Predicting ice accretion on wind turbine blades in high-latitude regions is of paramount importance. Traditional centralized approaches for icing prediction are often hampered by substantial storage and computational demands. Furthermore, multivariate time-series datasets acquired for this purpose frequently exhibit imbalance and redundancy. Moreover, conventional aggregation strategies in federated learning can lead to protracted convergence times for the global model to reach optimality. To overcome these challenges, this paper introduces FedBIP, a federated learning (FL) methodology. FedBIP integrates a feature selection mechanism to pinpoint the most informative features and an over-sampling technique to mitigate the impact of data imbalance. A serial architecture, comprising long short-term memory (LSTM) networks and a three-layer densely connected convolutional neural network (CNN), is developed to simultaneously capture temporal dynamics and latent representations. The proposed aggregation approach considers the significance of data volume, timestamps, and offsets for each participating client, effectively addressing issues arising from inactive or underperforming clients with heterogeneous data distributions. In addition, a knowledge distillation (KD) mechanism is incorporated into the local client training phase to bolster model performance and accelerate learning. Empirical evaluations reveal that FedBIP surpasses the strongest baseline model, achieving a 1.67% improvement in \(F_2\) -score and a 0.95% enhancement in balanced accuracy (BA). When applied to the over-sampled dataset, the proposed aggregation strategy yields improvements of 1.85% in \(F_2\) and 1.08% in BA compared to the best baseline. Notably, the integration of KD reduces the communication rounds required to reach the target BA in FedBIP by 5.2%.