Industrial electrical load prediction method and system based on multi-scale feature fusion

Abstract

The invention relates to an industrial electrical load prediction method and system based on multi-scale feature fusion. The method comprises the following steps: collecting multi-source heterogeneous data to construct a comprehensive data set, splitting according to industry types, and further performing enterprise-level splitting according to an operation mode under each industry type to obtain a plurality of industry data subsets; designing a universal backbone training network; wherein the feature extraction network adopts a plurality of parallel branch network structures to extract features of different scales, and outputs universal feature vectors; designing a plurality of operation type sub-layers according to operation modes, and extracting mode features in different operation modes from the general feature vector; and the mode characteristics output by each operation type sub-layer output industrial load prediction of a future time step through a corresponding output layer. According to the invention, through multi-scale feature fusion and industry adaptation design, the precision of industrial load prediction and the industry applicability are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power load prediction, and particularly relates to an industrial power load prediction method and system based on multi-scale feature fusion. BACKGROUND

[0002] Industrial power load prediction is a process of estimating the power consumption of an industrial scene. Compared with residential or commercial loads, industrial power loads have the characteristics of strong fluctuation, obvious periodicity, various prediction factors, and many abnormal points.

[0003] The patent application with the publication number CN108846528A fuses 15-minute power data into daily power, processes the abnormal values, extracts seasonal coefficients based on time series analysis, selects historical features based on autocorrelation coefficients, constructs a model using gradient ascent decision trees, and predicts the monthly load result by rolling window day by day. The patent application with the publication number CN116780524A extracts strong correlation features based on time series data, builds an LSTM architecture after scaling the data set, and optimizes the parameters to improve the accuracy of short-term load prediction and realize adaptive feature learning, providing fine support for load scheduling and energy management. The patent application with the publication number CN114970939A trains a model using unsupervised learning, predicts similar sequences by sliding window, takes the mean of multiple models as the result, improves the generalization ability and accuracy of short-term prediction, and enhances the robustness.

[0004] However, the existing technology also has limitations: (1) Industrial load data contains multi-scale features such as second-level device start-stop transient, minute-level process switching period, hour-level production shift fluctuation, and day / week-level industrial policy influence. The existing technology is difficult to capture both high-frequency transient and low-frequency trend; (2) The industrial production process has strong time sequence, and the existing technology fails to fully model the time sequence dependence; (3) The load influencing factors of different industries are significantly different, and the existing technology lacks the ability to dynamically focus on key features according to the industry type. The above problems comprehensively affect the prediction accuracy. SUMMARY

[0005] To solve the problems in the prior art, the application provides an industrial power load prediction method and system based on multi-scale feature fusion, which collects multi-source data and performs time alignment and preprocessing, splits the data set according to the industry and operation mode, and performs standardization. A plurality of parallel branch network structures are used to extract features of different time scales, and the optimal convolution parameters are selected by an optimization algorithm to generate a general feature vector. According to the operation mode, a sub-layer is designed to extract mode features from the general features, and finally the industrial load prediction at the future time step is output.

[0006] The first aspect of the application discloses an industrial power load prediction method based on multi-scale feature fusion, which adopts the following technical scheme: Collecting power data of smart meters, operation parameters of equipment SCADA system and production work order data of MES system to build a comprehensive data set; time aligning and pre-processing the data with different sampling frequencies in the comprehensive data set; Splitting the pre-processed comprehensive data set by industry type, and further splitting at the enterprise level according to the operation mode under each industry type; finally, splitting the pre-processed comprehensive data set into multiple industry data subsets, and performing secondary standardization on each industry data subset; Designing a general backbone training network; for the feature extraction network in the general backbone training network, a plurality of parallel branch network structures are adopted, each branch network corresponds to an industry data subset and extracts features of different scales; wherein the convolutional parameters of each branch network are iteratively selected by an optimization algorithm to obtain the best parameter combination; the best parameter combination is applied to the general backbone training network to output a general feature vector; Designing multiple operation type sub-layers according to the operation mode, for extracting mode features under different operation modes from the general feature vector; the mode features output by each operation type sub-layer are output through the corresponding output layer to predict the industrial load at the future time step.

[0007] Further, the feature extraction network adopts a plurality of parallel branch network structures, including: Each branch network is composed of a convolutional network for feature extraction, an algorithm for extracting time sequence features, and an attention mechanism; Each branch network corresponds to a time scale, and the Bayesian search range of the convolutional parameters of each branch network is set at the time scale; based on the industry data subset, the best parameter combination of each branch network parameter is output through Bayesian iterative training.

[0008] Further, applying the best parameter combination to the general backbone training network includes: Applying the best parameter combination of each branch network to the convolutional layer of the general backbone network to perform feature extraction of multiple scales; the features output by the multiple branch networks are spliced and input into a bidirectional long short-term memory network to extract time sequence information, and then an attention mechanism is used to aggregate and output a general feature vector.

[0009] Further, splicing the features output by the multiple branch networks includes: The output of each branch network is , is the batch size, is the time step, is the feature dimension, is the branch index; Splicing the features extracted by all branch networks in the feature dimension The concatenated tensor has a shape of , is the feature dimension of the branch , is the number of branches.

[0010] Further, a plurality of job type sub-layers are designed according to the job modes, including: A plurality of job type sub-layers are constructed according to the job mode list, each job type sub-layer corresponding to a job mode; according to the data characteristics under different job modes, a feature enhancement mode is designed in each job type sub-layer; the general feature vector is output to each sub-layer, and each sub-layer outputs different mode features.

[0011] Further, the mode features output by each job type sub-layer pass through a corresponding output layer, including: An output head is designed for each job type sub-layer, and each output head outputs the mode features output by the corresponding job type sub-layer through an output layer including a full connection layer and an activation function, and outputs the future time step load prediction; wherein in the full connection layer, the mode features and the general feature vector are fused.

[0012] Further, the data of different sampling frequencies in the overall data set are time-aligned, including: The power data of the smart meter, the operating parameters of the device SCADA system, and the MES system production work order data of different sampling frequencies are set with a unified timestamp and the missing data is completed; A type of data is selected from the comprehensive data set as a reference sequence, and two types of data in the comprehensive data set except the reference sequence are used as the sequences to be aligned; The time alignment algorithm is called to align the sequences to be aligned with the reference sequence; the time scale of the time-aligned comprehensive data set is consistent with the reference sequence.

[0013] The second aspect of the application discloses an industrial load prediction system with multi-scale feature fusion, which executes the technical solution of the industrial power load prediction method as described in the first aspect of the application, and the system includes: A data acquisition and processing module is used to acquire smart meter power data, device SCADA system operating parameters, and MES system production work orders, construct a comprehensive data set, and perform time alignment and preprocessing on the data of different sampling frequencies in the overall data set; A data set splitting module is used to split the preprocessed comprehensive data set according to industry types, further split each industry type according to job modes, and finally split the preprocessed comprehensive data set into a plurality of industry data subsets, and perform secondary standardization on each industry data subset; Branch training module; for designing a general-purpose main trunk training network; for the feature extraction network in the general-purpose main trunk training network, a plurality of parallel branch network structures are adopted, each branch network corresponds to an industry data subset, and features of different scales are extracted; wherein the convolutional parameters of each branch network are iteratively selected by an optimization algorithm to obtain the best parameter combination; the best parameter combination is applied to the general-purpose main trunk training network to output a general-purpose feature vector; Prediction output module; for designing a plurality of job type sub-layers according to the job mode, and extracting mode features under different job modes from the general-purpose feature vector; the mode features output by each job type sub-layer are combined with the general-purpose feature vector, and the industrial load prediction at the future time step is output through the corresponding output layer.

[0014] The beneficial effects of the present application are that, compared with the prior art, 1. Simple structure, strong industry applicability; the present application collects multi-source data, aligns and preprocesses time, adopts a simple multi-branch network structure, and extracts features of different time scales. This method can adapt to the needs of multiple industries (such as steel, cement, and charging and battery replacement), automatically split the data set according to the characteristics of the industry, and ensure the wide applicability of the method in different industries. At the same time, the designed general-purpose main network and industry adaptation mechanism make the model have strong flexibility and scalability.

[0015] 2. Multi-scale feature fusion, improve prediction accuracy; the present application extracts short-term, medium-term and long-term features through parallel branch networks, and selects the best convolutional parameter combination using the Bayesian optimization algorithm, which can effectively capture the load change law at different time scales. In addition, the design of the job type sub-layer enhances the adaptability of the model to different job modes, so that the accuracy of load prediction in complex industrial scenarios is improved, especially in the case of handling different scales (periodic fluctuations and long-term trends). BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the industrial power load prediction method of the method embodiment; Figure 2 The multi-branch CNN feature extraction network structure diagram. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. The embodiments described in the present application are only a part of the embodiments of the present application, not all embodiments. Based on the spirit of the present application, all other embodiments obtained by those skilled in the art without creative labor shall fall within the protection scope of the present application.

[0018] As an embodiment of the present application, a specific embodiment of a multi-scale feature fusion industrial electricity load prediction method is provided. The flowchart of the method embodiment is shown in FIG. 1. Figure 1 .

[0019] S1: As an embodiment of the method embodiment, the collection and time alignment of multi-source heterogeneous data are performed. This step is mainly used for collecting and processing data from different systems for comprehensive analysis on a unified time scale.

[0020] 1.1: First, 15-minute level power data of smart meters, second level operation parameters (such as current, speed) of equipment SCADA, and minute level production work orders of MES system are collected to build a comprehensive data set containing more than 20 dimensions of electrical parameters, equipment status, production plan, etc. Then, the dynamic time warping algorithm is used to align the data of different sampling frequencies to a unified time scale, solving the asynchronous problem of second level and 15-minute level data. Finally, wavelet packet transform is used to eliminate high frequency noise such as power grid harmonics, and sliding median filter is used to repair abnormal data points such as short circuit current, improving the data quality.

[0021] The 15-minute level power data of smart meters, the second level operation parameters of equipment SCADA system, and the minute level production work orders of MES system (Manufacturing Execution System) are collected to build a comprehensive data set containing more than 20 dimensions of electrical parameters, equipment status, production plan, etc. Table 1 shows some examples in the comprehensive data set.

[0022] Table 1 Part of the collected data examples

[0023] 1.2: For the time asynchronous problem of second level (SCADA), minute level (MES), and 15-minute level (smart meter) data, the dynamic time warping (DTW) algorithm is used to realize the alignment of unified time scale, and the specific process is as follows: 1.2.1: Add a unified precision timestamp to all data, accurate to seconds.

[0024] For the 15-minute level data of smart meters, set the end time of each 15-minute period power data as the timestamp. For example, the power data timestamp of 8:00-8:15 is set to 8:15:00.

[0025] For the minute level data of MES system, set the end time of each minute work order data as the timestamp. For example, the work order data timestamp of 8:01 is set to 8:01:00.

[0026] As a specific optional implementation, the forward filling method is used to complete the control in the data. For example, the missing second-level data of the SCADA system due to communication interruption is filled with the latest valid data before the missing time.

[0027] 1.2.2: Define the reference sequence and the sequence to be aligned of the DTW algorithm; Select the 15-minute-level data of the smart meter as the reference sequence, wherein the sequence length is N, N = 96, representing the number of 15-minute periods in a day.

[0028] The sequence to be aligned is defined as: SCADA second-level data, length M, M = 86400, representing the number of seconds in a day; MES minute-level data, length P, P = 1440, representing the number of minutes in a day.

[0029] Call the DTW algorithm to align the SCADA second-level data and the MES minute-level data with the 15-minute-level data of the smart meter as the reference sequence, and output the 15-minute-level comprehensive data set in a unified time scale.

[0030] S2: As an embodiment of the method, the pre-processing of the time-aligned data output by S1 is performed to provide more standardized and high-quality data input for subsequent modeling and analysis.

[0031] Normalize the numerical data (such as planned production, actual production, unit energy consumption, etc.) to a unified dimension; for Boolean data, convert it to numerical data for easy calculation, such as mapping TRUE to 1 and FALSE to 0.

[0032] As an optional specific implementation, for text data, i.e. the sub-type text data in this application, such as product model, equipment number, production workshop, etc., the size of the text is unified, and the text is kept for classification and distinction, and does not participate in modeling calculation.

[0033] Perform pre-processing steps on the aligned data output by step 1, including but not limited to denoising, signal reconstruction and anomaly point repair.

[0034] Split the pre-processed comprehensive data set by industry type (such as steel, cement, charging and battery replacement, etc.), then under each industry category, according to the job mode list, further split at the enterprise level; finally, split the comprehensive data set into multiple industry data subsets. For example, for the steel industry data set, it can be split into all-day operation type enterprises, peak-avoiding operation type enterprises, and daytime operation type enterprises.

[0035] And independently perform secondary standardization on each industry data subset, so that the data of each industry adapts to its unique distribution characteristics.

[0036] S3: As an implementation of the method embodiment, refer to Figure 2 , a general backbone training network is designed. The network backbone can adapt to various types of data in various industries for training and learning. Each branch network is composed of a convolutional network for feature extraction, an algorithm for extracting time sequence features, and an attention mechanism; In further embodiments, a CNN-BiLSTM-Attention network is preferably used as the backbone training network to extract and fuse multi-scale features. Among them, CNN uses convolution to extract features of different time scales, BiLSTM is used to capture temporal dependencies, and the Attention mechanism is used to focus on important time steps.

[0037] Specifically, since the data output by S2 above is independent data subsets of different industries, these data exhibit different characteristics in feature extraction due to differences in production methods, work patterns, and load characteristics of each industry. For example, for production methods: the load change of the steel industry is usually affected by production rhythm, equipment start-stop, etc., and its load fluctuation has strong periodicity; while the load change of the charging and changing industry may be more related to battery charging and discharging cycles, charging station usage frequency, etc. Each industry's data subset can reflect the unique operation rules and behavior patterns of the industry itself.

[0038] For example, for job types: the load data of all-day operation enterprises will exhibit relatively stable operation characteristics, while the data of peak-avoiding operation enterprises may show significant load decline or fluctuation in certain periods.

[0039] 3.1: Considering the above differences, using a general convolutional network structure for feature extraction may not accurately capture the differences between industries, thereby affecting the prediction performance. Therefore, this application proposes a parallel convolutional network structure feature extraction method, which processes the feature convolutional layer in the backbone training network as multiple parallel network structures corresponding to different industry and different job type data subsets. This embodiment takes CNN as an example, and designs three parallel CNN network structures for the three industry data subsets output by S2.

[0040] In further embodiments, in order to fully capture the load change law at different time scales, the CNN part of the general backbone is designed as three parallel branch networks, including a short-term branch, a regular branch and a long-term branch, and a Bayesian algorithm is called to select the best convolution kernel parameter combination for each branch network. Specifically: First, define the parameter space of the convolution kernel as the search range for Bayesian optimization. The convolution kernel size of each branch must be iteratively selected within a reasonable range and meet its specific time scale requirements.

[0041] Short-term branch is used to capture the transient characteristics of equipment start-stop, current mutation, etc., reflecting the short-period fluctuations of SCADA signals. The convolution kernel size range of this embodiment is set to 3-5, corresponding to a time scale of 45-75 minutes; Regular branch is used to model the production rhythm and shift switching rules, reflecting the influence of MES shift information. The convolution kernel size range of this embodiment is set to 12-24, corresponding to a time scale of 3-6 hours; Long-term branch is used to extract long-term trend information such as peak avoidance strategy and circadian rhythm, reflecting differences in enterprise operation strategy. The convolution kernel size range of this embodiment is set to 48-96, corresponding to a time scale of 12-24 hours.

[0042] In a further optional implementation, the number of convolution kernels of the three branches is uniformly set to 1-3 in this embodiment.

[0043] In a further implementation, based on the 3 industry data subsets output by S2, Bayesian optimization iteration training is called to output the optimal convolution kernel parameters of each branch network, including size and number.

[0044] In a further optional implementation, the number of convolution kernels of the three branches is uniformly set to 1-3 in this embodiment.

[0045] Based on the optimal convolution kernel parameters of each branch output, short-term branch CNN, regular branch CNN, and long-term branch CNN are initialized, and based on the industry data subsets output by S2, multi-scale feature extraction is completed.

[0046] 3.2: After multiple CNNs complete multi-scale feature extraction, the features output by multiple branches are spliced. The output of each branch is , is the batch size, is the time step, is the feature dimension, is the branch index. The three features of different scales are spliced in the feature dimension The shape of the spliced tensor is , is the feature dimension of the branch , is the number of branches.

[0047] As an example of this embodiment, the number of branch networks is 3, and the shape of the spliced tensor is: ; The spliced feature tensor can be used as the input of BiLSTM, which learns the time series dependency based on these multi-scale features and extracts more rich time series information.

[0048] BiLSTM captures the temporal dependencies by taking both directions, and outputs . Then, an attention mechanism is applied on the output of BiLSTM to calculate the weight of each time step, and the aggregated general feature vector is obtained by weighted summation, which represents the common rules captured by the model.

[0049] S4: As an implementation of the method embodiment, independent sub-layers are constructed according to the operation mode (such as all-day operation enterprises, peak-avoiding operation enterprises, and daytime operation enterprises), and each sub-layer will enhance the general feature vector to capture the unique data rules of each operation type enterprise.

[0050] The above branch CNN-BiLSTM-Attention backbone network provides a unified basic feature for all industries, which contains the common information of all industries and provides general feature support for all operation modes.

[0051] 4.1: In a further implementation, operation type sub-layers are constructed, and multiple sub-layers are constructed according to the operation mode list, each of which is responsible for processing and outputting data of a specific operation type. For example, corresponding to the above implementation steps, there are all-day operation enterprise sub-layer, peak-avoiding operation enterprise sub-layer, daytime operation enterprise sub-layer, etc.

[0052] Each sub-layer processes the input feature and performs targeted enhancement processing. Taking the three operation mode sub-layers of this embodiment as an example: All-day operation enterprise sub-layer: All-day operation enterprises usually have relatively stable load fluctuations without obvious peaks or troughs, so when processing features, attention is focused on stability and long-term trends while preventing overfitting. The feature enhancement method of this sub-layer is: first, input the input feature to the fully connected layer to adjust the feature space, ensuring that the input feature can be reasonably converted; then perform batch normalization to standardize the data of each batch, ensuring that the data is more stable during training. Finally, execute the ReLU activation function to enable the feature to capture more nonlinear relationships and improve the performance of the model.

[0053] Peak-avoiding operation enterprise sub-layer: Peak-avoiding operation enterprises usually experience periodic fluctuations, so special attention needs to be paid to periodic features to enhance their periodicity and trend capture ability. The feature processing method of this sub-layer is similar to that of the all-day operation enterprise sub-layer, with the difference being that batch normalization is replaced by Dropout, which randomly discards part of the features during training to avoid overfitting.

[0054] Daytime operation enterprise sub-layer: the load of daytime operation enterprise fluctuates greatly, and the characteristic enhancement focuses on capturing periodic fluctuations and short-term peaks, and has sensitivity to day-night periodicity. The feature processing method of this sub-layer is: the same architecture as the all-day operation enterprise sub-layer is applied, the difference is that the periodic feature coding (sinusoidal and cosine function coding can be used to reflect the day-night periodicity of time) is added, which can capture the periodic changes of the load.

[0055] 4.2: In a further embodiment, the output enhanced features of each sub-layer will output the final load prediction result of the future time step through the output layer, which includes a fully connected layer and an activation function, and the enhanced features and general feature vectors are fused in the fully connected layer, and the prediction result is output through the activation function.

[0056] As an embodiment of the present application, a multi-scale feature fusion industrial load prediction system is disclosed, which adopts the specific embodiments of the industrial power load prediction method described above, and the system comprises: A data acquisition and processing module is used to collect smart meter power data, equipment SCADA system operating parameters and MES system production work orders, construct a comprehensive data set, time align the data of different sampling frequencies in the overall data set and perform preprocessing; A data set splitting module is used to split the preprocessed comprehensive data set according to industry type, and further split the enterprises at each industry type according to the operation mode; finally, the preprocessed comprehensive data set is split into multiple industry data subsets, and each industry data subset is standardized again; A branch training module is used to design a general backbone training network; the feature extraction network in the general backbone training network adopts a plurality of parallel branch network structures, each branch network corresponds to an industry data subset, and different scale features are extracted; wherein the convolutional parameters of each branch network are iteratively selected by an optimization algorithm to select the best parameter combination; the best parameter combination is applied to the general backbone training network to output a general feature vector; A prediction output module is used to design multiple operation type sub-layers according to the operation mode, and extract mode features under different operation modes from the general feature vector; the mode features output by each operation type sub-layer are combined with the general feature vector, and the industrial load prediction of the future time step is output through the corresponding output layer.

[0057] As an embodiment of the present application, an electronic device is provided, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the computer program is loaded into the processor, the specific embodiments of the industrial power load prediction method described above are adopted.

[0058] As an embodiment of the present application, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to adopt the specific embodiments of the industrial power load prediction method.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application rather than limit them, and although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the specific embodiments of the present application can still be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or equivalent replacement without departing from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A method for predicting industrial electricity load through multi-scale feature fusion, characterized in that, include: Collect power data from smart meters, operating parameters of equipment SCADA systems, and production work order data from MES systems to construct a comprehensive dataset; Time alignment and preprocessing are performed on data from different sampling frequencies in the comprehensive dataset. The preprocessed comprehensive dataset is split by industry type, and further split at the enterprise level according to the operation mode within each industry type; finally, the preprocessed comprehensive dataset is split into multiple industry data subsets, and each industry data subset is standardized twice. Design a general backbone training network; The feature extraction network in the general backbone training network adopts a structure of multiple parallel branch networks. Each branch network corresponds to a subset of industry data and extracts features at different scales. The convolution parameters of each branch network are iteratively selected to obtain the best parameter combination through an optimization algorithm. The best parameter combination is applied to the general backbone training network to output a general feature vector. Multiple sub-layers for different work types are designed based on the work patterns to extract pattern features under different work patterns from the general feature vector; the pattern features output by each sub-layer for different work types are output through the corresponding output layer to predict the industrial load for future time steps.

2. The industrial electricity load forecasting method based on multi-scale feature fusion according to claim 1, characterized in that, The feature extraction network employs a multi-parallel branch network structure, including: Each branch network consists of a convolutional network for feature extraction, an algorithm for extracting temporal features, and an attention mechanism; Each branch network corresponds to a time scale, and the Bayesian search range of the convolution parameters of each branch network is set within the time scale. Based on a subset of industry data, the optimal parameter combination of each branch network is output through Bayesian iterative training.

3. The industrial electricity load forecasting method based on multi-scale feature fusion according to claim 1, characterized in that, Applying the optimal parameter combination to a general backbone training network includes: The optimal parameter combination of each branch network is applied to the convolutional layer of the general backbone network to extract features at multiple scales. The features output by multiple branch networks are concatenated and input into a bidirectional long short-term memory network to extract temporal information. Then, a general feature vector is aggregated and output through an attention mechanism.

4. The industrial electricity load forecasting method based on multi-scale feature fusion according to claim 1, characterized in that, The features output from multiple branch networks are concatenated, including: The output of each branch network is , For batch size, For time steps, For feature dimension, For branch index; Features extracted from all branch networks in the feature dimension The above are spliced ​​together, and the shape of the spliced ​​tensor is as follows: , For branches Feature dimensions, This represents the number of branches.

5. The industrial electricity load forecasting method based on multi-scale feature fusion according to claim 1, characterized in that, Based on the work mode, multiple sub-layers of work types are designed, including: Multiple job type sub-layers are constructed based on the job mode list, with each job type sub-layer corresponding to a job mode; feature enhancement methods are designed in each job type sub-layer according to the data characteristics of different job modes; the general feature vector is output to each sub-layer, and each sub-layer outputs different mode features.

6. The industrial electricity load forecasting method based on multi-scale feature fusion according to claim 1, characterized in that, The pattern features output by each job type sub-layer are passed through the corresponding output layer, including: An output head is designed for each job type sub-layer. Each output head passes the pattern features output by the corresponding job type sub-layer through an output layer containing a fully connected layer and an activation function to output the load prediction for the future time step. In the fully connected layer, the pattern features and the general feature vector are fused.

7. The industrial electricity load forecasting method based on multi-scale feature fusion according to claim 1, characterized in that, Time alignment is performed on data from different sampling frequencies in the overall dataset, including: Set a unified timestamp and fill in missing data for power data from smart meters with different sampling frequencies, operating parameters of equipment SCADA systems, and production work order data from MES systems; Select one type of data from the comprehensive dataset as the reference sequence, and use the other two types of data from the comprehensive dataset as the sequences to be aligned. The time warping algorithm is invoked to align the sequence to be aligned with the reference sequence; the time scale of the time-aligned composite dataset is consistent with that of the reference sequence.

8. An industrial load forecasting system with multi-scale feature fusion, executing the industrial electricity load forecasting method as described in any one of claims 1-7, characterized in that, The system includes: Data acquisition and processing module; used to collect power data from smart meters, operating parameters of the equipment's SCADA system, and production work orders from the MES system to build a comprehensive dataset; time alignment and preprocessing of data with different sampling frequencies in the overall dataset. The dataset splitting module is used to split the preprocessed comprehensive dataset by industry type, and further split each industry type by enterprise level according to the operation mode; finally, the preprocessed comprehensive dataset is split into multiple industry data subsets, and each industry data subset is standardized twice. Branch training module; used to design a general backbone training network; for the feature extraction network in the general backbone training network, multiple parallel branch network structures are adopted, each branch network corresponds to a subset of industry data, and features at different scales are extracted; the convolution parameters of each branch network are iteratively selected through an optimization algorithm to select the best parameter combination; the best parameter combination is applied to the general backbone training network to output a general feature vector; The prediction output module is used to design multiple sub-layers for different work types based on the work mode, and extract the mode features under different work modes from the general feature vector. The mode features output by each sub-layer of work type are combined with the general feature vector and output the industrial load prediction for the future time step through the corresponding output layer.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the industrial power load forecasting method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the industrial power load forecasting method according to any one of claims 1-7.

Images