Industrial load prediction method and system

Abstract

The invention provides an industrial load prediction method and system, and the method specifically comprises the steps: obtaining original multi-source industrial data, carrying out the preprocessing of the original multi-source industrial data, obtaining a preprocessed industrial data feature set, carrying out the data enhancement of the preprocessed industrial data feature set through building a denoising diffusion model, and obtaining an enhanced sequence; data features are extracted from the preprocessed industrial data set and then are processed, the processed data features are subjected to weight sorting, and input features are screened out; and constructing a time sequence Transform model, and inputting the enhanced sequence and the input features into the model to obtain a load prediction result. Compared with the prior art, data enhancement is performed by introducing the diffusion model, the ability of the model to extract the associated features of the complex working conditions is enhanced while the training sample data is expanded, and the precision of the industrial load prediction result is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of load assessment, and in particular to an industrial load forecasting method and system. Background Art

[0002] Current load forecasting methods are primarily divided into traditional methods based on statistical models and modern methods based on artificial intelligence. Traditional methods such as exponential smoothing, multivariate linear regression, and ARIMA models have simple structures but struggle to effectively handle nonlinear fluctuations in load data, resulting in limited forecasting accuracy. To overcome the limitations of traditional methods, artificial intelligence algorithms are widely used in load forecasting. While methods such as support vector machines (SVM) and LightGBM have improved forecasting speed and accuracy, they remain deficient in handling large-scale fluctuations and capturing complex load patterns. Artificial neural networks (ANNs) have performed well in load forecasting, but they still face difficulties in parameter adjustment and reduced accuracy when dealing with long-term data.

[0003] With the development of deep learning technology, attention mechanisms have been introduced into neural networks, giving rise to the Transformer model. Through its self-attention mechanism, the Transformer model can capture long-range dependencies in sequences, providing new insights for load forecasting. However, a challenge with neural network models for time series forecasting is that model complexity makes their performance dependent on the correlation between input variables and load. Compared to conventional load forecasting, industrial loads are more susceptible to complex patterns such as instantaneous spikes caused by equipment startups and shutdowns and multi-frequency oscillations caused by overlapping processes. Therefore, the variability of power loads often exhibits nonlinear characteristics. Existing models often suffer from bottlenecks such as insufficient feature decoupling and limited ability to capture multi-period dependencies. They require the ability to capture and enhance existing data features. Current Transformer models for load forecasting suffer from limited accuracy and general applicability.

[0004] Therefore, providing an industrial load forecasting method and system for solving the above problems is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of the present invention is to provide an industrial load forecasting method with clear logic, high efficiency and simple operation. By introducing a diffusion model for data enhancement, the model's ability to extract complex working condition-related features is enhanced while expanding the training sample data, thereby significantly improving the accuracy of industrial load forecasting results.

[0006] Based on the above objectives, the technical solutions provided by the present invention are as follows: An industrial load forecasting method comprises the following steps: Acquire original multi-source industrial data, and preprocess the original multi-source industrial data to obtain a preprocessed industrial data feature set; Constructing a denoising diffusion model to perform data enhancement on the preprocessed industrial data feature set to obtain an enhanced sequence; Extracting data features from the preprocessed industrial data set and processing the data, and weight-sorting the processed data features to screen out input features; A time series Transformer model is constructed, and the enhanced sequence and the input features are input into the time series Transformer model to obtain a load forecast result.

[0007] Preferably, the obtaining of original multi-source industrial data and preprocessing of the original multi-source industrial data to obtain a preprocessed industrial data feature set is specifically as follows: The electrical data characteristics, meteorological characteristics and economic activity characteristics are obtained through the collection tool, and the missing values of the electrical data characteristics, the meteorological characteristics and the economic activity characteristics are filled respectively, and the preprocessed electrical data characteristics, the preprocessed meteorological characteristics and the preprocessed economic activity characteristics are integrated into the preprocessed industrial data feature set.

[0008] Preferably, the step of constructing a denoising diffusion model to perform data enhancement on the preprocessed industrial dataset to obtain an enhanced sequence comprises the following steps: In a forward diffusion process, the denoising diffusion model gradually superimposes a variety of noises on the preprocessed industrial data set within a preset multiple time steps, so that the preprocessed industrial data set is converted into noisy data; Extracting features of the noisy data at each time step using the denoising diffusion model, retaining spatial information based on skip connections to obtain noise estimates corresponding to each time step; The noise estimation of each time step is started from random noise that conforms to the standard normal distribution by the denoising diffusion model in the reverse diffusion process, and denoising is continuously performed iteratively to obtain the enhanced sequence.

[0009] Preferably, the post-processing of extracting data features from the pre-processed industrial data set comprises the following steps: After extracting data features from the preprocessed industrial data set, the data are processed and filtered according to feature continuity and feature category to obtain multiple industrial data features; The plurality of industrial data features are encoded and data aggregated according to one-hot encoding and sliding windows to obtain continuous features and encoded features.

[0010] Preferably, the weight sorting of the processed data features to screen out the input features comprises the following steps: updating the preset weights of the continuous features and the coded features according to the RReliefF algorithm according to the preset parameters, sorting the updated weights, and after the sorting is completed, screening out the coded features or the continuous features with the highest weights; The preset parameters include: operating mode, temperature and holidays.

[0011] Preferably, the constructing of the temporal Transformer model is specifically as follows: Construct a Transformer model consisting of a feature input module, a feature fusion module, and a temporal encoder; The temporal encoder includes: a gated linear unit and a multi-head attention unit.

[0012] Preferably, the enhanced sequence and the input features are both input into the time series Transformer model to obtain a load forecast result, comprising the following steps: The feature input module inputs the acquired enhanced sequence and the encoding of the input feature into the feature fusion module; The feature fusion module performs feature fusion on the feature code of the enhanced sequence and the feature code of the input feature to obtain a fused feature code, and sends the fused feature code to the temporal encoder; The temporal encoder performs feature extraction on the fusion feature encoding according to the gated linear unit and the multi-head attention unit to obtain long-term and short-term dependency features, and obtains the load forecast result after normalization.

[0013] An industrial load forecasting system, comprising: an acquisition and processing module, configured to acquire original multi-source industrial data and pre-process the original multi-source industrial data to obtain a pre-processed industrial data feature set; A data enhancement module is used to build a denoising diffusion model to perform data enhancement on the preprocessed industrial data feature set to obtain an enhanced sequence; A feature screening module is used to extract data features from the preprocessed industrial data set and then process them, and weight-sort the processed data features to screen out input features; The load forecasting module is used to construct a time series Transformer model, and input the enhanced sequence and the input features into the time series Transformer model to obtain a load forecast result.

[0014] The industrial load forecasting method provided by the present invention specifically obtains original multi-source industrial data, preprocesses the original multi-source industrial data to obtain a preprocessed industrial data feature set, and enhances the preprocessed industrial data feature set by constructing a denoising diffusion model to obtain an enhanced sequence; extracts data features from the preprocessed industrial data set and then processes it, sorts the processed data features by weights, and filters out input features; constructs a time series Transformer model, inputs both the enhanced sequence and the input features into the model, and obtains a load forecasting result.

[0015] Compared with existing technologies, by introducing a diffusion model for data enhancement, the model's ability to extract related features of complex working conditions is enhanced while expanding the training sample data, significantly improving the accuracy of industrial load forecasting results.

[0016] The present invention also provides an industrial load forecasting system, which has the same technical concept as the method, solves the same technical problem, and should have the same beneficial effects, so it will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0018] Figure 1 A flowchart of an industrial load forecasting method provided by an embodiment of the present invention; Figure 2 A flowchart of step S2 provided in an embodiment of the present invention; Figure 3 A functional schematic diagram of a denoising diffusion model provided by an embodiment of the present invention; Figure 4 A flowchart of post-processing of extracting data features from a pre-processed industrial data set in step S3 provided in an embodiment of the present invention; Figure 5 A functional diagram of a temporal Transformer model provided by an embodiment of the present invention; Figure 6 A flow chart for obtaining load forecast results provided by an embodiment of the present invention; Figure 7 A diagram of the GLU structure provided by an embodiment of the present invention; Figure 8 A structural diagram of an industrial load forecasting system provided in an embodiment of the invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] The embodiments of the present invention are written in a progressive manner.

[0021] The embodiments of the present invention provide an industrial load forecasting method and system, which mainly solve the technical problems of the existing load forecasting Transformer model, namely, insufficient accuracy and poor universal applicability.

[0022] like Figure 1 As shown, an industrial load forecasting method includes the following steps: S1. Obtaining original multi-source industrial data and preprocessing the original multi-source industrial data to obtain a preprocessed industrial data feature set; S2. Construct a denoising diffusion model to perform data enhancement on the preprocessed industrial data feature set to obtain an enhanced sequence; S3. extracting data features from the preprocessed industrial dataset and processing them, and weighting and ranking the processed data features to select input features; S4. Build a time series Transformer model and input both the enhanced sequence and input features into the time series Transformer model to obtain load forecast results.

[0023] In step S1, industrial data such as electrical data, meteorological data, and economic activities are acquired through a collection tool, integrated to generate a corresponding industrial data set, and the corresponding data in the industrial data set is preprocessed to obtain a preprocessed industrial data feature set; In practical applications, data preprocessing includes: data cleaning; data transformation; feature selection; and feature construction. Data cleaning specifically involves: handling missing values: Missing values are common in datasets. Methods for handling missing values include deleting missing records, filling missing values (such as the mean, median, mode, and interpolation), and filling missing values based on predictive models. Outlier handling: Outliers are observations that deviate significantly from the majority of data points. Methods for handling outliers include deleting, scaling, or replacing them. The interquartile range (IQR) is often used for outlier detection. Data transformation specifically involves: data standardization: converting data to a distribution with a mean of 0 and a variance of 1, suitable for data of different dimensions. Data normalization: scaling data to a fixed range (such as [0, 1]), suitable for distance-sensitive algorithms such as KNN. Categorical data encoding: converting categorical data into numerical form. Common methods include one-hot encoding and label encoding. Feature selection specifically involves: filtering: selecting important features through statistical methods, such as the Pearson correlation coefficient, to select features with a high correlation with the target variable. Feature construction involves creating new features by transforming or combining existing features, such as extracting year, month, and day from a timestamp, or constructing word frequency features using text data.

[0024] In step S2, a denoising diffusion model is constructed. This denoising diffusion model is a generative AI method based on "first scrambling the data and then learning to restore it bit by bit". It can learn to reconstruct real data from pure noise, and then generate new samples that are "indistinguishable from the real thing". The preprocessed industrial dataset feature set is input into the denoising diffusion model for data enhancement to obtain an enhanced sequence. In step S3, after extracting data features from the pre-processed industrial data set, multiple data features are processed and weighted to sort the processed data features to select the best input features; It should be noted that after step S1 is completed, step S2 of obtaining the enhanced sequence and step S3 of screening out the best input features can be performed simultaneously.

[0025] In step S4, a time series Transformer model is constructed. The time series Transformer model is a deep learning model architecture based on the attention mechanism. It is good at processing sequence data and is widely used in tasks such as natural language processing, time series modeling, and power system behavior recognition. The enhanced sequence and input features are input into the model to calculate the industrial load forecast results.

[0026] Preferably, step S1 is specifically: The electrical data characteristics, meteorological characteristics and economic activity characteristics are obtained through collection tools, and the missing values of the electrical data characteristics, meteorological characteristics and economic activity characteristics are filled respectively. The preprocessed electrical data characteristics, preprocessed meteorological characteristics and preprocessed economic activity characteristics are integrated into the preprocessed industrial data feature set.

[0027] In actual application, the collected electrical data characteristics, meteorological characteristics and economic activity characteristics data are used to delete missing value records, fill missing values (such as mean, median, mode, interpolation method), and complete missing values based on the prediction model. After the missing values are completed, the integrated preprocessed electrical data characteristics, preprocessed meteorological characteristics and preprocessed economic activity characteristics are used as the preprocessed industrial data feature set.

[0028] like Figure 2 As shown, preferably, step S2 includes the following steps: A1. Using a denoising diffusion model, during the forward diffusion process, multiple noise levels are gradually added to the preprocessed industrial dataset over multiple preset time steps, transforming the preprocessed industrial dataset into noisy data. A2. Extract the features of the noisy data at each time step using a denoising diffusion model. Preserve the spatial information using skip connections to obtain a noise estimate for each time step. A3. In the reverse diffusion process, the noise estimation of each time step is started from random noise that conforms to the standard normal distribution through the denoising diffusion model, and denoising is continuously performed iteratively to obtain the enhanced sequence.

[0029] In steps A1 to A3, a denoising diffusion model specifically designed for industrial power consumption was constructed to achieve data enhancement. Specifically, a one-dimensional UNet architecture was employed to preserve time-dependent characteristics during synthetic data generation. Unlike the general network structure used in traditional diffusion methods, a residual learning mechanism was introduced to make the denoising process more consistent with historical load evolution patterns. This allows for consistent physical representation of key power load behaviors in synthetic samples, effectively capturing the multi-scale dependencies present in industrial processes. like Figure 3 As shown, in this embodiment, during the forward diffusion process, the original load data The exponentially scaled Gaussian noise is gradually superimposed over multiple time steps, so that the data evolves from an ordered time series to a nearly random noise distribution. , then, add noise to the data It is input into the 1D-UNET model, which extracts the features of the noise data at each time step through the convolution layer and retains the key spatial information with the help of jump connections. In order to ensure the quality of the generated sequence, in addition to using MSE, the model also uses and Constraints are optimized to obtain a noise estimate for the corresponding time step. Backward diffusion begins with random noise sampled from a standard normal distribution. Data is then iteratively denoised and gradually restored to a high-quality time series. Through a gradual disentanglement process, latent feature data is gradually extracted. The resulting new dataset more accurately reflects the inherent patterns of load data, helping to improve the accuracy and robustness of load forecasting models.

[0030] like Figure 4 As shown, preferably, in step S3, the post-processing of extracting data features from the pre-processed industrial data set includes the following steps: B1. After extracting data features from the preprocessed industrial dataset, process and filter them based on feature continuity and feature category to obtain multiple industrial data features; B2. Encode and aggregate multiple industrial data features based on one-hot encoding and sliding windows to obtain continuous features and encoded features.

[0031] In steps B1 and B2, after extracting data features from the preprocessed industrial data set, the data features are processed and screened according to feature continuity and feature category to obtain multiple industrial data features. The multiple industrial data features are encoded according to one-hot encoding, and the context information of the encoded features is obtained through a sliding window to screen out continuous features and encoded features.

[0032] Preferably, in step S3, the processed data features are weighted and sorted to select the input features, which includes the following steps: According to the RReliefF algorithm, the preset weights of continuous features and coded features are updated according to the preset parameters, and the updated weights are sorted. After the sorting is completed, the coded features or continuous features with the highest weights are selected; Among them, the preset parameters include: operating mode, temperature and holidays.

[0033] In actual application, the RReliefFs algorithm is a feature selection algorithm for regression problems. It determines the importance of features by evaluating the dependency between features and prediction targets. It is particularly suitable for high-dimensional, nonlinear, and redundant feature scenarios. In this embodiment, the RReliefFs algorithm is used to update and calculate the weights of continuous features and coded features according to the operating mode, temperature, and holidays, and the coded feature or continuous feature with the highest weight (i.e., the highest influence) is selected and output; The calculation process of weight update is as follows: for each time point, the algorithm first calculates the Euclidean distance in the feature space to find its k nearest neighbor samples to determine which samples are "similar" in the feature space. The formula is as follows: ;

[0034] in and The time points and The eigenvector of , A represents the feature, is the total number of features.

[0035] For each feature, the algorithm calculates the difference between the target sample and the neighboring samples. Suppose the algorithm randomly selects the value of an instance on feature A , the category value of its nearest neighbor on feature A is For continuous and categorical features, the difference is calculated as follows: ; ; It should be noted that the above difference calculation method evaluates the influence of a feature on the prediction target by calculating the "difference" between the sample and its nearest neighbor on feature A. The calculation method of the difference varies depending on the feature type. in, It is the difference calculation of continuous features (such as temperature, power, etc.). is the value of a randomly selected sample on feature A; It is the value of its nearest neighbor sample in the feature space on feature A. The denominator is used for normalization to avoid unfair comparison caused by different dimensions of different features. It is a difference calculation for category features (such as operating mode, holidays, etc.). If the category values of two samples on feature A are consistent, the difference is 0, indicating "no difference"; if they are inconsistent, the difference is 1, indicating "completely different".

[0036] Finally, the feature weights are updated by the following formula: ; in, is the cumulative difference of feature A, is the cumulative difference between feature A and the target variable, and m is the number of samples. The parameter formula is as follows: ; ; It should be noted that in In the calculation of , diff is a general variable, which represents the difference calculated value of feature A at the jth neighbor. or Depends on the type of A.

[0037] in, is the current index, is the distance scaling factor. The contribution of each feature to the target variable is calculated by calculating the difference between the random instance and each of its nearest neighbors and accumulating the differences for each feature.

[0038] Choose from The largest encoded feature or continuous feature is output.

[0039] Preferably, a time series Transformer model is constructed, specifically as follows: Construct a Transformer model consisting of a feature input module, a feature fusion module, and a temporal encoder; The temporal encoder includes: gated linear units and multi-head attention units.

[0040] In practical applications, a Transformer model was constructed that leverages the temporal characteristics of industrial power data to enhance its temporal context modeling capabilities. This framework introduces a temporal encoder to model inter-sequence associations, utilizes a gated linear unit (GLU) to focus on key time periods within the sequence, and incorporates a multi-head attention mechanism to extract multi-level features of long-term trends and short-term disturbances. This Transformer model effectively improves the ability to express and dynamically respond to industrial load behavior, thereby enabling highly accurate deduction of electricity consumption behavior.

[0041] like Figure 5 As shown, in this embodiment, the temporal Transformer model is provided with a feature input module, a feature fusion module and a temporal encoder; the temporal encoder includes a gated linear unit (structured as Figure 7 shown) and multi-head attention units.

[0042] like Figure 6 As shown, preferably, in step S4, both the enhanced sequence and the input features are input into the time series Transformer model to obtain the load forecast result, which includes the following steps: C1. The feature input module inputs the acquired enhanced sequence and the encoded input features into the feature fusion module; C2. The feature fusion module fuses the feature codes of the enhanced sequence and the input features to obtain a fused feature code, which is then sent to the temporal encoder. C3. The temporal encoder performs multiple feature extractions on the fused feature encoding based on the gated linear unit and the multi-head attention unit to obtain long-term and short-term dependency features, and obtains the load forecast result after normalization.

[0043] In steps C1 to C3, the feature input module obtains the enhanced sequence and input features, encodes the enhanced sequence and input features, and inputs them into the feature fusion module; the feature fusion module fuses the feature codes of the two to obtain the corresponding fused feature code, and sends the fused feature code to the temporal encoder; the temporal encoder performs multiple feature extractions on the fused feature code based on the gated linear unit and multi-head attention unit set therein to obtain the corresponding long-term and short-term dependency features, and obtains the load forecast result after normalization.

[0044] In actual application, the input matrix in the feature fusion module Integrated target load sequence and N covariate time series Covers electrical parameters and environmental factors: ; in and Represent the load vector and covariate matrix respectively. In order to maintain the time order and equipment state transition, a staged position encoding is injected into the embedding feature: ; It should be noted that Expressed as an embedding weight matrix, is the time stage encoding function, Expressed as a staged position encoding function of the stacking time.

[0045] The embedding sequence is initially processed through an LSTM encoder, which captures the local temporal dynamics: ; in, and Represent the hidden state and unit state respectively. Encoder output Short-term device operation modes are preserved while irrelevant features are attenuated through its inherent gating mechanism.

[0046] Next, the gated linear unit (GLU) regulates the information flow between attention layers to suppress the propagation of invalid fluctuations and cycles in the data, helping to identify valid start and stop information and avoid interference from invalid information. GLU transforms the encoder output in the following way: ; in, represents element-wise multiplication, is the sigmoid function, where represents a linear transformation. Element-wise multiplication of these outputs enables the gating branch to control the information flow of the main branch and dynamically filter irrelevant features, ensuring accurate extraction of valid information from both payload and synthetic data. The multi-head attention layer then computes global dependencies. For each head, the query, key, and value are the same as in the Transformer: ; Then, the softmax function is used to calculate the attention weight for each time step , a scaled dot product is applied to the query and key, which is to capture the complex relationships within the sequence and extract valuable covariate information.

[0047] Weighted output It is obtained by summing the product of the attention weights and the value vector, thus achieving accurate integration of historical data. The final attention output is obtained by combining the outputs of multiple heads through concatenation and linear projection to form a comprehensive representation: ; in, represents the scaling factor, used to stabilize the gradient, is the transposed matrix.

[0048] The attention output is further processed through a GLU-based decoder stack to align multi-scale patterns: ; This residual structure ensures gradient stability when integrating local and global temporal features. The final load forecast for: ; like Figure 8 As shown, an industrial load forecasting system includes: An acquisition and processing module is used to acquire original multi-source industrial data and pre-process the original multi-source industrial data to obtain a pre-processed industrial data feature set; The data enhancement module is used to build a denoising diffusion model to enhance the preprocessed industrial data feature set to obtain an enhanced sequence; The feature screening module is used to extract data features from the preprocessed industrial data set and then perform post-processing, sorting the processed data features by weight to screen out input features; The load forecasting module is used to build a time series Transformer model and input the enhanced sequence and input features into the time series Transformer model to obtain the load forecast results.

[0049] In actual application, an industrial load forecasting system is also disclosed, which is provided with an acquisition and processing module, a data enhancement module, a feature screening module and a load forecasting module.

[0050] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of modules is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0051] In addition, all functional modules in the embodiments of the present invention may be integrated into one processor, or each module may be a separate device, or two or more modules may be integrated into one device; the functional modules in the embodiments of the present invention may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0052] Those skilled in the art will understand that all or part of the steps of the above-mentioned method embodiment can be completed by program instructions and related hardware. The aforementioned program instructions can be stored in a computer-readable storage medium. When the program instructions are executed, the steps of the above-mentioned method embodiment are executed; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM), magnetic disks or optical disks, and other media that can store program codes.

[0053] It should be understood that the use of "system," "device," "unit," and / or "module" in this application is merely a method for distinguishing different components, elements, parts, portions, or assemblies at different levels. However, if other terms can achieve the same purpose, the terms may be replaced by other expressions.

[0054] As used in this application and the claims, unless the context clearly indicates an exception, the terms "a," "an," "an," and / or "the" are not intended to refer to the singular and may include the plural, unless the context clearly indicates otherwise. Generally speaking, the terms "comprises" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list. A method or apparatus may also include other steps or elements. The phrase "comprises a..." does not preclude the presence of additional identical elements in the process, method, product, or apparatus that includes the elements.

[0055] If a flow chart is used in this application, the flow chart is used to illustrate the operations performed by the system according to the embodiments of the application. It should be understood that the previous or subsequent operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. At the same time, other operations can be added to these processes, or one or more operations can be removed from these processes.

[0056] The above describes in detail the industrial load forecasting method and system provided by the present invention. The above description of the disclosed embodiments will enable those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein, but is intended to be embodied in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. An industrial load forecasting method, characterized in that: The steps include: Acquire original multi-source industrial data, and preprocess the original multi-source industrial data to obtain a preprocessed industrial data feature set; Constructing a denoising diffusion model to perform data enhancement on the preprocessed industrial data feature set to obtain an enhanced sequence; Extracting data features from the preprocessed industrial data set and processing the data, and weight-sorting the processed data features to screen out input features; A time series Transformer model is constructed, and the enhanced sequence and the input features are input into the time series Transformer model to obtain a load forecast result.

2. The industrial load forecasting method according to claim 1, wherein: The obtaining of original multi-source industrial data and preprocessing of the original multi-source industrial data to obtain a preprocessed industrial data feature set is specifically as follows: The electrical data characteristics, meteorological characteristics and economic activity characteristics are obtained through the collection tool, and the missing values of the electrical data characteristics, the meteorological characteristics and the economic activity characteristics are filled respectively, and the preprocessed electrical data characteristics, the preprocessed meteorological characteristics and the preprocessed economic activity characteristics are integrated into the preprocessed industrial data feature set.

3. The industrial load forecasting method according to claim 1, wherein: The step of constructing a denoising diffusion model to perform data enhancement on the preprocessed industrial dataset to obtain an enhanced sequence includes the following steps: In a forward diffusion process, the denoising diffusion model gradually superimposes a variety of noises on the preprocessed industrial data set within a preset multiple time steps, so that the preprocessed industrial data set is converted into noisy data; Extracting features of the noisy data at each time step using the denoising diffusion model, retaining spatial information based on skip connections to obtain noise estimates corresponding to each time step; The noise estimation of each time step is started from random noise that conforms to the standard normal distribution by the denoising diffusion model in the reverse diffusion process, and denoising is continuously performed iteratively to obtain the enhanced sequence.

4. The industrial load forecasting method according to claim 3, characterized in that: The post-processing of extracting data features from the pre-processed industrial data set comprises the following steps: After extracting data features from the preprocessed industrial data set, the data are processed and filtered according to feature continuity and feature category to obtain multiple industrial data features; The plurality of industrial data features are encoded and data aggregated according to one-hot encoding and sliding windows to obtain continuous features and encoded features.

5. The industrial load forecasting method according to claim 4, characterized in that: The weight sorting of the processed data features to filter out the input features includes the following steps: updating the preset weights of the continuous features and the coded features according to the RReliefF algorithm according to the preset parameters, sorting the updated weights, and after the sorting is completed, screening out the coded features or the continuous features with the highest weights; The preset parameters include: operating mode, temperature and holidays.

6. The industrial load forecasting method according to claim 5, characterized in that: The construction of the temporal Transformer model is specifically as follows: Construct a Transformer model consisting of a feature input module, a feature fusion module, and a temporal encoder; The temporal encoder includes: a gated linear unit and a multi-head attention unit.

7. The industrial load forecasting method according to claim 6, wherein: Inputting the enhanced sequence and the input features into the time series Transformer model to obtain a load forecast result includes the following steps: The feature input module inputs the acquired enhanced sequence and the encoding of the input feature into the feature fusion module; The feature fusion module performs feature fusion on the feature code of the enhanced sequence and the feature code of the input feature to obtain a fused feature code, and sends the fused feature code to the temporal encoder; The temporal encoder performs feature extraction on the fusion feature encoding according to the gated linear unit and the multi-head attention unit to obtain long-term and short-term dependency features, and obtains the load forecast result after normalization.

8. An industrial load forecasting system, characterized in that: include: an acquisition and processing module, configured to acquire original multi-source industrial data and pre-process the original multi-source industrial data to obtain a pre-processed industrial data feature set; A data enhancement module is used to build a denoising diffusion model to perform data enhancement on the preprocessed industrial data feature set to obtain an enhanced sequence; A feature screening module is used to extract data features from the preprocessed industrial data set and then process them, and weight-sort the processed data features to screen out input features; The load forecasting module is used to construct a time series Transformer model, and input the enhanced sequence and the input features into the time series Transformer model to obtain a load forecast result.

Images