Power load prediction method, device and medium based on quadratic modal decomposition and double model parallelism

Abstract

The application discloses a power load prediction method and device based on secondary modal decomposition and double-model parallelism, and a medium. The method finely decomposes an original load sequence through a secondary modal decomposition method, and denoises and reconstructs in combination with a wavelet threshold method, so that a series of stable modal components are finally obtained. In the prediction stage, a parallel prediction architecture of Informer and BiLSTM is constructed, all the modal components are synchronously input, the multi-head probability sparse self-attention mechanism of Informer is used to capture global long-term dependence, and BiLSTM is used to capture local short-term dynamics. The heterogeneous features extracted from the two are spliced and nonlinearly fused to obtain a final prediction value. Compared with the prior art, the application realizes the collaborative capture and accurate prediction of the multi-scale features of the non-stationary power load sequence through the secondary decomposition of the coarse-to-fine and the double-model parallel architecture.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to a power load prediction method, device and medium based on quadratic modal decomposition and double model parallelism. BACKGROUND

[0002] With the proposal of the "double carbon" target and the promotion of global energy transformation, the power system is undergoing profound changes. The high proportion of new energy access and the diversity of user behavior make the power load present high volatility, strong randomness and nonlinear characteristics. Accurate power load prediction is a key link to ensure the safe and stable operation of the power grid, optimize dispatching decisions and reduce operating costs.

[0003] Existing power load prediction methods are mainly divided into traditional prediction methods, machine learning methods and deep learning methods. Traditional prediction methods, such as regression analysis and time series method, are mostly linear models, which are difficult to capture the complex nonlinear characteristics of modern power system load, resulting in insufficient prediction accuracy; machine learning methods, such as support vector machines and random forests, have improved in handling nonlinear data, but have limited flexibility in model construction and feature extraction capabilities.

[0004] Deep learning methods, especially recurrent neural networks RNN and long short-term memory networks LSTM, are widely used due to their strong time series feature learning capabilities. However, a single deep learning model often struggles to fully capture the complex characteristics of multi-scale and multi-frequency in power load sequences with strong randomness and volatility. For example, long sequence prediction models such as Informer are good at capturing long-term dependencies, but may not be sensitive enough to short-term dramatic fluctuations; while BiLSTM and other models can better capture local time series features, but may face information forgetting and computational efficiency issues when dealing with ultra-long sequences. In addition, the original power load sequence often contains a large amount of high-frequency noise, and directly using the original data for prediction will interfere with the model's feature learning process and affect the prediction accuracy.

[0005] After searching, Chinese patent publication CN117592593A discloses a short-term power load prediction method based on improved quadratic modal decomposition and WOA optimized BiLSTM-attention. This method performs CEEMDAN and VMD quadratic decomposition on the load sequence, and uses the whale optimization algorithm WOA to optimize the hyperparameters of VMD and BiLSTM, and finally predicts each subsequence and superimposes the results to achieve load prediction.

[0006] However, after modal decomposition, this scheme predicts each subsequence component separately and then simply superimposes the prediction results of each component to obtain the final result. This severs the inherent intrinsic relationship between modal components of different frequencies. Furthermore, using a single BiLSTM-Attention model to process all components in the final prediction stage may make it difficult for the model to learn and utilize the complex nonlinear interaction relationship between features at different scales, thereby limiting further improvement in prediction accuracy and potentially affecting the stability of the final result due to the independent accumulation of errors.

[0007] Therefore, how to effectively decompose the multi-scale characteristics of power load and simultaneously leverage the advantages of different deep learning models to achieve accurate prediction of non-stationary and noisy power load is a technical problem that needs to be solved. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a power load forecasting method based on quadratic mode decomposition and dual-model parallelism.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] According to a first aspect of the present invention, a power load forecasting method based on quadratic mode decomposition and dual-model parallelism is provided, the method comprising:

[0011] The original power load sequence is subjected to initial mode decomposition and denoising to obtain the denoised reconstructed signal;

[0012] The denoised and reconstructed signal is input into the variational mode decomposition algorithm for secondary mode decomposition to obtain a set of mode components with different center frequencies;

[0013] All modal components obtained from the secondary decomposition are synchronously and in parallel input into a hybrid prediction model consisting of an Informer model and a bidirectional long short-term memory network. The Informer model uses a multi-head probabilistic sparse self-attention mechanism to extract global long-term dependency features from all modal components, while the bidirectional long short-term memory network extracts local short-term dynamic features from all modal components.

[0014] The extracted global long-term dependency features and local short-term dynamic features are concatenated along the feature dimension to form a fused feature vector;

[0015] The fused feature vector is mapped to the final power load prediction value through a fully connected layer.

[0016] As a preferred technical solution, the initial mode decomposition and denoising specifically includes:

[0017] The original power load sequence is decomposed into a set of intrinsic mode function (IMF) components and a residual term using an adaptive noise complete empirical mode decomposition algorithm.

[0018] The high-frequency IMF components containing noise are identified, and the high-frequency IMF components are denoised using the wavelet thresholding method.

[0019] The denoised high-frequency IMF components, the unprocessed low-frequency IMF components, and the residual terms are accumulated and reconstructed to obtain the denoised reconstructed signal.

[0020] As a preferred technical solution, when using the wavelet thresholding method for denoising, the threshold λ is adaptively determined based on the noise level, and the calculation method is as follows:

[0021] ,

[0022] Where σ is the noise standard deviation, n is the signal length, and ln is the natural logarithm function.

[0023] As a preferred technical solution, the number K of modal components in the variational mode decomposition algorithm is adaptively determined by the center frequency method. Specifically, it includes: through iterative search, using the separation degree of the center frequencies of the modal components as the evaluation index, when the minimum difference between the center frequency of the newly added modal component and the center frequency of the existing modal component is lower than a preset threshold, the current K value is determined as the optimal number of decompositions.

[0024] As a preferred technical solution, extracting the global long-term dependency features based on the Informer model specifically includes:

[0025] All modal components are integrated into a multi-channel time series tensor to characterize the composite fluctuations of power load at different time scales;

[0026] By utilizing its multi-head probabilistic sparse self-attention mechanism, the key time points and modal components that contribute the most to long-term load forecasting are screened from multi-channel time series tensor information.

[0027] Based on the filtered results, a fixed-dimensional global long-term dependency feature vector is output to characterize the long-term trend and periodicity of the power load sequence.

[0028] As a preferred technical solution, the bidirectional long short-term memory network is a BiLSTM model, and the extraction of the local short-term dynamic features based on the BiLSTM model specifically includes:

[0029] All the modal components are used as parallel inputs and input to their forward and backward long short-term memory network layers respectively;

[0030] For each modal component sequence, past and future contextual information is cyclically fused through its gating mechanism to capture transient fluctuation patterns at each time point in its local context;

[0031] The bidirectional hidden states of the final time step of all modal components are aggregated to output a fixed-dimensional local short-term dynamic feature vector that characterizes the short-term fluctuations and local mutation details in the power load sequence.

[0032] As a preferred technical solution, the method further includes a multi-feature screening step before performing the first mode decomposition and denoising on the original power load sequence: using Pearson correlation analysis to evaluate the linear correlation strength between each relevant feature and the power load sequence, and screening out features whose linear correlation strength is greater than a preset threshold for subsequent model training.

[0033] As a preferred technical solution, the method further includes a dynamic iteration process, which employs a multi-step rolling prediction strategy, including:

[0034] Iterative forecasting using a sliding time window is employed to generate future load forecast sequences.

[0035] During the iterative prediction process, the hybrid prediction model is retrained or its parameters are fine-tuned using newly acquired actual load data to dynamically correct the model's prediction error.

[0036] According to a second aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.

[0037] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] 1. This invention transforms complex load sequences into a set of stable modal components through quadratic mode decomposition, providing high-quality input for subsequent models. Through the parallel design of Informer and BiLSTM, it can simultaneously and fully capture the long-term global patterns and short-term local dynamics hidden in the modal components, realizing in-depth mining and synergistic utilization of multi-scale time-series characteristics of power load, and effectively improving prediction accuracy.

[0040] 2. The coarse-to-fine secondary mode decomposition strategy proposed in this invention uses CEEMDAN combined with wavelet thresholding for initial decomposition and coarse-grained denoising, effectively removing random noise interference and providing a cleaner input signal for subsequent decomposition, thereby enhancing the stability and reliability of the features.

[0041] 3. This invention synchronously inputs all modal components into the parallel architecture of Informer and BiLSTM, and extracts features from both long-term and short-term perspectives in parallel. These heterogeneous features are then spliced ​​and nonlinearly fused at the deep network level, enabling the model to autonomously learn the complex interaction relationships between features at different scales. This avoids the cognitive bias and error accumulation problems of a single model, and effectively improves the accuracy and generalization performance of the prediction model. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the overall technical process of the present invention.

[0043] Figure 2 This is a schematic diagram of the Informer algorithm of the present invention;

[0044] Figure 3 This is a schematic diagram of the BiLSTM neural network structure of the present invention;

[0045] Figure 4 This is a diagram showing the overall structure of the Informer-BiLSTM prediction model based on quadratic mode decomposition in this invention.

[0046] Figure 5 This is a graph of the original load data of the first dataset used in this embodiment of the invention.

[0047] Figure 6 This is a graph of the original load data of the second dataset used in this embodiment of the invention;

[0048] Figure 7 This is a diagram showing the result of the first dataset being decomposed once using CEEMDAN.

[0049] Figure 8 This is a graph showing the result of the second dataset being decomposed once by CEEMDAN.

[0050] Figure 9 This is a diagram showing the result of the first dataset after undergoing secondary mode decomposition.

[0051] Figure 10 The image shows the result of the second dataset after quadratic mode decomposition.

[0052] Figure 11 This is a comparison chart of the prediction results for the first dataset under different modality decomposition strategies.

[0053] Figure 12 This is a comparison chart of the prediction results for the second dataset under different modality decomposition strategies.

[0054] Figure 13 This is a comparison chart of the prediction results of the model of this invention with other comparative models under the first dataset;

[0055] Figure 14 This is a comparison chart of the prediction results of the model of this invention with other comparative models under the second dataset. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0057] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0058] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent.

[0059] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0060] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0061] Example 1:

[0062] like Figure 1 As shown, this invention provides a power load forecasting method based on quadratic mode decomposition and dual-model parallelism. The overall technical process mainly consists of four parts: data processing, quadratic mode decomposition, parallel forecasting, and result fusion, specifically including:

[0063] Step S1: Data Acquisition and Preprocessing

[0064] Step S101: Collect historical power load data and related multidimensional feature data of the study area, such as maximum temperature, minimum temperature, relative humidity, date features, etc., and preprocess the data, including handling missing values ​​(e.g., using Lagrange interpolation) and outliers.

[0065] Step S102: Use Pearson correlation analysis to screen features, calculate the linear correlation strength between each feature data and the power load, set a threshold of 0.3, retain features with a correlation greater than the threshold as input to the model, and eliminate redundant features to reduce the computational complexity of the model and improve the prediction accuracy. For example, temperature features usually have the strongest correlation with load and should be used as the core prediction feature.

[0066] Step S2: Quadratic Mode Decomposition Based on CEEMD

[0067] This step aims to denoise and extract multi-scale features from the original power load sequence;

[0068] Step S201: The adaptive noise complete empirical mode decomposition algorithm CEEMDAN is used for initial decomposition and wavelet thresholding for denoising;

[0069] Step S2011: The original power load signal is decomposed using CEEMDAN to obtain N Intrinsic Mode Functions (IMFs) and one residual component. The high-frequency IMF components, such as IMF1-IMF4, mainly contain noise and short-term fluctuations, while the low-frequency IMFs and residual components contain long-term trends and main signal characteristics.

[0070] Step S2012: To remove high-frequency noise, the identified high-frequency IMF components are processed using wavelet thresholding technology. Wavelet transform is performed on each high-frequency IMF to obtain wavelet coefficients. A threshold function is applied to denoise the wavelet coefficients, and then the denoised IMF is reconstructed through inverse transform, retaining significant signal components.

[0071] ,

[0072] Where λ is the threshold, which is adaptively determined based on the noise level and expressed as:

[0073] ,

[0074] in, The standard deviation of noise. The signal length;

[0075] Preserving the original form of the low-frequency IMF and residual terms, the denoised high-frequency components are added to the unprocessed low-frequency components to obtain the first decomposed signal:

[0076] ,

[0077] in, This is a decomposition signal. The high-frequency IMF components after noise reduction. These are unprocessed low-frequency IMF components. For residual terms;

[0078] This operation can remove high-frequency noise while retaining low-frequency effective signals, providing a cleaner input for subsequent VMD decomposition;

[0079] Step S202: Using the Variational Mode Decomposition (VMD) algorithm, the clean signal reconstructed in S201 is input into the VMD model for secondary decomposition. VMD is a non-recursive signal processing method that can adaptively decompose a signal into multiple stationary components with definite frequency characteristics and bandwidth limitations. The specific steps for secondary feature extraction using VMD are as follows:

[0080] Step S021: Construct a variational model, and... Decomposed into K modal components Each mode Corresponding to a center frequency And bandwidth, which is determined by the range supported by the frequency domain;

[0081] Each mode It can be regarded as an amplitude-frequency modulation signal. Its frequency characteristics are analyzed by performing the Hilbert-Huang transform (HHT). To ensure the accuracy of the frequency information, exponential modulation is used to convert it to the vicinity of the fundamental frequency.

[0082] The mathematical expression for constructing the unconstrained variational problem is shown below. First, the sum of the bandwidths of each mode is minimized to achieve frequency domain localization. At the same time, the signal reconstruction constraint needs to be satisfied to ensure that the sum of the decomposed modal components is equal to the original signal.

[0083] ,

[0084] Where K is the preset total number of modal decompositions; k is the index of the modal component. ; This is the k-th modal component; Unit impact function; Together they constitute the Hilbert converter; For exponential modulation terms;

[0085] Step S2022: To transform the constrained optimization problem into an unconstrained optimization problem and obtain the modified augmented Lagrange function, Lagrange multipliers need to be introduced. and secondary penalty factor ;

[0086] When the signal contains noise or decomposition error, the second-order penalty factor It can balance bandwidth constraints and reconstruction errors, avoiding overfitting or underfitting;

[0087] Step S2023: Iteratively optimize each modal component using the Alternating Direction Multiplier Method (ADMM). Center frequency and Lagrange multipliers ;

[0088] Step S2024: After the iteration meets the convergence condition, K modal components are obtained. These modal components have the characteristics of non-overlapping frequency domain components and strong stationarity. Furthermore, different modes can correspond to components at different time scales in the original signal, which is beneficial for subsequent feature analysis.

[0089] Step S2025: By setting the number of decomposition layers K (the optimal value can be determined by the center frequency method), VMD decomposes the signal into K intrinsic mode functions. These decomposed mode components correspond to the characteristic components of different time scales in the load, such as daily cycle, weekly cycle, seasonal trend and random fluctuation. They have good spectral characteristics and stability, and are suitable as inputs for subsequent deep learning models.

[0090] Step S3: As Figure 4 As shown, the Informer-BiLSTM parallel prediction model is constructed:

[0091] Step S301: Input the low-frequency modal components (representing long-term trends and seasonal variations) obtained from VMD decomposition in step S202 into the Informer model. The Informer is an improved model based on the Transformer architecture, such as... Figure 2 As shown, its core is the multi-head probabilistic sparse self-attention mechanism, which can effectively filter key information and significantly reduce computational complexity, making it particularly good at processing long sequence inputs and capturing long-term dependencies. Through this branch, the model can accurately grasp the overall trend and periodic pattern of power load. The operation logic of the self-attention mechanism is transformed into a probabilistic expression form, and the focus and efficient processing of key information are achieved through the probabilistic formula.

[0092] ,

[0093] Where Q*, K, and V are the query matrix, key matrix, and value matrix, respectively;

[0094] Step S302: Input the high-frequency modal components (representing short-term random fluctuations and local features) obtained from VMD decomposition in step S202 into the BiLSTM model, such as... Figure 3 As shown, the BiLSTM network model consists of a forward LSTM and a backward LSTM. This bidirectional structure enables it to utilize both past and future contextual information to analyze data at the current time point. Therefore, it performs well in capturing short-term, rapidly changing local dependencies in sequences. Through this branch, the model can accurately predict short-term load fluctuations caused by sudden weather changes, user behavior, etc.

[0095] Step S4: Model Training and Result Fusion

[0096] The outputs of the two parallel branches are weighted and fused or fused through a fully connected layer to obtain the final power load forecast result;

[0097] During the model training phase, the Adam optimizer is used, with appropriate learning rate, batch size, and number of training rounds set. A seven-day rolling prediction strategy is adopted, which uses data from the past seven days to predict one or more days in the future and rolls the prediction results into the historical window to achieve long-term prediction. To control error accumulation, the model can be retrained with real load data after a certain number of iterations for correction.

[0098] The method of this invention first uses CEEMDAN combined with wavelet thresholding to perform initial decomposition and denoising of the original load sequence, and then uses VMD to perform secondary fine decomposition of the denoised signal to obtain a set of stationary mode components. Subsequently, all components are synchronously input into a parallel hybrid model composed of Informer and BiLSTM to extract global and local features respectively. Finally, the prediction results are output through feature concatenation and fusion with a fully connected layer, which significantly improves the accuracy and generalization ability of load prediction.

[0099] Example 2:

[0100] To verify the effectiveness of the method of the present invention, experiments were conducted using two sets of publicly available power load datasets with different characteristics.

[0101] Dataset 1: Daily average load data for a certain region from 2012 to 2015, totaling 1107 sets. The original load data curves are shown below. Figure 5 As shown;

[0102] Dataset 2: Hourly load data for a certain region from 2016 to 2019, totaling 128,544 sets. The original load data curves are shown below. Figure 6 As shown.

[0103] The CEEMDAN-VMD-Informer-BiLSTM model was compared with multiple benchmark models, including single models (BiLSTM, Informer) and different combinations of models (VMD-Informer, Informer-BiLSTM, etc.). Mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²) were used. 2 () as an evaluation indicator.

[0104] To verify the key role of the proposed secondary mode decomposition CEEMDAN-VMD in improving prediction accuracy, an ablation experiment was designed. First, its effectiveness was demonstrated through visualization of the decomposition results, such as... Figure 7 and Figure 8As shown, after the original sequence is first decomposed by CEEMDAN, the noise is effectively separated into the high-frequency IMF component; furthermore, as Figure 9 and Figure 10 As shown, after performing a second VMD decomposition on the denoised signal, modal components with clear frequency boundaries and significantly enhanced stationarity are obtained. Based on this, the prediction performance of the complete model CEEMD-VMD-Informer-BiLSTM of this invention is compared with the baseline model Informer-BiLSTM without modal decomposition and the model VMD-Informer-BiLSTM with only one modal decomposition. The comparison results on the two datasets are shown in Tables 1 and 2. Table 1 shows the accuracy comparison results of the three modal decomposition prediction models under dataset 1; Table 2 shows the accuracy comparison results of the three modal decomposition prediction models under dataset 2.

[0105] Table 1

[0106]

[0107] Table 2

[0108]

[0109] Introducing VMD to perform preliminary decomposition of the original data, compared to directly using the Informer-BiLSTM model, significantly reduces all error metrics, including R. 2 The performance also improved. Based on VMD, the use of CEEMDAN for pre-decomposition resulted in a quadratic mode decomposition strategy, leading to a significant performance leap. This invention employs a quadratic decomposition model, and on the first dataset, MAPE was reduced by 23.199% and 12.57%, respectively, with a goodness-of-fit R-value of [value missing]. 2 The first two datasets showed improvements of 0.125 and 0.099 respectively; on the second dataset, MAPE decreased by 26.831% and 17.805% respectively, with a goodness-of-fit R-value of [value missing]. 2 Increased by 0.163 and 0.134 respectively. For example... Figure 11 and Figure 12 As shown, the predicted curve of the model in this invention has a better fit with the true value curve. Experiments demonstrate that the coarse-to-fine decomposition strategy of quadratic mode decomposition allows it to focus more on the intrinsic laws of each component, thereby achieving a significant improvement in prediction accuracy and goodness of fit.

[0110] To verify the overall prediction performance and generalization ability of the proposed Informer-BiLSTM multi-feature power load prediction model based on quadratic mode decomposition, the model was first used to predict loads on two datasets. The prediction performance of single models BiLSTM and Informer, and combined models VMD-Informer, VMD-BiLSTM, Informer-BiLSTM, CEEMDAN-VMD-Informer, and CEEMDAN-VMD-BiLSTM were used as control groups. The prediction results are shown in Tables 3 and 4. Table 3 shows the accuracy comparison between the CEEMDAN-VMD-Informer-BiLSTM model and multiple benchmark models in dataset 1; Table 4 shows the accuracy comparison between the CEEMDAN-VMD-Informer-BiLSTM model and multiple benchmark models in dataset 2. The prediction curves are shown below. Figure 13 and Figure 14 As demonstrated, the predicted curves of the model presented in this invention fit the true value curves better: on both datasets, the CEEMDAN-VMD-Informer-BiLSTM model proposed in this invention achieved the best results. Compared to the worst-performing single BiLSTM model, on the first dataset, the MAPE of the model presented in this invention was reduced by approximately 92.8%, and R... 2 The performance was improved by approximately 24.6%, and even compared to the second-best performing CEEMD-VMD-Informer model, the MAPE of this invention was reduced by approximately 83.6%. 2 There was also a significant improvement; on the second dataset, similar to the results of the first dataset, the model of this invention had an overwhelming advantage over all control models in all four evaluation metrics, which fully demonstrates that the model proposed in this invention has strong predictive performance and good generalization ability, and can adapt to power load forecasting tasks with different data characteristics.

[0111] Table 3

[0112]

[0113] Table 4

[0114]

[0115] Example 3:

[0116] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0117] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0118] The processing unit executes the various methods and processes described above, such as methods S1 to S4. For example, in some embodiments, methods S1 to S4 may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of methods S1 to S4 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute methods S1 to S4 by any other suitable means (e.g., by means of firmware).

[0119] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0120] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0121] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0122] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0123] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this disclosure can be achieved, and this is not limited herein.

[0124] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A power load forecasting method based on quadratic mode decomposition and dual-model parallelism, characterized in that, include: The original power load sequence is subjected to initial mode decomposition and denoising to obtain the denoised reconstructed signal; The denoised and reconstructed signal is input into a variational mode decomposition algorithm for secondary mode decomposition, resulting in a set of mode components with different center frequencies. All mode components obtained from the secondary decomposition are then synchronously and in parallel input into a hybrid prediction model consisting of an Informer model and a bidirectional long short-term memory network. The Informer model uses a multi-head probabilistic sparse self-attention mechanism to extract global long-term dependency features from all mode components, while the bidirectional long short-term memory network extracts local short-term dynamic features from all mode components. The extracted global long-term dependency features and local short-term dynamic features are concatenated along the feature dimension to form a fused feature vector. This fused feature vector is then mapped to the final power load prediction value through a fully connected layer. Extracting the global long-term dependency features based on the Informer model specifically includes: All modal components are integrated into a multi-channel time series tensor to characterize the composite fluctuations of power load at different time scales; By utilizing its multi-head probabilistic sparse self-attention mechanism, the key time points and modal components that contribute the most to long-term load forecasting are screened from multi-channel time series tensor information. Based on the filtered results, a fixed-dimensional global long-term dependency feature vector is output to characterize the long-term trend and periodicity of the power load sequence. The bidirectional long short-term memory network is a BiLSTM model, and the extraction of the local short-term dynamic features based on the BiLSTM model specifically includes: All the modal components are used as parallel inputs and input to their forward and backward long short-term memory network layers respectively; For each modal component sequence, past and future contextual information is cyclically fused through its gating mechanism to capture transient fluctuation patterns at each time point in its local context; The bidirectional hidden states of the final time step of all modal components are aggregated to output a fixed-dimensional local short-term dynamic feature vector that characterizes the short-term fluctuations and local mutation details in the power load sequence.

2. The power load forecasting method based on quadratic mode decomposition and dual-model parallelism as described in claim 1, characterized in that, The initial mode decomposition and denoising specifically include: The original power load sequence is decomposed into a set of intrinsic mode function (IMF) components and a residual term using an adaptive noise complete empirical mode decomposition algorithm. The high-frequency IMF components containing noise are identified, and the high-frequency IMF components are denoised using the wavelet thresholding method. The denoised high-frequency IMF components, the unprocessed low-frequency IMF components, and the residual terms are accumulated and reconstructed to obtain the denoised reconstructed signal.

3. The power load forecasting method based on quadratic mode decomposition and dual-model parallelism as described in claim 2, characterized in that, When using the wavelet thresholding method for denoising, the threshold λ is adaptively determined based on the noise level, and is calculated as follows: ， Where σ is the noise standard deviation, n is the signal length, and ln is the natural logarithm function.

4. The power load forecasting method based on quadratic mode decomposition and dual-model parallelism as described in claim 1, characterized in that, The number of modal components K in the variational mode decomposition algorithm is adaptively determined by the center frequency method. Specifically, it includes: through iterative search, using the separation degree of the center frequencies of the modal components as the evaluation index, when the minimum difference between the center frequency of the newly added modal component and the center frequency of the existing modal component is lower than a preset threshold, the current K value is determined as the optimal number of decompositions.

5. The power load forecasting method based on quadratic mode decomposition and dual-model parallelism according to claim 1, characterized in that, Before performing the first mode decomposition and denoising on the original power load sequence, the method also includes a multi-feature screening step: using Pearson correlation analysis to evaluate the linear correlation strength between each relevant feature and the power load sequence, and screening out features whose linear correlation strength is greater than a preset threshold for subsequent model training.

6. The power load forecasting method based on quadratic mode decomposition and dual-model parallelism according to claim 1, characterized in that, The method further includes a dynamic iterative process, which employs a multi-step rolling prediction strategy, including: Iterative forecasting using a sliding time window is employed to generate future load forecast sequences. During the iterative prediction process, the hybrid prediction model is retrained or its parameters are fine-tuned using newly acquired actual load data to dynamically correct the model's prediction error.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.

Images