Background: Accurate real-time forecasting of energy demand and electricity prices is at the core of modern smart grid infrastructure. Traditional statistical techniques, such as ARIMA and exponential smoothing, fail to capture the nonlinear, multivariate dynamics that are inherent in today’s power systems.
Objective: In this paper, a hybrid machine learning and deep learning forecasting framework based on the integration of the XGBoost and LSTM-based architectures is proposed for improving the prediction accuracy of energy demand and electricity price in the smart grid environment.
Methods: Temporal and weather-driven feature engineering, principal component analysis (PCA) and sliding window sequence generation were used to preprocess historical electricity consumption, market price and meteorological datasets. The evaluated models are: XGBoost, standard LSTM, Stacked LSTM and Encoder-Decoder LSTM. Performance was evaluated using RMSE, MAE and R².
Results: The Encoder-Decoder LSTM achieved the lowest RMSE of 7.12 MWh and highest R2 of 0.968, which outperforms the ARIMA baseline (RMSE = 18.74) and standalone XGBoost (RMSE = 11.23). Weather and temporal features were key to the accuracy improvements.
Conclusion: The proposed hybrid framework shows the excellent forecasting ability and practical utility for intelligent energy management in smart grids. Future works should also consider real-time deployment and transformer-based extensions.