1. Introduction
Lightning is an electrical discharge phenomenon that occurs in nature between clouds (C-C flash) or between clouds and the ground (C-G flash). Along with the continuous development of economy and rapid progress of society, the personal safety and property damage caused by lightning has attracted people’s attention. Lightning warning, as an important measure in active lightning protection, is of great significance to reduce the harm caused by lightning [
1,
2,
3,
4].
Generally, the equipment used to detect lightning are atmospheric electric field meters, weather radars and lightning locators [
5]. Moon et al. [
6] used machine learning to generate binary predictions of lightning occurrence within a specific location and time interval based on weather variables from the European Centre for Medium-Range Weather Forecasts and compared the results with lightning reports from a region including the Korean Peninsula and found equitable threat scores of 0.0885 and 0.0828 for support vector machines and random forests, respectively. Mostajabi et al. [
7] developed a four-parameter model based on four common surface weather variables (air pressure at station level (QFE), air temperature, relative humidity, and wind speed) and validated it using the data from lightning location systems. The evaluation results show that the model has a fairly high predictive capability for lead times of up to 30 min. Gharaylou et al. [
8] utilized idealized Weather Research and Forecasting explicit charging/discharge module (WRF-ELEC) to simulate two types of thunderstorm clouds to investigate the influence of tilting effect on charge density and lightning flash density. A new idea for inversion of the charge structure of airborne thunderstorms based on numerical weather models is provided. Yang et al. [
9] proposed a thunderstorm identification method combining the area of graupel distribution region and weather radar reflectivity, 312 thunderstorms from 17 weather processes in Nanjing, China, were tested for identification, and the optimal identification parameters were obtained, which can gain probability of detection of 91%, false alarm rate of 6.9% and critical success index of 85.3%, providing an effective means for thunderstorm nowcasting.
However, the atmospheric electric field meter is favored by the majority of weather forecasters due to its relatively low price (a few tenths of weather radar) and easy installation [
10]. During the development of a thunderstorm, ice crystals, shrapnel, and other particles in the clouds are constantly charged by friction, resulting in a strong electric field between the atmosphere and the ground. Therefore, the atmospheric electric field meter, as a measurement device reflecting the most fundamental cause of lightning flashes, can be used to deduce the process of thunderstorm formation, development and dissipation.
With the above principles in mind, scholars at home and abroad have conducted in-depth research on it, such as the use of atmospheric electric field amplitude, atmospheric electric field differential threshold, electric field fast-varying jitter, and other characteristics, combined with lightning locators, weather radars, or other devices to predict the occurrence of lightning. The WT (Wavelet transform)-LSSVM (Least squares support vector machine) method was used by Zhang et al. [
11] to develop a prediction model for the atmospheric electric field time series. It is important for indicating climate change and also plays an important role in lightning forecasting. Xing et al. [
12], in order to accurately obtain the location of thunderstorm clouds, established an electric field measurement model and proposed a thunderstorm cloud localization algorithm based on three-dimensional atmospheric electric field, and the results show that the method has a good localization performance with a ranging error rate of about 5% and a direction-finding error rate of about 3%. Wang et al. [
13] developed an intelligent lightning warning system (LWS) based on electromagnetic field and artificial neural network in order to improve the lightning prediction accuracy. The neural network was constructed using the change rate of electric field, temperature, and humidity acquired 2 min before the lightning strike, and the rationality of the proposed model was verified based on lightning strike observation and prediction for up to six months.
Most of the existing methods for lightning warning based on atmospheric electric field do not deeply explore the characteristics of atmospheric electric field oscillation, and there are problems of poor applicability and low accuracy of warning. In practical application, due to the complex and changing situation of the surrounding environment, the change of the atmospheric electric field presents nonlinear oscillation characteristics. More importantly, the measurement range of a single atmospheric electric field meter is limited (about 15 km), and it is a difficult task in the field of lightning prediction to use multiple stations for joint prediction to improve the warning performance and expand the warning range.
Based on the above shortcomings, combining the improved ResNet50 model and MLP neural network, a lightning spatio-temporal localization method is proposed. First, SAE is used to extract the features of the electric field temporal data from multiple sites to obtain their representations in low dimensions. Then the above extracted features are composed into visual images and fed into the improved ResNet50 model for recognition, resulting in the discriminative result of whether lightning will occur. Finally, if lightning will strike in the future, the center of future lightning flashes is predicted based on the MLP model, and the approximate area where lightning is likely to occur is known. The above methods passed accuracy and robustness tests and were compared with other methods. The results show that it can give more reliable lightning warning results.
The remainder structure of this paper is organized as follows:
Section 2.1 describes the data sources and the method of database formation.
Section 2.2,
Section 2.3 and
Section 2.4 list the feature extraction and the lightning spatio-temporal localization methods in detail.
Section 3 contains the experimental validation, and the excellent performance of the proposed model is fully demonstrated by a series of experiments with the constructed weather sample database.
Section 4 describes three case studies of the system in actual operation.
Section 5 concludes the paper.
4. Discussion
The construction of lightning warning system is basically completed after the above methods passed the accuracy and robustness tests. The following are a few case studies of the actual operation of the model. These three cases are three presentations of the complete life cycle of thunderstorms. Since the lightning flashes are too intensive to be easily read if they are all put in one figure, a complete life cycle of a thunderstorm is divided into three figures for display, which can better reveal the prediction performance and the pattern of thunderstorm evolution.
As shown in
Figure 21a, Case 1 is a thunderstorm that occurred on 2 May 2021 from 13:37 to 14:07. The model can first give a judgment whether there will be flashes based on the electric field measurements in the past hour, and then the location of the lightning center.
Table 4 shows the calculated results given by the system per minute and the comparison with the real values. According to Equation (
13), the points of all lightning flashes are equated to a point called the geometric center (blue numbers in
Figure 21), and the location that the system predicts lightning to occur is called the prediction point (red numbers in
Figure 21). It is worth noting that that lightning may not occur every minute, nor does the system make the determination that there will be lightning each minute. As described in
Table 1, there will be four cases (TP, FP, FN, TN) based on the predicted values and the actual observations. In this example, 13:37 is taken as the first minute of the start and the number of lightning flashes that occurred during each minute is given. The geometric center, the predicted point and the distance between the two are also listed. For example, in the first minute (13:37) there were two lightning flashes detected, the geometric center is (113.37, 23.51), and the predicted point is (113.32, 23.41), and the distance between them can be calculated by the following equation.
where
a and
b are the difference in longitude and latitude of the two points, respectively;
and
represent the latitudes of the two points;
R is the radius of the Earth, which is about 6378 km.
According to Equation (
29), the distance between the prediction point and geometric center at 13:37 is about 12.65 km, indicating that the system does indicate the approximate location of lightning flashes. At the same time, from the whole thunderstorm process, the accuracy of the prediction declined for the small number of flashes with more scattered distribution (“1”, “3” in
Figure 21a), which may be due to the low charge of the thunderstorm clouds and the relatively weak discharge phenomenon, resulting in the electric field meter not detecting the intense fluctuation changes. However, normally, the horizontal range of a single thunderstorm is about a few kilometers to 20 km [
35], and the average motion speed of a thunderstorm is 13.22 m/s [
36]. The model proposed in this paper is able to give predictions at every minute, which means that the location and motion of discharge areas can be inferred from the predictions made a few minutes before and after. As can be seen from
Table 4, the distance between the predicted points and the geometric centers are mostly in the range of 5–15 km. Combined with
Figure 21a, even if the model’s prediction deviates widely at a certain moment, it is still possible to know the area where lightning will occur by subsequent predictions.
Case 2 occurred on 10 June 2021 from 15:29 to 15:59. As shown in
Figure 21b, it can be clearly seen that the flashes were more concentrated. From 15:29 to 15:41, the number “1” represents the lightning occurred in the first minute. From
Table 5, we know that only two flashes occurred, and the distance between the predicted point and the geometric center is 16.9 km. However, during the periods of 15:42–15:51 and 15:52–15:59, most of the subsequent flashes were found to be concentrated in the square area of longitude (113.2, 113.3) and latitude (23.3, 23.5). From the spatio-temporal localization results, it is obvious that the model is able to locate lightning flashes, and a relatively large locational error at a given moment does not affect the overall determination of the massive discharge area.
Presented in
Figure 21c, Case 3 was recorded on 17 August 2021 from 18:32 to 19:02. At 18:32–18:43, the distribution of lightning was more scattered, but was mainly found in the northeastern part of the Conghua district. However, the subsequent lightning events were more concentrated (18:44–18:53), when there may have been a merger of multiple thunderstorm clouds [
37,
38,
39]. In addition, as can be seen from
Table 6, at this time the thunderstorm weather was particularly intense, with up to 45 lightning flashes per minute. Finally, at 18:54–19:02, the distribution of lightning was more dispersed, implying that the thunderstorm clouds started to develop toward the dissipation process. Thus, the algorithm proposed in this paper can not only locate lightning in time and space, but can also judge the process of lightning formation, development and dissipation based on the results of multiple warnings, which is of great significance for grasping the mechanism of thunderstorms and lightning protection.
Sometimes, during the development of a thunderstorm, ice crystals, shrapnel, and other particles in the clouds are constantly charged by friction, resulting in a strong electric field between the atmosphere and the ground. Although the electric field meter detects a strong electric field at this time and predicts that lightning will occur next, no lightning actually occurs at this time. However, this is usually reflected in the actual observations over the next few minutes. Similarly, when lightning occurs, the charge in the clouds is released, so the electric field measured on the ground drops accordingly, which may cause the model to give a result of no lightning occurring. However, in general, as shown in
Table 3, the model is able to maintain an accuracy of more than 80%.
5. Conclusions
In this paper, a deep learning framework for lightning prediction is proposed. First, the features of the original EF time series data are extracted using a SAE. Then, the above extracted features are used to construct a visual picture of EF measurement data of multiple stations. Next, the weather samples are classified based on the improved ResNet50 model to conclude whether a thunderstorm will occur in the next one minute. Finally, if it is judged that lightning will occur in the future, the center of lightning flashes is predicted based on the MLP model. In practice, the model can determine the trend of lightning with multiple predictions made in a few minutes, while also improving fault tolerance.
The composition of lightning prediction model is divided into two parts, i.e., making predictions about the time and location of lightning occurrence, which were tested for accuracy and robustness respectively. Due to the chaotic nature of lightning, it is challenging to forecast it. Taking into account the computational cost and prediction accuracy, more reliable lightning spatio-temporal localization results are given and analyzed with practical cases. The experimental results show that the proposed model has statistically quite high prediction capability for short-time proximity warning, which helps to improve the lightning protection level in the region, and active lightning protection measures can be taken to further reduce the property and safety losses caused by lightning according to the relevant results.