Deep Learning-Based Prediction of Ice Accretion Thickness Using Multi-layer Perceptron and MultiSource Meteorological Data

This study develops an ice accretion thickness prediction model using a Multilayer Perceptron (MLP) deep learning algorithm based on multi-source meteorological and online icing monitoring data, systematically analyzing the influence mechanisms between ice accretion and meteorological factors. By integrating micro-meteorological and equipment monitoring data, a high-dimensional feature space for ice accretion is constructed. A graded early warning module for icing risk tailored to power grid dispatching is developed, and the meteorological thresholds for different disaster levels are quantified using probability density methods. The results indicate that icing disasters in Yunnan exhibit distinct meteorological threshold characteristics, with the combination of low temperature ($\leq 0^{\circ} \mathrm{C}$), high humidity ($\geq 60 \%$), and weak wind ($\leq 2$ $\mathrm{m} / \mathrm{s}$) being the key conditions for ice accretion. High-risk icing events require stricter thresholds, including humidity $\geq 70 \%$, temperature $\leq 0^{\circ} \mathrm{C}$, and moderate wind speed ($\leq 6 \mathrm{~m} / \mathrm{s}$), while highaltitude terrain significantly exacerbates disaster risks. The study reveals the nonlinear regulatory mechanisms of meteorological factors in the icing process: temperature governs phase transitions, humidity controls moisture supply, and wind speed regulates the rate of ice growth. The research establishes a comprehensive early warning indicator system for icing disasters, providing accurate decision support for power grid anti-icing and disaster mitigation.