Multifeature Fusion-Based Thunderstorm Prediction System With Switchable Patterns

Atmospheric electric field signal (AEFS) features can be characterized by their average value (AV), standard deviation (SD), and entropy value (EV). How to mine and fully utilize AEFS features to ensure reliable and efficient thunderstorm detection has not been considered so far. In this article, based on the stacked autoencoder (SAE) and extreme gradient boosting (XGBoost) model, extracted deep-seated features of AEFS are used to obtain its predicted value (PV). It fuses three regular features plus one PV feature and proposes a thunderstorm moving path (TMP) prediction system with switchable patterns among the applied three AEFS prediction models based on the convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM). This fully considers that a single model is difficult to meet AEFS predictions with different weather attributes. Specifically, AEF data measured by a self-made AEF apparatus are adopted to determine feature values (FVs). According to FV intervals (FVIs) in sunny and thunderstorm weathers, the proportion of each feature satisfying FVIs is taken as the weighting factors of corresponding feature terms. A switchable pattern function with different switching conditions is formed by combining weightings and feature variables. Optimal AEFS prediction models are fixed under the same switching condition and applied to corresponding patterns. Empirical results confirm that the proposed system effectively predicts TMPs, with an average determination coefficient of 95.58%. This is the first study to design switchable patterns to detect thunderstorms from a new perspective of multiple AEFS feature fusion, which provides promising solutions to the refinement and intelligent prediction of thunderstorms.

I. Introduction

Ground type thunderstorm monitoring is under rapid development [1], [2]. The 3-D atmospheric electric field (AEF) near the ground is directly related to charges carried by thunderstorm clouds, which is suitable for detecting thunderstorm activities at close ranges. This exhibits better application prospects in early small-scale regional thunderstorm warning [3], [4], [5]. In recent years, thunderstorm warning methods based on 3-D AEF apparatuses (3DAEFAs) are mainly to set AEF thresholds and change rates. Related research involves the principal component analysis, decision trees, region recognition, and other algorithms [6], [7]. However, they generally come with low early warning rate problems. With the rise of artificial intelligence (AI), the dependence on conventional methods becomes limited. This accelerates AEF feature in-depth mining and utilization, and AI-based thunderstorm detection [8], [9]. It is of great practical significance for thunderstorm protection and disaster reduction to establish prediction models for thunderstorm activities in different regions. Particularly, it is advisable to change single models or patterns of most predictions. Nonlinear and nonstationary AEF itself fluctuates greatly, especially in thunderstorm weather. This puts forward higher requirements for prediction activities.