Optimization processing method and system for fusion data of multi-energy system

Abstract

The invention relates to an optimization processing method and system for fusion data of a multi-energy system, and the method comprises the following steps: S1, obtaining the data of the multi-energy system, including a wind power system and a photovoltaic system, and carrying out the preprocessing; s2, for the processed multi-energy system data, performing noise elimination and state smoothing processing on the preprocessed data by using a Kalman filtering technology to obtain stable state representation of the multi-source data; s3, constructing a fractal attention depth separable convolution GRU fusion network, and representing the stable state of the multi-source data to obtain prediction compensation data; s4, reconstructing and correcting the prediction compensation data based on a time domain analysis and probability statistics method; and S5, according to the reconstructed and corrected supplementary data based on the time domain and probability statistics, fine fusion of the multi-source data is realized through adaptive weight adjustment, and optimized fusion data is obtained. According to the invention, effective fusion and compensation of multi-source data are realized, and the real-time performance, the accuracy and the system safety of fault prediction of new energy equipment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data optimization, in particular to a multi-energy system fusion data optimization processing method and system. BACKGROUND

[0002] With the continuous development and application of new energy technology, the multi-energy development productivity level of wind power, light and other energies in Qinghai region is continuously improved. Qinghai has become an important new energy base in China, and Qinghai power grid has become the provincial power grid with the highest proportion of clean energy and new energy installed capacity in China. However, with the continuous increase of the proportion of new energy grid connection, the data generated in the operation process of various new energy equipment in the power grid shows diversity. The existing multi-system energy data has low data utilization rate due to complex structure type, high redundancy, data missing and other problems. However, the traditional data processing method has problems such as insufficient data preprocessing and limited feature extraction. In view of the above problems, the existing technology is mainly concentrated in single data source or fixed data fusion method, and a complete system based on multi-source heterogeneous data fusion fault prediction and compensation technology scheme has not been formed. Especially in the aspects of data denoising, missing compensation and multi-time scale representation, the traditional method cannot balance the deep nonlinear feature extraction of various data and the real-time feedback demand, thereby affecting the effect of power grid multi-energy system in operation scheduling and fault warning. SUMMARY

[0003] In order to solve the above problems, the purpose of the present application is to provide a multi-energy system fusion data optimization processing method and system, which solves the problems of data noise, missing, redundancy and complex structure existing in the current multi-energy system fusion data acquisition and fusion process.

[0004] To achieve the above purpose, the technical scheme adopted by the present application is as follows: A multi-energy system fusion data optimization processing method, comprising the following steps: S1: acquiring multi-energy system data, including wind power, photovoltaic system, and preprocessing; S2: for the processed multi-energy system data, using Kalman filter technology to eliminate noise and smooth the state of the preprocessed data, to obtain a stable state representation of multi-source data; S3: constructing a fractal attention deep separable convolution GRU fusion network, obtaining predicted compensation data from the stable state representation of multi-source data; S4: reconstructing and correcting the predicted compensation data based on time domain analysis and probability statistics method; S5: based on the reconstructed and corrected supplementary data based on time domain and probability statistics, realizing fine fusion of multi-source data through adaptive weight adjustment, to obtain optimized fusion data.

[0005] Further preprocessing, including time alignment, data normalization, and missing value imputation, is as follows: The time alignment maps data with different sampling frequencies to a unified time axis; the normalization process maps each data point to a preset interval according to the minimum-maximum normalization method; and the missing value filling uses a linear interpolation method to compensate for missing data.

[0006] Furthermore, Kalman filtering is specifically divided into two steps: prediction and update. The physical model is used to predict the state variables and error covariance, and sensor measurement data is used to correct the predicted values, resulting in a more accurate state estimate. The state prediction equation is: ; in, It is a moment Prior state estimation, Here is the state transition matrix. For posterior state estimation, For external control input; The error covariance equation is: ; in, Let be the prior error covariance matrix. The posterior covariance matrix is... The process noise covariance matrix; During the update phase, the Kalman gain needs to be determined first. The Kalman gain is calculated as follows: ; in, For Kalman gain, For the measurement matrix, To measure the noise covariance matrix; After obtaining the Kalman gain, the state update equation and the error covariance update equation can be obtained; The state update equation is: ; in, for Time sensor measurement value, The measured estimate is for the predicted state. This is a posterior estimate; The error covariance update equation is: ; in, The updated state error covariance matrix is ​​the posterior error. The identity matrix is ​​used, and the above process is repeated after entering the next time step.

[0007] Furthermore, the fractal attention-based depthwise separable convolutional GRU fusion network, through recursive generation... A depthwise separable convolutional branch extracts multi-scale local spatiotemporal features, and some branches are randomly masked during training to enhance regularization. Then, the outputs of each branch are concatenated and subjected to 1×1 pointwise convolution for dimensionality reduction. At the same time, a channel / spatial attention module is incorporated to adaptively weight and fuse the dimensionality-reduced features. Subsequently, the fused feature map is activated by ReLU and downsampled through a pooling layer to obtain a refined spatiotemporal feature representation. Finally, the refined spatiotemporal feature representation is flattened and input into a GRU unit to model the state evolution at each time step. Feature selection and classification are completed through a fully connected layer, and the final prediction result is output through a Softmax function.

[0008] Furthermore, multi-scale local spatiotemporal features are extracted through recursively generated strip depthwise separable convolutional branches, and some branches are randomly masked during the training phase to enhance regularization, as detailed below: First, the data from wind power and photovoltaic power, after Kalman filtering, are organized as follows: ;

[0009] in, Indicates spatial dimension, Indicates the number of channels; Multi-scale features are extracted using fractal depthwise separable convolutional branches, and for each branch... Define depth ; the first branch Deep convolution calculation: First, deep convolution is achieved by independently convolving each input channel. The specific calculation process is as follows: ; Among them, the kernel size The channels are not mixed, and the output remains the same. ; After depthwise convolution, its projection is input into a pointwise convolutional layer, and the output after pointwise convolution is: ; After linear combination, the local features of each channel are converted into global combined features, generating new feature representations; After each depthwise convolution and pointwise convolution, the convolution output is first batch normalized, and then ReLU activation is applied. The detailed calculation process is as follows: ; in These are the mean and standard deviation of the channel; By repeatedly calling itself using the same set of basic convolutional units, multiple self-similar parallel paths are constructed, achieving recursive stacking: ; First, the input , then the result as input, through recursive iteration, finally use as intermediate results, ultimately get output ; ; Drop-Path, fractal network for each parallel branch will be sampled a Bernoulli mask: , ensure , and calculate:

[0010] Here divide to ensure the expectation does not change, , reasoning when .

[0011] Further, the output of each branch is spliced and reduced dimension by 1×1 point convolution, and the channel / space attention module is integrated to adaptively weight and fuse the reduced dimension features, as follows: Using multi-branch fusion and channel dimension reduction, after completing Drop-Path, the network obtains normalized outputs of branches, first do simple average: ; On , apply 1×1 convolution for linear projection:

[0012] Among them , , the number of channels is compressed from to , and finally output as the input of the next channel / space attention fusion.

[0013] Further, the fused feature map is then activated by ReLU and down-sampled by the pooling layer to obtain refined spatio-temporal feature representation: Using channel / space attention fusion, after obtaining the reduced dimension feature map , channel attention and spatial attention are applied in turn: Channel attention: first perform pooling statistics operation, respectively do global average pooling and global maximum pooling along the spatial dimension, get two channel description vectors: ; ​Further, the two vectors are respectively input into a two-layer fully connected network, and then dimensionality reduction and dimensionality increase are performed:

[0014] wherein is the reduction ratio; Further, the two outputs are summed and passed through a sigmoid function to obtain a channel attention weight vector:

[0015] Further, the two vectors are respectively input into a two-layer fully connected network, and then dimensionality reduction and dimensionality increase are performed: extended to and multiplied with the original features channel by channel to realize channel recalibration, and the specific formula is as follows:

[0016] Spatial attention: to the recalibrated Along the channel dimension, maximum pooling and average pooling are performed to generate two single-channel feature maps:

[0017] Further, the two are spliced in the channel dimension to obtain a feature with a shape Then, a 7x7 convolution is performed and activated: ; wherein, an output single-channel map is output; Finally, the is broadcast multiplied back ;

[0018] Thus, a fused feature map is generated.

[0019] Further, the refined spatiotemporal feature representation is flattened and input into a GRU unit to model the state evolution at each time, and a fully connected layer is used to complete feature selection and classification decision, and the final prediction result is output through a Softmax function, and the specific formula is as follows: After obtaining the attention fusion feature map, a nonlinear activation is applied, first, a RELU function is activated:

[0020] Then, spatial down-sampling is performed on Global pooling, average pooling and global average pooling methods are used, and the specific formula is as follows:

[0021] In the Fractal-Attention-DSCNN-GRU, first, a local down-sampling is performed by using a max-pooling or average-pooling to reduce the dimension of the feature vector from to , which not only preserves the spatial structure but also reduces the computational complexity. Finally, a global average pooling is used to completely flatten the dimension of the feature vector from to , forming a feature vector of the time , which is then sent to the GRU for time series modeling. In this way, the activation and down-sampling together complete the transition from high-dimensional spatial features to one-dimensional time series input. The GRU time series modeling takes as the input of the current step GRU and recursively propagates along the time series: ; The GRU determines how much new information to update through , thereby modeling long-term dependencies while suppressing gradient vanishing, and outputs the hidden state . Further, a fully connected linear mapping is performed on the hidden state of the GRU to obtain a task-related intermediate vector:

[0022] wherein is the output dimension. Finally, the is sent to an activation function to generate the final prediction:

[0023] wherein, is the final prediction result.

[0024] An optimization processing system for multi-energy system fusion data, comprising a processor, a memory and a computer program stored in the memory, wherein the processor executes the computer program to specifically execute the steps of the optimization processing method for multi-energy system fusion data.

[0025] A computer storage medium, characterized in that the computer storage medium stores a plurality of instructions, the instructions being adapted to be loaded and executed by a processor to perform the above method steps.

[0026] The present application has the following beneficial effects: Firstly, the original data is smoothed by Kalman filtering and state estimation, then the nonlinear and spatiotemporal multiscale features in the data are extracted by combining a multiscale spatiotemporal feature extraction network, and finally, the effective fusion and compensation of multi-source data are realized through an adaptive weight mechanism and an improved data fusion strategy (for example, combining improved SVM, random forest and D-S evidence theory), thereby improving the real-time performance, accuracy and system safety of new energy equipment fault prediction, solving the problems of data redundancy, noise and missing in multi-energy system fusion data, and providing more refined and data-driven decision support for energy optimization scheduling and fault diagnosis, and having obvious technical advantages and application prospects. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a flowchart of the present application; Figure 2 is a working principle diagram of DS-CNN in an embodiment of the present application; Figure 3 is a principle diagram of GRU module in an embodiment of the present application; Figure 4 is a FractalNet model structure diagram in an embodiment of the present application; Figure 5 is a Fractal-Attention-DSCNN-GRU network structure diagram in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The present application will be further described in detail below in combination with the drawings and specific embodiments: REFERENCE Figure 1 In the embodiment, a multi-energy system fusion data optimization processing method is provided, which comprises the following steps: S1: Obtain multi-energy system data, including wind power and photovoltaic systems, and perform time alignment, normalization and data missing filling and other preprocessing processes to ensure clean data input for the later filtering process; S2: For the processed multi-energy system data, use Kalman filtering technology to eliminate noise and smooth the state of the preprocessed data, to obtain a stable state representation of the multi-source data; S3: Construct a Fractal-Attention-DSCNN-GRU fusion network, and obtain prediction compensation data from the stable state representation of the multi-source data; S4: Based on time domain analysis and probability statistical method, the predicted compensation data is reconstructed and corrected; after Fractal-Attention-DSCNN-GRU processing, time domain analysis is needed, and the time sequence characteristics of the output features of Fractal-Attention-DSCNN-GRU are further processed and analyzed. Part of the data changing with time after Fractal-Attention-DSCNN-GRU processing, such as the features output by Fractal-Attention-DSCNN-GRU of the wind power system, can include wind speed, unit speed and other data changing with time; the features extracted by Fractal-Attention-DSCNN-GRU of the photovoltaic system can include component temperature, current, voltage and other features changing with time; the time domain features can be obtained by mean and standard deviation, skewness and kurtosis, autocorrelation and other time domain analysis methods. Through the above operation, the dynamic characteristics of the time sequence features can be captured. Further, through the method of probability statistics, the distribution and statistical law of the time sequence features are described and quantified. The probability statistical method can use probability density function and cumulative distribution function, anomaly detection and confidence interval, hypothesis testing and statistical inference to provide quantitative description on probability for equipment state evaluation and fault prediction, and the time-varying characteristics in time domain analysis are qualitatively described; S5: According to the supplementary data reconstructed and corrected based on time domain and probability statistics, the fine fusion of multi-source data is realized through adaptive weight adjustment, and the optimized fusion data is obtained; first, the features reflecting the time dynamics and statistical distribution are extracted from the data, and then the gradient of the loss function with respect to each feature weight is calculated by gradient descent method, and the weight of each feature group is further updated dynamically. Not only solves the limitations that may be caused by fixed weight, but also can continuously optimize the fusion strategy according to the characteristics of real-time data, while retaining the effective information of different data sources, solving the problems of data redundancy and missing, realizing the real-time mapping of physical model and digital model.

[0029] In the embodiment, the way of preprocessing data mainly includes three parts of time alignment, data normalization and missing value filling.

[0030] 1. Time alignment The purpose of time alignment is to ensure that the data from each data source in the wind power and photovoltaic system has consistency on the unified time axis. Linear interpolation method is adopted.

[0031] (1) Core principle Linear interpolation: The sampling frequencies of various sensors may be different or there may be timestamp deviation when collecting data. If unified sampling data is to be obtained at a certain time, the data value at the target time can be calculated by using the linear relationship between the known data points. Specifically, if the sampling frequency of the target time is different from that of the known data points, the data value at the target time can be calculated by linear interpolation. Obtain the value At time Obtain the value Then for Satisfy The data value of the target moment, its data value Can be calculated by the following formula:

[0032] Step 1: Determine the unified time axis According to the overall application requirements, determine a unified time range and sampling interval. For example, set a unified sampling interval of once per second, from the start time To the end time Generate an equally spaced time sequence.

[0033] Step 2: Collect the original timestamps and corresponding data of each data source. For each data source in the wind power and photovoltaic system, obtain the original timestamp and data value collected. For example, wind speed, power, and irradiance data correspond to their respective time points.

[0034] Step 3: Interpolate each data source. For each data source, according to the unified time axis, within the original timestamp range, calculate the data value corresponding to the target time point by linear interpolation. If the target time point is earlier or later than the original data range, you can choose to extrapolate or assign missing values.

[0035] Step 4: Generate the aligned data set. Combine the interpolation results of all data sources according to the unified time axis to form a multi-source data matrix with the same row time label, providing consistent data input for subsequent Kalman filtering and feature extraction.

[0036] 2. Data normalization The role of data normalization is to convert data of different magnitudes and units to the same range, so as to eliminate the influence between different indicators, making the subsequent processing and model training more stable and faster convergence. Here, the minimum-maximum normalization is used, which is specifically represented by the following formula:

[0037] In the formula, is the original data, is the minimum value, is the maximum value, is the data mapped to the range of 0, 1.

[0038] Step 1: Determine the data range. For each feature that needs to be normalized, first calculate the minimum value and the maximum value .

[0039] Step 2: Apply the normalization formula. Subtract the original data... Substituting into the normalization formula, we obtain the normalized data. .

[0040] Step 3: Handling exceptional cases. If the maximum and minimum values ​​in the data are equal (i.e., the feature value is constant), special handling is usually required, such as directly assigning the feature a constant value (e.g., 0 or 0.5) to prevent the denominator from being zero.

[0041] 3. Missing value imputation The missing value filling part is also filled using linear interpolation.

[0042] Step 1: Identify missing values. Mark missing values ​​in the data (usually represented by NaN).

[0043] Step 2: Determine the valid data points before and after the missing data. For each missing value, find the previous and next valid data points.

[0044] Step 3: Apply the linear interpolation formula. Use the above formula to calculate the estimated value of the target time and replace the missing positions with the estimated value.

[0045] Step 4: Handling boundary cases. If missing values ​​appear at the beginning or end of the sequence, interpolation cannot be performed due to the lack of preceding and following data. In this case, methods such as forward padding, backward padding, or padding with the global mean can be used.

[0046] In this embodiment, Kalman filtering can be divided into two steps: prediction and update (predicting state variables and error covariance using a physical model and correcting the predicted values ​​using sensor measurement data to obtain a more accurate state estimate). The state prediction equation is:

[0047] in, It is a moment Prior state estimation, Here is the state transition matrix. For posterior state estimation, For external control input.

[0048] The error covariance equation is:

[0049] in, Let be the prior error covariance matrix. The posterior covariance matrix is... Let be the process noise covariance matrix.

[0050] The prediction stage is based on historical estimates to calculate the state at this time, but no correction is made, and the predicted state needs to be corrected through the update stage to make the data more reliable and more real. In the update stage, the Kalman gain needs to be determined first, and the calculation of the Kalman gain is as follows:

[0051] wherein, is the Kalman gain, is the measurement matrix, is the measurement noise covariance matrix.

[0052] The Kalman gain plays a key role in weighting and balancing between the measurement value and the predicted value. If the noise is large, the gain is small, and the influence of the measurement data is weak; if the prediction error is large, the gain is large, and the influence of the measurement data is strong. After obtaining the Kalman gain, the state update equation and the error covariance update equation can be obtained.

[0053] The state update equation is:

[0054] wherein, is the measurement value at time t, is the measurement estimated value of the predicted state, is the posterior estimated value. The error covariance update equation is:

[0055] wherein,

[0056] is the updated state error covariance matrix, i.e., the posterior error, is the unit matrix. After entering the next time section, the above process is repeated. This process in the state estimation of wind power and photovoltaic systems, in the presence of noise and incomplete data, smoothly estimates key parameters (such as wind speed, component temperature, current, voltage, etc.), providing high-quality input data for subsequent Fractal-Attention-DSCNN-GRU fusion network. In this embodiment, after the wind power and photovoltaic multi-source data smoothed and denoised by Kalman filtering are constructed into a two-dimensional matrix, they are first sent to the Fractal-Attention-DSCNN-GRU fusion network: the network generates

[0057] ​The strip depth separable convolution branch extracts multi-scale local space-time features, and part of the branch is randomly shielded (Drop-Path) during the training stage to enhance regularization; then, the outputs of each branch are spliced and reduced in dimension through 1*1 point convolution, and the channel / space attention module is incorporated to adaptively weight and fuse the reduced features; then, the fused feature map is activated through ReLU and down-sampled through the pooling layer to obtain a refined space-time feature representation; finally, the flattened space-time feature representation is input into the GRU unit to model the state evolution at each time, the feature selection and classification decision are completed through the fully connected layer, and the final prediction result is output through the Softmax function Step 1: First, the data from wind power and photovoltaic after Kalman filtering is arranged as .

[0058]

[0059] wherein, represents the spatial dimension, represents the number of channels.

[0060] Step 2: As an optimization scheme of the above embodiment, multi-scale features are extracted by fractal depth separable convolution branches. For each branch define the depth . The first layer convolution calculation of the branch: first, independent convolution is performed on each input channel to realize deep convolution, and the specific calculation process is as follows:

[0061] wherein the kernel size , the channels are not mixed, and the output is still .

[0062] After deep convolution, it can be projected into a point convolution layer, and the output after point convolution is:

[0063] After linear combination, the local features of each channel are converted into global combined features to generate a new feature representation. After each "deep convolution + point convolution", the convolution output is first batch normalized (BatchNorm) and then ReLU activated, and the detailed calculation process is as follows:

[0064] wherein is the mean and standard deviation of the channel. Use the same set of basic convolution units to call themselves constantly to build "self-similar" multiple parallel paths to realize recursive stacking. The specific process is as follows:

[0065] First, the result is taken as input, through recursive iteration, and finally use as intermediate results, the final output . Step 3, as an optimization scheme of the above example, Drop-Path (path level drop and normalization) is adopted. Fractal network will sample a Bernoulli mask for each parallel branch: , ensure

[0066] , and calculate:

[0067] Here divide to ensure the expectation does not change, , reasoning when .

[0068] Step 4, as an optimization scheme of the above example, multi-branch fusion and channel dimension reduction are adopted. After completing Drop-Path, the network obtains normalized outputs of branches. In order to converge the features of these multiple paths in the channel dimension without changing the scale, we first do a simple average:

[0069] This not only preserves the multi-scale information of each branch, but also maintains the unbiased expectation of the output. Then, in order to control the input dimension of the subsequent GRU module and further fuse the features between channels, a 1x1 convolution is applied on to perform linear projection:

[0070] Among them , , the number of channels is compressed from to . This operation is not only a dimension reduction method, but also realizes cross-channel feature recombination through learnable weights, which helps to generate attention weights and capture key information in subsequent timing modules. Finally, the output is taken as the input of the next channel / space attention fusion.

[0071] Step 5, as an optimization scheme of the above example, channel / space attention fusion is adopted. After obtaining the reduced feature map ​​After that, CBAM (Convolutional Block Attention Module) applies channel attention and spatial attention in turn, and the specific process and formula are as follows: 1. Channel attention First, perform the pooling statistical operation on Along the spatial dimension, respectively do global average pooling and global maximum pooling to get two channel description vectors:

[0072] Further, input the two vectors into a two-layer fully connected network (shared weights) respectively, and then reduce the dimension and then increase the dimension:

[0073] Among them, is the reduction ratio.

[0074] Further, sum the two-way output and pass through the sigmoid function to get the channel attention weight vector.

[0075]

[0076] Further, is extended to and multiplied by the original feature channel by channel to realize channel recalibration. The specific formula is as follows:

[0077] 2. Spatial attention On the recalibrated Along the channel dimension, respectively do maximum pooling and average pooling to generate two single-channel feature maps:

[0078] Further, splice the two in the channel dimension to get a feature with a shape , then pass through a 7x7 convolution and activation:

[0079] Among them, output a single-channel map.

[0080] Finally, broadcast multiply back .

[0081]

[0082] Thus, the fused feature map is generated.

[0083] Step 6, as an optimization scheme of the above examples, activation and down-sampling are used. After obtaining the attention fusion feature map, a nonlinear activation is applied. First, the RELU function is used for activation:

[0084] Then, spatial down-sampling (Pooling) is performed on using global pooling, average pooling, and global average pooling methods. The specific formula is as follows:

[0085] In Fractal-Attention-DSCNN-GRU, first, a max-pooling or average-pooling is used for local down-sampling, reducing to , which preserves the spatial structure and reduces the computational complexity; finally, global average pooling is used to completely flatten the dimension to form the feature vector at time , which is then sent to GRU for time series modeling. In this way, activation and down-sampling together complete the transition from high-dimensional spatial features to one-dimensional time series input.

[0086] Step 7, as an optimization scheme of the above examples, time series modeling (GRU) is used. Take as the input of the current step GRU, and recursively along the time series:

[0087] GRU determines how much new information to "update" through , thereby modeling long-term dependencies while suppressing gradient vanishing, and outputs the hidden state .

[0088] Further, a fully connected linear mapping is performed on the hidden state of GRU to obtain a task-related intermediate vector:

[0089] where is the output dimension.

[0090] Finally, send to the activation function to generate the final prediction:

[0091] where is the final prediction result.

[0092] In the embodiment, time domain analysis and probability statistics method are used to model and statistically describe the output features of Fractal-Attention-DSCNN-GRU.

[0093] As an optimization scheme of the above embodiment, first, time domain analysis is used to describe the dynamic characteristics of the sequence data in the time dimension. Let the feature sequence output by Fractal-Attention-DSCNN-GRU be: The time domain indicators include mean, standard deviation, skewness, kurtosis and autocorrelation coefficient. The above indicators can reveal the average trend, fluctuation degree, symmetry, sharpness and time correlation of the feature sequence, thereby representing the dynamic characteristics of the time sequence from various aspects.

[0094] As an optimization scheme of the above embodiment, in the probability statistics stage, the application uses the estimated time domain statistical parameters to model the distribution of the feature sequence and performs the following quantitative processing: 1. Adopting probability density function estimation, the sequence probability density is expressed as:

[0095] Where k is the kernel function and h is the bandwidth.

[0096] 2. As an optimization scheme of the above embodiment, further cumulative distribution function estimation

[0097] Where is the indicator function.

[0098] 3. Based on the above distribution information, the application constructs the confidence interval, and assuming that the data obeys normal distribution, there is the following formula:

[0099] 4. As an optimization scheme of the above embodiment, Z-Score method is used for data anomaly correlation detection, as follows:

[0100] When the absolute value of exceeds the set threshold value, it is detected as abnormal data.

[0101] In step S4, through the above time domain analysis and probability statistical method, not only the qualitative description of the time dynamic characteristics of the Fractal-Attention-DSCNN-GRU output feature sequence is realized, but also the quantitative modeling of the probability density, cumulative distribution and confidence interval is realized, so that the reconstructed data not only makes up for the original missing and noise problems, but also achieves more reasonable and more real performance in statistical distribution.

[0102] In the embodiment, an adaptive feature weight adjustment mechanism based on time domain analysis and probability statistical feedback is introduced, and the weight setting of feature fusion is dynamically updated in a data-driven manner, so as to realize fine fusion and redundancy suppression, and to complete efficient mapping of the physical-digital model. Mainly includes the following steps: Step 1: Extract multi-source feature groups, respectively: Time domain features:

[0103] Probability statistical features:

[0104] Deep features extracted from Fractal-Attention-DSCNN-GRU:

[0105] Spliced into a fusion feature vector:

[0106] Step 2: Design a weighted fusion mechanism The final representation form of multi-source heterogeneous data fusion features is:

[0107] Among them, are the fusion weights of the three types of features respectively. At the same time, satisfy the normalization constraint.

[0108] Step 3: Define the loss function and feedback optimization The loss function is defined as:

[0109] Among them, The predicted value obtained after the output passes through the prediction model is , and the actual label is .

[0110] As an optimization scheme of the above embodiment, the chain rule is used to calculate the gradient of the loss function with respect to each fusion weight:

[0111] As an optimization scheme of the above embodiment, the gradient descent method is used to dynamically update the weights:

[0112] wherein is the learning rate.

[0113] After each update, the softmax normalization is used:

[0114] The process is iterated in each training batch, adjusting the weight configuration of feature fusion in real time.

[0115] Step 4, fusion of optimization results and physical-digital model mapping As an optimization scheme of the above embodiment, after optimization, The feature information in the time domain statistics is retained, while the redundant features are suppressed. The behavior prediction of the physical model variables reflected in the digital model through feature mapping realizes the real-time mapping of the unified digital feature space of the multi-source heterogeneous physical data.

[0116] Those skilled in the art will understand that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0117] The present application is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The means for implementing the functions specified in a flow or multiple flows and / or blocks.

[0118] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0119] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, so that the instructions executed on the computer or other programmable data processing devices provide processes for implementing the flow Figure 1 one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0120] The above description is only the preferred embodiment of the present application, not other forms of the present application, any skilled in the art can use the above disclosed technical content to change or modify as equivalent embodiments of equivalent changes. But any simple modification, equivalent change and modification of the above embodiments without departing from the technical solution of the present application, according to the technical essence of the present application, still belongs to the protection scope of the technical solution of the present application.

Claims

1. A method for optimizing processing of multi-energy system fusion data, characterized in that, The method comprises the following steps: S1: obtaining multi-energy system data including wind power and photovoltaic systems, and preprocessing; S2: using Kalman filtering technology to remove noise and smooth the state of the preprocessed data to obtain a stable state representation of the multi-source data; S3: constructing a fractal attention deep separable convolution GRU fusion network to obtain prediction compensation data from the stable state representation of the multi-source data; S4: reconstructing and correcting the prediction compensation data based on time domain analysis and probability statistics methods; S5: realizing fine fusion of multi-source data based on adaptive weight adjustment to obtain optimized fused data.

2. The method of claim 1, wherein The preprocessing includes time alignment, data normalization and missing value filling, and specifically as follows: The time alignment maps data of different sampling frequencies to a unified time axis; the normalization processing maps each data to a preset interval according to the minimum-maximum normalization method; and the missing value filling compensates for the missing part of the data by using the linear interpolation method.

3. The method of claim 1, wherein the method further comprises: The Kalman filtering is specifically divided into two steps, namely prediction and update, the state quantity and error covariance are predicted using a physical model, and the predicted value is corrected using sensor measurement data to obtain more accurate state estimation, the state prediction equation is: ； wherein, is the prior state estimate at time , is the state transition matrix, is the posterior state estimate, is the external control input; The error covariance equation is: ； wherein, is a prior error covariance matrix, is a posterior covariance matrix, is a process noise covariance matrix; In the update stage, the Kalman gain needs to be determined, and the Kalman gain calculation is as follows: ； wherein, is the Kalman gain, is the measurement matrix, is the measurement noise covariance matrix; After obtaining the Kalman gain, the state update equation and the error covariance update equation are obtained; The state update equation is: ； wherein is the time instant sensor measurement, is the measurement estimate of the predicted state, is the posterior estimate; The error covariance update equation is: ； wherein, is the updated state error covariance matrix, i.e., the posterior error, is the identity matrix, and the above procedure is repeated after entering the next time segment.

4. The method of claim 1, wherein the method further comprises: The fractal attention depth separable convolution GRU fusion network extracts multi-scale local space-time features through a depth separable convolution branch The depth separable convolution branch extracts multi-scale local space-time features, and part of the branches is randomly shielded in the training stage to enhance regularization; then, the outputs of the branches are spliced and reduced in dimension through 1×1 point-by-point convolution, and a channel / space attention module is simultaneously incorporated to adaptively weight and fuse the reduced features; subsequently, the fused feature map is subjected to ReLU activation and down-sampling through a pooling layer to obtain a refined space-time feature representation; finally, the refined space-time feature representation is flattened and input into a GRU unit to model the state evolution at each time, complete feature selection and classification decision through a fully connected layer, and output the final prediction result through a Softmax function.

5. The method of claim 4, wherein the method further comprises: The multi-scale local spatio-temporal features are extracted by the recursively generated bar deep separable convolution branch, and part of the branch is randomly shielded during the training stage to enhance regularization, and specifically as follows: First, the data from wind power and photovoltaic after Kalman filter processing is arranged as ; ; wherein, denotes the spatial dimension, denotes the number of channels; A fractal depth separable convolution branch is adopted to extract multi-scale features, and a first layer of the branch is defined as a depth define depth ; the first layer of the branch is defined as a depth Layer convolution calculation: first, deep layer convolution is realized by independent convolution on each input channel, and the specific calculation process is as follows: ； where the core size , the channels do not mix, and the output is still ; After deep convolution, the projection input is input into the pointwise convolution layer, and the output after pointwise convolution is: ； After linear combination, the local features of each channel are converted into global combined features to generate new feature representations; After each deep convolution and pointwise convolution, the convolution output is first normalized in batches, and then connected with the ReLU activation, and the detailed calculation process is as follows: ； wherein is the mean and standard deviation of the channel; The same set of basic convolution units are called recursively to build self-similar multiple parallel paths to realize recursive stacking: ； First, the input is done , the result is taken as input, through recursive iteration, finally with as intermediate results, the final output ; With Drop-Path, the Fractal network samples a Bernoulli mask for each parallel branch: , ensuring , and computing: ; Here divided by Guarantee desired invariants, , reasoning time .

6. The method of claim 5, wherein the method further comprises: After the outputs of the branches are spliced, the dimensionality is reduced by 1x1 pointwise convolution, and the channel / space attention module is simultaneously incorporated to adaptively weight and fuse the reduced features, and specifically as follows: With multi-branch fusion and channel dimension reduction, the network obtains Normalized output of the i-th branch First, do a simple average: ； In Apply 1x1 convolution for linear projection: ; wherein , , the number of channels is compressed from to , and the final output is taken as the input of the next step of channel / space attention fusion.

7. The method of claim 6, wherein the method further comprises: The fused feature map is then activated by ReLU and down-sampled by the pooling layer to obtain refined spatio-temporal feature representations: The channel / space attention fusion is adopted, and the feature map after dimension reduction is obtained After that, the channel attention and the space attention are applied in sequence Channel attention: first, the pooling statistics operation is performed on Global average pooling and global max pooling are performed along the spatial dimension, respectively, to obtain two channel description vectors: ； The two vectors are further input into a two-layer fully connected network, first reduced in dimension and then increased in dimension: ; wherein r is the reduction ratio; The sum of the two outputs is further calculated and passed through the sigmoid function to obtain a channel attention weight vector: ; Further to extended to and multiplied by the original features channel by channel to achieve channel re-scaling, with the specific formula as follows: ; Spatial attention: on the re-scaled Max-pooling and average-pooling are performed along the channel dimension, respectively, to generate two single-channel feature maps: ; Further, both are concatenated in the channel dimension to get a feature map of shape and then passed through a 7x7 convolution and activated: ； wherein, outputting a single channel map; Eventually will Broadcast times back ; ; to generate a fused feature map .

8. The method of claim 7, wherein the method further comprises: After the refined spatio-temporal feature representations are flattened, they are input into the GRU unit to model the state evolution at each time, the feature selection and classification decision are completed through the fully connected layer, and the final prediction result is output through the Softmax function, and specifically as follows: After obtaining the attention fusion feature map, a nonlinear activation is applied, first, the RELU function is activated: ; Then, to Spatial downsampling is performed using global pooling, average pooling, and global average pooling methods, with the specific formulas as follows: ; In the Fractal-Attention-DSCNN-GRU, first, a local down-sampling is performed using a max-pooling or average-pooling to reduce the size of the input image from to , which preserves the spatial structure and reduces the computational cost. ​ Finally, global average pooling is used to flatten the dimensions, resulting in a feature vector at time which is fed into the GRU for temporal modeling; in this way, the activation and downsampling jointly complete the transition from high-dimensional spatial features to one-dimensional temporal inputs; GRU time-series modeling is adopted to take as the input of the current step GRU and recursively along the time series. ； The GRU passes through decides how much new information to update, thereby modeling long-term dependencies while suppressing gradient vanishing, outputting a hidden state ; Further, the hidden state of the GRU is linearly mapped by full connection to obtain a task-related intermediate vector: ; wherein is the output dimension; Finally, the output of the last layer is passed into an activation function to generate the final prediction: the output of the last layer is passed into an activation function to generate the ; wherein, is the final prediction result.

9. An optimization processing system of multi-energy system fusion data, characterized by, The processor executes the computer program, and specifically executes the steps of the optimization processing method for multi-energy system fusion data according to any one of claims 1-8.

10. A computer storage medium, characterized in that, The computer storage medium stores a plurality of instructions, which are suitable for being loaded and executed by the processor, and the method steps of any one of claims 1-8.

Images