Enabling Predictive Analytics in the Utilities: Power Generation and Consumption Forecasting

Abstract

In the face of increasing energy demands and the integration of variable renewable energy sources, accurate forecasting of power generation and consumption has become essential for the utilities sector. This paper evaluates the efficacy of various ML and DL models, including XGBoost, Random Forest, and LSTM networks, for predicting power generation from solar and wind sources and power consumption patterns. Our results reveal that the LSTM model achieved a MAPE of 2.7% and a RMSE of 10.1 MW for power generation forecasting, outperforming XGBoost and Random Forest. For power consumption, the hybrid STL + LSTM model achieved a MAPE of 4.1% and RMSE of 15.4 MW, showcasing superior performance compared to traditional methods like ARIMA (MAPE of 6.1%). The comparative analysis also highlighted the models' performance across different time horizons, with the LSTM model consistently providing the most accurate forecasts, particularly in daily predictions with a MAPE of 2.5%. This study underscores the potential of advanced predictive analytics in optimizing energy management and enhancing the reliability of power systems.