Intelligent power prediction method considering dynamic load change

Abstract

The invention discloses an intelligent power prediction method considering dynamic load change, and relates to the technical field of power grid load prediction, and the method comprises the steps: collecting original load data, carrying out the preprocessing, constructing a VMD constraint optimization model, carrying out the four-stage improvement of optimization parameters through employing an improved dung beetle optimization algorithm, so as to generate IMF components, reconstructing the IMF component by calculating a sample entropy to obtain a low-frequency component and a high-frequency component; establishing a Kalman filtering state space model based on the low-frequency component, and decomposing the low-frequency component into a residual component and a pseudo trend component through a Kalman filtering recursive algorithm; external features are obtained, the high-frequency component, the residual component and the pseudo trend component are aligned and spliced with the external features, multi-component collaborative prediction is carried out through a local-global interactive attention mechanism, and a final load prediction result is obtained; and generating a power demand visualization chart based on the final load prediction result. And reliable decision support is provided for power dispatching and energy management.

Description

Technical Field

[0001] This invention relates to the field of power grid load forecasting technology, and in particular to an intelligent power forecasting method that takes into account dynamic load changes. Background Technology

[0002] In the development of modern power systems, load forecasting technology has always been an important research direction for the safe and stable operation and economic dispatch of power grids. With the rapid growth of social electricity demand and the large-scale integration of new energy sources, the characteristics of power load have gradually shown strong nonlinearity, randomness and multi-time-varying characteristics. In the era of big data and artificial intelligence, neural network-based methods have been widely used. Their advantage lies in their ability to automatically learn nonlinear mapping relationships from massive historical data. However, these methods usually only rely on the directly input original load and external features, and lack in-depth processing of the multi-scale and non-stationary characteristics of the load sequence itself. As a result, when dealing with sudden disturbances, short-term fluctuations and long-term trends, the forecasting performance is often difficult to guarantee.

[0003] The shortcomings of existing technologies can be summarized as follows: First, traditional decomposition methods suffer from mode aliasing and endpoint effects, resulting in components lacking stable frequency separation characteristics, which affects subsequent prediction performance; Second, existing attention mechanisms in load prediction focus more on the internal dependencies of a single time series and do not adequately consider the interactions between different components, resulting in the failure to fully integrate cross-scale information. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent power prediction method that takes into account dynamic load changes, solving the problem of insufficient modeling of dynamic load fluctuations and multi-scale characteristics in existing load prediction methods.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent power prediction method that considers dynamic load changes, comprising,

[0008] Raw load data is collected and preprocessed to construct a VMD-constrained optimization model. An improved dung beetle optimization algorithm is used to perform a four-stage improvement and optimization of parameters to generate IMF components. The IMF components are reconstructed by calculating sample entropy to obtain low-frequency and high-frequency components.

[0009] A Kalman filter state-space model is established based on low-frequency components, and the low-frequency components are decomposed into residual components and pseudo-trend components through the Kalman filter recursive algorithm.

[0010] External features are acquired, and high-frequency components, residual components, and pseudo-trend components are aligned and concatenated with the external features. Multi-component collaborative prediction is performed through a local-global interactive attention mechanism to obtain the final load prediction result.

[0011] A power demand visualization chart is generated based on the final load forecast results.

[0012] As a preferred embodiment of the intelligent power prediction method considering dynamic load changes described in this invention, the following steps are included: collecting raw load data, preprocessing it, constructing a VMD-constrained optimization model, and using an improved dung beetle optimization algorithm to perform a four-stage improved optimization of parameters to generate IMF components, including...

[0013] Collect historical load sequences and preprocess them;

[0014] Based on the preprocessed historical load sequence, the preprocessed historical load sequence is decomposed into the following by minimizing the objective function of VMD and using ADMM iterative updates: One IMF component;

[0015] The fitness function is defined by combining the reconstruction error and the modal spectrum overlap.

[0016] An improved dung beetle optimization algorithm is used to set the hyperparameters of VMD as a search variable vector. To find the optimal value, where Indicates the number of IMF components. This represents the penalty factor, used to control the bandwidth of the modal components, sets the upper and lower bounds for each search variable, and sets the population size to [value missing]. The largest algebra is ;

[0017] For each individual, a uniform sample is taken within the value range to generate an initial population matrix. For each individual in the initial population matrix... Generate the corresponding lens reflection object through lens imaging. ;

[0018] Calculate the fitness of the initial population matrix and the corresponding lens inversion body, and retain the individuals with lower fitness as the official first generation individuals. The individual with the lowest fitness among all individuals is selected as the initial globally optimal individual. ;

[0019] For each generation of individuals, the process of reproduction-spiral search, foraging-optimal guidance, theft-dynamic weighting and evaluation-elite retention is performed iteratively until the maximum number of iterations is reached. Output the final globally optimal individual. ;

[0020] use Perform VMD on the preprocessed historical load sequence to generate IMF components .

[0021] As a preferred embodiment of the intelligent power prediction method considering dynamic load changes described in this invention, the step of reconstructing the IMF components by calculating sample entropy to obtain low-frequency and high-frequency components includes:

[0022] For each IMF component Construction length is Embedded vector ;

[0023] The distance between vectors is defined as the Chebyshev distance. Set threshold Statistical match ratio ;

[0024] structure dimensional vector calculation obtained And calculate the sample entropy. ;

[0025] By setting a low-frequency threshold through cross-validation, IMF components that are less than or equal to the low-frequency threshold are classified as low-frequency components. ,Will IMF components exceeding the low-frequency threshold are classified as high-frequency components. .

[0026] As a preferred embodiment of the intelligent power prediction method considering dynamic load changes described in this invention, the method involves establishing a Kalman filter state-space model based on low-frequency components, and decomposing the low-frequency components into residual components and pseudo-trend components using a Kalman filter recursive algorithm, including:

[0027] The state equations and observation equations of the Kalman filter state-space model are established based on low-frequency components;

[0028] The Kalman filter recursive algorithm is used to iterate the prediction and update steps, and output the posterior state estimate.

[0029] Residual calculation is performed based on posterior state estimation, and pseudo-trend components are output. and residual components .

[0030] As a preferred embodiment of the intelligent power prediction method considering dynamic load changes described in this invention, the step of acquiring external features and aligning and splicing high-frequency components, residual components, and pseudo-trend components with the external features includes:

[0031] Set the history window length By calculating the time step interval, the external features at the corresponding time point are obtained. These external features are then concatenated with the high-frequency components, residual components, and pseudo-trend components to construct the input vector for each component. , ;

[0032] The external features include temperature, humidity, weekdays and holidays as binary identifiers.

[0033] As a preferred embodiment of the intelligent power prediction method considering dynamic load changes described in this invention, the step of performing multi-component collaborative prediction through a local-global interactive attention mechanism to obtain the final load prediction result includes:

[0034] For each One-dimensional convolution and pooling are performed to obtain the encoding. ;

[0035] By using local attention branches, for each Calculate the query, key, and value matrix for each time point. The correlation between various time points within the internal timeframe is output for each... Local dependencies within;

[0036] Calculate different components using global attention branches. The interaction relationships between them are analyzed and weighted aggregation is performed.

[0037] The outputs of the two attention points are concatenated and linearly fused to obtain the context vector representation of each component;

[0038] The context vector of each component is input into the gated recurrent unit decoder, which outputs the future vector of each component. The load forecast values ​​at each time step are summed to obtain the final load forecast result.

[0039] Historical data is used to train the local attention branch, global attention branch, and gated recurrent unit decoder. The trained local attention branch, global attention branch, and gated recurrent unit decoder are then used to output the final load prediction result.

[0040] As a preferred embodiment of the intelligent power prediction method considering dynamic load changes described in this invention, the step of training the local attention branch, global attention branch, and gated recurrent unit decoder using historical data includes:

[0041] Using historical data as the training set, mean squared error is used as the loss function, and the prediction loss of the three components is calculated simultaneously. The Adam optimizer is used to pre-train the sub-network parameters of the three prediction components respectively. If the calculated loss no longer decreases significantly during continuous iteration, the iteration stops and the pre-trained sub-network is output. The pre-trained sub-networks are stacked into a parallel structure, and the loss of the final total output is calculated using mean squared error. The Adam optimizer is used for secondary fine-tuning training. After training, the trained local attention branch, global attention branch, and gated recurrent unit decoder are output.

[0042] As a preferred embodiment of the intelligent power forecasting method considering dynamic load changes described in this invention, the step of generating a power demand visualization chart based on the final load forecast result includes:

[0043] Based on the final load forecast results, basic statistical characteristics are calculated, including peak load, valley load, average load, and peak-valley difference.

[0044] By combining the prediction results of high-frequency components, residual components, and pseudo-trend components, a load trend curve is plotted.

[0045] Visual charts are generated based on basic statistical characteristics and load trend curves.

[0046] As a preferred embodiment of the intelligent power prediction method considering dynamic load changes described in this invention, the preprocessing of the historical load sequence includes cleaning outliers, filling in missing values, and normalization.

[0047] As a preferred embodiment of the intelligent power prediction method considering dynamic load changes described in this invention, the generation of visualization charts includes load trend superposition curves, peak-valley distribution histograms, and time-period average load bar charts.

[0048] The beneficial effects of this invention are as follows: by using the VMD constrained optimization model, the improved dung beetle optimization algorithm, sample entropy component reconstruction, and Kalman filter decomposition, the complex, non-stationary, multi-scale load signal is transformed into a structured and predictable component system. The pseudo-trend component focuses on long-term slow changes, the residual component focuses on medium-term fluctuations, and the high-frequency component focuses on instantaneous details. Ultimately, the total prediction output by the local-global interactive attention mechanism is both smooth and jitter-free and can quickly respond to short-term changes. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart of the intelligent power prediction method that takes into account dynamic load changes in Example 1.

[0051] Figure 2 This is a flowchart of VMD parameter optimization and IMF component decomposition in Example 1.

[0052] Figure 3 This is a flowchart of the local-global attention mechanism prediction in Example 1. Detailed Implementation

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0055] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0056] Example 1, referring to Figure 1 , Figure 2 and Figure 3 This is the first embodiment of the present invention, which provides a smart power prediction method that takes into account dynamic load changes, including the following steps:

[0057] S1. Collect raw load data and preprocess it to construct a VMD constrained optimization model. Use the improved dung beetle optimization algorithm to perform four-stage improved optimization parameters to generate IMF components. Reconstruct the IMF components by calculating sample entropy to obtain low-frequency and high-frequency components.

[0058] Specifically, historical load sequences are collected and preprocessed.

[0059] Based on the preprocessed historical load sequence, the preprocessed historical load sequence is decomposed into the following by minimizing the objective function of VMD and using ADMM iterative updates: For each IMF component, the objective function of VMD is expressed as:

[0060] ;

[0061] ;

[0062] in, and They represent the first Each IMF component sequence and its corresponding center frequency, Represents the Dirac function, Represents convolution. Represents the time derivative. Represents the imaginary unit. Represents a complex exponential function. The base of the natural logarithm. This represents the preprocessed historical load sequence;

[0063] The fitness function is defined by combining the reconstruction error and the modal spectrum overlap, and is expressed as:

[0064] ;

[0065] ;

[0066] ;

[0067] in, Represents the fitness function. and The weights are determined through cross-validation. Indicates reconstruction error, This indicates the time window length for the IMF components and the preprocessed historical load sequence. Indicates the modal spectrum overlap. Represents frequency domain transformation, and Indicates IMF components;

[0068] An improved dung beetle optimization algorithm is used to set the hyperparameters of VMD as a search variable vector. To find the optimal value, where Indicates the number of IMF components. This represents the penalty factor, used to control the bandwidth of the modal components, sets the upper and lower bounds for each search variable, and sets the population size to [value missing]. The largest algebra is ;

[0069] For each individual, a uniform sample is taken within the value range to generate an initial population matrix. For each individual in the initial population matrix... Generate the corresponding lens reflection object through lens imaging. , represented as:

[0070] ;

[0071] in, Let j represent the search variable for the i-th individual, and let j represent the search variable for the 1-th individual. The second search variable is , and These represent the upper and lower bounds of the j-th search variable, respectively. The lens inversion of the j-th search variable for the i-th individual;

[0072] Calculate the fitness of the initial population matrix and the corresponding lens inversion body, and retain the individuals with lower fitness as the official first generation individuals. The individual with the lowest fitness among all individuals is selected as the initial globally optimal individual. ;

[0073] For each generation of individuals, the process of reproduction-spiral search, foraging-optimal guidance, theft-dynamic weighting and evaluation-elite retention is performed iteratively until the maximum number of iterations is reached. Output the final globally optimal individual. ;

[0074] The aforementioned reproductive-spiral search refers to performing spiral updates centered on the current globally optimal individual, expressed as:

[0075] ;

[0076] ;

[0077] in, Indicates the first spiral update. The i-th individual of the generation, Indicates the first The i-th individual of the generation, Indicates the first The globally optimal individual of the generation. Indicates the dynamic radius parameter. Indicates the current iteration number. Represents a random number, with values ​​ranging from 1 to 2. between, The midpoint of the boundary is indicated to ensure a double-reference spiral search around the optimum and the center;

[0078] The foraging-optimal guidance refers to updating each search variable with upper and lower bounds and a random value, and guiding it towards the current optimal solution, expressed as:

[0079] ;

[0080] ;

[0081] in, Represents the first after optimal guidance The j-th search variable for the i-th individual in the generation, Indicates the first spiral update. The j-th search variable for the i-th individual in the generation, and Indicates the value in Random weights between them Indicates the first The j-th search variable for the globally optimal individual in the generation. This indicates the strength of the influence of the optimal position;

[0082] The "steal-dynamic weighting" refers to using two dynamic weights to control the contributions of global optimum and local perturbation, expressed as:

[0083] ;

[0084] ;

[0085] ;

[0086] in, Indicates the first The i-th individual of the generation, Represents the first after optimal guidance The i-th individual of the generation, and This represents the dynamic weights that change gradually with the number of iterations. Represents the normal distribution vector. This represents dimension-wise multiplication;

[0087] The evaluation-elite retention refers to calculating the fitness of all individuals in the current generation, and selecting the individual with the lowest fitness that is also lower than the fitness of the global best individual in the previous generation as the new best individual in the current generation; otherwise... ;

[0088] use Perform VMD on the preprocessed historical load sequence to generate IMF components .

[0089] Further preprocessing includes cleaning outliers, filling in missing values, and normalization.

[0090] By preprocessing historical load sequences, minimizing the VMD bandwidth objective function, and using ADMM iteration to decompose the preprocessed sequences into IMFs, a fitness function is jointly defined by reconstruction error and modal spectrum overlap. The VMD hyperparameters are encapsulated as search vectors. An improved dung beetle optimization is used to perform continuous iterations of "lens imaging reverse body" initialization, reproduction-spiral search, foraging-optimal guidance, stealing-dynamic weights, and elite retention. This achieves automatic determination of the optimal decomposition order and bandwidth penalty, and obtains an IMF set with clearer frequency band separation and lower reconstruction error. This is used to reduce modal aliasing and endpoint effects from the source and provide separable and interpretable multi-scale subsequences for subsequent modeling. Mirror pair stereo and spiral update enable the algorithm to both jump globally and search around the optimal vicinity in detail. Dynamic weights balance global expansion and local convergence, reducing noise coupling and feature redundancy from the source, making subsequent model inputs more interpretable and stable. Ultimately, this achieves the beneficial effects of eliminating manual parameter tuning of decomposition parameters, more stable component physical properties, and easier convergence of subsequent model training.

[0091] Furthermore, for each IMF component Construction length is Embedded vector , represented as:

[0092] ;

[0093] ;

[0094] in, Indicates the first indivual 3D embedding vector, Indicates the first The value of each IMF component at the i-th time point. , express Length, express Length;

[0095] The distance between vectors is defined as the Chebyshev distance. Set threshold Statistical match ratio , represented as:

[0096] ;

[0097] ;

[0098] in, Indicates the threshold. Indicates the threshold coefficient. express Standard deviation;

[0099] structure dimensional vector calculation obtained And calculate the sample entropy. , represented as:

[0100] ;

[0101] By setting a low-frequency threshold through cross-validation, IMF components that are less than or equal to the low-frequency threshold are classified as low-frequency components. ,Will IMF components exceeding the low-frequency threshold are classified as high-frequency components. .

[0102] For each IMF, an embedding vector is constructed to calculate the sample entropy SE. A low-frequency threshold is set through cross-validation, which enables the grouping and reconstruction of components based on complexity rather than ordinal number (SE≤threshold is low frequency, SE>threshold is high frequency). This achieves adaptive structure / noise partitioning of data, rather than subjective or fixed rule segmentation. It is used to clearly separate low frequencies that reflect long-term patterns from high frequencies that represent transient disturbances, providing a cleaner low-frequency carrier for subsequent Kalman filter input. It also provides three types of information sources with clear division of labor for multi-component collaborative prediction, ultimately achieving the beneficial effects of reducing information redundancy, improving component specificity and feature purity, and improving the signal-to-noise ratio and generalization ability of downstream models.

[0103] S2. Establish a Kalman filter state-space model based on low-frequency components, and decompose the low-frequency components into residual components and pseudo-trend components using the Kalman filter recursive algorithm.

[0104] Specifically, the state equations and observation equations of the Kalman filter state-space model based on low-frequency components are expressed as follows:

[0105] ;

[0106] ;

[0107] in, Indicates time The trend status, Represents the state transition matrix. The process noise is represented by a mean of 0 and a covariance of . The normal distribution The observed value at time t, i.e. , This represents the observation matrix, usually 1. The observation noise is represented by a mean of 0 and a covariance of . The normal distribution;

[0108] The Kalman filter recursive algorithm is used for iterative prediction and update steps, outputting a posterior state estimate. The prediction step is represented as follows:

[0109] ;

[0110] ;

[0111] in, This represents the prior state estimate. Indicates at time Posterior state estimation, initial state Pick The average value of the window, This represents the prior state error covariance. This represents the covariance of the posterior state error at the previous time step. Value , Represents process covariance;

[0112] The update step is represented as:

[0113] ;

[0114] ;

[0115] ;

[0116] in, Indicates Kalman gain, Represents the transpose of the observation matrix. Indicates time Posterior state estimation, Indicates time The posterior state error covariance, Represents the identity matrix;

[0117] Residual calculation is performed based on posterior state estimation, and pseudo-trend components are output. and residual components , represented as:

[0118] ;

[0119] ;

[0120] in, Indicates time The residual components, This indicates a pseudo-trend component.

[0121] By establishing state and observation equations using low-frequency components and running KF recursion (prediction-update), the posterior state estimate is output as a pseudo-trend. Simultaneously, residuals are constructed from the difference between observations and estimates, achieving smooth trend extraction and mid-term dynamic separation in the statistically optimal sense. This is used to remove the adverse effects of non-stationarity on the learner and to provide a more concentrated source of dynamic information for residual modeling. The trend responds more linearly to temperature and seasonal changes with less noise. The residuals capture mid-term behaviors such as equipment switching and load shifting, ultimately making the overall prediction both smooth and jitter-free and able to respond quickly to short-term changes. This reduces the overfitting of the feedforward model to outliers, ultimately achieving the beneficial effects of strict decoupling of trend and dynamics, faster convergence of subsequent models, and avoiding the "contamination" of short-term predictions by the trend term.

[0122] S3. Obtain external features, align and stitch high-frequency components, residual components, and pseudo-trend components with external features, and perform multi-component collaborative prediction through a local-global interactive attention mechanism to obtain the final load prediction result.

[0123] Specifically, set the history window length. By calculating the time step interval, the external features at the corresponding time point are obtained. These external features are then concatenated with the high-frequency components, residual components, and pseudo-trend components to construct the input vector for each component. , ;

[0124] The external features include temperature, humidity, weekdays and holidays as binary identifiers.

[0125] Based on the historical window length and time step, external features such as temperature, humidity, weekdays, and holidays are sequentially aligned and stitched with high-frequency components, residual components, and pseudo-trend components to form three input channels with external factors. This allows externally interpretable drivers to be injected into the modeling channel of each component in a time-aligned manner, enabling the model to learn the "differential impact of external factors on components at different time scales" (e.g., temperature mainly affects the trend level, while holidays mainly affect residuals and high-frequency spikes). This explicitly introduces "causal clues," reducing the possibility of the model misjudging "changes truly driven by external factors" as "internal autocorrelation," while improving business interpretability and usability.

[0126] Furthermore, for each One-dimensional convolution and pooling are performed to obtain the encoding. ;

[0127] By using local attention branches, for each Calculate the query, key, and value matrix for each time point. The correlation between various time points within the internal timeframe is output for each... Local dependencies within;

[0128] Calculate different components using global attention branches. The interaction relationships between them are analyzed and weighted aggregation is performed.

[0129] The outputs of the two attention points are concatenated and linearly fused to obtain the context vector representation of each component;

[0130] The context vector of each component is input into the gated recurrent unit decoder, which outputs the future vector of each component. The load forecast values ​​at each time step are summed to obtain the final load forecast result.

[0131] Historical data is used to train the local attention branch, global attention branch, and gated recurrent unit decoder. The trained local attention branch, global attention branch, and gated recurrent unit decoder are then used to output the final load prediction result.

[0132] For each input component, 1D-CNN local pattern extraction and denoising are performed separately. In the local attention branch, Q, K, and V are calculated to obtain the autocorrelation weights between time points, capturing the short-term dependencies within the component. In the global attention branch, cross-attention between components is used to characterize the cross-component modulation of trend ↔ residual ↔ high frequency, and weighted fusion is performed. Finally, GRU is used to decode the next τ steps and the three components are added together to reconstruct the total load. The attention and decoder are jointly trained with historical data, which realizes the simultaneous inclusion of "short-term structure within the component" and "long-term / cross-scale coupling between components" into the same information flow. It preserves local details and grasps the global modulation relationship. GRU models long-term dependencies with fewer parameters, maintaining computational controllability. The local attention branch makes the model more sensitive to local morphology before and after peaks. The global attention branch makes the model explicit the causal clues of "residual shift when the trend is upward" or "high frequency is prone to burst in certain residual stages". GRU autoregression ensures the coherence of multi-step prediction.

[0133] Furthermore, historical data is used as the training set, and mean squared error is used as the loss function. The prediction loss of the three components is calculated simultaneously. The Adam optimizer is used to pre-train the sub-network parameters of the three prediction components. If the calculated loss no longer decreases significantly during continuous iteration, the iteration stops and the pre-trained sub-network is output. The pre-trained sub-networks are stacked into a parallel structure, and the loss of the final total output is calculated using mean squared error. The Adam optimizer is used for secondary fine-tuning training. After training, the trained local attention branch, global attention branch, and gated recurrent unit decoder are output.

[0134] Using historical data as the training set and MSE as the loss, the three component subnetworks are first pre-trained with Adam and stopped early. Then, the three subnetworks are stacked in parallel, and the MSE of the total output is used as the secondary fine-tuning target for unified optimization. In the pre-training stage, each subnetwork learns the representation of its own domain first. In the fine-tuning stage, attention weights and decoder parameters are coordinated under the unified target, avoiding gradient interference and instability caused by end-to-end training from the beginning. This significantly reduces training variance and convergence difficulty, and improves robustness to hyperparameters.

[0135] S4. Generate a power demand visualization chart based on the final load forecast results;

[0136] Specifically, basic statistical characteristics are calculated based on the final load forecast results, including peak load, valley load, average load, and peak-valley difference;

[0137] By combining the prediction results of high-frequency components, residual components, and pseudo-trend components, a load trend curve is plotted.

[0138] Visual charts are generated based on basic statistical characteristics and load trend curves.

[0139] Furthermore, the visualization charts include load trend overlay curves, peak-valley distribution histograms, and time-period average load bar charts.

[0140] After obtaining the final forecast, basic statistics such as peak value, valley value, mean value, and peak-valley difference are calculated. The trend-dynamic overlay curve is drawn by combining the three-component forecast results, and a power demand visualization chart is generated. This realizes an intuitive mapping from model output to operational understanding, providing intuitive basis for load scheduling, demand response, reserve capacity configuration, and peak-valley electricity pricing strategies. It can also be used for abnormal early warning / alarm threshold triggering, forming an auditable and communicable intelligent forecasting system closed loop.

[0141] This embodiment also provides a computer device suitable for intelligent power prediction methods that consider dynamic load changes, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the intelligent power prediction method considering dynamic load changes as proposed in the above embodiment.

[0142] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0143] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the intelligent power prediction method considering dynamic load changes as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0144] It should be noted that the above 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A smart power prediction method considering dynamic load changes, characterized in that: include, Raw load data is collected and preprocessed to construct a VMD-constrained optimization model. An improved dung beetle optimization algorithm is used to perform a four-stage improvement and optimization of parameters to generate IMF components. The IMF components are reconstructed by calculating sample entropy to obtain low-frequency and high-frequency components. A Kalman filter state-space model is established based on low-frequency components, and the low-frequency components are decomposed into residual components and pseudo-trend components through the Kalman filter recursive algorithm. External features are acquired, and high-frequency components, residual components, and pseudo-trend components are aligned and concatenated with the external features. Multi-component collaborative prediction is performed through a local-global interactive attention mechanism to obtain the final load prediction result. A power demand visualization chart is generated based on the final load forecast results.

2. The intelligent power prediction method considering dynamic load changes as described in claim 1, characterized in that: The raw load data is collected and preprocessed to construct a VMD-constrained optimization model. An improved dung beetle optimization algorithm is then used for a four-stage optimization of parameters to generate IMF components. include, Collect historical load sequences and preprocess them; Based on the preprocessed historical load sequence, the preprocessed historical load sequence is decomposed into the following by minimizing the objective function of VMD and using ADMM iterative updates: One IMF component; The fitness function is defined by combining the reconstruction error and the modal spectrum overlap. An improved dung beetle optimization algorithm is used to set the hyperparameters of VMD as a search variable vector. To find the optimal value, where Indicates the number of IMF components. This represents the penalty factor, used to control the bandwidth of the modal components, sets the upper and lower bounds for each search variable, and sets the population size to [value missing]. The largest algebra is ; For each individual, a uniform sample is taken within the value range to generate an initial population matrix. For each individual in the initial population matrix... Generate the corresponding lens reflection object through lens imaging. ; Calculate the fitness of the initial population matrix and the corresponding lens inversion body, and retain the individuals with lower fitness as the official first generation individuals. The individual with the lowest fitness among all individuals is selected as the initial globally optimal individual. ; For each generation of individuals, the process of reproduction-spiral search, foraging-optimal guidance, theft-dynamic weighting and evaluation-elite retention is performed iteratively until the maximum number of iterations is reached. Output the final globally optimal individual. ; use Perform VMD on the preprocessed historical load sequence to generate IMF components .

3. The intelligent power prediction method considering dynamic load changes as described in claim 2, characterized in that: The process of reconstructing the IMF components by calculating sample entropy to obtain low-frequency and high-frequency components includes, For each IMF component Construction length is Embedded vector ; The distance between vectors is defined as the Chebyshev distance. Set threshold Statistical match ratio ; structure dimensional vector calculation obtained And calculate sample entropy ; By setting a low-frequency threshold through cross-validation, IMF components that are less than or equal to the low-frequency threshold are classified as low-frequency components. ,Will IMF components exceeding the low-frequency threshold are classified as high-frequency components. .

4. The intelligent power prediction method considering dynamic load changes as described in claim 3, characterized in that: The Kalman filter state-space model is established based on low-frequency components. The low-frequency components are decomposed into residual components and pseudo-trend components using a Kalman filter recursive algorithm. include, The state equations and observation equations of the Kalman filter state-space model are established based on low-frequency components; The Kalman filter recursive algorithm is used to iterate the prediction and update steps, and output the posterior state estimate. Residual calculation is performed based on posterior state estimation, and pseudo-trend components are output. and residual components .

5. The intelligent power prediction method considering dynamic load changes as described in claim 4, characterized in that: The process involves acquiring external features, aligning and concatenating high-frequency components, residual components, and pseudo-trend components with the external features. include, Set the history window length By calculating the time step interval, the external features at the corresponding time point are obtained. These external features are then concatenated with the high-frequency components, residual components, and pseudo-trend components to construct the input vector for each component. , ; The external features include temperature, humidity, weekdays and holidays as binary identifiers.

6. The intelligent power prediction method considering dynamic load changes as described in claim 5, characterized in that: The multi-component collaborative prediction using a local-global interactive attention mechanism yields the final load prediction result, including: For each One-dimensional convolution and pooling are performed to obtain the encoding. ; By using local attention branches, for each Calculate the query, key, and value matrix for each time point. The correlation between various time points within the system is output for each... Local dependencies within; Calculate different components using a global attention branch. The interaction relationships between them are analyzed and weighted aggregation is performed. The outputs of the two attention points are concatenated and linearly fused to obtain the context vector representation of each component; The context vector of each component is input into the gated recurrent unit decoder, which outputs the future vector of each component. The load forecast values ​​at each time step are summed to obtain the final load forecast result. Historical data is used to train the local attention branch, global attention branch, and gated recurrent unit decoder. The trained local attention branch, global attention branch, and gated recurrent unit decoder are then used to output the final load prediction result.

7. The intelligent power prediction method considering dynamic load changes as described in claim 6, characterized in that: The training of the local attention branch, global attention branch, and gated recurrent unit decoder using historical data includes, Using historical data as the training set, mean squared error is used as the loss function, and the prediction loss of the three components is calculated simultaneously. The Adam optimizer is used to pre-train the sub-network parameters of the three prediction components respectively. If the calculated loss no longer decreases significantly during continuous iteration, the iteration stops and the pre-trained sub-network is output. The pre-trained sub-networks are stacked into a parallel structure, and the loss of the final total output is calculated using mean squared error. The Adam optimizer is used for secondary fine-tuning training. After training, the trained local attention branch, global attention branch, and gated recurrent unit decoder are output.

8. The intelligent power prediction method considering dynamic load changes as described in claim 7, characterized in that: The generation of power demand visualization charts based on the final load forecast results includes, Based on the final load forecast results, basic statistical characteristics are calculated, including peak load, valley load, average load, and peak-valley difference. By combining the prediction results of high-frequency components, residual components, and pseudo-trend components, a load trend curve is plotted. Visual charts are generated based on basic statistical characteristics and load trend curves.

9. The intelligent power prediction method considering dynamic load changes as described in claim 8, characterized in that: The preprocessing of the historical load sequence includes cleaning outliers, filling in missing values, and normalization.

10. The intelligent power prediction method considering dynamic load changes as described in claim 9, characterized in that: The generated visualization charts include load trend overlay curves, peak-valley distribution histograms, and time-period average load bar charts.

Images