Performance Assessment of LSTM Networks for Short-Term Load Forecasting Based on Temporal Feature Engineering

Abstract

Accurate short-term load forecasting (STLF) is essential for maintaining grid stability, reducing operational costs, and optimizing energy dispatch within modern power systems. The article discusses Long Short-Term Memory (LSTM) neural network to forecast hourly load and compares the results of the neural network with four popular machine learning tools, namely Support Vector Regression (SVR), Gradient Boosting (GB), Extra Trees (ET), and Random Forest (RF). The relevant temporal features were also engineered using a large-scale, real-world dataset that comprised of over 145,000 hourly observations, and seasonal, weekly, and daily variations. All models were tuned using a structured optimization process and evaluated using standard error and correlation-based metrics. The empirical findings demonstrate that the LSTM model significantly outperforms all competing approaches, achieving an R² of 0.9977 and an error margin below 1% in terms of MAPE, representing a substantial improvement over ensemble- and kernel-based regressors. Visual diagnostic analysis further confirms the LSTM’s ability to accurately reproduce peak and off-peak load behavior while maintaining low predictive dispersion. The results suggest that deep sequence-based models, when properly configured, provide exceptional advantages for STLF applications. The study concludes by outlining opportunities to improve performance through the integration of exogenous features and advanced hybrid architectures.