The effect of driver variables on the estimation of bivariate probability density of peak loads in long-term horizon

A new paradigm in the energy sector is MES, which captures the interactions among various energy carriers, e.g. electricity, heating, and cooling to improve the performance of the system [1, 2]. To design robust multi-energy systems, forecasting is of paramount importance; therefore, it is of significance to conduct novel and accurate forecasting methods in the multi-energy systems to arrange the operation mode of integrated energy system efficiently and economically [3].

Load forecasting as a dominant field of study in designing the multi-energy systems draws a lot of interest [2‐5]. Conventional load forecasting approaches mainly concerned with only one type of loads, such as power loads, cooling loads, or heating loads. However, multi-energy load forecasting, as an ensemble forecasting approach considers the aggregated load, which has the intrinsic characteristics of the single load type, as well as the relevance among the series [4].

Long-term load forecasting is an indispensable tool for an effective planning of power systems. In long-term forecasting, inaccurate forecasts result in excessive investment, not fully utilized generating facilities, or insufficient generation and unfulfilled demand [6, 7]. Nevertheless, only few researchers have ever proposed new methods for long-term load forecasting in comparison with short-term forecasting [7].

The current load forecasting literatures have mainly focused on point forecasting, in which the expected value of the future load is forecasted through different techniques These forecasting techniques can be categorized as (1) statistical techniques, such as regression models, and time series models, (2) artificial intelligence techniques, such as neural networks and support vector machines, or (3) hybrid methods which are the combination of both statistical and artificial intelligence techniques. The point forecasting is mainly applied for very short-term and short-term forecasting, however, in medium-term and long-term forecasting, point forecast is not reliable since the inputs of forecasting models, which are mainly weather data, suffer from high uncertainty in long terms. Instead, probabilistic forecasting is applied for long-term forecasting where the possibility of having a demand is presented by a probabilistic value [7].

In spite of the importance of medium-term and long-term forecasting in operation and planning of power system, most of studies have focused on point forecasting in the short-term horizon, and few studies have been only conducted on probabilistic forecasting. Nonetheless, among these few studies on probabilistic load forecasting, most of them have focused on short-term forecasting. In [8], a review on probabilistic load forecasting is presented. Table 1 provides an overview of studies carried out in the literature of forecasting, taking into account inherent uncertainties in different contexts.

Fuzzy intervals are defined based on the covariance of data in different operating points, which are characterized by linear regression models. In this context, a fuzzy regression method is presented in [12] to predict the aesthetic quality of a new product or service considering all uncertain objective drivers. In this method, a genetic programming is used to develop nonlinear structures of the models while model coefficients are determined by optimizing the fuzzy criteria. In short-term and medium-term load forecasting context, a fuzzy interaction regression is applied by [13] to forecast electric load in the short-term horizon with the help of fuzzy intervals. Moreover, a prediction interval construction model based on linear programming is presented in [14] to quantify the variability and uncertainty of the output of photovoltaic generating units for very short-term forecasting purposes (i.e., 5-min). This model is based on extreme learning machine and quantile regression. Apart from the considerations concerning the methods and applications of probabilistic forecasting, provided in Table 1, a fuzzy interval model, which is suitable for forecasting of electric demand and the output power of weather-dependent renewable energy sources that have limited dispatchability, is also presented in [22].

The authors of [23] present a practical methodology for probabilistic load forecasting based on a set of predictions, called sister point forecasts, generated from the same family of models. This approach performs the quantile regression on the average of sister point forecasts and generates prediction intervals of future electric loads. Ref. [24] also presents a data-driven framework for probabilistic peak demand estimation using smart meter data of the consumers. This approach proposes four main steps including load modeling, customer grouping, maximum diversified demand estimation, and peak load estimation, to finally address both challenges of the unknown data of future loads and the influence of demand diversity among different customers. References [9, 10], among others, introduce a comprehensive class of time series models to precisely forecast the electric demand of industrial corporations. A simple procedure is also proposed to classify load profiles and present a probabilistic medium-term load forecasting tool for special types of industrial loads.

Among recent research works, the authors of [21] have presented a day-ahead electricity peak load interval forecasting that can easily convert an interval forecasting problem into a classification forecasting problem. The authors have applied a semi-supervised feature selection algorithm called group method of data handling (GMDH) to address an electricity load classification forecasting issue. From a computational point of view, [17] has proposed a hybrid method for probabilistic load forecasting, including a generalized learning machine to train an improved wavelet neural network, and wavelet preprocessing as well as bootstrapping. This hybrid method provides a load forecasting with high reliability, accuracy and speed so that it would be more profitable for practical applications in the electricity market. However, as far as authors’ knowledge concerned, this method has not been used for long-term forecasting purposes.

In a long-term context, [7] has presented a practical methodology for density forecast of the long-term peak electricity demand instead of common point-forecast approaches. The solution proposed by this methodology can hedge the financial risk caused by uncertain demand. At the first stage, the authors have used semi-parametric additive models to estimate the relationships between demand and the most influential driver variables such as temperature, calendar effects, and some economic variables. Then, they have forecasted the probability distribution of annual and weekly peak electricity demand up to 10 years ahead by using a mixture of temperature simulation, future economic scenarios, and residual bootstrapping. This methodology captures the complex nonlinear effect of temperature and also other possible drivers such as calendar effects, price changes, and economic growth.

Another method which is recently proposed in this area is the probabilistic wind power ramp forecasting, which is presented in [19]. The authors have applied an ensemble machine learning technique to generate wind power scenarios and calculate the historical forecasting errors. Then, they used Gaussian mixture model to fit the probability distribution function (PDF) of forecasting errors. This method has not been used for demand forecasting purposes, although it is able to predict with a high level of accuracy.

In this paper, we develop the method proposed by [25] for long-term peak load forecasting considering different driver variables. In fact, we estimate the PDF of peak load in long-term horizons taking into account the most important drivers, like peak load in similar seasons in past years, peak load in the last season, population, and GDP. We apply a SOMN to estimate the PDF, for the reason explained in [25], which allows much more accurate estimates to be obtained with a rapid convergence. The results show good forecasting capability of the proposed methodology at predicting the forecast PDF.

The paper is organized as follows. “Proposed method” section presents the model and its concepts as well as the SOMN for estimating the PDF. The application and high performance of the proposed approach for a real case study are demonstrated in “Results” section. Finally, the conclusions are drawn in “Conclusion and further research” section.

In this section, the proposed approach is applied on a real data in Queensland, Australia to derive long-term probabilistic forecasting. Half-hourly demand data during 2001 to 2016 were obtained from Australian Energy Market Operator (AEMO) [35]. The studied case is examined from two points of view including annual peak load and seasonal peak load. The yearly peak load data between 2001 and 2016 is illustrated in Fig. 2. The seasonal peak load data between 2007 and 2016 is also depicted in Fig. 3. It should be noted that a part of seasonal peak load data is ignored to avoid unnecessary historical data. The rest of seasonal data are presented in Table 2.

It is worth mentioning that the SOMN algorithm is implemented in MATLAB© and executed on a Windows-based PC with a Core™ i5 processor clocking at 3.2 GHz and 4 GB of RAM. In addition, all simulations for comparison and statistical tests are implemented in R-Studio© 3.4.2.

To derive long-term probabilistic forecasting, the univariate density estimation for two cases are studied separately for both training and learning purposes. Although the initial data do not show a normal distribution, they can easily transform to the normal distribution. The histogram graph of the initial and normalized annual and seasonal peak loads along with normal density function curve is illustrated in Fig. 4.

The seasonal and annual peak loads are subjected to different social, economic, and calendar drivers, such as population growth, changing technology, changing the economic condition, and so on [7].

The values of the Pearson correlation between the seasonal peak load and some influential drivers are provided in Table 3. As seen, the highest correlation is between the peak load and that of the similar season in the last year.

Besides, we conduct a principal components analysis (PCA) for different driver variables considering data provided in [36]. PCA aims to maximize the variance of a linear combination of the variables, and forms new variables which are linear composites of the original variables, and the new variables are uncorrelated among themselves [37, 38]. The results for data provided in [36] are illustrated in Figs. 5 and 6. Figure 5 illustrates the coefficients of each variable in principal components. In Fig. 6, the first two eigenvalues form a steep curve as a bend at the beginning and then a straight-line trend with shallow slope. Accordingly, we need to keep those eigen-values in the steep curve before the first one on the straight line, here upon, two components can be retained as follows:

However, due to the lack of data for a wide array of variables in case of Australia, we inevitably avoided conducting PCA and only relied on available data.

Parameters of the component densities including mean vectors and variance–covariance matrices, as well as prior probabilities are the learning weights. Hence, the initial mean vectors of a four-node Gaussian SOMN are set to small random vectors around the mean of standard normal distribution [0, 0]. Besides, the initial variance–covariance matrices are defined as matrices equal to the initial sample variance plus a random coefficient (a random value between 1 and 30) of the initial sample variance, and the initial probabilities are equally set to 1/3.

At each iteration, one data point was randomly taken from the 120-point training set. The learning rates for means and for variances and mixing priors were decreasing from 0.5 and 0.05, respectively.

Three possible scenarios for seasonal peak load forecasting in the most probable range of values are illustrated in Fig. 7. To analyze the fitting performance, the error metric of root-mean-square error (RMSE) in (13) is applied.

RMSE is more preferable in comparison with other measures like Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE). For example, MAPE is a poor-accuracy indicator although it is a quite well-known measure among business managers. With reference to its mathematical formulation, MAPE divides each error individually by the demand, so it is clearly skewed. It means that high errors during low-demand periods will have a significant impact on MAPE. For this reason, optimizing MAPE will result in a strange forecast that will most likely some undershoots may be seen in the demand profile poor-accuracy indicator [39]. On the other hand, compared to MAE and MAPE, the indicator RMSE is more accurate and it does not treat each error as the same. It gives more importance (i.e., weight) on the most significant errors which means that one big error is enough to get a very bad RMSE. Thus, taking the square root of the average squared errors might have some interesting implications for RMSE because the errors are squared before they are averaged and thus, the RMSE gives a relatively high weight to large errors. This proves that the RMSE could be more useful when large errors are particularly undesirable [40].

To evaluate the average of the obtained RMSEs, the proposed algorithm is carried out ten times. The averages of RMSEs for these ten times considering each of driver variables as the second variable for seasonal and annual peak load are illustrated in Tables 4 and 5, respectively. The driver variables for seasonal peak load are population, GDP, peak load in last season, and peak load in similar season in the last year. However, the driver variables for annual peak load are population, GDP. Furthermore, to evaluate the fitted PDFs, the RMSEs of our proposed PDF estimation method are compared with those of the non-central multivariate ‘t’ distribution in the “mvtnorm” package of R.

According to Table 4 for PDF estimation of the seasonal peak load, considering the driver variables lead to a lower RMSE. Accordingly, the lowest RMSE is obtained considering GDP as the driver variable in the case with 150 hidden neurons. However, for some number of the hidden neurons, other driver variables lead to a lower RMSE. For instance, average RMSE associated with the PDF estimation of the seasonal peak load for 30 and 100 hidden neurons are decreased considering the population as the driver variable. Table 4 for PDF estimation of the annual peak load provides similar results.

The key feature of our proposed methodology is provided a full density forecast for the peak demand with quantifiable probabilistic uncertainty, which captures the complex nonlinear effect of possible drivers. The RMSE results illustrate that the proposed method performs well on the historical data.

In light of the results presented in Tables 4 and 5 and the RMSE values in Fig. 8, we can conclude that the proposed method for PDF estimation is more effective than the commonly used method in “mvtnorm” package. In addition, if we consider GDP as the second variable (see Fig. 8 and the last row in Tables 4 and 5), the result will lead to the lowest RMSE.

Nevertheless, according to the overall pattern of RMSEs, there is no hard evidence to define a relationship between “correlation between the dependent variable and each of driver variables” and “RMSE”; this constitutes the ground for future research work.

In this paper, as a preliminary study, we aimed to find an appropriate forecasting approach for multi-energy systems. In this research, we focused on the importance of our proposed method on multi-energy systems. According to [1], the aim of multi-energy systems is considering the interaction among electricity, heat, cooling, fuels, transport at various levels to improve technical, economical, and environmental performance at the operational and planning stage in comparison with classical energy systems whose sectors are treated separately [1].

Multi-energy systems have two aspects; first, these systems as integrated energy systems are known as robust systems due to their ability to stand various types of disturbances by increasing the system responsibility and decreasing the system volatility through providing various alternatives.

On the other hand, the main issues in integrated energy systems are the uncertainty and scalability [41]. To practically implement multi-energy systems, their uncertainty parameters and their uncertainty sets should be first defined. For instance, according to the robust design methodology proposed by [41, 42], the noise factors beyond the control of the designer should be considered in the multi-energy system design.

Therefore, it stands to reason that considering different driver variables via a comprehensive forecasting method, which deals with uncertainties in multi-energy systems, is of paramount importance. However, due to the inherent randomness of the underlying energy resource (e.g., wind speed, solar radiation) alongside economic and social impacts, there will be definitely a high uncertainty associated with the load forecasts especially over the long term.

In addition, to cope with the bottleneck of performance improvement, a practical methodology for density forecast of the long-term peak electricity demand instead of common point-forecast approaches is highly needed. Applying such approaches can hedge the financial risk imposed by uncertain demands. Such approaches also capture the complex nonlinear effect of different possible drivers. Besides, the applied method of PDF estimation is also necessary. For example, here, we have applied a SOMN algorithm to estimate the PDF, which produces accurate estimations with rapid convergence.

In this paper, an unsupervised learning method called SOMN was proposed for estimating the bivariate density functions of the annual and seasonal peak load. The major contribution of this paper is presenting a novel systematic methodology for forecasting the density of long-term peak electricity demand in multi-energy system. Using the measure RMSE, the performance of the proposed method was compared with the non-central multivariate ‘t’ distribution. The simulation results demonstrated that the proposed method outperforms the non-central multivariate ‘t’ distribution.

According to the values of RMSE, it can be inferred that a high correlation between two variables does not necessarily lead to a low RMSE. In other words, there is no hard evidence to define a relationship between these concepts.

The results show that making a relationship between the “correlation between dependent variable and driver variables” and “RMSE” in bivariate probability density function still needs further research. Furthermore, the method proposed in this paper would be developed from several aspects. The most important one is the improvement of the proposed algorithm through introducing ensemble method by combining several artificial intelligent algorithms.