Dynamic sensing and analysis processing method and system for supply-demand double-side high-frequency multi-dimensional load information

Abstract

The invention discloses a dynamic sensing and analysis processing method and system for supply and demand double-side high-frequency multi-dimensional load information. The method comprises the steps that supply and demand double-side power load data are collected, preprocessing operation is conducted on the data, abnormal data are recognized and corrected, and missing data are filled up; the acquired data information is integrated, expanded and decomposed by adopting an ensemble empirical mode decomposition method and a variational mode decomposition method, and then the data is subjected to stabilization processing. Wherein the EEMD method is suitable for processing signals with non-linear and non-stationary characteristics, and the VMD method is more suitable for processing signals requiring accurate frequency control and modal aliasing reduction; and an improved Transform model is adopted to carry out modeling on the short-term power load information so as to predict the future development trend of the load information. The method can effectively reduce the calculation amount of the prediction model, improves the precision of power load prediction, and has high robustness and adaptivity.

Description

Technical Field

[0001] The present invention relates to the technical field of electric load forecasting, and in particular, to a method and system for dynamically perceiving and analyzing and processing high-frequency multi-dimensional load information on both the supply and demand sides. Background Art

[0002] The electric power industry shoulders the important mission of delivering stable, high-quality, safe and reliable electric power resources to the country, and its booming development is related to the long-term strategy of national security and sustainable development. With the rapid development of the economy and the continuous improvement of people's living standards, the electricity demand is also increasing day by day. However, the main challenge in realizing the reasonable planning and efficient utilization of electric power resources lies in how to accurately extract and analyze the electricity consumption characteristics of the load information on both the supply and demand sides, and predict the future development trend of the load information, so as to provide a premise for the operation state monitoring of the power system, the formulation of real-time electricity price strategies, and the dispatching of electric power resources. However, electric load information usually has the characteristics of high frequency, many dimensions, strong non-stationarity, and numerous influencing factors. Traditional methods based on deterministic analysis are difficult to accurately perceive and predict electric load information.

[0003] In the practical application of the power system situation awareness technology, the "state" in the situation awareness can be regarded as the real-time operation state of the power system, while the "trend" is regarded as the fundamental trend of the system operation and development. Through situation awareness, the high-frequency multi-dimensional load information on both the supply and demand sides in a dynamic environment can be perceived, analyzed, and the future situation can be predicted, thus providing a crucial basis for the planning, construction, and transformation of the power system.

[0004] At present, the existing situation awareness technology based on the load side focuses on the quantification of the adjustable potential of flexible loads and is a control strategy serving the power grid dispatching. However, the situation awareness of the load side for electric load forecasting has not been deeply studied. At the same time, in view of the randomness and volatility risks of the power system operation, it is necessary to enhance the observability of the situation awareness on the time scale. Along with the increase in the depth and breadth of the load data that can be collected in the power system, the traditional single load forecasting model has been difficult to meet the high-precision requirements of electric load forecasting. How to use a high-precision and low-complexity load forecasting model in combination with historical load data to obtain accurate electric load forecasting results is also an urgent problem to be solved. Summary of the Invention

[0005] Object of the Invention: The present invention aims to provide a method for dynamically perceiving and analyzing and processing high-frequency multi-dimensional load information on both the supply and demand sides, which can reduce the prediction calculation amount, improve the prediction accuracy, and has strong robustness and self-adaptability; another object of the present invention is to provide a system for dynamically perceiving and analyzing and processing high-frequency multi-dimensional load information on both the supply and demand sides.

[0006] Technical solution: The dynamic perception and analysis processing method for high-frequency multi-dimensional load information on both the supply and demand sides of the present invention includes the following steps:

[0007] (1) According to the demand for power load information situation awareness, collect power load data, perform preprocessing, identify and correct abnormal data, fill in missing data, and achieve situation perception;

[0008] (2) Use the Ensemble Empirical Mode Decomposition (EEMD) method and the Variational Mode Decomposition (VMD) method to integrate and decompose the preprocessed data, and perform stationary processing to complete situation understanding;

[0009] (3) Based on short-term power load information, construct an improved Transformer model for situation prediction of load information; among them, in the improved Transformer model, the time series mask is stripped from the decoder and moved to the encoder, and a cyclic multi-point prediction mechanism is introduced into the decoder.

[0010] Further, step (1) is specifically as follows:

[0011] (14) Collect power load data, use the subtractive clustering algorithm to process abnormal data in the experimental data, screen the main factors affecting the change of power load, remove the redundant amount in the data, and determine the key attributes between information;

[0012] (15) Process missing data, use the average value of adjacent sampling point data to fill in the data before and after the missing data; when the data is continuously missing, according to expert experience, refer to the load day data similar to the missing data and the external environmental influencing factors at that time, and manually correct and fill in the data;

[0013] (16) Perform data normalization processing.

[0014] Further, step (11) is specifically as follows:

[0015] (111) Initialize k = 1, and let the sample set Y = {y 1 ,y 2 ,…,y n} be a load data set on R n , and calculate the density index D i of the i-th point y i in the sample set:

[0016]

[0017] where γ a is the domain radius of y i , indicating that with y i as the center and γ aA circular region with a radius, and the data points outside the region have little influence on the density index of y i The definition of γ a is as follows:

[0018]

[0019] where y k represents the mean of the k-th class of samples; then, select the data point with the highest density index as the first clustering center, denoted as whose density index is

[0020] (112) For the k-th selected clustering and its density index it is necessary to correct the density index of each data point y i The corrected density index of the i-th point y i is

[0021]

[0022] where γ b is a positive number, take γ b =(1.2 - 1.5)γ a The purpose is to reduce the density index of the data points near the clustering center to avoid the occurrence of clustering centers that are very close to each other.

[0023] (113) Select the point with the highest density index among the corrected data points as the new clustering center, whose density index is

[0024] (114) If the condition is satisfied, then stop the iteration and output the number of clusters k max ; otherwise, return to step (112) to continue the iteration;

[0025] (115) Calculate the mean square error corresponding to the clustering center for each data point y i (116) Determine whether there are abnormal data in the load data, specifically as follows:

[0026]

[0027] If the above formula is satisfied, then this data is abnormal data;

[0028]

[0029] If the above formula is satisfied, then this data is abnormal data;

[0030] (117) Suppose a total of k maxA characteristic curve, and abnormal data exists in the curve X to be inspected d from point p to point q, and its characteristic

[0031] curve is Y t , and the correction curve is Y r . The abnormal data is corrected by the following formula:

[0032]

[0033] where Y d is historical load data, and Y t is current load data.

[0034] Furthermore, step (13) is specifically as follows:

[0035] (131) The extreme value method is used to normalize the historical load data, and the normalized historical load data Y' d is

[0036]

[0037] where Y' d takes values in [0, 1], Y max and Y min are the maximum and minimum values in the sample set Y;

[0038] (132) The meteorological data is normalized, and the normalized meteorological data T' m is

[0039]

[0040] where T' m takes values in [0, 1], T m is the original temperature and humidity data, T min is the minimum temperature and humidity, and T max is the maximum temperature and humidity;

[0041] (133) The time data is normalized, taking weekdays and non - weekdays as the objects of investigation, and quantifying the week as a type.

[0042] Furthermore, step (2) is specifically as follows:

[0043] (23) For signals with non - linear and non - stationary characteristics, the ensemble empirical mode decomposition method EEMD is used to complete the situation understanding;

[0044] (24) For signals with high - precision control frequency and low modal aliasing, the variational mode decomposition method VMD is used to complete the situation understanding.

[0045] Furthermore, step (21) is specifically as follows:

[0046] (211) Set the decomposition times \(m\), add Gaussian white noise with a standard deviation \(z\) times that of the original sequence to the original sequence \(y(t)\) to obtain the sequence formula: \(y'(t)=y(t)+\varepsilon\) z (t);

[0047] (212) Obtain all the intrinsic mode function values (t, where \(i\) is the order of the IMF and \(n = 1, 2, \ldots, m\));

[0048] (213) Repeat steps (211) and (212) \(m\) times to obtain all the IMFs;

[0049] (214) Calculate the average value of the IMFs \(m\) times to eliminate noise interference.

[0050] Further, step (22) is specifically as follows:

[0051] (221) Perform Hilbert transform on the signal to obtain the analytical signal of the mode function \(u\) k (t), that is, its single-sided spectrum where \(\delta(t)\) is the unit impulse signal;

[0052] (222) Add an exponential term to correct the central estimated frequency of each mode function and modulate its spectrum to a fundamental frequency bandwidth of

[0053]

[0054] where \(j\) is a complex number and \(\omega\) k is the central frequency;

[0055] (223) Calculate the squared norm \(L\) of the gradient of the demodulated signal, and obtain the estimated fundamental frequency bandwidth of each mode function, satisfying the following constraint conditions: 2 , and \(\{u\)

[0056]

[0057] where \(\{u\) k}\) is the set of mode function components, \(\{\omega\) k}\) is the set of central frequencies, is the derivative with respect to \(t\);

[0058] (224) Introduce the Lagrange operator \(\lambda\) and the quadratic penalty factor \(\alpha\) to convert the constrained variational problem into an unconstrained problem, and the set of function components is

[0059]

[0060] (225) Obtain the Lagrange operator \(\lambda\) and the mode component \(u\) through the alternating direction of the penalty algorithmk and modal frequency ω k :

[0061]

[0062] where f(ω) is the function of the original signal sequence with respect to the modal frequency, u i (ω) is the function of the modal component with respect to frequency, λ(ω) is the function of the operator with respect to frequency, and γ is the noise tolerance.

[0063] Furthermore, in step (3), the encoder of the improved Transformer model includes a CNN feature extractor, a position information generator, a mask matrix, and a multi-layer multi-head attention unit;

[0064] Among them, the CNN feature extractor performs sliding convolution on the m-dimensional convolution kernel, and the pixel value of the feature map at the position (u, v) is

[0065]

[0066] In the formula, σ is a hyperparameter, is the input, is the weight, b (l) is the bias term;

[0067] The position encoding generated by the position information generator is:

[0068]

[0069] where PE represents the position encoding, pos represents the offset position, and d represents the word embedding dimension;

[0070] The mask matrix is used to automatically mask the information after the current processing point during encoding, and is stripped from the decoder by the time series mask and moved to the encoder;

[0071] The multi-head multi-layer self-attention unit is used to perform residual correction and layer normalization on the input result; input the processed data into the multi-layer perceptron; perform residual correction and layer normalization on the output result of the multi-layer perceptron and then output the processing result of one layer of the multi-head self-attention unit.

[0072] Furthermore, in step (3), an LSTM network is provided between the decoder and the encoder of the improved Transformer model.

[0073] The dynamic perception and parsing processing system for the supply and demand bilateral high-frequency multi-dimensional load information of the present invention includes:

[0074] The situation awareness module is used to collect power load data according to the situation awareness requirements of power load information, perform preprocessing, identify and correct abnormal data, fill in missing data, and achieve situation awareness;

[0075] The situation understanding module is used to integrate and decompose the preprocessed data by using the ensemble empirical mode decomposition method (EEMD) and the variational mode decomposition method (VMD), and perform stationary processing to complete situation understanding;

[0076] The situation prediction module is used to construct an improved Transformer model based on short-term power load information for situation prediction of load information; among them, the improved Transformer model strips the time series mask from the decoder and moves it to the encoder, and introduces a cyclic multi-point prediction mechanism in the decoder.

[0077] Beneficial effects: Compared with the prior art, the significant advantages of the present invention are as follows: 1. For load signals with different characteristic requirements, the present invention respectively uses the ensemble empirical mode decomposition method and the variational mode decomposition method to achieve signal decomposition, and uses an improved Transformer load prediction model to reconstruct the power load situation awareness combined prediction model, model the short-term power load information to predict the future development trend of load information, reduce the computational amount of the prediction model, improve the prediction accuracy, and has strong robustness and self-adaptability; 2. According to the situation awareness requirements of power load information on both the supply and demand sides, the present invention collects power load data, uses the subtractive clustering algorithm to process the experimental data for anomalies to screen out the main factors affecting the change of power load, remove the redundant amount in the data, and at the same time mine the key attributes between information, so as to reduce the prediction calculation amount without affecting the prediction accuracy; 3. The present invention strips the time series mask from the decoding layer and moves it to the encoding layer, which can automatically mask the information after the current processing time point during encoding, making the input space closer to the real application scenario. Description of the Drawings

[0078] Figure 1 It is a flowchart of the present invention. Detailed Embodiment

[0079] The following further describes the present invention with reference to the drawings.

[0080] The dynamic perception and analysis processing method for high-frequency multi-dimensional load information on both the supply and demand sides of the present invention includes the following steps:

[0081] Step 1: According to the situation awareness requirements of power load information on both the supply and demand sides, first collect power load data, perform preprocessing operations on the data, identify and correct abnormal data, fill in missing data, and achieve situation awareness:

[0082] (11) Collect power load data, and use the subtraction clustering algorithm to process the experimental data for anomalies, so as to screen out the main factors affecting the change of power load, remove the redundant amount in the data, and at the same time mine the key attributes between information, so as to reduce the prediction calculation amount without affecting the prediction accuracy.

[0083] (111) Initialize k = 1, and let the sample set Y = {y 1 , y 2 , …, y n} be a load data set on R n . Calculate the density index D i of the i-th point y i in the sample set:

[0084]

[0085] where γ a is the neighborhood radius of y i , representing a circular area centered on y i with a radius of γ a . The data points outside the area have little influence on the density index of y i . The definition of γ a is as follows

[0086]

[0087] where y k represents the mean value of the k-th class of samples; then, select the data point with the highest density index as the first clustering center, denoted as and its density index is

[0088] (112) For the k-th selected clustering and its density index , it is necessary to correct the density index of each data point y i . The corrected density index of the i-th point y i is

[0089]

[0090] where γ b is a positive number, and take γ b = (1.2~1.5)γ a . The purpose is to reduce the density index of the data points near the clustering center and avoid the occurrence of clustering centers that are very close to each other.

[0091] (113) Select the point with the highest density index among the corrected data points as the new clustering center, and its density index is

[0092] (114) If the condition is satisfied then stop the iteration and output the number of clusters k at this time max ; otherwise, return to step (112) to continue the iteration;

[0093] (115) Calculate the mean square error of each data point y i corresponding to the cluster center

[0094]

[0095] (116) Determine whether there is abnormal data in the load data, specifically as follows:

[0096]

[0097] If the above formula is satisfied, this data is abnormal data;

[0098] (117) Suppose a total of k max characteristic curves are generated, and the abnormal data exists from point p to point q of the curve X to be inspected d whose characteristic curve is Y t , and the correction curve is Y r , and correct the abnormal data through the following formula:

[0099]

[0100] where Y d is the historical load data, and Y t is the current load data.

[0101] (12) Missing data processing: Due to factors such as network delay and system lag, the probability of single data missing increases. Therefore, in this case, the average value of adjacent sampling point data is usually used to fill the data before and after the missing data; in the case of power outage maintenance and sudden accident tripping, it will cause the missing of continuous data. At this time, it is necessary to draw on expert experience, refer to the load day data similar to the missing data, consider the influencing factors of the external environment at that time, and make artificial experience correction to fill the data;

[0102] (13) Data normalization processing: Power load is affected by many factors such as historical data, meteorological data (such as temperature, humidity, wind speed), time period, etc. Therefore, it is necessary to normalize these data:

[0103] (131) Adopt the extreme value method to normalize the historical load data and substitute it into the formula:

[0104]

[0105] Y' d is the normalized historical load data, with values in the range of [0, 1]. Y max and Y min are the maximum and minimum values in the sequence Y;

[0106] (132) Normalize the meteorological data: Meteorological data mainly includes the maximum temperature, minimum temperature, humidity, etc. To ensure the accuracy of model training and prediction, the meteorological data also needs to be normalized. By analogy with the method of normalizing historical load data, the temperature and humidity normalization formula is obtained:

[0107]

[0108] T' m is the normalized meteorological data, with values in the range of [0, 1];

[0109] (133) Normalize the time data: For a comprehensive urban power consumption system, on weekdays, measures with larger assigned values are usually adopted to adjust the pressure of the power system, and on rest days, measures with smaller assigned values are adopted to relieve the pressure of the power system. Therefore, taking weekdays and non-weekdays as the objects of investigation, the week is quantified as a type to reduce the expected deviation and increase the prediction accuracy. The quantification scale is as follows:

[0110]

[0111] Step 2: For signals with different characteristic requirements, the ensemble empirical mode decomposition method and the variational mode decomposition method are respectively used to integrate and decompose the acquired data information, and then the data is smoothed, that is, on the basis of the existing power load information, the operation state information of the power system is fully explored to complete the situation understanding:

[0112] (21) For signals with non-linear and non-stationary characteristics, the ensemble empirical mode decomposition method is used to complete the situation understanding. The steps are as follows:

[0113] (211) Set the decomposition times m. Add Gaussian white noise with a standard deviation of 0.2 times the standard deviation of the original sequence to the original sequence y(t) to obtain the sequence formula: y’(t) = y(t) + ε n (t);

[0114] (212) Obtain all the intrinsic mode function values i is the order of the IMF;

[0115] (213) Repeat the above two steps m times to obtain all the IMFs;

[0116] (214) Calculate the average value of the m IMFs to eliminate noise interference.

[0117] (22) For signals with high-precision control frequency and low modal aliasing, the variational mode decomposition method is adopted, and the steps are as follows:

[0118] (221) Perform Hilbert transform on the signal to obtain the analytic signal of the modal function u k (t), that is, its single-sided spectrum

[0119] (222) Add an exponential term to correct the central estimated frequency of each modal function and modulate its spectrum to a base frequency bandwidth of

[0120] (223) Calculate the square L 2 norm of the demodulated signal gradient, obtain the estimated base frequency bandwidth of each modal function, and construct a constrained variational problem, the expression of which is as follows:

[0121]

[0122] In the formula, {u k} is the set of modal function components, and {ω k} is the set of central frequencies;

[0123] (224) To solve the above variational constraint problem, introduce the Lagrange operator λ and the quadratic penalty factor α for solution. The formula is rewritten as the set of function components:

[0124]

[0125] (225) Obtain the operator λ, the modal component u k and the modal frequency ω k through the alternating direction of the penalty algorithm:

[0126]

[0127] λ n+1 (ω) = λ n (ω) + γ

[0128] where f(ω) is the function of the original signal sequence with respect to the modal frequency, u i (ω) is the function of the modal component with respect to the frequency, λ(ω) is the function of the operator with respect to the frequency, and γ is the noise tolerance.

[0129] Step 3: Evaluate and predict the future situation information of the system: The power load is a non-white noise time series with characteristics of monotonicity, periodicity, step change, normality, non-stationarity, long-distance dependence, and "large nearby and small far away". A single statistical learning method can only focus on a certain characteristic in the power load change trend, while the Transformer model with extremely strong long-distance non-linear fitting ability is not suitable for the load prediction task. To solve this problem, an improved Transformer model is adopted to adapt to the long-distance non-linear dependence characteristics of the power load to predict the future development trend of the load information:

[0130] (31) Construct the encoder of the improved Transformer model. The encoding layer consists of a CNN feature extractor, a position information generator, a mask matrix, and a multi-layer multi-head attention unit. The steps are as follows:

[0131] (311) Construct a CNN feature extractor. Use a convolutional kernel of 3 rows and 1 column, with 1 unit of edge padding; the number of convolutional kernels is an adjustable hyperparameter. Perform sliding convolution on the m-dimensional convolutional kernel, and calculate the pixel value of the feature map at position (u, v) through formula (13);

[0132]

[0133] In the formula, σ is a hyperparameter, is the input, is the weight, and b (l) is the bias term;

[0134] (312) Improve the time series mask. The improved algorithm strips the time series mask from the decoding layer and moves it to the encoding layer. Based on this design, the model can automatically mask the information after the current processing time point during encoding, making the input space closer to the real application scenario;

[0135] (313) Position encoding. The encoding position enables the input sequence to carry position information, enabling the model to automatically capture local dependencies related to the position. The calculation formula for position encoding is:

[0136]

[0137]

[0138] where PE represents the position encoding, pos represents the offset position, and d represents the word embedding dimension.

[0139] (314) Multi-Head Multi-Layer Self-Attention Unit. First, input the data into the multi-head self-attention mechanism unit; then perform residual correction and layer normalization on the output result; input the processed data into the multi-layer perceptron; perform another residual correction and layer normalization on the output result of the multi-layer perceptron and then output the processing result of one layer of the multi-head self-attention unit; after the input data is operated by the multi-layer processing structure, output the final result.

[0140] (32) Construct the decoding layer. To solve the problem of too high computational complexity when the decoder of the native Transformer model processes power data, simplify the decoder structure and introduce a cyclic multi-point prediction mechanism. The steps are as follows:

[0141] (321) After the encoder output, input it into the LSTM network. The LSTM network realizes the long-term and short-term memory mechanism through the gate structure. The gate structure includes an input gate, a forget gate, and an output gate. The LSTM network allows information to selectively affect the data state at each time point through these "gates";

[0142] (322) Introduce a cyclic multi-point prediction mechanism. Assume that the model predicts the load values of the next s time points based on the observed values of the previous r time points at the current time t (r >= s). During training, concatenate the load values of the previous r time points at time t with the relevant factors at time t + 1 as the input at time t for predicting the load at time t + 1: X (t) = (P t-r+1 , …, P t-1 , P t , θ t+1 ). During prediction, concatenate the output of the previous time point to the end of the previous r load values at the current time point, intercept the data of the last r time points, and concatenate it with the relevant features of the next time point to form the input at the current time point. The above process is executed cyclically s times to output the results of continuous multi-point prediction;

[0143] (33) Output the power load prediction value. After training is completed, output the prediction value of the power load after being processed by the encoding layer and the decoding layer.

Claims

1. A dynamic perception and analytical processing method for high-frequency multi-dimensional load information on both the supply and demand sides, characterized in that: The following steps are involved: (1) According to the needs of power load information situation awareness, collect power load data, perform preprocessing, identify and correct abnormal data, fill in missing data, and achieve situation awareness; (2) The EEMD and VMD methods are used to integrate and decompose the preprocessed data, and then stabilize the data to complete the situation understanding. (3) An improved Transformer model is constructed based on short-term power load information to perform load information situation prediction; wherein, the improved Transformer model removes the timing mask from the decoder and moves it to the encoder, and introduces a cyclic multi-point prediction mechanism in the decoder.

2. According to claim 1, the dynamic perception and analytical processing method of high-frequency multi-dimensional load information on both the supply and demand sides is characterized by: Step (1) is as follows: (11) Collect power load data, use subtractive clustering algorithm to handle anomalies in experimental data, screen the main factors affecting power load changes, remove redundancy in data, and determine the key attributes between information; (12) Missing data processing: use the average value of the data from adjacent sampling points to fill in the data before and after the missing data; when the data is missing continuously, refer to the load day data similar to the missing data and the external environmental influencing factors at that time based on expert experience, and manually correct and fill in the data; (13) Data normalization processing.

3. The method for dynamic perception and analytical processing of high-frequency multi-dimensional load information on both the supply and demand sides according to claim 2 is characterized in that: Step (11) is as follows: (111) Initialize k = 1, let the sample set Y = {y1, y2, …, y n } is R n A load data set on the sample set, calculate the i-th point y i Density index D i : Among them, γ a for y i The radius of the field, expressed in y i Centered on γ a is the circular area with radius, γ a is defined as follows: Among them, y k represents the mean of the k-th class samples; (112) For the kth selected cluster Its density index For each data point y i The density index is corrected, and the i-th point y i The modified density index is Among them, γ b is a positive number, take γ b =(1.2~1.5)γ a ; (113) Select the point with the highest density index among the corrected data points As the new cluster center, its density index is (114) If satisfied Then stop the iteration and output the number of clusters k at this time max ; Otherwise, return to step (112) to continue iteration; (115) Calculate each data point y i The mean square error corresponding to the cluster center (116) Determine whether the load data contains abnormal data, as follows: If the above formula is satisfied, the data is abnormal data; (117) Let k be the total number of max characteristic curves, abnormal data exists in the detected curve X d From point p to point q, its characteristic curve is Y t , the correction curve is Y r , and correct the abnormal data by the following formula: Among them, Y d is the historical load data, Y t is the current load data.

4. The method for dynamic perception and analytical processing of high-frequency multi-dimensional load information on both the supply and demand sides according to claim 3 is characterized in that: Step (13) is as follows: (131) The extreme value method is used to normalize the historical load data. The normalized historical load data Y' d for Among them, Y' d The value is [0,1], Y max and Y min is the maximum and minimum value in the sample set Y; (132) The meteorological data is normalized, and the normalized meteorological data T' m for Among them, T' m The value is [0,1], T m is the original temperature and humidity data, T min is the minimum temperature and humidity, T max is the maximum value of temperature and humidity; (133) The time data is normalized, and the working days and non-working days are taken as the objects of investigation, and the week is quantified as the type.

5. The method for dynamic perception and analytical processing of high-frequency multi-dimensional load information on both the supply and demand sides according to claim 4 is characterized in that: Step (2) is as follows: (21) For signals with nonlinear and non-stationary characteristics, the ensemble empirical mode decomposition method (EEMD) is used to achieve situation understanding; (22) For signals with high-precision control frequency and low modal aliasing, the variational mode decomposition method (VMD) is used to achieve situation understanding.

6. The method for dynamic perception and analytical processing of high-frequency multi-dimensional load information on both the supply and demand sides according to claim 5 is characterized in that: Step (21) is as follows: (211) Set the decomposition times m, add Gaussian white noise with a standard deviation z times the standard deviation of the original sequence to the original sequence y(t), and obtain the sequence formula: y'(t) = y(t) + ε z (t); (212) Get all intrinsic mode function values i is the order of IMF, n = 1, 2, …, m; (213) Repeat steps (211) and (212) m times to obtain all IMFs; (214) Calculate the average value of the IMF for m times to eliminate noise interference.

7. The method for dynamic perception and analytical processing of high-frequency multi-dimensional load information on both the supply and demand sides according to claim 6 is characterized in that: Step (22) is as follows: (221) Perform Hilbert transform on the signal and obtain the modal function u k (t) is the analytical signal, i.e. its one-sided spectrum Among them, δ(t) is the unit impulse signal; (222) Add exponential terms to correct the central estimated frequency of each mode function and modulate its spectrum to a baseband bandwidth of Where j is a complex number, ω k is the center frequency; (223) Calculate the square norm L of the demodulated signal gradient 2 , obtain the estimated fundamental bandwidth of each mode function, satisfying the following constraints: Among them, {u k } is the set of modal function components, {ω k } is the set of center frequencies, is the derivative with respect to t; (224) The Lagrange operator λ and the quadratic penalty factor α are introduced to transform the constrained variational problem into an unconstrained problem. The set of function components is (225) The Lagrange operator λ and the modal component u are obtained by alternating the direction of the penalty algorithm. k and modal frequency ω k : l n+1 (ω)=λ n (ω)+γ Among them, f(ω) is the function of the original signal sequence with respect to the modal frequency, u i (ω) is the function of the modal component with respect to frequency, λ(ω) is the function of the operator with respect to frequency, and γ is the noise tolerance.

8. The method for dynamic perception and analytical processing of high-frequency multi-dimensional load information on both the supply and demand sides according to claim 7 is characterized in that: In step (3), the encoder of the improved Transformer model includes a CNN feature extractor, a position information generator, a mask matrix, and a multi-layer multi-head attention unit; Among them, the CNN feature extractor performs sliding convolution on the m-dimensional convolution kernel, and the pixel value of the feature map at position (u, v) for In the formula, σ is a hyperparameter, For input, is the weight, b (l) is the bias term; The position code generated by the position information generator is: Among them, PE represents position encoding, pos represents offset position, and d represents word embedding dimension; The mask matrix is ​​used to automatically mask the information after the current processing time point during encoding, and is stripped from the decoder and moved to the encoder by the timing mask; The multi-head and multi-layer self-attention units are used to perform residual correction and layer standardization on the input results; the processed data is input into the multi-layer perceptron; the output results of the multi-layer perceptron are further subjected to residual correction and layer standardization, and then the processing results of a layer of multi-head self-attention units are output.

9. The method for dynamic perception and analytical processing of high-frequency multi-dimensional load information on both the supply and demand sides according to claim 1 is characterized in that: In step (3), an LSTM network is provided between the decoder and the encoder of the improved Transformer model.

10. A dynamic perception and analytical processing system for high-frequency multi-dimensional load information on both the supply and demand sides, characterized in that: include: The situation awareness module is used to collect power load data, perform preprocessing, identify and correct abnormal data, fill in missing data, and realize situation awareness according to the needs of power load information situation awareness; The situation understanding module is used to integrate and decompose the pre-processed data using the ensemble empirical mode decomposition method EEMD and the variational mode decomposition method VMD, and perform stabilization processing to complete the situation understanding; The situation prediction module is used to construct an improved Transformer model based on short-term power load information to perform situation prediction of load information; wherein, the improved Transformer model removes the timing mask from the decoder by the Transformer model, moves it to the encoder, and introduces a cyclic multi-point prediction mechanism in the decoder.

Images