Industrial load prediction method and device based on time sequence structure maintenance, and medium

Abstract

The invention provides an industrial load prediction method and device based on time sequence structure maintenance and a medium, and relates to the technical field of industrial load prediction. The method comprises the following steps: acquiring industrial load time series data and exogenous factor time series data of a target industrial park in a historical time period; preprocessing, standardizing and blocking the industrial load time series data and the exogenous factor time series data, inputting a pre-trained industrial load prediction model, and inversely standardizing a predicted value output by the industrial load prediction model to obtain a predicted industrial load of the target industrial park in a set time period in the future; wherein the industrial load prediction model is a hybrid model obtained by performing deep integration on a Transform encoder and a large model decoder. According to the method provided by the invention, an accurate industrial load prediction result can be efficiently generated, and a reliable basis is provided for efficient management and resource optimization of an industrial park.

Description

Technical Field

[0001] The present application relates to the technical field of industrial load forecasting, and in particular to an industrial load forecasting method, device, electronic device and storage medium based on time series structure preservation. Background Art

[0002] Electricity load, also known as "power consumption load," is the total amount of electrical power drawn from the power system by a user's electrical equipment at a given moment. Based on the different load characteristics of electricity users, electricity load can be categorized into various types, including industrial load, agricultural load, transportation load, and household load.

[0003] Accurately forecasting power loads in industrial parks plays a crucial role in optimizing energy management and improving operational efficiency. However, this area faces unique technical challenges: the multidimensional nature of load series stems not only from their inherent dynamic evolution but also from complex exogenous factors. Industrial loads exhibit complex temporal characteristics, with dependencies spanning different time scales, cross-series correlations between different building types, and are significantly influenced by external variables such as volatile weather patterns, rapidly changing production schedules, and varying efficiency metrics. These complexities require forecasting methods that go beyond simple pattern recognition and can adapt to diverse behaviors and changing operating conditions.

[0004] Previous industrial load forecasting methods suffered from the following shortcomings: 1. Structural flaws in traditional forecasting models: Classic time series models, such as the autoregressive moving average (ARMA) model and exponential smoothing, are constrained by stationary assumptions and linear modeling paradigms, making them difficult to capture the nonlinear dynamic evolution of industrial loads. Core issues include insufficient adaptability to nonstationary operating conditions, failing to capture sudden load fluctuations caused by production equipment startups and shutdowns; incomplete feature space construction, and a lack of ability to effectively mine the diverse and heterogeneous information contained in exogenous variables (such as meteorological parameters, production schedules, and energy efficiency indicators). 2. Adaptability limitations of deep learning models: Deep learning technology still exhibits significant limitations when processing time series data. Remote dependency modeling and heterogeneous integration attempt to uniformly source data from heterogeneous sources. Furthermore, these methods typically treat all data as homogeneous. Real industrial load data is much more complex, and this lack of diversity can lead to suboptimal results. 3. Domain adaptability limitations of the Transformer architecture: Existing Transformer models are primarily designed for single-variable or homogeneous multivariate sequences. There is usually a lack of mechanisms to effectively decompose multi-objective sequences or effectively integrate various exogenous variables, which are crucial for industrial load forecasting. Therefore, how to accurately forecast industrial load is a technical problem that needs to be solved in this field. Summary of the Invention

[0005] The purpose of this application is to provide an industrial load forecasting method, device, electronic device and storage medium based on time series structure maintenance, which can improve the accuracy of industrial load forecasting in industrial parks.

[0006] In a first aspect, an embodiment of the present application provides an industrial load forecasting method based on time series structure preservation, comprising:

[0007] Obtain the industrial load time series data and exogenous factor time series data of the target industrial park within the historical period;

[0008] After preprocessing, standardizing, and segmenting the industrial load time series data and the exogenous factor time series data, the data are input into a pre-trained industrial load forecasting model. After inverse normalization of the predicted value output by the industrial load forecasting model, the predicted industrial load of the target industrial park within a set time period in the future is obtained.

[0009] The industrial load forecasting model is a hybrid model obtained by deeply integrating the Transformer encoder and the large model decoder.

[0010] In some embodiments of the present application, the industrial load forecasting model is pre-trained in the following manner:

[0011] Obtain historical industrial load data and historical exogenous factor data of multiple industrial parks, and statistically convert them into time series within the same time interval, thereby obtaining historical industrial load time series data samples and historical exogenous factor time series data samples of multiple industrial parks;

[0012] After preprocessing, standardizing and segmenting the historical industrial load time series data samples and the historical exogenous factor time series data samples, the samples are input into the initial industrial load forecasting model for training, and a trained industrial load forecasting model is obtained after a preset training cutoff condition is reached.

[0013] In some embodiments of the present application, the segmentation includes: dividing the time series into sequence blocks of fixed length.

[0014] In some embodiments of the present application, the preprocessing includes data consistency processing, maximum value processing and missing value processing.

[0015] In some embodiments of the present application, the industrial load forecasting model includes a feature extraction unit, a feature enhancement unit, and a forecasting unit;

[0016] The feature extraction unit is used to linearly encode the sequence blocks obtained after the block division to map them into a high-dimensional linear space to obtain block embeddings; and to splice all the obtained block embeddings in relative order using a position embedding matrix to obtain a feature splicing result;

[0017] The feature enhancement unit is used to perform feature enhancement on the block embedding in the feature splicing result to obtain an enhanced feature splicing result;

[0018] The prediction unit is used to decode according to the enhanced feature splicing result to obtain a prediction value.

[0019] In a second aspect, an embodiment of the present application provides an industrial load forecasting device based on time series structure maintenance, comprising:

[0020] An acquisition module is used to obtain the industrial load time series data and exogenous factor time series data of the target industrial park within a historical period;

[0021] A prediction module is used to preprocess, standardize, and block the industrial load time series data and the exogenous factor time series data, input them into a pre-trained industrial load prediction model, and denormalize the predicted values ​​output by the industrial load prediction model to obtain the predicted industrial load of the target industrial park within a set time period in the future;

[0022] The industrial load forecasting model is a hybrid model obtained by deeply integrating the Transformer encoder and the large model decoder.

[0023] In some embodiments of the present application, the segmentation includes: dividing the time series into sequence blocks of fixed length.

[0024] In some embodiments of the present application, the preprocessing includes data consistency processing, maximum value processing and missing value processing.

[0025] In a third aspect, an embodiment of the present application provides an electronic device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect.

[0026] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having computer-readable instructions stored thereon, and the computer-readable instructions can be executed by a processor to implement the method described in the first aspect.

[0027] Compared with the existing technology, the industrial load forecasting method based on time series structure preservation provided by the present application obtains the industrial load time series data and exogenous factor time series data of the target industrial park in the historical time period; after preprocessing, standardizing and blocking the industrial load time series data and the exogenous factor time series data, the data are input into a pre-trained industrial load forecasting model, and the predicted value output by the industrial load forecasting model is inversely standardized to obtain the predicted industrial load of the target industrial park in the future set time period; wherein, the industrial load forecasting model is a hybrid model obtained by deeply integrating the Transformer encoder and the large model decoder. It can be seen that the industrial load forecasting model of the present application deeply integrates the global perception capability of the Transformer encoder and the sequence generation advantage of the large model decoder, and can effectively combine the exogenous factor data affecting the industrial load to generate accurate industrial load forecast results. Compared with the traditional industrial load forecasting method, the method proposed in the present application can efficiently generate accurate industrial load forecast results, providing a reliable basis for efficient management and resource optimization of industrial parks. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present application. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0029] Figure 1 A flowchart of an industrial load forecasting method based on time series structure preservation provided by an embodiment of the present application is shown;

[0030] Figure 2 A flowchart of an industrial load forecasting model training method provided by an embodiment of the present application is shown;

[0031] Figure 3 A schematic diagram of the prediction process of an industrial load prediction model provided by an embodiment of the present application is shown;

[0032] Figure 4 A schematic diagram of the prediction process of another industrial load prediction model provided in an embodiment of the present application is shown;

[0033] Figure 5 A schematic diagram of an industrial load forecasting device based on time series structure maintenance provided by an embodiment of the present application is shown;

[0034] Figure 6 A schematic diagram of an electronic device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0035] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0036] It should be noted that, unless otherwise specified, the technical or scientific terms used in this application should have the common meanings understood by those skilled in the art to which this application belongs.

[0037] In addition, the terms "first" and "second" are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements, but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0038] In order to better model multi-scale temporal patterns and fully utilize the parallel processing capabilities of modern hardware for multi-objective forecasting, this application provides an industrial load forecasting method based on temporal structure preservation. This method achieves parallel processing and effective integration of exogenous variables by decomposing multi-objective load sequences into univariate target sequences. Each univariate sequence can be modeled independently while still integrating relevant external data. This method achieves efficient and accurate forecasting of industrial power loads by integrating the Transformer encoder and GPT-2 architecture, combined with an innovative temporal structure preservation mechanism.

[0039] Please refer to Figure 1 , Figure 1 This is a flow chart of an industrial load forecasting method based on time series structure preservation provided by an embodiment of the present application, including the following S101 to S102:

[0040] S101. Obtain industrial load time series data and exogenous factor time series data of a target industrial park within a historical time period.

[0041] The industrial load time series data is the time series data of multiple industrial loads in the target industrial park at set time intervals. The exogenous factor time series data is the time series data of multiple exogenous factors affecting the industrial load in the target industrial park at set time intervals.

[0042] The time interval can be set to a specific time interval such as hour, day, week or month. Exogenous factors can include changing weather patterns and rapidly changing production schedules.

[0043] S102. After preprocessing, standardizing and segmenting the industrial load time series data and the exogenous factor time series data, the data are input into a pre-trained industrial load forecasting model. The predicted value output by the industrial load forecasting model is inversely standardized to obtain the predicted industrial load of the target industrial park within a set time period in the future.

[0044] The industrial load forecasting model is a hybrid model that deeply integrates a Transformer encoder with a large-scale decoder. The Transformer encoder's core function is to process heterogeneous data from multiple sources (such as meteorological data and process parameters) in parallel, decoupling features at different time scales through multi-head attention (12-24 heads). The large-scale decoder's core function is to transform its autoregressive generation mechanism into a prediction-oriented architecture, which has the advantage of sequence generation.

[0045] Exemplarily, the large model can be a GPT-2 model, whose key components include autoregressive mechanism, scaled dot product attention, and feedforward network, etc.

[0046] Exemplarily, the above-mentioned data preprocessing operations include: data consistency processing, maximum value processing and missing value processing.

[0047] The primary purpose of data consistency processing is to identify and resolve data consistency issues and ensure data consistency across different systems or environments. Data consistency processing can promptly identify data errors, data loss, data duplication, and other issues, allowing appropriate measures to be taken to repair them, thereby ensuring data accuracy and integrity.

[0048] Exemplarily, the above-mentioned operation of data segmentation includes: segmenting the time series into sequence blocks of fixed length.

[0049] Specifically, the above industrial load forecasting model is obtained by training with historical industrial load time series data samples and historical exogenous factor time series data samples of multiple industrial parks.

[0050] Specifically, the above industrial load forecasting model includes a feature extraction unit, a feature enhancement unit and a forecasting unit.

[0051] The feature extraction unit is used to linearly encode the sequence blocks obtained after the block division to map them into a high-dimensional linear space to obtain block embeddings; and to splice all the obtained block embeddings in relative order using a position embedding matrix to obtain a feature splicing result;

[0052] The feature enhancement unit is used to perform feature enhancement on the block embedding in the feature splicing result to obtain an enhanced feature splicing result;

[0053] The prediction unit is used to decode according to the enhanced feature splicing result to obtain a prediction value.

[0054] This application first standardizes and preprocesses the raw industrial load data to ensure consistent statistical characteristics across variables. It then segments the time series into fixed-length blocks to extract local temporal patterns. Finally, it captures complex temporal dependencies using an improved Transformer-GPT-2 hybrid model, while preserving key temporal structural features using feature enhancement units and multi-scale representations. Compared to existing technologies, this application significantly improves prediction accuracy, computational efficiency, and scalability, making it particularly suitable for industrial load forecasting scenarios with complex nonlinear dynamics and multiple exogenous variables.

[0055] The following describes how to pre-train an industrial load forecasting model. Specifically, the industrial load forecasting model can be pre-trained in the following ways: Figure 2 As shown, it includes steps S201 to S203:

[0056] S201. Obtain historical industrial load data and historical exogenous factor data for multiple industrial parks, and compile them into time series within the same time interval to obtain historical industrial load time series data samples and historical exogenous factor time series data samples for the multiple industrial parks;

[0057] S202. After preprocessing, standardizing and segmenting the historical industrial load time series data samples and the historical exogenous factor time series data samples, the samples are input into an initial industrial load forecasting model for training, and a trained industrial load forecasting model is obtained after a preset training cutoff condition is reached.

[0058] First, real-world industrial load data are collected, and the raw data are preprocessed and standardized to ensure that all variables (including industrial load and exogenous factors) have consistent statistical characteristics.

[0059] Consider time t i The historical input segment is composed of a multi-objective power load matrix and the exogenous factor matrix Indicates. Where T is the length of the historical observation window (number of time steps), N T is the number of industrial load time series, N E If not standardized, those time series with Larger magnitude variables may dominate model behavior, complicate parameter optimization, and potentially distort the learned representations.

[0060] Therefore, this application adopts a reversible normalization procedure that stabilizes the variance, enhances convergence, and is consistent with common practices in neural network-based prediction models. Represents a general input matrix with variable dimension N T or N E , depending on whether you are standardizing industrial load data or exogenous data. For each variable dimension n, calculate the mean μ n and standard deviation σ n :

[0061]

[0062] in, represents the time step t j The value of the nth variable at position ε is a small constant used to prevent division by zero and improve numerical stability.

[0063] Using these statistics, each entry is normalized to:

[0064]

[0065] Based on the normalized input, this application segments the time series into fixed-length chunks. The motivation for this approach is that, under a mild stationarity assumption, local segments of the time series may exhibit relatively stable correlations and low noise. By focusing on these local structures, the model can more effectively capture short-term dependencies.

[0066] Please refer to Figure 3 ,The following is a detailed introduction to the industrial load forecasting model.

[0067] 1. By introducing a position parameter, subsequent linear transformations can more easily identify and reconstruct deviations from the baseline, potentially improving the model's representational power. Without the mean, certain affine transformations would require more complex parameterizations to represent baseline shifts. By adding the mean, a single linear layer can more easily account for these changes, ensuring it operates in a space that explicitly incorporates baseline information.

[0068] To further enhance these segmented sequence blocks, this application uses a linear encoder: Map the segmented sequence blocks into a high-dimensional linear space. Apply f row by row enc The encoded block embedding can be obtained:

[0069]

[0070] Among them, the output dimension is P+1×d model , where P is the block length, d model is the model embedding representation, This transformation can be understood as feature expansion, where the time series block is embedded into a space that is more conducive to recognizing complex nonlinear temporal relationships.

[0071] By isolating short-term dependencies and providing a baseline reference, this application can more effectively utilize attention mechanisms and multi-scale integration to capture local and global temporal patterns. This structured approach ensures that the framework can robustly handle complex temporal dependencies and reliably scale to various real-world prediction scenarios.

[0072] 2. This application must incorporate strict temporal order information before inputting the block embeddings into the encoder based on Transformer and GPT-2. To make the temporal order clear, this application adds a position embedding matrix to encode the relative order of each position in the sequence, and obtains the concatenation result H of all block embeddings:

[0073]

[0074] Wherein, L=M(P+1) is the total sequence length after connecting all blocks (number of blocks M×length of each block P+1).

[0075] The core of Transformer is the attention mechanism, which can focus on different parts of the sequence at the same time. represents the universal input of the attention layer. The model computes the query Q, key K, and value V through the learned projection:

[0076] Q=XW Q ,K=XW K ,V=XW V ;

[0077] Among them, the learnable projection matrix And the dimension of each attention head is For h heads, the attention weights are calculated by softmax and the weighted value matrix is ​​calculated. The scaled dot product attention mechanism is calculated as follows:

[0078]

[0079] The heads are concatenated and projected back:

[0080] H attn =[head1,head2,…head h ]W o ;

[0081] Among them, the output projection matrix

[0082] This mechanism can be understood as the method's ability to learn multiple temporal "views" of the data, each focusing on different types of dependencies. From a theoretical perspective, multi-head attention mechanisms can approximate a variety of sequence-to-sequence functions. These capabilities are crucial for capturing both short-range and long-range dependencies in time series.

[0083] 3. Position Feedforward Network (FFN). Each Transformer module also contains a position feedforward network, which is defined as:

[0084] H ffn =max(H attn W1+b1,0)W2+b2;

[0085] Where W1, W2, b1, and b2 are parameters. FFN refines the representation by applying nonlinear transformations independently at each location, thereby enhancing the model's ability to capture complex nonlinear temporal interactions.

[0086] 4. The computational complexity of the self-attention mechanism is approximately O(L 2 d model ), where L is the sequence length. By splitting the input into chunks, we ensure that L remains manageable, thus balancing representational power and computational feasibility.

[0087] 5. Using GPT-2 for prediction requires treating block embeddings as linguistic tokens and temporal order as word order. Positional encoding ensures that temporal information is explicitly modeled. The encoder's attention layer learns not only correlations within a single block but also correlations across multiple blocks, thereby capturing both local and global temporal patterns. This approach builds on the intuition that attention mechanisms can also reveal and exploit structure in time series data.

[0088] 6. This application preserves the relative similarity or distance between points in the time embedding by applying some transformations. This ensures that structurally similar time periods remain comparable after the transformation. By preserving these relationships, the model can effectively model the data and adapt to changes at different time scales. We use a set of basis functions derived from mature time series expansions to represent complex temporal patterns. These functions (such as sine functions, polynomial functions, etc.) provide a theoretical basis for approximating various temporal patterns while still explicitly preserving their underlying structural information.

[0089] make is the encoder output, where L and d model Following the notation established previously, each line Where l = 1, ..., L is a block-level embedding used to capture complex dependencies. Define a family of transformations Each transformation is designed to model a specific time scale or structural pattern.

[0090] For example, consider the sine transform:

[0091] C i (Z)=[sin(ω i W i z),cos(ω i W i z)];

[0092] in, is a learnable parameter matrix, where d out is the output dimension, ω i is a frequency parameter used to capture certain periodic behaviors. Polynomial transformations or other kernel expansions can be defined similarly. By choosing an appropriate set of functions (e.g., for sinusoidal components, a limited set of frequencies; for trend modeling, a polynomial degree), the general term Z after multiscale concatenation is constructed. TSM :

[0093]

[0094] Among them, d TSM is the dimension obtained after concatenation.

[0095] 7. The choice of sine and polynomial functions is not arbitrary. The universal approximation theorem shows that combinations of trigonometric or polynomial functions can approximate a wide range of continuous-time functions with arbitrary accuracy.

[0096] This concept can be formalized as follows:

[0097] set up is a continuous function, which means that in the compact domain For any ε>0, there exists an integer M, a parameter matrix {W i} and the frequency or scale parameter {ω i}, making the family (defined by sine or polynomial expansions) satisfying:

[0098]

[0099] 8. This application introduces the feature enhancement unit FEU as an operator to emphasize the correlation deviation in the learned block embedding. Let represents the multi-scale representation of TSM, where L represents the length of the patched sequence, d TSM represents the embedding dimension.

[0100] Define FEU as:

[0101]

[0102] For each block embedding FEU first calculates a center point:

[0103]

[0104] This center c is used as a baseline reference. The deviation of each block embedding from c is then measured:

[0105] r l =Z l -c,d l =||r l ||2;

[0106] And calculate the mean deviation:

[0107]

[0108] The strength of emphasizing these deviations is controlled by the power parameter p>0. The enhanced block embedding is defined as:

[0109]

[0110] in, is the output of the feature enhancement unit, c is the embedding center point, r l is the deviation between the embedding and the center, and p is a parameter that controls the strength of the deviation emphasis.

[0111] The industrial load forecasting model learns the industrial load time series data and exogenous factor time series data through the above steps, and outputs the industrial load forecast value to be processed. Restoring it to its original scale gives the final predicted industrial load:

[0112]

[0113] The use of reversible normalization, which maintains interpretability and facilitates practical deployment, ensures that the adaptive feature coupling and correlation extraction encoder modules receive well-conditioned inputs, allowing them to focus on uncovering complex temporal and cross-variable patterns rather than coping with scale differences.

[0114] like Figure 4 As shown, this application can collect real-time industrial load data and exogenous factor data from industrial parks, preprocess the data, train industrial load forecasting models, and provide real-time forecasts. The forecast results are then fed back to the industrial park, providing a reliable basis for efficient management and resource optimization. This is particularly suitable for scenarios requiring accurate forecasting of complex industrial loads, such as power system optimization, industrial energy management, and intelligent manufacturing.

[0115] The industrial load forecasting method based on time series structure preservation provided by the embodiment of the present application has the following beneficial effects:

[0116] 1. This method achieves parallel processing and efficient integration of exogenous variables by decomposing multi-objective load series into univariate target series. This allows each univariate series to be modeled independently while still integrating relevant external data. This decomposition enables more efficient and effective modeling of complex industrial load patterns.

[0117] 2. This method employs a temporal structure-preserving mechanism to learn and extract multi-scale temporal patterns within each univariate target series by explicitly preserving the relationships between these patterns. This adapts to the different patterns observed at different time scales, simultaneously capturing both short-term and long-term dependencies, extracting rich temporal information, and enhancing the model's ability to simulate the dynamic characteristics of industrial loads.

[0118] 3. By combining this structure-preserving approach with multi-scale representation and attention framework, our method is able to better capture short-term and long-term dependencies while maintaining scalability and interpretability for industrial load forecasting.

[0119] In the above embodiment, an industrial load forecasting method based on time series structure maintenance is provided. Correspondingly, the present application also provides an industrial load forecasting device based on time series structure maintenance. The industrial load forecasting device based on time series structure maintenance provided in the embodiment of the present application can implement the above industrial load forecasting method based on time series structure maintenance. The industrial load forecasting device based on time series structure maintenance can be implemented through software, hardware, or a combination of software and hardware. For example, the industrial load forecasting device based on time series structure maintenance can include integrated or separate functional modules or units to perform the corresponding steps in the above methods. Please refer to Figure 5 As shown, the industrial load forecasting device 10 based on time series structure maintenance provided by the present application includes:

[0120] An acquisition module 101 is used to acquire industrial load time series data and exogenous factor time series data of a target industrial park within a historical period;

[0121] The forecasting module 102 is configured to preprocess, standardize, and block the industrial load time series data and the exogenous factor time series data, input the data into a pre-trained industrial load forecasting model, and denormalize the predicted values ​​output by the industrial load forecasting model to obtain the predicted industrial load of the target industrial park within a set time period in the future;

[0122] The industrial load forecasting model is a hybrid model obtained by deeply integrating the Transformer encoder and the large model decoder.

[0123] In some embodiments of the present application, the device further comprises:

[0124] A training module is used to pre-train the industrial load forecasting model in the following manner:

[0125] Obtain historical industrial load data and historical exogenous factor data of multiple industrial parks, and statistically convert them into time series within the same time interval, thereby obtaining historical industrial load time series data samples and historical exogenous factor time series data samples of multiple industrial parks;

[0126] After preprocessing, standardizing and segmenting the historical industrial load time series data samples and the historical exogenous factor time series data samples, the samples are input into the initial industrial load forecasting model for training, and a trained industrial load forecasting model is obtained after a preset training cutoff condition is reached.

[0127] In some embodiments of the present application, the industrial load forecasting model includes a feature extraction unit, a feature enhancement unit, and a forecasting unit;

[0128] The feature extraction unit is used to linearly encode the sequence blocks obtained after the block division to map them into a high-dimensional linear space to obtain block embeddings; and to splice all the obtained block embeddings in relative order using a position embedding matrix to obtain a feature splicing result;

[0129] The feature enhancement unit is used to perform feature enhancement on the block embedding in the feature splicing result to obtain an enhanced feature splicing result;

[0130] The prediction unit is used to decode according to the enhanced feature splicing result to obtain a prediction value.

[0131] The industrial load forecasting device based on time series structure maintenance provided in the embodiment of the present application and the industrial load forecasting method based on time series structure maintenance provided in the embodiment of the present application are based on the same inventive concept and have the same beneficial effects as the methods adopted, operated or implemented therein.

[0132] An embodiment of the present application also provides an electronic device corresponding to the method provided in the aforementioned embodiment, and the electronic device may be an electronic device for a client, such as a mobile phone, a laptop computer, a tablet computer, a desktop computer, etc., to execute the above-mentioned industrial load forecasting method based on time series structure maintenance.

[0133] Please refer to Figure 6 , which shows a schematic diagram of an electronic device provided by some embodiments of the present application. Figure 6As shown, the electronic device 20 includes: a processor 200, a memory 201, a bus 202 and a communication interface 203, and the processor 200, the communication interface 203 and the memory 201 are connected via the bus 202; the memory 201 stores a computer program that can be run on the processor 200, and when the processor 200 runs the computer program, it executes the industrial load forecasting method based on time series structure maintenance provided by any of the aforementioned embodiments of the present application.

[0134] The memory 201 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage. The system network element communicates with at least one other network element via at least one communication interface 203 (which may be wired or wireless), and may utilize the Internet, a wide area network, a local area network, a metropolitan area network, or the like.

[0135] The bus 202 may be an ISA bus, a PCI bus, or an EISA bus. The bus may be divided into an address bus, a data bus, a control bus, and the like. The memory 201 is used to store programs, and the processor 200 executes the programs upon receiving execution instructions. The industrial load forecasting method based on time series structure preservation disclosed in any of the aforementioned embodiments of the present application may be applied to or implemented by the processor 200.

[0136] The processor 200 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by hardware integrated logic circuits in the processor 200 or by software instructions. The above processor 200 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in conjunction with the embodiments of this application can be directly implemented and executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory 201 , and the processor 200 reads the information in the memory 201 and completes the steps of the above method in combination with its hardware.

[0137] The electronic device provided in the embodiment of the present application and the industrial load forecasting method based on time series structure maintenance provided in the embodiment of the present application are based on the same inventive concept and have the same beneficial effects as the methods adopted, operated or implemented by them.

[0138] An embodiment of the present application also provides a computer-readable storage medium corresponding to the industrial load forecasting method based on time series structure maintenance provided by the aforementioned embodiment, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it will execute the industrial load forecasting method based on time series structure maintenance provided by any of the aforementioned embodiments.

[0139] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical or magnetic storage media, which are not listed here one by one.

[0140] The computer-readable storage medium provided in the above-mentioned embodiments of the present application and the industrial load forecasting method based on time series structure maintenance provided in the embodiments of the present application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application programs stored therein.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application, and they should all be included in the scope of the claims and description of the present application.

Claims

1. An industrial load forecasting method based on time series structure preservation, characterized in that: include: Obtain the industrial load time series data and exogenous factor time series data of the target industrial park within the historical period; After preprocessing, standardizing, and segmenting the industrial load time series data and the exogenous factor time series data, the data are input into a pre-trained industrial load forecasting model. After inverse normalization of the predicted value output by the industrial load forecasting model, the predicted industrial load of the target industrial park within a set time period in the future is obtained. The industrial load forecasting model is a hybrid model obtained by deeply integrating the Transformer encoder and the large model decoder.

2. The method according to claim 1, characterized in that The industrial load forecasting model is pre-trained as follows: Obtain historical industrial load data and historical exogenous factor data of multiple industrial parks, and statistically convert them into time series within the same time interval, thereby obtaining historical industrial load time series data samples and historical exogenous factor time series data samples of multiple industrial parks; After preprocessing, standardizing and segmenting the historical industrial load time series data samples and the historical exogenous factor time series data samples, the samples are input into the initial industrial load forecasting model for training, and a trained industrial load forecasting model is obtained after a preset training cutoff condition is reached.

3. The method according to claim 1 or 2, characterized in that The blocking includes: dividing the time series into sequence blocks of fixed length.

4. The method according to claim 1 or 2, characterized in that The preprocessing includes data consistency processing, maximum value processing and missing value processing.

5. The method according to claim 1 or 2, characterized in that The industrial load forecasting model includes a feature extraction unit, a feature enhancement unit and a forecasting unit; The feature extraction unit is used to linearly encode the sequence blocks obtained after the block division to map them into a high-dimensional linear space to obtain block embeddings; and to splice all the obtained block embeddings in relative order using a position embedding matrix to obtain a feature splicing result; The feature enhancement unit is used to perform feature enhancement on the block embedding in the feature splicing result to obtain an enhanced feature splicing result; The prediction unit is used to decode according to the enhanced feature splicing result to obtain a prediction value.

6. The method according to claim 1, characterized in that The large model includes the GPT-2 model.

7. An industrial load forecasting device based on time series structure maintenance, characterized in that: include: An acquisition module is used to obtain the industrial load time series data and exogenous factor time series data of the target industrial park within a historical period; A prediction module is used to preprocess, standardize, and block the industrial load time series data and the exogenous factor time series data, input them into a pre-trained industrial load prediction model, and denormalize the predicted values ​​output by the industrial load prediction model to obtain the predicted industrial load of the target industrial park within a set time period in the future; The industrial load forecasting model is a hybrid model obtained by deeply integrating the Transformer encoder and the large model decoder.

8. The device according to claim 7, characterized in that The preprocessing includes data consistency processing, maximum value processing and missing value processing.

9. An electronic device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 6 when executing the computer program.

10. A computer-readable storage medium, characterized in that Computer-readable instructions are stored thereon, and the computer-readable instructions can be executed by a processor to implement the method according to any one of claims 1 to 6.

Images