Machine learning models for forecasting power electricity consumption using a high dimensional dataset

We use regularized machine learning models to forecast Brazilian power electricity consumption for short and medium terms. We compare our models to benchmark specifications such as Random Walk and Autoregressive Integrated Moving Average. Our results show that machine learning methods, especially Random Forest and Lasso Lars, give more accurate forecasts for all horizons. Random Forest and Lasso Lars managed to keep up with the trend and the seasonality for various time horizons. The gain in predicting PEC using machine learning models relative to the benchmarks is considerably higher for the very short-term. Machine learning variable selection further shows that lagged consumption values are extremely important for very short-term forecasting due to the series high autocorrelation. Other variables such as weather and calendar variables are important for longer time horizons.