Typical industry power consumption prediction method, device, equipment and medium

Abstract

The application discloses a typical industry power consumption prediction method, device, equipment and medium, and relates to the technical field of data processing. First, the time series data of daily power consumption of a target typical industry is acquired, time sequence characteristic sequences are generated through feature extraction and fusion, and the time sequence characteristic sequences are decomposed into multiple intrinsic mode components through modal decoupling processing. Then, multiple track difference sliding window sequences of each component are constructed based on a preset memory length, and are respectively input into preset mixed prediction models corresponding to each component to obtain difference prediction results. Finally, the final power consumption prediction value is generated through physical inverse difference reconstruction and superposition fusion. Through the fine processing procedure of adaptive decomposition, space-time feature enhancement, modal independent modeling and physically interpretable reconstruction, the model's capturing ability, prediction accuracy and cross-industry generalization for non-stationary, multi-period and strong fluctuation time series data of a typical industry are significantly improved, thereby effectively improving the accuracy and robustness of typical industry power consumption prediction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, equipment and medium for predicting electricity consumption in a typical industry. Background Technology

[0002] Currently, short-term electricity consumption forecasting methods typically employ a unified feature engineering and model architecture to process electricity consumption data from multiple industries. Specifically, this involves concatenating historical electricity consumption sequences with calendar features and then inputting them into frameworks such as temporal convolutional networks, long short-term memory networks, or variational mode decomposition combined with a single prediction model for end-to-end modeling to achieve rolling electricity consumption forecasts. However, existing methods fail to design adaptive differentiated modeling strategies to address the fundamental differences in electricity consumption patterns across industries when dealing with typical industry electricity consumption data. Furthermore, the parameter selection for variational mode decomposition is highly dependent on human experience, and fixed parameters are difficult to adapt to the vastly different frequency domain features of different industries. Single deep learning models are limited by fixed receptive fields or gradient instability, failing to simultaneously meet the dual requirements of capturing local fluctuations and modeling long-distance periodic dependencies. In addition, the simple concatenation of calendar features and historical sequences lacks an adaptive weighting mechanism for feature importance. Most methods employ a decomposition-overall prediction-reconstruction process, leading to the amplification of prediction errors from different modes through mutual superposition. The combined effect of the above problems results in significant deficiencies in the existing methods in terms of non-stationary sequence modeling ability, multi-industry generalization performance, and prediction stability. Ultimately, this leads to the fact that the accuracy and robustness of electricity consumption prediction results for typical industries cannot meet the needs of practical applications. Summary of the Invention

[0003] In view of this, the present invention provides a method, apparatus, equipment and medium for predicting electricity consumption in typical industries, in order to solve the technical problem that the existing technology for predicting electricity consumption in typical industries uses a unified model and has insufficient prediction accuracy and robustness.

[0004] Firstly, a typical industry electricity consumption forecasting method is provided, which includes: Obtain time-series data of daily electricity consumption for typical target industries over a preset historical period; Feature extraction is performed on daily electricity consumption time series data to generate time series feature sequences of typical target industries within a preset historical time period; Modal decoupling is performed on the temporal feature sequence to generate multiple intrinsic mode components; Based on a preset memory length, a multi-track differential sliding window sequence for each intrinsic mode component is constructed. The multi-track differential sliding window sequence of each intrinsic mode component is input into the preset hybrid prediction model corresponding to each intrinsic mode component to obtain the differential prediction result of each intrinsic mode component. Physical inverse differential reconstruction is performed on the differential prediction results of each intrinsic mode component to generate the single-mode power consumption prediction value corresponding to each intrinsic mode component; By superimposing single-mode electricity consumption forecasts from multiple intrinsic mode components, electricity consumption forecasts for typical target industries are generated.

[0005] Secondly, a typical industry electricity consumption forecasting device is provided, the device comprising: The acquisition module is used to acquire time-series data of daily electricity consumption of a typical target industry over a preset historical period. The first generation module is used to extract features from daily electricity consumption time series data and generate time series feature sequences of target typical industries in a preset historical time period. The second generation module is used to perform modal decoupling processing on the temporal feature sequence to generate multiple intrinsic mode components. The module is used to construct multi-track differential sliding window sequences of each intrinsic mode component based on a preset memory length; The third generation module is used to input the multi-track differential sliding window sequence of each intrinsic mode component into the preset hybrid prediction model corresponding to each intrinsic mode component, so as to obtain the differential prediction result of each intrinsic mode component. The fourth generation module is used to perform physical inverse differential reconstruction on the differential prediction results of each intrinsic mode component to generate the single-mode power consumption prediction value corresponding to each intrinsic mode component. The fifth generation module is used to superimpose single-mode electricity consumption prediction values ​​of multiple intrinsic mode components to generate electricity consumption prediction results for typical target industries.

[0006] Thirdly, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described typical industry electricity consumption prediction method.

[0007] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described typical industry electricity consumption forecasting method.

[0008] The above-mentioned typical industry electricity consumption forecasting method, device, equipment, and medium first acquire daily electricity consumption time series data of the target typical industry, generate time series feature sequences through feature extraction and fusion, and decompose them into multiple intrinsic modal components using modal decoupling processing. Then, based on a preset memory length, multi-track differential sliding window sequences for each component are constructed and input into the preset hybrid prediction model corresponding to each component to obtain differential prediction results. Finally, the final electricity consumption forecast value is generated through physical inverse differential reconstruction and superposition fusion. This process transforms the typical industry electricity consumption forecasting problem from traditional manual parameter dependence and coarse hybrid modeling to a refined processing paradigm of adaptive decomposition, spatiotemporal feature enhancement, modal independent modeling, and physically interpretable reconstruction. This significantly improves the model's ability to capture non-stationary, multi-period, and highly fluctuating time series data of typical industries, as well as its prediction accuracy and cross-industry generalization. It effectively solves the problems of parameter selection dependence on manual methods, error accumulation in unified modal modeling, and poor cross-industry adaptability in existing technologies, significantly improving the accuracy and robustness of typical industry electricity consumption forecasting. Attached Figure Description

[0009] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating a typical industry electricity consumption forecasting method in one embodiment of the present invention; Figure 2 This is a schematic diagram of a typical industry electricity consumption prediction device in one embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in the present invention are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of the present invention.

[0011] Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this invention illustrate operations implemented according to some embodiments of the invention. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or performed simultaneously. Moreover, those skilled in the art, guided by the content of this invention, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0012] Furthermore, the embodiments described herein are merely some, not all, of the embodiments of the invention. The components of the embodiments of the invention described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0013] It should be noted that the term "comprising" will be used in the embodiments of the present invention to indicate the presence of a feature subsequently declared, but does not exclude the addition of other features. It should also be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0014] The following is a detailed description of this case, in conjunction with the relevant accompanying drawings in the instruction manual.

[0015] Please see Figure 1 This application provides a typical industry electricity consumption forecasting method, which specifically includes the following steps: S10: Obtain time series data of daily electricity consumption of the target typical industry within a preset historical period.

[0016] It is understood that the executing entity of this invention can be a typical industry electricity consumption forecasting device, or it can be a terminal or a server; no specific limitation is made here. This embodiment of the invention will be described using a server as an example.

[0017] In this step, typical industries refer to industrial or service sectors with significantly identifiable electricity consumption patterns, substantial impact on regional load, and high forecasting difficulty. Daily electricity consumption records of the target typical industry within a preset historical time window are extracted from the power grid dispatch system or power data acquisition platform. The acquired daily electricity consumption time series data are then arranged chronologically to form a continuous discrete sequence.

[0018] Optionally, the preset historical time period can be set according to the actual needs of the forecasting task. For example, all historical daily electricity consumption data within one or two years before the preset forecasting time can be selected to ensure that the sample size is sufficient to cover the multi-cycle change pattern of industry electricity consumption.

[0019] S20: Extract features from daily electricity consumption time series data to generate time series feature sequences of typical target industries within a preset historical time period.

[0020] In this step, periodic features are constructed from the original daily electricity consumption sequence. The original electricity consumption values ​​are fused and encoded with the corresponding time attribute information, so that the electricity consumption data at each historical moment carries contextual information reflecting its time position and periodic attributes. Finally, the original one-dimensional electricity consumption sequence is transformed into a multi-dimensional time-series feature sequence containing electricity consumption values ​​and time attribute semantics. This encoding vector not only retains the amplitude information of the original electricity consumption but also incorporates the inherent periodicity prior knowledge of the time series, providing a structured input representation for subsequent modal decomposition and deep modeling.

[0021] In one embodiment of this application, a specific scheme for generating time-series feature sequences is provided. In S20, feature extraction is performed on daily electricity consumption time-series data to generate time-series feature sequences of a target typical industry within a preset historical time period. This specifically includes the following steps S21-S25: S21: The daily electricity consumption time series data is filled with missing data using linear interpolation.

[0022] In this step, the data points in the daily electricity consumption time series data are first examined to determine if any data points are missing. Each data point refers to the electricity consumption record for each historical day in the daily electricity consumption time series, arranged chronologically. Each data point contains the electricity consumption value for the corresponding date and its associated time attribute information. For each missing data point in the daily electricity consumption time series, the nearest valid observation day before and after each missing data point and its corresponding electricity consumption value are identified. Then, based on the temporal position relationship, linear interpolation is performed between the preceding and following valid observations to estimate the electricity consumption value for the missing day.

[0023] The above methods can maintain the continuity and integrity of the daily electricity consumption sequence without introducing complex external assumptions, providing a complete and reliable data foundation for subsequent periodic feature construction and mode decomposition.

[0024] S22: Based on the month, weekday and date corresponding to each data point in the completed daily electricity consumption time series data, construct the triangular periodic features of each data point.

[0025] In this step, given that typical industry electricity consumption often contains multiple time patterns such as monthly, weekly, and daily cycles, for each data point in the daily electricity consumption time series data, a corresponding triangular periodic code is constructed using sine and cosine transforms based on its month, week number, and date. This maps the discrete month, week, and date information to a continuous periodic numerical space. This transforms the originally discrete time attributes into periodic and continuous numerical features, allowing the transitions between months, weeks, and dates to be expressed in a smooth numerical form. This facilitates the subsequent model's identification and learning of periodic patterns in the electricity consumption series.

[0026] S23: Based on the date attributes corresponding to each data point in the completed daily electricity consumption time series data, construct the external calendar features corresponding to each data point.

[0027] The date attribute includes at least one of the following: weekend, public holiday, and weekday.

[0028] In this step, the date corresponding to each data point is classified into at least one category: weekend, public holiday, or weekday by attribute discrimination. The category information is then converted into numerical features using a preset encoding format. This allows the model to explicitly perceive the social production and work schedule at each historical moment, enhancing the model's ability to respond to events such as sudden load drops during holidays and low electricity consumption on weekends, thereby improving the adaptability of the prediction results to actual electricity consumption scenarios.

[0029] S24: The completed daily electricity consumption time series data, the triangular periodic features of each data point and its external calendar features are spliced ​​together to obtain the fusion feature vector sequence of the target typical industry in the preset historical time period.

[0030] In this step, for each data point within a preset historical time period, the daily electricity consumption value, triangular periodic feature, and external calendar feature corresponding to that data point are sequentially concatenated, so that the electricity consumption record of each data point carries a joint representation of three types of information: electricity consumption amplitude, periodic time position, and social production and work status. By performing the above concatenation operation on all data points within the preset historical time period, a fused feature vector sequence of the target typical industry within the preset historical time period is obtained. Each fused feature vector contains a structured combination of electricity consumption value and time context information, providing an information-rich input foundation for subsequent modal decoupling.

[0031] S25: Normalize the fused feature vector sequence to generate a time-series feature sequence of the target typical industry within a preset historical time period.

[0032] In this step, because the features of different dimensions in the fused feature vector sequence differ in numerical units and amplitude ranges, directly inputting them into the model would affect the convergence speed and stability of the training process. Therefore, a preset normalization method is used to transform each dimension of the features to the same numerical range, resulting in a temporal feature sequence. After normalization, the features of each dimension are comparable in numerical range, thus providing a uniform-scale input representation for subsequent modal decoupling and deep modeling.

[0033] S30: Perform modal decoupling on the time-series feature sequence to generate multiple intrinsic mode components.

[0034] In this step, modal decoupling is performed on the time-series feature sequence, separating the original composite time-series feature sequence according to its inherent different frequency scales to reduce the non-stationarity of the original sequence. After modal decoupling, the mutually coupled trend components, periodic fluctuation components, and random disturbance components in the original sequence are decomposed into different intrinsic modal components. Each intrinsic modal component has its own center frequency in the frequency domain and exhibits relatively stable fluctuation characteristics. The resulting multiple intrinsic modal components characterize the evolution of the original electricity consumption sequence at different time scales, facilitating subsequent refined modeling of each component.

[0035] In one embodiment of this application, a specific intrinsic mode component generation scheme is provided. In S30, that is, performing mode decoupling processing on the temporal feature sequence to generate multiple intrinsic mode components, the following steps S31-S38 are specifically included: S31: Obtain the combination of parameters to be optimized in variational mode decomposition.

[0036] The combination of parameters to be optimized includes the number of mode decompositions and the penalty factor.

[0037] In this step, since the electricity consumption sequences of different typical industries have different frequency domain characteristics, fixed combinations of manual experience parameters are difficult to adapt to the data distribution of all industries. Therefore, the combination of parameters to be optimized by variational mode decomposition is obtained as the optimization object for subsequent automatic search of the optimal parameter combination applicable to the target typical industry using optimization algorithms.

[0038] Optionally, the combination of parameters to be optimized includes the number of mode decompositions and a penalty factor. The number of mode decompositions controls the number of intrinsic mode components after variational mode decomposition. Its value directly affects the fineness of the decomposition; if the value is too small, components at different frequency scales will not be effectively separated; if the value is too large, spurious modes may be introduced or over-decomposition may occur. The penalty factor controls the fidelity constraint strength during variational mode decomposition. Its value affects the bandwidth characteristics of each intrinsic mode component. If the value is too large, each intrinsic mode component will be too narrow, thus losing useful information; if the value is too small, mode aliasing or noise residue may occur.

[0039] S32: The population of the improved spotted kingfisher optimization algorithm is initialized using SPM chaotic mapping to generate multiple initial individuals.

[0040] Each initial individual corresponds to a set of parameters to be optimized.

[0041] In this step, the SPM (Sine-Piecewise Linear Chaotic Map) chaotic map is a one-dimensional chaotic map based on a composite design of the Sine map and the piecewise linear chaotic map. The chaotic sequence it generates iteratively exhibits good ergodicity and randomness. During population initialization, the SPM chaotic map is used to generate a uniformly distributed chaotic sequence in the search space, and each component in the chaotic sequence is mapped to the range of values ​​for the number of mode decompositions and the penalty factor, thereby generating multiple initial individuals. Each initial individual is a set of determined combinations of the number of mode decompositions and the penalty factor.

[0042] In practical applications, for each typical industry, the number of modes to be optimized in variational mode decomposition and the penalty factor are combined into a parameter vector to be optimized. The iterative equation of the SPM chaotic mapping is: ; in, Let g be the chaotic state variable at the g-th iteration, with a value range of [0, 1). The next chaotic state variable is obtained after SPM mapping; g is the iteration step; Let η be the segmented threshold parameter, η∈(0,0.5); is the sinusoidal disturbance intensity parameter, μ∈(0,1); This is a random phase perturbation term.

[0043] The chaotic state variables are mapped to the range of values ​​of the parameters to be optimized using a mapping formula, generating multiple initial individuals. The mapping formula is: ; in, Let be the initial position of the nth individual in the dth dimension; This is the lower bound for the d-th parameter to be optimized; This is the upper bound of the d-th parameter to be optimized; Let be the chaotic state variable of the nth individual in the dth dimension; n is the index number of the individual in the population; d is the dimension index, d=1, 2, which correspond to the two parameters to be optimized: the number of mode decompositions and the penalty factor, respectively.

[0044] By employing the above methods, the initial population can cover different regions of the search space as much as possible, avoiding premature aggregation of individuals in local areas during the initialization phase, thus providing a good initial distribution foundation for subsequent iterative searches.

[0045] S33: Perform variational mode decomposition based on the combination of parameters to be optimized for each individual in the current population, and calculate the joint fitness function value of permutation entropy and mode orthogonality.

[0046] In this step, for each individual in the current population, variational mode decomposition is performed on the temporal feature sequence using its corresponding number of mode decompositions and penalty factor as parameters, resulting in a set of intrinsic mode components (EMFs). Subsequently, permutation entropy and mode orthogonality are calculated based on each EMF obtained from the decomposition, and the weighted sum of these two values ​​yields the joint fitness function value for that individual. Permutation entropy measures the complexity and randomness of each EMF; a lower permutation entropy indicates purer and more regular mode components. Mode orthogonality measures the degree of spectral overlap between different EMFs; stronger orthogonality indicates more thorough mode separation and less redundant information.

[0047] In practical applications, to enhance the adaptive optimization capability of the improved kingfisher optimization algorithm for electricity consumption sequences in different industries, a joint fitness function integrating permutation entropy and modal orthogonality is constructed. For each intrinsic modal component obtained under each set of candidate parameter combinations, its normalized permutation entropy value is calculated to measure the temporal regularity and complexity of each modal component. Simultaneously, the orthogonality index between modal components is calculated to evaluate the degree of spectral overlap between modes. Specifically, the closer the permutation entropy value is to 0, the stronger the regularity of the modal components and the easier they are to predict; the lower the orthogonality index, the more thorough the mode separation.

[0048] Based on this, for each candidate parameter combination, the normalized permutation entropy average of each intrinsic modal component corresponding to the parameter combination is first calculated. Then, the average permutation entropy is weighted and summed with the modal orthogonality index according to preset weights. The result is used as the joint fitness function value of the parameter combination. The preset weight coefficients range from 0 to 1; the smaller the joint fitness function value, the better the modal decomposition effect.

[0049] To avoid the population getting trapped in local optima during the iteration process, an adaptive t-distribution perturbation strategy is introduced. The degrees of freedom parameters of the t-distribution are adaptively calculated based on the ratio of the current iteration number to the maximum iteration number. Let the current iteration number be g, and the maximum iteration number be T. max Then the adaptive degree of freedom parameter v g Calculate using the following formula: ; in, Here, represents the adaptive degree of freedom parameter at the g-th iteration; g is the current iteration number. This represents the maximum number of iterations.

[0050] In the early stages of iteration, the smaller degrees of freedom result in larger perturbation amplitudes, which helps to enhance global search capabilities and effectively escape local optima. As iteration progresses, the degrees of freedom gradually increase, and the perturbation amplitude decreases accordingly, thereby improving the accuracy of local exploration.

[0051] In each iteration, a random perturbation vector following an adaptive t-distribution is applied to the best individual in the current population. This perturbation vector is multiplied element-wise by the position of the best individual and then superimposed onto the original position to generate a new perturbed individual. ; in, The new individual position after the perturbation is applied; This represents the optimal position of an individual in the current population. To obey the degree of freedom The random perturbation vector generated by the t-distribution; represents the adaptive degree of freedom parameter at the g-th iteration.

[0052] Meanwhile, a fixed-capacity elite memory is constructed to store the individuals with the best fitness in each iteration; in each iteration, elite individuals are randomly selected from the elite memory to guide the current individual's position update, so that the population gradually converges to the high-quality solution region.

[0053] After the iteration terminates, the number of mode decompositions and the penalty factor corresponding to the individual with the smallest joint fitness function value in the current population are determined as the optimal combination of variational mode decomposition parameters for this typical industry.

[0054] In this embodiment, the hyperparameters of the improved kingfisher optimization algorithm are set as follows: population size is set to 15, elite pool capacity is set to 5, and maximum number of iterations is set to 20. The elite guidance coefficient is not a fixed constant, but is adaptively generated within the interval (0, 1) based on the ratio of the current individual fitness to the optimal individual fitness. The population size determines the search breadth of the algorithm, the elite pool capacity is used to retain good solutions from previous iterations, the elite guidance coefficient is used to achieve a balance between convergence speed and search stability, and the maximum number of iterations is used to control the computational load of the optimization process.

[0055] The above method enables a comprehensive evaluation of the decomposition effect of each parameter combination from two dimensions: modal purity and modal separation, thereby providing a fitness evaluation basis for subsequent population iteration and updates.

[0056] S34: Based on the joint fitness function value, select elite individuals from the current population and store them in the elite memory bank.

[0057] In this step, individuals are sorted in ascending order of their joint fitness function values, and a predetermined number of individuals with the best fitness values ​​are selected as elite individuals. These elite individuals and their corresponding fitness values ​​are then stored in the elite memory. The elite memory is used to store historical individuals with excellent fitness performance in previous iterations. Its size is fixed. When new elite individuals are added, if the memory is full, individuals with relatively poor fitness are removed to maintain the representativeness of the elite population.

[0058] By employing the above methods, the information of excellent solutions during the iteration process can be effectively preserved, providing a directional reference for subsequent elite individuals to guide population updates, which helps to accelerate the convergence speed of the algorithm and avoid the degradation of the search direction.

[0059] S35: In each iteration, the position update of the current individual is guided by the elite individuals in the elite memory, and an adaptive t-distribution perturbation is applied to the updated individual.

[0060] In this step, in each iteration, an elite individual is randomly selected from the elite memory. Using the position of this elite individual as a reference, a position update operation is performed on every individual in the current population, moving them closer to the high-quality solution region where the elite individual is located, thus guiding the population to gradually converge towards the global optimum. After the position update, an adaptive t-distribution perturbation is applied to each updated individual. The degrees of freedom of the adaptive t-distribution are dynamically adjusted with the number of iterations. In the early stages of iteration, the degrees of freedom are small, and the perturbation amplitude is large, which helps enhance the population's global exploration ability to escape local optima. In the later stages of iteration, the degrees of freedom gradually increase, and the perturbation amplitude decreases accordingly, which helps enhance the local exploitation ability to finely search for the optimal solution. Through the synergistic effect of the elite guidance and adaptive perturbation, a dynamic balance can be achieved between global exploration and local exploitation, improving the algorithm's convergence efficiency and optimization accuracy.

[0061] S36: Perform boundary constraint processing on the individuals after position update, and round down the number of modal decompositions.

[0062] In this step, for each individual after position update, it is checked whether its modality decomposition number and penalty factor exceed the preset search space boundary. If the value of a certain dimension exceeds the upper bound, it is truncated to the upper bound value; if it is below the lower bound, it is truncated to the lower bound value. This ensures that all individuals are within the pre-defined feasible search space and avoids generating invalid or meaningless parameter combinations during the search process. Since the modality decomposition number must be a positive integer in practical applications, after completing the boundary constraints, a rounding operation is further performed on the dimension corresponding to the modality decomposition number, rounding it to the nearest integer to ensure the rationality of this parameter in subsequent variational modality decomposition.

[0063] By employing the above methods, the legality and rationality of individual parameters can be effectively maintained during the iteration process, thereby ensuring the smooth execution of variational mode decomposition and the reliability of the optimization results.

[0064] S37: Iterate until the preset termination condition is met, and determine the parameter combination corresponding to the individual with the smallest joint fitness function value in the current population as the optimal parameter combination.

[0065] In this step, after each iteration completes population update and boundary constraint processing, it is determined whether the current iteration meets the termination condition (such as reaching the maximum number of iterations or the optimal fitness function value no longer decreasing within a preset number of generations). If the current iteration meets the termination condition, the iteration process is terminated, and the parameter combination corresponding to the individual with the smallest joint fitness function value in the current population is determined as the optimal parameter combination. Since a smaller joint fitness function value indicates a better variational mode decomposition effect, the number of mode decompositions and the penalty factor corresponding to the individual with the smallest fitness value are the optimal parameter combination for the electricity consumption sequence of the target typical industry.

[0066] To address the issues of traditional variational mode decomposition (VMD) relying on manually preset parameters and easily causing mode aliasing, this application improves the Kingfisher optimization algorithm by using SPM chaotic mapping to initialize the population, adaptive t-distribution perturbation, and an elite memory strategy. It constructs a joint fitness function based on permutation entropy and mode orthogonality to achieve adaptive search for optimal parameters for VMD in various industries. This accurately decomposes non-stationary electricity consumption sequences into physically meaningful stationary modal components such as trend terms, seasonal cycle terms, and high-frequency noise terms, thereby reducing sequence non-stationarity at its source and eliminating human intervention.

[0067] S38: Using the optimal parameter combination as control parameters, perform variational mode decomposition on the time series feature sequence to obtain multiple intrinsic mode components.

[0068] In this step, the number of mode decompositions and the penalty factor in the optimal parameter combination are used as control parameters for variational mode decomposition. Variational mode decomposition is performed on the time-series feature sequence, separating the original composite time-series feature sequence into multiple intrinsic mode components according to their inherent different frequency scales. After the above decomposition process, the multiple intrinsic mode components are arranged sequentially from low frequency to high frequency in the frequency domain, corresponding to the trend component, periodic component, and noise component in the original time-series feature sequence, respectively, and each intrinsic mode component exhibits relatively stable fluctuation characteristics.

[0069] The above method enables the variational mode decomposition parameters to be specifically adapted to the electricity consumption series of typical target industries, providing a decomposition basis with a clear structure and explicit physical meaning for the subsequent independent modeling of each intrinsic mode component.

[0070] In one embodiment of this application, a specific variational mode decomposition scheme is provided. In S38, the optimal parameter combination is used as the control parameter to perform variational mode decomposition on the time-series feature sequence to obtain multiple intrinsic mode components. Specifically, it includes the following steps S381-S383: S381: Based on the optimal number of mode decompositions and the optimal penalty factor in the optimal parameter combination, a constrained variational model of variational mode decomposition is constructed, and Lagrange multipliers are introduced to transform the constrained variational model into an unconstrained optimization problem.

[0071] In this step, the optimal number of mode decompositions is used as the total number of intrinsic mode components to be decomposed, the optimal penalty factor is used as the coefficient of the quadratic penalty term, and the time-series feature sequence is used as the input signal to construct a constrained variational model. The goal of the constrained variational model is to find the set of modes that minimizes the sum of the estimated bandwidths of all intrinsic mode components under constraints, namely, the sum of all intrinsic mode components equals the original input signal. Subsequently, Lagrange multipliers and quadratic penalty terms are introduced based on the constrained variational model, transforming the original constrained optimization problem into an augmented Lagrange form, making the alternating updating of each intrinsic mode component, center frequency, and Lagrange multiplier feasible in subsequent iterative solutions. The construction of the constrained variational model and the introduction of Lagrange multipliers provide a mathematical foundation for the subsequent iterative solution of unconstrained optimization problems using the alternating direction multiplier method.

[0072] S382: In the frequency domain, the time-series feature sequence is iteratively decomposed by the alternating direction multiplier method. In each iteration, each intrinsic mode component and its corresponding center frequency are alternately updated until the preset convergence condition is met.

[0073] In this step, the alternating direction multiplier method decomposes the original problem into multiple subproblems and solves them alternately. In the frequency domain, each intrinsic mode component, its corresponding center frequency, and the Lagrange multiplier are updated alternately. The updates for each intrinsic mode component are achieved through Wiener filtering, the updates for each center frequency are achieved through weighted averaging, and the updates for each Lagrange multiplier are achieved through gradient ascent. In each iteration, the above three update steps are executed sequentially. After each complete iteration, it is determined whether a preset convergence condition is met: the relative difference between the intrinsic mode components obtained in the current iteration and those obtained in the previous iteration is less than a preset tolerance, or the number of iterations reaches a preset maximum number of iterations. When the preset convergence condition is met, the iteration terminates, and the intrinsic mode components obtained in the current iteration are output as the final decomposition result of the variational mode decomposition.

[0074] The above method can efficiently separate time-series feature sequences into multiple intrinsic mode components with different center frequencies, and each component has limited bandwidth characteristics, thereby achieving accurate decoupling of the original sequence at different frequency scales.

[0075] S383: Output the intrinsic modal components obtained after iterative convergence as the modal decoupling result.

[0076] In this step, after determining that each intrinsic mode component satisfies the preset convergence condition, the intrinsic mode components obtained after iterative convergence are output as the modal decoupling result. The output intrinsic mode components are arranged sequentially from low frequency to high frequency in the frequency domain, corresponding to different physical components in the original time-series characteristic sequence. The low-frequency components represent the long-term evolution trend of electricity consumption, the mid-frequency components represent periodic fluctuations caused by production schedules, and the high-frequency components mainly carry random disturbances and statistical noise.

[0077] The above method enables clear separation of trend information, periodic information and noise information in the original time series feature sequence at the modal level. Each intrinsic modal component has a limited bandwidth and a different center frequency, thus providing a structurally clear and physically meaningful decomposition basis for subsequent independent modeling and prediction of each component.

[0078] S40: Based on the preset memory length, construct a multi-track differential sliding window sequence for each intrinsic mode component.

[0079] In this step, by performing differential transformation on the original modal sequence from different time interval dimensions within a time window of preset memory length, a multi-track differential sliding window sequence is constructed to separate the short-term fluctuation information and periodic change information in each intrinsic modal component. This allows each intrinsic modal component to present a clearer and more predictable pattern in the time dimension, thereby providing a more easily learned feature representation for subsequent deep modeling.

[0080] Optionally, the preset memory length refers to the number of days included in the historical time window used to construct a single prediction sample. Its value range is set according to the cycle length corresponding to the prediction task and the electricity consumption pattern changes of the target typical industry. For example, using 14 days as the preset memory length can effectively cover two complete weekly cycles, thereby effectively capturing the regular fluctuations in industry electricity consumption between weekdays and weekends. Through a 14-day historical sliding window, daily and inter-weekly differences are constructed as dual-track time-series features for each modal component. The daily difference characterizes short-term production fluctuations and inter-day continuity changes, while the weekly difference captures weekly cycle patterns and holiday difference effects. Together, they effectively remove the inherent cyclical trends and local drift components of the sequence. Furthermore, external calendar features are integrated, explicitly introducing time attribute information such as weekdays, holidays, and workdays to enhance the model's perception and response capabilities to industry production cycles and sudden fluctuation events.

[0081] In one embodiment of this application, a specific scheme for constructing a multi-track differential sliding window sequence is provided. In S40, that is, based on a preset memory length, a multi-track differential sliding window sequence for each intrinsic mode component is constructed, specifically including the following steps S41-S45: S41: For any intrinsic mode component, a time window of a preset memory length is extracted forward from the preset prediction time to obtain the historical sequence fragment of the intrinsic mode component within the time window.

[0082] In this step, for each intrinsic mode component, a time window of a preset memory length is truncated forward from the preset prediction time, so that each intrinsic mode component obtains a continuous historical sequence segment before the preset prediction time. This historical sequence segment contains the mode component values ​​for several days before the preset prediction time, providing basic data for subsequently constructing differential sliding window sequences from different time interval dimensions, thereby ensuring the temporal alignment and structural integrity of the prediction input information.

[0083] S42: Calculate the first difference value of each data pair separated by a first preset interval in the historical sequence segment, and arrange them in chronological order to obtain the first difference orbit sequence.

[0084] The first preset interval is one day.

[0085] In this step, the first preset interval is a one-day time span. For each pair of adjacent data points separated by one day in the historical sequence segment, the modal component value of the subsequent data point is subtracted from the modal component value of the preceding data point to obtain the first difference value reflecting the amplitude of the change between adjacent days. Subsequently, the first difference values ​​corresponding to all adjacent day data pairs in the historical sequence segment are arranged in chronological order to form the first difference orbit sequence.

[0086] By using the above methods, the long-term trend accumulation effect in the intrinsic mode components can be effectively removed, so that the short-term fluctuation information presents a more stable distribution characteristic in the difference domain, thereby providing the subsequent prediction model with an input representation that focuses on short-term local changes.

[0087] S43: Calculate the second difference value of each data in the historical sequence segment that is separated by a second preset interval, and arrange them in chronological order to obtain the second difference orbit sequence.

[0088] The second preset interval is one week.

[0089] In this step, the second preset interval is a one-week time span. For each pair of corresponding data points separated by seven days in the historical sequence segment, the modal component value of the subsequent data point is subtracted from the modal component value of the previous data point to obtain a second difference value reflecting the magnitude of change between positions within the same week. Subsequently, the second difference values ​​corresponding to all data pairs separated by seven days in the historical sequence segment are arranged in chronological order to form a second difference orbital sequence. Through the above operations, the periodic variation information caused by the weekly production schedule in the intrinsic modal components can be effectively extracted, highlighting the differences in electricity consumption patterns before and after weekdays and holidays, thereby providing an input representation for subsequent prediction models that focuses on periodic patterns.

[0090] The above methods can effectively characterize the regular fluctuations caused by the periodicity in the intrinsic modal components, highlighting the differences in electricity consumption between weekdays and weekends, and between holidays and non-holidays, and providing more periodic-identifiable change inputs for the prediction model.

[0091] S44: Align and combine the first and second differential orbit sequences in chronological order to generate dual-track differential feature vectors corresponding to each data point in the historical sequence segment.

[0092] In this step, for each data point covered by the historical sequence segment, the first difference value and the second difference value belonging to the same data point are combined to form a dual-track difference feature vector indexed by that data point. The dual-track difference feature vector contains features in two dimensions: short-term intraday difference information and weekly difference information for that data point.

[0093] S45: Based on the dual-track differential feature vectors of all data points within the time window, construct the multi-track differential sliding window sequence corresponding to the intrinsic mode components.

[0094] In this step, the dual-track differential feature vectors of all data points within the time window are arranged sequentially according to time, forming a complete differential feature sequence of the intrinsic mode components within a preset memory length time window, i.e., a multi-track differential sliding window sequence. This multi-track differential sliding window sequence integrates the short-term fluctuation information of each intrinsic mode component on adjacent daily scales and the periodic variation information on weekly scales, providing a joint input representation for subsequent prediction models that simultaneously includes local short-term variation patterns and weekly periodic variation patterns.

[0095] In practical applications, for each intrinsic mode component (EMC), the difference between two adjacent days is first calculated as the daily difference orbit reflecting short-term fluctuations. Simultaneously, the difference between the EMC and the same day a week prior is calculated as the weekly difference orbit reflecting the weekly cycle. Based on this, with the day before the prediction date as the end of the sliding window, a 14-day historical window is extracted. The daily difference, weekly difference, and the corresponding triangular cycle feature and external calendar feature for each day within the window are concatenated chronologically to form a multi-dimensional feature vector, which serves as the input sample for that EMC at the corresponding sliding window time. Through this method, each EMC can obtain a dual-track difference feature sequence integrating short-term changes, cycle differences, and temporal attribute information at each historical sliding window position, which is then used as input for the subsequent prediction model.

[0096] By constructing a dual-track differential sliding window, the pre-set hybrid prediction model can simultaneously learn the local variation patterns between consecutive days and the periodic difference patterns between the same week positions, thereby better adapting to the diverse electricity consumption characteristics of typical industries, such as concentrated production on weekdays, significant fluctuations on weekends or holidays, and the sensitivity of the charging and swapping service industry to holidays.

[0097] S50: Input the multi-track differential sliding window sequence of each intrinsic mode component into the preset hybrid prediction model corresponding to each intrinsic mode component to obtain the differential prediction result of each intrinsic mode component.

[0098] In this step, each intrinsic mode component (EMF) corresponds to an independently constructed pre-defined hybrid prediction model. Each pre-defined hybrid prediction model takes the multi-track differential sliding window sequence of the corresponding mode component as input, and outputs the differential prediction result of that mode component at a pre-defined prediction time by mining the time dependency within the sequence. Since each EMF differs in frequency characteristics and fluctuation patterns, configuring an independent pre-defined hybrid prediction model for each mode component allows each model to learn the evolution patterns of its own mode component in a targeted manner, thereby avoiding mutual interference between different frequency characteristics and improving the overall prediction accuracy and stability.

[0099] In one embodiment of this application, a specific scheme for obtaining differential prediction results is provided. In S50, the multi-track differential sliding window sequences of each intrinsic mode component are respectively input into the preset hybrid prediction model corresponding to each intrinsic mode component to obtain the differential prediction results of each intrinsic mode component. Specifically, it includes the following steps S51-S58: S51: Construct a temporal convolutional-linear recurrent hybrid model based on attention enhancement mechanism.

[0100] In this step, a temporal convolutional-linear recurrent hybrid model based on an attention enhancement mechanism is constructed. The hybrid model consists of temporal convolutional units, attention enhancement units, and linear recurrent units connected sequentially. The temporal convolutional units extract local temporal features from the input sequence; the attention enhancement units filter key information and adaptively weight channels within these local temporal features; the linear recurrent units model long-distance temporal dependencies based on the attention-enhanced features; and the fully connected layers fuse external calendar features and output the differential prediction results. These four components work together to form a multimodal differential prediction architecture that balances local sensitivity and global stability. The cascaded structure of temporal convolution and linear recurrent allows the hybrid model to address both local pattern capture and long-term periodic tracking needs. Simultaneously, the attention enhancement mechanism adaptively adjusts the importance of different feature channels, enhancing the model's ability to focus on key temporal information. This provides a modeling architecture for subsequent modal differential prediction that combines multi-scale perception and key information selection capabilities.

[0101] S52: Obtain the differential labels corresponding to the multi-track differential sliding window sequences of each intrinsic mode component.

[0102] In this step, the difference label refers to the actual difference value at the corresponding position in the historical sequence of the intrinsic mode components, corresponding to the multi-track difference sliding window sequence. This value is calculated by the actual values ​​of the mode components at each data point according to the same rules as those used to construct the difference track sequence. For each intrinsic mode component, the difference label corresponding to its multi-track difference sliding window sequence is obtained. The difference label includes a first difference label corresponding to the first difference track sequence and a second difference label corresponding to the second difference track sequence.

[0103] The above method provides a supervision signal for the training of each preset hybrid prediction model, enabling the model to learn parameters with the optimization objective of minimizing the deviation between the predicted difference value and the true difference value, thereby improving the accuracy of the differential prediction of each modal component.

[0104] S53: Based on the multi-track differential sliding window sequence of each intrinsic modal component and its corresponding differential label, the temporal convolution-linear recurrent hybrid model corresponding to each intrinsic modal component is independently trained to obtain the preset hybrid prediction model corresponding to each intrinsic modal component.

[0105] In this step, the multi-track differential sliding window sequence of each intrinsic mode component (EMF) is used as input, and the differential label corresponding to that mode component is used as the supervision signal. The optimization objective is to minimize the deviation between the predicted differential value and the differential label. A backpropagation algorithm is used to independently update the parameters of the hybrid model corresponding to each mode component. This allows each hybrid model to adapt its parameters to the frequency characteristics and fluctuation patterns of its respective mode component, avoiding mutual interference between the feature differences of different mode components during the training process. This results in a preset hybrid prediction model suitable for differential prediction of each mode component. After training, each EMF corresponds to a preset hybrid prediction model with optimized parameters, which is used for subsequent modal differential prediction.

[0106] By adopting the above method, the traditional process of decomposition-overall prediction-reconstruction is abandoned. Instead, the corresponding SE-TCN-LRU sub-model is trained independently for each modal component, so that it can adapt to the differentiated evolution law of different frequency domain modes.

[0107] S54: For any intrinsic mode component, the multi-track differential sliding window sequence is input into the preset hybrid prediction model, and the local temporal features of the multi-track differential sliding window sequence are extracted through the temporal convolutional network unit to obtain the local temporal feature map.

[0108] In this step, for each intrinsic modal component, a multi-track differential sliding window sequence is input into a pre-defined hybrid prediction model. The temporal convolutional network unit in the pre-defined hybrid prediction model employs a structure combining causal convolution and dilated convolution. While ensuring temporal causality, dilated convolution expands the receptive field layer by layer, enabling the network to cover the complete historical window with a limited parameter scale. The multi-track differential sliding window sequence passes sequentially through multiple convolutional layers of the temporal convolutional network unit. Each convolutional layer performs local neighborhood aggregation on the input features through dilated convolution to extract short-term change patterns and local fluctuation features within the time window. After being processed by stacking multiple layers of dilated convolution, the temporal convolutional network unit outputs a local temporal feature map. This local temporal feature map preserves the temporal structure information of the input sequence in the time dimension, while encoding local temporal patterns at different scales in the channel dimension, providing a local feature foundation for subsequent attention enhancement and long-distance dependency modeling.

[0109] S55: The attention enhancement unit performs global compression on each channel of the local temporal feature map, and generates adaptive weights for each channel through a fully connected layer. The adaptive weights are then multiplied with the local temporal feature map channel by channel to obtain the attention enhancement feature map.

[0110] In this step, the local temporal feature map is input to the attention enhancement unit. The attention enhancement unit first performs global average pooling on the local temporal feature map along the time dimension, compressing the two-dimensional temporal information of each feature channel into a single value, obtaining a channel description vector reflecting the global response intensity of each channel. Subsequently, the channel description vector is nonlinearly transformed sequentially through a first fully connected layer, an activation function layer, and a second fully connected layer, and then normalized to generate adaptive weights between 0 and 1 for each channel, representing the importance of each feature channel under the current input. Finally, the adaptive weights are element-wise multiplied with the corresponding channels of the local temporal feature map to complete the adaptive recalibration of the feature channels. After attention enhancement, the attention-enhanced feature map, while retaining the original local temporal features, enhances the representational ability of key channels and suppresses the responses of redundant or irrelevant channels, thereby improving the model's ability to focus on important temporal patterns and providing a more refined feature representation for subsequent long-distance dependency modeling.

[0111] S56: Through linear cyclic units, long-distance temporal dependency modeling of attention-enhanced feature maps is performed in the complex domain by linear state recursion, resulting in temporal dependency feature vectors.

[0112] In this step, the attention-enhanced feature map is input into a linear recurrent unit (LRU). The LRU maps the attention-enhanced feature map to the complex domain space and sequentially performs linear state updates and output transformations in the complex domain. At each time step, the hidden state of the previous time step is linearly recursively derived from the complex domain state transition matrix, while simultaneously generating the hidden state of the current time step by combining the input features of the current time step. The output features of the current time step are then extracted via linear output projection. The eigenvalues ​​of the complex domain state transition matrix are constrained to within the unit circle to ensure the long-term stability of the recursive process. Subsequently, after completing the complex domain recursive processing of the attention-enhanced feature map for all time steps, the hidden state output by the LRU at the last time step, or the aggregated result of the outputs from all time steps, is used as the temporal dependency feature vector. This temporal dependency feature vector integrates the long-distance periodic variation information of the intrinsic modal components within a preset memory length time window with global temporal dependencies, providing a global feature representation containing complete historical periodic information for subsequent differential prediction.

[0113] S57: Obtain the external calendar features corresponding to the preset prediction time, and concatenate the time-dependent feature vector with the external calendar features corresponding to the preset prediction time to obtain the fusion prediction feature vector of the intrinsic mode components.

[0114] In this step, the preset prediction time refers to the date on which electricity consumption forecasting needs to be performed, i.e., a future day corresponding to the model output. External calendar features include triangular periodic features and date attribute features corresponding to the preset prediction time. The triangular periodic features are generated by sine and cosine transforms of the month, weekday, and date of the prediction time. The date attribute features include at least one of weekend identifiers, public holiday identifiers, and weekday identifiers. The temporal dependency feature vector and external calendar features are sequentially concatenated, so that the fused prediction feature vector contains both the temporal dependency information of the intrinsic modal components within the historical window and the periodic position and social production and work-rest status information of the prediction time point. This provides a joint input representation that integrates historical patterns and future time context information for subsequent differential prediction.

[0115] S58: Input the fused prediction feature vector into the output fully connected layer of the preset hybrid prediction model for regression mapping, and generate the differential prediction results of the feature modal components at the preset prediction time.

[0116] The differential prediction results include the first differential prediction value and the second differential prediction value.

[0117] In this step, after obtaining the fused prediction feature vector of the intrinsic modal components, the fused prediction feature vector is input to the output fully connected layer of the preset hybrid prediction model. The output fully connected layer, as the final output layer of the preset hybrid prediction model, linearly maps the fused prediction feature vector from the feature space to the two-dimensional output space. The first dimension of the output corresponds to the first difference prediction value, representing the change of the intrinsic modal component at the preset prediction time relative to the previous day; the second dimension corresponds to the second difference prediction value, representing the periodic change of the intrinsic modal component at the preset prediction time relative to the same day a week prior. Through the regression mapping of the output fully connected layer, the preset hybrid prediction model integrates the multi-scale temporal features extracted step-by-step by the aforementioned temporal convolutional units, attention enhancement units, and linear recurrent units with the temporal context information provided by the external calendar features, ultimately transforming them into a dual-channel prediction output of the intrinsic modal component in the difference domain, providing a difference domain input basis for subsequent physical inverse difference reconstruction.

[0118] In practical applications, for each intrinsic mode component's multi-track differential sliding window sequence, it is first input into a temporal convolutional network unit. This unit uses a combination of causal convolution and dilated convolution to extract local temporal features from the sequence. The dilation factor of the dilated convolution increases layer by layer with the number of network layers, so as to gradually expand the receptive field under controllable parameter scale, cover the complete historical window, and output a local temporal feature map.

[0119] Subsequently, the local temporal feature map is input into the compressed-excitation attention unit. This unit first performs global average pooling along the time dimension on each feature channel, compressing the temporal information of each channel into a single global descriptive value. Then, a two-layer fully connected network is used to perform a nonlinear transformation on the descriptive value, and a sigmoid activation function is used to map and generate adaptive weights for each channel between 0 and 1. Finally, the obtained weights are multiplied channel-by-channel with the original local temporal feature map to obtain the attention-enhanced feature map. Through the above mechanism, the model can adaptively enhance key feature channels and suppress irrelevant noise channels according to different industry types, modal frequency characteristics, and prediction time periods.

[0120] Next, the attention-enhanced feature map is input into the linear recurrent unit (LRU). This unit performs linear state recursion in the complex domain, updating the hidden state of the previous time step using a complex diagonal state transition matrix, and generating the output of the current time step by combining the input features of the current time step. Simultaneously, the eigenvalues ​​of the state transition matrix are constrained within the unit circle to ensure numerical stability during the long sequence recursion process. After the LRU completes recursion at all historical time steps, its output sequence is pooled in the time dimension, concatenated with the triangular periodic features and external calendar features corresponding to the prediction day, and finally regressed through a fully connected layer to obtain the differential prediction result of this modality component on the prediction day.

[0121] S60: Perform physical inverse differential reconstruction on the differential prediction results of each intrinsic mode component to generate the single-mode power consumption prediction value corresponding to each intrinsic mode component.

[0122] In this step, the differential prediction results of each intrinsic mode component are physically reconstructed using inverse differential reconstruction. Based on historical electricity consumption values, the differential prediction results are inversely superimposed onto the corresponding historical reference points. This restores the differential prediction results of each intrinsic mode component from the differential domain to the original electricity consumption scale, thereby recovering the electricity consumption amplitude of each intrinsic mode component at the preset prediction time. After physical inverse differential reconstruction, the prediction results of each intrinsic mode component are all below the original electricity consumption dimension, preserving the variation patterns of each intrinsic mode component and ensuring the interpretability of the reconstruction results in a physical sense. This provides a dimensionally consistent input basis for the subsequent superposition and fusion of each intrinsic mode component.

[0123] In one embodiment of this application, a specific scheme for generating single-mode power consumption prediction values ​​is provided. In S60, that is, physical inverse differential reconstruction is performed on the differential prediction results of each intrinsic mode component to generate single-mode power consumption prediction values ​​corresponding to each intrinsic mode component, specifically including the following steps S61-S64: S61: Obtain the historical reference power consumption value corresponding to each intrinsic mode component at the preset prediction time.

[0124] The historical reference electricity consumption values ​​include the first historical electricity consumption value of the day before the preset forecast time and the second historical electricity consumption value of the same day of the week before the preset forecast time.

[0125] In this step, for each intrinsic mode component, based on the preset prediction time, one day and one week are traced back respectively, and the first historical electricity consumption value corresponding to the day before the preset prediction time and the second historical electricity consumption value corresponding to the same day of the week before the preset prediction time are extracted from the historical sequence of the intrinsic mode component. These are used as the historical reference benchmarks for subsequent inverse differential reconstruction of daily and weekly differential orbits.

[0126] S62: Determine the first restored electricity consumption value based on the sum of the first historical electricity consumption value and the first differential prediction value of each intrinsic mode component on the day before the preset prediction time.

[0127] In this step, for any intrinsic mode component, the first historical power consumption value is used as the reference point for the inverse differential reconstruction of the daily differential orbit, and added to the first differential prediction value to obtain the first restored power consumption value of the intrinsic mode component based on the daily differential orbit.

[0128] In this way, the predicted values ​​representing the changes between adjacent days in the difference domain are reversed back to the original electricity consumption scale, so that the first restored electricity consumption value not only retains the actual electricity consumption level of the intrinsic mode components on the day before the prediction, but also incorporates the short-term fluctuation prediction information characterized by the daily difference orbit.

[0129] S63: Determine the second restored electricity consumption value based on the sum of the second historical electricity consumption value and the second differential prediction value of each intrinsic mode component on the same day one week before the preset prediction time.

[0130] In this step, for any intrinsic mode component, the second historical power consumption value is used as the reference point for the inverse differential reconstruction of the cyclic differential orbit, and added to the second differential prediction value to obtain the second restored power consumption value of the intrinsic mode component based on the cyclic differential orbit.

[0131] In this way, the predicted value representing the change during the week in the difference domain is reversed back to the original electricity consumption scale, so that the second restored electricity consumption value not only retains the actual electricity consumption level of the intrinsic mode components on the same day of the week before the prediction, but also incorporates the periodic change prediction information characterized by the weekly difference orbit.

[0132] S64: Based on the preset fusion weights, the first restored electricity consumption value and the second restored electricity consumption value are weighted and summed to generate the single-mode electricity consumption prediction value corresponding to each intrinsic mode component.

[0133] In this step, the preset fusion weights include a first weight corresponding to the first restored power consumption value and a second weight corresponding to the second restored power consumption value. The first restored power consumption value is multiplied by the first weight, and the second restored power consumption value is multiplied by the second weight. The products of the two are added together to obtain the single-mode power consumption prediction value of the intrinsic mode component at the preset prediction time.

[0134] Optionally, preset fusion weights are used to adjust the contribution of daily and weekly differential orbits in inverse differential reconstruction. Their values ​​can be preset according to the electricity consumption pattern characteristics of typical target industries. The sum of the first weight and the second weight is 1.

[0135] In practical applications, for each intrinsic mode component, the daily difference prediction value output by the preset hybrid prediction model is superimposed on the actual mode value of the day before the prediction date to obtain the first restored electricity consumption value based on the daily difference orbit. Simultaneously, the weekly difference prediction value output by the preset hybrid prediction model is superimposed on the actual mode value of the same day of the week before the prediction date to obtain the second restored electricity consumption value based on the weekly difference orbit. Subsequently, based on the electricity consumption fluctuation characteristics of the target typical industry, a preset fusion weighting coefficient is used to weight and sum the two restored values ​​to obtain the single-mode electricity consumption prediction value for that mode component on the prediction date. Specifically, for smelting industries with stable production rhythms and significant weekly cycles, the weight of the weekly difference orbit in the fusion is increased; for industries such as charging and swapping services with strong daily fluctuations and sensitivity to short-term changes, the weight of the daily difference orbit is increased to achieve the optimal reconstruction effect under different industry scenarios. After completing the independent reconstruction of all mode components, the single-mode electricity consumption prediction values ​​of each intrinsic mode component on the same prediction date are directly added together to obtain the final daily electricity consumption prediction value for the target typical industry. For multi-day continuous forecasting scenarios, the forecast value of the previous day is used as the historical input for the next forecasting step for recursion, and the historical reference benchmark values ​​of daily difference and weekly difference are updated simultaneously to iteratively generate the electricity consumption forecast sequence for the next multiple days.

[0136] The above method enables the single-mode electricity consumption forecast to combine the sensitivity of the daily differential orbit to short-term fluctuations with the stability of the weekly differential orbit to periodic changes, thus taking into account the information advantages of the two types of differential orbits under different electricity consumption scenarios.

[0137] S70: Superimpose single-mode electricity consumption forecasts from multiple intrinsic mode components to generate electricity consumption forecasts for typical target industries.

[0138] In this step, since each intrinsic mode component is a subsequence at different frequency scales obtained after modal decoupling of the original electricity consumption sequence, and the predicted electricity consumption values ​​of each single mode are all under the same electricity consumption dimension and have orthogonal characteristics, the single-mode electricity consumption predicted values ​​of all mode components can be directly summed to reconstruct the complete electricity consumption prediction value of the target typical industry at the preset prediction time. Through the above superposition operation, the prediction contribution of each intrinsic mode component at its respective frequency domain scale is effectively integrated, so that the final output electricity consumption prediction result not only retains the complete amplitude information of the original sequence, but also integrates the frequency-division refined prediction advantage brought by the independent modeling of each intrinsic mode component.

[0139] By using the above method, the single-mode power consumption prediction value corresponding to each mode is restored, and then the prediction results of all mode components are superimposed to obtain the final power consumption prediction value. This effectively avoids the gradual accumulation of prediction errors of each mode, while ensuring the interpretability and consistency of the reconstruction results at the physical level.

[0140] As can be seen, the above scheme first acquires daily electricity consumption time series data of typical target industries, generates time series feature sequences through feature extraction and fusion, and decomposes them into multiple intrinsic modal components using modal decoupling processing. Then, based on a preset memory length, a multi-track differential sliding window sequence for each component is constructed and input into the preset hybrid prediction model corresponding to each component to obtain differential prediction results. Finally, the final electricity consumption prediction value is generated through physical inverse differential reconstruction and superposition fusion. This process transforms the electricity consumption prediction problem of typical industries from traditional manual parameter dependence and coarse modeling of hybrid modes to a refined processing paradigm of adaptive decomposition, spatiotemporal feature enhancement, modal independent modeling, and physically interpretable reconstruction. This significantly improves the model's ability to capture non-stationary, multi-period, and highly volatile time series data of typical industries, as well as its prediction accuracy and cross-industry generalization. It effectively solves the problems of parameter selection dependence on manual methods, error accumulation in unified modal modeling, and poor cross-industry adaptability in existing technologies, significantly improving the accuracy and robustness of electricity consumption prediction for typical industries.

[0141] In one embodiment, a typical industry electricity consumption forecasting device is provided, which corresponds one-to-one with the typical industry electricity consumption forecasting methods described in the above embodiments. For example... Figure 2 As shown, the typical industry electricity consumption forecasting device 100 includes: an acquisition module 101, a first generation module 102, a second generation module 103, a construction module 104, a third generation module 105, a fourth generation module 106, and a fifth generation module 107. Detailed descriptions of each functional module are as follows: The acquisition module 101 is used to acquire time series data of daily electricity consumption of a target typical industry within a preset historical period. The first generation module 102 is used to extract features from the daily electricity consumption time series data and generate a time series feature sequence of the target typical industry in a preset historical time period. The second generation module 103 is used to perform modal decoupling processing on the temporal feature sequence to generate multiple intrinsic mode components. Module 104 is used to construct a multi-track differential sliding window sequence of each intrinsic mode component based on a preset memory length; The third generation module 105 is used to input the multi-track differential sliding window sequence of each intrinsic mode component into the preset hybrid prediction model corresponding to each intrinsic mode component, so as to obtain the differential prediction result of each intrinsic mode component. The fourth generation module 106 is used to perform physical inverse differential reconstruction on the differential prediction results of each intrinsic mode component to generate the single-mode power consumption prediction value corresponding to each intrinsic mode component. The fifth generation module 107 is used to superimpose single-mode electricity consumption prediction values ​​of multiple intrinsic mode components to generate electricity consumption prediction results for typical target industries.

[0142] In one embodiment, the first generation module 102 is specifically used for: Missing data in the time series of daily electricity consumption were filled in using linear interpolation. Based on the month, weekday and date corresponding to each data point in the completed daily electricity consumption time series data, a triangular periodic feature of each data point is constructed. Based on the date attributes corresponding to each data point in the completed daily electricity consumption time series data, an external calendar feature corresponding to each data point is constructed. The date attributes include at least one of the following: weekends, statutory holidays, and weekdays. The completed daily electricity consumption time series data, the triangular periodic features of each data point and its external calendar features are spliced ​​together to obtain the fusion feature vector sequence of the target typical industry in the preset historical time period. The fused feature vector sequence is normalized to generate a time-series feature sequence of the target typical industry within a preset historical time period.

[0143] In one embodiment, the second generation module 103 is specifically used for: Obtain the combination of parameters to be optimized in the variational mode decomposition, wherein the combination of parameters to be optimized includes the number of mode decompositions and the penalty factor; The population of the improved spotted kingfisher optimization algorithm is initialized using SPM chaotic mapping to generate multiple initial individuals, where each initial individual corresponds to a set of parameters to be optimized. Based on the combination of parameters to be optimized for each individual in the current population, perform variational mode decomposition and calculate the joint fitness function value of permutation entropy and mode orthogonality. Based on the joint fitness function value, elite individuals are selected from the current population and stored in the elite memory bank; In each iteration, the position update of the current individual is guided by the elite individuals in the elite memory, and an adaptive t-distribution perturbation is applied to the updated individual; Boundary constraints are applied to the individuals after position updates, and the number of mode decompositions is rounded down. Iterate until the preset termination condition is met, and determine the parameter combination corresponding to the individual with the smallest joint fitness function value in the current population as the optimal parameter combination; Using the optimal parameter combination as control parameters, variational mode decomposition is performed on the time-series feature sequence to obtain multiple intrinsic mode components.

[0144] In one embodiment, the second generation module 103 is further configured to: Based on the optimal number of mode decompositions and the optimal penalty factor in the optimal parameter combination, a constrained variational model of variational mode decomposition is constructed, and Lagrange multipliers are introduced to transform the constrained variational model into an unconstrained optimization problem. In the frequency domain, the time-series feature sequence is iteratively decomposed using the alternating direction multiplier method. In each iteration, each intrinsic mode component and its corresponding center frequency are alternately updated until the preset convergence condition is met. The intrinsic modal components obtained after iterative convergence are output as the modal decoupling results.

[0145] In one embodiment, the construction module 104 is specifically used for: For any intrinsic mode component, a time window of a preset memory length is extracted forward from the preset prediction time to obtain the historical sequence fragment of the intrinsic mode component within the time window; Calculate the first difference value of each data pair separated by a first preset interval in the historical sequence segment, and arrange them in chronological order to obtain the first difference orbit sequence, wherein the first preset interval is one day; Calculate the second difference value of each data point in the historical sequence segment that is separated by a second preset interval, and arrange them in chronological order to obtain the second difference orbit sequence, wherein the second preset interval is one week; The first and second differential orbit sequences are aligned and combined in chronological order to generate dual-track differential feature vectors corresponding to each data point in the historical sequence segment. Based on the dual-track differential feature vectors of all data points within the time window, a multi-track differential sliding window sequence corresponding to the intrinsic mode components is constructed.

[0146] In one embodiment, the third generation module 105 is specifically used for: Construct a temporal convolutional-linear recurrent hybrid model based on an attention enhancement mechanism; Obtain the difference labels corresponding to the multi-track difference sliding window sequences of each intrinsic mode component; Based on the multi-track differential sliding window sequence of each intrinsic modal component and its corresponding differential label, the temporal convolution-linear recurrent hybrid model based on attention enhancement mechanism corresponding to each intrinsic modal component is trained independently to obtain the preset hybrid prediction model corresponding to each intrinsic modal component. For any intrinsic mode component, the multi-track differential sliding window sequence is input into a temporal convolutional network, and local temporal features are extracted from the multi-track differential sliding window sequence through dilated convolution to obtain a local temporal feature map; The local temporal feature map is input into the attention module, and each channel of the local temporal feature map is globally compressed. Adaptive weights for each channel are generated through a fully connected layer. The adaptive weights are multiplied with the local temporal feature map channel by channel to obtain the attention-enhanced feature map. The attention-enhanced feature map is input into a linear recurrent unit, and long-distance temporal dependency modeling is performed on the attention-enhanced feature map in the complex domain through linear state recursion to obtain the temporal dependency feature vector. Obtain the external calendar features corresponding to the preset prediction time, and concatenate the time-dependent feature vector with the external calendar features corresponding to the preset prediction time to obtain the fusion prediction feature vector of the intrinsic mode components; The fused prediction feature vector is input into the output fully connected layer of the preset hybrid prediction model for regression mapping, generating differential prediction results of the feature modal components at a preset prediction time. The differential prediction results include a first differential prediction value and a second differential prediction value.

[0147] In one embodiment, the fourth generation module 106 is specifically used for: Obtain the historical reference electricity consumption values ​​of each intrinsic mode component at the preset prediction time. The historical reference electricity consumption values ​​include the first historical electricity consumption value of the day before the preset prediction time and the second historical electricity consumption value of the same day of the week before the preset prediction time. The first restored electricity consumption value is determined based on the sum of the first historical electricity consumption value and the first differential prediction value of each intrinsic mode component on the day before the preset prediction time. The second restored electricity consumption value is determined based on the sum of the second historical electricity consumption value and the second differential prediction value of each intrinsic mode component on the same day one week before the preset prediction time. Based on preset fusion weights, the first restored electricity consumption value and the second restored electricity consumption value are weighted and summed to generate the single-mode electricity consumption prediction value corresponding to each intrinsic mode component.

[0148] The typical industry electricity consumption prediction device 100 provided by this invention first acquires the daily electricity consumption time series data of the target typical industry, generates a time series feature sequence through feature extraction and fusion, and decomposes it into multiple intrinsic modal components using modal decoupling processing. Then, based on a preset memory length, a multi-track differential sliding window sequence for each component is constructed and input into the preset hybrid prediction model corresponding to each component to obtain differential prediction results. Finally, the final electricity consumption prediction value is generated through physical inverse differential reconstruction and superposition fusion. This process transforms the typical industry electricity consumption prediction problem from traditional manual parameter dependence and coarse modeling of hybrid modes to a refined processing paradigm of adaptive decomposition, spatiotemporal feature enhancement, modal independent modeling, and physically interpretable reconstruction. This significantly improves the model's ability to capture non-stationary, multi-period, and highly fluctuating time series data of typical industries, as well as its prediction accuracy and cross-industry generalization. It effectively solves the problems of parameter selection dependence on manual methods, error accumulation in unified modal modeling, and poor cross-industry adaptability in existing technologies, significantly improving the accuracy and robustness of typical industry electricity consumption prediction.

[0149] Specific limitations regarding the electricity consumption forecasting device for typical industries can be found in the limitations of the electricity consumption forecasting method for typical industries mentioned above, and will not be repeated here. Each module in the aforementioned electricity consumption forecasting device for typical industries can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of the electronic device in hardware form or independently of the processor, or stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of each module.

[0150] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described typical industry electricity consumption prediction method.

[0151] In one embodiment, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the above-described typical industry electricity consumption forecasting method.

[0152] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or electronic device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0153] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0154] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0155] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A typical industry electricity consumption forecasting method, characterized in that, include: Obtain time-series data of daily electricity consumption for typical target industries over a preset historical period; Feature extraction is performed on the daily electricity consumption time series data to generate a time series feature sequence of the target typical industry in a preset historical time period; The temporal feature sequence is subjected to modal decoupling processing to generate multiple intrinsic mode components; Based on a preset memory length, a multi-track differential sliding window sequence for each intrinsic mode component is constructed. The multi-track differential sliding window sequence of each intrinsic mode component is input into the preset hybrid prediction model corresponding to each intrinsic mode component to obtain the differential prediction result of each intrinsic mode component. Physical inverse differential reconstruction is performed on the differential prediction results of each intrinsic mode component to generate the single-mode power consumption prediction value corresponding to each intrinsic mode component; The electricity consumption prediction results for the target typical industry are generated by superimposing the single-mode electricity consumption prediction values ​​of multiple intrinsic mode components.

2. The method for predicting electricity consumption in typical industries according to claim 1, characterized in that, The step of extracting features from the daily electricity consumption time series data to generate a time series feature sequence of the target typical industry over a preset historical period specifically includes: The daily electricity consumption time series data was filled with missing data using linear interpolation. Based on the month, weekday and date corresponding to each data point in the completed daily electricity consumption time series data, a triangular periodic feature of each data point is constructed. Based on the date attributes corresponding to each data point in the completed daily electricity consumption time series data, an external calendar feature corresponding to each data point is constructed, wherein the date attributes include at least one of the following: weekends, statutory holidays and weekdays; The completed daily electricity consumption time series data, the triangular periodic features of each data point and its external calendar features are spliced ​​together to obtain the fusion feature vector sequence of the target typical industry in the preset historical time period. The fused feature vector sequence is normalized to generate the time-series feature sequence of the target typical industry within the preset historical time period.

3. The method for predicting electricity consumption in typical industries according to claim 1, characterized in that, The step of performing modal decoupling processing on the temporal feature sequence to generate multiple intrinsic modal components specifically includes: Obtain the combination of parameters to be optimized for variational mode decomposition, wherein the combination of parameters to be optimized includes the number of mode decompositions and a penalty factor; The population of the improved spotted kingfisher optimization algorithm is initialized using SPM chaotic mapping to generate multiple initial individuals, where each initial individual corresponds to a set of parameters to be optimized. Based on the combination of parameters to be optimized for each individual in the current population, perform variational mode decomposition and calculate the joint fitness function value of permutation entropy and mode orthogonality; Based on the joint fitness function value, elite individuals are selected from the current population and stored in the elite memory bank; In each iteration, the position update of the current individual is guided by the elite individuals in the elite memory, and an adaptive t-distribution perturbation is applied to the updated individual; Boundary constraints are applied to the individuals after position updates, and the number of mode decompositions is rounded down. Iterate until the preset termination condition is met, and determine the parameter combination corresponding to the individual with the smallest joint fitness function value in the current population as the optimal parameter combination; Using the optimal parameter combination as control parameters, variational mode decomposition is performed on the time-series feature sequence to obtain the multiple intrinsic mode components.

4. The method for predicting electricity consumption in typical industries according to claim 3, characterized in that, The step of using the optimal parameter combination as control parameters to perform variational mode decomposition on the time-series feature sequence to obtain the multiple intrinsic mode components specifically includes: Based on the optimal number of mode decompositions and the optimal penalty factor in the optimal parameter combination, a constrained variational model of variational mode decomposition is constructed, and Lagrange multipliers are introduced to transform the constrained variational model into an unconstrained optimization problem. In the frequency domain, the time-series feature sequence is iteratively decomposed using the alternating direction multiplier method. In each iteration, each intrinsic mode component and its corresponding center frequency are alternately updated until the preset convergence condition is met. The intrinsic modal components obtained after iterative convergence are output as the modal decoupling results.

5. The method for predicting electricity consumption in typical industries according to claim 1, characterized in that, The step of constructing a multi-track differential sliding window sequence for each intrinsic mode component based on a preset memory length specifically includes: For any intrinsic mode component, a time window of a preset memory length is extracted forward from the preset prediction time to obtain the historical sequence fragment of the intrinsic mode component within the time window; Calculate the first difference value of each data pair separated by a first preset interval in the historical sequence segment, and arrange them in chronological order to obtain the first difference orbit sequence, wherein the first preset interval is one day; Calculate the second difference value of each data point in the historical sequence segment that is separated by a second preset interval, and arrange them in chronological order to obtain the second difference orbit sequence, wherein the second preset interval is one week; The first differential orbit sequence and the second differential orbit sequence are aligned and combined in chronological order to generate dual-track differential feature vectors corresponding to each data point in the historical sequence segment; Based on the dual-track differential feature vectors of all data points within the time window, a multi-track differential sliding window sequence corresponding to the intrinsic mode components is constructed.

6. The method for predicting electricity consumption in typical industries according to claim 1, characterized in that, The step of inputting the multi-track differential sliding window sequence of each intrinsic mode component into the preset hybrid prediction model corresponding to each intrinsic mode component to obtain the differential prediction result of each intrinsic mode component specifically includes: Construct a temporal convolutional-linear recurrent hybrid model based on an attention enhancement mechanism; Obtain the difference labels corresponding to the multi-track difference sliding window sequences of each intrinsic mode component; Based on the multi-track differential sliding window sequence of each intrinsic modal component and its corresponding differential label, the temporal convolution-linear recurrent hybrid model based on attention enhancement mechanism corresponding to each intrinsic modal component is trained independently to obtain the preset hybrid prediction model corresponding to each intrinsic modal component. For any intrinsic mode component, the multi-track differential sliding window sequence is input into a temporal convolutional network, and local temporal features are extracted from the multi-track differential sliding window sequence through dilated convolution to obtain a local temporal feature map; The local temporal feature map is input into the attention module, and each channel of the local temporal feature map is globally compressed. Adaptive weights for each channel are generated through a fully connected layer. The adaptive weights are multiplied with the local temporal feature map channel by channel to obtain the attention-enhanced feature map. The attention-enhanced feature map is input into a linear recurrent unit, and long-distance temporal dependency modeling is performed on the attention-enhanced feature map in the complex domain through linear state recursion to obtain the temporal dependency feature vector. Obtain the external calendar features corresponding to the preset prediction time, and concatenate the time-dependent feature vector with the external calendar features corresponding to the preset prediction time to obtain the fusion prediction feature vector of the intrinsic mode components; The fused prediction feature vector is input to the output fully connected layer of the preset hybrid prediction model for regression mapping, generating differential prediction results of the eigenmode components at a preset prediction time, wherein the differential prediction results include a first differential prediction value and a second differential prediction value.

7. The method for predicting electricity consumption in typical industries according to claim 1, characterized in that, The step of performing physical inverse differential reconstruction on the differential prediction results of each intrinsic mode component to generate single-mode power consumption prediction values ​​corresponding to each intrinsic mode component specifically includes: Obtain historical reference power consumption values ​​for each intrinsic mode component at a preset prediction time, wherein the historical reference power consumption values ​​include a first historical power consumption value on the day before the preset prediction time and a second historical power consumption value on the same day of the week before the preset prediction time. The first restored electricity consumption value is determined based on the sum of the first historical electricity consumption value and the first differential prediction value of each intrinsic mode component on the day before the preset prediction time. The second restored electricity consumption value is determined based on the sum of the second historical electricity consumption value and the second differential prediction value of each intrinsic mode component on the same day one week before the preset prediction time. Based on preset fusion weights, the first restored power consumption value and the second restored power consumption value are weighted and summed to generate the single-mode power consumption prediction value corresponding to each intrinsic mode component.

8. A typical industry electricity consumption forecasting device, characterized in that, include: The acquisition module is used to acquire time-series data of daily electricity consumption of a typical target industry over a preset historical period. The first generation module is used to extract features from the daily electricity consumption time series data and generate the time series feature sequence of the target typical industry in a preset historical time period. The second generation module is used to perform modal decoupling processing on the temporal feature sequence to generate multiple intrinsic modal components; The module is used to construct multi-track differential sliding window sequences of each intrinsic mode component based on a preset memory length; The third generation module is used to input the multi-track differential sliding window sequence of each intrinsic mode component into the preset hybrid prediction model corresponding to each intrinsic mode component, so as to obtain the differential prediction result of each intrinsic mode component. The fourth generation module is used to perform physical inverse differential reconstruction on the differential prediction results of each intrinsic mode component to generate the single-mode power consumption prediction value corresponding to each intrinsic mode component. The fifth generation module is used to superimpose single-mode electricity consumption prediction values ​​of multiple intrinsic mode components to generate electricity consumption prediction results for the target typical industry.

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 processor executes the computer program, it implements the steps of the typical industry electricity consumption forecasting method as described in any one of claims 1 to 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 steps of the typical industry electricity consumption forecasting method as described in any one of claims 1 to 7.

Images