The dynamics of the growth of electricity consumption have been maintained in the world for more than 30 years (Figure 1). There are no prerequisites for reducing electricity consumption in the future, since at the present stage of human development, electricity is a key resource—professional and household human activity are impossible without the use of electricity. According to statistics on the world energy and climate portal Enerdata for 2021, electricity consumption in that year amounted to 24,877 TWh, which is 5.5% and 4.8% more than in 2020 and 2019, respectively. The growth of electricity consumption is also confirmed by statistics in the field of global electrification of final consumption. The trend towards an increase in electrification in the world continues to be traced: in 2021, the indicator reached 20.4% (+1 point compared to 2019).

According to Enerdata’s global energy and climate data, in 2021, Russia was among the top five countries in terms of electricity consumption. The highest electricity consumption was in China (7714 TWh), followed by the USA (3869 TWh), India (1355 TWh), Russia (963 TWh TWh), and Japan (916 TWh). According to the electricity consumption schedule for the period from 1990 to 2021, shown in Figure 1, there is a noticeable trend in the growth of electricity consumption in the world.

In Russia, from 2017 to 2021, electricity consumption increased by 58 TWh. The following dynamics of electricity consumption can be observed in a number of countries. For instance, in China, the consumption increased by 1834 TWh; in the USA, it decreased by 18 TWh; in India, there was an increase of 206 TWh; in Japan, consumption reduced by 64 TWh. The share of using electricity as an energy source is growing, and in 2021, it made up 10% of the world consumption of all types of energy sources (29% oil, 24% natural gas, 27% coal, 10% biomass).

The dynamic pattern of electricity consumption in the five countries with the highest electricity demand from 2000 to 2021 is shown in the Figure 2.

In this regard, energy saving issues are extremely important to maintaining the uninterrupted operation of electric power systems and providing consumers with electricity of proper quality. Energy saving refers to a set of measures for the rational use of energy resources, increasing the share of renewable energy sources in total electricity production, and other measures aimed at reducing the use of energy resources and contributing to solving environmental problems. Without the availability of energy-saving measures, it is not possible to manage the growing demand for electricity every year. To ensure the uninterrupted operation of electric power systems (EES), it is necessary to maintain a balance of power and consumption in the EES. This implies the need to ensure the proper level of frequency and voltage in the EES. With sudden changes (increase or decrease) in electricity consumption, there is a violation of the balance of power and electricity consumption, which leads to failures and accidents in power plants. In addition, it is impossible to operate the wholesale electricity and capacity market without managing power consumption modes. Therefore, the management of operating modes in the EES is a complex task, and for its effective solution, detailed planning of electricity consumption is necessary.

One of the possible solutions to the problem of load planning of electric power systems is the forecasting of electricity consumption. The presence of a reliable forecast of electricity consumption contributes to the validity of decision making when managing the operating modes of power facilities.

Managing the process of electrical energy consumption is efficient due to the functioning of various incentive mechanisms that operate on the wholesale electricity and power market (WECM). This operates on the mechanism of economic management of consumer demand, known as the “Demand response”, or DR. This mechanism includes a set of measures to reduce electrical energy consumption, including during peak hours, thereby contributing to a uniform and more efficient use of the capacities of generation facilities. DR can be referred to as the technology of price-dependent consumption, which implies the influence of consumers on the demand and electricity price in different periods of time (days). Therefore, for example, during peak hours on the WECM, consumers are offered lower-price electricity in return for reducing electricity consumption. The WECM uses such tools as a balancing market, a day-ahead market, bilateral agreements, competitive power take-off procedures, etc. The conditions created on the WECM encourage consumers to switch to economically advantageous conditions, assuming the existence of an accurate plan for electricity consumption. Therefore, forecasting is an urgent and important task whose solution will allow WECM participants (electricity buyers) to receive the opportunity to purchase electricity at favorable rates. The forecasting of electricity consumption by WECM participants in particular and the WECM’s functioning as a whole contribute to the balance between supply and demand for the WECM and, as a result, improve the efficiency of managing the process of electricity production and consumption.

It is worth noting the key features of the electricity market, which are the difference between electricity and other types of goods and services. Firstly, the process of generating and transmitting electricity is a complex technological process, and the final consumer does not know electricity’s cost as a commodity. Secondly, the transmission of electricity through the network occurs in accordance with the laws of electrical engineering; the logistics of electricity are different from those of other goods and services. Thus, electricity is a unique product, the rational use of which affects all spheres of society.

2. Operative Forecasting

Operational forecasting is the research subject of many scientists; it touches upon the issues of operational management of the operating modes of power facilities. The day-ahead market provides for hourly differentiation of the electricity tariff within a day, so intraday forecasting is especially important when the WECM operates. The papers in Refs. [26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44]^{[1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19]} present approaches to solving the forecasting problem through various methods. In Ref. [26]^[1], hourly forecasting is presented using an artificial neural network of direct propagation (perceptron) with reference to the regional power system. As a result of a series of experiments, the authors chose a three-layer perceptron with a sigmoidal activation function as a network configuration. In total, 24 variables were fed to the network input as retrospective data of the electric power load. The hidden layer contains five neurons; the output layer contains one predicted value Ref. [26]^[1]. A detailed description of the network configuration and learning algorithm is given in Ref. [26]^[1]. The authors Ref. [26]^[1] developed adaptive feedback integrated into an artificial neural network model, which made it possible to reduce the root mean square prediction error by about 1.5%. In Ref. [27]^[2], an approach was proposed that uses a long short-term memory (LSTM) network using the “Butterfly” optimization algorithm and preliminary feature normalization. The algorithm developed in Ref. [27]^[2] was tested by the authors on two sets of data from the electrical energy consumption by households. Using the metaheuristic approach, a more accurate forecast was built compared to using the linear regression algorithm, support vector regression (SVR), and neural networks of various architectures—bi-directional long short-term memory (Bi-LSTM), LSTM, and convolutional neural network (CNN)—for both datasets. The average absolute forecast errors were 9% and 5%, respectively, which allows uresearchers to consider this method to be effective. It should be noted that the method proposed in Ref. [27]^[2] is also productive from the point of view of the forecast execution time since it allows the creation of a forecast faster than any of the algorithms considered by researchers. The use of artificial neural networks made it possible to predict the peak load of the building of justice in the United States Ref. [28]^[3]. The researchers developed an artificial neural network, the use of which made it possible to obtain a fairly accurate forecast of peak electricity consumption (the average absolute percentage error was 3.9%). The developed model was compared to classical approaches to modeling intraday load curves: moving average, linear regression, and a multidimensional adaptive regression splines (MARSplines) model. The study Ref. [29]^[4] presents the results of implementing the system project for monitoring and predicting peak loads of mining industry electricity. During the implementation of this project, the following results were achieved. The electricity consumption at the selected site (the Ben Guerir quarry, Morocco) was studied using machine learning tools. Several power-demand-forecasting models based on historical data were created (artificial neural network model, neuro-fuzzy inference time series model, SVM support vector regression model, and fast forest quantile regression), among which the best forecast results were obtained using the forest fast quantile regression model (FFQR). In the considered project, a new infrastructure for the energy management system for a mining facility (quarry) is proposed, from which data on the quality of electricity and the state of the electrical network are obtained in real time. The authors note that such infrastructure will make it possible to apply methods for optimizing and planning loads on the energy system, promote decentralized electricity production using renewable sources and energy storage systems, and simplify energy audit procedures. The implementation of the project will allow the mining enterprise to increase productivity through accurate planning of electricity consumption and improve the efficiency of production management as a whole. The authors note the need for additional research, including the study of various neural network architectures for predicting peak loads and comparing them with the fast forest quantile regression model already developed in Ref. [29]^[4]. Of particular interest are the works in which the study of flexible approaches necessary for real-time forecasting models was carried out Refs. [40,41,42,43]^{[15][16][17][18]}. WResearchers decided to attribute them also to operational methods in accordance with the classification by lead period described earlier. Thus, the paper Ref. [40]^[15] describes an approach that makes it possible to improve the process of managing electricity demand by introducing optimization procedures for the aggregator of small prosumers with participation in the energy market. Two optimization procedures are proposed in Ref. [40]^[15]: a two-stage stochastic optimization model for determining supply and demand and optimization of the predictive control model (MPC) for managing aggregated flexible loads in real time. Under conditions of uncertainty, the proposed strategies (called smart) surpass other theoretical guidelines, such as deterministic, flexible, and inflexible approaches. The two-stage stochastic optimization model increases the reliability of energy applications by taking into account uncertainty and flexibility in the optimization process. This reduces the costs to the aggregator, particularly the cost of regulation. Table 1 summarizes the analysis of the methods used by researchers to solve the problem of operational forecasting of power consumption. Based on the analyzed works, it can be concluded that neural networks are used for operational forecasting. Machine learning algorithms, including deep learning, are used to train these networks. The following neural network architectures were used in the analyzed works: multilayer perceptron in Refs. [26,28]^[1][3] and recurrent LSTM networks with the Butterfly hyperparameter optimization algorithm in Ref. [27]^[2]. According to the researchers, the developed models of forecasting electricity consumption based on neural networks allow us to build a more accurate forecast compared to traditional approaches to forecasting electricity consumption—linear regression, moving average models, and others. If we compareing them to each other, then recurrent LSTM networks are a more optimal option since they cope much better with a large amount of data (require fewer computing resources), have a memory mechanism, and effectively cope with the problem of gradient attenuation. Using this architecture, a forecast can be obtained faster than by using a perceptron. However, it should be emphasized that for each specific task, the process of choosing the optimal architecture and selecting hyperparameters of the neural network model is individual. Therefore, the adaptation of the models recommended in the literature takes into account the composition and structure of the source data. Thus, the use of a fully connected neural network is recommended for intraday forecasting of power consumption in the regional dispatching department Ref. [26]^[1]. To predict the power consumption of individual objects (including several administrative buildings Ref. [28]^[3] or households Ref. [27]^[2]), the use of both perceptron and recurrent LSTM networks is recommended. It is worth noting that these recommendations do not exclude the possibility of using other regression algorithms in forecasting. For example, the researchers in Ref. [29]^[4] suggest using the fast forest quantile regression algorithm, which allows one to obtain a reliable forecast of electricity consumption for the mining industry with a developed infrastructure for monitoring electricity consumption. At the same time, the authors of Ref. [29]^[4] consider the construction of neural network models and the comparison of their prediction accuracy with the developed FFQR model to be a promising study. It can be argued that the studies reviewed confirm the effectiveness and prospects of using neural network algorithms since they often allow us to better solve the problem of operative predicting power consumption. Thus, all the methods used for operational forecasting can be conditionally divided into time series forecasting methods and probabilistic forecasting methods. As a result of the analysis of the use of various methods for operational forecasting, it was revealed that in most cases, the use of artificial neural networks is an effective solution to the problem of operational forecasting of power consumption. Moreover, of the large diversity of different neural network architectures, perceptron and recurrent neural networks (including LSTM) are usually used. The nonlinearity of the time series of power consumption is approximated fairly accurately in neural network models, and this distinguishes them as a powerful and most accurate method of operational forecasting of power consumption. However, it is worth considering the fact that neural network models are poorly interpreted. That is, with a sufficiently accurate forecast, it is difficult to explain exactly how the model builds it. The construction of surrogate models based on classical well-interpreted algorithms for neural networks is one of the significant prospects for further research in this area. On the other hand, probabilistic forecasting is sometimes a more reliable solution than neural networks, since it allows you to make a forecast taking into account the estimate of the probability of the value of electricity consumption for the previous period. It is also worth noting that in operational forecasting, it is important to take into account external factors affecting the amount of electricity consumption. Therefore, in most cases, data from weather sensors and data on power consumption for previous points in time are used. For intraday forecasting, it is important to obtain operational information about these factors in real time, which requires additional technical solutions when implementing predictive analytics systems.