**Preface to "Advanced Methods of Power Load Forecasting"**

Advanced societies are characterized by the intensive use of energy produced, distributed and consumed in an uninterrupted, reliable and safe manner. The liberalization of the markets has meant that the entities participating in the electricity system do not have a vertical structure, differentiating in this way between the entities that produce energy from the sellers, all going to an energy pool. Consequently, price competition has increased. In other words, energy systems aim to produce and consume energy efficiently and economically.

A case of special importance is the electricity obtained as a result of a combination of fossil, nuclear and renewable energy sources. The current trend at a global level is to have more and more resources of clean and non-polluting, renewable energy, but with the limitations that this entails, since on many occasions, production does not depend exclusively on human planning, but also on environmental conditions. Both planning, programming and the cost of energy are based on a correct estimate of the electrical load, both in the short term (STLF) and in the medium and long term (MTLF and LTLF).

For the proper functioning of the energy system within today's liberalized energy markets, predicting the energy load has therefore become a crucial task. Improving the accuracy of forecasting energy load, as well as peak loads to ensure energy supply from the energy system to final consumers, has been of increasing interest to researchers in recent years. Responsibility for the system rests with the transmission system operators (TSO). The system works based on the work carried out by the TSO, which manages the transmission of energy in the countries (and the continents, like the ENTSO), and is therefore in charge of making the predictions that the electricity market will use to establish not only the price of energy, but also the units of production.

This concept of forecasting is not exclusive to the TSO. Production units, lower systems such as local and domestic, or any entity participating in the electrical system must have a good load forecasting system. Even industries with high energy needs. The perception of a consumer who is oblivious to electricity prices, and whose consumption is established, has changed radically. High-energy consumers pay much more attention to their consumption and the way they consume. Their studies allow consumption to be improved while accompanying contracts with electricity marketers, with planning according to anticipated energy consumption and needs.

The prediction methodology for power load used in the literature has been advancing over time. From the use of the simplest methods to the most modern ones based on artificial intelligence, always going through statistical methods. Their evolution has always been gradual, and in general with few advances in strategies to deal with the problem of accuracy and safety in predictions. However, as a result of deregulation, the system has radically changed for everyone, from the generator to the consumer. Monitoring consumption and prices is necessary for better management of resources. This has caused an explosion of proposals that try to channel efforts to improve the ways of predicting and achieve more accurate predictions for both the STLF and the LTLF.

The objective of this Special Issue is to present new emerging methodologies that improve the traditional tools used in load forecasting. Artificial intelligence, machine learning, deep learning and hybrid models are some of the new methods that can help improve decision-making in today's energy markets, characterized by high uncertainty and volatility.

The articles presented here are very interesting and innovative. We have made a brief summary of the most important points of all of them.

Cai et al. present a prediction method based on multi-layer stacked bidirectional LSTM for the STLF. The proposed methodology uses a typical Deep Learning structure such as the LSTM neural network. It introduces the combination of two layers, one to compute the hidden vector from front to back, and one in the opposite direction combined, in order to reduce the accumulated error. This structure is repeated in multiple layers connected sequentially that allow information to be filtered and predictions to be made. The efficiency of the predictions is checked using data from an AC power station in southwest China and is measured in terms of MAPE, RMSE and MAE. The results show efficiency around 0.45% of MAPE.

Helsenmeyer and Grzegorzek present an application for electricity load forecasting (STLF) at the NYISO. They use LSTM neural networks with the SESDA architecture (sequential encoder-stacked decoder architecture). The encoder-decoder structure is especially suitable for time series prediction, where the encoder shows a compressed representation of the input information that the decoder will later use to make predictions. The results of the predictions are compared with other methods, and the MAPE reported is 1.52%.

Andriopoulos et al. present a new prediction method based on convolutional neural networks, where the use of statistical methods allow to optimize the obtaining of the hyperparameters of the network. It makes its application to three different types of load (household demand, from a local electricity company in northern Italy, and from an office) with different frequencies. The results show that the use of the LSTM-CNN models improve the results with respect to the LSTM, of proven efficiency.

Trull et al. present Holt-Winters models with discrete moving interval seasonalities (DIMS) applied to a forecast of electricity demand with irregular seasonality. Specifically, it is the demand for a galvanized steel production factory located in Spain. The use of DIMS makes it possible to circumvent existing problems by not having regular seasonality. A comparison is made with other common methods, such as neural networks, ARIMA models, exponential smoothing models. The results show that the prediction efficiency is better than all of them, and that it is at the same level as those made by NARX neural networks.

Almazrouee et al. proposes the use of Facebook's Prophet method for the prediction of long-term electricity load in Kuwait. In his paper, he describes the Prophet method and applies it to electricity demand data from the Kuwait National Control Center. It uses a seasonality of P=365.25 days and uses the holiday function h(t)=Z(t)κ with the matrix of regressors Z(t)=[1(t infoNumberD1),. . . ,1(t infoNumberD L ) ] considering D as the holidays, κ N(0, ˆ2) , where is the holiday smoothing parameter and t measured in days.

Almazrouee et al. uses the Prophet method to determine the maximum long-term peak electrical load, and applies it to the consumption data. It uses 354.25 days as annual seasonality and 7 days for weekly seasonality. He then compares the results with a Holt-Winters model with annual seasonality. The comparison is made using different indicators (MAPE, MAE, RMSE, CVRMSE and R2). In all cases, it is observed that the new proposal improves the results, which are in the order of 1.75% in ASM.

The results shown by the articles presented here contribute to the improvement of electricity load predictions in many areas, and encourage research in all the areas worked on. In addition, the coexistence between modern methods, based on artificial intelligence, with methods considered traditional is verified. With the publication of this Special Issue, a new range of possibilities opens up that the electricity load forecaster can use.

Future research will be more focused on combining the fundamental aspects of traditional

models with artificial intelligence models, creating a symbiosis between both models. One of the fundamental aspects to take into account is the volatility of demand due to the introduction of renewable energies, as well as the need for a fast calculation that can deal with sudden and unexpected changes in the power load. This aspect will be developed in future editions of this Special Issue.

We want to thank all the authors who have contributed to this Special Issue and congratulate them for their good work.

> **J. Carlos Garc´ıa-D´ıaz and Oscar Trull ´** *Editors*
