Energy storage power station monitoring data missing complementing method and system based on feature fusion

Abstract

According to the energy storage power station monitoring data missing complementing method and system based on feature fusion, all historical monitoring data and actual monitoring data of an energy storage power station are obtained, and multiple groups of monitoring data with different data missing rates are generated based on a probability model; each group of monitoring data comprises complete random missing data, random missing data, non-random missing data and actual monitoring data; forming a monitoring data set of the energy storage power station by a plurality of groups of monitoring data with different data missing rates; based on the network structure of the noise reduction auto-encoder, an asymmetric noise reduction auto-encoder model is established, and the asymmetric noise reduction auto-encoder model takes the monitoring data set as input and takes the preliminary complementation data set of the monitoring data set as output; outputting a preliminary complementation data set of the monitoring data set of the energy storage power station by using the trained model; and filtering the preliminary complementation data set by using a moving average filter to obtain a complementation data set, thereby realizing accurate complementation of the monitoring data of the energy storage power station.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy storage power station monitoring data processing, and in particular, relates to a method and system for completing missing energy storage power station monitoring data based on feature fusion. Background Art

[0002] With the rapid transition of energy storage power station operation and maintenance from manual regular inspections to preventive status inspection and maintenance, a large number of intelligent sensors and communication devices deployed in the battery management system have realized the real-time collection, transmission and cloud storage of key parameters of energy storage power stations. The massive monitoring data obtained has provided a basis for battery evaluation and trend prediction. However, problems such as equipment power outages, communication fluctuations or sensor failures during the monitoring process will lead to normal operation trends of missing coverage data of various types and degrees in the time series data. Once the missing data is used for status analysis, it will cause significant deviations in the subsequent perception model judgment or prediction, threatening the safe and stable operation of the power system.

[0003] Due to the complex deployment environment of battery energy storage systems, the large number of data parameters that need to be monitored, and the diverse data transmission methods, there are various types of missing data in the actual energy storage system monitoring data, the missing rates vary, and the overall missing rate varies in a wide range, making it difficult to complete. In the prior art, a variety of methods have been proposed to complete the missing data of battery energy storage system monitoring data, which can be mainly divided into statistical methods and data-driven methods. Statistical methods include interpolation methods, low-rank matrix methods, and data-driven methods. However, while simplifying the system model, the existing statistical methods will cause insufficient mining of relationships between data, and it is difficult to cope with data completion in mixed missing scenarios. Data-driven algorithms have better adaptability to various missing data, but due to the low frequency of missing data in actual data and the lack of data volume, data-driven methods still have problems such as insufficient completion accuracy and low stability, and there is a lot of room for optimization. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the present invention provides a method for completing missing data of energy storage power station monitoring based on feature fusion. The method realizes accurate completion of the monitoring data of the energy storage power station by constructing data and completing it twice through an asymmetric encoder and a sliding average filter.

[0005] The present invention adopts the following technical solution.

[0006] The present invention proposes a method for completing missing data of energy storage power station monitoring based on feature fusion, comprising:

[0007] Obtain all historical monitoring data and actual monitoring data of the energy storage power station, and generate multiple sets of monitoring data with different data missing rates based on the probability model. Each set of monitoring data includes completely random missing data, random missing data, non-random missing data, and actual monitoring data. Multiple sets of monitoring data with different data missing rates together constitute the monitoring data set of the energy storage power station.

[0008] Based on the network structure of the denoising autoencoder, an asymmetric denoising autoencoder model is established. The asymmetric denoising autoencoder model takes the monitoring dataset as input and the preliminary complement dataset of the monitoring dataset as output.

[0009] When training the asymmetric denoising autoencoder model, all historical monitoring data are used as the benchmark output. Based on the back propagation method, the loss of the monitoring data set is first calculated. When the loss of the monitoring data set is not greater than the first set threshold, the loss of all missing data is calculated. When the sum of the loss of the monitoring data set and the loss of all missing data is not greater than the second set threshold, the model training is terminated; the trained model is used to output the preliminary completion data set of the monitoring data set of the energy storage power station;

[0010] The preliminary completed data set is filtered using a sliding average filter to obtain the completed data set.

[0011] Preferably, all historical monitoring data and actual monitoring data of the energy storage power station are obtained, and multiple groups of monitoring data with different data missing rates are generated based on the probability model, including:

[0012] Randomly generate a completely random missing rate p within a set range 1 , random missing rate p 2 and the missing non-random rate p 3 , and the sum of the completely random missing rate, random missing rate and non-random missing rate meets the upper limit of the data missing rate;

[0013] Using all the historical monitoring data of the energy storage power station, according to the probability model Generate completely random missing data, where Y is all the historical monitoring data of the energy storage power station, Missing completely random data were generated;

[0014] Using all historical monitoring data and actual monitoring data of the energy storage power station, according to the probability model Generate randomly missing data where is the randomly missing data generated, Z is the actual monitoring data, and p 2 (Z) is the random missing rate of actual monitoring data;

[0015] Using all the historical monitoring data of the energy storage power station, according to the probability model Generate nonrandom missing data where, For the generated non-random missing data, p 3 (Y) is the non-random missing rate of all historical monitoring data;

[0016] Using the completely random missing rate, the random missing rate of actual monitoring data, and the non-random missing rate of all historical monitoring data, the data missing rate of each group of monitoring data is determined using the following relationship:

[0017] P=p 1 +(1-p 1 ) 2 (Z)+(1-p 1 )(1-p 2 (Z))p 3 (Y)

[0018] Where P is the data missing rate of each group of monitoring data.

[0019] Preferably, the setting range is [0, 20%], and the upper limit of the data missing rate is no more than 60%.

[0020] Preferably, As the data missing identifier of the ith data monitoring point in the mth group of monitoring data, When , it means that the i-th data monitoring point is missing in the m-th group of monitoring data, so the position of the i-th data monitoring point in the m-th group of monitoring data is set to 0, otherwise Indicates that the i-th data monitoring point has no missing data in the m-th group of monitoring data, i=1,2,…,N 1 ,m=1,2,…,N 2 , N 1 is the total number of data monitoring points, N 2 is the total number of monitoring data groups in the monitoring data set.

[0021] Preferably, the asymmetric denoising autoencoder model includes: an input layer, an encoder, a jump connection module, a decoder and an output layer; the monitoring data set is input into the encoder via the input layer, the encoder extracts the periodic features and timing features of the monitoring data set, and the jump connection module extracts the element features of the monitoring data set, wherein the periodic features and timing features are both first-dimensional matrices, and the element features are second-dimensional matrices, the decoder fuses the periodic features, timing features and element features, and outputs a feature fused data set; the feature fused data set obtains a preliminary completed data set via the output layer.

[0022] Preferably, based on the back propagation method, the double cosine logarithmic loss function is used to calculate the loss of the monitoring data set, satisfying the following relationship:

[0023]

[0024] Where, L 1 To monitor the loss of the dataset, is the data error of the ith data monitoring point in the mth group of monitoring data, which is the difference between the data of the ith data monitoring point in the mth group of monitoring data and the corresponding historical detection data, N 1 is the total number of data monitoring points, N 2 is the total number of monitoring data groups in the monitoring data set;

[0025] The first threshold is set to [e-ln2,0.5e 2 ].

[0026] Preferably, based on the back propagation method, the MSELoss function is used to calculate the loss of all missing data, satisfying the following relationship:

[0027]

[0028] Where, L 2 is the loss for all missing data, is the data error of the ith data monitoring point in the mth group of monitoring data, As the data missing identifier of the ith data monitoring point in the mth group of monitoring data, When , it means that the i-th data monitoring point is missing in the m-th group of monitoring data, so the position of the i-th data monitoring point in the m-th group of monitoring data is set to 0, otherwise It means that the data of the i-th data monitoring point is not missing in the m-th group of monitoring data.

[0029] Preferably, the loss L of the monitoring dataset is 1 The loss L with all missing data 2 The sum is taken as the total loss of the training model. When the total loss of the training model is less than the second set threshold, the model training is terminated; wherein the second set threshold is 1.85.

[0030] Preferably, the sliding average filtering process is as follows:

[0031]

[0032] In the formula, is the first Initial completion data, is the nth completed data obtained after sliding average filtering, k is the filter window length, which is a positive odd number.

[0033] The present invention also proposes a system for completing missing data of energy storage power station monitoring based on feature fusion, comprising:

[0034] The data processing module is used to obtain all historical monitoring data and actual monitoring data of the energy storage power station, and generate multiple groups of monitoring data with different data missing rates based on the probability model. Each group of monitoring data includes completely random missing data, random missing data, non-random missing data and actual monitoring data. Multiple groups of monitoring data with different data missing rates together constitute the monitoring data set of the energy storage power station.

[0035] The data completion module is used to establish an asymmetric denoising autoencoder model based on the network structure of the denoising autoencoder. The asymmetric denoising autoencoder model takes the monitoring data set as input and the preliminary completed data set of the monitoring data set as output. When training the asymmetric denoising autoencoder model, all historical monitoring data are used as the benchmark output. Based on the back propagation method, the loss of the monitoring data set is first calculated. When the loss of the monitoring data set is not greater than the first set threshold, the loss of all missing data is calculated. When the sum of the loss of the monitoring data set and the loss of all missing data is not greater than the second set threshold, the model training is terminated. The trained model is used to output the preliminary completed data set of the monitoring data set of the energy storage power station. The preliminary completed data set is filtered by a sliding average filter to obtain a completed data set.

[0036] A terminal comprises a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method.

[0037] A computer-readable storage medium stores a computer program, which implements the steps of the method when executed by a processor.

[0038] The beneficial effect of the present invention lies in that, compared with the prior art, at least the method for completing missing data of energy storage power station monitoring by combining an asymmetric denoising autoencoder and a sliding average filter proposed in the present invention performs preliminary completion of missing data in the monitoring data by using an asymmetric denoising autoencoder, and then uses a sliding average filter for secondary optimization, thereby achieving accurate completion of missing data.

[0039] Based on the existing historical monitoring data of energy storage power stations, the present invention fully considers the actual occurrence of missing data, adds completely random missing data, random missing data and non-random missing data according to the missing ratio, and constructs a data set for model training, thereby ensuring that the asymmetric denoising autoencoder can learn as many data missing rules as possible and improve the generalization performance of the network.

[0040] The present invention constructs an asymmetric denoising autoencoder. The encoder is constructed by using convolution and gated recurrent units, where convolution obtains high-dimensional periodic features and gated recurrent units obtain high-dimensional temporal features; jump connections are constructed layer by layer between the encoder and the decoder to extract low-dimensional features; a fully connected layer is used as a decoder to fuse multi-dimensional features and complete the missing part values. The model can effectively improve the model completion accuracy by fusing multi-dimensional features.

[0041] The present invention uses a sliding average filter to perform secondary optimization on the data after the asymmetric denoising autoencoder has completed the completion, which effectively suppresses the high-frequency fluctuations in the data caused by extreme missing completion, and further improves the model's completion accuracy for data with high missing rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of a method for completing missing monitoring data of an energy storage power station based on feature fusion proposed by the present invention;

[0043] Figure 2 is a schematic diagram of an energy storage power station capacity monitoring data set generated in an embodiment of the present invention;

[0044] Figure 3 is a structural diagram of an asymmetric denoising autoencoder model in an embodiment of the present invention;

[0045] Figure 4 is a schematic diagram of the structure of an asymmetric denoising autoencoder in an embodiment of the present invention;

[0046] Figure 5 is a schematic diagram of the structure of a temporal feature extraction module in an embodiment of the present invention;

[0047] Figure 6 is a schematic diagram of sliding average filtering in an embodiment of the present invention;

[0048] Figure 7 4 is a diagram of data completion results in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only embodiments of a part of the present invention, rather than all embodiments. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the protection scope of the present invention.

[0050] The present invention proposes a method for completing missing data of energy storage power station monitoring based on feature fusion, such as Figure 1 As shown, including:

[0051] Step 1: Obtain all historical monitoring data and actual monitoring data of the energy storage power station, and generate multiple groups of monitoring data with different data missing rates based on the probability model. Each group of monitoring data includes completely random missing data, random missing data, non-random missing data and actual monitoring data; multiple groups of monitoring data with different data missing rates together constitute the monitoring data set of the energy storage power station.

[0052] Specifically, step 1 includes:

[0053] Step 1.1: Randomly generate the missing completely at random rate p within the set range 1 , random missing rate p 2 and the missing non-random rate p 3 , and the sum of the completely random missing rate, random missing rate and non-random missing rate meets the upper limit of the data missing rate;

[0054] In the embodiment, the setting range is [0, 20%], the upper limit of the data missing rate is no more than 60%, and the sum of the completely random missing rate, the random missing rate and the non-random missing rate satisfies the following relationship:

[0055] p 1 +p 2 +p 3 ≤60%

[0056] Step 1.2: Using all the historical monitoring data of the energy storage power station, according to the probability model Generate completely random missing data, where Y is all the historical monitoring data of the energy storage power station, Missing completely random data were generated;

[0057] According to the method proposed by the present invention, the probability of completely random missing data in all historical monitoring data is is the missing completely at random rate p 1 , indicating that the generated completely random missing data is independent of any observed or unobserved data, and each data monitoring point has the same probability of data missing. Compared with the prior art which has not considered the existence of hidden dangers of data missing introduced by the data monitoring device, the present invention can more accurately simulate the data missing situation introduced by the data monitoring device in the energy storage power station, and further improve the comprehensiveness and reliability of the monitoring data of the energy storage power station.

[0058] Using all historical monitoring data and actual monitoring data of the energy storage power station, according to the probability model Generate randomly missing data where is the randomly missing data generated, Z is the actual monitoring data, and p 2 (Z) is the random missing rate of actual monitoring data;

[0059] According to the method proposed by the present invention, the probability of random missing data in all historical monitoring data and actual monitoring data is The random missing rate p of the actual monitoring data 2(Z) is the same, indicating that the generated randomly missing data depends on the actual monitoring data. The larger the amount of actual monitoring data, the larger the amount of randomly missing data. This provides a large number of samples for subsequent data analysis and avoids the accuracy of data analysis being affected by a small number of samples. At the same time, the randomly missing data are distributed in the actual monitoring data to dynamically and adaptively increase the number of data samples, and also avoid affecting the efficiency of data analysis when the total amount of data increases.

[0060] Using all the historical monitoring data of the energy storage power station, according to the probability model Generate nonrandom missing data where, For the generated non-random missing data, p 3 (Y) is the non-random missing rate of all historical monitoring data;

[0061] According to the method proposed by the present invention, the probability of non-random missing data in all historical monitoring data is The non-random missing rate is the same as the non-random missing rate of all historical monitoring data, which means that the non-random missing data produced based on all historical monitoring data is unobserved data. The generated non-random data is used to make up for the unobserved data and improve the coverage rate of the monitoring data of the energy storage power station.

[0062] In the embodiment, the random missing data is fitted using a linear or nonlinear function of the actual monitoring data, and the non-random data is fitted using a linear or nonlinear function of the historical monitoring data.

[0063] The prior art generally only considers a single type of data missing, while the present invention considers the situation where data missing under multiple different conditions exist simultaneously, which is more in line with the actual missing situation of monitoring data of energy storage power stations.

[0064] Step 1.3, using the completely random missing rate, the random missing rate of actual monitoring data, and the non-random missing rate of all historical monitoring data, determine the data missing rate of each group of monitoring data using the following relationship:

[0065] P=p 1 +(1-p 1 ) 2 (Z)+(1-p 1 )(1-p 2 (Z))p 3 (Y)

[0066] Where P is the data missing rate of each group of monitoring data.

[0067] Step 1.4, looping through steps 1.1 to 1.3, and when the set number of monitoring data sets is reached, the loop execution ends, and multiple sets of monitoring data with different data missing rates are obtained;

[0068] In the embodiment, As the data missing identifier of the ith data monitoring point in the mth group of monitoring data, When , it means that the i-th data monitoring point is missing in the m-th group of monitoring data, so the position of the i-th data monitoring point in the m-th group of monitoring data is set to 0, otherwise It means that the data of the i-th data monitoring point is not missing in the m-th group of monitoring data.

[0069] Each set of monitoring data includes not only completely random missing data, random missing data, and non-random missing data, but also actual monitoring data, such as Figure 2 As shown, the capacity of the energy storage power station is monitored, and a monitoring data set is obtained according to step 1. In the figure, the solid line represents the actual monitoring data, and the dotted line represents the missing data. The non-uniform spacing of the missing data represents different data missing rates and different data missing types, thereby more accurately simulating the situation of missing monitoring data occurring in an actual energy storage power station.

[0070] The method proposed in the present invention is based on the existing historical monitoring data of energy storage power stations, fully considers the actual occurrence of missing data, adds completely random missing data, random missing data and non-random missing data according to the missing ratio, and constructs a data set for model training, thereby ensuring that the asymmetric denoising autoencoder can learn as many data missing rules as possible and improve the generalization performance of the network.

[0071] Step 2: Based on the network structure of the denoising autoencoder, an asymmetric denoising autoencoder model is established. The asymmetric denoising autoencoder model takes the monitoring dataset as input and a preliminary completed dataset of the monitoring dataset as output.

[0072] Specifically, the asymmetric denoising autoencoder model includes: an input layer, an encoder, a skip connection module, a decoder and an output layer; the monitoring data set is input into the encoder through the input layer, the encoder extracts the periodic features and time series features of the monitoring data set, and the skip connection module extracts the element features of the monitoring data set, wherein the periodic features and time series features are both first-dimensional matrices, and the element features are second-dimensional matrices, the decoder fuses the periodic features, time series features and element features, and outputs a feature fused data set; the feature fused data set obtains a preliminary completed data set through the output layer.

[0073] In the embodiment, the first dimensional matrix is ​​a matrix of two dimensions or above, and the second dimensional matrix is ​​a one-dimensional matrix.

[0074] Specifically, Figure 3The structure of the asymmetric denoising autoencoder model shown in the figure, the input data is a monitoring data set with mixed missing features, the monitoring data set is input into the asymmetric denoising autoencoder model through the input layer, the input layer processes the monitoring data x into an input data matrix x′ for subsequent calculations, wherein the single-dimensional time series data in the monitoring data set is converted into a data matrix form according to the acquisition frequency; the encoder includes a periodic feature extraction module and a time series feature extraction module, in an embodiment, the periodic feature extraction module includes a plurality of convolutional layers connected in series, the time series feature extraction module includes a plurality of gated cyclic units connected in series, and the periodic feature extraction module outputs a periodic feature matrix y under a first set dimension 1 The temporal feature extraction module outputs the temporal feature matrix y under the first set dimension 2 ; The skip connection module is constructed by the global average pooling method, and the skip connection module outputs the feature matrix s under the second set dimension; the decoder is constructed by multi-layer fully connected layers, and the periodic feature matrix y under the first set dimension 1 , the time series feature matrix y under the first set dimension 2 The feature matrix s under the second set dimension is input into the decoder for fusion, and the decoder outputs the feature fusion matrix h; the feature fusion matrix h passes through the output layer to obtain the preliminary completed data matrix z.

[0075] Specifically, in the periodic feature extraction module, the sizes of each convolution layer are different. Multi-size convolution kernel sliding convolution is used and channel weights are shared to obtain the horizontal time series features and vertical periodic features of the data in the input data matrix x′, thereby efficiently extracting the first dimension periodic feature matrix y 1 , the process is shown as follows:

[0076] y 1 =f(x′)=f Relu (f Conv (x′))

[0077] Where f(x′) is the function for extracting the periodic features of the input data matrix x′, and f Relu () is the Relu activation function, f Conv () is the convolution operation function.

[0078] In the embodiment, Figure 4 As shown in Figure 1, the period feature extraction module is constructed by stacking multiple convolutional layers of different scales. Figure 4 The horizontal arrows in the figure represent the convolution operation with a stride of 1 using a 1×1 convolution kernel. The feature map size is not reduced but the number of channels is changed. Through continuous 1×1 convolutions in feature maps of sizes such as 48×48, 24×24, and 13×13, cross-channel features can be efficiently fused with only a relatively low number of parameters. Figure 4The vertical arrow in the figure represents the convolution operation with a step size of 2 using a 4×4 or 2×2 convolution kernel. The feature map size is reduced to half of the original size without changing the number of feature map channels. The large-size convolution kernel can effectively remove the noise caused by data loss during the feature map reduction process. Finally, a 4×6×6 high-dimensional periodic feature matrix y is extracted through multi-layer convolution. 1 , and resize it into a one-dimensional feature vector of 144×1 for decoder data reconstruction.

[0079] In the embodiment, Figure 5 As shown in the timing feature extraction module, each gated recurrent unit includes an update gate R and a reset gate Z. The gated recurrent unit can effectively capture the timing data H t-1 With H t The calculation process shown in the following relationship is used to gate the input data matrix x′, selectively update and forget information, so as to better handle the long-term dependencies in the time series data and extract a high-dimensional time series feature matrix y with a size of 48×1 2 :

[0080] Z t =σ(W z ·[H t-1 ,x t ′])

[0081] R t =σ(W r ·[H t-1 ,x t ′])

[0082] A t =f tanh (W h ·[R t *H t-1 ,x′ t ])

[0083] y 2t =H t =(1-Z t )*H t-1 +Z t *A t

[0084] In the formula, Z t is the value of the update gate at time t, R t is the value of the update gate at time t, W z is the weight matrix of the reset gate, Wr is the weight matrix of the update gate, W h is the candidate activation weight matrix, f tanh () is the tanh activation function, H t-1 and H tare the hidden states at time t-1 and time t, respectively, x t ′ is the input data matrix at time t, A t is the candidate activation value at time t, σ is the activation function of the gated recurrent unit, y 2t is the high-dimensional time series feature matrix at time t.

[0085] In the embodiment, Figure 4 As shown in Figure 1, the skip connection module constructed using the global average pooling method includes multiple global average pooling layers. As the depth of the convolutional layer increases, the size of the feature map gradually decreases. The encoder effectively focuses on high-dimensional features, but some low-dimensional features are discarded due to their weak representation ability. To further enrich the types of features, the following is taken before each feature map reduction in the encoder convolutional layer: Figure 4 The global average pooling operation shown by the horizontal arrow in the middle is to find the average value of each channel of the m×i×j feature map to obtain an m×1×1 low-dimensional feature matrix s, which satisfies the following relationship:

[0086]

[0087] In the formula, s m is the element feature matrix in the mth channel, is the total number of elements in the rectangular region of the feature map with length i and width j in the mth channel, x′ pq Represents the element at (p,q) in the input data rectangular area, R ij is the number of elements in the rectangular region of the feature map with length i and width j respectively.

[0088] In the embodiment, Figure 4 As shown by the right arrow in , the decoder uses multiple fully connected layers to gradually expand the vector length, reconstruct missing data and retain the continuity of the reconstructed data. At the same time, the low-dimensional feature matrices s with sizes of 7×1×1, 13×1×1, 24×1×1, and 48×1×1 are successively integrated into the corresponding fully connected layers of the decoder. The input of each fully connected layer is composed of the concatenation of the corresponding low-dimensional element feature matrix s and the output of the previous fully connected layer decoding. The decoder uses a multi-scale fully connected layer to combine the low-dimensional element feature matrix s and the high-dimensional periodic feature matrix y 1 and high-dimensional time series feature matrix y 2 They are fused in sequence to form a feature fusion matrix h, which satisfies the following relationship:

[0089] h=g(y 1 ,y 2 ,s)=f Rule (f FC (y 1 ,y 2 ,s))=f Rule (W'[y1 ,y 2 ,s]+b')

[0090] In the formula, f Relu () is the Relu activation function, f FC () is the activation function of the fully connected layer, W′ is the weight matrix of the fully connected layer, and b′ is the bias matrix of the fully connected layer.

[0091] The output layer consists of a single fully connected layer, which transforms the feature fusion matrix h into the preliminary completed data matrix z.

[0092] Step 3, when training the asymmetric denoising autoencoder model, all historical monitoring data are used as the benchmark output. Based on the back propagation method, the loss of the monitoring data set is first calculated. When the loss of the monitoring data set is not greater than the first set threshold, the loss of all missing data is calculated. When the sum of the loss of the monitoring data set and the loss of all missing data is not greater than the second set threshold, the model training is terminated; the trained model is used to output a preliminary completed data set of the monitoring data set of the energy storage power station.

[0093] All historical monitoring data are non-missing monitoring data, so they are used as the benchmark output when training the asymmetric denoising autoencoder model;

[0094] Based on the back propagation method, the double cosine logarithmic loss function is used to calculate the loss of the monitoring data set, satisfying the following relationship:

[0095]

[0096] Where, L 1 To monitor the loss of the dataset, is the data error of the ith data monitoring point in the mth group of monitoring data, which is the difference between the data of the ith data monitoring point in the mth group of monitoring data and the corresponding historical detection data, N 1 is the total number of data monitoring points, N 2 is the total number of monitoring data groups in the monitoring data set.

[0097] The first threshold is set to [e-ln2,0.5e 2 ].

[0098] In the embodiment, when When it is small, the value of the double cosine logarithmic loss function is approximately 0.5e 2 , as the upper limit of the first set threshold, when When it is larger, the value of the double cosine logarithmic loss function is approximately e-ln2. As the lower limit of the first set threshold, the use of the double cosine logarithmic loss function to calculate the loss of the monitoring data set will not cause error amplification, and the second order is differentiable everywhere, which can accelerate the convergence of the model. Therefore, when the loss of the monitoring data set is not greater than the first set threshold, the loss of all missing data is calculated, which ensures the convergence of the model while avoiding error amplification.

[0099] Based on the back propagation method, the MSELoss function is used to calculate the loss of all missing data, satisfying the following relationship:

[0100]

[0101] Where, L 2 is the loss for all missing data, is the data error of the ith data monitoring point in the mth group of monitoring data, As the data missing identifier of the ith data monitoring point in the mth set of monitoring data.

[0102] In the embodiment, the MSELoss function is used to calculate the loss of all missing data. With the help of the error amplification characteristics of the MSELoss function, the asymmetric denoising autoencoder model can more effectively focus on the missing data completion effect. Moreover, since the model convergence has been ensured by monitoring that the loss of the data set is not greater than the first set threshold, amplifying the error of all missing data at this time will not affect the convergence of the model. In addition, the data missing identifier is introduced into the MSELoss function calculation, which effectively improves the calculation efficiency and facilitates the calculation through programming.

[0103] To monitor the loss L of the dataset 1 The loss L with all missing data 2 The sum is taken as the total loss of the training model. While controlling the overall running trend of the data, the focus is on the missing data. When the total loss of the training model is less than the second set threshold, the model not only converges quickly but also has the best missing data completion effect, so the model training is terminated. In the embodiment, the second set threshold is 1.85.

[0104] The trained asymmetric denoising autoencoder model is used to preliminarily complete the monitoring dataset to obtain a preliminary completed dataset.

[0105] Step 4: Use a sliding average filter to filter the preliminary completed data set z to obtain the completed data set z′.

[0106] After obtaining the preliminary completed data set, a sliding average filter with a set filter window length is used to perform secondary filtering optimization on the data at the completed position to further improve the completion accuracy. The obtained preliminary completed data is optimized by adding a window to the completed missing data and calculating the mean within the window to suppress abnormal fluctuations in the preliminary completed data. The sliding average filtering process is shown in the following formula:

[0107]

[0108] In the formula, is the first Initial completion data, is the nth complementary data obtained after sliding average filtering, k is the filter window length, which is a positive odd number, forming a symmetrical window to prevent phase deviation.

[0109] In the embodiment, a sliding average filter is used to filter the preliminary completed data set to obtain a completed data set such as Figure 6 As shown in the figure, the "21" in the monitoring data set x is processed as the missing data "nan" in the input data matrix x′, and the filter window length is set to k = 7. "26" is the completed data generated by the asymmetric autoencoder, and 21.9 is the completed data obtained after the sliding average filter is optimized, which is closer to the actual monitoring value 21.

[0110] The energy storage power station monitoring data actually collected with more than 60% mixed missing data is completed twice by an asymmetric denoising autoencoder and a sliding average filter to obtain a completed data set.

[0111] Since the data in the actual energy storage power station monitoring data is mixed and missing, the location is unknown and it is easy to be confused with the actual abnormal status data of the power station. For this reason, in the embodiment, the isolation forest method is first used to identify the outliers in the energy storage power station monitoring data. The outliers include: missing values ​​and real abnormal data, among which the missing values ​​should be completed and the real abnormal values ​​should be retained; then the correlation analysis method is used to classify the outliers and determine the specific attributes of the data outliers. The position of the missing value is marked as Pm, and the abnormal status data is retained. According to the missing value position Pm, the corresponding position data is selected from the completed data set to replace the missing value in the actual monitoring data, so as to achieve the effective completion of the missing data of the actual energy storage power station monitoring data. Figure 7 As shown, Figure 7 The middle dotted line is the actual monitoring data, which is implemented as a monitoring data set with missing values. The dotted line is the completed monitoring data obtained by the method proposed in the present invention. It can be seen that the completed monitoring data is close to the real data without missing values, and the accuracy is significantly improved compared with the existing method.

[0112] The present invention also proposes a system for completing missing data of energy storage power station monitoring based on feature fusion, comprising:

[0113] The data processing module is used to obtain all historical monitoring data and actual monitoring data of the energy storage power station, and generate multiple groups of monitoring data with different data missing rates based on the probability model. Each group of monitoring data includes completely random missing data, random missing data, non-random missing data and actual monitoring data. Multiple groups of monitoring data with different data missing rates together constitute the monitoring data set of the energy storage power station.

[0114] The data completion module is used to establish an asymmetric denoising autoencoder model based on the network structure of the denoising autoencoder. The asymmetric denoising autoencoder model takes the monitoring data set as input and the preliminary completed data set of the monitoring data set as output. When training the asymmetric denoising autoencoder model, all historical monitoring data are used as the benchmark output. Based on the back propagation method, the loss of the monitoring data set is first calculated. When the loss of the monitoring data set is not greater than the first set threshold, the loss of all missing data is calculated. When the sum of the loss of the monitoring data set and the loss of all missing data is not greater than the second set threshold, the model training is terminated. The trained model is used to output the preliminary completed data set of the monitoring data set of the energy storage power station. The preliminary completed data set is filtered by a sliding average filter to obtain a completed data set.

[0115] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0116] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0117] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0118] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for completing missing data of energy storage power station monitoring based on feature fusion, characterized in that: include: Obtain all historical monitoring data and actual monitoring data of the energy storage power station, and generate multiple sets of monitoring data with different data missing rates based on the probability model. Each set of monitoring data includes completely random missing data, random missing data, non-random missing data, and actual monitoring data. A monitoring data set of an energy storage power station is composed of multiple sets of monitoring data with different data missing rates; Based on the network structure of the denoising autoencoder, an asymmetric denoising autoencoder model is established. The asymmetric denoising autoencoder model takes the monitoring dataset as input and the preliminary complement dataset of the monitoring dataset as output. When training the asymmetric denoising autoencoder model, all historical monitoring data are used as the benchmark output. Based on the back propagation method, the loss of the monitoring data set is first calculated. When the loss of the monitoring data set is not greater than the first set threshold, the loss of all missing data is calculated. When the sum of the loss of the monitoring data set and the loss of all missing data is not greater than the second set threshold, the model training is terminated; the trained model is used to output the preliminary completion data set of the monitoring data set of the energy storage power station; The preliminary completed data set is filtered using a sliding average filter to obtain the completed data set.

2. The method for completing missing data of energy storage power station monitoring based on feature fusion according to claim 1 is characterized in that: Obtain all historical monitoring data and actual monitoring data of the energy storage power station, and generate multiple sets of monitoring data with different data missing rates based on the probability model, including: Randomly generate the completely random missing rate p1, random missing rate p2 and non-random missing rate p3 within the set range, and the sum of the completely random missing rate, random missing rate and non-random missing rate meets the upper limit of the data missing rate; Using all the historical monitoring data of the energy storage power station, according to the probability model Generate completely random missing data, where Y is all the historical monitoring data of the energy storage power station, Missing completely random data were generated; Using all historical monitoring data and actual monitoring data of the energy storage power station, according to the probability model Generate randomly missing data where is the randomly missing data generated, Z is the actual monitoring data, and p2(Z) is the random missing rate of the actual monitoring data; Using all the historical monitoring data of the energy storage power station, according to the probability model Generate nonrandom missing data where, is the generated non-random missing data, and p3(Y) is the non-random missing rate of all historical monitoring data; Using the completely random missing rate, the random missing rate of actual monitoring data, and the non-random missing rate of all historical monitoring data, the data missing rate of each group of monitoring data is determined using the following relationship: P=p1+(1-p1)p2(Z)+(1-p1)(1-p2(Z))p3(Y) Where P is the data missing rate of each group of monitoring data.

3. The method for completing missing data of energy storage power station monitoring based on feature fusion according to claim 2 is characterized in that: The setting range is [0, 20%], and the upper limit of the data missing rate is no more than 60%.

4. The method for completing missing data of energy storage power station monitoring based on feature fusion according to claim 2 is characterized in that: by As the data missing identifier of the i-th data monitoring point in the m-th group of monitoring data, when When , it means that the i-th data monitoring point is missing in the m-th group of monitoring data, so the position of the i-th data monitoring point in the m-th group of monitoring data is set to 0, otherwise It means that the data of the i-th data monitoring point in the m-th group of monitoring data is not missing, i = 1, 2, ..., N1, m = 1, 2, ..., N2, N1 is the total number of data monitoring points, and N2 is the total number of monitoring data groups in the monitoring data set.

5. The method for completing missing data of energy storage power station monitoring based on feature fusion according to claim 1 is characterized in that: The asymmetric denoising autoencoder model includes: an input layer, an encoder, a skip connection module, a decoder and an output layer; the monitoring data set is input into the encoder through the input layer, the encoder extracts the periodic features and time series features of the monitoring data set, and the skip connection module extracts the element features of the monitoring data set, wherein the periodic features and time series features are both first-dimensional matrices, and the element features are second-dimensional matrices; the decoder fuses the periodic features, time series features and element features, and outputs a feature fused data set; the feature fused data set is initially completed through the output layer.

6. The method for completing missing data of energy storage power station monitoring based on feature fusion according to claim 1 is characterized in that: Based on the back propagation method, the double cosine logarithmic loss function is used to calculate the loss of the monitoring data set, satisfying the following relationship: Where L1 is the loss of the monitoring data set, is the data error of the ith data monitoring point in the mth group of monitoring data, is the difference between the data of the ith data monitoring point in the mth group of monitoring data and the corresponding historical detection data, N1 is the total number of data monitoring points, and N2 is the total number of monitoring data groups in the monitoring data set; The first threshold is set to [e-ln2,0.5e 2 ].

7. The method for completing missing data of energy storage power station monitoring based on feature fusion according to claim 6 is characterized in that: Based on the back propagation method, the MSELoss function is used to calculate the loss of all missing data, satisfying the following relationship: Where L2 is the loss of all missing data, is the data error of the ith data monitoring point in the mth group of monitoring data, As the data missing identifier of the i-th data monitoring point in the m-th group of monitoring data, when When , it means that the i-th data monitoring point is missing in the m-th group of monitoring data, so the position of the i-th data monitoring point in the m-th group of monitoring data is set to 0, otherwise It means that the data of the i-th data monitoring point is not missing in the m-th group of monitoring data.

8. The method for completing missing data of energy storage power station monitoring based on feature fusion according to claim 7 is characterized in that: The sum of the loss L1 of the monitoring data set and the loss L2 of all missing data is taken as the total loss of the training model. When the total loss of the training model is less than the second set threshold, the model training is terminated; wherein the second set threshold is 1.

85.

9. The method for completing missing data of energy storage power station monitoring based on feature fusion according to claim 1 is characterized in that: The sliding average filtering process is shown as follows: In the formula, is the first Initial completion data, is the nth completed data obtained after sliding average filtering, k is the filter window length, which is a positive odd number.

10. A system for completing missing data of energy storage power station monitoring based on feature fusion, characterized in that: include: The data processing module is used to obtain all historical monitoring data and actual monitoring data of the energy storage power station, and generate multiple groups of monitoring data with different data missing rates based on the probability model. Each group of monitoring data includes completely random missing data, random missing data, non-random missing data and actual monitoring data. A monitoring data set of an energy storage power station is composed of multiple sets of monitoring data with different data missing rates; The data completion module is used to establish an asymmetric denoising autoencoder model based on the network structure of the denoising autoencoder. The asymmetric denoising autoencoder model takes the monitoring data set as input and the preliminary completed data set of the monitoring data set as output. When training the asymmetric denoising autoencoder model, all historical monitoring data are used as the benchmark output. Based on the back propagation method, the loss of the monitoring data set is first calculated. When the loss of the monitoring data set is not greater than the first set threshold, the loss of all missing data is calculated. When the sum of the loss of the monitoring data set and the loss of all missing data is not greater than the second set threshold, the model training is terminated. The trained model is used to output the preliminary completed data set of the monitoring data set of the energy storage power station. The preliminary completed data set is filtered by a sliding average filter to obtain a completed data set.

Images