Multi-source data fusion method, system and equipment of active power distribution network and medium

Abstract

A multi-source data fusion method, system, device and medium for an active power distribution network, the method comprising: using a target training denoising diffusion implicit model of a denoising diffusion probability model, generating a conditional denoising diffusion implicit model by modifying a sampling method of a back diffusion process of the denoising diffusion probability model; and then realizing self-supervised training of the conditional denoising diffusion implicit model introducing conditional information by using a minimization loss function, and finally deploying the trained conditional denoising diffusion implicit model into an actual active power distribution network to perform real-time processing and fusion of multi-source data of the active power distribution network. According to the method, a conditional information denoising mechanism is introduced, a conditional denoising diffusion implicit model is deployed in an actual active power distribution network, and high-quality active power distribution network measurement sample data is generated in fewer steps, so that the generation efficiency and quality of the data are improved, the intelligent level of a data fusion system is improved, and the data fusion efficiency is improved. And efficient and accurate fusion of the multi-source heterogeneous data of the power distribution network is realized.

Description

Technical Field

[0001] The present invention belongs to the field of active distribution network and deep learning technology, and specifically relates to a multi-source data fusion method, system, equipment and medium for an active distribution network. Background Art

[0002] In the context of new power systems, the transformation and upgrading of traditional distribution networks to new active distribution networks has become an inevitable trend. High-performance intelligent measurement equipment and traditional terminal equipment generally coexist, and distributed power sources are generally connected. This leads to the generation of a large amount of data from different sources during the operation of the new active distribution network, which in turn brings challenges to the accurate perception of the operating status of the distribution network. These data include synchronous phasor unit data PMU, data acquisition and monitoring system data SCADA, and advanced measurement system data AMI, etc. They come from different systems and are of great significance for the state estimation, stable operation, optimized scheduling and fault prediction of the new active distribution network power system. However, due to the inconsistency of timestamps and frequency standards of various systems, and the easy generation of bad data with large measurement deviations, traditional physical methods can no longer meet the actual needs of distribution network data fusion.

[0003] In recent years, deep generative models have made significant progress in the field of data fusion. Among them, the denoising diffusion probability model DDPM and the denoising diffusion implicit model DDIM have attracted widespread attention due to their ability to generate high-quality data samples. Compared with the denoising diffusion probability model, the denoising diffusion implicit model significantly improves the sample generation efficiency by introducing non-Markov processes while maintaining high sample quality.

[0004] Compared with traditional physical fusion methods, deep generative models driven by neural network data can process multi-source measurement data from different sources, formats and precisions. Neural networks have powerful nonlinear mapping capabilities and can organically and efficiently integrate data to extract valuable information without complex pre-processing. Data that has been preliminarily processed can be used as input for the model, which greatly reduces the difficulty and workload of data analysis and modeling and improves data processing efficiency. Through deep learning analysis, a more accurate prediction of the future state of the distribution network can be made. This prediction capability is of great significance for the dispatching, optimization and fault prevention of power grids. Summary of the invention

[0005] The object of the present invention is to provide a multi-source data fusion method, system, device and medium for an active power distribution network in view of the above-mentioned problems existing in the prior art.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention provides a multi-source data fusion method for an active power distribution network, comprising:

[0008] S1. Using the target of the denoising diffusion probability model to train the denoising diffusion implicit model, and by modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model, a conditional denoising diffusion implicit model is generated;

[0009] S2. realizing self-supervised training of a conditional denoising diffusion implicit model with conditional information by minimizing a loss function, wherein the conditional information is a part of the direct observation value of the original measurement data in the time series of the active distribution network, including real-time measurement data of synchronized phasor unit data, data acquisition and monitoring system data, and advanced measurement system data;

[0010] S3. Deploy the trained conditional denoising diffusion implicit model to the actual active distribution network to perform real-time fusion of multi-source data of the active distribution network.

[0011] The S1 includes:

[0012] S11. Construct the following denoising diffusion probability model for the reverse diffusion process:

[0013] p(x t-1 |x t , x0)=p(x t-1 |x t , t)~N(μ(x t , t), σ 2 (x t , t));

[0014]

[0015] In the above formula, x t is the data after adding noise t times, x0 is the original data of the denoised diffusion probability model, p(x t-1 |x t , x0) is the current data x t Under the condition of the initial data x0, the posterior probability of the data at time t-1, p(x t-1 |x t , t) is the current data x t and the current time t, the posterior probability of the data at time t-1, μ(x t , t) is the mean of all noise data, σ 2 (x t , t) is the variance of all noise data, α t is the noise intensity, for a t *a t-1 *...a2*a1, is the result of iterative calculation, ε θ (x t , t) is the denoising function;

[0016] S12, modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model, and generating the conditional denoising diffusion implicit model comprises:

[0017] S121, denoising diffusion probability model p(x t-1 |, x t , x0) satisfies the following normal distribution:

[0018] p(x t-1 |, x t ,x0)~N(kx0+mx t , σ 2 );

[0019] x t-1 =kx0+mx t +σε,ε~N(0,1);

[0020] In the above formula, k is the initial data x0 in generating x t-1 The contribution degree when m is the current noise data x t In generating x t-1 , σ is the noise scale parameter related to time step t, and ε is the added Gaussian noise;

[0021] S122, combined with the denoising diffusion probability model forward diffusion process x t Iterate to get:

[0022]

[0023] In the above formula, ε′ is a newly introduced random variable associated with the original noise;

[0024] S123, x in the denoising diffusion probability model t-1 Iterative formula Corresponding to the iterative formula described in S122, we get:

[0025]

[0026] Then p(x t-1 |x t , x0) obeys the normal distribution:

[0027]

[0028] S124, use ε t Replace x0 in the normal distribution, get:

[0029]

[0030] In the above formula, ε t is the noise predicted by the model;

[0031] S125, from p(x t-1 |x t , x0) obeys the normal distribution. It can be seen that the reverse process of generating the conditional denoising diffusion implicit model does not use the Markov process and does not use σ 2 , so let σ 2 = 0, and let x p Replace x t-1 Generate the following conditional denoising diffusion implicit model:

[0032]

[0033] In the above formula, x p is the data after adding noise p times, is the calculation result of the pth iteration of the noise intensity; x t and x p Multiple noise addition times can be spaced apart to increase the sampling speed of the model.

[0034] The S2 includes:

[0035] S21, the mean and variance of the conditional denoising diffusion probability model are represented as μ DDPM (x t ,t,ε θ (x t , t)) and σ 2DDPM (x t , t), define the conditional denoising function as ε θ :(X ta ×R|X co )→X ta , where X ta is the missing or to-be-interpolated part of the time series data, R is a real number set or real number space, X co is the known observation portion of the time series data;

[0036] S22, conditional observation value As input, a conditional denoising function is used parameterized as follows:

[0037]

[0038] In the above formula, is the conditional mean, is the learned conditional denoising function, is the conditional variance, is the missing value that needs to be generated or interpolated at a certain time step t, is the conditional observation value known in sample x0 and used to estimate the missing value;

[0039] S23. Set all observations in x0 as conditional observations All missing values ​​are set as imputation targets While sampling the following noise targets, the conditional denoising function ε is trained by minimizing the loss function θ :

[0040]

[0041] In the above formula, are the missing values ​​that need to be generated or imputed in the initial time period;

[0042] The minimization loss function is:

[0043]

[0044] In the above formula, L is the loss function target to be minimized, θ is the variable to be learned and updated during the optimization process, and ε θ is the noise vector sampled from the standard normal distribution, q(x0) is the actual data distribution, is the expected operation, and the dimension of ε corresponds to the generated target Dimension.

[0045] The S3 includes:

[0046] S31, obtaining the measurement data at each moment within a certain measurement time range, and using the following discrimination formula to perform preliminary identification on the measurement data at each moment. If the measurement data at a certain moment within the time range does not meet the discrimination requirements, the measurement data is considered to be abnormal data:

[0047] |a t -μ|<δσ;

[0048] In the above formula, a t is the measured data at time t, μ is the mean of the measured data at n times, δ is the set confidence threshold, and σ is the standard deviation of the measured data at n times;

[0049] S32. Use the following autocorrelation function to describe the correlation between abnormal data and measured data at other times, and eliminate bad measurement data:

[0050]

[0051] In the above formula, f k is the correlation coefficient, a t +k is the measurement data at time t+k;

[0052] S33, inputting the conditional information and the bad data set at a specific position represented by the mask matrix into the trained conditional denoising diffusion implicit model, obtaining the measurement data after replacing the bad data through the reverse denoising process, and obtaining a complete measurement data set;

[0053] S34. Based on the obtained complete measurement data set, different frequency fusion strategies are adopted for different types of data, including:

[0054] For the synchronous phasor unit data measurement, the following formula is used to calculate the fused measurement value of the synchronous phasor unit data at time t:

[0055]

[0056] In the above formula, Z t，P is the fused measurement value of the synchronized phasor unit data at time t, N is the number of synchronized phasor unit data measured before time t, n is the number of current measurements, and Z (t+nT)，P is the fused measurement value of the synchronized phasor unit data at time t+nT, where T is the sampling period of the synchronized phasor unit data;

[0057] For the data measurement of the data collection and monitoring system, the following formula is used to calculate the fusion measurement value of the data collection and monitoring system data after the fusion prediction value at time t:

[0058]

[0059] In the above formula, Z t，S is the fusion measurement value of the data acquisition and monitoring system data at time t, c is the sampling time before time t, a is the sampling time after time t, Z a，Y , Z c，Y They are the fused measurement values ​​of the data acquisition and monitoring system data predicted at time a and time c respectively;

[0060] For advanced measurement system data measurement, the following formula is used to calculate the fusion measurement value of the advanced measurement system data at time t:

[0061] Z A，t =Z A，i +x(Z A，i+1 -Z A，i )+y(Z A，i+1 -Z A，i ) 2 +z(Z A，i+1 -Z A，i ) 3 ;

[0062] In the above formula, Z A，t is the fused measurement value of the advanced measurement system data at time t, and t∈(i, i+1), ZA，i is the fusion measurement value of the advanced measurement system data at time i, Z A，i+1 is the fused measurement value of the advanced measurement system data predicted at time i+1, i is the initial time of cubic spline interpolation, and x, y, and z are the weights of each interpolation term.

[0063] In a second aspect, the present invention proposes a multi-source data fusion system for an active power distribution network, comprising a conditional denoising diffusion implicit model generation module, a conditional denoising diffusion implicit model training module, and a multi-source data fusion module;

[0064] The conditional denoising diffusion implicit model generation module is used to train the denoising diffusion implicit model using the target of the denoising diffusion probability model, and to generate the conditional denoising diffusion implicit model by modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model;

[0065] The conditional denoising diffusion implicit model training module is used to realize the self-supervised training of the conditional denoising diffusion implicit model introducing conditional information by minimizing the loss function, wherein the conditional information is a part of the direct observation value of the original measurement data in the active distribution network time series, including the real-time measurement data of the synchronous phasor unit data, the data acquisition and monitoring system data, and the advanced measurement system data;

[0066] The multi-source data fusion module is used to deploy the trained conditional denoising diffusion implicit model to the actual active distribution network to perform real-time fusion of multi-source data of the active distribution network.

[0067] The denoising diffusion implicit model generation module includes a denoising diffusion probability model construction unit and a sampling method modification unit;

[0068] The denoising diffusion probability model building unit is used to build the denoising diffusion probability model of the following reverse diffusion process:

[0069]

[0070] In the above formula, x t is the data after adding noise t times, x0 is the original data of the denoised diffusion probability model, p(x t-1 |x t , x0) is the current data x t Under the condition of the initial data x0, the posterior probability of the data at time t-1, p(x t-1 |x t , t) is the current data x t and the current time t, the posterior probability of the data at time t-1, μ(x t , t) is the mean of all noise data, σ 2 (x t , t) is the variance of all noise data, αt is the noise intensity, for a t *a t-1 *...a2*a1, is the result of iterative calculation, ε θ (x t , t) is the denoising function;

[0071] The sampling method modification unit is used to modify the sampling method of the reverse diffusion process of the denoising diffusion probability model to generate a conditional denoising diffusion implicit model, including a normal distribution model construction subunit, an iteration subunit, a corresponding subunit, a replacement subunit, and a model generation subunit;

[0072] The normal distribution model building subunit is used to build a denoising diffusion probability model p(x t-1 |x t , x0):

[0073] p(x t-1 |x t ,x0)~N(kx0+mx t , σ 2 );

[0074] x t-1 =kx0+mx t +σε,ε~N(0,1);

[0075] In the above formula, k is the initial data x0 in generating x t-1 The contribution degree when m is the current noise data x t In generating x t-1 , σ is the noise scale parameter related to time step t, and ε is the added Gaussian noise;

[0076] The iterative subunit is used to combine the denoising diffusion probability model forward diffusion process x t Iterate to get:

[0077]

[0078] In the above formula, ε′ is a newly introduced random variable associated with the original noise;

[0079] The corresponding subunit is used to transform x in the denoising diffusion probability model t-1 Iterative formula Corresponding to the iterative formula described in the iterative subunit, we get:

[0080]

[0081] Then p(x t-1 |xt , x0) obeys the normal distribution:

[0082]

[0083] The replacement subunit is used to replace ε t Replace x0 in the normal distribution, get:

[0084]

[0085] In the above formula, ε t is the noise predicted by the model;

[0086] The model generation subunit is used to generate the model from p(x t-1 |x t , x0) obeys the normal distribution. It can be seen that the reverse process of generating the conditional denoising diffusion implicit model does not use the Markov process and does not use σ 2 , so let σ 2 = 0, and let x p Replace x t-1 Generate the following conditional denoising diffusion implicit model:

[0087]

[0088] In the above formula, x p is the data after adding noise p times, is the calculation result of the pth iteration of the noise intensity; x t and x p Multiple noise addition times can be spaced apart to increase the sampling speed of the model.

[0089] The conditional denoising diffusion implicit model training module includes a parameter definition unit, a function parameterization unit, and a loss function training unit;

[0090] The parameter definition unit is used to represent the mean and variance of the conditional denoising diffusion probability model as μ DDPM (x t ,t,ε θ (x t , t)) and σ 2DDPM (x t , t), define the conditional denoising function as ε θ :(X ta ×R|X co )→X ta , where X ta is the missing or to-be-interpolated part of the time series data, R is a real number set or real number space, X co is the known observation portion of the time series data;

[0091] The function parameterization unit is used to transform the conditional observation value As input, a conditional denoising function is used parameterized as follows:

[0092]

[0093] In the above formula, is the conditional mean, is the learned conditional denoising function, is the conditional variance, is the missing value that needs to be generated or interpolated at a certain time step t, is the conditional observation value known in sample x0 and used to estimate the missing value;

[0094] The loss function training unit is used to set all observations in x0 as conditional observations. All missing values ​​are set as imputation targets While sampling the following noise targets, the conditional denoising function ε is trained by minimizing the loss function θ :

[0095]

[0096] In the above formula, are the missing values ​​that need to be generated or imputed in the initial time period;

[0097] The minimization loss function is:

[0098]

[0099] In the above formula, L is the loss function target to be minimized, θ is the variable to be learned and updated during the optimization process, and ε θ is the noise vector sampled from the standard normal distribution, q(x0) is the actual data distribution, is the expected operation, and the dimension of ε corresponds to the generated target Dimension.

[0100] The multi-source data fusion module includes an abnormal data identification unit, a bad data elimination unit, a bad data replacement unit, and a data frequency fusion unit;

[0101] The abnormal data identification unit is used to obtain the measurement data at each moment within a certain measurement time range, and use the following identification formula to perform preliminary identification on the measurement data at each moment. If the measurement data at a certain moment within the time range does not meet the identification requirements, the measurement data is considered to be abnormal data:

[0102] |a t -μ|<δσ;

[0103] In the above formula, a t is the measured data at time t, μ is the mean of the measured data at n times, δ is the set confidence threshold, and σ is the standard deviation of the measured data at n times;

[0104] The bad data elimination unit is used to use the following autocorrelation function to describe the correlation between abnormal data and measured data at other times, and eliminate bad measurement data:

[0105]

[0106] In the above formula, f k is the correlation coefficient, a t+k is the measurement data at time t+k;

[0107] The bad data replacement unit is used to input the conditional information and the bad data set at a specific position represented by the mask matrix into the trained conditional denoising diffusion implicit model, and obtain the measurement data after the bad data is replaced through the reverse denoising process to obtain a complete measurement data set;

[0108] The data frequency fusion unit is used to adopt different frequency fusion strategies for different types of data based on the obtained complete measurement data set, including:

[0109] For the synchronous phasor unit data measurement, the following formula is used to calculate the fused measurement value of the synchronous phasor unit data at time t:

[0110]

[0111] In the above formula, Z t，P is the fused measurement value of the synchronized phasor unit data at time t, N is the number of synchronized phasor unit data measured before time t, n is the number of current measurements, and Z (t+nT)，P is the fused measurement value of the synchronized phasor unit data at time t+nT, where T is the sampling period of the synchronized phasor unit data;

[0112] For the data measurement of the data collection and monitoring system, the following formula is used to calculate the fusion measurement value of the data collection and monitoring system data after the fusion prediction value at time t:

[0113]

[0114] In the above formula, Z t，S is the fusion measurement value of the data acquisition and monitoring system data at time t, c is the sampling time before time t, a is the sampling time after time t, Z a，Y , Z c，Y They are the fused measurement values ​​of the data acquisition and monitoring system data predicted at time a and time c respectively;

[0115] For advanced measurement system data measurement, the following formula is used to calculate the fusion measurement value of the advanced measurement system data at time t:

[0116] Z A，t =Z A，i +x(Z A，i+1 -Z A，i )+y(Z A，i+1 -Z A，i ) 2 +z(Z A，i+1 -Z A，i ) 3 ;

[0117] In the above formula, Z A，t is the fused measurement value of the advanced measurement system data at time t, and t∈(i, i+1), Z A，i is the fusion measurement value of the advanced measurement system data at time i, Z A，i+1 is the fused measurement value of the advanced measurement system data predicted at time i+1, i is the initial time of cubic spline interpolation, and x, y, and z are the weights of each interpolation term.

[0118] In a third aspect, the present invention provides a multi-source data fusion device for an active power distribution network, comprising a processor and a memory;

[0119] The memory is used to store computer program code and transmit the computer program code to the processor;

[0120] The processor is used to execute the aforementioned multi-source data fusion method for an active power distribution network according to the instructions in the computer program code.

[0121] In a fourth aspect, the present invention provides a computer storage medium having a computer program stored thereon;

[0122] When the computer program is executed by a processor, the steps of the aforementioned multi-source data fusion method for an active power distribution network are implemented.

[0123] Compared with the prior art, the present invention has the following beneficial effects:

[0124] The present invention proposes a multi-source data fusion method, system, device and medium for an active distribution network. The method first uses a denoising diffusion implicit model to train a target denoising diffusion probability model, generates a conditional denoising diffusion implicit model by modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model, and then uses a minimization loss function to realize self-supervised training of the conditional denoising diffusion implicit model that introduces conditional information. Finally, the trained conditional denoising diffusion implicit model is deployed in an actual active distribution network to perform real-time processing and fusion of multi-source data of the active distribution network. On the one hand, this method uses the target training of the denoising diffusion implicit model of the denoising diffusion probability model, and generates high-quality active distribution network sample data in fewer steps by modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model. It not only improves the data generation efficiency and quality, enhances the data fusion effect, and expands the data application scenarios, but also helps to improve the intelligence level of the data fusion system and provide strong support for the intelligent management of the active distribution network; on the other hand, this method introduces the conditional information denoising mechanism and deploys the conditional denoising diffusion implicit model in the actual active distribution network, so that the complete measurement data set is no longer completely unconstrained, but needs to be carried out under the premise of satisfying the conditional information constraints, which helps to reduce the diversity and uncertainty of the measurement data, improve the accuracy and reliability of the generated data, and utilize the efficient generation capability of the conditional denoising diffusion implicit model to achieve efficient and accurate fusion of multi-source heterogeneous data of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0125] Figure 1 This is a structural diagram of the active distribution network of Example 1 described in the present invention.

[0126] Figure 2 The figure is an overall flow chart of the method of the present invention.

[0127] Figure 3 The structure diagram of the system of the present invention is shown in FIG.

[0128] Figure 4 This is a structural diagram of the device described in Example 3. DETAILED DESCRIPTION

[0129] The present invention is further described in detail below in conjunction with specific implementations and drawings.

[0130] The present invention proposes a multi-source data fusion method, system, device and medium for an active distribution network. The original measurement data within a certain measurement time range is collected from the data source of the active distribution network for preprocessing, including synchronous phasor unit data, data acquisition and monitoring system data, and advanced measurement system data. The denoising diffusion implicit model is trained using the target of the denoising diffusion probability model. The conditional denoising diffusion implicit model is generated by modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model. High-quality active distribution network sample data is generated in fewer steps. The self-supervised training of the denoising diffusion implicit model introducing conditional information is realized by minimizing the loss function. Finally, the trained conditional denoising diffusion implicit model is deployed in the actual active distribution network to perform real-time processing and fusion of multi-source data of the active distribution network.

[0131] Embodiment 1:

[0132] This embodiment is as follows Figure 1 The active distribution network structure shown in the figure is the research object. The active distribution network structure contains 30 nodes, which are distributed in different geographical locations and interconnected by transmission lines. The active distribution network also contains traditional power transmission and distribution facilities, equipped with synchronous phasor units, data acquisition and monitoring systems, and advanced measurement systems to achieve comprehensive monitoring of the status of the active distribution network. It is connected to a variety of distributed power sources, such as solar photovoltaic panels, wind turbines, etc. These distributed power sources, as active parts, improve the flexibility of the distribution network and the utilization rate of renewable energy.

[0133] like Figure 2 As shown, a multi-source data fusion method for an active distribution network is performed in the following steps:

[0134] Firstly, the original measurement data within a certain measurement time range is collected from the data source of the active distribution network, including synchronous phasor unit data, data acquisition and monitoring system data, and advanced measurement system data. The original measurement data is then cleaned, outliers and missing values ​​are removed, and normalization is performed.

[0135] S1. Use the target training of the denoising diffusion probability model to train the denoising diffusion implicit model. By setting key parameters such as model noise intensity and number of iterations, modify the sampling method of the reverse diffusion process of the denoising diffusion probability model, generate a conditional denoising diffusion implicit model, and improve the efficiency and quality of data generation.

[0136] S11. The reverse diffusion process is the core part of the denoising diffusion probability model. The model adds noise t times to the data x t Step by step denoising, we get x t-1 , x t-2 Wait until the data is restored to the original data x0 of the denoised diffusion probability model;

[0137] In the denoising process, the last moment x t-1 The value of is unknown and needs to be determined based on x t To make predictions, we can only use the probability form and introduce the auxiliary quantity x0, and use the following Bayesian formula to calculate the posterior probability:

[0138]

[0139] In the above formula, p(x t-1 |x t , x0) is the current data x t Under the condition of the initial data x0, the posterior probability of the data at time t-1, p(x t |x t-1 , x0) is the current data x under the conditions of the data at time t-1 and the initial data x0 t The conditional probability, p(x t-1 |x0) is the posterior probability of the data at time t-1 when only the initial data x0 is known, p(x t |x0) is the current data x under the condition of initial data x0 t The conditional probability of

[0140] Using the denoised diffusion probability model to predict the forward diffusion process In turn, they represent p(x t |x t-1 , x0), p(x t-1 |x0), p(x t |x0):

[0141]

[0142] In the above formula, α t is the noise intensity, x t-1 is the data after adding noise t-1 times, ε is the added Gaussian noise, x0 is the initial data, for a t *a t-1 *…a2*a1 is the result of iterative calculation;

[0143] Combining all the above formulas, the following denoising diffusion probability model of the reverse diffusion process is constructed:

[0144] p(x t-1 |x t , x0)=p(x t-1 |x t , t)~N(μ(x t , t), σ 2 (x t , t));

[0145]

[0146] In the above formula, x t is the data after adding noise t times, p(x t-1 |x t , t) is the current data x t and the current time t, the posterior probability of the data at time t-1, μ(x t , t) is the mean of all noise data, σ 2 (x t , t) is the variance of all noise data, ε θ (x t , t) is the denoising function;

[0147] S12, modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model to generate a conditional denoising diffusion implicit model;

[0148] The denoising diffusion implicit model breaks the Markov chain process through mathematical reasoning. Without changing the forward noise addition, the sampling speed can be greatly increased by simply modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model, including:

[0149] S121, assuming the denoising diffusion probability model p(x t-1 |, x t , x0) satisfies the following normal distribution:

[0150] p(x t-1 |x t ,x0)~N(kx0+mx t , σ 2 );

[0151] x t-1 =kx0+mx t +σε,ε~N(0,1);

[0152] In the above formula, k is the initial data x0 in generating x t-1 The contribution degree when m is the current noise data x t In generating x t-1 σ is the contribution degree of the sampling process, and σ is the noise scale parameter related to the time step t, which controls the randomness of the sampling process;

[0153] S122, combined with the denoising diffusion probability model forward diffusion process x t Iterate to get:

[0154]

[0155] In the above formula, ε does not come directly from the noise of the initial data, but is a newly introduced random variable associated with the original noise;

[0156] S123, x in the denoising diffusion probability model t-1 Iterative formula With x t Iterate to get x t-1 The corresponding formula is:

[0157]

[0158] Then p(x t-1| x t , x0) obeys the normal distribution:

[0159]

[0160] S124, use ε t Replace x0 in the normal distribution, get:

[0161]

[0162] In the above formula, ε t The noise predicted by the model is part of the model output and is used to gradually recover the initial data x0 during the back diffusion process;

[0163] S125, from p(x t-1 |x t , x0) obeys the normal distribution. It can be seen that the reverse process of generating the conditional denoising diffusion implicit model does not use the Markov process and does not use σ 2 And x t-1 It does not need to be strictly x t Calculate, so let σ 2 = 0, and by x p Replace x t-1 Generate the following conditional denoising diffusion implicit model:

[0164]

[0165] In the above formula, x p is the data after adding noise p times, is the calculation result of the pth iteration of the noise intensity; x t and x p Multiple noise addition times can be spaced to increase the sampling speed of the model;

[0166] This is a denoising diffusion implicit model DDIM trained using the denoising diffusion probability model DDPM objective. In the process of generating the conditional denoising diffusion implicit model, the model predicts the noise εt The previous coefficient becomes zero, x t The fixed process to x0 is generated by adding different amounts of Gaussian noise data samples at different time steps, i.e., a series of noisy data generated in the forward diffusion process.

[0167] S2, dividing the preprocessed original measurement data into a training set and a validation set, using the training set data to implement self-supervised training of the denoising diffusion implicit model that introduces conditional information by minimizing the loss function;

[0168] On the basis of the standard denoising diffusion implicit model, a conditional information input channel is added. The conditional information refers to some direct observation values ​​of the original measurement data in the time series of the active distribution network, including synchronous phasor unit data, data acquisition and monitoring system data, and real-time measurement data of advanced measurement system data. It guides the generation of the conditional denoising diffusion implicit model in the diffusion and denoising process of each time step, and learns to map from the current state, that is, noisy data and observations, to the next state, that is, the state closer to the real data.

[0169] In each step t, the reverse process uses part of the direct observations from the time series as conditional information input, starts interpolation from random noise, and predicts the state of the next step based on the output and observations of the previous step, that is, the current state and conditional information, thereby constructing a conditional denoising diffusion implicit model combined with conditional information input, allowing the model to use the information in the observations for denoising, and gradually convert the noise into a reasonable time series through the inverse process of the conditional denoising diffusion implicit model;

[0170] S21, the mean and variance of the denoised diffusion probability model are represented as μ DDPM (x t ,t,ε θ (x t , t)) and σ 2DDPM (x t , t), define the conditional denoising function as ε θ :(X ta ×R|X co )→X ta , where ε θ is a learned conditional denoising function, with the missing or to-be-interpolated part X in the time series data ta , time step t and the known observation part of the time series data, i.e., the observation condition X co is the input, predicts the denoised target measurement value, adjusts the mean and standard deviation obtained from the unconditional model, R is a real number set or real number space, which is the mathematical representation space of model parameters and input data, or represents parameters and variables as real numbers or real number vectors;

[0171] S22, conditional observation value As input, the conditional denoising function is parameterized as follows. The purpose of parameterization is to extend the unconditional diffusion model to the conditional case, so that the model can predict the data based on the conditional observations, that is, the known part of the data;

[0172]

[0173] In the above formula, is the conditional mean, is the learned conditional denoising function, is the conditional variance, is the missing value that needs to be generated or interpolated at a certain time step t, is the conditional observation value known in sample x0 and used to estimate the missing value;

[0174] S23. Set all observations in x0 as conditional observations All missing values ​​are set as imputation targets While sampling the following noise targets, the conditional denoising function ε is trained by minimizing the loss function θ :

[0175]

[0176] In the above formula, are the missing values ​​that need to be generated or imputed in the initial time period;

[0177] The minimization loss function is:

[0178]

[0179] In the above formula, L is the loss function target to be minimized, θ is the variable to be learned and updated during the optimization process, and ε θ is the noise vector sampled from the standard normal distribution, q(x0) is the actual data distribution, This expectation is achieved through random sampling approximation, usually using the average value of a large number of samples to approximate. The dimension of ε corresponds to the generated target Dimensions;

[0180] When generating the conditional denoising diffusion implicit model, a self-supervised training method is designed to separate the observed values ​​and the missing values, that is, to regard part of the observed values ​​as new conditional information and the other part of the observed values ​​as missing values ​​to be generated. Sampling is performed and ε is trained by minimizing the loss function θ , which solves the problem that effective training cannot be performed in practical applications due to unclear missing values.

[0181] S3, deploy the trained conditional denoising diffusion implicit model to the actual active distribution network to perform real-time processing and fusion of multi-source data of the active distribution network;

[0182] S31, obtaining measurement data at each moment within a certain measurement time range, and using a standard score determination method to perform preliminary identification on the measurement data at each moment;

[0183] For measurement a, the measurement data at each moment within a certain measurement time range are a1, a2, ..., a n If the measured data at a certain moment within the time range does not meet the following judgment requirements, the measured data is considered to be abnormal data:

[0184] |a t -μ|<δσ;

[0185] In the above formula, a t is the measurement data at time t, μ is the mean of the measurement data at n times, δ is the set confidence threshold, usually 3-5 times, and σ is the standard deviation of the measurement data at n times;

[0186] S32. Considering the intermittency of urban and rural active distribution networks connected to distributed power sources, the abnormal data may be fluctuating data caused by intermittency. The following autocorrelation function is used to identify fluctuating data and bad data:

[0187]

[0188] In the above formula, f k is the correlation coefficient, 1≤k≤n-1, f k The closer it is to 1, the greater the correlation between the abnormal data and the measured data at other times. The threshold of the correlation coefficient is set to 0.6. k When ≥0.6, the abnormal data is fluctuating data, otherwise it is bad data that needs to be eliminated;

[0189] S33, inputting the conditional information and the bad data set at a specific position represented by the mask matrix into the trained conditional denoising diffusion implicit model, obtaining the measurement data after replacing the bad data through the reverse denoising process, and obtaining a complete measurement data set;

[0190] S34. In a certain period of time, different types of data have very different upload frequencies. Based on the complete measurement data set obtained, different frequency fusion strategies are adopted for different types of data, including:

[0191] For the measurement of synchronized phasor unit data at time t, considering that the sampling time is dense and the sampling accuracy is required to be higher, a periodic recursive mean method is used to achieve the frequency fusion of synchronized phasor unit data;

[0192] The fused measurement value of the synchronized phasor unit data at time t is calculated using the following formula:

[0193]

[0194] In the above formula, Z t，P is the fused measurement value of the synchronized phasor unit data at time t, N is the number of synchronized phasor unit data measured before time t, n is the number of current measurements, and Z (t+nT)，P is the fused measurement value of the synchronized phasor unit data at time t+nT, where T is the sampling period of the synchronized phasor unit data;

[0195] For the measurement of the data acquisition and monitoring system data at time t, the measurement value of the data acquisition and monitoring system data at the last moment closest to this moment and the prediction information at this moment are used to achieve the fusion of the measurement value and the prediction value;

[0196] The following formula is used to calculate the fusion measurement value of the data acquisition and monitoring system data after the fusion prediction value at time t:

[0197]

[0198] In the above formula, Z t，S is the fusion measurement value of the data acquisition and monitoring system data at time t, c is the sampling time before time t, a is the sampling time after time t, Z a，Y , Z c，Y They are the fused measurement values ​​of the data acquisition and monitoring system data predicted at time a and time c respectively;

[0199] For the measurement of the advanced measurement system data at time t, a more accurate cubic spline interpolation method is used to fuse the measurement data of the advanced measurement system data;

[0200] The fused measurement value of the advanced measurement system data at time t is calculated using the following formula:

[0201] Z A，t =Z A，i +x(Z A，i+1 -Z A，i )+y(Z A，i+1 -Z A，i ) 2 +z(Z A，i+1 -Z A，i ) 3 ;

[0202] In the above formula, Z A，t is the fused measurement value of the advanced measurement system data at time t, and t∈(i, i+1), Z A，i is the fusion measurement value of the advanced measurement system data at time i, ZA，i+1 is the fused measurement value of the advanced measurement system data predicted at time i+1, i is the initial time of cubic spline interpolation, and x, y, and z are the weights of each interpolation term respectively;

[0203] For the generation of predicted measurement values ​​in the above-mentioned data acquisition and monitoring system data and advanced measurement system data, the conditional information and the data set with a mask matrix representing a specific interpolation position are input into the trained conditional denoising diffusion implicit model, and the reverse denoising process is used to obtain the generated results to achieve data fusion.

[0204] Embodiment 2:

[0205] like Figure 3 As shown, a multi-source data fusion system for an active power distribution network includes a conditional denoising diffusion implicit model generation module, a conditional denoising diffusion implicit model training module, and a multi-source data fusion module;

[0206] The conditional denoising diffusion implicit model generation module is used to train the denoising diffusion implicit model using the target of the denoising diffusion probability model, and to generate the conditional denoising diffusion implicit model by modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model;

[0207] The conditional denoising diffusion implicit model training module is used to realize the self-supervised training of the conditional denoising diffusion implicit model introducing conditional information by minimizing the loss function, wherein the conditional information is a part of the direct observation value of the original measurement data in the active distribution network time series, including the real-time measurement data of the synchronous phasor unit data, the data acquisition and monitoring system data, and the advanced measurement system data;

[0208] The multi-source data fusion module is used to deploy the trained conditional denoising diffusion implicit model to the actual active distribution network to perform real-time fusion of multi-source data of the active distribution network.

[0209] The denoising diffusion implicit model generation module includes a denoising diffusion probability model construction unit and a sampling method modification unit;

[0210] The denoising diffusion probability model building unit is used to build the denoising diffusion probability model of the following reverse diffusion process:

[0211] p(x t-1 |, x t , x0)=p(x t-1 |x t , t)~N(μ(x t , t), σ 2 (x t , t));

[0212]

[0213] In the above formula, x t is the data after adding noise t times, x0 is the original data of the denoised diffusion probability model, p(x t-1 |x t , x0) is the current data x t Under the condition of the initial data x0, the posterior probability of the data at time t-1, p(x t-1 |x t , t) is the current data x t and the current time t, the posterior probability of the data at time t-1, μ(x t , t) is the mean of all noise data, σ 2 (x t , t) is the variance of all noise data, α t is the noise intensity, for a t *a t-1 *...a2*a1, is the result of iterative calculation, ε θ (x t , t) is the denoising function;

[0214] The sampling method modification unit is used to modify the sampling method of the reverse diffusion process of the denoising diffusion probability model to generate a conditional denoising diffusion implicit model, including a normal distribution model construction subunit, an iteration subunit, a corresponding subunit, a replacement subunit, and a model generation subunit;

[0215] The normal distribution model building subunit is used to build a denoising diffusion probability model p(x t-1 |x t , x0):

[0216] p(x t-1 |x t ,x0)~N(kx0+mx t , σ 2 );

[0217] x t-1 =kx0+mx t +σε,ε~N(0,1);

[0218] In the above formula, k is the initial data x0 in generating x t-1 The contribution degree when m is the current noise data x t In generating x t-1 , σ is the noise scale parameter related to time step t, and ε is the added Gaussian noise;

[0219] The iterative subunit is used to combine the denoising diffusion probability model forward diffusion process xt Iterate to get:

[0220]

[0221] In the above formula, ε′ is a newly introduced random variable associated with the original noise;

[0222] The corresponding subunit is used to transform x in the denoising diffusion probability model t-1 Iterative formula Corresponding to the iterative formula described in the iterative subunit, we get:

[0223]

[0224] Then p(x t-1 |x t , x0) obeys the normal distribution:

[0225]

[0226] The replacement subunit is used to replace ε t Replace x0 in the normal distribution, get:

[0227]

[0228] In the above formula, ε t is the noise predicted by the model;

[0229] The model generation subunit is used to generate the model from p(x t-1 |x t , x0) obeys the normal distribution. It can be seen that the reverse process of generating the conditional denoising diffusion implicit model does not use the Markov process and does not use σ 2 , so let σ 2 = 0, and let x p Replace x t-1 Generate the following conditional denoising diffusion implicit model:

[0230]

[0231] In the above formula, x p is the data after adding noise p times, is the calculation result of the pth iteration of the noise intensity; x t and x p Multiple noise addition times can be spaced apart to increase the sampling speed of the model.

[0232] The conditional denoising diffusion implicit model training module includes a parameter definition unit, a function parameterization unit, and a loss function training unit;

[0233] The parameter definition unit is used to represent the mean and variance of the conditional denoising diffusion probability model as μ DDPM (x t ,t,ε θ (x t , t)) and σ 2DDPM (x t , t), define the conditional denoising function as ε θ :(X ta ×R|X co )→X ta , where X ta is the missing or to-be-interpolated part of the time series data, R is a real number set or real number space, X co is the known observation portion of the time series data;

[0234] The function parameterization unit is used to transform the conditional observation value As input, a conditional denoising function is used parameterized as follows:

[0235]

[0236] In the above formula, is the conditional mean, is the learned conditional denoising function, is the conditional variance, is the missing value that needs to be generated or interpolated at a certain time step t, is the conditional observation value known in sample x0 and used to estimate the missing value;

[0237] The loss function training unit is used to set all observations in x0 as conditional observations. All missing values ​​are set as imputation targets While sampling the following noise targets, the conditional denoising function ε is trained by minimizing the loss function θ :

[0238]

[0239] In the above formula, are the missing values ​​that need to be generated or imputed in the initial time period;

[0240] The minimization loss function is:

[0241]

[0242] In the above formula, L is the loss function target to be minimized, θ is the variable to be learned and updated during the optimization process, and ε θ is the noise vector sampled from the standard normal distribution, q(x0) is the actual data distribution, is the expected operation, and the dimension of ε corresponds to the generated target Dimension.

[0243] The multi-source data fusion module includes an abnormal data identification unit, a bad data elimination unit, a bad data replacement unit, and a data frequency fusion unit;

[0244] The abnormal data identification unit is used to obtain the measurement data at each moment within a certain measurement time range, and use the following identification formula to perform preliminary identification on the measurement data at each moment. If the measurement data at a certain moment within the time range does not meet the identification requirements, the measurement data is considered to be abnormal data:

[0245] |a t -μ|<δσ;

[0246] In the above formula, a t is the measured data at time t, μ is the mean of the measured data at n times, δ is the set confidence threshold, and σ is the standard deviation of the measured data at n times;

[0247] The bad data elimination unit is used to use the following autocorrelation function to describe the correlation between abnormal data and measured data at other times, and eliminate bad measurement data:

[0248]

[0249] In the above formula, f k is the correlation coefficient, a t+k is the measurement data at time t+k;

[0250] The bad data replacement unit is used to input the conditional information and the bad data set at a specific position represented by the mask matrix into the trained conditional denoising diffusion implicit model, and obtain the measurement data after the bad data is replaced through the reverse denoising process to obtain a complete measurement data set;

[0251] The data frequency fusion unit is used to adopt different frequency fusion strategies for different types of data based on the obtained complete measurement data set, including:

[0252] For the synchronous phasor unit data measurement, the following formula is used to calculate the fused measurement value of the synchronous phasor unit data at time t:

[0253]

[0254] In the above formula, Z t，P is the fused measurement value of the synchronized phasor unit data at time t, N is the number of synchronized phasor unit data measured before time t, n is the number of current measurements, and Z (t+nT)，Pis the fused measurement value of the synchronized phasor unit data at time t+nT, where T is the sampling period of the synchronized phasor unit data;

[0255] For the data measurement of the data collection and monitoring system, the following formula is used to calculate the fusion measurement value of the data collection and monitoring system data after the fusion prediction value at time t:

[0256]

[0257] In the above formula, Z t，S is the fusion measurement value of the data acquisition and monitoring system data at time t, c is the sampling time before time t, a is the sampling time after time t, Z a，Y , Z c，Y They are the fused measurement values ​​of the data acquisition and monitoring system data predicted at time a and time c respectively;

[0258] For advanced measurement system data measurement, the following formula is used to calculate the fusion measurement value of the advanced measurement system data at time t:

[0259]

[0260] In the above formula, Z A，t is the fused measurement value of the advanced measurement system data at time t, and t∈(i, i+1), Z A，i is the fusion measurement value of the advanced measurement system data at time i, Z A，i+1 is the fused measurement value of the advanced measurement system data predicted at time i+1, i is the initial time of cubic spline interpolation, and x, y, and z are the weights of each interpolation term.

[0261] Embodiment 3:

[0262] like Figure 4 As shown, a multi-source data fusion device for an active power distribution network includes a processor and a memory;

[0263] The memory is used to store computer program code and transmit the computer program code to the processor;

[0264] The processor is used to execute the multi-source data fusion method of an active power distribution network described in Example 1 according to the instructions in the computer program code.

[0265] Embodiment 4:

[0266] A computer storage medium having a computer program stored thereon;

[0267] When the computer program is executed by a processor, the steps of a multi-source data fusion method for an active power distribution network described in this solution are implemented.

Claims

1. A multi-source data fusion method for an active distribution network, characterized in that: The method comprises: S1. Using the target of the denoising diffusion probability model to train the denoising diffusion implicit model, and by modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model, a conditional denoising diffusion implicit model is generated; S2. realizing self-supervised training of a conditional denoising diffusion implicit model with conditional information by minimizing a loss function, wherein the conditional information is a part of the direct observation value of the original measurement data in the time series of the active distribution network, including real-time measurement data of synchronized phasor unit data, data acquisition and monitoring system data, and advanced measurement system data; S3. Deploy the trained conditional denoising diffusion implicit model to the actual active distribution network to perform real-time fusion of multi-source data of the active distribution network.

2. The multi-source data fusion method of an active power distribution network according to claim 1, characterized in that: The S1 includes: S11. Construct the following denoising diffusion probability model for the reverse diffusion process: p(x t-1 |x t ,x0)=p(x t-1 |x t ,t)~N(μ(x t ,t),σ 2 (x) t ,t)); In the above formula, x t is the data after adding noise t times, x0 is the original data of the denoised diffusion probability model, p(x t-1 |x t , x0) is the current data x t Under the condition of the initial data x0, the posterior probability of the data at time t-1, P(x t-1 |x t , t) is the current data x t and the current time t, the posterior probability of the data at time t-1, μ(x t , t) is the mean of all noise data, σ 2 (x t , t) is the variance of all noise data, α t is the noise intensity, for a t *a t-1 *...a2*a1, is the result of iterative calculation, ε θ (x t , t) is the denoising function; S12, modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model, and generating the conditional denoising diffusion implicit model comprises: s121, denoising diffusion probability model p(x t-1 |x t , x0) satisfies the following normal distribution: P(x t-1 |x t ,x0)~N(kx0+mx t ,s 2 ); x t-1 =kx0+mx t +in, e~N(0,1); In the above formula, k is the initial data x0 in generating x t-1 The contribution degree when m is the current noise data x t In generating x t-1 , σ is the noise scale parameter related to time step t, and ε is the added Gaussian noise; S122, combined with the denoising diffusion probability model forward diffusion process x t Iterate to get: In the above formula, ε′ is a newly introduced random variable associated with the original noise; S123, x in the denoising diffusion probability model t-1 Iteration formula Corresponding to the iterative formula described in S122, we get: Then p(x t-1 |x t , x0) obeys the normal distribution: S124, use ε t Replace x0 in the normal distribution, get: In the above formula, ε t is the noise predicted by the model; S125, from p(x t-1 |x t , x0) obeys the normal distribution. It can be seen that the reverse process of generating the conditional denoising diffusion implicit model does not use the Markov process and does not use σ 2 , so let σ 2 = 0, and let x p Replace x t-1 Generate the following conditional denoising diffusion implicit model: In the above formula, x p is the data after adding noise p times, is the calculation result of the pth iteration of the noise intensity; x t and x p Multiple noise addition times can be spaced apart to increase the sampling speed of the model.

3. The multi-source data fusion method of an active power distribution network according to claim 1, characterized in that: The S2 includes: s21, the mean and variance of the conditional denoising diffusion probability model are represented as μ DDPM (x t ,t,ε θ (x t , t)) and σ 2DDPM (x t , t), define the conditional denoising function as ε θ :(X ta ×R|X co )→X ta , where X ta is the missing or to-be-interpolated part of the time series data, R is a real number set or real number space, X co is the known observation portion of the time series data; S22, conditional observation value As input, a conditional denoising function is used parameterized as follows: In the above formula, is the conditional mean, is the learned conditional denoising function, is the conditional variance, is the missing value that needs to be generated or interpolated at a certain time step t, is the conditional observation value known in sample x0 and used to estimate the missing value; S23. Set all observations in x0 as conditional observations All missing values ​​are set as imputation targets While sampling the following noise targets, the conditional denoising function ε is trained by minimizing the loss function θ : In the above formula, are the missing values ​​that need to be generated or imputed in the initial time period; The minimization loss function is: In the above formula, L is the loss function target to be minimized, θ is the variable to be learned and updated during the optimization process, and ε θ is the noise vector sampled from the standard normal distribution, q(x0) is the actual data distribution, is the expected operation, and the dimension of ε corresponds to the generated target Dimension.

4. The multi-source data fusion method of an active power distribution network according to claim 1, characterized in that: The S3 includes: S31, obtaining the measurement data at each moment within a certain measurement time range, and using the following discrimination formula to perform preliminary identification on the measurement data at each moment. If the measurement data at a certain moment within the time range does not meet the discrimination requirements, the measurement data is considered to be abnormal data: |a t -μ|<δσ; In the above formula, a t is the measured data at time t, μ is the mean of the measured data at n times, δ is the set confidence threshold, and σ is the standard deviation of the measured data at n times; S32. Use the following autocorrelation function to describe the correlation between abnormal data and measured data at other times, and eliminate bad measurement data: In the above formula, f k is the correlation coefficient, a t+k is the measurement data at time t+k; S33, inputting the conditional information and the bad data set at a specific position represented by the mask matrix into the trained conditional denoising diffusion implicit model, obtaining the measurement data after replacing the bad data through the reverse denoising process, and obtaining a complete measurement data set; S34. Based on the obtained complete measurement data set, different frequency fusion strategies are adopted for different types of data, including: For the synchronous phasor unit data measurement, the following formula is used to calculate the fused measurement value of the synchronous phasor unit data at time t: In the above formula, Z t，P is the fused measurement value of the synchronized phasor unit data at time t, N is the number of synchronized phasor unit data measured before time t, n is the number of current measurements, and Z (t+nT)，P is the fused measurement value of the synchronized phasor unit data at time t+nT, where T is the sampling period of the synchronized phasor unit data; For the data measurement of the data collection and monitoring system, the following formula is used to calculate the fusion measurement value of the data collection and monitoring system data after the fusion prediction value at time t: In the above formula, Z t，S is the fusion measurement value of the data acquisition and monitoring system data at time t, c is the sampling time before time t, a is the sampling time after time t, and Z a，Y , Z c，Y They are the fused measurement values ​​of the data acquisition and monitoring system data predicted at time a and time c respectively; For advanced measurement system data measurement, the following formula is used to calculate the fusion measurement value of the advanced measurement system data at time t: WITH A，t =Z A，i +x(Z A，i+1 -WITH A，i )+y(Z A，i+1 -WITH A，i ) 2 +z(Z A，i+1 -WITH A，i ) 3 ; In the above formula, Z A，t is the fused measurement value of the advanced measurement system data at time t, and t∈(i, i+1), Z A，i is the fusion measurement value of the advanced measurement system data at time i, Z A，i+1 is the fused measurement value of the advanced measurement system data predicted at time i+1, i is the initial time of cubic spline interpolation, and x, y, and z are the weights of each interpolation term.

5. A multi-source data fusion system for an active distribution network, characterized in that: The system includes a conditional denoising diffusion implicit model generation module, a conditional denoising diffusion implicit model training module, and a multi-source data fusion module; The conditional denoising diffusion implicit model generation module is used to train the denoising diffusion implicit model using the target of the denoising diffusion probability model, and to generate the conditional denoising diffusion implicit model by modifying the sampling method of the reverse diffusion process of the denoising diffusion probability model; The conditional denoising diffusion implicit model training module is used to realize the self-supervised training of the conditional denoising diffusion implicit model introducing conditional information by minimizing the loss function, wherein the conditional information is a part of the direct observation value of the original measurement data in the active distribution network time series, including the real-time measurement data of the synchronous phasor unit data, the data acquisition and monitoring system data, and the advanced measurement system data; The multi-source data fusion module is used to deploy the trained conditional denoising diffusion implicit model to the actual active distribution network to perform real-time fusion of multi-source data of the active distribution network.

6. The multi-source data fusion system of an active power distribution network according to claim 5, characterized in that: The denoising diffusion implicit model generation module includes a denoising diffusion probability model construction unit and a sampling method modification unit; The denoising diffusion probability model building unit is used to build the denoising diffusion probability model of the following reverse diffusion process: p(x t-1 |x t ,x0)=p(x t-1 |x t ,t)~N(μ(x t ,t),σ 2 (x) t ,t)); In the above formula, x t is the data after adding noise t times, x0 is the original data of the denoised diffusion probability model, p(x t-1 |x t , x0) is the current data x t Under the condition of the initial data x0, the posterior probability of the data at time t-1, p(x t-1 |x t , t) is the current data x t and the current time t, the posterior probability of the data at time t-1, μ(x t , t) is the mean of all noise data, σ 2 (x t , t) is the variance of all noise data, α t is the noise intensity, is α t *α t-1 *...a2*a1, is the result of iterative calculation, ε θ (x t , t) is the denoising function; The sampling method modification unit is used to modify the sampling method of the reverse diffusion process of the denoising diffusion probability model to generate a conditional denoising diffusion implicit model, including a normal distribution model construction subunit, an iteration subunit, a corresponding subunit, a replacement subunit, and a model generation subunit; The normal distribution model building subunit is used to build a denoising diffusion probability model p(x t-1 |x t , x0): p(x t-1 |x t ,x0)~N(kx0+mx t ,s 2 ); x t-1 =kx0+mx t +in, e~N(0,1); In the above formula, k is the initial data x0 in generating x t-1 The contribution degree when m is the current noise data x t In generating x t-1 , σ is the noise scale parameter related to time step t, and ε is the added Gaussian noise; The iterative subunit is used to combine the denoising diffusion probability model forward diffusion process x t Iterate to get: In the above formula, ε′ is a newly introduced random variable associated with the original noise; The corresponding subunit is used to transform x in the denoising diffusion probability model t-1 Iterative formula Corresponding to the iterative formula described in the iterative subunit, we get: Then p(x t-1 |x t , x0) obeys the normal distribution: The replacement subunit is used to replace ε t Replace x0 in the normal distribution, get: In the above formula, ε t is the noise predicted by the model; The model generation subunit is used to generate the model from p(x t-1 |x t , x0) obeys the normal distribution. It can be seen that the reverse process of generating the conditional denoising diffusion implicit model does not use the Markov process and does not use σ 2 , so let σ 2 = 0, and let x p Replace x t-1 Generate the following conditional denoising diffusion implicit model: In the above formula, x p is the data after adding noise p times, is the calculation result of the pth iteration of the noise intensity; x t and x p Multiple noise addition times can be spaced apart to increase the sampling speed of the model.

7. The multi-source data fusion system of an active power distribution network according to claim 5, characterized in that: The conditional denoising diffusion implicit model training module includes a parameter definition unit, a function parameterization unit, and a loss function training unit; The parameter definition unit is used to represent the mean and variance of the conditional denoising diffusion probability model as μ DDPM (x t ,t,ε θ (x t , t)) and σ 2DDPM (x t , t), define the conditional denoising function as ε θ :(x ta ×R|X co )→X ta , where X ta is the missing or to-be-interpolated part of the time series data, R is a real number set or real number space, X co is the known observation portion of the time series data; The function parameterization unit is used to transform the conditional observation value As input, a conditional denoising function is used parameterized as follows: In the above formula, is the conditional mean, is the learned conditional denoising function, is the conditional variance, is the missing value that needs to be generated or interpolated at a certain time step t, is the conditional observation value known in sample x0 and used to estimate the missing value; The loss function training unit is used to set all observations in x0 as conditional observations. All missing values ​​are set as imputation targets While sampling the following noise targets, the conditional denoising function ε is trained by minimizing the loss function θ : In the above formula, are the missing values ​​that need to be generated or imputed in the initial time period; The minimization loss function is: In the above formula, L is the loss function target to be minimized, θ is the variable to be learned and updated during the optimization process, and ε θ is the noise vector sampled from the standard normal distribution, q(x0) is the actual data distribution, is the expected operation, and the dimension of ε corresponds to the generated target Dimension.

8. The multi-source data fusion system of an active power distribution network according to claim 5, characterized in that: The multi-source data fusion module includes an abnormal data identification unit, a bad data elimination unit, a bad data replacement unit, and a data frequency fusion unit; The abnormal data identification unit is used to obtain the measurement data at each moment within a certain measurement time range, and use the following identification formula to perform preliminary identification on the measurement data at each moment. If the measurement data at a certain moment within the time range does not meet the identification requirement, the measurement data is considered to be abnormal data; |a t -μ|<δσ; In the above formula, a t is the measured data at time t, μ is the mean of the measured data at n times, δ is the set confidence threshold, and σ is the standard deviation of the measured data at n times; The bad data elimination unit is used to use the following autocorrelation function to describe the correlation between abnormal data and measured data at other times, and eliminate bad measurement data: In the above formula, f k is the correlation coefficient, a t+k is the measurement data at time t+k; The bad data replacement unit is used to input the conditional information and the bad data set at a specific position represented by the mask matrix into the trained conditional denoising diffusion implicit model, and obtain the measurement data after the bad data is replaced through the reverse denoising process to obtain a complete measurement data set; The data frequency fusion unit is used to adopt different frequency fusion strategies for different types of data based on the obtained complete measurement data set, including: For the synchronous phasor unit data measurement, the following formula is used to calculate the fused measurement value of the synchronous phasor unit data at time t: In the above formula, Z t，P is the fused measurement value of the synchronized phasor unit data at time t, N is the number of synchronized phasor unit data measured before time t, n is the number of current measurements, and Z (t+nT)，P is the fused measurement value of the synchronized phasor unit data at time t+nT, where T is the sampling period of the synchronized phasor unit data; For the data measurement of the data collection and monitoring system, the following formula is used to calculate the fusion measurement value of the data collection and monitoring system data after the fusion prediction value at time t: In the above formula, Z t，S is the fusion measurement value of the data acquisition and monitoring system data at time t, c is the sampling time before time t, a is the sampling time after time t, and Z a，Y , Z c，Y They are the fused measurement values ​​of the data acquisition and monitoring system data predicted at time a and time c respectively; For advanced measurement system data measurement, the following formula is used to calculate the fusion measurement value of the advanced measurement system data at time t: WITH A，t =Z A，i +x(Z A，i+1 -WITH A，i )+y(Z A，i+1 -WITH A，i ) 2 +z(Z A，i+1 -WITH A，i ) 3 ; In the above formula, Z A，t is the fused measurement value of the advanced measurement system data at time t, and t∈(i, i+1), Z A，i is the fusion measurement value of the advanced measurement system data at time i, Z A，i+1 is the fused measurement value of the advanced measurement system data predicted at time i+1, i is the initial time of cubic spline interpolation, and x, y, and z are the weights of each interpolation term.

9. A multi-source data fusion device for an active power distribution network, characterized in that: including a processor and a memory; The memory is used to store computer program code and transmit the computer program code to the processor; The processor is used to execute the multi-source data fusion method for an active power distribution network according to any one of claims 1 to 4 according to the instructions in the computer program code.

10. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a multi-source data fusion method for an active power distribution network described in any one of claims 1 to 4 are implemented.

Images