Intelligent load forecasting of power grid based on Spark-LSTM

To address the dual challenges of massive time-series data processing and complex nonlinear feature capture in power grids, this study proposes a distributed intelligent load forecasting approach grounded on Apache Spark's Long Short-Term Memory Neural Network (Spark-LSTM). By integrating Spark's distributed computing framework with LSTM's deep learning model, we developed a five-layer system architecture covering data acquisition, processing, feature engineering, model training, and prediction services, achieving efficient processing and accurate prediction of terabyte-level power grid data. Experimental results demonstrate that Spark-LSTM achieves 11.5 RMS error, 9.8 average absolute error, and 6.4% average absolute percentage error at a 5-hour prediction horizon, significantly outperforming single-machine LSTM's 13.6, 10.9, and 7.1% in prediction accuracy. The distributed processing time for 1TB data is reduced to 235 minutes, with the acceleration ratio reaching 7.37 when data scale increases. Extreme scenario tests show the model maintains the lowest RMSE and MAPE growth rates, validating its superior performance and stability. The research provides an efficient and accurate load forecasting solution for smart grid, which significantly improves the economy and reliability of power grid dispatching, and has important practical significance for promoting the intelligent transformation of power system.