Power Spot Market Supply and Demand Forecasting and Indicator Anomaly Detection Method Integrating Autoformer and Bayesian Optimization

Abstract

Accurate forecasting of electricity spot market supply and demand and timely anomaly detection are essential for intelligent energy management and communication-assisted monitoring systems operating over distributed electromagnetic information networks. To improve prediction accuracy under high-frequency non-stationary fluctuations a nd a ccelerate h yperparameter c onvergence i n multi-source feature spaces, this study proposes a power market forecasting and indicator anomaly detection framework integrating Autoformer with particle swarm optimization and Bayesian optimization (PSO-BO). The Autoformer model exploits trend–seasonal decomposition to capture long- and short-term temporal dependencies, while PSO performs global exploration and Bayesian optimization refines local hyperparameters for efficient model tuning. A Bayesian probabilistic anomaly detection model based on a normal-inverse-gamma prior is further established to quantify residual uncertainty and achieve adaptive online warning. Experimental results on a 24-hour rolling forecasting task demonstrate MAE/RMSE values of 1.23/1.75 GW for load prediction and 1.85/2.53 GW for generation prediction, with anomaly detection accuracy reaching 93.87%. The proposed framework exhibits strong sensitivity to rapid market fluctuations and provides an effective solution for realtime intelligent scheduling and secure monitoring. Moreover, its distributed forecasting and adaptive decision mechanisms offer valuable support for communication-enabled energy systems and electromagnetic sensing infrastructures requiring reliable information transmission and operational awareness.