An artificial intelligence-based power load forecasting system

The complexity of power load data is solved by using overlapping feature extraction and decomposition, time-varying filters and variational mode decomposition, wavelet transformation and convolutional time feature extraction in the power load prediction system. Feature extraction problem improves prediction effect and system performance.

CN118627929BActive Publication Date: 2024-11-15BEIJING GUONENG GUOYUAN ENERGY TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410828840.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-25
Publication Date
2024-11-15
Estimated Expiration
2044-06-25

AI Technical Summary

Technical Problem

In the existing power load prediction system, the frequency band data of power load data overlaps with complex patterns, resulting in low feature extraction quality and affecting the prediction effect; traditional methods have poor performance when processing nonlinear and nonsmooth signals and cannot Effectively handle complex high-frequency, low-frequency and overlapping features.

Method used

The dual-path feature extraction idea of ​​overlapping feature extraction and overlapping feature decomposition is used, and the feature extraction performance is optimized by combining time-varying filters and wavelet transformations of variational modal decomposition, smoothing function and marginal spectrum analysis; the multi-purpose feature extraction is extracted using convolution time The path cyclic bidirectional long and short-term memory network performs classification prediction, processing high-frequency, low-frequency and overlapping data.

Benefits of technology

Improve the quality of feature data, optimize system performance, improve prediction effect, and more effectively handle complex power load characteristics, improve the overall architecture of the prediction model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118627929B_ABST
    Figure CN118627929B_ABST
Patent Text Reader

Abstract

The present invention discloses an artificial intelligence-based power load forecasting system, which includes a data acquisition module, a feature processing module, a prediction drive module, a prediction application module and a monitoring feedback module. The present invention belongs to the technical field of power load forecasting, and specifically is an artificial intelligence-based power load forecasting system. The present invention adopts a dual-path feature extraction idea of ​​overlapping feature extraction and overlapping feature decomposition to extract overlapping features and decompose features, thereby optimizing the overall performance of the system; adopts a time-varying filter to improve the traditional empirical mode decomposition, thereby optimizing the performance of feature extraction; adopts an improved wavelet transform method combining variational mode decomposition, smoothing function and marginal spectrum analysis, thereby further improving the quality of features; adopts a multi-path cyclic bidirectional long short-term memory network combined with convolution time feature extraction to perform classified load forecasting, thereby optimizing the overall architecture of the forecasting model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of power load forecasting, and specifically relates to a power load forecasting system based on artificial intelligence. Background Art

[0002] An AI-based power load forecasting system uses machine learning algorithms, such as deep learning and time series analysis, to model and analyze historical load data, weather conditions, economic indicators and other factors to predict future power demand. This system can significantly improve the efficiency and stability of the power system, optimize power generation and transmission and distribution plans, thereby reducing operating costs and improving energy utilization.

[0003] However, in the existing power load forecasting system, the quality of feature extraction of power load determines the final effect of prediction to a large extent, and the frequency band data of power load forecasting data overlaps with each other and has complex rules. Therefore, the feature engineering design for overlapping power load data needs to consider both overlap and separation. In the existing overlapping feature extraction methods, there are frequency band data with aliasing modes. In the traditional empirical mode decomposition method, the performance of processing nonlinear and non-smooth signals is poor, resulting in low extraction efficiency and affecting the subsequent prediction effect. Technical problems; in the existing overlapping feature decomposition methods, there is a technical problem that only using time-varying filters to change the empirical mode decomposition method cannot solve the fundamental problem of mutual series connection and interference between frequency bands, and the traditional wavelet transform method for decomposition feature extraction has the technical problem that the spectrum segmentation accuracy is insufficient and the smoothness and meticulousness need to be improved; in the existing specific methods of power load forecasting, there is a simple model mixture that cannot effectively handle complex high-frequency, low-frequency and overlapping features, so it is necessary to design an appropriate hybrid model for effective classification prediction. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an artificial intelligence-based power load forecasting system. In view of the fact that in the existing power load forecasting system, the quality of power load feature extraction largely determines the final prediction effect, and the frequency band data of the power load forecasting data overlaps with each other and has complex rules, the feature engineering design of overlapping power load data needs to consider both overlap and separation. The present solution creatively adopts a dual-path feature extraction approach of overlapping feature extraction and overlapping feature decomposition to perform overlapping feature extraction and decomposition feature extraction. By designing adaptive extraction algorithms respectively, the feature data quality is effectively improved and the overall performance of the system is optimized. In view of the fact that in the existing overlapping feature extraction method, there is a technical problem that the performance of processing nonlinear and non-smooth signals in the traditional empirical mode decomposition method of frequency band data with aliasing modes is poor, resulting in low extraction efficiency and affecting subsequent prediction effects, the present solution creatively adopts a time-varying filter to improve the traditional empirical mode decomposition and optimizes the performance of feature extraction. In view of the fact that in the existing overlapping feature decomposition method, there is only The use of time-varying filters to change the empirical mode decomposition method cannot solve the fundamental problem of mutual connection and interference between frequency bands, while the traditional wavelet transform method for decomposition feature extraction has technical problems such as insufficient spectrum segmentation accuracy and the need to improve smoothness and meticulousness. This solution creatively adopts an improved wavelet transform method combining variational mode decomposition, smoothing function and marginal spectrum analysis to optimize the overall process of decomposition feature extraction and further improve the quality of features. In view of the fact that in the existing specific methods of power load forecasting, a simple model mixture cannot effectively handle complex high-frequency, low-frequency and overlapping features, and thus it is necessary to design an appropriate hybrid model for effective classification prediction, this solution creatively adopts a multi-path cyclic bidirectional long short-term memory network combined with convolutional time feature extraction to perform classification load forecasting, further combines time features through convolutional subnets to reduce the neglect of environmental conditions and other objective conditions, and uses cyclic subnets to process high-frequency data and bidirectional long short-term subnets to process low-frequency and overlapping data, thereby optimizing the overall architecture of the prediction model and providing a practical basis and exploration experience for further refinement of power load forecasting.

[0005] The technical solution adopted by the present invention is as follows: an artificial intelligence-based power load forecasting system provided by the present invention comprises a data acquisition module, a feature processing module, a forecasting driving module, a forecasting application module and a monitoring feedback module;

[0006] The data acquisition module is used for data collection and processing, and obtains the original power load forecast data and the optimized power load forecast data through data collection and processing, and sends the optimized power load forecast data to the feature processing module, and sends the original power load forecast data to the forecast driving module;

[0007] The feature processing module is used for overlapping feature extraction and overlapping feature decomposition, and obtains overlapping frequency-division feature data and decomposed feature data through overlapping feature extraction and overlapping feature decomposition, and sends the overlapping frequency-division feature data and decomposed feature data to the prediction driving module;

[0008] The prediction driving module is used for load classification prediction, obtains power load prediction data through load classification prediction, and sends the power load prediction data to the prediction application module and the monitoring feedback module;

[0009] The prediction application module is used for predictive value visualization, obtains predictive value visualization image analysis data through predictive value visualization, and sends the predictive value visualization image analysis data to the monitoring feedback module;

[0010] The monitoring feedback module is used for model accuracy monitoring, and obtains a power load prediction reference through model accuracy monitoring.

[0011] Furthermore, the data collection process is used to collect raw data required for power load forecasting, specifically, to obtain raw data for power load forecasting from power system and environmental sensor records through data collection, and to obtain optimized power load forecasting data by optimizing the raw data for power load forecasting;

[0012] The power load forecast raw data specifically includes historical power load data, weather environment data and special date data;

[0013] The optimization process specifically includes outlier processing and time series decomposition operations.

[0014] Furthermore, the overlapping feature extraction is used to extract frequency-division features from power load data that overlap in different frequency bands, specifically, based on the optimized power load forecast data, a time-varying filtering modal decomposition method combined with sample entropy is used to extract overlapping features to obtain overlapping frequency-division feature data;

[0015] The time-varying filtering modal decomposition method combined with sample entropy specifically includes a time series sample entropy operator and a time-varying filtering modal decomposition operator;

[0016] The step of extracting overlapping features by using a time-varying filtering modal decomposition method combined with sample entropy to obtain overlapping frequency-division feature data includes:

[0017] Construct a time-varying filter modal decomposition operator, decompose the original data, and obtain the time-varying filter decomposition data. The specific steps include:

[0018] Instantaneous data extraction, specifically using the Hilbert transform method, extracts the instantaneous amplitude and instantaneous phase of the optimized power load forecast data. The calculation formula is:

[0019] ;

[0020] Where A(t) is the instantaneous amplitude of the optimized power load forecast data, t is the time index, is the instantaneous phase, X(t) is the original signal data in the optimized power load forecasting data, It is the data after the optimized power load forecast data has been Hilbert transformed;

[0021] Constructing a local extreme value sequence for searching for local maximum values ​​and local minimum values ​​in the original data, specifically identifying a local maximum value sequence and a local minimum value sequence according to the instantaneous amplitude A(t) of the optimized power load forecast data;

[0022] The average envelope value is calculated by interpolating the local maximum value sequence and the local minimum value sequence to obtain the upper envelope sequence interpolation data and the lower envelope sequence interpolation data, and obtaining the upper envelope sequence instantaneous amplitude and the lower envelope sequence instantaneous amplitude. The calculation formula is:

[0023] ;

[0024] In the formula, is the instantaneous amplitude of the upper envelope sequence, is the instantaneous amplitude of the lower envelope sequence, is the upper envelope sequence interpolation data, is the lower envelope sequence interpolation data;

[0025] Calculate the instantaneous frequency value to distinguish the frequency components of the data signal. The calculation formula is:

[0026] ;

[0027] In the formula, is the first instantaneous frequency, is the second instantaneous frequency, is the frequency change rate of the upper envelope sequence, is the frequency change rate of the lower envelope sequence, which is obtained by interpolating the square and phase inverse of the envelope sequence interpolation data. is the instantaneous amplitude of the upper envelope sequence, is the instantaneous amplitude of the lower envelope sequence;

[0028] The local cutoff frequency is rearranged, specifically, the local cutoff frequency is calculated and rearranged, and the calculation formula is:

[0029] ;

[0030] In the formula, is the rearranged local cutoff frequency, is the first instantaneous frequency, is the second instantaneous frequency;

[0031] Calculate the filtered data, specifically, calculate the filtered data signal according to the rearranged local cutoff frequency, and the calculation formula is:

[0032] ;

[0033] In the formula, is the filtered data signal, is the rearranged local cutoff frequency;

[0034] Smoothing filtering, specifically, performing smoothing filtering according to the filtered data signal, the local maximum value sequence and the local minimum value sequence to obtain a smoothing filtered data approximate signal m(t);

[0035] Construct a decomposition threshold, specifically, set a bandwidth threshold, and calculate the decomposition threshold. When the decomposition threshold is not greater than the bandwidth threshold, the original signal data in the optimized power load forecast data is directly used as the output of the intrinsic mode function. When the decomposition threshold is greater than the bandwidth threshold, the difference X(t)-m(t) between the original signal data in the optimized power load forecast data and the approximate signal of the smoothing filter data is used as the new original signal input, and the above-mentioned steps of instantaneous data extraction to smoothing filtering are repeated. The calculation formula of the decomposition threshold is:

[0036] ;

[0037] In the formula, is the decomposition threshold, Bw(t) is the instantaneous bandwidth, is the weighted average instantaneous frequency;

[0038] Overlapping data integration is specifically to recursively apply the above instantaneous data extraction to the operation of constructing the decomposition threshold to obtain the integrated eigenmode function output, and the calculation formula is:

[0039] ;

[0040] In the formula, is the integrated IMF output, n is the total number of IMF outputs, i is the IMF index, is the output of the i-th eigenmode function;

[0041] Construct a sample entropy operator to calculate the sample entropy of the time-varying filtering decomposition data, specifically including the following steps:

[0042] Set sample entropy parameters, specifically, set the initial values ​​of reconstruction dimension and similarity tolerance;

[0043] Calculate the sequence matching probability, specifically, calculate the probability that the upper envelope sequence and the lower envelope sequence match in the first reconstruction dimension and the second reconstruction dimension,

[0044] Calculate sample entropy to represent sequence complexity. The calculation formula is:

[0045] ;

[0046] Where SE is the sample entropy, is the probability that the upper envelope sequence and the lower envelope sequence match in the second reconstruction dimension, is the probability that the upper envelope sequence and the lower envelope sequence match in the first reconstruction dimension, v is the reconstruction dimension, is the similarity tolerance;

[0047] Overlapping feature extraction is specifically to construct the sample entropy operator to represent the temporal complexity of the time-varying filtering decomposition data, and to construct the time-varying filtering modal decomposition operator in combination with the sample entropy operator to extract overlapping signal data features to obtain overlapping frequency-division feature data.

[0048] Furthermore, the overlapping feature decomposition is used to decompose independent features of different frequency bands from the overlapping power load data, specifically by using an empirical wavelet transformation method combining variational mode decomposition and high-order window smoothing to perform overlapping feature decomposition, and extract decomposed feature data from the optimized power load forecast data;

[0049] The empirical wavelet transformation method combining variational mode decomposition and high-order window smoothing specifically includes a variational mode decomposition adaptive segmentation operator, a high-order window smoothing wavelet filtering operator and a marginal spectrum feature extraction operator;

[0050] The variational mode decomposition adaptive segmentation operator is used to improve the adaptability of the frequency band segmentation of the traditional method and optimize the adaptability and accuracy of the local characteristics of the signal;

[0051] The high-order window smoothing wavelet filter operator is used to perform frequency band filtering transition using a smoothly transitioned cosine window function to optimize the structure of a traditional filter;

[0052] The marginal spectrum feature extraction operator is used to calculate the instantaneous amplitude and frequency by combining the Hilbert transform and the marginal spectrum analysis method to optimize the meticulousness of feature extraction;

[0053] The step of performing overlapping feature decomposition by using an empirical wavelet transformation method combining variational mode decomposition and high-order window smoothing includes:

[0054] Adaptive spectrum segmentation, specifically constructing a variational mode decomposition adaptive segmentation operator to perform adaptive spectrum segmentation to obtain frequency band division data, specifically includes the following steps:

[0055] Variational mode decomposition is used to decompose the optimized power load forecast data into intrinsic mode function outputs, and the calculation formula is:

[0056] ;

[0057] In the formula, is the decomposition target intrinsic mode function, Is to seek The joint minimization function on is used to extract the optimal intrinsic mode function and the corresponding center frequency, where is the kth intrinsic mode function in the set of intrinsic mode functions, is the kth center frequency in the set of center frequencies of all modes, K is the total number of eigenmode functions in the set, k is the eigenmode function index, is the variational operator of the Fourier transform, is the Dickera function, t is the time index, j is the imaginary unit, is the output of the kth eigenmode function at time t, is the rotation factor, ||·||2 is the L2 norm operator;

[0058] The spectrum segmentation is adaptively selected, specifically, the center frequency set {w k}, perform spectrum segmentation adaptive selection to obtain frequency band division data;

[0059] The wavelet filter smoothing reconstruction is specifically to construct a high-order window smoothing wavelet filter operator, optimize the structure of the wavelet change filter according to the frequency band division data, and obtain a smoothing wavelet filter, which specifically includes the following steps:

[0060] The basis function is constructed based on the center frequency set {w k}, for each frequency band [w k-1 , w k ]Construct wavelet basis function and use square cosine function for smooth transition. The calculation formula is:

[0061] ;

[0062] In the formula, is the frequency domain representation of the empirical wavelet basis function, f is the frequency variable, is the kth center frequency in the set of center frequencies of all modes, k is the eigenmode function index;

[0063] The filter construction is specifically to construct a smooth transition filter based on the frequency domain representation of the empirical wavelet basis function obtained by the basis function construction to obtain a smooth wavelet filter. The calculation formula is:

[0064] ;

[0065] In the formula, is the frequency domain representation of the smoothing wavelet filter, f is the frequency variable, is the kth center frequency in the set of center frequencies of all modes, k is the eigenmode function index;

[0066] The empirical wavelet transform is specifically to use the smoothing wavelet filter to perform wavelet transform to obtain decomposed empirical wavelet signal data, and the calculation formula is:

[0067] ;

[0068] In the formula, is the frequency band signal in the kth decomposed empirical wavelet signal data, F -1 is the inverse Fourier transform function, It is the frequency domain representation of the signal data in the optimized power load data. is the frequency domain representation of the smoothing wavelet filter, is the decomposed empirical wavelet signal data, f is the frequency variable;

[0069] Decomposition feature extraction is specifically to construct a marginal spectrum feature extraction operator, and perform decomposition feature extraction in combination with the decomposition empirical wavelet signal data to obtain decomposition feature data, which specifically includes the following steps:

[0070] Hilbert transform, specifically, uses Hilbert transform to calculate the instantaneous amplitude and instantaneous frequency of each frequency band signal in the empirical wavelet signal data. The calculation formula is:

[0071] ;

[0072] In the formula, is the instantaneous amplitude of each frequency band signal in the empirical wavelet signal data, is the instantaneous frequency of each frequency band signal in the empirical wavelet signal data, H(·) is the Hilbert transform operation representation function, is the frequency band signal in the kth decomposed empirical wavelet signal data, arg(·) is the phase angle calculation function;

[0073] Marginal spectrum analysis, specifically, calculates the frequency characteristics of the band signal through marginal spectrum analysis. The calculation formula is:

[0074] ;

[0075] In the formula, is the frequency characteristic of the band signal corresponding to the kth center frequency, T is the total duration of the band signal, is the instantaneous amplitude of each frequency band signal in the empirical wavelet signal data, is the instantaneous frequency of each frequency band signal in the empirical wavelet signal data;

[0076] The energy calculation of the frequency band signal is specifically to calculate the energy characteristics of the frequency band signal. The calculation formula is:

[0077] ;

[0078] In the formula, is the energy characteristic of the frequency band signal corresponding to the kth center frequency, is the frequency band signal in the k-th decomposed empirical wavelet signal data, |·| is the modulus operator;

[0079] The average frequency calculation is specifically to calculate the average frequency characteristics of the frequency band signal. The calculation formula is:

[0080] ;

[0081] In the formula, is the average frequency characteristic of the band signal corresponding to the kth center frequency, f is the frequency variable, is the frequency domain representation of the band signal corresponding to the kth center frequency;

[0082] Decomposition and extraction, specifically, performing decomposition feature extraction in combination with the decomposition empirical wavelet signal data through the Hilbert transform, the marginal spectrum analysis, the frequency band signal energy calculation and the average frequency calculation, to obtain decomposition feature data;

[0083] The decomposition feature data specifically refers to high- and low-frequency decomposition feature data.

[0084] Furthermore, the load classification prediction is used to construct a frequency band domain model to perform power load classification prediction, specifically, based on the power load prediction original data, original overlapping frequency division feature data and decomposed feature data, a multi-path cyclic bidirectional long short-term memory network combined with convolution time feature extraction is used to perform load classification prediction to obtain power load prediction data;

[0085] The multi-path recurrent bidirectional long short-term memory network combined with convolutional temporal feature extraction specifically includes a convolutional subnet, a recurrent neural subnet and a bidirectional long short-term memory subnet;

[0086] The convolutional subnet is used to extract time features, weather features and special features from the original data;

[0087] The recurrent neural subnet is used to process high-frequency feature data in the decomposed features;

[0088] The bidirectional long short-term memory network is used to process low-frequency feature data and overlapping frequency-divided feature data in the decomposed features;

[0089] The step of using a multi-path cyclic bidirectional long short-term memory network combined with convolutional time feature extraction to perform load classification prediction and obtain power load prediction data includes:

[0090] Convolutional feature extraction, specifically, constructing a convolutional feature extraction subnet including only a convolutional layer, a pooling layer, and a fully connected layer, and using one-dimensional convolution for feature extraction, and extracting time series comprehensive feature data from the power load forecast raw data by constructing the convolutional feature extraction subnet;

[0091] The time series comprehensive characteristic data specifically includes time series time characteristic data, weather time characteristic data and special date time characteristic data;

[0092] Recurrent network high-frequency feature prediction, specifically, constructing a recurrent neural network subnet including an input layer, a hidden layer, and an output layer, and using the recurrent neural network subnet, combined with the time series comprehensive feature data, to process the feature data in the high-frequency feature data set to obtain high-frequency feature data prediction data;

[0093] Bidirectional long short-term memory subnetwork prediction, specifically, constructing two layers of bidirectional long short-term memory subnetworks, respectively used to process the feature data in the low-frequency feature data set and the overlapping frequency-divided feature data, and combining the time series comprehensive feature data to perform hierarchical prediction to obtain low-frequency feature data prediction data and overlapping feature data prediction data;

[0094] The long short-term memory subnet specifically includes an input gate, an output gate, a forget gate and a storage unit;

[0095] The model output integration decision is specifically to weight the prediction results of the high-frequency feature data prediction data, the low-frequency feature data prediction data and the overlapping feature data prediction data to obtain the model integrated prediction data, and the calculation formula is:

[0096] ;

[0097] In the formula, Y is the model integrated prediction data, is the high-frequency prediction weight, is the high-frequency feature data prediction data, is the low-frequency prediction weight, is the low-frequency feature data prediction data, is the overlap prediction weight, It is the overlapping feature data prediction data;

[0098] The power load classification prediction model training is specifically carried out through the convolution feature extraction, the recurrent network high-frequency feature prediction, the bidirectional long short-term memory subnet prediction and the model output integrated decision, to obtain the power load prediction model Model ECP ;

[0099] Power load forecasting, specifically using the power load forecasting model Model ECP , combining the original power load prediction data, the original overlapping frequency-division feature data and the decomposed feature data, to perform power load prediction and obtain power load prediction data.

[0100] Furthermore, the predicted value visualization is used to construct a visual data reference for the load forecasting results, specifically by using a data visualization module to draw the power load forecasting data into an interactive visual image to obtain predicted value visualization image analysis data.

[0101] Furthermore, the model accuracy monitoring is used to implement model prediction and monitor the prediction accuracy of the model, and optimize the model prediction performance. Specifically, the predicted value visualization image analysis data is manually analyzed, and the accuracy of the model is optimized based on the comparison between the predicted situation and the actual situation. The optimal prediction model is obtained through accuracy optimization, and the optimal prediction model is used to perform the power load forecasting task to obtain the final power load forecasting reference.

[0102] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0103] (1) In the existing power load forecasting system, the quality of power load feature extraction largely determines the final prediction effect. The frequency band data of power load forecasting data overlaps with each other and has complex rules. Therefore, the feature engineering design for overlapping power load data needs to consider both overlap and separation. This solution creatively adopts the dual-path feature extraction idea of ​​overlapping feature extraction and overlapping feature decomposition to extract overlapping features and decompose features. By designing adaptive extraction algorithms respectively, the quality of feature data is effectively improved and the overall performance of the system is optimized.

[0104] (2) In view of the technical problem that in the existing overlapping feature extraction methods, the frequency band data with aliased modes has poor performance in processing nonlinear and non-smooth signals in the traditional empirical mode decomposition method, resulting in low extraction efficiency and affecting the subsequent prediction effect, this scheme creatively uses a time-varying filter to improve the traditional empirical mode decomposition and optimize the performance of feature extraction;

[0105] (3) In the existing overlapping feature decomposition methods, there is a fundamental problem that the empirical mode decomposition method cannot solve the mutual connection and interference between frequency bands by only using time-varying filters, and the traditional wavelet transform method for decomposition feature extraction has technical problems such as insufficient spectrum segmentation accuracy and the need to improve smoothness and meticulousness. This scheme creatively adopts an improved wavelet transform method that combines variational mode decomposition, smoothing function and marginal spectrum analysis, optimizes the overall process of decomposition feature extraction, and further improves the quality of features;

[0106] (4) In view of the technical problem that the existing specific methods of power load forecasting cannot effectively handle complex high-frequency, low-frequency and overlapping features, it is necessary to design an appropriate hybrid model for effective classification prediction. This scheme creatively adopts a multi-path recurrent bidirectional long short-term memory network combined with convolutional time feature extraction to perform classification load forecasting. The convolutional subnet further combines the time features to reduce the neglect of environmental conditions and other objective conditions. At the same time, the recurrent subnet is used to process high-frequency data, and the bidirectional long short-term subnet is used to process low-frequency and overlapping data. The overall architecture of the prediction model is optimized, and at the same time, a practical basis and exploration experience are provided for the further refinement of power load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 A schematic diagram of the structure of an artificial intelligence-based power load forecasting system provided by the present invention;

[0108] Figure 2 A schematic flow chart of steps executed by an artificial intelligence-based power load forecasting system provided by the present invention;

[0109] Figure 3 A flowchart of the steps performed for overlapping feature extraction in the feature processing module;

[0110] Figure 4 A flowchart of the steps performed by overlapping feature decomposition in the feature processing module;

[0111] Figure 5 A schematic diagram of the flow chart of the steps executed by the prediction driving module;

[0112] Figure 6 A visualization diagram of the predicted values ​​of high-frequency feature data predicted by the prediction application module;

[0113] Figure 7 A visualization diagram of the predicted values ​​of low-frequency feature data predicted by the prediction application module;

[0114] Figure 8 A visualization diagram of the predicted values ​​of the overlapping feature data predicted data drawn for the prediction application module;

[0115] Fig. 9 A visualization of the predicted values ​​of the model-integrated prediction data drawn for the prediction application module.

[0116] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0117] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0118] Example 1, see Figure 1 and Figure 2 , the present invention provides an artificial intelligence-based power load forecasting system, including a data acquisition module, a feature processing module, a forecasting driving module, a forecasting application module and a monitoring feedback module;

[0119] The data acquisition module is used for data collection and processing, and obtains the original power load forecast data and the optimized power load forecast data through data collection and processing, and sends the optimized power load forecast data to the feature processing module, and sends the original power load forecast data to the forecast driving module;

[0120] The feature processing module is used for overlapping feature extraction and overlapping feature decomposition, and obtains overlapping frequency-division feature data and decomposed feature data through overlapping feature extraction and overlapping feature decomposition, and sends the overlapping frequency-division feature data and decomposed feature data to the prediction driving module;

[0121] The prediction driving module is used for load classification prediction, obtains power load prediction data through load classification prediction, and sends the power load prediction data to the prediction application module and the monitoring feedback module;

[0122] The prediction application module is used for predictive value visualization, obtains predictive value visualization image analysis data through predictive value visualization, and sends the predictive value visualization image analysis data to the monitoring feedback module;

[0123] The monitoring feedback module is used for model accuracy monitoring, and obtains a power load prediction reference through model accuracy monitoring.

[0124] By performing the above operations, in the existing power load forecasting system, the quality of power load feature extraction largely determines the final prediction effect, and the frequency band data of the power load forecasting data has a special problem of overlapping and complex rules. Therefore, the feature engineering design for overlapping power load data needs to consider both overlapping and separable technical issues. This scheme creatively adopts the dual-path feature extraction idea of ​​overlapping feature extraction and overlapping feature decomposition to perform overlapping feature extraction and decomposed feature extraction. By designing adaptive extraction algorithms respectively, the feature data quality is effectively improved and the overall performance of the system is optimized.

[0125] Embodiment 2, this embodiment is based on the above embodiment, see Figure 1 and Figure 2 The data collection and processing is used to collect the original data required for power load forecasting, specifically, to obtain the original data for power load forecasting from the power system and environmental sensor records through data collection, and to obtain optimized power load forecasting data by optimizing the original data for power load forecasting;

[0126] The power load forecast raw data specifically includes historical power load data, weather environment data and special date data;

[0127] The historical power load data specifically includes hourly, daily, weekly and monthly historical power load time series data;

[0128] The weather environment data specifically includes temperature data, humidity data and wind speed data;

[0129] The special date data specifically refers to the date record data of special holidays and special large-scale events, which is used to record the power fluctuations on special dates;

[0130] The optimization process specifically includes outlier processing and time series decomposition operations;

[0131] The outlier processing specifically refers to processing missing values ​​and outliers in the data and performing data standardization;

[0132] The time series decomposition specifically refers to decomposing the data into time series according to trend, seasonality and residual components.

[0133] Embodiment 3, this embodiment is based on the above embodiment, see Figure 2 and Figure 3 The overlapping feature extraction is used to extract frequency-division features from power load data that overlap in different frequency bands. Specifically, based on the optimized power load forecast data, a time-varying filtering modal decomposition method combined with sample entropy is used to extract overlapping features to obtain overlapping frequency-division feature data;

[0134] The time-varying filtering modal decomposition method combined with sample entropy specifically includes a time series sample entropy operator and a time-varying filtering modal decomposition operator;

[0135] The step of extracting overlapping features by using a time-varying filtering modal decomposition method combined with sample entropy to obtain overlapping frequency-division feature data includes:

[0136] Construct a time-varying filter modal decomposition operator, decompose the original data, and obtain the time-varying filter decomposition data. The specific steps include:

[0137] Instantaneous data extraction, specifically using the Hilbert transform method, extracts the instantaneous amplitude and instantaneous phase of the optimized power load forecast data. The calculation formula is:

[0138] ;

[0139] Where A(t) is the instantaneous amplitude of the optimized power load forecast data, t is the time index, is the instantaneous phase, X(t) is the original signal data in the optimized power load forecasting data, It is the data after the optimized power load forecast data has been Hilbert transformed;

[0140] Constructing a local extreme value sequence for searching for local maximum values ​​and local minimum values ​​in the original data, specifically identifying a local maximum value sequence and a local minimum value sequence according to the instantaneous amplitude A(t) of the optimized power load forecast data;

[0141] The average envelope value is calculated by interpolating the local maximum value sequence and the local minimum value sequence to obtain the upper envelope sequence interpolation data and the lower envelope sequence interpolation data, and obtaining the upper envelope sequence instantaneous amplitude and the lower envelope sequence instantaneous amplitude. The calculation formula is:

[0142] ;

[0143] In the formula, is the instantaneous amplitude of the upper envelope sequence, is the instantaneous amplitude of the lower envelope sequence, is the upper envelope sequence interpolation data, is the lower envelope sequence interpolation data;

[0144] Calculate the instantaneous frequency value to distinguish the frequency components of the data signal. The calculation formula is:

[0145] ;

[0146] In the formula, is the first instantaneous frequency, is the second instantaneous frequency, is the frequency change rate of the upper envelope sequence, is the frequency change rate of the lower envelope sequence, which is obtained by interpolating the square and phase inverse of the envelope sequence interpolation data. is the instantaneous amplitude of the upper envelope sequence, is the instantaneous amplitude of the lower envelope sequence;

[0147] The local cutoff frequency is rearranged, specifically, the local cutoff frequency is calculated and rearranged, and the calculation formula is:

[0148] ;

[0149] In the formula, is the rearranged local cutoff frequency, is the first instantaneous frequency, is the second instantaneous frequency;

[0150] Calculate the filtered data, specifically, calculate the filtered data signal according to the rearranged local cutoff frequency, and the calculation formula is:

[0151] ;

[0152] In the formula, is the filtered data signal, is the rearranged local cutoff frequency;

[0153] Smoothing filtering, specifically, performing smoothing filtering according to the filtered data signal, the local maximum value sequence and the local minimum value sequence to obtain a smoothing filtered data approximate signal m(t);

[0154] Construct a decomposition threshold, specifically, set a bandwidth threshold, and calculate the decomposition threshold. When the decomposition threshold is not greater than the bandwidth threshold, the original signal data in the optimized power load forecast data is directly used as the output of the intrinsic mode function. When the decomposition threshold is greater than the bandwidth threshold, the difference X(t)-m(t) between the original signal data in the optimized power load forecast data and the approximate signal of the smoothing filter data is used as the new original signal input, and the above-mentioned steps of instantaneous data extraction to smoothing filtering are repeated. The calculation formula of the decomposition threshold is:

[0155] ;

[0156] In the formula, is the decomposition threshold, Bw(t) is the instantaneous bandwidth, is the weighted average instantaneous frequency;

[0157] The bandwidth threshold is specifically set to 0.1;

[0158] Overlapping data integration is specifically to recursively apply the above instantaneous data extraction to the operation of constructing the decomposition threshold to obtain the integrated eigenmode function output, and the calculation formula is:

[0159] ;

[0160] In the formula, is the integrated IMF output, n is the total number of IMF outputs, i is the IMF index, is the output of the i-th eigenmode function;

[0161] Construct a sample entropy operator to calculate the sample entropy of the time-varying filtering decomposition data, specifically including the following steps:

[0162] Set sample entropy parameters, specifically, set the initial values ​​of reconstruction dimension and similarity tolerance;

[0163] Calculate the sequence matching probability, specifically, calculate the probability that the upper envelope sequence and the lower envelope sequence match in the first reconstruction dimension and the second reconstruction dimension,

[0164] Calculate sample entropy to represent sequence complexity. The calculation formula is:

[0165] ;

[0166] Where SE is the sample entropy, is the probability that the upper envelope sequence and the lower envelope sequence match in the second reconstruction dimension, is the probability that the upper envelope sequence and the lower envelope sequence match in the first reconstruction dimension, v is the reconstruction dimension, is the similarity tolerance;

[0167] Overlapping feature extraction is specifically to construct the sample entropy operator to represent the temporal complexity of the time-varying filtering decomposition data, and to construct the time-varying filtering modal decomposition operator in combination with the sample entropy operator to extract overlapping signal data features to obtain overlapping frequency-division feature data.

[0168] By performing the above operations, in order to solve the technical problem that in the existing overlapping feature extraction methods, there are frequency band data with aliased modes in the traditional empirical mode decomposition method, and the performance of processing nonlinear and non-smooth signals is poor, resulting in low extraction efficiency and affecting subsequent prediction effects, this scheme creatively uses a time-varying filter to improve the traditional empirical mode decomposition and optimize the feature extraction performance.

[0169] Embodiment 4, this embodiment is based on the above embodiment, see Figure 2 and Figure 4The overlapping feature decomposition is used to decompose independent features of different frequency bands from the overlapping power load data, specifically using an empirical wavelet transformation method combining variational mode decomposition and high-order window smoothing to perform overlapping feature decomposition, and extract decomposed feature data from the optimized power load forecast data;

[0170] The empirical wavelet transformation method combining variational mode decomposition and high-order window smoothing specifically includes a variational mode decomposition adaptive segmentation operator, a high-order window smoothing wavelet filtering operator and a marginal spectrum feature extraction operator;

[0171] The variational mode decomposition adaptive segmentation operator is used to improve the adaptability of the frequency band segmentation of the traditional method and optimize the adaptability and accuracy of the local characteristics of the signal;

[0172] The high-order window smoothing wavelet filter operator is used to perform frequency band filtering transition using a smoothly transitioned cosine window function to optimize the structure of a traditional filter;

[0173] The marginal spectrum feature extraction operator is used to calculate the instantaneous amplitude and frequency by combining the Hilbert transform and the marginal spectrum analysis method to optimize the meticulousness of feature extraction;

[0174] The step of performing overlapping feature decomposition by using an empirical wavelet transformation method combining variational mode decomposition and high-order window smoothing includes:

[0175] Adaptive spectrum segmentation, specifically constructing a variational mode decomposition adaptive segmentation operator to perform adaptive spectrum segmentation to obtain frequency band division data, specifically includes the following steps:

[0176] Variational mode decomposition is used to decompose the optimized power load forecast data into intrinsic mode function outputs, and the calculation formula is:

[0177] ;

[0178] In the formula, is the decomposition target intrinsic mode function, Is to seek The joint minimization function on is used to extract the optimal intrinsic mode function and the corresponding center frequency, where is the kth intrinsic mode function in the set of intrinsic mode functions, is the kth center frequency in the set of center frequencies of all modes, K is the total number of eigenmode functions in the set, k is the eigenmode function index, is the variational operator of the Fourier transform, is the Dickera function, t is the time index, j is the imaginary unit, is the output of the kth eigenmode function at time t, is the rotation factor, ||·||2 is the L2 norm operator;

[0179] The spectrum segmentation is adaptively selected, specifically, the center frequency set {w k}, perform spectrum segmentation adaptive selection to obtain frequency band division data;

[0180] The wavelet filter smoothing reconstruction is specifically to construct a high-order window smoothing wavelet filter operator, optimize the structure of the wavelet change filter according to the frequency band division data, and obtain a smoothing wavelet filter, which specifically includes the following steps:

[0181] The basis function is constructed based on the center frequency set {w k}, for each frequency band [w k-1 , w k ]Construct wavelet basis function and use square cosine function for smooth transition. The calculation formula is:

[0182] ;

[0183] In the formula, is the frequency domain representation of the empirical wavelet basis function, f is the frequency variable, is the kth center frequency in the set of center frequencies of all modes, k is the eigenmode function index;

[0184] The filter construction is specifically to construct a smooth transition filter based on the frequency domain representation of the empirical wavelet basis function obtained by the basis function construction to obtain a smooth wavelet filter. The calculation formula is:

[0185] ;

[0186] In the formula, is the frequency domain representation of the smoothing wavelet filter, f is the frequency variable, is the kth center frequency in the set of center frequencies of all modes, k is the eigenmode function index;

[0187] The empirical wavelet transform is specifically to use the smoothing wavelet filter to perform wavelet transform to obtain decomposed empirical wavelet signal data, and the calculation formula is:

[0188] ;

[0189] In the formula, is the frequency band signal in the kth decomposed empirical wavelet signal data, F -1 is the inverse Fourier transform function, It is the frequency domain representation of the signal data in the optimized power load data. is the frequency domain representation of the smoothing wavelet filter, is the decomposed empirical wavelet signal data, f is the frequency variable;

[0190] Decomposition feature extraction is specifically to construct a marginal spectrum feature extraction operator, and perform decomposition feature extraction in combination with the decomposition empirical wavelet signal data to obtain decomposition feature data, which specifically includes the following steps:

[0191] Hilbert transform, specifically, uses Hilbert transform to calculate the instantaneous amplitude and instantaneous frequency of each frequency band signal in the empirical wavelet signal data. The calculation formula is:

[0192] ;

[0193] In the formula, is the instantaneous amplitude of each frequency band signal in the empirical wavelet signal data, is the instantaneous frequency of each frequency band signal in the empirical wavelet signal data, H(·) is the Hilbert transform operation representation function, is the frequency band signal in the kth decomposed empirical wavelet signal data, arg(·) is the phase angle calculation function;

[0194] Marginal spectrum analysis, specifically, calculates the frequency characteristics of the band signal through marginal spectrum analysis. The calculation formula is:

[0195] ;

[0196] In the formula, is the frequency characteristic of the band signal corresponding to the kth center frequency, T is the total duration of the band signal, is the instantaneous amplitude of each frequency band signal in the empirical wavelet signal data, is the instantaneous frequency of each frequency band signal in the empirical wavelet signal data;

[0197] The energy calculation of the frequency band signal is specifically to calculate the energy characteristics of the frequency band signal. The calculation formula is:

[0198] ;

[0199] In the formula, is the energy characteristic of the frequency band signal corresponding to the kth center frequency, is the frequency band signal in the k-th decomposed empirical wavelet signal data, |·| is the modulus operator;

[0200] The average frequency calculation is specifically to calculate the average frequency characteristics of the frequency band signal. The calculation formula is:

[0201] ;

[0202] In the formula, is the average frequency characteristic of the band signal corresponding to the kth center frequency, f is the frequency variable, is the frequency domain representation of the band signal corresponding to the kth center frequency;

[0203] Decomposition and extraction, specifically, performing decomposition feature extraction in combination with the decomposition empirical wavelet signal data through the Hilbert transform, the marginal spectrum analysis, the frequency band signal energy calculation and the average frequency calculation, to obtain decomposition feature data;

[0204] The decomposition feature data specifically refers to high- and low-frequency decomposition feature data, and the calculation formula is:

[0205] ;

[0206] In the formula, IMF HF is a set of high-frequency feature data, N is the total number of high-frequency feature data, IMF LF is the low-frequency feature data set, K is the total number of eigenmode function sets, which is used to represent the total number of decomposed feature data. is the first high-frequency feature data, It is the first low-frequency feature data;

[0207] The total number K in the intrinsic mode function set is specifically 7, the total number N of high-frequency feature data is specifically 4, and the total number of low-frequency feature data is 3.

[0208] By performing the above operations, in the existing overlapping feature decomposition methods, there is a fundamental problem that only using time-varying filters to change the empirical mode decomposition method cannot solve the mutual series connection and interference between frequency bands, and the traditional wavelet transform method for decomposition feature extraction has technical problems such as insufficient spectrum segmentation accuracy and the need to improve smoothness and meticulousness. This scheme creatively adopts an improved wavelet transform method that combines variational mode decomposition, smoothing function and marginal spectrum analysis, optimizes the overall process of decomposition feature extraction, and further improves the quality of features.

[0209] Embodiment 5, this embodiment is based on the above embodiment, see Figure 1 , Figure 2 and Figure 5 The load classification prediction is used to construct a frequency band domain model to perform power load classification prediction, specifically, based on the power load prediction original data, original overlapping frequency division feature data and decomposed feature data, a multi-path cyclic bidirectional long short-term memory network combined with convolution time feature extraction is used to perform load classification prediction to obtain power load prediction data;

[0210] The multi-path recurrent bidirectional long short-term memory network combined with convolutional temporal feature extraction specifically includes a convolutional subnet, a recurrent neural subnet and a bidirectional long short-term memory subnet;

[0211] The convolutional subnet is used to extract time features, weather features and special features from the original data;

[0212] The recurrent neural subnet is used to process high-frequency feature data in the decomposed features;

[0213] The bidirectional long short-term memory network is used to process low-frequency feature data and overlapping frequency-divided feature data in the decomposed features;

[0214] The step of using a multi-path cyclic bidirectional long short-term memory network combined with convolutional time feature extraction to perform load classification prediction and obtain power load prediction data includes:

[0215] Convolutional feature extraction, specifically, constructing a convolutional feature extraction subnet including only a convolutional layer, a pooling layer, and a fully connected layer, and using one-dimensional convolution for feature extraction, and extracting time series comprehensive feature data from the power load forecast raw data by constructing the convolutional feature extraction subnet;

[0216] The time series comprehensive characteristic data specifically includes time series time characteristic data, weather time characteristic data and special date time characteristic data;

[0217] Recurrent network high-frequency feature prediction, specifically, constructing a recurrent neural network subnet including an input layer, a hidden layer, and an output layer, and using the recurrent neural network subnet, combined with the time series comprehensive feature data, to process the feature data in the high-frequency feature data set to obtain high-frequency feature data prediction data;

[0218] Bidirectional long short-term memory subnetwork prediction, specifically, constructing two layers of bidirectional long short-term memory subnetworks, respectively used to process the feature data in the low-frequency feature data set and the overlapping frequency-divided feature data, and combining the time series comprehensive feature data to perform hierarchical prediction to obtain low-frequency feature data prediction data and overlapping feature data prediction data;

[0219] The long short-term memory subnet specifically includes an input gate, an output gate, a forget gate and a storage unit;

[0220] The specific steps of constructing the bidirectional long short-term memory subnetwork include:

[0221] Construct the forward long short-term memory subnet, and the calculation formula is:

[0222] ;

[0223] In the formula, is the final hidden state of the forward LSTM subnetwork, LSTM(·) is the LSTM subnetwork representation function, is the forward weight, is the previous hidden state of the forward long short-term memory subnetwork, and t is the time index;

[0224] Construct the backward long short-term memory subnet, and the calculation formula is:

[0225] ;

[0226] In the formula, is the final hidden state of the backward LSTM subnetwork, is the backward weight, is the previous hidden state of the backward long short-term memory subnetwork;

[0227] Construct a bidirectional LSTM subnetwork. Specifically, the final hidden state of the forward LSTM subnetwork and the final hidden state of the backward LSTM subnetwork are weighted summed to obtain the predicted output of the bidirectional LSTM subnetwork. The calculation formula is:

[0228] ;

[0229] In the formula, y t is the prediction output of the bidirectional long short-term memory subnet, a f is the forward weight, a b is the backward weight, is the final hidden state of the forward LSTM subnetwork, is the final hidden state of the backward LSTM subnetwork, b t is the bias term;

[0230] The model output integration decision is specifically to weight the prediction results of the high-frequency feature data prediction data, the low-frequency feature data prediction data and the overlapping feature data prediction data to obtain the model integrated prediction data, and the calculation formula is:

[0231] ;

[0232] In the formula, Y is the model integrated prediction data, is the high-frequency prediction weight, is the high-frequency feature data prediction data, is the low-frequency prediction weight, is the low-frequency feature data prediction data, is the overlap prediction weight, It is the overlapping feature data prediction data;

[0233] The power load classification prediction model training is specifically carried out through the convolution feature extraction, the recurrent network high-frequency feature prediction, the bidirectional long short-term memory subnet prediction and the model output integrated decision, to obtain the power load prediction model Model ECP;

[0234] Power load forecasting, specifically using the power load forecasting model Model ECP , combining the original power load prediction data, the original overlapping frequency-division feature data and the decomposed feature data, to perform power load prediction and obtain power load prediction data.

[0235] By performing the above operations, in view of the technical problem that in the existing specific methods of power load forecasting, simple model mixing cannot effectively handle complex high-frequency, low-frequency and overlapping features, and therefore it is necessary to design an appropriate hybrid model for effective classification prediction, this solution creatively adopts a multi-path recurrent bidirectional long short-term memory network combined with convolutional time feature extraction to perform classification load forecasting, and further combines the time features through the convolutional subnet to reduce the neglect of environmental conditions and other objective conditions. At the same time, a recurrent subnet is used to process high-frequency data, and a bidirectional long short-term subnet is used to process low-frequency and overlapping data, which optimizes the overall architecture of the prediction model and provides a practical basis and exploration experience for further refinement of power load forecasting.

[0236] Embodiment 6, this embodiment is based on the above embodiment, see Figure 1 and Figure 2 The predicted value visualization is used to construct a visual data reference for the load forecasting results. Specifically, a data visualization module is used to draw the power load forecasting data into an interactive visual image to obtain predicted value visualization image analysis data.

[0237] Embodiment 7, this embodiment is based on the above embodiment, see Figure 6 , Figure 6 A visualization diagram of the predicted values ​​of high-frequency feature data prediction data drawn for the prediction application module, where the horizontal coordinate is the time index and the vertical coordinate is the specific high-frequency feature predicted power load prediction value.

[0238] Embodiment 8, this embodiment is based on the above embodiment, see Figure 7 , Figure 7 A visualization diagram of the predicted values ​​of low-frequency feature data prediction data drawn for the prediction application module, where the horizontal coordinate is the time index and the vertical coordinate is the specific low-frequency feature predicted power load prediction value.

[0239] Embodiment 9, this embodiment is based on the above embodiment, see Figure 8 , Figure 8 A visualization diagram of the predicted values ​​of overlapping feature data prediction data drawn for the prediction application module, where the horizontal coordinate is the time index and the vertical coordinate is the specific overlapping feature prediction power load prediction value.

[0240] Embodiment 10, this embodiment is based on the above embodiment, see Fig. 9 , Fig. 9 A visualization diagram of the predicted values ​​of the model integrated prediction data drawn for the prediction application module, where the horizontal coordinate is the time index and the vertical coordinate is the specific model integrated prediction power load prediction value.

[0241] Embodiment 11, this embodiment is based on the above embodiment, see Figure 1 and Figure 2 The model accuracy monitoring is used to implement model prediction and monitor the prediction accuracy of the model, and optimize the model prediction performance. Specifically, the predicted value visualization image analysis data is manually analyzed, and the accuracy of the model is optimized based on the comparison between the predicted situation and the actual situation. The optimal prediction model is obtained through accuracy optimization, and the optimal prediction model is used to perform the power load prediction task to obtain the final power load prediction reference.

[0242] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process and method including a series of elements include not only those elements, but also other elements not explicitly listed, or also include elements inherent to such process and method.

[0243] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0244] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. An artificial intelligence-based power load forecasting system, Features: It includes data acquisition module, feature processing module, prediction drive module, prediction application module and monitoring feedback module; The data acquisition module is used to collect the original data required for power load prediction, obtain the power load prediction original data and optimized power load prediction data through data collection and processing, and send the optimized power load prediction data to the feature processing module, and send the power load prediction original data to the prediction driving module; The power load forecast raw data specifically includes historical power load data, weather environment data and special date data; The feature processing module is used to extract frequency-division features from power load data that overlap in different frequency bands and decompose independent features of different frequency bands from the overlapping power load data, obtain overlapping frequency-division feature data and decomposed feature data through overlapping feature extraction and overlapping feature decomposition, and send the overlapping frequency-division feature data and decomposed feature data to the prediction driving module; The overlapping feature extraction specifically adopts a time-varying filtering modal decomposition method combined with sample entropy; The overlapping feature extraction is used to extract frequency-division features from power load data that overlaps in different frequency bands. Specifically, overlapping feature extraction is performed based on the optimized power load forecast data using a time-varying filtering modal decomposition method combined with sample entropy to obtain overlapping frequency-division feature data. The time-varying filtering modal decomposition method combined with sample entropy specifically includes a time series sample entropy operator and a time-varying filtering modal decomposition operator; The step of extracting overlapping features by using a time-varying filtering modal decomposition method combined with sample entropy to obtain overlapping frequency-division feature data includes: Construct a time-varying filter modal decomposition operator to decompose the original data and obtain the time-varying filter decomposition data; Construct a sample entropy operator to calculate the sample entropy of the time-varying filtering decomposition data; Overlap feature extraction, specifically, by constructing the sample entropy operator, performing temporal complexity representation of the time-varying filtering decomposition data, and by constructing the time-varying filtering modal decomposition operator, combined with the sample entropy operator, performing feature extraction of overlapping signal data, and obtaining overlapping frequency-division feature data; The time-varying filtering modal decomposition operator is constructed to decompose the original data to obtain the time-varying filtering decomposition data, and the specific steps include: Instantaneous data extraction, specifically using the Hilbert transform method, extracts the instantaneous amplitude and instantaneous phase of the optimized power load forecast data. The calculation formula is: Where A(t) is the instantaneous amplitude of the optimized power load forecast data, t is the time index, ψ(t) is the instantaneous phase, X(t) is the original signal data in the optimized power load forecast data, It is the data after the optimized power load forecast data has been Hilbert transformed; Constructing a local extreme value sequence for searching for local maximum values ​​and local minimum values ​​in the original data, specifically identifying a local maximum value sequence and a local minimum value sequence according to the instantaneous amplitude A(t) of the optimized power load forecast data; The average envelope value is calculated by interpolating the local maximum value sequence and the local minimum value sequence to obtain the upper envelope sequence interpolation data and the lower envelope sequence interpolation data, and obtaining the upper envelope sequence instantaneous amplitude and the lower envelope sequence instantaneous amplitude. The calculation formula is: In the formula, α 1 (t) is the instantaneous amplitude of the upper envelope sequence, α 2 (t) is the instantaneous amplitude of the lower envelope sequence, ε upper (t) is the upper envelope sequence interpolation data, ε lower (t) is the lower envelope sequence interpolation data; Calculate the instantaneous frequency value to distinguish the frequency components of the data signal. The calculation formula is: In the formula, ψ inst1 (t) is the first instantaneous frequency, ψ inst2 (t) is the second instantaneous frequency, Δ upper (t) is the frequency change rate of the upper envelope sequence, Δ lower (t) is the frequency change rate of the lower envelope sequence, which is obtained by interpolating the square and phase inverse of the envelope sequence interpolation data, α 1 (t) is the instantaneous amplitude of the upper envelope sequence, α 2 (t) is the instantaneous amplitude of the lower envelope sequence; The local cutoff frequency is rearranged, specifically, the local cutoff frequency is calculated and rearranged, and the calculation formula is: In the formula, ψ bias (t) is the rearranged local cutoff frequency, ψ inst1 (t) is the first instantaneous frequency, ψ inst2 (t) is the second instantaneous frequency; Calculate the filtered data, specifically, calculate the filtered data signal according to the rearranged local cutoff frequency, and the calculation formula is: θ(t)=cos[∫ψ bias (t)dt]; Where θ(t) is the filtered data signal, ψ bias (t) is the rearranged local cutoff frequency; Smoothing filtering, specifically, performing smoothing filtering according to the filtered data signal, the local maximum value sequence and the local minimum value sequence to obtain a smoothing filtered data approximate signal m(t); Construct a decomposition threshold, specifically, set a bandwidth threshold, and calculate the decomposition threshold. When the decomposition threshold is not greater than the bandwidth threshold, the original signal data in the optimized power load forecast data is directly used as the output of the intrinsic mode function. When the decomposition threshold is greater than the bandwidth threshold, the difference X(t)-m(t) between the original signal data in the optimized power load forecast data and the approximate signal of the smoothing filter data is used as the new original signal input, and the above-mentioned steps of instantaneous data extraction to smoothing filtering are repeated. The calculation formula of the decomposition threshold is: Where ρ(t) is the decomposition threshold, Bw(t) is the instantaneous bandwidth, φ avg (t) is the weighted average instantaneous frequency; Overlapping data integration is to recursively apply the above instantaneous data extraction to construct the decomposition threshold operation to obtain the integrated eigenmode function output, and the calculation formula is: In the formula, IMF A (t) is the integrated IMF output, n is the total number of IMF outputs, i is the IMF index, IMF i (t) is the output of the i-th eigenmode function; The sample entropy operator is constructed to calculate the sample entropy of the time-varying filtering decomposition data, specifically comprising the following steps: Set sample entropy parameters, specifically, set the initial values ​​of reconstruction dimension and similarity tolerance; Calculate the sequence matching probability, specifically, calculate the probability that the upper envelope sequence and the lower envelope sequence match in the first reconstruction dimension and the second reconstruction dimension, Calculate sample entropy to represent sequence complexity. The calculation formula is: Where SE is the sample entropy, B v+1∈ (t) is the probability that the upper envelope sequence and the lower envelope sequence match in the second reconstruction dimension, B v∈ (t) is the probability that the upper envelope sequence and the lower envelope sequence match in the first reconstruction dimension, v is the reconstruction dimension, and ∈ is the similarity tolerance; The overlapping feature decomposition specifically adopts an empirical wavelet transformation method combining variational mode decomposition and high-order window smoothing; The empirical wavelet transformation method combining variational mode decomposition and high-order window smoothing specifically includes a variational mode decomposition adaptive segmentation operator, a high-order window smoothing wavelet filtering operator and a marginal spectrum feature extraction operator; The variational mode decomposition adaptive segmentation operator is used to improve the adaptability of the frequency band segmentation of the traditional method and optimize the adaptability and accuracy of the local characteristics of the signal; The high-order window smoothing wavelet filter operator is used to perform frequency band filtering transition using a smoothly transitioned cosine window function to optimize the structure of a traditional filter; The marginal spectrum feature extraction operator is used to calculate the instantaneous amplitude and frequency by combining the Hilbert transform and the marginal spectrum analysis method to optimize the meticulousness of feature extraction; The step of performing overlapping feature decomposition by using an empirical wavelet transformation method combining variational mode decomposition and high-order window smoothing includes: Adaptive spectrum segmentation, specifically constructing a variational mode decomposition adaptive segmentation operator to perform adaptive spectrum segmentation to obtain frequency band division data, specifically includes the following steps: Variational mode decomposition is used to decompose the optimized power load forecast data into intrinsic mode function outputs, and the calculation formula is: In the formula, IMF B (t) is the decomposition target intrinsic mode function, is to find {u k ), {w k } is used to extract the optimal intrinsic mode function and the corresponding center frequency, where u k is the kth intrinsic mode function in the set of intrinsic mode functions, w k is the kth center frequency in the set of center frequencies of all modes, K is the total number of eigenmode functions in the set, k is the eigenmode function index, is the variational operator of the Fourier transform, δ(t) is the Dickler function, t is the time index, j is the imaginary unit, and u k (t) is the output of the kth eigenmode function at time t, is the twiddle factor, ||·|| 2 is the L2 norm operator; The spectrum segmentation is adaptively selected, specifically, the center frequency set {w k }, perform spectrum segmentation adaptive selection to obtain frequency band division data; The wavelet filter smoothing reconstruction is specifically to construct a high-order window smoothing wavelet filter operator, optimize the structure of the wavelet change filter according to the frequency band division data, and obtain a smoothing wavelet filter, which specifically includes the following steps: The basis function is constructed based on the center frequency set {w k }, for each frequency band [w k-1 ,w k ]Construct wavelet basis function and use square cosine function for smooth transition. The calculation formula is: In the formula, ψ k (f) is the frequency domain representation of the empirical wavelet basis function, f is the frequency variable, w k is the kth center frequency in the set of center frequencies of all modes, k is the eigenmode function index; The filter construction is specifically to construct a smooth transition filter based on the frequency domain representation of the empirical wavelet basis function obtained by the basis function construction to obtain a smooth wavelet filter. The calculation formula is: In the formula, H k (f) is the frequency domain representation of the smoothing wavelet filter, f is the frequency variable, w k is the kth center frequency in the set of center frequencies of all modes, k is the eigenmode function index; The empirical wavelet transform is specifically to use the smoothing wavelet filter to perform wavelet transform to obtain decomposed empirical wavelet signal data, and the calculation formula is: In the formula, x k (t) is the frequency band signal in the k-th decomposed empirical wavelet signal data, F -1 is the inverse Fourier transform function, It is the frequency domain representation of the signal data in the optimized power load data, H k (t) is the frequency domain representation of the smoothing wavelet filter, is the decomposed empirical wavelet signal data, f is the frequency variable; Decomposition feature extraction is specifically to construct a marginal spectrum feature extraction operator, and perform decomposition feature extraction in combination with the decomposition empirical wavelet signal data to obtain decomposition feature data, which specifically includes the following steps: Hilbert transform, specifically, uses Hilbert transform to calculate the instantaneous amplitude and instantaneous frequency of each frequency band signal in the empirical wavelet signal data. The calculation formula is: In the formula, A k (t) is the instantaneous amplitude of each frequency band signal in the empirical wavelet signal data, f k (t) is the instantaneous frequency of each frequency band signal in the empirical wavelet signal data, H(·) is the Hilbert transform operation representation function, x k (t) is the frequency band signal in the k-th decomposed empirical wavelet signal data, arg(·) is the phase angle calculation function; Marginal spectrum analysis, specifically, calculates the frequency characteristics of the band signal through marginal spectrum analysis. The calculation formula is: M k (f)=∫ 0 T A k (t)·(f-f k (t))dt; Where M k (f) is the frequency characteristic of the band signal corresponding to the kth center frequency, T is the total duration of the band signal, A k (t) is the instantaneous amplitude of each frequency band signal in the empirical wavelet signal data, f k (t) is the instantaneous frequency of each frequency band signal in the empirical wavelet signal data; The energy calculation of the frequency band signal is specifically to calculate the energy characteristics of the frequency band signal. The calculation formula is: E k (f)=∫ 0 T |x k (t)| 2 dt; In the formula, E k (f) is the energy characteristic of the frequency band signal corresponding to the kth center frequency, x k (t) is the frequency band signal in the k-th decomposed empirical wavelet signal data, |·| is the modulus operator; The average frequency calculation is specifically to calculate the average frequency characteristics of the frequency band signal. The calculation formula is: In the formula, is the average frequency characteristic of the band signal corresponding to the kth center frequency, f is the frequency variable, is the frequency domain representation of the band signal corresponding to the kth center frequency; Decomposition and extraction, specifically, performing decomposition feature extraction in combination with the decomposition empirical wavelet signal data through the Hilbert transform, the marginal spectrum analysis, the frequency band signal energy calculation and the average frequency calculation, to obtain decomposition feature data; The decomposed feature data specifically refers to high- and low-frequency decomposed feature data, including a high-frequency feature data set and a low-frequency feature data set; The prediction driving module is used to construct a frequency band domain model to perform power load classification prediction. Through load classification prediction, a multi-path cyclic bidirectional long short-term memory network combined with convolution time feature extraction is used to perform load classification prediction, obtain power load prediction data, and send the power load prediction data to the prediction application module and the monitoring feedback module; The prediction application module is used to construct a visual data reference for the load prediction result, obtain the prediction value visualization image analysis data through prediction value visualization, and send the prediction value visualization image analysis data to the monitoring feedback module; The monitoring feedback module is used to implement model prediction and monitor the prediction accuracy of the model, optimize the model prediction performance, and obtain a power load prediction reference through model accuracy monitoring.

2. According to the artificial intelligence-based power load forecasting system of claim 1, Features: The data collection process is used to collect the original data required for power load forecasting, specifically, to obtain the original data for power load forecasting from the power system and environmental sensor records through data collection, and to obtain optimized power load forecasting data by optimizing the original data for power load forecasting; The optimization process specifically includes outlier processing and time series decomposition operations.

3. According to the artificial intelligence-based power load forecasting system of claim 2, Features: The multi-path recurrent bidirectional long short-term memory network combined with convolutional temporal feature extraction specifically includes a convolutional subnet, a recurrent neural subnet and a bidirectional long short-term memory subnet; The convolutional subnet is used to extract time features, weather features and special features from the original data; The recurrent neural subnet is used to process high-frequency feature data in the decomposed features; The bidirectional long short-term memory network is used to process low-frequency feature data and overlapping frequency-divided feature data in the decomposed features; The step of using a multi-path cyclic bidirectional long short-term memory network combined with convolutional time feature extraction to perform load classification prediction and obtain power load prediction data includes: Convolutional feature extraction, specifically, constructing a convolutional feature extraction subnet including only a convolutional layer, a pooling layer, and a fully connected layer, and using one-dimensional convolution for feature extraction, and extracting time series comprehensive feature data from the power load forecast raw data by constructing the convolutional feature extraction subnet; The time series comprehensive characteristic data specifically includes time series time characteristic data, weather time characteristic data and special date time characteristic data; Recurrent network high-frequency feature prediction, specifically, constructing a recurrent neural network subnet including an input layer, a hidden layer, and an output layer, and using the recurrent neural network subnet, combined with the time series comprehensive feature data, to process the feature data in the high-frequency feature data set to obtain high-frequency feature data prediction data; Bidirectional long short-term memory subnetwork prediction, specifically, constructing two layers of bidirectional long short-term memory subnetworks, respectively used to process the feature data in the low-frequency feature data set and the overlapping frequency-divided feature data, and combining the time series comprehensive feature data to perform hierarchical prediction to obtain low-frequency feature data prediction data and overlapping feature data prediction data; The long short-term memory subnet specifically includes an input gate, an output gate, a forget gate and a storage unit; The model output integration decision is specifically to weight the prediction results of the high-frequency feature data prediction data, the low-frequency feature data prediction data and the overlapping feature data prediction data to obtain the model integrated prediction data, and the calculation formula is: In the formula, Y is the model integrated prediction data, α is the high-frequency prediction weight, is the high-frequency feature data prediction data, β is the low-frequency prediction weight, is the low-frequency feature data prediction data, γ is the overlap prediction weight, It is the overlapping feature data prediction data; The power load classification prediction model training is specifically carried out through the convolution feature extraction, the recurrent network high-frequency feature prediction, the bidirectional long short-term memory subnet prediction and the model output integrated decision, to obtain the power load prediction model Model ECP ; Power load forecasting, specifically using the power load forecasting model Model ECP , combining the original power load prediction data, the original overlapping frequency-division feature data and the decomposed feature data, to perform power load prediction and obtain power load prediction data.

4. According to the artificial intelligence-based power load forecasting system of claim 3, Features: The predicted value visualization is used to construct a visual data reference for the load forecasting results, specifically using a data visualization module to draw the power load forecasting data into an interactive visual image to obtain predicted value visualization image analysis data.

5. According to claim 4, an artificial intelligence-based power load forecasting system, Features: The model accuracy monitoring is used to implement model prediction and monitor the prediction accuracy of the model, and optimize the model prediction performance. Specifically, it is to manually analyze the predicted value visualization image analysis data, and optimize the model accuracy based on the comparison between the predicted situation and the actual situation, and obtain the optimal prediction model through accuracy optimization. The optimal prediction model is used to perform the power load prediction task and obtain the final power load prediction reference.