Accurate identification of residential air conditioning load is essential for leveraging their regulation potential and enabling effective demand response. To overcome the limitations of existing residential air conditioning power estimation methods, which often suffer from insufficient accuracy and high computational complexity, this paper proposes a novel non-intrusive neural network model that combines a variational autoencoder (VAE) with an enhanced efficient channel attention (ECA) mechanism. The improved ECA incorporates a dual-pooling strategy-combining global average pooling and global maximum pooling-to capture rich statistical information while highlighting prominent local responses. Additionally, a compression-reconstruction mechanism is introduced: after dimensionality reduction, fast dynamic convolutional kernels adaptively model local channel interactions, focusing on key features and assigning appropriate channel weights. This enhanced ECA module is embedded within the VAE decoder to improve feature reconstruction for air conditioning load estimation. Furthermore, a multi-task learning framework jointly optimizes power disaggregation and state recognition tasks, promoting effective information sharing and complementarity to boost overall identification accuracy. An output module with post-processing state threshold constraints is employed to suppress interference from non-air conditioning loads. The proposed model is validated on real-world residential electricity datasets, showing a mean absolute error (MAE) reduction of 59.71% and 9.22%, respectively, compared to two baseline models across three regions, while achieving an air conditioner state recognition F1 score of 84.58%. Ablation studies reveal that the improved ECA contributes to MAE reductions of 56.23% and 12.47% in two regions, and the multi-task learning framework further improves identification accuracy by 3.17% and 5.90%. Moreover, the minimally intrusive measurement approach-training with intrusive data from only 30% of users-significantly reduces reliance on extensive user data while maintaining high accuracy, demonstrating strong potential for practical deployment.