A short-term net load probability forecasting method based on high-frequency enhanced denoising diffusion model

Abstract

The present invention discloses a short-term net load probability prediction method based on a high-frequency enhanced denoising diffusion model, and relates to the technical field of power load prediction. The present invention includes: receiving historical net load data of the distribution network and multi-dimensional meteorological collection data, screening conditional features through the XGBoost algorithm and the MIC maximum information coefficient, and constructing a conditional feature set based on the screened conditional features; forming noisy net load data based on the historical load data and the forward diffusion process of the diffusion model; constructing a high-frequency enhanced denoising diffusion model, and training the high-frequency enhanced denoising diffusion model using the conditional feature set and the noisy net load data. The present invention can improve the ability to accurately characterize complex load fluctuation patterns and the stability of prediction results through adaptive enhancement of high-frequency signals and effective extraction of multi-scale features, and provide a more reliable decision-making basis for distribution network planning, production, and operation scheduling.

Description

Technical Field

[0001] The present invention relates to the technical field of electric load forecasting, and in particular to a short-term net load probability forecasting method based on a high-frequency enhanced denoising diffusion model. Background Art

[0002] As the penetration rate of distributed resources continues to increase, the uncertainty and volatility of the net load of the distribution network are becoming more and more prominent due to the influence of intermittent power generation and unstable user load demand, and the difficulty of power load forecasting is gradually increasing. Therefore, in order to fully capture the uncertainty of future loads and improve the safety and stability of power system operation, accurate and reliable short-term net load probability forecasting is particularly important.

[0003] Existing load probability forecasting methods can be divided into parametric and non-parametric methods. Among them, the parametric analysis method that relies on some a priori assumed probability distribution cannot effectively characterize the time series evolution characteristics of the net load, and has poor practical adaptability in the scenario where the data distribution drifts. Although some sub-parametric methods represented by quantile regression and kernel density have higher flexibility and reliability, they have problems such as high computational cost and insufficient ability to analyze complex load fluctuation patterns. For this reason, we propose a short-term net load probability forecasting method based on a high-frequency enhanced denoising diffusion model. Summary of the invention

[0004] The purpose of the present invention is to provide a short-term net load probability prediction method based on a high-frequency enhanced denoising diffusion model to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a short-term net load probability prediction method based on a high-frequency enhanced denoising diffusion model, comprising the following steps:

[0006] Receive the historical net load data of the distribution network and the multi-dimensional meteorological data, screen the conditional features through the XGBoost algorithm and the MIC maximum information coefficient, and construct a conditional feature set based on the screened conditional features;

[0007] Construct a high-frequency enhanced denoising diffusion model to form noisy net load data based on historical load data and the forward diffusion process of the diffusion model;

[0008] The high-frequency enhanced denoising diffusion model is trained using the conditional feature set and the noisy net load data. After reaching the usage accuracy, the noise data and conditional features are input to achieve accurate noise prediction, and the initial load probability prediction result is generated through the reverse denoising process.

[0009] The initial load prediction value at the confidence interval quantile is obtained based on kernel density estimation, and then the parameters of the high-frequency enhancement denoising diffusion model are fine-tuned and updated through the improved loss function;

[0010] The updated high-frequency enhanced denoising diffusion model is used to obtain the final net load probability prediction result through back diffusion.

[0011] Furthermore, a conditional feature set is constructed, specifically including the following:

[0012] (1) Taking 6 a.m. and 6 p.m. as the boundary, the input distribution network historical net load data and multi-dimensional meteorological data are divided into two scenes: daytime and nighttime;

[0013] (2) For different scenarios, the XGBoost algorithm is used to quantify the importance of each meteorological feature on load changes, and the MIC maximum information coefficient is used to calculate the linear / nonlinear correlation between features. The importance and correlation of features are comprehensively considered to form a meteorological feature set;

[0014] (3) The obtained meteorological feature set is combined with the historical load data and the one-hot encoded date variable to form a complete conditional input feature set for the diffusion model.

[0015] Furthermore, the historical meteorological data characteristics and historical load data characteristics are normalized as subsequent inputs. The specific processing formula is as follows:

[0016]

[0017] In the formula, y * is the normalized feature data, y * ∈[0,1], y is the original feature data; y max and y min are the maximum and minimum values ​​of the corresponding data in the total sample data.

[0018] Furthermore, the high-frequency enhanced denoising diffusion model includes an encoder and a decoder. The encoder is a CNN-LSTM network and a Linear layer based on fused wavelet attention, which is used to encode input noise data and conditional features; the decoder is a multi-channel convolutional neural network, which is used to accurately predict noise.

[0019] Furthermore, the cross-attention mechanism is selected as the conditional guidance architecture, and the correlation between the input noise value and the conditional feature is mined based on the conditional guidance architecture, and the information is fused and used as the decoder input.

[0020] Furthermore, the construction is based on a CNN-LSTM network fused with wavelet attention, as follows:

[0021] (1) Combining discrete wavelet transform with self-attention mechanism, performing multi-level wavelet decomposition on the features of linear layer mapping;

[0022] (2) In the frequency domain, a self-attention mechanism is used to adaptively enhance the high-frequency detail components. After high-frequency enhancement, CNN-LSTM is used to extract local key and long-term dependency information.

[0023] Furthermore, the discrete wavelet transform is calculated as follows:

[0024]

[0025] Where: α 0 is the initial scale parameter; τ is the translation parameter; Ψ * (•) is the wavelet basis function; m is the scaling constant (indicating the order of decomposition); n is the translation constant; m , n All are integers; t is the time variable of the original signal.

[0026] Furthermore, the high-frequency component enhancement formula of the self-attention mechanism is as follows:

[0027]

[0028] Where: softmax represents normalization processing; Q , K , V is the matrix after the coefficients are linearly transformed, and the calculation formula is as follows:

[0029]

[0030] Where: , and is the parameter matrix to be learned; For the m The detail coefficient matrix of the layer.

[0031] Furthermore, the improved loss function is as follows:

[0032]

[0033]

[0034]

[0035] In the formula, is the improved loss function; is the original loss function calculated by the MSE indicator; is the designed auxiliary loss function; λ is the weighted weight; For the t The actual noise value added in the first step; The noise neural network predicts t step noise value, c is the conditional input, x t is the historical load value Forward Diffusion t The noise data obtained in the first step; and For m ub and m lb diffusion t The noise data obtained by step, where m ub and m lb is the load value of the initial prediction result at the upper and lower limits of the confidence interval, which can be obtained by kernel density estimation.

[0036] Furthermore, the calculation formulas for the kernel density estimation and prediction interval are as follows:

[0037]

[0038]

[0039]

[0040] Where: x is the probability density function f ( x )’s independent variable, x ∈R; F (•) is the cumulative distribution function obtained by kernel density estimation; n 1 and h are the estimated load quantity and bandwidth respectively; x η For the estimation of η elements; K (•) is the kernel function; q α is the quantile value at the nominal quantile level α; γ is the nominal coverage probability of the probability prediction interval; and is the load probability prediction interval P 1 , γ Nominal quantile levels for the upper and lower boundaries.

[0041] The present invention has at least the following beneficial effects:

[0042] 1. The short-term net load probability prediction method based on the high-frequency enhanced denoising diffusion model provided by the present invention aims at the situation that the net load forms during the day and at night are quite different in the scenario of high proportion of clean energy access. In order to avoid the important features of individual time periods being masked, the XGBoost algorithm and the MIC maximum information coefficient are used to extract key meteorological features, and the conditional input feature set is formed by combining historical load data with date variables, which more concisely and comprehensively refines the conditional features that have a greater impact on the net load in different scenarios.

[0043] 2. The short-term net load probability prediction method based on the high-frequency enhanced denoising diffusion model provided by the present invention adopts the high-frequency enhanced denoising diffusion model to generate prediction results in view of the non-stationarity and strong volatility of the net load data. Mainly for the encoding and decoding links, a CNN-LSTM network and a multi-channel convolutional neural network integrating wavelet attention are designed respectively. Through the adaptive enhancement of high-frequency signals and the effective extraction of multi-scale features, the model's ability to accurately characterize complex load fluctuation patterns is improved.

[0044] 3. The short-term net load probability prediction method based on the high-frequency enhanced denoising diffusion model provided by the present invention takes into account the possible defects in the quality of the single model prediction results. Based on the improved loss function, the pre-trained model parameters are updated within the limited computational complexity and time consumption range. It can appropriately shrink the interval width and improve the stability of the prediction results, providing a more reliable decision-making basis for distribution network planning, production, and operation scheduling.

[0045] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the overall process of the method of the present invention;

[0047] Figure 2 This is a schematic diagram of the probability prediction results of a typical day 1 interval in the present invention;

[0048] Figure 3 This is a schematic diagram of the probability prediction results of the typical day 2 interval in the present invention;

[0049] Figure 4 It is a schematic diagram of the probability prediction results of the three intervals on a typical day in the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0051] See also Figure 1-Figure 4 The present invention provides a technical solution: a short-term net load probability prediction method based on a high-frequency enhanced denoising diffusion model, comprising the following steps:

[0052] S1. Receive the historical net load data of the distribution network and the multi-dimensional meteorological data, screen the conditional features through the XGBoost algorithm and the MIC maximum information coefficient, and construct the conditional feature set based on the screened conditional features, as follows:

[0053] S11. Taking 6:00 am and 6:00 pm as the boundary (it can be set according to the actual situation, and can separate the time periods with large differences in net load forms), the input distribution network historical net load data and multi-dimensional meteorological collection data are divided into two scenes: daytime and nighttime;

[0054] S12. For different scenarios, the XGBoost algorithm is used to quantify the importance of each meteorological feature on load change, and the MIC maximum information coefficient is used to calculate the linear / nonlinear correlation between features. The importance and correlation of features are comprehensively considered to form a meteorological feature set;

[0055] S13. Combining the obtained meteorological feature set with the historical load data and the one-hot encoded date variable to form a complete diffusion model condition input feature set;

[0056] S14. Normalize the historical meteorological data characteristics and historical load data characteristics as subsequent input. The specific processing formula is as follows:

[0057]

[0058] In the formula, y * is the normalized feature data, y * ∈[0,1], y is the original feature data; y max and y min is the maximum and minimum value of the corresponding data in the total sample data;

[0059] S2. Construct a high-frequency enhanced denoising diffusion model, based on the historical load data and the forward diffusion process of the diffusion model, to form the noisy net load data, as follows:

[0060] Noise is obtained by randomly adding Gaussian noise. The noise addition formula is:

[0061]

[0062]

[0063] In the formula, It is a preset value that decreases gradually, indicating the size of the noise added in the corresponding step. The value range can be set to 0-0.5, and finally the noise data tends to Gaussian distribution. is Gaussian noise, t is the set number of forward diffusion steps, and I is the identity matrix of Gaussian distribution;

[0064] S3. Use the conditional feature set and the noisy net load data to train the high-frequency enhanced denoising diffusion model. After reaching the use accuracy, input the noise data and conditional features to achieve accurate prediction of the noise, and generate the initial load probability prediction result through the reverse denoising process;

[0065] According to the technical solution of this embodiment, the high-frequency enhanced denoising diffusion model includes an encoder and a decoder. The encoder is a CNN-LSTM network and a Linear layer based on fusion wavelet attention, which is used to encode input noise data and conditional features; the decoder is a multi-channel convolutional neural network, which is used to accurately predict noise.

[0066] It should be noted that based on the Inception v2 architecture, multi-view features are extracted, and a multi-channel convolutional neural network integrating the maximum, average and attention pooling layers is constructed to decode the noise and output the noise prediction value. The attention pooling layer captures features through a multi-head attention mechanism, and its expression is as follows:

[0067]

[0068] Where: head h is the single-head self-attention value; W o is a linear transformation matrix used to integrate multi-head information;

[0069] Furthermore, the cross-attention mechanism is selected as the conditional guidance architecture, based on which the correlation between the input noise value and the conditional features is mined, and the information is fused as the decoder input;

[0070] The input of the cross attention mechanism is the classic attention mechanism, which embeds the noise data and diffusion steps and generates the key vector (K) and value vector (V) through the self-attention mechanism, inputs the conditional features and generates the query vector (Q) through the self-attention mechanism, and inputs Q, K, and V into the cross attention for weighted input;

[0071] Furthermore, the encoding embedding part combines discrete wavelet transform and self-attention mechanism, performs multi-level wavelet decomposition on the features of linear layer mapping, and uses self-attention mechanism in frequency domain to adaptively enhance high-frequency detail components. After high-frequency enhancement, CNN-LSTM is used to extract local key and long-term dependency information.

[0072] In order to reduce the computational burden, discrete wavelet transform is used, and its calculation formula is as follows:

[0073]

[0074] Where: α 0 is the initial scale parameter; τ is the translation parameter; Ψ * (•) is the wavelet basis function; m is the scaling constant (indicating the order of decomposition); n is the translation constant; m , n All are integers; t is the original signal time variable;

[0075] The formula for high-frequency component enhancement based on the self-attention mechanism is as follows:

[0076]

[0077] Where: softmax represents normalization processing; Q , K , V is the matrix after the coefficients are linearly transformed, and the calculation formula is as follows:

[0078]

[0079] Where: , and is the parameter matrix to be learned; For the m The detail coefficient matrix of the layer;

[0080] S4. Obtain the initial load prediction value at the confidence interval quantile based on kernel density estimation, and then fine-tune and update the parameters of the high-frequency enhancement denoising diffusion model through the improved loss function;

[0081] Among them, the improved loss function is as follows:

[0082]

[0083]

[0084]

[0085] In the formula, is the improved loss function; is the original loss function calculated by the MSE indicator; is the designed auxiliary loss function; λ is the weighted weight; For the t The actual noise value added in the first step; The noise neural network predicts t step noise value, c is the conditional input, x t is the historical load value Forward Diffusion t The noise data obtained in the first step; and For m ub and m lb diffusion t The noise data obtained by step, where m ub and m lb is the load value of the initial prediction result at the upper and lower limits of the confidence interval, which can be obtained by kernel density estimation;

[0086] The calculation formulas for kernel density estimation and prediction interval are as follows:

[0087]

[0088]

[0089]

[0090] Where: x is the probability density function f ( x )’s independent variable, x ∈R; F (•) is the cumulative distribution function obtained by kernel density estimation; n 1 and h are the estimated load quantity and bandwidth respectively; x η For the estimation of η elements; K (•) is the kernel function; q α is the quantile value at the nominal quantile level α; γ is the nominal coverage probability of the probability prediction interval; and is the load probability prediction interval P 1 , γ Nominal quantile levels for upper and lower boundaries;

[0091] S5. Using the updated high-frequency enhanced denoising diffusion model, the final net load probability prediction result is obtained through back diffusion.

[0092] Next, the present invention is further described in conjunction with specific embodiments:

[0093] To verify the effectiveness of the above prediction method, this embodiment performs a case analysis based on the actual net load data of the distribution network in a certain area in East China. The data time resolution is 5 minutes, the load unit is MW, and the time span is 8 months. At the same time, the local weather data of the corresponding granularity is obtained through SolcastAPI. Based on the above data, the short-term net load probability prediction of the day ahead is realized through Python coding in the Pycharm development environment. The following are the specific implementation steps:

[0094] Step 1) The historical net load data of the distribution network and the multi-dimensional meteorological data are divided into a data set in a ratio of 8:2, with a training set of 196 days and a validation set of 49 days. XGBoost is used to analyze the differential importance of the impact of various meteorological elements on the net load in different scenarios and the correlation between features, and the historical load data and date variables are combined to form a conditional input feature set;

[0095] Step 2) The input distribution network historical net load data and multi-dimensional meteorological data are divided into two scenes: daytime and nighttime, with 6 am and 6 pm as the boundary. For different scenes, the XGBoost algorithm is used to quantify the importance of each meteorological feature on the load change, and the MIC maximum information coefficient is used to calculate the linear / nonlinear correlation between the features.

[0096] Step 3) Comprehensively consider the importance and relevance of features to form a meteorological feature set, and combine historical load data and one-hot encoded date variables to form a complete diffusion model condition input feature set;

[0097] Step 4) Normalize the selected temperature, radiation, and historical load data characteristics of the previous five days. The normalization formula is as follows:

[0098]

[0099] In the formula, y is the original data; y max and y minis the maximum and minimum value of the corresponding data in the total sample;

[0100] Step 5) Generate noise data based on the forward diffusion process. During forward diffusion, the noise is added at the beginning. β 0 is 1e-3, and finally β T is 0.5, and the number of diffusion steps is 200;

[0101] The CNN-LSTM network and Linear layer integrated with wavelet attention are used to encode the input noise data and conditional features, and then a multi-channel convolutional network is used to accurately predict the noise, and the initial load probability prediction result is generated through the reverse denoising process;

[0102] The initial learning rate of the model is set to 0.001, the batch size is set to 16, and the epoch is set to 200. During reverse diffusion, the number of random noise groups is 200 and the noise dimension is 288;

[0103] Step 6) Obtain the initial load forecast value at the confidence interval quantile based on kernel density estimation, and then fine-tune the model parameters through the improved loss function;

[0104] Step 7) Based on the updated model, reverse diffusion generates the final net load probability prediction result. The probability prediction result is as follows: Figure 2-Figure 4 As shown, the red line represents the actual value, the black line represents the maximum probability, and the gray area represents the 95% confidence space; at the same time, a variety of load probability prediction methods with superior performance are selected for comparative analysis. The model proposed in this embodiment is a high-frequency enhanced diffusion model (High-frequency enhanced diffusion model, HFEDM), and the comparative models include a Dropout-based uncertainty estimation prediction model, a deep ensemble (Deep ensembles, DE) model, and an improved quantile long short-term memory (IQLSTM) neural network;

[0105] Step 8) The prediction results were evaluated using multi-dimensional indicators such as mean absolute error (MAE), root mean square error (RMSE), prediction interval coverage error (PICE), prediction interval average width (PIAW), Winkler score Sw, and continuous ranked probability score (CRPS). The comparative evaluation results are shown in Table 1. At the same time, the time consumption of model training was analyzed, and the results are shown in Table 2.

[0106] Table 1 Comparison of multidimensional prediction performance indicators of different models

[0107]

[0108] Table 2 Comparison of training time of different models

[0109]

[0110] Figure 2-Figure 4 The load probability prediction intervals for different typical days and the prediction curves of the maximum probability points obtained by kernel density estimation for the prediction method described in this embodiment are shown. It can be seen from the figure that for typical day 2 with sunny weather as the main weather tone, only a small prediction deviation occurs in local time periods, and the true values ​​for the rest of the day basically fall within the 50% confidence interval and are consistent with the point prediction values. For typical days 1 and 3 under cloudy and overcast weather conditions, the increase in prediction uncertainty for the entire time period leads to a significant expansion of the interval range, and the volatility between adjacent time points also increases, but HFEDM can still dynamically adjust the prediction interval to achieve effective coverage of the true load values ​​under a variety of mutation scenarios.

[0111] Table 1 shows the prediction performance evaluation results of different models based on multi-dimensional indicators. It can be seen from Table 1 that for the average performance of multiple days, HFEDM gives MAE of 0.9166, RMSE of 1.1644, PIAW of 4.1290, and S wThe MAE and CRPS are 4.9112 and 0.6545, respectively. Although the index values ​​fluctuate in different daily scenarios, they are still significantly better than the other comparison models overall. DE, which performs better in the comparison models, shows more stable performance in MAE and CRPS indicators, but the mean values ​​of various indicators are still quite different from those of HFEDM. This is because HFEDM improves the uncertainty aggregation ability under conditional constraints through adaptive high-frequency enhancement, multi-scale feature capture and model fine-tuning, and takes into account the accurate characterization of the prediction single point and probability value distribution while having a narrow prediction interval.

[0112] Table 2 is a comparison of the training time of different prediction models, among which IQLSTM only trains the model parameters of 5 quantiles. It can be seen from Table 2 that the HFEDM model proposed in this embodiment takes longer to train compared with methods such as Dropout due to the increased complexity and expansion of the network. However, the time of less than 3 minutes is acceptable for the day-ahead load probability forecasting and most load forecasting scenarios, and there will be no high training cost problem similar to DE, so the method of this embodiment still has good applicability.

[0113] In summary, the present invention constructs a conditional diffusion model architecture based on the cross-attention mechanism, and uses the XGBoost algorithm and the maximum information coefficient to extract key features as input; for the encoding-decoding link, a CNN-LSTM network and a multi-channel convolutional neural network integrating the wavelet attention mechanism are designed respectively to enhance the ability to capture high-frequency features, and then based on the initial prediction value and the kernel density estimation method, the model is fine-tuned using an improved loss function, and the final net load prediction result is obtained by performing multi-step reverse denoising on the diffusion model after parameter update. The present invention can improve the ability to accurately characterize complex load fluctuation patterns and the stability of prediction results through adaptive enhancement of high-frequency signals and effective extraction of multi-scale features, thereby providing a more reliable decision-making basis for distribution network planning, production, and operation scheduling.

[0114] 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 a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0115] For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. When an element is referred to as being "assembled on", "installed on", "fixed on" or "set on" another element, it can be directly on the other element or there can also be a centered element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be a centered element at the same time. The terms "vertical", "horizontal", "up", "down", "left", "right" and similar expressions used herein are only for illustrative purposes and are not intended to be the only implementation method.

[0116] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0117] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

Claims

1. A short-term net load probability prediction method based on a high-frequency enhanced denoising diffusion model is characterized by: The following steps are involved: Receive the historical net load data of the distribution network and the multi-dimensional meteorological data, screen the conditional features through the XGBoost algorithm and the MIC maximum information coefficient, and construct a conditional feature set based on the screened conditional features; Construct a high-frequency enhanced denoising diffusion model to form noisy net load data based on historical load data and the forward diffusion process of the diffusion model; The high-frequency enhanced denoising diffusion model is trained using the conditional feature set and the noisy net load data. After reaching the usage accuracy, the noise data and conditional features are input to achieve accurate noise prediction, and the initial load probability prediction result is generated through the reverse denoising process. The initial load prediction value at the confidence interval quantile is obtained based on kernel density estimation, and then the parameters of the high-frequency enhancement denoising diffusion model are fine-tuned and updated through the improved loss function; The updated high-frequency enhanced denoising diffusion model is used to obtain the final net load probability prediction result through reverse diffusion; Construct a conditional feature set, including the following: (1) Taking 6 am and 6 pm as the boundary, the input distribution network historical net load data and multi-dimensional meteorological data are divided into two scenes: daytime and nighttime; (2) For different scenarios, the XGBoost algorithm is used to quantify the importance of each meteorological feature on load changes, and the MIC maximum information coefficient is used to calculate the linear / nonlinear correlation between features. The importance and correlation of features are comprehensively considered to form a meteorological feature set; (3) The obtained meteorological feature set is combined with the historical load data and the one-hot encoded date variable to form a complete diffusion model condition input feature set; The high-frequency enhancement denoising diffusion model includes an encoder and a decoder, wherein the encoder is a CNN-LSTM network and a Linear layer based on fusion wavelet attention, which is used to realize the encoding of input noise data and conditional features; The decoder is a multi-channel convolutional neural network, which is used to achieve accurate prediction of noise; The cross-attention mechanism is selected as the conditional guidance architecture. The correlation between the input noise value and the conditional features is mined based on the conditional guidance architecture, and the information is fused as the decoder input. Construct a CNN-LSTM network based on fusion wavelet attention, as follows: (4) Combine discrete wavelet transform and self-attention mechanism to perform multi-level wavelet decomposition on the features mapped by linear layers; (5) A self-attention mechanism is used in the frequency domain to adaptively enhance the high-frequency detail components. After high-frequency enhancement, CNN-LSTM is used to extract local key and long-term dependency information.

2. The short-term net load probability prediction method based on high-frequency enhanced denoising diffusion model according to claim 1 is characterized by: The historical meteorological data characteristics and historical load data characteristics are normalized as subsequent inputs. The specific processing formula is as follows: In the formula, y * is the normalized feature data, y * ∈[0,1], y is the original feature data; y max and min is the maximum and minimum value of the corresponding data in the total sample data.

3. The short-term net load probability prediction method based on high-frequency enhanced denoising diffusion model according to claim 2 is characterized by: The discrete wavelet transform is calculated as follows: Where: α0 is the initial scale parameter; τ0 is the translation parameter; Ψ * (·) is the wavelet basis function; m is the scaling constant; n is the translation constant; m and n are both integers; t is the time variable of the original signal.

4. The short-term net load probability prediction method based on high-frequency enhanced denoising diffusion model according to claim 3 is characterized by: The high-frequency component enhancement formula of the self-attention mechanism is as follows: Where: softmax represents normalization processing; Q, K, V are matrices after coefficient linear transformation, and the calculation formula is as follows: Q=X H,m ×W Q K=X H,m ×W X V=X H,m ×W V Where: W Q , W x and W v is the parameter matrix to be learned; X H，m is the detail coefficient matrix of the mth layer.

5. The short-term net load probability prediction method based on high-frequency enhanced denoising diffusion model according to claim 4 is characterized by: The improved loss function is as follows: THE refine =L o +λL r L o =MSE(∈ t ,∈ θ (x t ,c,t)) Where, L refine is the improved loss function; L o is the original loss function calculated by the MSE indicator; L r is the designed auxiliary loss function; λ is the weighted weight; ∈ t is the real noise value added in the tth step; ∈ θ (x t ,c,t) is the noise value of the t-th step predicted by the noise neural network, c is the conditional input, x t is the noise data obtained by forward diffusion of the historical load value X0 for t steps; and For m ub and m lb Diffusion t steps of noise data, where m ub and m lb is the load value of the initial prediction result at the upper and lower limits of the confidence interval, which can be obtained by kernel density estimation.

6. The short-term net load probability prediction method based on high-frequency enhanced denoising diffusion model according to claim 5 is characterized by: The calculation formulas for the kernel density estimation and prediction interval are as follows: what α =F(α) -1 Where: x is the independent variable of the probability density function f(x), x∈R; F(·) is the cumulative distribution function obtained by kernel density estimation; n1 and h are the estimated load quantity and bandwidth respectively; x η is the ηth element used for estimation; K(·) is the kernel function; q α is the quantile value at the nominal quantile level α; γ is the nominal coverage probability of the probability prediction interval; and α is the load probability prediction interval P 1,γ Nominal quantile levels for the upper and lower boundaries.

Images