Return stroke peak current remote estimation method based on semi-supervised learning

Abstract

The invention relates to a return stroke peak current remote estimation method based on semi-supervised learning, and belongs to the technical field of lightning discharge parameter inversion. The method comprises the following steps: obtaining an error correction coefficient of each observation station intensity by utilizing return strikes on a perpendicular bisector of an observation station, and completing the correction of station errors; fitting to obtain an overall attenuation curve of an electric field signal generated by return stroke by using the return stroke of 100km + / -10km around the observation station; semi-supervised learning is utilized to predict the offset of the real attenuation curve of each return stroke electric field signal received by each observation station relative to the obtained overall attenuation curve, and then the real attenuation curve of each return stroke electric field signal is obtained, so that an accurate normalized electric field peak value is obtained; and calculating the return stroke peak current by using the obtained normalized electric field peak value and the transmission line equation, namely the remotely estimated return stroke peak current. According to the method, more accurate remote estimation of the return stroke peak current is realized, a short-baseline lightning positioning network is not needed as a reference, and the method is easy to popularize and apply.

Description

Technical Field

[0001] The present invention belongs to the technical field of lightning discharge parameter inversion, and particularly relates to a method for remotely estimating the return stroke peak current based on semi-supervised learning, and more particularly to a method for remotely estimating the return stroke peak current based on semi-supervised learning applicable to a long-distance lightning location network. Background Art

[0002] At present, the measurement of the return stroke peak current is mainly carried out in two ways: direct measurement using Rogowski coils and coaxial shunts, and indirect inversion of the electromagnetic field or light measured by a ground-based lightning location system. Among them, when estimating the return stroke peak current by measuring the electric field using a ground-based lightning location system, the propagation of the electromagnetic wave excited by lightning discharge on a lossy surface is usually considered, the peak value of the electric field signal is normalized to a certain distance (usually 100 km), and then an empirical equation or a transmission line model is used to estimate the return stroke peak current. In particular, in a long-distance lightning location network, the signal received by a station is not a pure ground wave, but a composite wave of a ground wave and a sky wave reflected by the ionosphere, and its propagation attenuation is more difficult to accurately calculate. Therefore, the current technology tends to establish an empirical formula between the electrical parameters (such as the normalized peak value and normalized power of the signal) of the long-distance lightning location network and the peak current estimated by the short-baseline lightning location network. It should be noted that this method depends on the short-baseline lightning location network as a reference and cannot operate independently.

[0003] Therefore, how to overcome the deficiencies of the existing technology is an urgent problem to be solved in the current technical field of lightning discharge parameter inversion. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies of the existing technology and provide a method for remotely estimating the return stroke peak current based on semi-supervised learning.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] A method for remotely estimating the return stroke peak current based on semi-supervised learning, comprising the following steps:

[0007] Step (1), using the return strokes on the perpendicular bisector of the station to obtain the error correction coefficients of the intensities of each station, and completing the correction of the station errors;

[0008] Step (2), using the return strokes within 100 km ± 10 km around the station to fit and obtain the overall attenuation curve of the electric field signal generated by the return stroke;

[0009] Step (3): Using semi-supervised learning to predict the offset of the true attenuation curve of each return stroke electric field signal received by each station relative to the overall attenuation curve obtained in step (2), thereby obtaining the true attenuation curve of each return stroke electric field signal, and thus obtaining an accurate normalized electric field peak value.

[0010] Step (4): Using the normalized electric field peak value obtained in step (3) and the transmission line equation to calculate the return stroke peak current, which is the remotely estimated return stroke peak current.

[0011] Furthermore, preferably, the specific method of step (1) is as follows:

[0012] Step (1.1): Select all return strokes on the perpendicular bisector of the line connecting two stations, count the ratio of the ground wave peaks of the same return stroke measured by the two stations, and further calculate the average value of all ratios. Take this average value as the station gain ratio of the two stations.

[0013] Step (1.2): Using the gain ratios of all stations to the reference station obtained in step (1.1), take the gain ratio of each station to the reference station as the initial error correction coefficient of this station; it is required that the reference station is a national meteorological station and the surrounding area within 10 km is flat.

[0014] Use the least squares method to optimize the initial error correction coefficient to obtain the error correction coefficient of each station.

[0015] When optimizing, it is required to satisfy the following relationship:

[0016] Assume that the gain ratio of station A to station B is k A , the gain ratio of station C to station B is k B , the gain ratio of station A to station C is k C , the following relationship should be satisfied.

[0017]

[0018] Step (1.3): Divide the return stroke electric field signals measured by each station by the error correction coefficient of the corresponding station obtained in step (1.2), so as to correct the signal intensity measured by each station.

[0019] Furthermore, preferably, the specific method of step (2) is as follows:

[0020] For each station, select the return strokes within 100 km ± 10 km around this station, divide the ground wave peak of the return stroke electric field measured by other stations by the ground wave peak measured by this station to obtain the peak ratio, and take this peak ratio as the attenuation ratio y.

[0021] For each return stroke, there exists a set of data (x, y) for any other measuring station M; where x is the distance of the return stroke from the measuring station M, and y is the attenuation ratio calculated from the ground wave peak measured by the return stroke and the measuring station M.

[0022] Then, for all (x, y), fitting is performed in the manner of power function attenuation and exponential function attenuation respectively to obtain the overall attenuation curve of the ground wave peak ratio of the electric field signals generated by all return strokes with respect to the propagation distance.

[0023] Furthermore, preferably, the formula for power function attenuation is:

[0024] y = ax -b

[0025] The formula for exponential function attenuation is:

[0026]

[0027] where a represents the scaling factor, b represents the attenuation exponent, and L represents the attenuation distance.

[0028] Furthermore, preferably, in step (3), a deep learning model is used to predict the offset of the true attenuation curve of each return stroke electric field signal received by each measuring station relative to the overall attenuation curve obtained in step (2).

[0029] The deep learning model includes a waveform feature extraction branch, a fusion layer, a fifth batch normalization layer, a dense block, an attention layer, and a fully connected layer connected in sequence; the waveform feature extraction branch includes a first convolutional layer, a first batch normalization layer, an activation layer of the first Relu function, a second convolutional layer, a second batch normalization layer, an activation layer of the second Relu function, a first multi-scale residual block, a second multi-scale residual block, a third multi-scale residual block, a fourth multi-scale residual block, a max pooling layer, a third convolutional layer, a third batch normalization layer, an activation layer of the third Relu function, a fourth convolutional layer, a fourth batch normalization layer, an activation layer of the fourth Relu function, a fifth multi-scale residual block, and a global average pooling layer connected in sequence.

[0030] The input of the deep learning model is the return stroke waveform, the measuring station name, the distance from the return stroke to the measuring station, the ground wave peak measured by the measuring station, time, azimuth angle, and the propagation speed of the lightning electromagnetic wave, and the output is the offset of the true attenuation curve of each return stroke electric field signal relative to the overall attenuation curve obtained in step (2).

[0031] Furthermore, preferably, a semi-supervised learning method is used for training, and the specific method of step (3) is:

[0032] Step (3.1), pre-training stage: Construct a pseudo-label to train the deep learning model.

[0033] The output of the deep learning model is the offsets Δb and ΔL of the true attenuation curve of each return stroke electric field signal relative to the overall attenuation curve obtained in step (2), where Δb corresponds to the offset of the power function attenuation curve and ΔL corresponds to the offset of the exponential function attenuation curve;

[0034] The formula for the pseudo label is:

[0035]

[0036] where RNSS represents the ground wave peak value of the electric field signal normalized to 100 km, SS represents the ground wave peak value of the return stroke electric field signal measured at the station; r represents the propagation distance of the electric field signal, i.e., the distance of the return stroke from the station; b represents the attenuation exponent, L represents the attenuation distance; pseudo label represents the pseudo label; n represents the number of stations that detected the return stroke;

[0037] Train the model. The final effect is that the RNSS calculated after the signals of the same return stroke measured at different stations pass through the deep learning model should be consistent;

[0038] Step (3.2), fine-tuning stage: Freeze all layers of the model except the fully connected layer, and train the deep learning model again; the pseudo label used in the training process is replaced with the ground wave peak value of the signal measured at stations 100 ± 10 km away from the return stroke;

[0039] In step (3.3), after the training is completed, use the deep learning model trained in step (3.2) for prediction to obtain the offsets Δb and ΔL relative to the overall attenuation curve, and then calculate RNSS to obtain the accurate normalized electric field peak values of all return strokes.

[0040] Furthermore, preferably, in step (3.2), the return strokes used in the training dataset are measured at stations 100 ± 10 km away from the return stroke, and the return strokes with a signal-to-noise ratio greater than 10.

[0041] Furthermore, preferably, the specific method of step (4) is:

[0042] Calculate using the following formula:

[0043]

[0044] where ε 0 represents the vacuum permittivity, c represents the speed of light, r represents the distance of the return stroke from the station, v represents the return stroke speed, E represents the normalized field electric field peak value, and I represents the return stroke peak current.

[0045] Furthermore, preferably, v is set to 1.3×10 8 m / s.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] The present invention overcomes the problems in the estimation of the return stroke peak current in the current long-distance lightning location network, where it is difficult to accurately calculate the propagation attenuation, it depends on the short-baseline lightning location network as a reference and cannot operate independently. It realizes a more accurate remote estimation of the return stroke peak current and gets rid of the dependence on the short-baseline lightning location network. Compared with the prior art, the solution proposed by the present invention can operate independently. After being corrected by semi-supervised learning, the coefficient of determination of the long-distance estimated return stroke peak current of the present invention and the return stroke peak current estimated by the artificial triggered lightning relationship at close range increases from 0.8362 to 0.9178, and the mean square error decreases from 21.51% to 10.67%. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic flow chart of the method for remotely estimating the return stroke peak current based on semi-supervised learning of the present invention;

[0049] Figure 2 It is a schematic diagram of the site error correction of the present invention; (a) is a schematic diagram of the measuring station and the perpendicular bisector of the present invention; (b) is the ratio of the ground wave peak values of all return strokes on the perpendicular bisector measured by the SQ station and the NJ station; (c) is the ratio of the ground wave peak values of all return strokes on the perpendicular bisector measured by the HY station and the NJ station; (d) is the ratio of the ground wave peak values of all return strokes on the perpendicular bisector measured by the SQ station and the HY station.

[0050] Figure 3 It is an overall attenuation curve graph of the electric field signal generated by lightning of the present invention;

[0051] Figure 4 It is a schematic diagram of the structure of the deep learning model of the present invention;

[0052] Figure 5 It is a schematic diagram of the pre-training stage of the present invention;

[0053] Figure 6 It is a schematic diagram of the effect of remotely estimating the return stroke peak current of the present invention. (a) is a scatter diagram of the return stroke peak current estimated by combining the corrected electric field peak value of the overall attenuation curve in step 2 of the present invention with the transmission line equation and the return stroke peak current estimated by the artificial triggered lightning relationship at close range; (b) is a scatter diagram of the return stroke peak current estimated by the present invention and the return stroke peak current estimated by the artificial triggered lightning relationship at close range DETAILED DESCRIPTION OF THE INVENTION

[0054] The present invention will be further described in detail below in conjunction with the embodiments.

[0055] Those skilled in the art will understand that the following embodiments are only used to illustrate the present invention and should not be construed as limiting the scope of the present invention. For those without specific technical or conditions noted in the embodiments, the techniques or conditions described in the literature in this field or according to the product specifications shall be followed. For those materials or equipment without the manufacturer noted, they are all conventional products that can be obtained by purchase.

[0056] Embodiment 1

[0057] As Figure 1 shown, a method for remotely estimating the return stroke peak current based on semi-supervised learning, characterized by comprising the following steps:

[0058] Step (1), using the return strokes on the perpendicular bisector of the measuring station to obtain the error correction coefficients of the intensities of each measuring station, and completing the correction of the station errors;

[0059] Step (2), using the return strokes within 100 km ± 10 km around the measuring station to fit the overall attenuation curve of the electric field signal generated by the return stroke;

[0060] Step (3), using semi-supervised learning to predict the offset of the true attenuation curve of each return stroke electric field signal received by each measuring station relative to the overall attenuation curve obtained in step (2), and then obtaining the true attenuation curve of each return stroke electric field signal, so as to obtain an accurate normalized electric field peak;

[0061] Step (4), using the normalized electric field peak obtained in step (3) and the transmission line equation to calculate the return stroke peak current, which is the remotely estimated return stroke peak current.

[0062] Embodiment 2

[0063] As Figure 1 shown, a method for remotely estimating the return stroke peak current based on semi-supervised learning, characterized by comprising the following steps:

[0064] Step (1), using the return strokes on the perpendicular bisector of the measuring station to obtain the error correction coefficients of the intensities of each measuring station, and completing the correction of the station errors;

[0065] Step (2), using the return strokes within 100 km ± 10 km around the measuring station to fit the overall attenuation curve of the electric field signal generated by the return stroke;

[0066] Step (3), using semi-supervised learning to predict the offset of the true attenuation curve of each return stroke electric field signal received by each measuring station relative to the overall attenuation curve obtained in step (2), and then obtaining the true attenuation curve of each return stroke electric field signal, so as to obtain an accurate normalized electric field peak;

[0067] Step (4): Calculate the return stroke peak current by using the normalized electric field peak value obtained in step (3) and the transmission line equation, which is the remotely estimated return stroke peak current.

[0068] The specific method of step (1) is as follows:

[0069] Step (1.1): Select all return strokes on the perpendicular bisector of the line connecting two stations, count the ratio of the ground wave peak values of the same return stroke measured by the two stations, further calculate the mean value of all ratios, and use this mean value as the station gain ratio of the two stations.

[0070] Step (1.2): Use the gain ratios of all stations to the reference station obtained in step (1.1), and use the gain ratio of each station to the reference station as the initial error correction coefficient of this station; it is required that the reference station is a national meteorological station and the surrounding area within 10 km is flat.

[0071] Optimize the initial error correction coefficient by using the least squares method to obtain the error correction coefficient of each station.

[0072] During optimization, it is required to satisfy the following relationship:

[0073] Assume that the gain ratio of station A to station B is k A , the gain ratio of station C to station B is k B , and the gain ratio of station A to station C is k C , the following relationship should be satisfied.

[0074]

[0075] Step (1.3): Divide the return stroke electric field signal measured by each station by the error correction coefficient of the corresponding station obtained in step (1.2) to correct the signal intensity measured by each station.

[0076] The specific method of step (2) is as follows:

[0077] For each station, select the return strokes within 100 km ± 10 km around this station, divide the ground wave peak value of the return stroke electric field measured by other stations by the ground wave peak value measured by this station to obtain the peak ratio, and use this peak ratio as the attenuation ratio y.

[0078] For each return stroke, there is a set of data (x, y) for any other station M; where x is the distance of the return stroke from station M, and y is the attenuation ratio calculated from this return stroke and the ground wave peak value measured by this station M.

[0079] Then, for all (x, y), fit them in the form of power function attenuation and exponential function attenuation respectively to obtain the overall attenuation curve of the ground wave peak ratio of the electric field signal generated by all return strokes with the propagation distance.

[0080] The formula for power function attenuation is:

[0081] y = ax -b

[0082] The formula for exponential function attenuation is:

[0083]

[0084] where a represents the scaling factor, b represents the attenuation exponent, and L represents the attenuation distance.

[0085] In step (3), a deep learning model is used to predict the offset of the true attenuation curve of each return stroke electric field signal received by each station relative to the overall attenuation curve obtained in step (2);

[0086] The deep learning model includes a waveform feature extraction branch, a fusion layer, a fifth batch normalization layer, a dense block, an attention layer, and a fully connected layer connected in sequence; the waveform feature extraction branch includes a first convolutional layer, a first batch normalization layer, an activation layer of the first Relu function, a second convolutional layer, a second batch normalization layer, an activation layer of the second Relu function, a first multi-scale residual block, a second multi-scale residual block, a third multi-scale residual block, a fourth multi-scale residual block, a max pooling layer, a third convolutional layer, a third batch normalization layer, an activation layer of the third Relu function, a fourth convolutional layer, a fourth batch normalization layer, an activation layer of the fourth Relu function, a fifth multi-scale residual block, and a global average pooling layer connected in sequence;

[0087] The input of the deep learning model is the return stroke waveform, station name, distance from the return stroke to the station, peak value of the ground wave measured by the station, time, azimuth angle, and propagation speed of the lightning electromagnetic wave, and the output is the offset of the true attenuation curve of each return stroke electric field signal relative to the overall attenuation curve obtained in step (2).

[0088] Trained using a semi-supervised learning method, the specific method of step (3) is:

[0089] Step (3.1), pre-training stage: Construct a pseudo-label to train the deep learning model;

[0090] The output of the deep learning model is the offsets Δb and ΔL of the true attenuation curve of each return stroke electric field signal relative to the overall attenuation curve obtained in step (2), where Δb corresponds to the offset of the power function attenuation curve and ΔL corresponds to the offset of the exponential function attenuation curve;

[0091] The formula for the pseudo-label is:

[0092]

[0093]

[0094] Among them, RNSS represents the ground wave peak value of the electric field signal normalized to 100 km, and SS represents the ground wave peak value of the return stroke electric field signal measured at the station; r represents the propagation distance of the electric field signal, that is, the distance from the return stroke to the station; b represents the attenuation exponent, L represents the attenuation distance; pseudo label represents the pseudo label; n represents the number of stations detecting the return stroke;

[0095] Train the model, and the final effect is that the RNSS calculated after the signals of the same return stroke measured at different stations pass through the deep learning model should be consistent;

[0096] Step (3.2), fine-tuning stage: Freeze all layers of the model except the fully connected layer, and train the deep learning model again; the pseudo label used in the training process is replaced with the ground wave peak value of the signal measured at the station 100 ± 10 km away from the return stroke;

[0097] In step (3.3), after the training is completed, use the deep learning model trained in step (3.2) for prediction to obtain the offsets Δb and ΔL of the above relative overall attenuation curve, and then calculate RNSS to obtain the accurate normalized electric field peak value of all return strokes.

[0098] In step (3.2), the return strokes used in the training dataset are measured at stations 100 ± 10 km away from the return stroke, and the signal-to-noise ratio of the signal is greater than 10.

[0099] The specific method of step (4) is as follows:

[0100] The following formula is used for calculation:

[0101]

[0102] Among them, ε 0 represents the vacuum permittivity, c represents the speed of light, r represents the distance from the return stroke to the station, v represents the return stroke speed, E represents the normalized field electric field peak value, and I represents the return stroke peak current.

[0103] v is set to 1.3×10 8 m / s.

[0104] Application example

[0105] As Figure 1 shown, this example discloses a method for remotely estimating the return stroke peak current based on semi-supervised learning, which combines traditional estimation techniques and semi-supervised learning to achieve more accurate remote estimation of the return stroke peak current, and does not require a short baseline lightning location network as a reference, including the following steps:

[0106] Step (1), using the return strokes on the perpendicular bisector of the station to obtain the error correction coefficients of the intensities of each station, and complete the correction of the station errors;

[0107] Step (1) specifically includes:

[0108] In step (1.1), select all the return strokes on the perpendicular bisector of the line connecting two measurement stations, count the ratio of the ground wave peak values of the same return stroke measured by the two measurement stations, and further calculate the average value of all the ratios to obtain the station gain ratio of the two measurement stations;

[0109] Specifically: The lightning data is from the established very low frequency long-distance lightning location network, which consists of more than twenty very low frequency measurement stations with baseline lengths ranging from several hundred to several thousand kilometers distributed across China. The data range includes the return stroke data for one month in China and its surrounding areas. As Figure 2 shown in (a), for the three measurement stations SQ, NJ, and HY in the example, respectively select the return strokes (solid lines) on the perpendicular bisectors of each pair, count the ratio of the ground wave peak values of the same return stroke measured by the two measurement stations, and further calculate the average value of all the ratios. Take this average value as the station gain ratio of the two measurement stations; The results are as Figure 2 (b), Figure 2 (c) and Figure 2 (d) shown. As Figure 2 shown in (b) for SQ-NJ, SQ is the numerator and HJ is the denominator. For the average values of the ground wave peak ratios between each pair of the three measurement stations in this example, they are 0.491, 1.073, and 0.449 respectively, and the average values of the ground wave peak ratios basically do not change with the change of the lightning signal propagation distance. Then it can be considered that the station gain ratios of the three measurement stations in this example are the same.

[0110] In step (1.2), use the gain ratios of all measurement stations and the reference station obtained in step (1.1), and take the gain ratio of each measurement station and the reference station as the initial error correction coefficient of this measurement station; It is required that the reference station is a national meteorological station and the surrounding area within 10 km is flat;

[0111] Use the least squares method to optimize the initial error correction coefficient to obtain the error correction coefficient of each measurement station;

[0112] During optimization, it is required to satisfy the following relationship:

[0113] Assume that the gain ratio of measurement station A and measurement station B is k A , the gain ratio of measurement station C and measurement station B is k B , and the gain ratio of measurement station A and measurement station C is k C , then the following relationship should be satisfied.

[0114]

[0115] Specifically: In an ideal situation, assume that the gain ratio of SQ-NJ is k1, the gain ratio of HY-NJ is k2, and the gain ratio of SQ-HY is k3, then the following relationship should be satisfied.

[0116]

[0117] The SQ station in this embodiment is installed at a national meteorological station, and the surrounding area is flat. It is considered that there is no station error, that is, the gain ratio is 1. Further, the gain ratio obtained in step (1.1) is used as the error correction coefficient, and the error correction coefficient also satisfies the above relationship. Further, the initial error correction coefficients of 20 stations in this example are obtained, and the least squares method is used to iterate and optimize the above ratio so that the error correction coefficients of every three stations satisfy the above relational expression. The station error correction coefficients of the 20 stations obtained in this example are shown in Table 1.

[0118] In step (1.3), the original return stroke electric field signals measured at each station are divided by the error correction coefficients in Table 1, so as to correct the signal intensity measured at each station.

[0119] Table 1 Error correction coefficients of 20 stations obtained in this example

[0120] Station name for measurement SQ ALT CD CH DLS Error correction coefficient 1.000 10.081 1.373 1.831 1.684 Station name for measurement DM GY HM HY JQ Error correction coefficient 1.030 0.690 0.534 2.261 2.085 Station name for measurement KS NJ NX PE SFH Error correction coefficient 9.400 2.029 0.412 4.308 2.551 Station name for measurement TY WHai WHan WQX WZ Error correction coefficient 2.280 1.903 0.473 0.916 2.542

[0121] In step (2), the overall attenuation curve of the electric field signal generated by lightning is fitted by using the return strokes within 100 ± 10 km around the station;

[0122] Specific location: Select the 20 stations that have been corrected for station error in step (1). For each station, select the return strokes within 100 km ± 10 km around the station. Divide the ground wave peak value of the return stroke electric field signal measured at other stations by the ground wave peak value measured at this station to obtain the peak ratio, and use this peak ratio as the attenuation ratio y;

[0123] For each return stroke, there is a set of data (x, y) for any other station M; where x is the distance of the return stroke from station M, and y is the attenuation ratio calculated by this return stroke and the ground wave peak value measured at station M;

[0124] A total of 683,852 sets of data on the distance of the return stroke from the station and the attenuation ratio are obtained in the above steps. Further, the overall attenuation curves of the ground wave peak value of the electric field signal generated by the return stroke with the propagation distance (the propagation distance is the distance of the return stroke from the station) are fitted in the ways of power function attenuation and exponential function attenuation respectively. The formulas for power function attenuation and exponential function attenuation are as follows:

[0125] y = ax -b

[0126]

[0127] Among them, x represents the distance from the return stroke to the measuring station, y represents the attenuation ratio, a represents the scaling factor, b represents the attenuation exponent, and L represents the attenuation distance. The overall attenuation curve of the electric field signal generated by lightning obtained in this example is as Figure 3 shown. For the attenuation in the form of a power function, b = 1.109, and for the exponential function attenuation, L = 2615. The value of a can be cancelled out during the calculation process below.

[0128] Step (3), using semi-supervised learning to predict the offset of the true attenuation curve of each return stroke electric field signal received by each measuring station relative to the overall attenuation curve obtained in step (2), and then obtaining the true attenuation curve of each return stroke electric field signal, so as to obtain an accurate normalized electric field peak value;

[0129] Step (3) specifically includes:

[0130] Using a deep learning model to predict the offset of the true attenuation curve of each return stroke electric field signal received by each measuring station relative to the overall attenuation curve obtained in step (2);

[0131] Step (3.1), pre-training stage: constructing a deep learning model with pseudo-labels for training;

[0132] As Figure 4 , the deep learning model includes a waveform feature extraction branch, a fusion layer, a fifth batch normalization layer, a dense block, an attention layer, and a fully connected layer connected in sequence;

[0133] The waveform feature extraction branch includes a first convolutional layer, a first batch normalization layer, an activation layer of the first ReLU function, a second convolutional layer, a second batch normalization layer, an activation layer of the second ReLU function, a first multi-scale residual block, a second multi-scale residual block, a third multi-scale residual block, a fourth multi-scale residual block, a max pooling layer, a third convolutional layer, a third batch normalization layer, an activation layer of the third ReLU function, a fourth convolutional layer, a fourth batch normalization layer, an activation layer of the fourth ReLU function, a fifth multi-scale residual block, and a global average pooling layer connected in sequence;

[0134] The input of the deep learning model is the return stroke waveform, the measuring station name, the distance from the return stroke to the measuring station, the ground wave peak value measured by the measuring station, time, azimuth angle, and the propagation speed of the lightning electromagnetic wave, and the output is the offsets Δb and ΔL of the true attenuation curve of each return stroke electric field signal relative to the overall attenuation curve obtained in step (2), where Δb corresponds to the offset of the power function attenuation curve and ΔL corresponds to the offset of the exponential function attenuation curve. Further, calculate RNSS according to Δb and ΔL and construct pseudo-labels;

[0135] The formula for the pseudo-label is:

[0136]

[0137] Among them, RNSS represents the ground wave peak value of the electric field signal normalized to 100 km, and SS represents the ground wave peak value of the return stroke signal measured at the station; r represents the propagation distance of the electric field signal, that is, the distance of the return stroke from the station; b represents the attenuation exponent, and L represents the attenuation distance; pseudo label represents the pseudo label; n represents the number of stations that detected this return stroke.

[0138] Train the model. The final effect is that the RNSS calculated after the signals of the same return stroke measured at different stations pass through the deep learning model should be consistent with the pseudo label.

[0139] The model of the present invention adopts a branch fusion structure, in which the signal waveform features are extracted by the feature extraction branch ( Figure 4 ), and then concatenated and fused with the remaining features in Table 2, and finally the prediction result is obtained through the attention module and the output layer. The feature extraction branch is composed of an input layer, a convolutional layer, a batch normalization layer, an activation layer, a pooling layer and a residual block stacked in a certain order. Except for the traditional convolutional layer used to increase (decrease) the number of feature channels, all convolutional layers adopt multi-scale convolution with residual blocks.

[0140] All the features in the above step (3.1) are shown in Table 2, including the return stroke waveform, station name, distance from the return stroke to the station, ground wave peak value measured at the station, time, azimuth angle between the return stroke and the station, and propagation speed of the lightning electromagnetic wave. In this example, all features are normalized and limited to [-1, 1].

[0141] Table 2 Features and descriptions used in this example

[0142] Feature Feature size Description Waveform feature 16 Waveform feature extracted by the feature extraction branch Propagation distance 1 Distance from the lightning strike point to the measurement station Propagation speed 1 Propagation speed of the signal in the Earth-ionosphere waveguide Time 1 Time when the return stroke occurs Azimuth angle 1 Azimuth angle from the lightning strike point to the measurement station Peak value 1 Ground wave peak value of the signal Station identification 4 ASCII code converted from the first four pinyin characters of the station name

[0143] For a return stroke, there may be n stations that measure the electric field signal generated by it. Calculate RNSS using power function attenuation and exponential function attenuation respectively, then 2n RNSS are obtained. The pseudo label in the pre-training stage is the mean value of all RNSS of the same return stroke, aiming to make the RNSS calculated after the signals of the same return stroke measured at different stations pass through the model consistent.

[0144] The dataset used in the training process of the above step (3.1) is all return strokes within the detection range, with a total of 6,461,277 return strokes.

[0145] The loss function in the above step (3.1) specifically includes: The loss function of the deep learning model in this example during the training process is composed of an error percentage loss function and an interval loss function, and a distribution loss is set for the prediction result distribution of each batch. The specific formula is as follows

[0146]

[0147] Among them, x is the model prediction result (Δb and ΔL), x d and x u are the upper and lower limits of Δb and ΔL, P and N are the numbers of positive and negative prediction results in a batch, and the purpose is to limit the convergence direction of the model within a reasonable range, especially in the pre-training stage. This is because the scattering is distributed on both sides of the entire attenuation curve (see Figure 3 ). E p is the error percentage loss value; E i is the interval loss function; E d is the distribution loss value.

[0148] Step (3.2), fine-tuning stage: Freeze all layers of the model except the fully connected layer, and train the deep learning model again; the pseudo-labels used in the training process are replaced with the signal ground wave peaks measured by the stations at a distance of 100 ± 10 km from the return stroke;

[0149] The return strokes used in the training dataset are measured by the stations at a distance of 100 ± 10 km from the return stroke, and the return strokes with a signal signal-to-noise ratio greater than 10. At this time, there are 27,271 return stroke data for training.

[0150] Step (3.3), after the training is completed, use the deep learning model trained in step (3.2) for prediction to obtain the offsets Δb and ΔL, and further obtain the accurate normalized electric field peak (RNSS) of all return strokes.

[0151] Step (4), according to the normalized field electric field peak obtained in step (3), calculate the return stroke peak current according to the following transmission line model:

[0152]

[0153] Among them, ε 0 represents the vacuum permittivity, c represents the speed of light, r represents the distance of the return stroke from the station, v represents the return stroke speed, set to 1.3 × 10 8 m / s, E represents the normalized field electric field peak, and I represents the return stroke peak current.

[0154] In order to evaluate and demonstrate the effect of remote estimation of the return stroke peak current, this example sorted out 5,485 return strokes that occurred in the Foshan area, and the distances of these return strokes from the CH station in this example were between 60 and 120 km. This example used the ground wave peak measured by the CH station combined with the relationship between the electric field and the return stroke peak current obtained from artificial triggered lightning in this area in the literature to calculate the return stroke peak current as the true value, and evaluated the accuracy of the remotely estimated return stroke peak current measured by other stations. In addition, the return stroke speed measured by artificial triggered lightning is basically the same as that used in this example. AsFigure 5 As shown, compared with the result obtained by directly remotely estimating the return stroke peak current from the overall attenuation curve obtained in step (2), after being corrected by step (3), the overall intensity of the return stroke peak current has increased by 22%. The coefficient of determination has increased from 0.8362 to 0.9178, and the mean square error has decreased from 21.51% to 10.67%.

[0155] The above has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will also have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A remote estimation method of return stroke peak current based on semi-supervised learning, characterized in that: The steps include: Step (1), using the backstroke on the vertical line of the measuring station, the error correction coefficient of the strength of each measuring station is obtained to complete the correction of the station error; Step (2), using the return stroke within 100km±10km around the measuring station, fitting the overall attenuation curve of the electric field signal generated by the return stroke; Step (3), using semi-supervised learning to predict the offset of the true attenuation curve of each return electric field signal received by each measuring station relative to the overall attenuation curve obtained in step (2), and then obtain the true attenuation curve of each return electric field signal, thereby obtaining an accurate normalized electric field peak value; Step (4), using the normalized electric field peak value obtained in step (3) and the transmission line equation, calculate the return stroke peak current, which is the remote estimated return stroke peak current.

2. The method for remotely estimating return stroke peak current based on semi-supervised learning according to claim 1, characterized in that: The specific method of step (1) is: Step (1.1), select all the return strokes on the perpendicular median of the line connecting the two measuring stations, count the ratio of the ground wave peaks of the same return stroke measured by the two measuring stations, further calculate the mean of all the ratios, and use the mean as the station gain ratio of the two measuring stations; Step (1.2), using step (1.1) to obtain the gain ratio of all measuring stations to the reference station, and taking the gain ratio of each measuring station to the reference station as the initial error correction coefficient of the measuring station; the reference station is required to be a national meteorological station and the surrounding 10km is a plain; The initial error correction coefficient is optimized using the least square method to obtain the error correction coefficient of each measuring station; During optimization, the following relationship must be satisfied: Assume that the gain ratio between station A and station B is k A , the gain ratio between station C and station B is k B , the gain ratio between station A and station C is k C The following relationship should be satisfied: Step (1.3), divide the return electric field signal measured at each measuring station by the error correction coefficient of the corresponding measuring station obtained in step (1.2), so as to correct the signal strength measured at each measuring station.

3. The method for remotely estimating return stroke peak current based on semi-supervised learning according to claim 1, characterized in that: The specific method of step (2) is: For each station, select the return stroke within 100km±10km around the station, divide the ground wave peak value of the return stroke electric field measured by other stations by the ground wave peak value measured by the station, and obtain the peak ratio, which is used as the attenuation ratio y; For each return stroke, any other station M has a set of data (x, y); where x is the distance of the return stroke from the station M, and y is the attenuation ratio calculated from the return stroke and the ground wave peak value measured by the station M; Then, all (x, y) are fitted in the way of power function decay and exponential function decay respectively, and the overall attenuation curve of the ground wave peak ratio of the electric field signal generated by all return strokes with the propagation distance is obtained.

4. The method for remotely estimating return stroke peak current based on semi-supervised learning according to claim 3, characterized in that: The formula for power function decay is: y=ax -b The formula for exponential decay is: Among them, a represents the scaling factor, b represents the attenuation exponent, and L represents the attenuation distance.

5. The method for remotely estimating return stroke peak current based on semi-supervised learning according to claim 4, characterized in that: In step (3), a deep learning model is used to predict the offset of the actual attenuation curve of each return electric field signal received by each measuring station relative to the overall attenuation curve obtained in step (2); The deep learning model includes a sequentially connected waveform feature extraction branch, a fusion layer, a fifth batch normalization layer, a dense block, an attention layer, and a fully connected layer; The waveform feature extraction branch includes a first convolution layer, a first normalization layer, an activation layer of a first Relu function, a second convolution layer, a second normalization layer, an activation layer of a second Relu function, a first multi-scale residual block, a second multi-scale residual block, a third multi-scale residual block, a fourth multi-scale residual block, a maximum pooling layer, a third convolution layer, a third normalization layer, an activation layer of a third Relu function, a fourth convolution layer, a fourth normalization layer, an activation layer of a fourth Relu function, a fifth multi-scale residual block, and a global average pooling layer, which are sequentially connected; The input of the deep learning model is the return stroke waveform, the name of the measuring station, the distance from the return stroke to the measuring station, the peak value of the ground wave measured by the measuring station, the time, the azimuth and the propagation speed of the lightning electromagnetic wave. The output is the offset of the true attenuation curve of each return stroke electric field signal relative to the overall attenuation curve obtained in step (2).

6. The method for remotely estimating return stroke peak current based on semi-supervised learning according to claim 5, characterized in that: The semi-supervised learning method is used for training. The specific method of step (3) is: Step (3.1), pre-training phase: construct a pseudo-label training deep learning model; The output of the deep learning model is the offset Δb and ΔL of the real attenuation curve of each return electric field signal relative to the overall attenuation curve obtained in step (2), where Δb corresponds to the offset of the power function attenuation curve and ΔL corresponds to the offset of the exponential function attenuation curve. The pseudo label calculation formula is: Among them, RNSS represents the peak value of the electric field signal ground wave normalized to 100 kilometers, SS represents the peak value of the return electric field signal ground wave measured by the measuring station; r represents the propagation distance of the electric field signal, that is, the distance of the return stroke from the measuring station; b represents the attenuation index, L represents the attenuation distance; pseudo label represents the pseudo label; n represents the number of measuring stations that detected the return stroke; The training model has the final effect that the RNSS calculated by the deep learning model for the same return signal measured at different stations should be consistent; Step (3.2), fine-tuning stage: freeze all layers of the model except the fully connected layer, and train the deep learning model again; the pseudo labels used in the training process are replaced with the signal ground wave peak value measured by the station 100±10 km away from the return stroke; Step (3.3), after completing the training, use the deep learning model trained in step (3.2) to make predictions to obtain the offsets Δb and ΔL relative to the overall attenuation curve, and then calculate RNSS to obtain the accurate normalized electric field peaks of all return shots.

7. The method for remotely estimating return stroke peak current based on semi-supervised learning according to claim 6, characterized in that: In step (3.2), the return shots used in the training data set are those measured at a station 100±10 km away from the return shot, and the signal-to-noise ratio of the signal is greater than 10.

8. The method for remotely estimating return stroke peak current based on semi-supervised learning according to claim 1, characterized in that: The specific method of step (4) is: The following formula is used for calculation: Among them, ε0 represents the dielectric constant of vacuum, c represents the speed of light, r represents the distance from the return stroke measuring station, v represents the return stroke velocity, E represents the normalized electric field peak value, and I represents the return stroke peak current.

9. The method for remotely estimating return stroke peak current based on semi-supervised learning according to claim 8, characterized in that: v is set to 1.3×10 8 m / s.

Images