3DAEFA-Based Thunderstorm Prediction System With Higher Performance

Abstract

This article presents a thunderstorm prediction system with a 3-D atmospheric electric field (AEF) apparatus (3DAEFA), wherein the data source is the high-resolution 3-D AEF (3DAEF) values. The system is mainly composed of 3DAEF sensor design and calibration, 3DAEF acquisition and processing, and thunderstorm prediction, localization, and imaging. First, a self-made single-axis rotary vane 3DAEFA is employed to measure 3DAEFs. Second, a calibration device is developed to calibrate 3DAEF directions while completing intensity calibration. Considering that the AEF signal (AEFS) is disturbed by low-frequency noises, full-frequency-domain AEFS denoising is conducted through Savitzky–Golay (SG) filtering, and baseline estimation and denoising with sparsity (BEADS). A prediction model is built based on the bidirectional long short-term memory (BiLSTM) network. After inputting denoised AEFS spatial features into the model, which is extracted by a convolutional neural network (CNN), a CNN-BiLSTM model is formed. The proposed system is assessed in different weathers. There is a significant improvement in AEFS’s signal-to-noise ratio (SNR) measured by this system, especially in thunderstorm weather, compared with the original AEFS (OAEFS). Meanwhile, determining coefficients are more than 95%, showing better prediction effects. The deviation between predicted values and real values is smaller than that of OAEFS, which makes it easier to analyze AEFS characteristics. Comparisons with radar charts demonstrate that the proposed system can effectively predict thunderstorms.