Digital twin power prediction method and system based on adaptive multi-resolution wavelet filtering

Abstract

The invention discloses a digital twin power prediction method and system based on adaptive multi-resolution wavelet filtering, and relates to the technical field of power prediction. Comprising the following steps: acquiring multivariable power time sequence data, and performing reversible instance normalization processing; inputting the normalized power time sequence data into an inverted embedding layer, and mapping the normalized power time sequence data to a high-dimensional feature space to obtain an initial sequence; carrying out adaptive multi-resolution discrete wavelet transform on the initial sequence, and extracting low-frequency and multi-scale high-frequency components of the initial sequence; performing element-by-element product on the frequency domain component and the learnable filtering kernel to complete adaptive wavelet filtering; reconstructing a filtered time domain sequence through inverse wavelet transform, and performing cross-channel aggregation and feature redistribution through a star aggregation-redistribution module; and inputting the extracted feature sequence into a feedforward neural network to obtain a power index prediction result of a future time step. According to the method, the modeling capability and prediction robustness of the non-stationary power data are effectively enhanced, and the method is high in adaptability and easy to deploy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power prediction, in particular to a digital twin power prediction method and system based on adaptive multi-resolution wavelet filtering. BACKGROUND

[0002] In recent years, with the continuous evolution of artificial intelligence technology, the application value of time series prediction in digital twin systems has become increasingly prominent. As a core technology supporting key links such as intelligent energy management, load scheduling optimization, and renewable energy access in digital twin systems, time series prediction has become an important part of power informatization and digital transformation. Especially under the background of increasing penetration of new energy and increasing load volatility, how to accurately predict high-frequency, long-period, and non-stationary power data has become a technical difficulty in intelligent power grid operation and dispatch. Traditional linear models such as autoregressive (AR) and difference integrated moving average model (ARIMA) have certain theoretical basis and interpretability, but when faced with the high dimensionality, multi-source heterogeneity, non-stationarity, and multi-scale change trend of power data, they are difficult to effectively capture the internal dynamic law, and the prediction accuracy and generalization ability are obviously insufficient.

[0003] The introduction of deep learning (DL) provides a new breakthrough for power time series prediction. From early recurrent neural networks (RNN), long short-term memory networks (LSTM), and gated recurrent units (GRU), to the widely used Transformer structure and its improved models (such as Informer, Autoformer, PatchTST, etc.) in recent years, these methods have shown strong representation ability in modeling complex temporal dependencies, long-term trends, and sudden responses, and have achieved excellent performance in multiple power load prediction and photovoltaic / wind power prediction tasks.

[0004] However, although the deep model-based method has made significant progress, it still faces multiple challenges in actual power applications. First, the training data in the power scenario is usually complex, covering multiple sites, cross-regional, and multiple time granularities, and there are problems such as data missing, inconsistent sampling, and distribution shift, making the deep model very sensitive to data quality and lacking stability and portability. Second, current mainstream methods mainly focus on time domain modeling, which is difficult to effectively mine the frequency domain structure features contained in power data, and the modeling ability of periodic changes, multi-scale trends, and high-frequency disturbances is insufficient, limiting its performance in high-dynamic loads or new energy scenarios.

[0005] In addition, although the series of models represented by the Transformer have strong long sequence modeling capability, the attention mechanism thereof has problems of quadratic growth of calculation complexity with sequence length and excessive memory consumption when processing large-scale power data (such as 10-minute-level sampling and cross-year sequences), and is difficult to efficiently run in resource-limited deployment environments such as edge terminals and micro-grid control. Therefore, a new prediction method that simultaneously considers time domain and frequency domain modeling capability, has multi-scale feature expression capability, and is suitable for the complexity of power data and resource-limited scenarios is urgently needed to improve the accuracy and practicality of time series modeling in power systems.

[0006] In order to further improve the accuracy and generalization capability of time series modeling, in recent years, some researches have attempted to introduce frequency domain processing ideas, such as preprocessing signals before model input through wavelet transform or Fourier transform, or embedding the transform process into the neural network structure to realize end-to-end feature extraction. However, most of the existing methods rely on fixed structures or artificially set scale parameters, lack the ability to dynamically adjust different frequency bands, and the frequency domain operation process is usually performed on the CPU, which is difficult to efficiently cooperate with the GPU training process, resulting in low training efficiency and difficulty in adapting to the needs of industrial data modeling. SUMMARY

[0007] The purpose of the present application is to propose a digital twin power prediction method and system based on adaptive multi-resolution wavelet filtering, which can effectively capture the local and global features of non-stationary time series, significantly improve the long-term prediction accuracy and computational efficiency, and is particularly suitable for high-frequency, large-dimensional power load prediction, distributed photovoltaic power station output modeling and other task scenarios, thereby improving the prediction accuracy while reducing the computational and deployment costs.

[0008] According to a first aspect of the embodiments of the present disclosure, a digital twin power prediction method based on adaptive multi-resolution wavelet filtering is provided, comprising the following steps:

[0009] Obtain multivariate power time series data and perform reversible instance normalization processing;

[0010] Input the normalized power time series data into an inverted embedding layer, map to a high-dimensional feature space, and obtain an initial sequence S0; perform adaptive multi-resolution discrete wavelet transform on the initial sequence S0 to extract low-frequency and multi-scale high-frequency components thereof; perform element-wise multiplication of the frequency domain components and a learnable filter kernel to complete adaptive wavelet filtering; and reconstruct the filtered time domain sequence through inverse wavelet transform;

[0011] The time domain sequence is further subjected to cross-channel aggregation and feature redistribution by a star aggregation-redistribution module;

[0012] The feature sequence extracted through the star polymerization-redistribution module is input into a feedforward neural network to obtain a power index prediction result at a future time step.

[0013] In one embodiment, the adaptive multi-resolution discrete wavelet transform is performed on an initial sequence X to obtain J layers of discrete wavelet decomposition:

[0014]

[0015] The low-frequency component y L and the high-frequency component are filtered by applying a learnable filter kernel:

[0016]

[0017] The filtered time-domain sequence is reconstructed by inverse wavelet transform:

[0018]

[0019] wherein ⊙ denotes element-wise multiplication, R denotes a real number field, C is the number of channels, and d is the sequence length.

[0020] In one embodiment, the inverted embedding layer is implemented using linear projection, and the embedding expression is:

[0021] S0=XW+b

[0022] wherein X is the input normalized power data matrix, W is the projection weight matrix, and b is the bias term.

[0023] In one embodiment, the input of the feedforward neural network is the sequence representation S filtered after feature extraction, and the output is the future H-step prediction value

[0024]

[0025] In one embodiment, the filter kernel is a trainable parameter that can adaptively adjust its frequency response characteristics according to the frequency structure of the input sequence in different power scenarios.

[0026] In one embodiment, the feedforward neural network is trained end-to-end through a loss function, which is the mean square error between the predicted value and the true value:

[0027]

[0028] In one of the embodiments, the wavelet basis used in the adaptive multi-resolution discrete wavelet transform includes the following wavelet families: Daubechies, Haar, Coiflet or Symlet, to adapt to the periodicity and abruptness characteristics of different power data.

[0029] According to a second aspect of the embodiments of the present disclosure, a digital twin power prediction system based on adaptive multi-resolution wavelet filtering is provided, comprising:

[0030] A preprocessing module acquires multivariate power time series data and performs reversible instance normalization processing.

[0031] A wavelet filtering module inputs the normalized power time series data into an inverted embedding layer, maps to a high-dimensional feature space, and obtains an initial sequence S0; performs adaptive multi-resolution discrete wavelet transform on the initial sequence S0, extracts low-frequency and multi-scale high-frequency components thereof; performs element-wise multiplication of the frequency domain components and a learnable filter kernel to complete adaptive wavelet filtering; and reconstructs the filtered time domain sequence through inverse wavelet transform.

[0032] An extraction module further performs cross-channel aggregation and feature redistribution of the time domain sequence through a star aggregation-redistribution module.

[0033] A prediction module inputs the feature sequence extracted through the star aggregation-redistribution module into a feedforward neural network to obtain a power index prediction result at a future time step.

[0034] According to a third aspect of the embodiments of the present disclosure, an electronic device is provided, comprising a memory, a processor, and a computer program stored on the memory and running on the memory, wherein the processor implements the adaptive multi-resolution wavelet filtering based digital twin power prediction method when executing the program.

[0035] According to a fourth aspect of the embodiments of the present disclosure, a computer readable storage medium is provided, which stores a computer program, and the program is executed by a processor to implement the adaptive multi-resolution wavelet filtering based digital twin power prediction method.

[0036] The above technical solutions adopted by the present application have the following advantages compared with the prior art:

[0037] 1. Long-term prediction optimization of non-stationary time series: the wavelet domain adaptive filtering mechanism is adopted to realize accurate separation of trend items and noise components through multi-resolution analysis, and the model network MSE index is improved by 12%-18%. This technology is particularly suitable for industrial time series scenarios such as power prediction, and can effectively capture the periodic and abrupt characteristics of signals.

[0038] 2. Digital twin real-time processing architecture: adaptive multi-resolution wavelet filtering can achieve end-to-end 2ms-level inference delay on the NVIDIA V100 GPU platform based on the lightweight design of the AMRW filter, reducing the computational overhead by 63% compared to traditional frequency domain methods. Its linear complexity architecture supports parallel processing of thousands of sensor channels, meeting the stringent requirements of intelligent manufacturing lines for millisecond-level real-time response.

[0039] 3. Multivariate coupling modeling enhancement: through the star-shaped aggregation-reassignment module, cross-variable correlation modeling is realized on the ECL dataset, and the prediction error is reduced by 22% compared to the independent channel model.

[0040] 4. Frequency domain multi-scale feature joint modeling: based on the frequency domain wavelet analysis framework, a multi-scale feature extraction network is constructed to realize the collaborative perception of transient mutations and long-term periodic trends in power time series data. This technology significantly enhances the modeling capability of complex power system dynamic characteristics, and simultaneously improves the prediction accuracy and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application, and of which are used to explain the present application, do not constitute an inappropriate limitation to the present application.

[0042] Figure 1 The overall architecture of the prediction model TWRNet is shown in the figure;

[0043] Figure 2 The implementation details of the adaptive multi-resolution wavelet filtering module are shown in the figure;

[0044] Figure 3 The flowchart of the star-shaped aggregation-reassignment module is shown in the figure;

[0045] Figure 4 The visualization comparison of different resolution wavelet coefficients is shown in the figure;

[0046] Figure 5 The prediction effect of industrial power equipment oil temperature (after normalization) is shown in the figure;

[0047] Figure 6 The prediction effect of photovoltaic power generation data (after normalization) is shown in the figure;

[0048] Figure 7 The wavelet transform time-frequency diagram of the model based on the cmor wavelet basis on the Weather dataset is shown in the figure;

[0049] Figure 8 The wavelet transform time-frequency diagram of the model based on the cmor wavelet basis on the Solar dataset is shown in the figure;

[0050] Figure 9Wavelet transform time-frequency map based on cmor wavelet basis for the model on the Traffic dataset. DETAILED DESCRIPTION

[0051] The present disclosure is further illustrated by the following examples and figures.

[0052] It should be noted that the following detailed description is merely exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0053] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0054] It should be noted that the flowchart and block diagrams in the drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of the present disclosure. It should also be noted that each block in the flowchart and block diagrams and / or combinations of blocks in the flowchart and block diagrams can be implemented by a combination of hardware and software, or hardware alone, as desired. Further, it should be noted that each block in the flowchart and block diagrams and / or combinations of blocks in the flowchart and block diagrams can be implemented by special-purpose hardware-based systems that perform the specified functions or acts, or combinations of any of the above. The flowcharts and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of the present disclosure. In this regard, each block in the flowcharts or block diagrams can represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical functions ("instructions"). It should also be noted that in some alternative implementations, the functions noted in the block can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the flowchart illustrations and / or block diagrams, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of any of the above. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising", when used in this specification, specify the presence of stated features, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0055] Embodiment One:

[0056] The embodiment provides a digital twin power prediction method based on adaptive multi-resolution wavelet filtering, including the following steps:

[0057] The digital twin power prediction method based on adaptive multi-resolution wavelet filtering includes the following steps:

[0058] S1. Obtain multivariate power time series data and perform reversible instance normalization processing;

[0059] Specifically, from the power system monitoring equipment, database and other channels, the power time series data containing voltage, current, power and other variables are collected; in order to unify the data dimension and improve the subsequent model training effect, reversible instance normalization processing is performed on the obtained original data.

[0060] S2. The normalized power time series data is input into the inverted embedding layer, mapped to a high-dimensional feature space to obtain an initial sequence S0; adaptive multi-resolution discrete wavelet transform is performed on the initial sequence S0 to extract its low-frequency and multi-scale high-frequency components; element-wise multiplication is performed between the frequency domain components and the learnable filter kernel to complete adaptive wavelet filtering; the filtered time domain sequence is reconstructed through inverse wavelet transform;

[0061] Specifically, the inverted embedding layer realizes high-dimensional mapping in a linear projection manner, S0=XW+b, where X is the normalized power data matrix input, the dimension is adapted to the multi-variable power time series data structure; W is the projection weight matrix, used to realize linear transformation from the original data dimension to the high-dimensional feature space, and its dimension is set according to the size of the target high-dimensional space; b is the bias term, used to adjust the overall offset of the mapped data. Through this linear projection operation, the initial sequence mapped to the high-dimensional feature space is obtained.

[0062] The initial sequence S0 is decomposed by J-layer discrete wavelet (J is the decomposition layer number set according to the characteristics of power data and prediction requirements), and the signal is decomposed in time and frequency by a wavelet basis function in the decomposition process to obtain low-frequency components and multi-scale high-frequency components. The low-frequency components reflect the overall trend of the power time series data, and the high-frequency components capture the detailed fluctuations at different scales (such as load mutation, periodic disturbance, etc.).

[0063] In this embodiment, the wavelet basis used in the adaptive multi-resolution discrete wavelet transform includes but is not limited to the following wavelet families: Daubechies, Haar, Coiflet or Symlet. The appropriate wavelet basis can be selected flexibly according to the periodicity (such as the daily-week-year cycle of photovoltaic output and the peak-valley cycle of load) and the mutation characteristics (such as power mutation caused by faults) of different power data. For example, when processing power fault transient data with obvious sharp mutation characteristics, Haar wavelet basis can be preferred; when processing continuous smooth and periodic regular load data, Daubechies wavelet family may be more suitable.

[0064] The filter kernel is a trainable parameter that can adaptively adjust its frequency response characteristics according to the frequency structure of the input sequence in different power scenarios. The extracted low-frequency components and multi-scale high-frequency components are respectively multiplied element-wise with the corresponding learnable filter kernel. Through this operation, adaptive filtering of different frequency components is realized, the effective features related to power prediction are highlighted, and noise and irrelevant interference are suppressed.

[0065] S3. The time domain sequence is further aggregated across channels by a star aggregation-reassignment module, and features are reassigned;

[0066] Specifically, the star aggregation-reassignment module first aggregates the features across channels to mine the potential correlation between different variable channels, and then reassigns the features according to the aggregation results to optimize the distribution of the features in each channel and improve the effectiveness of feature expression, thereby providing more valuable input for the subsequent prediction module and achieving the following implementation:

[0067] Cross-channel aggregation:

[0068]

[0069] Feature reassignment:

[0070]

[0071] where g represents the channel aggregation result; C is the total number of channels; S c to correspond to the channel sequence features; represents channel concatenation, and MLP is a two-layer perceptron with a GELU activation function.

[0072] S4. The feature sequence extracted by the star aggregation-reassignment module is input into a prediction model to obtain a prediction result of the power index at a future time step.

[0073] Specifically, the prediction model uses a feed-forward neural network (FFNN) whose input is the feature sequence extracted and optimized by the star aggregation-reassignment module. The feed-forward neural network performs layer-by-layer nonlinear transformation on the input features through a multi-layer perceptron structure, and finally outputs the predicted value of the power index at a future time step (such as 1 hour, 1 day, etc., which is set according to actual prediction requirements). To enable the prediction model to accurately learn the mapping relationship of power time series data, the feed-forward neural network is trained end-to-end through a loss function. In this embodiment, the mean squared error (MSE) between the predicted value and the true value is selected as the loss function.

[0074] Embodiment Two:

[0075] The embodiment provides a digital twin power prediction system based on adaptive multi-resolution wavelet filtering, which comprises:

[0076] A preprocessing module acquires multivariate power time series data and performs reversible instance normalization processing;

[0077] The wavelet filtering module inputs the normalized power time series data into an inverted embedding layer, maps to a high-dimensional feature space, and obtains an initial sequence S0; the initial sequence S0 is subjected to adaptive multi-resolution discrete wavelet transform, and low-frequency and multi-scale high-frequency components thereof are extracted; the frequency domain components are subjected to element-by-element multiplication with a learnable filter kernel, and adaptive wavelet filtering is completed; and the filtered time domain sequence is reconstructed through inverse wavelet transform;

[0078] The extraction module performs cross-channel aggregation and feature redistribution on the time domain sequence again through the star aggregation-redistribution module.

[0079] The prediction module inputs the feature sequence extracted through the star aggregation-redistribution module into a feedforward neural network, and obtains a power index prediction result at a future time step.

[0080] Embodiment three:

[0081] An electronic device includes a memory, a processor, and a computer program stored on the memory and running on the memory, and the processor executes the program to implement the above-mentioned digital twin power prediction method based on adaptive multi-resolution wavelet filtering, including:

[0082] Obtain multivariate power time series data and perform reversible instance normalization processing;

[0083] Input the normalized power time series data into an inverted embedding layer, map to a high-dimensional feature space, and obtain an initial sequence S0; the initial sequence S0 is subjected to adaptive multi-resolution discrete wavelet transform, and low-frequency and multi-scale high-frequency components thereof are extracted; the frequency domain components are subjected to element-by-element multiplication with a learnable filter kernel, and adaptive wavelet filtering is completed; and the filtered time domain sequence is reconstructed through inverse wavelet transform;

[0084] The time domain sequence is subjected to cross-channel aggregation and feature redistribution again through the star aggregation-redistribution module.

[0085] The feature sequence extracted through the star aggregation-redistribution module is input into a feedforward neural network, and a power index prediction result at a future time step is obtained.

[0086] Embodiment four:

[0087] A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the above-mentioned digital twin power prediction method based on adaptive multi-resolution wavelet filtering, including:

[0088] Obtain multivariate power time series data and perform reversible instance normalization processing;

[0089] The normalized power time series data is input to an inverted embedding layer, mapped to a high-dimensional feature space, and an initial sequence S0 is obtained; the initial sequence S0 is subjected to adaptive multi-resolution discrete wavelet transform, and low-frequency and multi-scale high-frequency components thereof are extracted; the frequency domain components are subjected to element-by-element multiplication with a learnable filter kernel, and adaptive wavelet filtering is completed; the filtered time domain sequence is reconstructed through inverse wavelet transform;

[0090] The time domain sequence is subjected to cross-channel aggregation and feature redistribution through a star aggregation-redistribution module.

[0091] The feature sequence extracted through the star aggregation-redistribution module is input to a feedforward neural network, and a power index prediction result at a future time step is obtained.

[0092] The application is particularly suitable for industrial-level time series prediction tasks and can maintain stable prediction performance under harsh conditions such as long period, large dimension, and low delay. The following embodiments are described in combination with three typical application scenarios:

[0093] Application Example 1: Medium and long-term power load prediction for a substation (based on the ETT dataset)

[0094] This embodiment is directed to medium and long-term load prediction tasks in a power system and is suitable for substation load management, daily / hourly load planning of a power grid dispatching center, and power market transaction prediction. The task requires the model to have both long-term trend modeling capability and robustness to short-term fluctuations.

[0095] In this scenario, the TWRNet model can access multivariate time series input composed of historical load records of a substation, power grid operation indicators, and related meteorological factors (such as temperature, humidity, holiday information, etc.). Inverse normalization is used in the model preprocessing stage for data preprocessing to ensure prediction accuracy and result interpretability. Then, the AMRW module in the model extracts multi-band signals to realize collaborative modeling of periodicity, growth trend, and abnormal disturbance of electricity consumption.

[0096] Evaluation results on the ETT (Electricity Transformer Temperature) series dataset show that TWRNet has an average improvement of more than 15% in MSE and MAE indicators in a multi-step prediction scenario (such as predicting the next 96 steps) compared to baseline models such as Informer and Autoformer. The model has the advantages of fast reasoning speed and low deployment overhead, is suitable for deployment on edge computing nodes, and helps to realize advance adjustment of power grid load and generation of energy optimization strategies.

[0097] Application Example 2: Prediction of electricity consumption mode in a typical industrial park (based on the ECL dataset)

[0098] This embodiment is suitable for industrial park, data center, large factory and mine enterprise electricity consumption trend prediction, and can serve intelligent power distribution management, energy budget formulation and real-time optimization of load transfer strategy.

[0099] The ECL (Electricity Consuming Load) dataset contains long-span electricity consumption data from 370 customers (such as enterprises and units), covering multiple regions and business types, with typical multi-source heterogeneous characteristics, and can represent industrial power scenarios.

[0100] In this task, TWRNet fully utilizes its multi-scale frequency domain modeling advantage to automatically separate the power fluctuations caused by intra-day load fluctuations, weekly periodicity and holiday changes. At the same time, with the STAR module, the model can also identify key external factors (such as temperature, workday information, etc.) that drive changes in consumption behavior, improving the model's interpretability and prediction adaptability.

[0101] The experimental results show that on the ECL dataset, TWRNet outperforms the mainstream Transformer-based method by more than 12% in terms of RMSE, MAE, etc., especially in holiday load mutation identification and daily trend modeling, and is suitable for deployment in energy consumption monitoring systems or power distribution management systems to provide prediction support for industrial power dispatching.

[0102] Application Example 3: Distributed Photovoltaic System Power Generation Prediction (based on Solar dataset)

[0103] This embodiment focuses on photovoltaic power generation prediction tasks in renewable energy scenarios, and is widely applicable to photovoltaic power station, park photovoltaic roof system, and distributed energy micro-grid power generation power real-time prediction and dispatching management.

[0104] The Solar dataset provides output power data of solar photovoltaic systems at different time points, reflecting the combined effects of weather, irradiance, cloud cover, and other factors on photovoltaic power generation. In this embodiment, TWRNet effectively separates the short-term fluctuations caused by weather disturbances from the periodic trend of sunlight through wavelet filtering, thereby accurately restoring the power variation law at multiple scales.

[0105] Compared with traditional LSTM and Informer models, TWRNet shows stronger short-term prediction accuracy and noise resistance in this task. Experiments on the Solar dataset show that its RMSE index is improved by an average of 18.6%, and the prediction stability is significantly enhanced, making it suitable for deployment in distributed energy management platforms to realize dynamic assessment of photovoltaic output, energy storage dispatch linkage, and auxiliary regulation of power supply and demand balance.

[0106] Those skilled in the art should understand that the modules or steps of the present disclosure described above can be realized by a general computer device, and alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device and executed by a computing device, or they can be respectively manufactured into individual integrated circuit modules, or a plurality of modules or steps among them can be manufactured into a single integrated circuit module. The present disclosure is not limited to any specific combination of hardware and software.

[0107] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0108] The above describes the specific embodiments of the present disclosure in conjunction with the accompanying drawings, but is not intended to limit the protection scope of the present disclosure. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present disclosure without creative labor are still within the protection scope of the present disclosure.

Claims

1. A digital twin power prediction method based on adaptive multi-resolution wavelet filtering, characterized in that, The method comprises the following steps: obtaining multivariate power time series data and performing reversible instance normalization processing; inputting the normalized power time series data into an inverted embedding layer to map to a high-dimensional feature space to obtain an initial sequence S0; performing adaptive multi-resolution discrete wavelet transform on the initial sequence S0 to extract low-frequency and multi-scale high-frequency components thereof; performing element-by-element multiplication of the frequency domain components and a learnable filter kernel to complete adaptive wavelet filtering; and reconstructing the filtered time domain sequence through inverse wavelet transform; the time domain sequence is further subjected to cross-channel aggregation and feature redistribution through a star aggregation-redistribution module; inputting the feature sequence extracted through the star aggregation-redistribution module into a feedforward neural network to obtain a power index prediction result at a future time step.

2. The digital twin power prediction method based on adaptive multi-resolution wavelet filtering according to claim 1, characterized in that, The adaptive multi-resolution discrete wavelet transform mode is: performing a J-level discrete wavelet decomposition on the initial sequence performing a J-level discrete wavelet decomposition: for low frequency components y L and high frequency components applying a learnable filter kernel: The filtered time domain sequence is reconstructed through inverse wavelet transform: wherein denotes element-wise multiplication, R denotes the field of real numbers, C is the number of channels, and d is the sequence length.

3. The digital twin power prediction method based on adaptive multi-resolution wavelet filtering according to claim 1, characterized in that, The inverted embedding layer is realized in a linear projection manner, and the embedding expression is: S0 = XW + b where X is the input normalized power data matrix, W is the projection weight matrix, and b is the bias term.

4. The digital twin power prediction method based on adaptive multi-resolution wavelet filtering according to claim 1, characterized in that, The feedforward neural network input is a sequence representation S after feature extraction filtered , and the output is a future H-step prediction value 5. The digital twin power prediction method based on adaptive multi-resolution wavelet filtering according to claim 1, characterized in that, The filter kernel is a trainable parameter that can adaptively adjust its frequency response characteristics according to the frequency structure of the input sequence in different power scenarios.

6. The digital twin power prediction method based on adaptive multi-resolution wavelet filtering according to claim 4, characterized in that, The feedforward neural network is trained end-to-end through a loss function, which is the mean square error between the predicted value and the true value.

7. The digital twin power prediction method based on adaptive multi-resolution wavelet filtering according to claim 1, characterized in that, The wavelet basis used in the adaptive multi-resolution discrete wavelet transform includes the following wavelet families: Daubechies, Haar, Coiflet or Symlet, to adapt to the periodicity and abruptness characteristics of different power data.

8. A digital twin power prediction system based on adaptive multi-resolution wavelet filtering, characterized in that, It comprises: a preprocessing module that obtains multivariate power time series data and performs reversible instance normalization processing; a wavelet filtering module that inputs the normalized power time series data into an inverted embedding layer to map to a high-dimensional feature space to obtain an initial sequence S0; performs adaptive multi-resolution discrete wavelet transform on the initial sequence S0 to extract low-frequency and multi-scale high-frequency components thereof; performing element-by-element multiplication of the frequency domain components and a learnable filter kernel to complete adaptive wavelet filtering; and reconstructing the filtered time domain sequence through inverse wavelet transform; an extraction module that performs cross-channel aggregation and feature redistribution on the time domain sequence through a star aggregation-redistribution module; a prediction module that inputs the feature sequence extracted through the star aggregation-redistribution module into a feedforward neural network to obtain a power index prediction result at a future time step.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the memory, wherein the processor implements the adaptive multi-resolution wavelet filtering-based digital twin power prediction method of any one of claims 1-7 when executing the program.

10. A computer-readable storage medium having a computer program stored thereon, wherein the program implements the adaptive multi-resolution wavelet filtering-based digital twin power prediction method of claims 1-7 when executed by a processor.

Images