A multi-scale bilateral GRU measurement error prediction method, system and device

Abstract

A multi-scale bilateral GRU method, system, and device for predicting metering errors are disclosed. This method acquires historical interaction data during the charging process of a DC charging pile and generates sequences at different time scales according to different sampling frequencies. Mapping features of each scale sequence are extracted and subjected to STL feature decomposition to obtain the decomposed trend term, seasonal term, and remainder term. A novel bilateral GRU network is constructed, and the trend term, seasonal term, and remainder term from the feature decomposition of each scale sequence are used for temporal feature extraction within the bilateral GRU network to obtain the metering error prediction results for each scale sequence. Finally, the metering error prediction results of each scale sequence are scaled down and integrated to obtain the metering error prediction result for the DC charging pile. This invention constructs a novel bilateral GRU network, selectively extracts temporal features for each scale sequence, and effectively integrates metering error prediction results from different time scales by continuously scaling down the large-scale metering error prediction results, thereby achieving accurate online prediction of charging pile metering errors.

Description

Technical Field

[0001] This invention belongs to the field of online power metering and monitoring, specifically relating to a method, system, and equipment for predicting metering errors using a multi-scale bilateral GRU. Background Technology

[0002] Charging piles, as key equipment for electric vehicle energy replenishment, primarily function to provide DC or AC power for charging electric vehicles. The metering system, as a core component of the charging pile, directly affects the user's charging costs and the energy replenishment efficiency of the electric vehicle. However, in actual operation, factors such as equipment aging, sensor failure, environmental interference, and software algorithm errors can all lead to discrepancies between the metering results and the actual charging volume, thus impacting user interests and potentially causing trade disputes.

[0003] To ensure the accuracy of charging pile metering, the metering system of charging piles needs to be calibrated regularly. However, traditional calibration methods require the charging piles to be offline, which not only affects the normal use of the charging piles but also leads to significant economic losses. Therefore, traditional offline calibration methods can no longer meet the requirements, and the research and application of charging pile metering error prediction algorithms is a key technical challenge that urgently needs to be solved. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems in the prior art by providing a method, system, and device for predicting metrological errors using a multi-scale bilateral GRU.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, this invention proposes a metrological error prediction method for a multi-scale bilateral GRU, comprising:

[0007] S1. Obtain historical interaction data during the charging process of DC charging piles, and generate sequences of different time scales from the historical interaction data according to different sampling frequencies;

[0008] S2. Extract the mapping features of each scale sequence, and perform STL feature decomposition on the extracted mapping features to obtain the trend term, seasonal term and remainder term after feature decomposition of each scale sequence.

[0009] S3. Construct a new bilateral GRU network, and feed the trend term, seasonal term and residual term after the feature decomposition of each scale sequence into the bilateral GRU network for time series feature extraction to obtain the measurement error prediction results of each scale sequence.

[0010] S4. The measurement error prediction results of each scale sequence are scaled down and integrated to obtain the measurement error prediction results of DC charging piles.

[0011] In S3, the new bilateral GRU network includes a gating unit and a linear prediction unit, and the specific steps include:

[0012] S31. In the gating unit of the new bilateral GRU network, the reset gate and update gate are used to extract the temporal features of the trend and seasonal terms after the feature decomposition of each scale sequence, and the output results of the trend and seasonal terms are obtained:

[0013]

[0014] In the above formula, This represents the hidden state of the seasonal term at time t, i.e., the output of the seasonal term. The update gate that the seasonal item passes through. Let be the hidden state of the seasonal term at time t-1. Let be the candidate hidden state of the seasonal term at time t, and σ(.) be the Sigmoid function. Update the gate weight matrix for the seasonal terms. Let be the temporal eigenvector corresponding to the seasonal term at time t, and tanh(.) be the hyperbolic tangent function. Here is the weight matrix for the candidate hidden states of the seasonal item. The reset gate that the seasonal item passes through. Reset the weight matrix of the gate for the seasonal term. This represents the hidden state of the trend term at time t, i.e., the output result of the trend term. The update gate that the trend item passes through. Let be the hidden state of the trend term at time t-1. Let be the candidate hidden state of the trend term at time t. Update the gate weight matrix for the trend term. Let be the time series feature vector corresponding to the trend term at time t. The weight matrix represents the candidate hidden states of the trend term. The reset gate that the trend item passes through. Reset the weight matrix of the gate for the trend term;

[0015] In the linear prediction unit of the new bilateral GRU network, the following formula is used to extract the temporal features of the remainder after feature decomposition, and the output of the remainder is obtained:

[0016]

[0017] In the above formula, The output of the remainder term, where α is the linear smoothing coefficient. Let be the temporal feature vector corresponding to the remainder term at time t. λ is the initial time-series feature vector of the remainder term, and θ are both coefficients of the exponential model.

[0018] S32. By superimposing the outputs of the trend term, seasonal term, and residual term, the measurement error prediction results for each scale series are obtained:

[0019]

[0020] In the above formula, Y m The measurement error prediction result for the m-th scale sequence is given, where T is the total time. This represents the measurement error prediction result for the m-th scale sequence at time t.

[0021] S4 includes:

[0022] S41. Arrange the measurement error prediction results of each scale sequence in descending order of sampling frequency, i.e., in ascending order of time scale, as {Y0, ..., Y...}. m , ..., Y M};

[0023] S42, Y M Perform dilated convolution and mean pooling operations to make Y M The timescale is reduced to that of Y M-1 At the same scale, the scale-reduced econometric error prediction results AverPool(Dconv) are obtained. M (Y M In this context, AverPool(.) is the mean pooling operation, and Dconv is the mean pooling operation. M(.) This is an expanded convolution operation at the Mth scale, and the scale-reduced econometric error prediction result AverPool(Dconv) is used. M (Y M )) and Y M-1 Summing yields the summed measurement error prediction result;

[0024] S43. Perform dilated convolution and mean pooling operations on the summed measurement error prediction results to reduce them to a value similar to Y. M-2 Same scale, and with Y M-2 Summation is performed, and this process is repeated until the econometric error prediction results for all scale sequences are integrated.

[0025] S2 includes:

[0026] S21. Use the following formula to extract the mapping features of each scale sequence in the multi-scale sequence dataset:

[0027]

[0028] In the above formula, χm Let X be the sequence at the m-th scale. m The mapping features, For multi-scale sequence sets,

[0029] S22. Use the following formula to perform STL feature decomposition on the mapping features of each scale sequence to obtain the trend term, seasonal term, and remainder term after feature decomposition for each scale sequence:

[0030] STL(χ m )=(L m S m P m );

[0031]

[0032] In the above formula, L m S is the trend term. m For seasonal terms, P m For the remaining items, Let be the time series feature vector corresponding to the trend term at time t. Let be the temporal feature vector corresponding to the seasonal term at time t. Let be the temporal feature vector corresponding to the remainder term at time t.

[0033] Secondly, this invention proposes a multi-scale bilateral GRU measurement error prediction system, including a time-scale sequence generation module, a feature extraction and decomposition module, a bilateral GRU network construction module, and a result integration module;

[0034] The time-scale sequence generation module is used to acquire historical interaction data during the charging process of DC charging piles, and generate sequences of different time scales from the historical interaction data according to different sampling frequencies;

[0035] The feature extraction and decomposition module is used to extract the mapping features of the sequence at each scale, and to perform STL feature decomposition on the extracted mapping features to obtain the trend term, seasonal term and remainder term after feature decomposition of the sequence at each scale.

[0036] The bilateral GRU network construction module is used to construct a new bilateral GRU network, which feeds the trend term, seasonal term and residual term after the feature decomposition of each scale sequence into the bilateral GRU network for time series feature extraction, and obtains the measurement error prediction results of each scale sequence.

[0037] The result integration module is used to scale down the measurement error prediction results of each scale sequence and integrate them to obtain the measurement error prediction results of DC charging piles.

[0038] The bilateral GRU network construction module includes a result output unit and an output result overlay unit;

[0039] The result output unit is used in the gating unit of the new bilateral GRU network to extract the temporal features of the trend and seasonal terms after feature decomposition at each scale using a reset gate and an update gate, and to obtain the output results of the trend and seasonal terms. In the linear prediction unit of the new bilateral GRU network, the unit extracts the temporal features of the remaining terms after feature decomposition, and to obtain the output results of the remaining terms.

[0040]

[0041]

[0042] In the above formula, This represents the hidden state of the seasonal term at time t, i.e., the output of the seasonal term. The update gate that the seasonal item passes through. Let be the hidden state of the seasonal term at time t-1. Let be the candidate hidden state of the seasonal term at time t, and σ(.) be the Sigmoid function. Update the gate weight matrix for the seasonal terms. Let be the temporal eigenvector corresponding to the seasonal term at time t, and tanh(.) be the hyperbolic tangent function. Here is the weight matrix for the candidate hidden states of the seasonal item. The reset gate that the seasonal item passes through. Reset the weight matrix of the gate for the seasonal term. This represents the hidden state of the trend term at time t, i.e., the output result of the trend term. The update gate that the trend item passes through. Let be the hidden state of the trend term at time t-1. Let be the candidate hidden state of the trend term at time t. Update the gate weight matrix for the trend term. Let be the time series feature vector corresponding to the trend term at time t. The weight matrix represents the candidate hidden states of the trend term. The reset gate that the trend item passes through. Reset the weight matrix of the gate for the trend term. The output of the remainder term, where α is the linear smoothing coefficient. Let be the temporal feature vector corresponding to the remainder term at time t. λ is the initial time-series feature vector of the remainder term, and θ are both coefficients of the exponential model.

[0043] The output result overlay unit is used to overlay the output results of the trend term, seasonal term, and remainder term to obtain the measurement error prediction results for each scale series:

[0044]

[0045]

[0046] In the above formula, Y m The measurement error prediction result for the m-th scale sequence is given, where T is the total time. This represents the measurement error prediction result for the m-th scale sequence at time t.

[0047] The result integration module includes a sequence arrangement unit, a scaling unit, and a cyclic repetition unit;

[0048] The sequence arrangement unit is used to arrange the measurement error prediction results of each scale sequence in descending order of sampling frequency, i.e., in ascending order of time scale, as {Y0, ..., Y...}. m , ..., Y M};

[0049] The scaling unit is used for Y M Perform dilated convolution and mean pooling operations to make Y M The timescale is reduced to that of Y M-1 At the same scale, the scale-reduced econometric error prediction results AverPool(Dconv) are obtained. M (Y M In this context, AverPool(.) is the mean pooling operation, and Dconv is the mean pooling operation. M (.) represents the dilated convolution operation at the M-th scale, and the scaled-down econometric error prediction result AverPool(Dconv) is used to convert the scaled-down econometric error prediction result into an AverPool(Dconv) value. M (Y M )) and Y M-1 Summation yields the summed measurement error prediction results.

[0050] The recurring unit is used to perform dilated convolution and mean pooling operations on the summed measurement error prediction results, reducing them to a value similar to Y. M-2 Same scale, and with Y M-2 Summation is performed, and this process is repeated until the econometric error prediction results for all scale sequences are integrated.

[0051] The feature extraction and decomposition module includes a mapping feature extraction unit and a feature decomposition unit;

[0052] The mapping feature extraction unit is used to extract the mapping features of each scale sequence in the multi-scale sequence dataset using the following formula:

[0053]

[0054] In the above formula, χ mLet Xm be the mapping feature of the sequence at the m-th scale. For multi-scale sequence sets,

[0055] The feature decomposition unit is used to perform STL feature decomposition on the mapping features of each scale sequence using the following formula, to obtain the trend term, seasonal term, and remainder term after feature decomposition for each scale sequence:

[0056] STL(χ n )=(L m S m P m );

[0057]

[0058]

[0059] In the above formula, L m S is the trend term. m For seasonal terms, P m For the remaining items, Let be the time series feature vector corresponding to the trend term at time t. Let be the temporal feature vector corresponding to the seasonal term at time t. Let be the temporal feature vector corresponding to the remainder term at time t.

[0060] Thirdly, the present invention proposes a measurement error prediction device for a multi-scale bilateral GRU, including a processor and a memory;

[0061] The memory is used to store computer program code and to transmit the computer program code to the processor;

[0062] The processor is used to execute the aforementioned multi-scale bilateral GRU measurement error prediction method according to the instructions in the computer program code.

[0063] Fourthly, the present invention provides a computer storage medium on which a computer program is stored;

[0064] When the computer program is executed by the processor, it implements the steps of the aforementioned multi-scale bilateral GRU measurement error prediction method.

[0065] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0066] 1. This invention proposes a multi-scale bilateral GRU method, system, and device for predicting measurement errors. The method first acquires historical interaction data during the charging process of a DC charging pile and generates sequences at different time scales based on different sampling frequencies. It then extracts the mapping features of each scale sequence and performs STL feature decomposition on these features to obtain the trend term, seasonal term, and remainder term after feature decomposition. Next, a novel bilateral GRU network is constructed, and the trend term, seasonal term, and remainder term from each scale sequence feature decomposition are fed into the bilateral GRU network for temporal feature extraction, yielding the measurement error prediction results for each scale sequence. Finally, the measurement error prediction results for each scale sequence are scaled down and integrated to obtain the measurement error prediction result for the DC charging pile. This method constructs a novel bilateral GRU network, specifically extracting the temporal features of the three decomposition terms for each scale sequence. Through multi-dimensional feature extraction, the output of the bilateral GRU network better matches the hidden state of the current task, thus obtaining more accurate measurement error prediction results.

[0067] 2. This invention proposes a multi-scale bilateral GRU metering error prediction method, system, and device. When integrating the metering error prediction results of DC charging piles, this method continuously reduces the scale of large-scale metering error prediction results through dilated convolution and mean pooling operations, effectively integrating metering error prediction results of different time scales, better simulating the fluctuation of charging pile metering errors in actual operation, and realizing online accurate prediction of charging pile metering errors. Attached Figure Description

[0068] Figure 1 This is an overall flowchart of the method described in this invention.

[0069] Figure 2 This is a schematic diagram of the structure of the new bilateral GRU network described in Example 1.

[0070] Figure 3 This is a flowchart illustrating the integration of the measurement error prediction results described in Example 1.

[0071] Figure 4 This is a structural diagram of the system described in this invention.

[0072] Figure 5 This is a structural diagram of the device described in Example 3. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0074] This invention proposes a multi-scale bilateral GRU method, system, and device for predicting metering errors. It acquires historical interaction data during the charging process of a DC charging pile, including the charging pile's output voltage, current, battery state of charge, charging duration, and cumulative energy indication error. The historical interaction data is then used to generate sequences at different time scales according to different sampling frequencies. Mapping features of each scale sequence are extracted, and these features are decomposed using STL (Standard Traversal Layer) to obtain the trend term, seasonal term, and remainder term after feature decomposition for each scale sequence. A novel bilateral GRU network is constructed to selectively extract the temporal features of these three decomposition terms for each scale sequence, yielding the metering error prediction results for each scale sequence. Finally, the metering error prediction results for each scale sequence are scaled down using dilated convolution and mean pooling operations, and then integrated to obtain the metering error prediction results for the DC charging pile, achieving accurate online prediction of charging pile metering errors.

[0075] Example 1:

[0076] like Figure 1 As shown, a method for predicting measurement errors using a multi-scale bilateral GRU is performed in the following steps:

[0077] 1. Acquire historical interaction data during the charging process of DC charging piles, and generate sequences of different time scales from the historical interaction data according to different sampling frequencies;

[0078] Historical interactive data during the charging process of a DC charging pile includes the charging pile output voltage U, current I, battery state of charge S, charging time T, and cumulative energy indication error Ep.

[0079] Since the units and orders of magnitude of various types of historical interaction data are different, the following formula is used to standardize the acquired data:

[0080]

[0081] In the above formula, Here, z represents the standardized data value, μ represents the original data value, and δ represents the standard deviation of the original dataset.

[0082] Define the standardized historical interaction data sequence X∈R T×C Where R is the set of real numbers, T is the length of sequence X, and C is the dimension of sequence X. Based on different sampling frequencies, sequence X is divided into sequences at different time scales using a sliding window averaging method, generating a multi-scale sequence dataset. in m∈{0, ..., M}, where m is the number of scale types, and X0 is the smallest time scale sequence. Its sampling frequency is determined based on the actual acquisition frequency of the acquisition equipment. If the original acquisition data of the equipment is minute-level data, then the smallest time scale sequence can take one data point per minute, containing the most refined time variation data. M For the largest time scale series, it includes the overall trend and macro changes of the series.

[0083] 2. Extract the mapping features of the sequences at each scale, and perform STL feature decomposition on the extracted mapping features to obtain the trend term, seasonal term and remainder term after feature decomposition of the sequences at each scale;

[0084] Multiscale sequence datasets The mapping is performed on the embedding layer for multi-scale deep feature representation. The following formula is used to extract the mapping features of each scale sequence in the multi-scale sequence dataset, so as to facilitate feature extraction in subsequent network structures:

[0085]

[0086] In the above formula, χ m Let X be the sequence at the m-th scale. m The mapping features, For multi-scale sequence sets,

[0087] To reduce the complexity of multi-scale sequences and make the sequence features more intuitive, the following formula is used to perform STL feature decomposition on the mapping features of sequences at each scale, obtaining the trend term, seasonal term, and remainder term after feature decomposition for each scale sequence:

[0088] STL(χ m )=(L m S m P m );

[0089]

[0090] In the above formula, L m S is the trend term. m For seasonal terms, P m For the remaining items, Let be the time series feature vector corresponding to the trend term at time t. Let be the temporal feature vector corresponding to the seasonal term at time t. Let be the temporal feature vector corresponding to the remainder term at time t.

[0091] 3. Construct a new bilateral GRU network, the structure of which is as follows: Figure 2As shown, the trend term, seasonal term and remainder term after feature decomposition at each scale are fed into a bilateral GRU network for time series feature extraction to obtain the measurement error prediction results for each scale sequence.

[0092] The new bilateral GRU network includes gating units and linear prediction units. In the gating units of the new bilateral GRU network, reset gates and update gates are used to extract the temporal features of the trend and seasonal terms after the feature decomposition of each scale sequence, and the output results of the trend and seasonal terms are obtained.

[0093]

[0094]

[0095] In the above formula, This represents the hidden state of the seasonal term at time t, i.e., the output of the seasonal term. The update gate that the seasonal item passes through. Let be the hidden state of the seasonal term at time t-1. Let be the candidate hidden state of the seasonal term at time t, and σ(.) be the Sigmoid function. Update the gate weight matrix for the seasonal terms. Let be the temporal eigenvector corresponding to the seasonal term at time t, and tanh(.) be the hyperbolic tangent function. Here is the weight matrix for the candidate hidden states of the seasonal item. The reset gate that the seasonal item passes through. Reset the weight matrix of the gate for the seasonal term. This represents the hidden state of the trend term at time t, i.e., the output result of the trend term. The update gate that the trend item passes through. Let be the hidden state of the trend term at time t-1. Let be the candidate hidden state of the trend term at time t. Update the gate weight matrix for the trend term. Let be the time series feature vector corresponding to the trend term at time t. This is the weight matrix of the candidate hidden states for the trend term. The reset gate that the trend item passes through. Reset the weight matrix of the gate for the trend term;

[0096] Since the residual information contained in the remainder also carries the local features of the original data, in order to utilize it in the final measurement error prediction result, a linear prediction unit is designed in the new bilateral GRU network to specifically extract the local features in the remainder. The linear prediction unit is different from the nonlinear relationship between the reset gate and the update gate in the gated unit. Instead, it uses a linear smoothing method to directly obtain the output result of the remainder, without considering complex global information, so that the extraction effect of local feature information is better.

[0097] In the linear prediction unit of the new bilateral GRU network, the following formula is used to extract the temporal features of the remainder after feature decomposition, and the output of the remainder is obtained:

[0098]

[0099] In the above formula, The output of the remainder term, where α is the linear smoothing coefficient. Let be the temporal feature vector corresponding to the remainder term at time t. λ is the initial time-series feature vector of the remainder term, and θ are both coefficients of the exponential model.

[0100] The outputs of the trend term, seasonal term, and remainder term are superimposed to obtain the measurement error prediction results for each scale series:

[0101]

[0102] In the above formula, Y m The measurement error prediction result for the m-th scale sequence is given, where T is the total time. This represents the measurement error prediction result for the m-th scale sequence at time t.

[0103] 4. The measurement error prediction results of each scale sequence are scaled down and integrated to obtain the measurement error prediction results of DC charging piles.

[0104] Multiscale sequence datasets The measurement error prediction results for each scale sequence are arranged in descending order of sampling frequency, i.e., in ascending order of time scale, as {Y0, ..., Y...}. m , ..., Y M}, where m is the number of time scale types, Y0 is the measurement error prediction result of the smallest time scale, and Y MThe measurement error prediction results at the largest time scale can reveal the characteristics of different frequencies. For high-frequency small-scale data, it is very sensitive to the instantaneous fluctuations of error, which helps to identify sudden changes in error. Large-scale data is more conducive to the prediction of long-term error trends. For time scale series with large error fluctuations, it is also sufficient to capture the changes in measurement error by adjusting minute-level data to hourly or daily frequency data.

[0105] Integration process such as Figure 3 As shown, firstly, for Y M Perform dilated convolution and mean pooling operations at the corresponding scales to make Y M The scale is reduced to that of Y M-1 At the same scale, the scale-reduced econometric error prediction results AverPool(Dconv) are obtained. M (Y M In this context, AverPool(.) is the mean pooling operation, and Dconv is the mean pooling operation. M(.) This is an expanded convolution operation at the Mth scale, and the scale-reduced econometric error prediction result AverPool(Dconv) is used. M (Y M )) and Y M-1 Summation yields the summed measurement error prediction results.

[0106] Then, the summed measurement error prediction results are subjected to dilated convolution and mean pooling operations to reduce them to a value similar to Y. M-2 Same scale, and with Y M-2 Summation is performed, and this process is repeated until the summed measurement error prediction result is reduced to the same scale as Y0. The summation is then repeated with Y0 to obtain the measurement error prediction result Y of the DC charging pile. out The advantage of doing this is that predictions on a large time scale often focus on the trend of errors, while merging and normalizing the output error values ​​to a smaller time scale will make the predictions more accurate.

[0107] In summary, the following formula is used to calculate the predicted metering error of DC charging piles:

[0108]

[0109] In practice, relevant interactive data during the charging process of the DC charging pile under test is obtained. If the obtained measurement error prediction result is within the accuracy level of the charger (i.e., ±1.0% of level 1), it indicates that the current charging pile is normal. If the measurement error prediction result exceeds this range, it indicates that the current error status of the charging pile is abnormal.

[0110] To verify the effectiveness of the method of this invention, RMSE, MAE, and MAPE were used as evaluation indicators. The deviations between the measurement error prediction results and the actual error values ​​obtained by mainstream time series prediction algorithms, including RNN, LSTM, GRU, and TCN, and the method described in this invention were compared. The comparison of the evaluation indicator results is shown in Table 1.

[0111] Table 1 Evaluation Index Results

[0112]

[0113] The above results show that the method of the present invention is significantly lower than other similar algorithms in terms of RMSE, MAE, and MAPE, indicating that the method of the present invention has the highest accuracy among similar algorithms and the measurement error prediction results obtained in practical applications are more accurate.

[0114] Example 2:

[0115] like Figure 4 As shown, a multi-scale bilateral GRU measurement error prediction system includes a time-scale series generation module, a feature extraction and decomposition module, a bilateral GRU network construction module, and a result integration module.

[0116] The time-scale sequence generation module is used to acquire historical interaction data during the charging process of DC charging piles, and generate sequences of different time scales from the historical interaction data according to different sampling frequencies;

[0117] The feature extraction and decomposition module is used to extract the mapping features of the sequence at each scale, and to perform STL feature decomposition on the extracted mapping features to obtain the trend term, seasonal term and remainder term after feature decomposition of the sequence at each scale.

[0118] The bilateral GRU network construction module is used to construct a new bilateral GRU network, which feeds the trend term, seasonal term and residual term after the feature decomposition of each scale sequence into the bilateral GRU network for time series feature extraction, and obtains the measurement error prediction results of each scale sequence.

[0119] The result integration module is used to scale down the measurement error prediction results of each scale sequence and integrate them to obtain the measurement error prediction results of DC charging piles.

[0120] The feature extraction and decomposition module includes a mapping feature extraction unit and a feature decomposition unit;

[0121] The mapping feature extraction unit is used to extract the mapping features of each scale sequence in the multi-scale sequence dataset using the following formula:

[0122]

[0123] In the above formula, χm Let X be the sequence at the m-th scale. m The mapping features, For multi-scale sequence sets,

[0124] The feature decomposition unit is used to perform STL feature decomposition on the mapping features of each scale sequence using the following formula, to obtain the trend term, seasonal term, and remainder term after feature decomposition for each scale sequence:

[0125] STL(χ m )=(L m S m P m );

[0126]

[0127] In the above formula, L m S is the trend term. m For seasonal terms, P m For the remaining items, Let be the time series feature vector corresponding to the trend term at time t. Let be the temporal feature vector corresponding to the seasonal term at time t. Let be the temporal feature vector corresponding to the remainder term at time t.

[0128] The bilateral GRU network construction module includes a result output unit and an output result overlay unit;

[0129] The result output unit is used in the gating unit of the new bilateral GRU network to extract the temporal features of the trend and seasonal terms after feature decomposition at each scale using a reset gate and an update gate, and to obtain the output results of the trend and seasonal terms. In the linear prediction unit of the new bilateral GRU network, the unit extracts the temporal features of the remaining terms after feature decomposition, and to obtain the output results of the remaining terms.

[0130]

[0131] In the above formula, This represents the hidden state of the seasonal term at time t, i.e., the output of the seasonal term. The update gate that the seasonal item passes through. Let be the hidden state of the seasonal term at time t-1. Let be the candidate hidden state of the seasonal term at time t, and σ(.) be the Sigmoid function. Update the gate weight matrix for the seasonal terms. Let be the temporal eigenvector corresponding to the seasonal term at time t, and tanh(.) be the hyperbolic tangent function. Here is the weight matrix for the candidate hidden states of the seasonal item. The reset gate that the seasonal item passes through. Reset the weight matrix of the gate for the seasonal term. This represents the hidden state of the trend term at time t, i.e., the output result of the trend term. The update gate that the trend item passes through. Let be the hidden state of the trend term at time t-1. Let be the candidate hidden state of the trend term at time t. Update the gate weight matrix for the trend term. Let be the time series feature vector corresponding to the trend term at time t. This is the weight matrix of the candidate hidden states for the trend term. The reset gate that the trend item passes through. Reset the weight matrix of the gate for the trend term. The output of the remainder term, where α is the linear smoothing coefficient. Let be the temporal feature vector corresponding to the remainder term at time t. λ is the initial time-series feature vector of the remainder term, and θ are both coefficients of the exponential model.

[0132] The output result overlay unit is used to overlay the output results of the trend term, seasonal term, and remainder term to obtain the measurement error prediction results for each scale series:

[0133]

[0134] In the above formula, Y m The measurement error prediction result for the m-th scale sequence is given, where T is the total time. This represents the measurement error prediction result for the m-th scale sequence at time t.

[0135] The result integration module includes a sequence arrangement unit, a scaling unit, and a cyclic repetition unit;

[0136] The sequence arrangement unit is used to arrange the measurement error prediction results of each scale sequence in descending order of sampling frequency, i.e., in ascending order of time scale, as {Y0, ..., Y...}. m , ..., Y M};

[0137] The scaling unit is used for Y M Perform dilated convolution and mean pooling operations to make Y M The timescale is reduced to that of Y M-1 At the same scale, the scale-reduced econometric error prediction results AverPool(Dconv) are obtained. M (Y M In this context, AverPool(.) is the mean pooling operation, and Dconv is the mean pooling operation. M(.)This is an expanded convolution operation at the Mth scale, and the scale-reduced econometric error prediction result AverPool(Dconv) is used. M (Y M )) and Y M-1 Summation yields the summed measurement error prediction results.

[0138] The recurring unit is used to perform dilated convolution and mean pooling operations on the summed measurement error prediction results, reducing them to a value similar to Y. M-2 Same scale, and with Y M-2 Summation is performed, and this process is repeated until the econometric error prediction results for all scale sequences are integrated.

[0139] Example 3:

[0140] like Figure 5 As shown, a measurement error prediction device for a multi-scale bilateral GRU includes a processor and a memory.

[0141] The memory is used to store computer program code and to transmit the computer program code to the processor;

[0142] The processor is used to execute the metrological error prediction method of a multi-scale bilateral GRU as described in Embodiment 1 according to the instructions in the computer program code.

[0143] Example 4:

[0144] A computer storage medium on which computer programs are stored;

[0145] When the computer program is executed by the processor, it implements the steps of the metrological error prediction method of a multi-scale bilateral GRU described in this scheme.

Claims

1. A method for predicting measurement errors using a multi-scale bilateral GRU, characterized in that, The method includes: S1. Obtain historical interaction data during the charging process of DC charging piles, and generate sequences of different time scales from the historical interaction data according to different sampling frequencies; S2. Extract the mapping features of the sequences at each scale, and perform STL feature decomposition on the extracted mapping features to obtain the trend term, seasonal term, and remainder term after feature decomposition of the sequences at each scale, including: S21. Extract the mapping features of each scale sequence in the multi-scale sequence dataset using the following formula: ； In the above formula, For the first Sequences at the seed scale The mapping features, For multi-scale sequence sets, ; S22. Use the following formula to perform STL feature decomposition on the mapping features of each scale sequence to obtain the trend term, seasonal term, and remainder term after feature decomposition for each scale sequence: ； ； ； ； In the above formula, For trend items, For seasonal items, For the remaining items, For trend items in The temporal feature vector corresponding to time step 1. For seasonal items The temporal feature vector corresponding to time step 1. For the remaining items The temporal feature vector corresponding to each moment; S3. Construct a new bilateral GRU network, and feed the trend term, seasonal term and residual term after the feature decomposition of each scale sequence into the bilateral GRU network for time series feature extraction to obtain the measurement error prediction results of each scale sequence. S4. Scale-reduction is performed on the measurement error prediction results of each scale sequence, and the results are integrated to obtain the measurement error prediction results for DC charging piles, including: S41. Arrange the measurement error prediction results of each scale sequence in descending order of sampling frequency, i.e., in ascending order of time scale. ; S42, to Perform dilated convolution and mean pooling operations to make The timescale is reduced to the same At the same scale, the measurement error prediction results after scale reduction are obtained. ,in, For mean pooling operation, For the first Dilated convolution operations at various scales, and the scale-reduced econometric error prediction results. and Summation yields the summed measurement error prediction results. S43. Perform dilated convolution and mean pooling operations on the summed measurement error prediction results to reduce them to a value similar to... Same scale, and with Summation is performed, and this process is repeated until the econometric error prediction results for all scale sequences are integrated.

2. The metrological error prediction method for a multi-scale bilateral GRU according to claim 1, characterized in that, In S3, the new bilateral GRU network includes a gating unit and a linear prediction unit, and the specific steps include: S31. In the gating unit of the new bilateral GRU network, the reset gate and update gate are used to extract the temporal features of the trend and seasonal terms after the feature decomposition of each scale sequence, and the output results of the trend and seasonal terms are obtained: ； ； ； ； ； ； ； ； In the above formula, For seasonal items The hidden state at any given time, i.e., the output of the seasonal term. The update gate that the seasonal item passes through. For seasonal items The hidden state at all times For seasonal items The candidate hidden state at time step. For the Sigmod function, Update the gate weight matrix for the seasonal terms. For seasonal items The temporal feature vector corresponding to time step 1. It is the hyperbolic tangent function. Here is the weight matrix for the candidate hidden states of the seasonal item. The reset gate that the seasonal item passes through. Reset the weight matrix of the gate for the seasonal term. For trend items in The hidden state at any given time, i.e., the output of the trend term. The update gate that the trend item passes through. For trend items in The hidden state at all times For trend items in The candidate hidden state at time step. Update the gate weight matrix for the trend term. For trend items in The temporal feature vector corresponding to time step 1. This is the weight matrix of the candidate hidden states for the trend term. The reset gate that the trend item passes through. Reset the weight matrix of the gate for the trend term; In the linear prediction unit of the new bilateral GRU network, the following formula is used to extract the temporal features of the remainder after feature decomposition, and the output of the remainder is obtained: ； In the above formula, The output result for the remainder term, The linear smoothing coefficient is... For the remaining items The temporal feature vector corresponding to time step 1. The initial temporal feature vector of the remainder term. , All are coefficients of the exponential model; S32. By superimposing the outputs of the trend term, seasonal term, and residual term, the measurement error prediction results for each scale series are obtained: ； ； In the above formula, For the first Measurement error prediction results for various scale sequences. For the total time, For the first Scale sequences in Predicted measurement error results at time.

3. A metrological error prediction system for a multi-scale bilateral GRU, characterized in that, The system includes a time-scale sequence generation module, a feature extraction and decomposition module, a bilateral GRU network construction module, and a result integration module. The time-scale sequence generation module is used to acquire historical interaction data during the charging process of DC charging piles, and generate sequences of different time scales from the historical interaction data according to different sampling frequencies; The feature extraction and decomposition module is used to extract the mapping features of the sequence at each scale, and to perform STL feature decomposition on the extracted mapping features to obtain the trend term, seasonal term and remainder term after feature decomposition of the sequence at each scale, including a mapping feature extraction unit and a feature decomposition unit. The mapping feature extraction unit is used to extract the mapping features of each scale sequence in the multi-scale sequence dataset using the following formula: ； In the above formula, For the first Sequences at the seed scale The mapping features, For multi-scale sequence sets, ; The feature decomposition unit is used to perform STL feature decomposition on the mapping features of each scale sequence using the following formula, to obtain the trend term, seasonal term, and remainder term after feature decomposition for each scale sequence: ； ； ； ； In the above formula, For trend items, For seasonal items, For the remaining items, For trend items in The temporal feature vector corresponding to time step 1. For seasonal items The temporal feature vector corresponding to time step 1. For the remaining items The temporal feature vector corresponding to each moment; The bilateral GRU network construction module is used to construct a new bilateral GRU network, which feeds the trend term, seasonal term and residual term after the feature decomposition of each scale sequence into the bilateral GRU network for time series feature extraction, and obtains the measurement error prediction results of each scale sequence. The result integration module is used to scale down the measurement error prediction results of each scale sequence and integrate them to obtain the measurement error prediction results of DC charging piles, including a sequence arrangement unit, a scale reduction unit, and a cyclic repetition unit. The sequence arrangement unit is used to arrange the measurement error prediction results of each scale sequence in descending order of sampling frequency, i.e., in ascending order of time scale. ; The scale reduction unit is used for... Perform dilated convolution and mean pooling operations to make The timescale is reduced to the same At the same scale, the measurement error prediction results after scale reduction are obtained. ,in, For mean pooling operation, For the first Dilated convolution operations at various scales, and the scale-reduced econometric error prediction results. and Summation yields the summed measurement error prediction results. The recurring unit is used to perform dilated convolution and mean pooling operations on the summed measurement error prediction results, reducing them to a value similar to... Same scale, and with Summation is performed, and this process is repeated until the econometric error prediction results for all scale sequences are integrated.

4. The metrological error prediction system for a multi-scale bilateral GRU according to claim 3, characterized in that, The bilateral GRU network construction module includes a result output unit and an output result overlay unit; The result output unit is used in the gating unit of the new bilateral GRU network to extract the temporal features of the trend and seasonal terms after feature decomposition at each scale using a reset gate and an update gate, and to obtain the output results of the trend and seasonal terms. In the linear prediction unit of the new bilateral GRU network, the unit extracts the temporal features of the remaining terms after feature decomposition, and to obtain the output results of the remaining terms. ； ； ； ； ； ； ； ； ； In the above formula, For seasonal items The hidden state at any given time, i.e., the output of the seasonal term. The update gate that the seasonal item passes through. For seasonal items The hidden state at all times For seasonal items The candidate hidden state at time step. For the Sigmod function, Update the gate weight matrix for the seasonal terms. For seasonal items The temporal feature vector corresponding to time step 1. It is the hyperbolic tangent function. Here is the weight matrix for the candidate hidden states of the seasonal item. The reset gate that the seasonal item passes through. Reset the weight matrix of the gate for the seasonal term. For trend items in The hidden state at any given time, i.e., the output of the trend term. The update gate that the trend item passes through. For trend items in The hidden state at all times For trend items in The candidate hidden state at time step. Update the gate weight matrix for the trend term. For trend items in The temporal feature vector corresponding to time step 1. This is the weight matrix of the candidate hidden states for the trend term. The reset gate that the trend item passes through. Reset the weight matrix of the gate for the trend term. The output result for the remainder term, The linear smoothing coefficient is... For the remaining items The temporal feature vector corresponding to time step 1. The initial temporal feature vector of the remainder term. , All are coefficients of the exponential model; The output result overlay unit is used to overlay the output results of the trend term, seasonal term, and remainder term to obtain the measurement error prediction results for each scale series: ； ； In the above formula, For the first Measurement error prediction results for various scale sequences. For the total time, For the first Scale sequences in Predicted measurement error results at time.

5. A measurement error prediction device for a multi-scale bilateral GRU, characterized in that, Including the processor and memory; The memory is used to store computer program code and to transmit the computer program code to the processor; The processor is configured to execute, according to instructions in the computer program code, a metrological error prediction method for a multi-scale bilateral GRU as described in any one of claims 1-2.

6. A computer storage medium storing a computer program thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the metrological error prediction method of a multi-scale bilateral GRU as described in any one of claims 1-2.

Images