Forecasting day-ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark

Abstract

While the field of electricity price forecasting has benefited from plenty of contributions in the last two decades, it arguably lacks a rigorous approach to evaluating new predictive algorithms. The latter are often compared using unique, not publicly available datasets and across too short and limited to one market test samples. The proposed new methods are rarely benchmarked against well established and well performing simpler models, the accuracy metrics are sometimes inadequate and testing the significance of differences in predictive performance is seldom conducted. Consequently, it is not clear which methods perform well nor what are the best practices when forecasting electricity prices. In this paper, we tackle these issues by performing a literature survey of state-of-the-art models, comparing state-of-the-art statistical and deep learning methods across multiple years and markets, and by putting forward a set of best practices. In addition, we make available the considered datasets, forecasts of the state-of-the-art models, and a specifically designed python toolbox, so that new algorithms can be rigorously evaluated in future studies.

Neural basis expansion analysis with exogenous variables: Forecasting electricity prices with NBEATSx

“…(iii) Time Series Forecasting Comparison: We showcase the use of NBEATSx model on five EPF tasks achieving state-of-the-art performance on all of the considered datasets. We obtain accuracy improvements of almost 20% in comparison to the original NBEATS and ESRNN architectures, and up to 5% over other well-established machine learning, EPF-tailored methods (Lago et al, 2021).…”

“…The Electricity Price Forecasting (EPF) task aims at predicting the spot (balancing, intraday, day-ahead) and forward prices in wholesale markets. Since the workhorse of shortterm power trading is the day-ahead market with its conducted once-per-day uniform-price auction (Mayer & Trück, 2018), the vast majority of research has focused on predicting electricity prices for the 24 hours of the next day, either in a point (Amjady, 2012;Weron, 2014;Lago et al, 2021) or a probabilistic setting (Nowotarski & Weron, 2018). There are, however, also studies on EPF for very short-term (Narajewski & Ziel, 2020), as well as midand long-term horizons (Ziel & Steinert, 2018).…”

“…Out of the numerous approaches to EPF developed over the last two decades (Weron & Ziel, 2020), two classes of models are of particular importance when predicting dayahead prices -statistical (also called econometric or technical analysis) and computational intelligence (also referred to as artificial intelligence, non-linear or machine learning); to provide a comprehensive test ground we will use two well-performing benchmarks, one from each class. Moreover, many -if not most -of the proposed methods are hybrid solutions, that typically comprise data decomposition, feature selection, clustering, forecast averaging and/or heuristic optimization to estimate the (hyper)parameters (Lago et al, 2021).…”

“…Inspired by the success of deep learning (DL) in other fields and the outcomes of recent forecasting competitions (Hong et al, 2016;Smyl, 2019;Makridakis et al, 2020), the EPF community has focused on deep neural networks. The study of Lago et al (2021) provides a comprehensive overview of the deep learning EPF papers published between 2016 and 2019. The consideration is broad and includes various neural network designs ranging from Multi Layer Perceptrons (MLPs; Luo & Weng 2019), Recurrent Neural Networks (RNNs; Ugurlu et al 2018;Zhou et al 2019), to Convolutional Neural Networks (CNNs; Afrasiabi et al 2019), and methods that combine RNNs and CNNs (Kuo & Huang, 2018) or RNNs used to simultaneously forecast price and electric load (Mujeeb et al, 2019).…”

“…Unfortunately, the majority of the neural network EPF related research seems to suffer from some of the following issues: too short and limited to one market test periods, lack of well performing and established benchmark methods, and incomplete descriptions of the pipeline and training methodology resulting in a poor reproducibility of the results. From the abundant literature we have found the empirical studies of Mujeeb et al (2019), Zhou et al (2019) and Lago et al (2021) to be perhaps the most rigorous and complete. Hence, our models will be compared against the highly competitive benchmarks recommended in the latter study: the Lasso Estimated Auto-Regressive (LEAR) model and a parsimonious Deep Neural Network (DNN).…”