Multidimensional electric power parameter time series forecasting and anomaly fluctuation analysis based on the AFFC-GLDA-RL method

With the advancement of big data and artificial intelligence, the intelligence level of power systems has significantly increased. Accurate prediction of power parameter time series fluctuations is crucial for the stable operation and efficient management of these systems. Traditional methods face challenges in capturing long-term dependencies and multi-scale features of multi-dimensional variables in nonlinear, non-stationary time series, limiting prediction accuracy. To address this, we propose an Adaptive Feature Fusion Convolution and Global-Local Dynamic Attention Reinforcement Learning (AFFC-GLDA-RL) model. The AFFC module enables multi-scale feature fusion and extraction from multi-dimensional variables, while the GLDA mechanism and RL strategy provide dynamic feature weighting and adaptive optimization, enhancing model focus on key features and robustness to noise. Additionally, a z-score-based anomaly detection method is incorporated to identify and analyze anomalous fluctuations in power parameter sequences. Experimental results demonstrate the model's superior performance in forecasting periodic power parameters, achieving R2 values above 0.99 and RMSE below 0.012. For more volatile parameters, the model maintains high accuracy with R2 above 0.85 and RMSE below 0.06. Furthermore, the model demonstrates strong predictive accuracy for 24-h ahead power parameter forecasts, providing robust support for stable operation and load forecasting in power systems.