Predicting signal blockages in millimetre waves (mmWave) and terahertz (THz) networks is a challenging task that requires anticipating environmental changes. One promising solution is to use multi-modal data, such as vision and wireless inputs, and deep learning. However, combining these data sources can lead to higher communication costs, inefficient bandwidth usage, and undesirable latency, making it challenging. This paper proposes a semantic aware federated blockage prediction (SFBP) framework for a vision-aided next-generation wireless network. This framework uses computer vision techniques to extract semantic information from images and performs distributed on-device learning to enhance blockage prediction. Federated learning enables collaborative model training without exposing private data. Our proposed framework achieves 97.5% accuracy in predicting signal blockages, which is very close to the performance of centralised training. By using semantic information to train models, SFBP reduces communication costs by 88.75% and 57.87% compared to centralised learning and federated learning without semantics, respectively. On-device inference further reduces latency by 23% and 18% compared to centralised and federated learning without semantics, respectively.

Short-term load forecasting (STLF) is essential for efficient power system operations, including unit commitment and energy market operations. Traditional deep learning models like Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks have shown promise but often struggle with capturing complex temporal dependencies and nonlinearities in electricity load data. This paper introduces SpikeNet, a novel neural network architecture incorporating integrate-and-fire (IF) spiking neurons with Recurrent Neural Network (RNN) and LSTM layers to improve temporal dynamics modelling in STLF. Spiking neurons process information through discrete events "spikes", offering enhanced efficiency and the ability to model temporal patterns more effectively. We evaluate SpikeNet's performance against state-of-the-art models on three diverse datasets: Islamabad Electric Supply Company (IESCO), the IEEE data portal, and Panama City. Our results demonstrate that SpikeNet outperforms traditional models, achieving up to 33% improvement in mean absolute percentage error (MAPE) and up to 42% reduction in inference time. These findings highlight SpikeNet's potential as a more accurate and efficient solution for STLF, with significant implications for real-world power system applications.

Human Activity Recognition (HAR) using channel state information (CSI) is crucial for advancing healthcare applications, enabling non-invasive monitoring for improved patient care. However, traditional HAR techniques mainly rely on centralised data processing, which necessitates the sharing of raw data. This approach raises privacy concerns, leading to inefficient bandwidth usage and incurs significant communication overhead and latency, thereby limiting real-time performance and scalability. This paper introduces FedFusionQuant (FFQ), a novel federated learning (FL) framework that integrates advanced signal processing, feature fusion, and model compression during the training process. The proposed framework utilises the federated distance (FedDist) algorithm to adapt parameter adjustments based on neuron dissimilarity measures, effectively mitigating overfitting. Additionally, the incorporation of quantisation-aware training (QAT) allows the model to maintain high accuracy while substantially reducing model size. Extensive empirical evaluations demonstrate the efficacy of FFQ, showing a significant improvement in classification accuracy and energy efficiency through feature fusion and QAT. Furthermore, model compression with QAT achieves a 47% reduction in communication overhead while maintaining accuracy comparable to state-of-the-art methods.