Abstract
Dynamic Trees are a tree-based machine learning technique specially designed for online environments where data are to be analyzed sequentially as they arrive. Our purpose is to test this methodology for the very first time for Electricity Price Forecasting (EPF) by using data from the Iberian market. For benchmarking the results, we will compare them against another tree-based technique, Random Forest, a widely used method that has proven its good results in many fields. The benchmark includes several versions of the Dynamic Trees approach for a very short term EPF (one-hour ahead) and also a short term (one-day ahead) approach but only with the best versions. The numerical results show that Dynamic Trees are an adequate method, both for very short and short term EPF—even improving upon the performance of the Random Forest method. The comparison with other studies for the Iberian market suggests that Dynamic Trees is a proper and promising method for EPF.
Highlights
Electricity has become an indispensable commodity for all societies
The numerical results show that Dynamic Trees are an adequate method, both for very short and short term Electricity Price Forecasting (EPF)—even improving upon the performance of the Random Forest method
Due to the fact that Dynamic Trees are designed to work in online environments, its approach to the problem of forecasting differs from those Artificial Neural Networks (ANN), Support Vector Machines (SVM) or Random Forests (RF) as the nature of these three last techniques forces them to work in batch mode, that is, each time we want to forecast we have to train a whole new model with previously available training data
Summary
Electricity has become an indispensable commodity for all societies. Commonly, its price is traded in a market in which, due to the deregulation process and the existence of competitive markets, the price of electric power fluctuates according to consumer demand and the offers of the different producers, so the prediction of that price has become a need, for energy companies (such as producers, power transmission operators and retailers) and for all kinds of market participants (such as investors or traders) and, the final consumers. In our work we want to use a machine learning technique that, as far as we know, has never been used to face this problem, the so-called Dynamic Trees [11]. Due to the fact that Dynamic Trees are designed to work in online environments, its approach to the problem of forecasting differs from those ANN, SVM or RF as the nature of these three last techniques forces them to work in batch mode, that is, each time we want to forecast we have to train a whole new model with previously available training data. Many more, are typical approximations to the EPF problem that provide good results, specially when combined with other machine learning techniques [17,18,19,20,21,22,23], but, again, we just want to know if the raw Dynamic. We shall describe the tests performed with the two machine learning methodologies and we shall discuss the results obtained
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
