; Continuously calculate the posterior probability of the k-th Gaussian component and update the parameters of the probability density function until the parameter change is less than the parameter threshold, and determine the final parameters of the probability density function. The number of time series subdomains is set to K, the first running data is divided into K time series subdomains, and the Bayesian information criterion coefficients corresponding to the final parameters of the probability density function for each K value are calculated. The K value with the smallest Bayesian information criterion coefficient is taken as the number of time series subdomains Q. The first running data is divided into Q time-series subdomains in the time domain.

[0010] The beneficial effects of this preferred technical solution are as follows: by combining the Gaussian mixture model with the Bayesian information criterion, the changes in the data distribution characteristics of the unit at different operating stages are identified, and the continuous operating data is divided into a specified number of time-series subdomains with statistical significance, providing a high-quality domain partitioning basis for domain generalization, and being able to cope with the problem of time-series distribution changes in single unit operation.

[0011] As a preferred embodiment of the short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network described in this invention, the step of constructing a domain adversarial neural network framework includes: Construct a domain adversarial neural network framework that includes a feature extractor, a predictor, and a domain classifier; Extract the hidden layer feature representations of each temporal subdomain from the feature extractor output in the domain adversarial neural network framework; Calculate the correlation weights between two temporal subdomains based on the hidden layer feature representation; The dynamic smoothing coefficient is calculated based on the number of time-series subdomains, and a soft label vector is constructed based on the relevant weights and the dynamic smoothing coefficient.

[0012] The beneficial effects of this preferred technical solution are as follows: by setting a smoothing coefficient that decreases with the training process, the common learning between domains is emphasized in the early stage of training, while the extraction of intra-domain characteristics is emphasized in the later stage. This achieves the adjustment of the contribution of each domain to the model training, ensures the balance and stability of the learning process, and improves the prediction accuracy of the model on unknown target units.

[0013] As a preferred embodiment of the short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network described in this invention, the step of determining the parameters of the domain adversarial neural network framework includes: Calculate the prediction loss function based on the output of the predictor in the domain adversarial neural network framework; Calculate the domain classification loss function based on the soft label vector and the output of the domain classifier in the domain adversarial neural network framework; Calculate the joint loss function based on the prediction loss function and the domain classification loss function; The parameters of the feature extractor, predictor, and domain classifier in the domain adversarial neural network framework are updated based on the joint loss function.

[0014] The beneficial effects of this preferred technical solution are as follows: by constructing a domain adversarial neural network framework and training the adversarial neural network framework with Q temporal subdomains, the feature extractor learns the domain-invariant feature representation that can confuse the domain classifier, thus overcoming the problem of domain distribution shift caused by differences in equipment parameters and operating strategies among different thermal power units; by quantifying the similarity of different temporal subdomains and constructing a correlation weight matrix through the maximum mean difference between the feature distributions of different temporal subdomains, the model focuses on different temporal subdomains with similar distributions.

[0015] As a preferred embodiment of the short-term load forecasting method for thermal power units based on Gaussian weighted domain generalization network described in this invention, the formula for constructing the probability density function in the time-domain partitioning operation of the first operating data is: ; in, For probability density, This is the first running data for the k-th time-series subdomain. The set number of time-series subdomains, The number of samples in the first run. For mixed weights, Let be the mean of the k-th time series subdomain. Let k be the covariance matrix; The formula for calculating the posterior probability of a data point in the first set of running data belonging to the k-th Gaussian component based on the probability density function is as follows: ; in, Let be the posterior probability that the j-th data point belongs to the k-th Gaussian component, and i be the sample index.

[0016] As a preferred embodiment of the short-term load forecasting method for thermal power units based on Gaussian weighted domain generalization network described in this invention, the formula for calculating the correlation weight between two time-series subdomains based on the hidden layer feature representation is as follows: ; in, The correlation weights between the k-th and j-th time series subdomains are given. It is the function of maximum mean difference. and The hidden layer feature representations obtained after the k-th and j-th temporal subdomains are processed by the feature extractor. This represents the number of time-series subdomains. The formula for calculating the dynamic smoothing coefficient based on the number of time-series subdomains is as follows: ; in, For dynamic smoothing coefficients, The number of time-series subdomains. To preset the total number of training steps, This represents the current training iteration number; The formula for constructing a soft label vector based on relevant weights and a dynamic smoothing coefficient is as follows: ; in, To construct a soft tag vector based on the k-th temporal subdomain, assign a value to the position (i.e., the k-th position) of the soft tag vector corresponding to its own domain. As the positive class confidence level, the rest... Each position is assigned a value. ⋅ , as the weighted confidence level.

[0017] This invention provides a short-term load forecasting system for thermal power units based on a Gaussian weighted domain generalization network.

[0018] To address the aforementioned technical problems, the present invention further provides the following technical solution: a short-term load forecasting system for thermal power units based on a Gaussian weighted domain generalization network, comprising: Data acquisition module: Collects historical operating data of thermal power units and preprocesses it to obtain the first operating data; Data partitioning module: The probability distribution of the first running data is fitted using a probabilistic graphical model to obtain the probability density function. Based on the probability density function, the number Q of time series subdomains is determined and the first running data is divided into Q time series subdomains. Framework building module: Constructs a domain adversarial neural network framework and trains the adversarial neural network framework through Q temporal subdomains, and determines the feature extractor parameters, predictor parameters and domain classifier parameters in the domain adversarial neural network framework; Data output module: Inputs the real-time operating data of the target unit into the trained domain adversarial neural network framework and outputs the short-term load forecast results of the target unit.

[0019] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the described method for short-term load forecasting of thermal power units based on a Gaussian weighted domain generalization network.

[0020] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the aforementioned method for short-term load forecasting of thermal power units based on a Gaussian weighted domain generalization network.

[0021] The beneficial effects of this invention are as follows: This invention uses a Gaussian mixture model combined with Bayesian information criteria to identify changes in the data distribution characteristics of a power unit at different operating stages, and segments continuous operating data into a specified number of statistically significant time-series subdomains, providing a high-quality domain partitioning foundation for domain generalization and addressing the problem of time-series distribution changes in single-unit operation; by constructing a domain adversarial neural network framework and training the framework with Q time-series subdomains, the feature extractor learns domain-invariant feature representations that can confuse the domain classifier, overcoming the problem of domain distribution shift caused by differences in equipment parameters and operating strategies among different thermal power units; by quantifying the similarity of different time-series subdomains and constructing a correlation weight matrix through the maximum mean difference between the feature distributions of different time-series subdomains, the model focuses on different time-series subdomains with similar distributions; by setting a smoothing coefficient that decreases with the training process, the invention emphasizes the learning of commonalities between domains in the early stage of training and the extraction of intra-domain characteristics in the later stage, thereby adjusting the contribution of each domain to the model training, ensuring the balance and stability of the learning process, and improving the prediction accuracy of the model on unknown target units. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 The following is an overall flowchart of a short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network, provided as an embodiment of the present invention.

[0024] Figure 2 This is a general framework diagram of a short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network, provided as an embodiment of the present invention.

[0025] Figure 3 This invention provides an embodiment of a method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization networks, which includes a time-series subdomain MMD heatmap of a certain unit after time-series segmentation.

[0026] Figure 4 This is a domain neural adversarial network framework diagram of a short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network, provided as an embodiment of the present invention.

[0027] Figure 5 The diagram shows the loss change before and after introducing a soft label vector into the loss function of the domain classifier in a short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network, as provided in an embodiment of the present invention.

[0028] Figure 6 This is a comparison chart of the prediction results of the proposed method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization network, which is an embodiment of the present invention, with prediction results from other models.

[0029] Figure 7 The diagram shows the alignment effect of feature distribution before and after using a short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network, as provided in an embodiment of the present invention. Detailed Implementation

[0030] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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 protection scope of the present invention.

[0031] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network, including: S100: Perform time-domain partitioning on the first operating data to obtain several time-series subdomains. The first operating data is the preprocessed historical operating data of the target thermal power unit. S200: Construct a domain adversarial neural network framework, which is used to obtain soft label vectors of several temporal subdomains; S300: Determine the parameters of the domain adversarial neural network framework, wherein the parameters of the domain adversarial neural network framework are determined based on the joint loss function, wherein the joint loss function is calculated based on the soft label vector and the output of the domain adversarial neural network framework; S400: Output the short-term load forecast result of the target unit, which is generated by inputting the real-time operating data of the target unit into the trained domain adversarial neural network framework.

[0032] It should be noted that, in response to the heterogeneity of load data distribution and temporal distribution offset, traditional load forecasting methods adopt the assumption of independent and identically distributed load data based on training and test data. This makes it difficult to capture the heterogeneity of load data distribution and temporal distribution offset, resulting in a sharp drop in the generalization performance of traditional forecasting methods during cross-unit migration or long-term operation. This makes it impossible to meet the flexible peak-shaving requirements of new power systems. Therefore, it is very important to construct a forecasting method that can adapt to distribution offset.

[0033] Therefore, to address the aforementioned problems in short-term load forecasting for thermal power units, this invention constructs a short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network through steps S100~S400. This invention uses a Gaussian mixture model combined with Bayesian information criteria to identify changes in the data distribution characteristics of the unit at different operating stages, and segments continuous operating data into a specified number of statistically significant time-series subdomains, providing a high-quality domain partitioning foundation for domain generalization and addressing the problem of time-series distribution changes during single-unit operation. Furthermore, it constructs a domain adversarial neural network framework and trains the framework using Q time-series subdomains. This approach enables the feature extractor to learn domain-invariant feature representations that can obfuscate the domain classifier, overcoming the domain distribution shift problem caused by differences in equipment parameters and operating strategies among different thermal power units. By quantifying the similarity between different time-series subdomains through the maximum mean difference between feature distributions, a correlation weight matrix is ​​constructed, allowing the model to focus on different time-series subdomains with similar distributions. Furthermore, a smoothing coefficient that decreases with the training process is set, emphasizing the learning of commonalities between domains in the early stage of training and emphasizing the extraction of intra-domain characteristics in the later stage, thereby adjusting the contribution of each domain to the model training, ensuring the balance and stability of the learning process, and improving the prediction accuracy of the model on unknown target units.

[0034] Example 2, refer to Figures 1-4 This is the second embodiment of the present invention, which provides a method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization network.

[0035] In this embodiment of the application, obtaining the preprocessed historical operating data of the target thermal power unit in step S100 includes the following steps A1-A3: A1: Collect historical operating data related to load forecasting from the thermal power unit's operating database; In this embodiment, historical operating data is collected from the distributed control system (DCS) of the thermal power unit. The collected historical operating data includes, but is not limited to, the following operating parameters: generator active power, boiler efficiency, main steam temperature, reheat steam temperature, flue gas oxygen content, nitrogen oxide concentration at the inlet of the selective catalytic reduction system, real-time coal flow of each coal mill, and corresponding load data. The time span of the collection is usually 1-2 years of historical data during the normal operation of the unit, and the sampling frequency is once every 5 minutes to ensure that the data has sufficient time length and representativeness. A2: Remove outliers from historical running data and normalize the data to obtain normalized running data; In this embodiment, the isnan function of MATLAB is used to identify and remove null values ​​in the historical running data; then, the mean and standard deviation of each running parameter sequence are calculated using the 3σ criterion, and data points that exceed the mean ± 3 times the standard deviation are regarded as outliers and removed; after removing outliers in the historical running data, the minimum-maximum normalization formula is used for normalization processing. For each running parameter in the historical running data, the minimum value of the running parameter in the entire dataset is subtracted, and then divided by the range of the running parameter, thereby mapping the parameter to the interval [0,1] to obtain the normalized running data; A3: Preprocess the normalized operating data, calculate the correlation coefficient between each operating parameter and load data in the normalized operating data, and filter the normalized operating data according to the correlation coefficient to obtain the first operating data.

[0036] In this embodiment, the Pearson correlation coefficient is used to calculate the correlation coefficient between each operating parameter in the normalized operating data and the corresponding load data. The formula for calculating the correlation coefficient between each operating parameter in the normalized operating data and the corresponding load data is as follows: ; in, The correlation coefficient is... Let i be the i-th sample value of the running parameter in the normalized running data. For the i-th sample value of the load data corresponding to the operating parameters in the normalized operating data, This is the average value of the operating parameters in the normalized operating data. is the average value of the load data corresponding to the operating parameters in the normalized operating data, where n is the total number of samples and i is the sample index; Calculate the correlation coefficient between each operating parameter and the corresponding load data in the normalized operating data. Set the correlation coefficient threshold to 0.5. Compare the absolute value of the calculated correlation coefficient with the correlation coefficient threshold. Only the operating parameters and their corresponding load data with an absolute value of correlation coefficient greater than the correlation coefficient threshold are retained as the first operating data. The operating parameters with an absolute value of correlation coefficient less than or equal to the correlation coefficient threshold are considered to be operating parameters with weak correlation and are not retained. It should be noted that by calculating the correlation coefficient, the operating parameters with high correlation are selected to participate in subsequent calculations, while the operating parameters with weak correlation are eliminated. This reduces the feature dimensions of the operating parameters, which not only improves the efficiency of training and shortens the training time, but also reduces the resource requirements of subsequent calculations. This allows us to focus more on learning the key operating parameters that are strongly correlated with the load data, which helps to improve the final prediction accuracy and generalization ability.

[0037] In one alternative implementation, dimensionality reduction of normalized operational data can also be achieved through principal component analysis. This involves standardizing historical operational data containing multiple operational parameters, calculating the covariance matrix and corresponding eigenvalues ​​and eigenvectors, sorting the eigenvalues ​​from largest to smallest, and selecting the top k eigenvectors as principal components. Finally, the original data is projected onto a new coordinate system composed of these principal components, thus achieving dimensionality reduction. Cross-validation can be used to determine the optimal number of principal components, avoiding information loss or overfitting.

[0038] In another alternative implementation, dimensionality reduction of normalized running data can also be achieved through recursive feature elimination. A base learner is trained using all normalized running parameters, such as linear regression, support vector regression, or random forest. Based on the importance of the output features of the base learner, the features are sorted and the features with the lowest importance are removed, for example, one feature at a time. Then, the model is retrained using the remaining features, and this process is repeated until a preset number of features is reached, thereby achieving dimensionality reduction of the data.

[0039] In this embodiment of the application, the time-domain partitioning operation of the first running data in step S100 includes the following steps B1-B5: B1: The probabilistic graphical model uses a Gaussian mixture model, which is used to fit the probability distribution of each data point in the first running data to construct the probability density function; In this embodiment of the application, the formula for constructing the probability density function is: ; in, For probability density, The first running data for the k-th Gaussian component. The set number of time-series subdomains, each corresponding to a Gaussian component. The number of samples in the first run. The mixing weights for the k-th Gaussian component. Let be the mean of the k-th Gaussian component. Let be the covariance matrix of the k-th Gaussian component; It should be noted that the K Gaussian distribution components in the probability density function represent dividing the first running data into K temporal subdomains. This can identify and quantify multiple different temporal subdomains hidden in historical data in an unsupervised manner, providing an accurate temporal domain division basis based on data distribution characteristics for subsequent domain adversarial training. B2: Calculate the posterior probability of a data point in the first running data belonging to the k-th Gaussian component based on the probability density function; In this embodiment of the application, the formula for calculating the posterior probability of a data point in the first running data belonging to the k-th Gaussian component is as follows: ; in, Let be the posterior probability that the j-th data point belongs to the k-th Gaussian component, and i be the sample index; It should be noted that in the formula for calculating the posterior probability, the numerator represents the probability density of the k-th Gaussian component, and the denominator is the overall probability density. Dividing the two yields the posterior probability that the j-th data point belongs to the k-th Gaussian component. This allocation method quantifies the confidence level of each data point belonging to each Gaussian component, i.e., the temporal subdomain, rather than making a rigid division that is either one or the other, thus providing a data basis for subsequent parameter updates. B3: Update the parameters of the probability density function based on the posterior probability of the k-th Gaussian component; In this embodiment, the parameters of the probability density function include the mixing weights, mean, and covariance matrix. The formula for updating the mixing weights of the k-th Gaussian component is: ; The formula for updating the mean of the k-th Gaussian component is: ; The formula for updating the covariance matrix of the k-th Gaussian component is: ; It should be noted that, based on the current posterior probability, the mixture weights, mean, and covariance matrix of the probability density function are updated. The mixture weights are updated to the average of the posterior probabilities of all data points with respect to the k-th Gaussian component, reflecting the proportion of this component in the overall data. Using the posterior probability of the k-th Gaussian component as the weight, the mean is updated to the weighted average of all data points, repositioning the center of this component in the feature space. Using the posterior probability of the k-th Gaussian component as the weight, the covariance matrix is ​​updated to the weighted sample covariance around the updated mean, characterizing the dispersion shape and correlation of the data within this component. By iteratively updating the parameters of the probability density function, the mixture distribution of the real data is gradually approximated, enabling each Gaussian component to more accurately represent a statistically consistent time-series subdomain, providing a data foundation for the partitioning of time-series subdomains. B4: Continuously calculate the posterior probability of the kth Gaussian component and update the parameters of the probability density function until the parameter change is less than the parameter threshold, and determine the final parameters of the probability density function. In this embodiment of the application, step B2 is repeated to calculate the posterior probability and step B3 is repeated to update the parameters until the parameters of the probability density function converge. Specifically, the parameter threshold is set to 10. -5Calculate the change in the parameters of the probability density function between the current iteration and the previous iteration. If the change in the parameters of the probability density function is greater than the parameter threshold, continue to the next iteration. If the change in the parameters of the probability density function is less than or equal to the parameter threshold, stop the iteration and use the parameters of the probability density function after the last update as the final parameters. It should be noted that when judging the magnitude of the change in the parameters of the probability density function relative to the parameter threshold, it is necessary to ensure that the changes in the mixed weights, mean, and covariance matrix of the probability density function are all less than or equal to the parameter threshold before it can be considered that the change in the parameters of the probability density function is less than or equal to the parameter threshold, and the iteration can be stopped. B5: Sequentially set the number of time series subdomains to K, divide the first running data into K time series subdomains, and calculate the Bayesian information criterion coefficient corresponding to the final parameter of the probability density function for each K value. Take the K value with the smallest Bayesian information criterion coefficient as the number of time series subdomains Q, and divide the first running data into Q time series subdomains in the time domain.

[0040] In this embodiment, the number of time series subdomains is sequentially set to K, K=2, 3, ..., 10; the first running data is divided into K time series subdomains, and steps B1 to B4 are executed respectively. After determining the final parameters, the log-likelihood value of the probability density function under the final parameters is determined. Combining the sample size of the first running data and the total number of parameters of all probability density functions, the Bayesian information criterion coefficient is calculated; the formula for calculating the Bayesian information criterion coefficient is: ; in, These are the coefficients of the Bayesian information criterion. The log-likelihood value is... The total number of parameters for all probability density functions. The number of samples in the first run; Calculate the Bayesian information criterion coefficients corresponding to the number K of all time series subdomains respectively, and select the K with the smallest Bayesian information criterion coefficient as the final number Q of time series subdomains. Divide the first running data into Q time series subdomains in the time domain. In one alternative implementation, the number of time-series subdomains to be divided into the first running data can also be determined by evaluating the clustering quality of silhouette coefficient pairs. Multiple K values ​​are set sequentially for the first running data, and the K-means clustering algorithm is executed for each K value to calculate the silhouette coefficient of each sample. Then, the average silhouette coefficient of all samples is calculated, and the K value that maximizes the average silhouette coefficient is selected as the optimal number of time-series subdomains.

[0041] In another alternative implementation, the number of time-series subdomains to be divided into the first running data can also be determined by gap statistics. Multiple K values ​​are set sequentially, and the clustering error under different K values ​​is calculated for the first running data. A set number of reference datasets are generated, and the clustering error under different K values ​​is calculated for each reference dataset. The gap statistics are calculated, and the K value that maximizes the gap statistics is selected as the optimal number of time-series subdomains.

[0042] By using the Bayesian information criterion to automatically balance model fit and complexity, the number of time-series subdomains that best represent the inherent distribution structure of the data can be objectively and quantitatively identified, thereby avoiding overfitting or underfitting problems that may be caused by subjectively setting the number of time-series subdomains. The number of time-series subdomains of the operating data of each unit over the past fourteen days is shown in Table 2: Table 2 Number of Time-Series Subdomains for 14-Day Operation Data of Each Unit

[0043] To demonstrate the existence of distribution shifts in each temporal subdomain after temporal segmentation, the MMD heatmap measuring the distribution shift of each subdomain is shown below. Figure 3 As shown: It can be seen that after temporal subdomain segmentation, there is a certain degree of temporal distribution shift in each subdomain.

[0044] In this embodiment of the application, the construction of the domain adversarial neural network framework in step S200 includes the following steps C1-C4: C1: Construct a domain adversarial neural network framework that includes a feature extractor, a predictor, and a domain classifier; In this embodiment, a domain adversarial neural network framework comprising a feature extractor, a predictor, and a domain classifier is constructed. The network architecture diagram of the domain adversarial neural network framework is shown below. Figure 4 As shown; the feature extractor uses a three-layer fully connected network with an input dimension of the feature dimension of the first running data and an output hidden layer feature dimension of 128; the predictor is a single-layer fully connected network with an output dimension of 1, used for regression prediction; the domain classifier is a two-layer fully connected network with an output dimension of Q, used for domain classification tasks, and includes the feature extractor, predictor and domain classifier connected by a gradient inversion layer; C2: Extract the hidden layer feature representations of each temporal subdomain from the feature extractor output in the domain adversarial neural network framework; In this embodiment of the application, the first running data is divided into Q temporal subdomains according to the number determined in step S200, and the Q temporal subdomains are respectively input into the feature extractor in the domain adversarial neural network framework to extract the hidden layer feature representation corresponding to each temporal subdomain. C3: Calculate the correlation weights between two temporal subdomains based on the hidden layer feature representation; In this embodiment of the application, the formula for calculating the correlation weight between two temporal subdomains based on the hidden layer feature representation is as follows: ; in, The correlation weights between the k-th and j-th time series subdomains are given. It is the function of maximum mean difference. and The hidden layer feature representations obtained after the k-th and j-th temporal subdomains are processed by the feature extractor; It should be noted that in the formula for calculating the correlation weights between two time series subdomains, the numerator is the reciprocal of the maximum mean difference between the k-th and j-th time series subdomains, which reflects the similarity of the characteristic distributions between the two time series subdomains. The smaller the value of the maximum mean difference and the larger the reciprocal, the higher the similarity. The denominator is the sum of the reciprocals of the maximum mean differences between the k-th time series subdomain and all other m+1 time series subdomains, ensuring that the sum of all correlation weights is 1. C4: Calculate the dynamic smoothing coefficient based on the number of time-series subdomains, and construct a soft label vector based on the relevant weights and the dynamic smoothing coefficient.

[0045] In this embodiment, the dynamic smoothing coefficient decreases with the number of training steps. The formula for calculating the dynamic smoothing coefficient based on the number of training steps and the number of temporal subdomains is as follows: ; in, For dynamic smoothing coefficients, The number of time-series subdomains. To preset the total number of training steps, This represents the current training iteration number; The formula for constructing a soft label vector based on relevant weights and a dynamic smoothing coefficient is as follows: ; in, To construct a soft tag vector based on the k-th temporal subdomain, assign a value to the position (i.e., the k-th position) of the soft tag vector corresponding to its own domain. As the positive class confidence level, the rest... Each position is assigned a value. ⋅ As a weighted confidence level; It should be noted that the formula for constructing soft-label vectors consists of dynamic smoothing coefficients and relevant weights. By replacing traditional hard labels with dynamic soft-label vectors, the problems of gradient conflict and pattern collapse in domain adversarial training are alleviated, thereby making more balanced use of multi-source domain information and improving the stability of the final load prediction.

[0046] In an alternative implementation, the soft label vector can also be constructed by adversarial sample perturbation. For each training sample, the gradient of the loss function corresponding to the hard label of the training sample is first calculated, and adversarial perturbation is generated based on the gradient of the loss function to generate perturbed samples. Then, the predicted probability distribution is obtained based on the perturbed samples, and the hard label of the original sample and the predicted distribution of the adversarial sample are weighted and fused to obtain the soft label vector.

[0047] In another alternative implementation, the soft label vector can also be constructed using a teacher-student soft label model. A teacher model is trained, and for training samples, the teacher model generates a soft prediction distribution. The student model fits the true labels during training, and also needs to fit the soft labels of the teacher model. In the domain adversarial training scenario, the teacher prediction of the domain classifier can be used as the soft label. For the sample of the k-th temporal subdomain, the domain classification output of the teacher model is used as the soft label target for the current training, and the soft label vector is constructed through the soft label target.

[0048] In this embodiment of the application, determining the parameters of the domain adversarial neural network framework in step S300 includes the following steps D1-D4: D1: Calculate the prediction loss function based on the output of the predictor in the domain adversarial neural network framework; In this embodiment, mean squared error is used as the prediction loss function. The formula for calculating the prediction loss function based on the output of the predictor in the domain adversarial neural network framework is as follows: ; in, For predicting the loss function, For the i-th sample of the running parameters in the first running data, To input the i-th sample of the running parameters in the first running data into the output of the predictor in the domain adversarial neural network framework, N is the total number of samples in the first running data; It should be noted that the mean squared error, the standard loss function of the regression task, is used as the prediction loss function. The mathematical property of the mean squared error being continuously differentiable facilitates the updating of parameters by the gradient optimization algorithm. At the same time, the squared term imposes a higher penalty on larger prediction errors, thereby prioritizing the correction of significantly deviated prediction results and providing a stable and reliable prediction accuracy monitoring signal for the joint loss function. D2: Calculate the domain classification loss function based on the soft label vector and the output of the domain classifier in the domain adversarial neural network framework; In this embodiment of the application, the formula for calculating the domain classification loss function is as follows: ; in, For the domain classification loss function, The number of time-series subdomains. This represents the soft domain label of the j-th sample in the i-th time-series subdomain. Let k be the total number of samples in the k-th time-series subdomain, where k is a fixed index specifically referring to the subdomain currently being computed. The output of the feature extractor for the j-th sample in the i-th subdomain. This represents the predicted probability distribution of features by the domain classifier. The core of this loss function lies in comparing the distribution differences between the current sample's domain k and all other source domains, rather than calculating the loss of each domain in isolation. Cross-entropy loss measures the difference between the domain classifier's predicted probability and the true soft label. By averaging across all source domains and their samples, the feature extractor is guided to generate domain-invariant features, ultimately achieving the goal of cross-domain feature alignment.

[0049] D3: Calculate the joint loss function based on the prediction loss function and the domain classification loss function; In this embodiment of the application, the formula for calculating the joint loss function is as follows: in, For the joint loss function, For predicting the loss function, For the domain classification loss function, These are the parameters of the feature extractor. For the parameters of the predictor, For the parameters of the domain classifier, For loss trade-off coefficients; It should be noted that the loss tradeoff coefficient is used to adjust the relative weights of the prediction loss function and the domain classification loss function in the joint loss function. The value of the loss tradeoff coefficient needs to be adjusted experimentally to balance prediction performance and domain generalization ability. By combining the prediction loss function and the domain classification loss function as parameters for the feature extractor, predictor, and domain classifier, clear and mutually restraining optimization objectives are set. The predictor's objective is to minimize the prediction loss to improve the accuracy of load prediction; the domain classifier's objective is to minimize the domain classification loss to accurately identify the data source; and the feature extractor needs to extract features that are effective for load prediction to minimize the prediction loss, and also needs to generate hidden feature representations for the confusing domain classifier to cooperate in minimizing the domain classification loss. In an alternative implementation, the joint loss function can also be calculated by regularizing the loss based on domain-invariant feature constraints. A regularization term that measures the cross-domain difference of feature distributions, such as maximum mean difference or correlation alignment, is introduced into the loss function. The regularization term is multiplied by a regularization coefficient and then added to the domain classification loss to obtain the joint loss function.

[0050] In another alternative implementation, the joint loss function can also be calculated through task importance awareness. Task importance metrics are evaluated using the performance improvement rate, loss descent rate, or gradient norm ratio on the validation set. Dynamic weights are calculated based on the importance metrics, and the joint loss function is obtained by weighted summing of the predictor's loss and the domain classifier's loss.

[0051] D4: Update the parameters of the feature extractor, predictor, and domain classifier based on the joint loss function.

[0052] In this embodiment of the application, the optimization objective expression for the parameters of the feature extractor and the predictor is: ; in, The optimization objective for the parameters of the feature extractor is... The optimization objective for the predictor's parameters; For and The parameters that minimize the joint loss function are variables; The objective expression for optimizing the parameters of the domain classifier is: ; in, The optimization objective is to optimize the parameters of the domain classifier; For the sake of peace The parameters that allow the joint loss function to reach its maximum value are variables; The backpropagation algorithm is used to calculate the gradient based on the joint loss function, and the Adam optimizer is used to update the parameters of the feature extractor, predictor, and domain classifier; a parameter convergence threshold of 10 is set. -5 If the change in the joint loss function is less than the parameter convergence threshold over five consecutive training epochs, the training is considered converged. The parameters of the feature extractor, predictor, and domain classifier at this point are the final determined parameters. It should be noted that this adversarial training is achieved through a gradient reversal layer (GRL). Specifically, during backpropagation, the domain classification loss gradient passed to the feature extractor is multiplied by a negative coefficient −λ, thereby achieving gradient reversal. This causes the parameter update direction of the feature extractor to be opposite to the target of the domain classifier, forcing the feature extractor to generate hidden feature representations that the domain classifier cannot distinguish.

[0053] The feature extractor processes input data from different temporal subdomains to generate high-level feature representations and reduces inter-domain bias by optimizing the parameters of the feature extractor. The predictor makes predictions based on the features extracted by the feature extractor and is trained under the supervision of the temporal subdomains. Accurate predictions are achieved by optimizing the parameters of the predictor. The domain classifier is used to determine the temporal subdomain to which the features extracted by the feature extractor belong. By updating the parameters of the domain classifier, the feature extractor is driven to generate hidden feature representations.

[0054] The parameters learned after training convergence will be the saddle points of the joint loss function in the minimax optimization.

[0055] In this embodiment of the application, the output of the short-term load forecast result of the target unit in step S500 includes: Set Unit 3 of a power plant as the target unit. Input the operating parameters of the target unit at the current moment, such as total coal consumption, plant power consumption, coal consumption for power generation, heat consumption, main steam temperature, real-time coal consumption of the coal mill, weather temperature, wind speed, precipitation, and reheat steam temperature, into a pre-trained domain adversarial neural network framework. The feature extractor of the domain adversarial neural network framework extracts the features of the input data, and the predictor performs load prediction based on the extracted features, outputting the short-term load prediction value of the target unit for the next hour.

[0056] Example 3, referring to Figures 1-4 This is the third embodiment of the present invention, which provides a short-term load forecasting system for thermal power units based on a Gaussian weighted domain generalization network, comprising: Data acquisition module: Collects historical operating data of thermal power units and preprocesses it to obtain the first operating data; Data partitioning module: The probability distribution of the first running data is fitted using a probabilistic graphical model to obtain the probability density function. Based on the probability density function, the number Q of time series subdomains is determined and the first running data is divided into Q time series subdomains. Framework building module: Constructs a domain adversarial neural network framework and trains the adversarial neural network framework through Q temporal subdomains, and determines the feature extractor parameters, predictor parameters and domain classifier parameters in the domain adversarial neural network framework; Data output module: Inputs the real-time operating data of the target unit into the trained domain adversarial neural network framework and outputs the short-term load forecast results of the target unit.

[0057] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that: if the function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0058] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0059] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0060] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0061] Example 5, refer to Figures 5-7 This is the fifth embodiment of the present invention, which provides a method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization network. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0062] like Figure 5 As shown, the loss changes before and after introducing soft label vectors into the loss function of the domain classifier. With increasing training epochs, the loss value of the proposed short-term load forecasting method for thermal power units exhibits a more stable trend, with significantly reduced fluctuations compared to the short-term load forecasting method using normal domain labels. This phenomenon indicates that the weighted smoothing label strategy employed in the proposed short-term load forecasting method for thermal power units can effectively alleviate the oscillation problem commonly encountered in adversarial training, thereby making model parameter updates more stable and reliable.

[0063] like Figure 6As shown in the figure, the prediction results using the proposed method are compared with those of other models. The short-term load prediction method for thermal power units proposed in this application shows the highest degree of agreement between the prediction curve and the actual load curve, and can track the actual fluctuation trend. Especially during rapid load changes in the peak and trough phases, models like domain adversarial neural networks exhibit system fluctuations at certain cycle nodes, particularly showing overfitting to source domain units during the stable operation phase. The No DG (no-domain generalization) baseline shows a significant deviation from the actual value throughout the prediction range, even showing directional errors at load inflection points. Its MAPE (mean absolute percentage error) is between 7.52% and 10.21%, MAE (mean absolute error) is between 26.27 and 36.45, and the minimum R² value (coefficient of determination) is only 0.573. This reflects the fundamental defects of traditional methods when facing distributional changes. Table 2 shows a comparison of the prediction accuracy of the proposed method with other models, further validating the superiority of the proposed short-term load forecasting method for thermal power units: its MAPE consistently ranges from 1.74% to 2.09%, MAE consistently ranges from 6.29 to 8.84, and R² consistently remains above 0.97. Compared with various benchmark models, the proposed method for short-term load forecasting of thermal power units achieves an average MAPE reduction of 18.9% compared to Gaussian domain generalization networks using normal domain labels, confirming the effectiveness of the weighted smoothing label strategy. Compared with domain adversarial neural networks, generative adversarial networks, and multi-domain learning generative adversarial networks, the accuracy is improved by 29.1% to 43.1%, 44.7% to 50.3%, and 30.7% to 45.3%, respectively. This fully demonstrates that its dual mechanism of Gaussian mixture model temporal distribution segmentation and weighted adversarial training can effectively solve the problems of inter-unit variation and temporal distribution shift.

[0064] Table 2 Comparison of prediction accuracy metrics with other models

[0065] like Figure 7 As shown, the t-SNE dimensionality reduction algorithm is used to conduct an in-depth analysis of the feature distribution alignment effect of the proposed short-term load forecasting method for thermal power units. The feature distribution processed by the proposed method exhibits an ideal continuous manifold structure, with samples from different domains fully fused, domain boundaries largely eliminated, and a highly consistent feature distribution formed. This feature alignment effect confirms that the proposed method for short-term load forecasting of thermal power units, through a dual mechanism of temporal distribution segmentation and weighted adversarial training, extracts more accurate domain-invariant features and eliminates domain-specific features related to the unit and time period.

[0066] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization networks, characterized in that, include, The first operating data is divided into several time-domain subdomains by performing a time-domain partitioning operation. The first operating data is the preprocessed historical operating data of the target thermal power unit. A domain adversarial neural network framework is constructed, which is used to obtain soft label vectors of several temporal subdomains; The parameters of the domain adversarial neural network framework are determined based on a joint loss function, which is calculated based on the soft label vector and the output of the domain adversarial neural network framework. Output the short-term load forecast results of the target unit, which are generated by inputting the real-time operating data of the target unit into a trained domain adversarial neural network framework; The steps to construct a domain adversarial neural network framework include: Construct a domain adversarial neural network framework that includes a feature extractor, a predictor, and a domain classifier; Extract the hidden layer feature representations of each temporal subdomain from the feature extractor output in the domain adversarial neural network framework; Calculate the correlation weights between two temporal subdomains based on the hidden layer feature representation; Calculate the dynamic smoothing coefficient based on the number of time-series subdomains, and construct a soft label vector based on the relevant weights and the dynamic smoothing coefficient; The formula for calculating the correlation weight between two temporal subdomains based on the hidden layer feature representation is as follows: ； in, The correlation weights between the k-th and j-th time series subdomains are given. It is the function of maximum mean difference. and The hidden layer feature representations obtained after the k-th and j-th temporal subdomains are processed by the feature extractor. Number of time-series subdomains: The formula for calculating the dynamic smoothing coefficient based on the number of time-series subdomains is as follows: ； in, For dynamic smoothing coefficients, The number of time-series subdomains. To preset the total number of training steps, This represents the current training iteration number; The formula for constructing a soft label vector based on relevant weights and a dynamic smoothing coefficient is as follows: ； in, To construct a soft tag vector based on the k-th temporal subdomain, the position corresponding to its own domain, i.e., the k-th position, is assigned a value. As the positive class confidence score, the remaining... Each position is assigned a value. , As a weighted confidence level.

2. The method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization network as described in claim 1, characterized in that, The steps for performing time-domain partitioning on the first running data include: The probabilistic graphical model uses a Gaussian mixture model, which is used to fit the probability distribution of each data point in the first running data to construct the probability density function. Calculate the posterior probability of a data point in the first running data belonging to the k-th Gaussian component based on the probability density function; Update the parameters of the probability density function based on the posterior probability of the k-th Gaussian component; Continuously calculate the posterior probability of the k-th Gaussian component and update the parameters of the probability density function until the parameter change is less than the parameter threshold, and determine the final parameters of the probability density function. The number of time series subdomains is set to K, the first running data is divided into K time series subdomains, and the Bayesian information criterion coefficients corresponding to the final parameters of the probability density function for each K value are calculated. The K value with the smallest Bayesian information criterion coefficient is taken as the number of time series subdomains Q. The first running data is divided into Q time-series subdomains in the time domain.

3. The method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization network as described in claim 2, characterized in that, The steps for determining the parameters of a domain adversarial neural network framework include: Calculate the prediction loss function based on the output of the predictor in the domain adversarial neural network framework; Calculate the domain classification loss function based on the soft label vector and the output of the domain classifier in the domain adversarial neural network framework; Calculate the joint loss function based on the prediction loss function and the domain classification loss function; The parameters of the feature extractor, predictor, and domain classifier in the domain adversarial neural network framework are updated based on the joint loss function.

4. The method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization network as described in claim 3, characterized in that, The steps to obtain the preprocessed historical operating data of the target thermal power unit include: Historical operating data related to load forecasting are collected from the operating database of thermal power units; Outliers in historical operational data are removed and normalized to obtain normalized operational data; The normalized operating data is preprocessed, the correlation coefficient between each operating data and load data in the normalized operating data is calculated, and the normalized operating data is filtered according to the correlation coefficient to obtain the first operating data.

5. The method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization network as described in claim 4, characterized in that, The formula for constructing the probability density function during the time-domain partitioning operation of the first running data is as follows: ； in, For probability density, This is the first running data for the k-th time-series subdomain. The set number of time-series subdomains, The number of samples in the first run. For mixed weights, Let be the mean of the k-th time series subdomain. Let k be the covariance matrix; The formula for calculating the posterior probability of a data point in the first set of running data belonging to the k-th Gaussian component based on the probability density function is as follows: ； in, Let be the posterior probability that the j-th data point belongs to the k-th Gaussian component, and i be the sample index.

6. A short-term load forecasting system for thermal power units based on a Gaussian weighted domain generalization network, employing the short-term load forecasting method for thermal power units based on a Gaussian weighted domain generalization network as described in any one of claims 1 to 5, characterized in that, include: Data acquisition module: Collects historical operating data of thermal power units and preprocesses it to obtain the first operating data; Data partitioning module: The probability distribution of the first running data is fitted using a probabilistic graphical model to obtain the probability density function. Based on the probability density function, the number Q of time series subdomains is determined and the first running data is divided into Q time series subdomains. Framework building module: Constructs a domain adversarial neural network framework and trains the adversarial neural network framework through Q temporal subdomains, and determines the feature extractor parameters, predictor parameters and domain classifier parameters in the domain adversarial neural network framework; Data output module: Inputs the real-time operating data of the target unit into the trained domain adversarial neural network framework and outputs the short-term load forecast results of the target unit.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization network as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for short-term load forecasting of thermal power units based on Gaussian weighted domain generalization network as described in any one of claims 1 to 5.

Images