A method for power system load forecasting based on conditional guided diffusion process

Abstract

The present invention discloses a method for load forecasting of a power system based on a conditional guided diffusion process, which belongs to the field of time series forecasting of a power system. The method designs a multi-stage progressively guided denoising process to solve the problem of uncontrollable denoising process; effectively discovers pattern features through similar period load curve mining algorithms and multi-resolution trend features, and uses a mixed feature condition generation network to guide the multi-stage denoising process; finally, a deep denoising network based on the Transformer architecture is designed, and the model's ability to restore load curves from Gaussian noise is improved by embedding a deep frequency domain decomposition module. The method realizes the controllability of the denoising process, ensures the precision and accuracy of load curve generation, enhances the interpretability of the model, significantly improves the accuracy of day-ahead load point prediction and deterministic prediction, and can provide more accurate prediction information for power system control and optimized operation.

Description

Technical Field

[0001] The invention relates to a method for predicting load of an electric power system based on a condition-guided diffusion process, and belongs to the field of electric power system time series prediction. Background Art

[0002] As the scale of new energy access to the grid continues to increase, the uncertainty of new energy output has an increasingly obvious impact on the stability of the power system. The technical difficulty of achieving accurate load forecasting continues to increase, posing great challenges to the optimal scheduling, safe operation and risk warning of the power system.

[0003] Load probability forecasting is a technical method that quantitatively describes the uncertainty of load in the forecast interval based on the external characteristic information and historical pattern of power load. Classical probability forecasting methods include statistical methods represented by linear regression model, autoregressive moving average model, and Bayesian model; machine learning methods represented by extreme learning machine ELM, long short-term memory neural network LSTM, graph neural network GNN and other methods. However, most of the above methods rely on historical sequences of forecast features and lookback windows to construct forecast inputs, and fail to fully capture the periodicity and morphological characteristics of load curves in load forecasting scenarios. This limitation makes it difficult to ensure the prediction stability of the model when facing the non-stationary and complex fluctuations of load curves. In contrast, the conditional diffusion model can learn more complex probability distributions during training to cope with the uncertainty and complex periodic patterns of load curves. It has advantages in capturing load fluctuations caused by changes in user electricity consumption behavior, and the samples it generates can well reflect the potential probability distribution of data.

[0004] In recent years, the application of conditional diffusion models in probabilistic forecasting tasks has developed rapidly, but there are still the following shortcomings: 1) The iterative denoising process of recovering load data from noise is uncontrollable, and the quality of curve generation is difficult to guarantee; 2) The construction of conditional information is mostly based on simple feature splicing, which cannot reflect the weight differences of different features and cannot effectively extract complex pattern information; 3) Prediction research based on diffusion models mostly focuses on improving prediction accuracy and generation quality, and the interpretability of the model itself needs to be improved. Summary of the invention

[0005] The purpose of the present invention is to provide a method for power system load forecasting based on conditional guided diffusion process in view of the deficiencies of the prior art.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A method for power system load forecasting based on conditional guided diffusion process comprises the following steps:

[0008] Step 1: Obtain historical temperature, humidity, year, month, day characteristics, and weather forecast information for the period to be predicted, and explore load curves for similar periods;

[0009] Step 2: The load curve of the similar period obtained in step 1 is used as a guiding condition and input into the denoising model guided by the condition generated by the hybrid feature network to perform multi-stage denoising to obtain the load curve of the period to be predicted;

[0010] Step 3: Repeat step 2 to obtain a set of load curves for the period to be predicted; use kernel density estimation to statistically analyze the set of load curves for the period to be predicted to obtain the load probability prediction result for the period to be predicted; select the median of the probability prediction interval at each moment as the deterministic prediction result for the period to be predicted.

[0011] In the above technical solution, further, step 1 is to discover the load curve of similar period based on the similar period load curve discovery algorithm of the subdivided scene matching, and the specific method is:

[0012] 1) Select historical temperature, humidity, year, month, and day features to form a feature matrix set for similarity measurement. Each feature vector corresponds to a historical load curve.

[0013] 2) Segment the feature matrix set according to different scenarios, specifically: divide the feature matrix set into working day / holiday subsets according to national statutory working days; divide the feature matrix set into sunny day / rainy day / cloudy day subsets according to weather scenarios; divide the feature matrix set into high temperature / normal / low temperature subsets according to extreme temperature scenarios;

[0014] 3) Construct a feature matrix based on the weather forecast information of the time period to be predicted, search for the most similar load curve in the subset obtained in step 2), and finally take the average of the most similar load curves searched from different subsets as the optimal load curve for the similar period.

[0015] Furthermore, the similar period load curve mining algorithm based on segmented scene matching is used to mine similar period load curves, and the specific steps are as follows:

[0016] First, the feature matrix set for similarity measurement is obtained based on the historical temperature, humidity, year, month, and day features. , represents the number of historical patterns (one similar period corresponds to one historical pattern); the feature matrix in the set F , is the number of historical features obtained; is the dimension of each eigenvector; each row of the eigenvector in the eigenmatrix corresponds to a historical load curve. Historical load curve The corresponding eigenvector The matching forms a pair of historical pattern vectors. Accordingly, a historical pattern set corresponding to the feature matrix set can be obtained. , which can be specifically expressed as follows:

[0017]

[0018] After determining the forecast period, based on the weather forecast information of the forecast period, the load scenario is matched for each subdivided scenario, and all relevant historical pattern subsets are selected (historical pattern subsets are grouped according to weather scenarios). Divided into multiple subsets), the load curve of a similar period that is closest to the load curve of the period to be predicted can be searched in the above subsets (find the corresponding historical load curve based on the feature vector). The KNN algorithm based on MIC coefficient weighting is used to discover the most similar curve. The specific method is as follows:

[0019] First, calculate the The maximum information coefficient of the feature :

[0020]

[0021] In the formula, is the feature number, ; express Middle Eigenvectors of rows; is the eigenvector With the historical load curve The mutual information coefficient between ; It is the upper limit of the gridded time series data.

[0022] Then, taking the maximum information coefficient of each feature as the weight, the KNN algorithm based on MIC coefficient weighting is applied to the historical pattern subsets of each matching scenario to screen out the load curve closest to the forecast feature of the forecast period. A collection of subdivided scene modes. A subset of historical patterns Calculate the feature distance based on MIC weighting , the specific calculation formula is as follows:

[0023]

[0024] In the formula, is the characteristic matrix of the forecast period obtained through weather forecast, Represents the feature matrix Middle Eigenvectors of rows; is the feature distance function based on MIC weighting;

[0025] According to the characteristic distance based on MIC weighting, the closest historical load curve is selected as the most similar load curve in each subset; thus, in the historical pattern subset The most similar load curve is obtained by filtering .

[0026] The most similar load curves of each sub-scenario are averaged to obtain the optimal load curve for similar periods. .

[0027] Furthermore, in step 2, multi-stage denoising is performed based on a denoising model guided by a hybrid feature network generation condition. Unlike the existing diffusion model that directly restores the original time series from random noise, the present invention designs a multi-stage denoising process guided by a hybrid feature condition generation network, which enhances the controllability of the diffusion process and realizes the refined generation of the load curve. The entire denoising process can be divided into In this stage, the historical curve characteristics and the load curves of similar periods are used as model inputs in the model training and inference process, guiding the model to restore the load curve of the period to be predicted from random noise from coarse to fine (i.e., the resolution is from small to large).

[0028] Furthermore, the method for constructing the denoising model guided by the hybrid feature network generation condition is:

[0029] The whole denoising process is divided into The denoising model guided by the hybrid feature network generation condition is designed as The same serial structure; multi-resolution feature extraction is performed on the historical load curve to obtain the historical curve characteristics of the corresponding stage; The historical curve characteristics of each stage and the load curve of similar periods are used as model input, and the load curve of the period to be predicted is used as model output; the training process of the denoising model guided by the generation condition of the hybrid feature network is as follows:

[0030] 1) The historical curve characteristics of each stage and the load curve of similar periods are taken as the The input of the structure is used to extract features using a hybrid feature condition generation network, and the feature extraction results are used as guiding conditions to input the deep denoising network for denoising;

[0031] 2) The denoising results of the structure -1 stage historical curve characteristics and similar period load curve as the first -1 structure input, use the hybrid feature condition generation network to extract features, and use the feature extraction results as guiding conditions to input the deep denoising network for denoising;

[0032] 3) For the - 2 structures, -3 structures... Repeat step 2) until the first structure;

[0033] 4) Repeat steps 1)-3) until the set number of times or the loss converges to the set threshold; during the training process, The losses of each stage are accumulated as the loss of one training.

[0034] Furthermore, the multi-resolution feature extraction of the historical load curve is performed to obtain the historical curve features of the corresponding stage. The specific method is: multi-resolution feature extraction of the historical load curve is performed using a sliding average pooling operation, from the first From the first stage to the second stage, the average pooling kernel changes from large to small, thus ensuring that the resolution of the extracted features changes from large to small.

[0035] Furthermore, the multi-stage denoising process is guided by a mixed feature condition generation network. The present invention uses multi-resolution historical curve features and similar period load curves as network input information to better guide the model to complete the reconstruction of the time series. Since the length of the historical curve feature is different from the length of the load curve of the period to be predicted, it is necessary to downsample it through linear mapping to obtain a feature vector representation of the historical curve feature. The prediction result of the previous stage is combined with the load curve of the similar period to obtain , the formula is as follows:

[0036]

[0037] In the formula, represents the Hadamard product, The parameter matrix is ​​randomly initialized and has a value range of [0,1]. Its weights can be continuously updated during training. is the optimal load curve for similar periods, and They represent the load curve prediction result and the real load curve of the predicted period in stage n+1 respectively. Considering the different model data flows in the training and reasoning processes, formula (a) is the feature fusion method in the model reasoning process; formula (b) is the feature fusion method in the model training process.

[0038] Furthermore, the deep denoising network retains the basic architecture of the Transformer model as a whole, which consists of multiple serial encoders and decoders. In order to further improve the model's ability to capture the periodic characteristics of the input sequence, a parallel deep frequency domain decomposition module is embedded in the feedforward neural network in the decoder, and the input sequence is restored to periodic terms and residual terms in the step-by-step decomposition of multiple serial decoders. The deep denoising network is specifically as follows: using Transformer as the basic structure, adding a parallel deep frequency domain decomposition module to the output part of the multi-head cross attention module, the deep denoising network takes randomly generated Gaussian noise as input, completes denoising under conditional guidance, and outputs phased prediction results; the working method of the deep denoising network is as follows: first, Transformer processes the input Gaussian noise in combination with the guiding conditions, and outputs it through the multi-head cross attention module; then, the deep frequency domain decomposition module is used to perform fast Fourier transform on the output of the multi-head cross attention module, obtain the periodic component from the output of the multi-head cross attention module through fast Fourier transform, select the basis with the most significant amplitude in the frequency domain as the main frequency component of the current sequence, and obtain the main periodic features of the output of the multi-head cross attention module in the time domain after inverse Fourier transform and output mapping layer; extract the mean from the output of the decoder's feedforward neural network, merge the mean with the main periodic features, and obtain the periodic term output; subtract the periodic term from the output of the Transformer feedforward neural network to obtain the residual term; the residual term and the periodic term are superimposed to obtain the load curve of the time period to be predicted in the current stage. The prediction result of the first stage is the prediction output of the model, that is, the load curve of the forecast period obtained by the model prediction.

[0039] Furthermore, the hybrid feature condition generation network specifically includes a feature fusion module, a linear mapping module and a feature channel convolution module; wherein the feature fusion module is used to perform feature fusion on the output result of the previous stage with the load curve of a similar time period; the linear mapping module is used to downsample the historical curve features of the corresponding stage, so that the downsampled result is aligned with the load curve of a similar time period; the feature channel convolution module is used to perform channel convolution on the splicing result of the output of the feature fusion module and the output of the linear mapping module, and the channel convolution result is aligned with the data dimension of the deep denoising network.

[0040] The beneficial effects of the present invention are:

[0041] The present invention proposes a load forecasting method for a power system based on a condition-guided diffusion process. The method realizes the process controllability of the conditional diffusion model in the generation of load curves through the mechanism design of multi-stage guided generation; and effectively discovers the pattern characteristics through the similar period load curve mining algorithm and multi-resolution trend characteristics (i.e., historical curve characteristics). The method generates a condition-guided denoising model based on a hybrid feature network for multi-stage progressive denoising to solve the problem of uncontrollable denoising process; the condition-guided denoising model generated by the hybrid feature network includes a hybrid feature condition generation network and a deep frequency domain decomposition module. The hybrid feature condition generation network is used to efficiently extract condition information from multi-resolution trend information and similar period load curves, thereby guiding the refined generation of load curves; the deep denoising network of the deep frequency domain decomposition module improves the model's ability to restore load curves from Gaussian noise, and enhances the transparency and interpretability of the denoising model. The method realizes the controllability of the denoising process, ensures the precision and accuracy of load curve generation, and significantly improves the accuracy of deterministic prediction of power system loads. The comparative experiment also verifies the excellent prediction ability of the present invention in the load probability prediction task. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of the algorithm for discovering similar period load curves based on segmented scenario matching;

[0043] Figure 2 It is a schematic diagram of the denoising process of the denoising model guided by the generation condition based on the hybrid feature network;

[0044] Figure 3 It is a schematic diagram of the multi-resolution feature extraction method;

[0045] Figure 4 It is a schematic diagram of the deep denoising network structure;

[0046] Figure 5 (a) and (b) are the day-ahead probability forecast results of regional load and 10kV medium voltage load, respectively. DETAILED DESCRIPTION

[0047] The following is further described based on the accompanying drawings and implementation examples.

[0048] An embodiment of the present invention provides a method for forecasting power system load based on a conditional guided diffusion process. The method is used to forecast the day-ahead load, specifically comprising the following steps:

[0049] Step 1: Obtain historical temperature, humidity, year, month, day characteristics, and weather forecast information for the day to be predicted, and discover similar day curves. The complete process is as follows Figure 1 shown.

[0050] First, data preprocessing is performed to remove unreasonable abnormal data, and adaptive interpolation is used to fill missing values. In the scene division part of similar daily curve mining, holidays include weekends and national statutory holidays and take into account the arrangement of holidays, and the rest are regarded as working days; the basis for weather scene division is the local historical weather forecast information; the judgment standard for high and low temperatures in extreme scenes is whether the meteorological bureau issues relevant warning information on the same day, and no warning information is issued. It is considered normal. For year characteristics, different year information is encoded as a two-digit One-Hot code; for month and day characteristics, sine and cosine codes are introduced to fully reflect the periodic characteristics of date changes:

[0051]

[0052]

[0053] In the formula, and Respectively represent the feature vectors formed by the monthly feature and the daily feature; Repeat function means repeating the sine and cosine calculation results multiple times until they are aligned with the length of the temperature and humidity feature vectors. After splicing the temperature, humidity, year, month, and day features, the feature vector of each historical scene can be obtained, and the feature vectors are integrated to form a feature matrix set corresponding to the historical load mode. Each feature vector corresponds to a historical load curve.

[0054] Then, the feature vector set is subdivided according to different scenarios, specifically: the feature matrix set is divided into working day / holiday subsets according to national statutory working days; the feature matrix set is divided into sunny day / rainy day / cloudy day subsets according to weather scenarios; the feature matrix set is divided into high temperature / normal / low temperature subsets according to extreme temperature scenarios;

[0055] According to the historical data obtained, the maximum information coefficient MIC between the characteristics (temperature, humidity, year, month, day) and the historical load curve is calculated to determine the correlation between each characteristic and the load; each historical load curve is matched with its corresponding characteristic vector to form a pair of historical pattern vectors, and accordingly, a historical pattern set corresponding to the characteristic matrix set can be obtained. . According to the national statutory working days, historical patterns are collected Divide into weekday / holiday historical pattern subsets; group historical patterns according to weather scenarios Divide into sunny / rainy / cloudy historical pattern subsets; group historical patterns according to extreme temperature scenarios Divided into high / normal / low temperature history mode subsets.

[0056] Finally, a feature matrix is ​​constructed based on the meteorological forecast information of the day to be predicted (specifically including forecast temperature, humidity, year, month, and day), and the matching of segmented scenarios is completed according to the meteorological information of the day to be predicted. For example, if the day to be predicted is a working day, cloudy day, or normal temperature, the KNN algorithm based on MIC weighting is used to search for similar day curves in the working day subset, cloudy day subset, and normal temperature subset respectively. The three most similar daily load curves obtained in the three subsets (the KNN algorithm based on MIC coefficient weighting is used to discover the most similar curves) are averaged to obtain the optimal similar day curve.

[0057] Step 2: The similar daily curve obtained in step 1 is used as a guiding condition and input into the denoising model guided by the conditional generation based on the hybrid feature network to perform multi-stage denoising to obtain the load curve of the day to be predicted.

[0058] like Figure 2 It is a schematic diagram of the denoising process of the denoising model guided by the conditional generation of the hybrid feature network (TG-DDPM for short).

[0059] by Figure 2 In Taking the stage as an example, the input information of the model is divided into two parts: historical information and forecast information. The former is the trend characteristics extracted from the historical load curve (i.e., historical curve characteristics). , the latter includes similar daily load curves and trend features extracted from the forecast values ​​(i.e., the denoising results of stage 3) , the input information is passed into the hybrid feature condition generation network to form the guiding condition. At the same time, the prediction target of this stage Perform forward diffusion to obtain the trend sequence after adding noise , the deep denoising network predicts and outputs more refined day-ahead load trend features from the input noise sequence under the guidance of the hybrid feature condition generation network , according to the deviation between the predicted value and the true value, The loss value predicted by the stage model Similarly, the prediction losses of the other two stages can be calculated and In single-step prediction, the calculation formula for the overall prediction loss of the model is:

[0060]

[0061] In the above multi-stage denoising process, a sliding average pooling operation is used to extract the multi-resolution trend features of the input sequence (i.e., historical load curve), such as Figure 3 As shown, as the trend-guided diffusion proceeds, the number of stages from Starts to gradually decrease to 1, the sliding core length and The positive correlation decreases with the deepening of guided denoising, so the model can extract features with different resolutions from coarse to fine. Similarly, for the predicted real load curve Repeat the same operation to obtain the actual load trend characteristics to be predicted. , and .by , and As the true value, the model training process is supervised to obtain a trained denoising model guided by the generation conditions of the hybrid feature network.

[0062] like Figure 4 The deep denoising network of the present invention is shown in FIG. To adapt to the task of restoring the original sequence features from the noisy sequence, the deep denoising network uses a diffusion step and forward diffusion sequence The original structure is replaced by a diffusion embedding layer, and the diffusion step is encoded by sinusoidal position embedding; the diffusion step is converted to The coded information injection The adaptive normalization process is expressed as follows:

[0063]

[0064] In the formula, represents the adaptive normalization layer, SiLU is the activation function between network layers, t represents the number of diffusion steps, Represents the Layer Norm layer normalization operation, is the diffusion step embedding representation of the input normalization layer, and the parameter and All are made by Linear transformation is obtained.

[0065] In the decoder, there is a complete multi-head attention module and a cross attention module (the two constitute a multi-head cross attention module). The multi-head attention module is used to extract multi-scale feature information from the input noise sequence; the cross attention module is used to output The characteristic information of trend guiding conditions is effectively combined in the calculation. The specific calculation process is as follows:

[0066]

[0067] In the formula, represents the cross attention layer, is the dimension of the model backbone network. The encoder output that incorporates the guided condition Parameter matrix of the mapping matrix used to generate the query matrix and key matrix , Multiply to get and Matrix, substitute it into the formula to calculate the updated cross attention weight, where The generation process is as follows:

[0068]

[0069] In the formula, is the encoder output feature, The output condition of the network is generated by the mixed feature condition, which is formed by concatenating Concat and channel convolution Convld , Output features for the encoder that contain conditional guidance information.

[0070] like Figure 4 As shown in the internal structure of the deep frequency domain decomposition module, the deep frequency domain decomposition module is used to perform fast Fourier transform on the output of the multi-head cross attention module, obtain the periodic component from the output of the multi-head cross attention module through fast Fourier transform, select the basis with the most significant amplitude in the frequency domain as the main frequency component of the current sequence, and obtain the main periodic features of the output of the multi-head cross attention module in the time domain after inverse Fourier transform and output mapping layer; extract the mean from the output of the decoder's feedforward neural network, merge the mean with the main periodic features, and obtain the periodic term output; the difference between the periodic term and the output of the Transformer feedforward neural network is the residual term. The extraction of the main frequency component by the deep frequency domain decomposition module in the multi-stage cascade decoder can ensure that the denoising model learns the inherent trends and details in the random noise sequence.

[0071] The hybrid feature condition generation network specifically includes a feature fusion module, a linear mapping module and a feature channel convolution module; wherein the feature fusion module is used to perform feature fusion on the output result of the previous stage with the similar daily curve; the linear mapping module is used to downsample the historical curve features of the corresponding stage, so that the downsampled results are aligned with the similar daily curve; the feature channel convolution module is used to combine the output of the feature fusion module The output of the linear mapping module The splicing result is subjected to channel convolution, and the channel convolution result is aligned with the data dimension of the deep denoising network.

[0072] The load data point resolution used in this embodiment is 30 minutes, and contains 48 data points per day. The backbone network dimension of the deep denoising network is set to 192, the output dimension of the hybrid feature condition generation network is set to 24, and the convolution input and output of the feature channel for concatenating the encoder output and the guide condition are set to 192 and 144 respectively. The trend guidance is divided into three stages, namely , the smoothing kernel is set to , the sliding step size is set to 1. To prevent overfitting and improve the training effect of the model, the Adam optimizer and the linear Warmup strategy are used to warm up the model. In the first 10% of iterations, the model training rate gradually increases from 8e-4 to 1e-3, and linearly decreases to 1e-5 in subsequent training. In the mixed feature conditional generation network, the input dimension of the historical curve feature is 96, the input dimension of the similar daily curve feature is 48, and the dimension of the input random noise is 48. And the noise obtained by random sampling can be used in three stages at the same time.

[0073] Based on the above model configuration, multi-resolution trend extraction is first completed according to the number of trend guidance stages and the setting of the smoothing kernel to obtain the trend characteristics of the historical load data: , and And the trend information of the real load to be predicted: , and At the same time, based on the similar daily curves discovered in step 1, The multi-resolution trend and similar daily curves are input into the hybrid feature generation network in the stage, and the randomly generated Gaussian noise is input into the denoising model. The model will spontaneously perform denoising operations step by step until The load curve generation result of the day to be predicted is output in this stage.

[0074] Repeat the above process of generating the load curve for the day to be predicted, input multiple groups of random Gaussian noise into the model, and you can get multiple groups of load curves for the day to be predicted. Apply kernel density estimation to the generated load curve set to get the probability prediction result of the day's load, and select the median of the probability prediction interval at each moment as the deterministic prediction result of the day to be predicted.

[0075] Table 1 Performance comparison of point prediction results obtained by different prediction methods

[0076]

[0077] Table 2 Performance indicators of day-ahead load probability forecasting under different nominal coverage ratios

[0078]

[0079] In the point prediction part, the model settings and parameter configurations used are the same as those of the probability prediction. The difference between the two is that the output of the point prediction is a fixed value at each moment, while the output of the probability prediction is a probability distribution at each moment. The two indicators of MAPE and RMSE are usually used, that is, the mean absolute percentage error RMS Error To measure the prediction effect:

[0080]

[0081]

[0082] In the formula, represents the number of test samples, is the model prediction value, is the model prediction value, is the true value. From the test results in Table 1, it can be seen that the method proposed in the present invention has good performance in the point prediction scenario, and its prediction error in the four scenarios of spring, summer, autumn and winter is kept within 1.43%. Compared with the comparison model, the average RMSE loss of the method proposed in the present invention is 20% lower than that of the best performing time series prediction Transformer model in the comparison model, and has the highest prediction accuracy, which verifies the effectiveness of the proposed method in the day-ahead load prediction scenario.

[0083] Commonly used performance evaluation indicators in probabilistic forecasting include the average coverage deviation (ACD) of the forecast interval, the average width (AW) of the forecast interval, and the Winkler score used to reflect the comprehensive performance of the forecast results, such as sharpness and reliability. The relevant calculation expressions are as follows:

[0084]

[0085]

[0086]

[0087] In the formula: superscript represents the nominal coverage; , and Indicates the corresponding indicators under the corresponding nominal coverage; is the total number of samples in the test set; is the true sample value; It is an indicator function, and takes the value 1 if the condition in the brackets is true, otherwise it takes the value 0; and The nominal coverage is The upper and lower quantile levels of and is the quantile of the corresponding quantile level; is the actual coverage of the probability prediction results, is the nominal coverage, Indicates nominal coverage The deviation between the actual coverage and the nominal coverage, the smaller the value, the more accurate the interval description of the probability prediction result; Indicates nominal coverage The average width of the prediction interval at each moment; Indicates nominal coverage The comprehensive performance of the probability prediction results is composed of three parts: Reflects the width of the prediction interval, is the loss when the true value exceeds the upper bound of the quantile, is the loss when the true value is lower than the lower bound of the quantile.

[0088] Table 2 summarizes the performance indicators of different probability prediction methods in the four seasons of spring, summer, autumn and winter. The ACD index of the TG-DDPM model proposed in the present invention is lower than 3.2% in the four seasons, and the AW index and Winkler Score index are better than other comparative models. The CEDM model, which is also based on the conditional diffusion model, adopts the network combination of iTransformer+Unet, and achieves an interval width and Winkler Score close to TG-DDPM, but the ACD index is higher, indicating that when the interval width is close, the interval coverage of the method of the present invention is closer to the nominal coverage. CGAN, which is also a generative model, has achieved prediction performance close to that of the method of the present invention in summer and autumn, but the probability prediction effect in spring and winter is worse. TG-DDPM and CEDM based on the diffusion model technology route are better than the CGAN model in the Winkler Score index. This is because compared with the adversarial neural network, the diffusion model does not have feature information loss caused by dimensionality reduction during training and reasoning, and can obtain richer generation patterns during the generation process, so it has stronger comprehensive prediction performance in the probability prediction scenario.

[0089] Figure 5 (a) and Figure 5 (b) in the figure shows the day-ahead load probability prediction results of the present invention for regional load and 10kV medium-voltage load data, including multiple prediction intervals from 90% to 10%. From the visualization results of the prediction, during the load peak and valley periods with higher uncertainty, the probability prediction results of the model give a relatively wider prediction interval; in the load ramp interval with obvious trends, the prediction results of the model show lower uncertainty and give a narrower prediction interval width. Compared with the medium-voltage load prediction results, the regional load prediction results have more obvious volatility and randomness in the power consumption behavior of the medium-voltage load, so the probability prediction interval at each moment is wider. Overall, the prediction interval width generated by the method of the present invention in different prediction scenarios is narrow and accurately surrounds the real load curve, indicating that the method proposed by the present invention can accurately and efficiently learn the complex morphological characteristics of the load curve in different seasons.

Claims

1. A method for power system load forecasting based on conditional guided diffusion process, characterized in that: The following steps are involved: Step 1: Obtain historical temperature, humidity, year, month, day characteristics, and weather forecast information for the period to be predicted, and explore load curves for similar periods; Step 2: The load curve of the similar period obtained in step 1 is used as a guiding condition and input into the denoising model guided by the condition generated by the hybrid feature network to perform multi-stage denoising to obtain the load curve of the period to be predicted; Step 3: Repeat step 2 to obtain a set of load curves for the period to be predicted; use kernel density estimation to perform statistics on the set of load curves for the period to be predicted to obtain a load probability prediction result for the period to be predicted; The median of the probability prediction period at each moment is selected as the deterministic prediction result of the period to be predicted; In step 2, the method for constructing the denoising model guided by the hybrid feature network generation condition is: The whole denoising process is divided into N stages, and the denoising model guided by the hybrid feature network generation condition is designed as N identical serial structures; multi-resolution feature extraction is performed on the historical load curve to obtain the historical curve features of the corresponding stage; the historical curve features of the Nth stage and the load curve of the similar period are used as model input, and the load curve of the period to be predicted is used as model output; the training process of the denoising model guided by the hybrid feature network generation condition is as follows: 1) The historical curve characteristics of the Nth stage and the load curve of similar periods are used as the input of the Nth structure, and the hybrid feature condition generation network is used to extract features. The feature extraction results are used as guiding conditions to input the deep denoising network for denoising; 2) The denoising results of the Nth structure, the historical curve characteristics of the N-1th stage and the load curve of the similar period are used as the input of the N-1th structure, and the hybrid feature condition generation network is used for feature extraction. The feature extraction results are used as the guiding conditions to input the deep denoising network for denoising; 3) Repeat step 2) for the N-2 structure, the N-3 structure, ... until the first structure; Repeat steps 1)-3) until the set number of times or the loss converges to the set threshold; during the training process, the losses of N stages are accumulated as the loss of one training.

2. The method for power system load forecasting based on conditional guided diffusion process according to claim 1, characterized in that: In step 1, a similar period load curve mining algorithm based on segmented scenario matching is used to mine similar period load curves, which specifically includes the following steps: 1) Select historical temperature, humidity, year, month, and day features to form a feature matrix set for similarity measurement. The feature vector of each row in the feature matrix corresponds to a historical load curve; 2) Segment the feature matrix set according to different scenarios, specifically: divide the feature matrix set into working day / holiday subsets according to national statutory working days; divide the feature matrix set into sunny day / rainy day / cloudy day subsets according to weather scenarios; divide the feature matrix set into high temperature / normal / low temperature subsets according to extreme temperature scenarios; 3) Construct a feature matrix based on the weather forecast information of the time period to be predicted, search for the most similar load curve in the subset obtained in step 2), and finally take the average of the most similar load curves searched from different subsets as the optimal load curve for the similar period.

3. The method for power system load forecasting based on conditional guided diffusion process according to claim 2, characterized in that: In step 3), the specific method of searching for the most similar load curve in the subset obtained in step 2) is: The KNN algorithm based on MIC coefficient weighting is used to search for the most similar curve, as follows: First, calculate the maximum information coefficient MIC of the f-th feature f : Where f is the feature number, f = 1, 2, ..., N f , N f is the number of historical features; Represents Y fea The eigenvector of the fth row in ; represents the feature matrix used for similarity measurement, d f is the dimension of each feature vector; is the eigenvector The mutual information coefficient between the load curve y and the historical load curve y; B is the upper limit of the gridded time series data; Then, the feature distance based on MIC weighting is calculated; Calculate the MIC-weighted feature distance D(Y fea ,Y fore ): In the formula, is the feature matrix of the time period to be predicted, Represents the feature matrix Y fore The eigenvector of the fth row in ; Finally, the closest historical load curve is selected as the most similar load curve in each subset according to the characteristic distance based on MIC weighting.

4. The method for power system load forecasting based on conditional guided diffusion process according to claim 1, characterized in that: The multi-resolution feature extraction of the historical load curve is performed to obtain the historical curve features of the corresponding stage. The specific method is: multi-resolution feature extraction of the historical load curve is performed using a sliding average pooling operation. From the Nth stage to the 1st stage, the average pooling kernel decreases from large to small, thereby ensuring that the extracted feature resolution changes from small to large.

5. The method for power system load forecasting based on conditional guided diffusion process according to claim 1, characterized in that: The deep denoising network is specifically as follows: using Transformer as the basic structure, adding a parallel deep frequency domain decomposition module to the output part of the multi-head cross attention module; the deep denoising network takes randomly generated Gaussian noise as input, and outputs a phased prediction result; the working method of the deep denoising network is as follows: first, Transformer processes the input Gaussian noise in combination with the guiding condition, and outputs it through the multi-head cross attention module; then, the deep frequency domain decomposition module is used to perform fast Fourier transform on the output of the multi-head cross attention module, obtain the periodic component from the output of the multi-head cross attention module through fast Fourier transform, select the basis with the most significant amplitude in the frequency domain as the main frequency component of the current sequence, and obtain the main periodic features of the output of the multi-head cross attention module in the time domain after inverse Fourier transform and output mapping layer; extract the mean from the output of the decoder's feedforward neural network, merge the mean with the main periodic features, and obtain the periodic term output; subtract the periodic term from the output of the Transformer feedforward neural network to obtain the residual term; the residual term and the periodic term are superimposed to obtain the load curve of the time period to be predicted in the current stage.

6. The method for power system load forecasting based on conditional guided diffusion process according to claim 1, characterized in that: The hybrid feature condition generation network specifically includes a feature fusion module, a linear mapping module and a feature channel convolution module; wherein the feature fusion module is used to perform feature fusion on the output result of the previous stage with the load curve of a similar period; the linear mapping module is used to downsample the historical curve features of the corresponding stage, so that the downsampled result is aligned with the load curve of a similar period; the feature channel convolution module is used to perform channel convolution on the splicing result of the output of the feature fusion module and the output of the linear mapping module, and the channel convolution result is aligned with the data dimension of the deep denoising network.

7. A power system load forecasting device based on conditional guided diffusion process, used to execute the method according to claim 1, characterized in that: It includes similar period load curve discovery module, multi-stage denoising module and deterministic prediction module. The similar period load curve mining module is used to obtain historical temperature, humidity, year, month, day characteristics, and weather forecast information for the period to be predicted, and to mine similar period load curves; The multi-stage denoising module is used to use the load curve of the similar period as a guiding condition, input the guiding condition into a denoising model guided by a hybrid feature network generation condition, perform multi-stage denoising, obtain the load curve of the period to be predicted, and repeatedly perform the multi-stage denoising process multiple times to obtain a set of load curves of the period to be predicted; The deterministic prediction module is used to obtain a load probability prediction result for the time period to be predicted by using kernel density estimation to perform statistics on the load curve set for the time period to be predicted; The median of the probability prediction period at each moment is selected as the deterministic prediction result of the period to be predicted.

8. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer instructions stored thereon, characterized in that: The computer instructions are used to enable a computer to execute the steps of the method according to any one of claims 1 to 6.

Images