Medellín - Ciencias - Maestría en Ciencias - Estadística · 2024
Probabilistic forecasting of electricity demand in Colombia
New approaches have emerged in the field of uncertainty measurement, offering ways to estimate models and their corresponding confidence levels for point predictions. Our first purpose is to compare the predictive capabilities of some models built for forecasting daily electricity demand in Colombia. Initially, we employ generalized linear models, followed by Machine Learning models such as ensemble learning models, support vector machines (SVM), and finally deep learning models. The goal is to determine which model demonstrates superior predictive accuracy in forecasting daily electricity demand in Colombia. In order to evaluate their performance, we mainly use Mean Absolute Percentage Error (MAPE) as a comprehensive measure, which allows us to evaluate their effectiveness in capturing the actual demand values. And also take into account the mean absolute error (MAE) and the root mean squared error (RMSE). Next, we turn our attention on the creation of prediction intervals to handle the uncertainty in our forecasts. We use techniques like Bootstrapping to figure out these intervals. We also incorporate conformal prediction to improve the reliability of our intervals. Our prediction intervals are evaluated primarily based on their coverage percentage. This will allow us to see how frequently our prediction intervals correspond to the actual demand from this data. Through this combination of methods, our goal is to establish a robust and user-friendly framework for forecasting daily electricity demand in Colombia. The results of this development suggest that (1) for the daily energy demand of Colombia, with the variables obtained at a daily frequency, a simple model such as a regularized model works better than an advanced and much more complex model such as a deep learning model. (2) Regarding feature selection concerns, the most important variables are the energy demand lags and demand structure variables for the Lasso model, which works as a feature selection method, due to its regularization nature. This confirms that the inclusion of lags or having an autocorrelated structure is important in this type of problem. Finally, for the forecast intervals, in which we used two methods, the first and most common was the bootstrap method and the second, whose development is more recent, is the conformal Prediction. The construction of our prediction intervals allowed us to give a 99 % confidence level to the point prediction and not just rely on the comparison between the actual and predicted values. (Tomado de la fuente)