电加热防冰系统是防止飞机结冰，保障飞行安全的关键技术之一，其防冰效率直接受加热元件的功率影响。开展电加热防冰系统优化设计的相关研究，对降低防冰系统能耗，提高系统的安全性具有重要意义。针对传统功率分布优化设计中，需多次计算稳态结冰、水膜和温度分布数据，导致耗时长、优化效率低等问题，提出基于改进渐进分层提取（PLE）网络的电加热防冰功率分布优化设计方法，利用多任务的联合学习能力，提高深度学习模型的预测精度。该方法采用本征正交分解对数值计算得到的机翼表面水膜、结冰和温度分布数据进行降维，以功率分布参数作为输入，降维后的模态拟合系数作为输出，以改进的渐进分层提取多任务网络为基础，构建多任务学习模型。该模型对门控网络进行改进，显式添加独有任务专家信息，并在共享专家输出端引入注意力机制，增强模型对特定信息的提取能力。实验结果表明，改进的多任务学习模型在水膜、结冰和温度分布的预测任务中均表现出较高的准确度，在测试集上的均方根误差分别为0.045 8、0.084 8和0.714 9。结合遗传算法对电加热防冰系统功率分布进行单目标优化设计，以最小化总功率为目标，设定约束条件求解出较优的功率分布，与基准结果相比能耗下降约34.3%。数值计算验证结果表明，优化后的功率分布能在降低能耗的同时满足防冰要求。

Electrothermal anti-icing systems are one of the key technologies for preventing aircraft icing and ensuring flight safety， with their anti-icing efficiency being directly influenced by the power distribution of heating elements. Research on the optimal design of such systems is of great significance for reducing energy consumption and enhancing system safety. To address the limitations of traditional power distribution optimization—namely， the requirement of repeated numerical simulations of steady-state ice accretion， water film， and temperature distributions， which result in high computational cost and low optimization efficiency—this study proposes an optimized power distribution design method for electrothermal anti-icing systems based on an improved Progressive Layered Extraction （PLE） network. A multi-task learning framework is constructed by employing Proper Orthogonal Decomposition （POD） to reduce the dimensionality of numerical simulation data， including water film， ice accretion， and temperature distributions on the wing surface. The power distribution parameters are used as inputs， and the POD modal coefficients as outputs. Based on the PLE architecture， the network is enhanced by explicitly integrating task-specific expert information into the gating mechanism and introducing an attention mechanism at the shared expert output， thereby improving the model’s capability to extract task-relevant features. Experimental results demonstrate that the improved multi-task learning model achieves high accuracy across all three prediction tasks， with root mean square errors of 0.045 8 for water film， 0.084 8 for ice accretion， and 0.714 9 for temperature on the test set. Finally， a single-objective optimization of the power distribution is performed using a genetic algorithm， aiming to minimize total power consumption under a set of physical constraints. Compared to the reference scheme， the optimized power distribution achieves approximately 34.3% reduction in energy consumption. Numerical simulation results further verify that the optimized scheme meets anti-icing requirements while significantly reducing power demand.