Blade icing in wind farms, especially in high-latitude regions, is a critical issue impacting power generation and safety due to its complex spatio-temporal dynamics. Traditional icing detection methods using physical sensors or model-based approaches often lack sensitivity and accuracy, particularly when dealing with high-dimensional and imbalanced sensor data from wind turbines. To address these limitations, this study proposes the Temporal Attention-based Convolutional Neural Network (TACNN) for improved blade icing estimation. TACNN incorporates a temporal attention (TA) module prior to convolutional layers, utilizing LSTM cells and hidden states to dynamically weigh sensor data and time steps based on their relevance. This mechanism facilitates the extraction of highly informative features and captures crucial temporal dependencies within the data. Benchmarking on ten public multivariate time-series datasets demonstrates TACNN’s competitive performance against established classification techniques. Furthermore, evaluations using real-world wind turbine SCADA data, encompassing operational and environmental parameters, highlight TACNN’s superior accuracy in icing detection, enhanced robustness to temporal variations, and effective handling of data imbalance. Ablation studies confirm the essential role of the TA module, and online estimation tests validate its real-time feasibility. TACNN presents a robust and practical solution for accurate blade icing estimation in diverse wind farm environments, offering the potential for significant reductions in maintenance expenses.