Online prediction algorithm for power data deviation extremes based on time series mining algorithm

Abstract The integration of large-scale distributed power sources complicates grid topology and intensifies spatiotemporal coupling in power data. This complicates the analysis of intrinsic correlations and makes predicting deviation extremes more challenging. Traditional models, which rely on fitting historical data, struggle to measure similarity between power sequences in high-dimensional spaces under real-time operating conditions, leading to significant prediction errors. To address this, an online prediction algorithm based on time series mining is proposed. Real-time operational data are acquired and processed using piecewise linear approximation and polynomial fitting to reduce dimensionality. The Minkowski distance is then applied to measure similarity between sequences in the reduced space, enabling accurate pattern analysis. Subsequently, a Long Short-Term Memory (LSTM) network, optimized with a genetic algorithm, captures subtle fluctuations and deviation extremes in the data. The LSTM updates its internal state dynamically, facilitating real-time online prediction. Experimental results demonstrate high accuracy, with high R² and correlation coefficient (CC) values, indicating excellent model fitting and effective prediction of power data deviation extremes.