Generative AI-based Short-Term Electricity Load Forecasting in 5G Mobile Edge Computing

This work proposes a generative diffusion model-base neural network architecture specifically designed for short-term electricity load forecasting. With the widespread adoption of smart meters, a large volume of high-frequency, noisy, and non-stationary household electricity data is now available for analysis. However, existing load forecasting models still face numerous challenges in handling the complexity of household or power plant electricity data, such as insufficient adaptability, limited generalization capability, and low real-time processing efficiency. Inspired by recent advents of generative artificial intelligence (AI) and mobile edge computing (MEC) developed in 5G mobile communication systems, we try to address these challenges by integrating multimodal feature fusion with time series decomposition to design an innovative framework for temporal decomposition and deep prediction networks. The framework includes a time series decomposition module, a DT module for residual prediction, as well as a DeepAR module, which can effectively capture trends, seasonality and uncertainty factors in the load data. Experimental results exhibit that as compared to traditional methods, proposed model performs well across multiple datasets, demonstrating strong robustness and high predictive accuracy, particularly in handling sudden fluctuations and uncertainties.