Abstract:
The increasing digitalization of electric power systems has created a growing demand for adaptive and spectrum-effcient wireless communication in smart grid environments. Conventional static spectrumallocation approaches may become ineffcient when spectrum occupancy and channel conditions vary over time. This thesis develops an artifcial intelligence-assisted spectrum-management framework for a 6G-oriented smart grid communication scenario by combining short-horizon spectrum-occupancy prediction with reinforcement-learning-based channel allocation. The empirical component of the study is derived from real-world spectrum measurements obtained through the ElectroSense platform. These measurements retain their original frequency-domain interpretation and are not presented as native 6G observations. Instead, they are used to learn temporal spectrum-occupancy behaviour, while an assumed 8 GHz carrier-frequency scenario and an effective communication bandwidth of 100 MHz are introduced separately for communication-level evaluation. A compact feature set consisting of empirical center frequency, received energy, short-term energy variation, rolling mean energy, and rolling energy variability is used to predict the occupancy state of the subsequent chronologically ordered observation. Support Vector Machine, Random Forest, Decision Tree, and K-Nearest Neighbors are evaluated under the same sequence-aware chronological train–test procedure. Random Forest achieves the strongest overall predictive performance, with an accuracy of approximately 91.53%, a probability of detection of 93.61%, and an F1-score of 88.59%. The Random Forest prediction output is subsequently used to generate candidate spectrum resources for Q-Learning and Deep Q-Network based channel-allocation mechanisms. Under the defned simulation assumptions, DQN achieves approximately 157.48 Mbps throughput, 3.405 ms communication latency, a reliability/PDR indicator of 0.7348, and a spectral effciency of 1.5748 bit/s/Hz. These results represent a modest improvement over Q-Learning, although DQN requires substantially greater training complexity. The study therefore demonstrates the feasibility of integrating empirical next-observation spectrum prediction with intelligent channel allocation under a clearly separated 6G-oriented communication scenario. The reported communication metrics should be interpreted as comparative simulation indicators rather than measurements from a deployed 6G smart grid network
