A power grid icing prediction method based on multi-source feature temporal convolution deep learning

Abstract

This invention discloses a power grid icing prediction method based on multi-source feature temporal convolutional deep learning, comprising: Step 1, constructing a historical icing force sequence dataset P, a force sequence Y to be predicted, a historical multi-environmental variable sequence dataset H related to meteorology, a meteorological forecast information dataset F, and a geographical location discretized dataset G; Step 2, constructing corresponding temporal convolutional network (TCN) modules for P, H, and F; Step 3, standardizing each dataset; Step 4, processing the standardized discretized influencing factors G′ and the temporal features extracted in Step 3, and simultaneously applying their respective learnable weight matrices W. i and offset b i Make empowerment X i Step 5: The weighted X from Step 4... i After being spliced ​​together, the data is simultaneously input into the fully connected layer to predict the tensile force data for the next 12 time steps, thus solving the technical problem of low prediction accuracy for power grid icing.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid icing prediction, and in particular relates to a power grid icing prediction method based on multi-source feature time convolution deep learning. Background Art

[0002] The temporal characteristics of power grid icing are affected by many factors, and there is a mapping relationship between them and the temporal data from various sources, and the data structure of the latter and the degree of influence on icing prediction are often different. In addition, in the same temporal data feature extraction network, discrete data in space, such as geographic information, is often difficult to use, resulting in low prediction accuracy of power grid icing. Summary of the invention

[0003] The technical problem to be solved by the present invention is: to provide a power grid icing prediction method based on multi-source feature time convolution deep learning, so as to solve the technical problems that when predicting power grid icing in the prior art, since the temporal characteristics of power grid icing are affected by various factors, there is a mapping relationship between it and the time-series data from various sources, and the data structure of the latter and the degree of influence on icing prediction are often different; in the same time-series data feature extraction network, discrete data in space, such as geographic information, is often difficult to use, resulting in low prediction accuracy of power grid icing.

[0004] The technical solution of the present invention is:

[0005] A power grid icing prediction method based on multi-source feature time convolution deep learning, which includes:

[0006] Step 1: Construct the historical ice tension sequence dataset P, the tension sequence Y to be predicted, the meteorological-related historical multi-environmental variable sequence dataset H, the meteorological forecast information dataset F and the geographic location discretization dataset G.

[0007] Step 2: Build the corresponding time convolution network TCN modules for P, H and F;

[0008] Step 3: Standardize each data set;

[0009] Step 4: For the discretized influencing factors G′ after standardization and the time series features extracted in step 3, use their respective learnable weight matrices W i and offset b i Do Empowerment X i ;

[0010] Step 5: The weighted X in step 4 i After concatenation, the data is input into the fully connected layer to predict the tension data Y for the next l2 time steps.

[0011] Step 6: Compare the prediction accuracy of the power grid icing prediction method based on multi-source feature time convolution deep learning and four multivariate time series prediction models; the four multivariate time series prediction models are ARIMA, RNN, LSTM and Seq2Seq.

[0012] The prediction accuracy is compared by comparing the mean absolute error (MAE), the root mean square error (RMSE) and the mean absolute percentage error (MAPE) as analysis indicators. The smaller the value of each analysis indicator, the higher the prediction accuracy.

[0013] The calculation formulas for the mean absolute error MAE, root mean square error RMSE and mean absolute percentage error MAPE as analysis indicators are:

[0014]

[0015]

[0016]

[0017] where y i and They represent the true value and the predicted value respectively, and n is the number of variables in the training sample.

[0018] Constructing a historical ice tension series dataset And the tension sequence that needs to be predicted: t represents the time step of the sequence data; historical multi-environmental variable sequence data set related to meteorology: i represents the i-th environmental variable; weather forecast information dataset: Geographic location discretization dataset: The length of each time window of the sequence data is l.

[0019] The method of building the corresponding time convolution network TCN module is as follows: the TCN network simultaneously uses one-dimensional convolution, causal convolution, dilated convolution and residual network technology to process serialized data in parallel; the calculation formula of the dilation coefficient d of each layer is:

[0020] d = b i

[0021] b is the expansion base number set by yourself, and i represents the number of layers;

[0022] Therefore, the relationship between the hyperparameters of each TCN module is:

[0023]

[0024] Where: w represents the receptive field width of a TCN residual block; k represents the width of the convolution kernel, the depth of the convolution kernel and the length of the input data feature dimension, that is, the contained features are the same; n is the number of layers of a TCN residual block;

[0025] In each TCN module, the hyperparameters, the number of convolution kernels, the number of TCN residual blocks, and whether to regularize and standardize the parameters are set individually according to the data structure, dimension, time window length, and data volume of each timing influencing factor.

[0026] The formula for the standardization process is:

[0027]

[0028] min and max represent the minimum and maximum values ​​of a feature;

[0029] The standardized time series influencing factors P′, H′, and F′ are input into their respective TCN modules to extract their respective time series features.

[0030] The size of nk is the same as the number of convolution kernels in the TCN module.

[0031] At the same time, through each learnable weight matrix W i and offset b i Do Empowerment X i The formula is:

[0032] X i =W i *x i +b i

[0033] x i Indicates p', h', f' or G'.

[0034] The activation function in the fully connected layer is the RELU function:

[0035] RELU(x)=max(0,x).

[0036] Beneficial effects of the present invention:

[0037] The present invention first designs multiple TCN modules with different structures to extract features of temporal influencing factors of different data dimensions, structures, and window lengths. After weighting and merging the output of each module and the pre-processed discretized influencing factors, the future ice cover changes are jointly predicted, so that the prediction process can simultaneously capture the different degrees of influence of multi-source data.

[0038] Compared with the existing technology, it has the following characteristics:

[0039] Corresponding TCN modules are set separately for different temporal influencing factors. Different TCN modules have different network structures and parameters, so that they can better adapt to different data structures, dimensions, time window lengths and data volumes of different temporal inputs, making the iterative process of each TCN network more targeted to their respective input data and enhancing their respective feature extraction capabilities.

[0040] Through their own learnable weight matrices and offsets, the discrete influencing factors and the features extracted from the temporal factors are weighted at the same time, and then merged to jointly predict the future ice cover. This can not only deeply mine the information of multi-source influencing factors at the same time, especially better utilize the discrete data in space, but also follow the different degrees of influence of different influencing factors on future ice cover changes, further improving the prediction effect.

[0041] The present invention solves the technical problems that the existing technology for predicting power grid icing has the following problems: since the temporal characteristics of power grid icing are affected by various factors, there is a mapping relationship between it and the temporal data from various sources, and the data structures of the latter and the degree of influence on icing prediction are often different; in the same temporal data feature extraction network, discrete data in space, such as geographic information, is often difficult to use, resulting in low prediction accuracy of power grid icing, etc. DETAILED DESCRIPTION

[0042] Specific implementation steps:

[0043] Step 1: Construct a training set. The tension signal collected by the sensor can directly reflect the ice information, so a historical ice tension sequence dataset is constructed: And the tension sequence that needs to be predicted: t represents the time step of the sequence data; historical multi-environmental variable sequence data set related to meteorology: i represents the i-th environmental variable; weather forecast information dataset: Geographic location discretization dataset: The lengths of the time windows of the sequence data are l1, l2, l3, and l4 respectively; there are n1 meteorological-related historical environmental variables recorded in H; there are n2 environmental variables included in the meteorological forecast recorded in F; and there are n3 geographical environment-related variables recorded in G.

[0044] Step 2: Build the corresponding time convolution network TCN module for each data set P, H, and F of the temporal influencing factors. The TCN network uses multiple CNN core technologies such as one-dimensional convolution, causal convolution, dilated convolution, and residual network to process serialized data in parallel. The calculation formula for the dilation coefficient d of each layer is:

[0045] d = b i

[0046] b is the expansion base set by yourself, and i represents the number of layers.

[0047] Therefore, the relationship between the hyperparameters of each TCN module is:

[0048]

[0049] Among them, w represents the receptive field width of a TCN residual block; k represents the width of the convolution kernel, the depth of the convolution kernel and the length of the input data feature dimension, that is, the contained features are the same; n is the number of layers of a TCN residual block.

[0050] In each TCN module, the above hyperparameters, as well as the number of convolution kernels, the number of TCN residual blocks, whether to regularize the parameters and the method of parameter standardization, are all set separately according to the data structure, dimension, time window length and data volume of each timing influencing factor.

[0051] Step 3: Standardize the data:

[0052]

[0053] min and max represent the minimum and maximum values ​​of a feature.

[0054] The standardized data sets of temporal influencing factors P′, H′, and F′ are input into the TCN modules built for each of them to extract their respective temporal features. The size of nk is the same as the number of convolution kernels in the TCN module.

[0055] Step 4: For the standardized discretized influencing factor dataset G′ and the time series features extracted in step 3, use their respective learnable weight matrices W i and offset b i Do empowerment:

[0056] X i =W i *x i +b i

[0057] x i Indicates the above p', h', f', G'.

[0058] Step 5: Multiple weighted X in step 4 i, after being spliced ​​in the same dimension direction, they are input into a multi-layer fully connected layer built for the merged feature matrix for nonlinear operation to predict the tension data Y of the next l2 time steps. This allows the prediction process to capture the different degrees of influence of multi-source data at the same time and improve the accuracy of power grid ice prediction.

[0059] The activation function in the fully connected layer is the RELU function:

[0060] RELU(x)=max(0,x)

[0061] In order to verify the effectiveness of the present invention, 1235 power grid monitoring terminals were selected to conduct experiments on the data set collected by China Southern Power Grid.

[0062] Experiment 1 compares the method of the present invention with a method that cuts each time-series influencing factor into the same time window and splices them on the feature dimension and then uniformly inputs them into a TCN module for prediction. The latter fails to utilize the discrete influencing factors in space, such as geographic information.

[0063] Experiment 2 is a comparative experiment with four multivariate time series prediction models. The comparison models are:

[0064] (1) ARIMA: Autoregressive Integrated Moving Average is a model widely used in time series analysis that combines moving average and autoregressive methods.

[0065] (2) RNN: This is the most traditional deep learning method for processing sequence tasks.

[0066] (3) LSTM: LSTM is a classic time recurrent neural network improved on the basis of RNN, which is good at capturing the temporal patterns in the input sequence.

[0067] (4) Seq2Seq: A classic sequence-to-sequence deep learning model that can be used for traffic flow prediction. This model selects LSTM as the encoder and decoder components.

[0068] The comparative experiment uses mean absolute error MAE, root mean square error RMSE, and mean absolute percentage error MAPE as analysis indicators. The calculation formula is:

[0069]

[0070]

[0071]

[0072] where y i and Represent the true value and predicted value respectively, and n is the number of variables in the training sample. The smaller the value of each analysis indicator, the higher the prediction accuracy.

[0073] The comparison results of Experiment 1 are shown in Table 1:

[0074] Table 1 Experiment 1 results

[0075]

[0076]

[0077] The experimental results show that the TDNet constructed by the present invention is compared with the method that does not consider the discretization influencing factors and uniformly inputs all the time series influencing factors into a single TCN module for prediction in the format of the same time window length. The values ​​of various indicators are significantly reduced and the prediction performance is significantly improved. It can be seen that discrete data such as geographic information in space has a significant impact on ice cover changes. At the same time, it is proved that building time convolutional network modules of different structures for multi-source heterogeneous time series data can make the feature extraction process of each module more targeted and adaptive, and improve the overall prediction effect of the model.

[0078] The comparison results of Experiment 2 are shown in Table 2:

[0079] Table 2 Experiment 2 results

[0080]

[0081] The experimental results show that the TDNet constructed by the present invention is better than the classic time series prediction algorithm. Among them, none of the comparative algorithms considers discrete data such as geographic information in space, which once again shows the importance of multi-source data factors related to ice cover on prediction performance and the correctness of the design of the present invention.

Claims

1. A power grid icing prediction method based on multi-source feature time convolution deep learning, which includes: Step 1: construct a historical ice tension sequence dataset P, a tension sequence Y to be predicted, a meteorological-related historical multi-environmental variable sequence dataset H, a meteorological forecast information dataset F, and a geographic location discretization dataset G; Step 2: Build the corresponding time convolution network TCN modules for P, H and F; Step 3: Standardize each data set; the formula for the standardization is: min and max represent the minimum and maximum values ​​of a feature; The standardized time series influencing factors P′, H′, and F′ are input into their respective TCN modules to extract their respective time series features. The size of nk is the same as the number of convolution kernels in the TCN module; Step 4: For the discretized influencing factors G′ after standardization and the time series features extracted in step 3, the learnable weight matrix W i and offset b i Do Empowerment X i ; Step 5: The weighted X in step 4 i After concatenation, the data is input into the fully connected layer to predict the tension data Y for the next l2 time steps.

2. According to claim 1, a method for predicting power grid icing based on multi-source feature time convolution deep learning is characterized in that: Step 6: Compare the prediction accuracy of the power grid icing prediction method based on multi-source feature time convolution deep learning and four multivariate time series prediction models; the four multivariate time series prediction models are ARIMA, RNN, LSTM and Seq2Seq.

3. The method for predicting power grid icing based on multi-source feature time convolution deep learning according to claim 2 is characterized in that: The prediction accuracy is compared by comparing the mean absolute error (MAE), the root mean square error (RMSE) and the mean absolute percentage error (MAPE) as analysis indicators. The smaller the value of each analysis indicator, the higher the prediction accuracy.

4. The method for predicting power grid icing based on multi-source feature time convolution deep learning according to claim 3 is characterized in that: The calculation formulas for the mean absolute error MAE, root mean square error RMSE and mean absolute percentage error MAPE as analysis indicators are: where y i and They represent the true value and predicted value respectively, and n is the number of variables in the training sample.

5. The method for predicting power grid icing based on multi-source feature time convolution deep learning according to claim 1 is characterized in that: Constructing a historical ice tension series dataset And the tension sequence that needs to be predicted: t represents the time step of the sequence data; historical multi-environmental variable sequence data set related to meteorology: i represents the i-th environmental variable; weather forecast information dataset: Geographic location discretization dataset:

6. The method for predicting power grid icing based on multi-source feature time convolution deep learning according to claim 1 is characterized in that: The method of building the corresponding time convolution network TCN module is as follows: the TCN network simultaneously uses one-dimensional convolution, causal convolution, dilated convolution and residual network technology to process serialized data in parallel; the calculation formula of the dilation coefficient d of each layer is: d=b i b is the expansion base number set by yourself, and i represents the number of layers; Therefore, the relationship between the hyperparameters of each TCN module is: Where: w represents the receptive field width of a TCN residual block; k represents the width of the convolution kernel, the depth of the convolution kernel and the length of the input data feature dimension, that is, the contained features are the same; n is the number of layers of a TCN residual block; In each TCN module, the hyperparameters, the number of convolution kernels, the number of TCN residual blocks, and whether to regularize and standardize the parameters are set individually according to the data structure, dimension, time window length, and data volume of each timing influencing factor.

7. The method for predicting power grid icing based on multi-source feature time convolution deep learning according to claim 1 is characterized in that: At the same time, through each learnable weight matrix W i and offset b i Do Empowerment X i The formula is: X i =W i *x i +b i x i Indicates p', h', f' or G'.

8. The method for predicting power grid icing based on multi-source feature time convolution deep learning according to claim 1 is characterized in that: The activation function in the fully connected layer is the RELU function: RELU(x)=max(0,x).

Images