Concurrent, Performance-Based Methodology for Increasing the Accuracy and Certainty of Short-Term Neural Prediction Systems

Accurate prediction of the short time series with highly irregular behavior is a challenging task found in many areas of modern science. Such data fluctuations are not systematic and hardly predictable. In recent years, artificial neural networks have widely been exploited for those purposes. Although it is possible to model nonlinear behavior of short time series by using ANNs, very often they are not able to handle all events equally well. Therefore, alternative approaches have to be applied. In this study, a new, concurrent, performance-based methodology that combines best ANN topologies in order to decrease the forecasting errors and increase the forecasting certainty is proposed. The proposed approach is verified on three different data sets: the Serbian Gross National Income time series, the municipal traffic flow for a particular observation point, and the daily electric load consumption time series. It is shown that the method can significantly increase the forecasting accuracy of the individual networks, regardless of their topologies, which makes the methodology more applicable. For quantitative comparison of the accuracy of the proposed methodology with that of similar methodologies, a series of additional forecasting experiments that include a state-of-the-art ARIMA modelling and a combination of ANN and linear regression forecasting have been conducted.


Introduction
Prediction is a process that uses data from the present and the past in order to estimate future. e result of this process is the information about probable events in the future and their effects and outcomes. Making good forecasts is essential for making good decisions and planning in all areas of life. Although it does not have to reduce uncertainties and difficulties of the future, it can increase the certainty and the level of the preparedness for challenges and environmental changes that future events bring. e need for development of prediction methods occurs in almost every area of life-technology, engineering, industry, science, politics, economy, business, sport, medicine, etc. Good forecasts can ensure lower cost of the services and products, increased customer/client satisfaction, and significant competitive advantage [1].
Every daily activity begins with planning. e planning begins with a prediction [2]. Prediction errors may have crucial implications on decision-making, profits and investment justification, risk assessment, alerting events, hard real-time systems' actions, timely handling of emergency health and medical conditions, etc. [3]. Because of that, decreasing the error of the prediction is an essential task for every forecasting expert, regardless of the applied prediction methods.
Prediction methods described in the literature can be roughly categorized into two large groups: traditional and modern. Each of them has advantages and disadvantages. None of them is superior to all others if we consider all possible criteria of evaluation [4]. Traditional prediction methods try to extrapolate time series data using different modelling: exponential smoothing [5], linear or nonlinear regression [6][7][8], simple (AR) or more complex autoregressive models (ARMA, ARIMA, and double seasonal ARIMA) [6,9]. On the other hand, modern prediction methods exploiting the artificial intelligence (AI) behavior can model both nonlinear and linear structures of time series [10] and can produce good accuracy of the forecasting. Such techniques use different topologies of artificial neural network, fuzzy modelling, vector machines, and genetic-simulated annealing algorithms to predict time series data [4,[11][12][13][14]. Different authors have shown that AI-based models frequently express better predictive characteristics compared to models using standard multilinear regression [14]. Finally, both theoretical and empirical findings in the literature show that combination of two or more different methods can be an effective and efficient way to improve forecasts and decrease the error [10]. Such hybrid methods are studied in [2,3,10,[15][16][17].
Despite numerous ways to predict the future mathematically, there are many cases of variables that could not reliably be predicted. Causes for this limitation could be found in the randomness of the events and the lack of significant relationship in data. When factors considered during forecasting of a certain variable are not well known or understood, prediction becomes imprecise or mistaken. Sometimes, there is simply not enough data about everything that affects the forecasted variable. e prediction process relies on some specific hypothesis. If they are set wrong, due to bad judgment, i.e., human error, the prediction will be mistaken. Although the forecasting is based on past events, no one can guarantee that the history will repeat every time in the same way. erefore, forecasts are subject to human error.
A time series can be defined as a sequence of numerical data occurring in regular intervals over a period of time, collected in a successive order. Short time series are characterized by a lack of trend information, randomness and periodicity, and demands for such forecasting represent a challenging problem [18]. Usually, time series cases where the sample length N is very small are not applicable for generating statistically reliable variants of forecasting. In this paper, we will focus on such time series and their forecasting. We will propose a new methodology that can be applied to irregular series. e methodology is applicable to all types and topologies of neural networks, or similar AI based forecasting methods, in order to improve their accuracy. e usual step in development of forecasting ANN is to train many networks, while changing the number of neurons in some particular layer. e ANN with the most accurate forecasting wins. Nevertheless, if we observe the forecasts of all obtained networks, we can conclude that sometimes different networks predict different directions of the trend change in the next forecasting step. In this point, one cannot determine which one is the correct. is is particularly noticeable when dealing with volatile data series. erefore, incorporating more than one network in the forecasting decision could make better predictions of the future events. e methodology that is suggested in this paper improves the forecasting accuracy of the ANN in the sense that it concurrently exploits several most accurate networks instead of the winner one. In this way, the forecasting accuracy can be significantly improved, as well as the confidence of the prediction. e performance of the proposed method is verified on an example of Serbian Gross National Income (GNI) data series, using Feed Forward Accommodated for Prediction (FFAP) neural networks' topology. e results demonstrate higher forecasting accuracy compared to individual FFAP networks. e rest of the paper is organized in the following manner. In Section 2, the structure of the FFAP neural network topology is presented. e section that follows describes in detail the concurrent best-performance-based methodology for increasing the accuracy of the short time series FFAP forecasting. ree case studies are performed and analyzed: Serbian Gross National Income time series, the municipal traffic flow for a particular observation point, and the daily based electric load consumption time series; the forecasting results of the proposed methodology and other state-of-the-art forecasting techniques and their combinations are given in Section 4. e obtained results are discussed in Section 5, while conclusions are summarized in the last section.

FFAP Neural Networks
In general, neural network-based computational and forecasting methods developed from the desire to reveal, realize, and emulate the capability of the brain to process information [14]. e entire brain is composed of many neural networks that receive information from the surroundings, extract and recombine their relevant parts, and make the decisions about the needs of the organism. Artificial neural networks (ANN) emulate such abilities of the brain in order to realize complex nonlinear input-output transformations.
Consider a time series denoted by y i , i � 1, 2, . . . , m. It is a set of observables of an undefined function y � f(t), that are taken at regular time intervals Δt, where t i+1 � t i + Δt. In the forecasting process, the historical data are used to determine the direction of future trends, while one-step-ahead forecasting implies the mathematical search for such a function f that can accurately perform the following mapping: where the term y m+1 represents the desired response, while ε is the acceptable error.
In the past decades, ANNs have been developed as a tool that has great capabilities for recognizing and modelling data patterns that are not easily identifiable by traditional methods. However, one may notice a common feature in all existing ANN applications in forecasting. It is the necessity for a relatively long time series in order to achieve high accuracy. Usually, there should be at least 50 data points to consider [19]. Because of this and due to previous research in short-term forecasting [20][21][22], we have chosen the FFAP neural network topology, as a base to be used throughout this study. is structure will briefly be explained next. 2 Computational Intelligence and Neuroscience General structure of a feed forward neural network is illustrated in Figure 1. It has just one hidden layer, since it is confirmed to be sufficient enough to solve univariate forecasting problems [23]. In this figure, indices "in," "h," and "o" denote input, hidden, and output layers of the ANN, respectively. Weights are labeled with w(k, l), where connections of the input and the hidden layer are designated with k � 1, 2, . . ., m in , l � 1, 2, . . ., m h , while connections of the hidden and output layer are designated with k � 1, 2, . . ., m h , l � 1, 2, . . ., m o . e thresholds are denoted with θ x,r , r � 1, 2, . . ., m h or m o , depending on the layer. e neurons in the input layer distribute the input signals, while neurons in the hidden layer are activated by a sigmoid function. Finally, linear function activates neurons in the output layer. A modified version of the steepest-descent minimization algorithm is applied as a learning method [24]. e problem of initialization was solved using the method described in [25].
In the case of short time series prediction problem, a set of observables (samples) is given (per time instant) meaning that only one input signal is available, the discretized time [18]. To enable mathematical solution for the forecasting problem, in most cases, both values for time variable and the response need to be transformed, as shown in the following equation: Having in mind that t * stands for the time instant, this reduction gives the value of 1 to the time of the first sample (t 0 ). Samples are normalized in the following way: where y * stands for the current value of the target function and M is a constant, selected so as to reduce the relative difference between output values, if necessary. When implementing the architecture in Figure 1, the following series would have to be learned: (t i , f(t i )), i � 1, . . ., m.
Exploiting the basic topology shown in Figure 1, in [26,27], better forecasting solutions were suggested for the problem of short prediction base period. is architecture is referred to as feed forward accommodated for prediction (FFAP) and depicted in Figure 2. e main idea during the FFAP architecture development was to force the neural network to learn the same mapping several times simultaneously but shifted in time. In that way, it is supposed that previous responses of the function will have larger influence on the f(t) mapping. is also forces the network to identify complex intertwined deterministic relations existing between phenomena that influence the observed variable.
e FFAP architecture is depicted in Figure 2. e input set (t i ) is brought to the input terminal. e future terminal at the Output 3 , approximates y i+1 . Output 3 may also be seen as a vector when a multiple-step prediction is required. e present value y i is obtained at terminal Output 2 . Finally, Output 1 should learn the past value, i.e., y i-1 . Although it is not explicitly stressed out, this may also be seen as a vector of past values of the response (since multiple samples from the past responses were used).
We can express the functionality of the network as where Output 1 � y i−1 , y i−2 , y i−3 , meaning that one future, one present, and three previous responses are to be learned. Our task here was to do one-step-ahead prediction. Using the already predicted value as input data for multistep-ahead prediction leads to accumulation of the prediction error what we demonstrated in [21].
In this way, an efficient network topology is created, that uses in the test phase, only matrix w in in combination with single column/row of matrix w o , and the outputs corresponding to different moments, for evaluation of weighs in w in and thresholds in hidden layer.

Concurrent Best-Performance-Based
Prediction Methodology e methodology for increasing the accuracy and certainty of short time series ANN forecasting that we are suggesting is depicted in Figure 3. e aim of this procedure is to establish tools and procedure that will increase the accuracy of the existing individual forecasting FFAP networks, exploiting the best of them, in a concurrent manner. Figure 2: FFAP-feed forward accommodated for prediction ANN structure. We start with a turbulent short time series, expressing a certain variable for a given period of time. In order to facilitate their learning, these data sets have to be properly adjusted (shifted and/or normalized). After that, we evaluate the effect of changing the number of neurons in the hidden layer of the FFAP on the forecast accuracy. We begin with generating and training the FFAP network that has 3 neurons in the hidden layer. e experiment repeats with increasing number of hidden neurons. e construction of networks ends when FFAP with 10 neurons in the hidden layer is obtained and trained. is gives 8 different neural networks.
When a single network is trained, it requires a minimal learning set of first 13 samples (in our case, this is shown to be enough) in order to create the first forecast at the output, that is the extrapolated value of the trend curve. Since this is still training, one can also calculate the first forecasting error for the particular network. By entering every further sample, the network can better learn and predict the trend and its change for the input variable. As already mentioned, this process repeats for the entire group of networks. At the end of this process, all networks are trained with the entire time series, representing the past and the present. e result of this process can be systematized in a form of a forecasting matrix, with columns representing the number of neurons in a particular network, while rows represent time instances for which the forecast were made. e number of columns goes from three to ten, while number of rows is equal to the number of input samples.
In the next step, we analyze the forecasting matrix and search for three topologies that have the best performance measures. e reason for choosing three topologies lies in the fact that the time series can be irregular. Two best networks can predict totally opposite trend change for the next predicting step.
Because no one can tell which one is more correct, we introduce the forecast of the third best network to perform a kind of arbitration. In this way, the confidence of the prediction as well as its accuracy can be improved. In this analysis, we calculate mean average error (MAE), mean square error (MSE), root-mean-squared error (RMSE), and mean absolute percentage error (MAPE), for each column, using equations (5)- (8). e least three values of a particular error correspond to three most accurate ANN forecasting topologies: where N denotes the number of sample cases, y i represents the individual predicted value, andŷ i is the target value for the sample i. e process ends by adopting the final prediction on a test set data calculated using three best topologies simultaneously. Four cases of concurrent forecasting were investigated. First, we simply calculate an average of the forecasts for three most accurate networks. Namely, if the two of three predictions are supporting each other in forecasting the future trend, they make the largest impact to the final prediction, decreasing the importance of the third that predicts the opposite change of the trend. In that way, none may be qualified as the better one. So, the average is the best representative.  Nevertheless, from our experience, creating a linear combination of three best forecasts, where weights or a multiplying factor for each network, corresponds to a reciprocal of its forecasting error, can further improve the forecasting accuracy. e one that was the closest to the correct value gets the largest weight in the equation for the final forecast, as shown in the following equations: where ε can be of any error types that are calculated (MAE, MSE, or RMSE). In these equations y final denotes the final forecast value, while y f1 , ε 1 ; y f2 , ε 2 ; and y f3 , ε 3 denote forecasts of the first, second, and third most accurate networks and their corresponding errors, respectively.

Prediction of the Serbian Gross National Income.
Gross National Income (GNI) is defined as an estimate of the income from goods and services produced by an economy and received by a country both domestically and from abroad. is measure of the size of an economy is highly important and has large economic, political, and social implications. For politically and economically turbulent developing countries, it is very difficult to predict it due to mutual influence of many known or unknown factors. e methodology that can be considered for such predictions exploits artificial neural networks. is kind of time series appears superb for verifying the methodology that we propose. In this study, we will try enhancing the neural network approaches, described in Section 2, for short -term forecasting of the Serbian GNI, by decreasing the forecasting error and increasing the confidents of the predictions in the next term, using its historical data. e same data set will be used for ARIMA forecasting in order to create a feeling about the order of accuracy of the proposed methodology.
For the analyzed series, time is discretized at year long periods and reduced by 1989, as given by (2), while the value of the GNI is given in billions of Dollars, like in original data. In this case, there was no need for normalization of data. However, in our experience, these transformations can make the training process numerically better conditioned.
Although this time series covers a long period of time (1990-2017), the GNI is the economic variable that is obtained annually. It was first introduced by economic scientists in 1990s. Because of that, this time series can be considered as a very short. e time series data are obtained from the World Bank National Accounts data and OECD National Accounts data files.
After the initial training of 8 networks using the test set that contains GNI data for a period 1990-2012, we have calculated their performance measures, i.e., MAE, RMSE, and MSE. ree most accurate FFAP networks have 7, 9, and 10 neurons in the hidden layer. Corresponding weights were calculated based on the values of the networks' errors in order to be used in combined predictions of the test set (2013)(2014)(2015)(2016)(2017). Values are listed in Table 1, while Figure 4 gives their graphical representation. Table 2 contains prediction results obtained for 5 successive time instants of the test set, starting from 2013, as well as the true value of the GNI. Figure 5 visualizes three best GNI predictions from Table 2, while Figure 6 illustrates different GNI predictions, i.e., true, averaged, MAE-weighted, RMSE-weighted, and MSE-weighted values.
As a final measure of performances, for the proposed methodology, we have calculated MAE, MSE, RMSE, and MAPE for all GNI predictions of the test set for the period 2013-2017, and these are shown in Table 3. All types of errors obtained after applying our forecasting algorithm are better than any particular FFAP. For example, the best FFAP that has 7 neurons in the hidden layer had a RMSE for the entire predicted period of 6.85, while the best improvement in forecasting is achieved when MSE-weighted concurrent linear combination of three best FFAP is applied. RMSE in this case is 2.60. e error reduction in this case is almost 65%.
An economic variable that is very similar to the GNI is the GNI per capita. It is an economy measure that is obtained when the value of the GNI is divided by the number of residents in a country. It should be emphasized that values for the GNI and the GNI per capita show very similar trends for particular time interval in the case of the Republic of Serbia. is could be explained by the fact that the natality for the Republic of Serbia is varying very slow over time. e authors have performed the similar forecasting procedure on a GNI per capita time series for the same time period, and it has shown very similar forecasting performance measures.
To the best of our knowledge, analyzing GNI time series in the case of Serbian economy is the task that has not been studied or published yet, and no comparative analysis of the forecasting accuracies with some alternative forecasting methods can be conducted. To overcome this problem alternative state-of-the-art, i.e., autoregressive integrated moving average forecasting methodology (ARIMA) was performed on the same data set. e theoretical background on this topic as well as its implementation strategies can be found in the literature [28,29]. e model fitting process for GNI training set will be very briefly explained here, since this task is out of the scope of this study. To achieve the adequate ARIMA(p, d, q) model, GNI series was tested for stationarity by applying the unit root tests: Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test and augmented Dickey-Fuller test (ADF). ey indicated that the GNI series is nonstationary. Since the null hypothesis of nonstationarity is not rejected, the series needs to be transformed. After the first difference, the autocorrelation coefficients led to a conclusion that this new series is now stationary. Second differencing over the data led to overdifferenced series and was discarded from further analysis. In this way, the order of I term, d, in ARIMA(p, d ,q), has been identified (here, 1). e analysis of autocorrelation We believed that it would be interesting to extend existing experiments with additional predictions that combine ANN weights with those obtained if an appropriate weight for linear regression of the observed time series is added into a final linear combination. We have constructed a new forecasting system that exploits the performance measures of both most accurate individual ANN topologies with most accurate linear regression model. According to their performances, new weights have been calculated in order to obtain modified forecasting linear combinations of four terms. e newly obtained forecasting results are shown in Figure 7.
Corresponding performance measures in GNI forecasting in the case of linearly combined best ANN topologies and their extensions with linear regression models, outperformed the traditional ARIMA modelling. is is systematized in Table 4.

Prediction of the Municipal Traffic Flow.
We define the traffic flow as the number of vehicles that pass a particular observation point per unit of time (usually 15 minutes).
ere can be various ways to "measure" the volume of the traffic and various sources of data such as simulations, sensors, taxi GPS, floating cars, and similar. In short-term prediction, which is our goal, the traffic is predicted in the next moments (usually 15 minutes) on the basis of real-time historical data. e traffic flow time series consists of about 70 measurements, collected in collaboration with the Faculty of Transport and Traffic Engineering of the University of Belgrade, for one of the Belgrade's busiest roundabouts. is time series can also be considered as short. Here, we assume that the traffic from the immediate past has the greatest impact on the future value which, in turn, is produced for a      moment in the near future. In that sense, a large series of consecutive values, we believe, can obscure the information needed for prediction. at stands especially for specific days such as state holidays (very low urban traffic) or football matches (very high urban traffic) for which the information older than several hours (at most 24 or so) has no significance. is is the reason for shortening the initial time series to 70 samples of interest.
Traffic data are accommodated and constant M (here, 140) selected so as to reduce the relative difference between output values. Although this transformation creates negative values in the training set, it is also the way to make the training process numerically better conditioned.
e initial training of 8 networks using the test set was performed. We have calculated their performance measures, i.e., MAE, RMSE, and MSE. ree most accurate FFAP networks have 5, 3, and 4 neurons in the hidden layer. Based on the values of the networks' errors, corresponding weights were calculated for the combined predictions of the test set. Combined weights of three best networks, corresponding to particular error types, are shown in Table 5, while Figure 8 gives their graphical representation. Table 6 contains the prediction results obtained for 11 successive time instants. Previous instants were used for ANN trainings and its forecasting of the first sample in the table. is matrix also required 8 FFAP-ANNs to be trained with rising number of hidden neurons. e best performance measures are obtained for ANNs with 3, 4, and 5 neurons in the hidden layer. Figure 9 visualizes three best traffic flow predictions from Table 6, while Figure 10 illustrates different predictions of the traffic density, i.e., true, averaged, MAE-weighted, RMSE-weighted, and MSEweighted values.
We have again calculated errors (MAE, MSE, RMSE, and MAPE), i.e., performance measures of most accurate networks and of the concurrent, performance-based predictions for the next 11 time instances, and these are shown in Table 7. All types of errors obtained after applying our forecasting algorithm are again better than any particular FFAP. For example, the best FFAP that has 5 neurons in the hidden layer had a MAE for the entire predicted period of 20.46, while the best improvement in forecasting is achieved when RMSE-weighted concurrent linear combination of three best FFAP is applied. MAE in this case is 16.44, which is the improvement of almost 20%.
We have again performed an ARIMA fitting process in order to get the best forecasting model. e ARIMA(8,1,2) model appeared to have minimal values of AIC and BIC.
In forecasting the municipal traffic flow for 11 samples from the test set with 95% confidence limits, using ARIMA (8,1,2) model, values obtained for MAE, RMSE, MSE, and MAPE were 47.20, 24.00, 575.83, and 389.60, respectively. Corresponding performance measures in the case of linearly combined best ANN topologies and their extensions with linear regression model performed better than the traditional ARIMA modelling, considering different types of forecasting errors. is is systematized in Table 8. Introducing the linear regression into the suggested ANN based forecasting model additionally improved the accuracy of the prediction. ese predictions are shown in Figure 11.

Prediction of the Daily Electric Load Consumption.
We define values of the electric load consumption as an average power consumed (in kWh) for a period of one day, at a particular suburban measuring point. Data for the implementation of the method are acquired from the EUNITE 2001 competition file [30]. e electric load consumption time series consists of about a hundred measurements and is considered short.   Computational Intelligence and Neuroscience Using the training set data, 8 FFAP ANNs were trained with rising number of hidden neurons. e best performance measures are obtained for ANNs with 3, 5, and 7 neurons in the hidden layer. Combined weights of three best networks, corresponding to particular error types, are listed in Table 9, while their graphical distributions are shown in Figure 12. Table 10 contains the prediction results obtained for 11 successive time instants of the test set. Figure 13 visualizes three best traffic flow predictions from Table 8  For this forecasting process, we have introduced additional accuracy metric, i.e., maximal error of estimate-M [30,31]. It can be determined using the following equation: where N again denotes the number of samples, y i is the individual predicted value, andŷ i is the true value for the sample i. Calculated errors (MAE, MSE, RMSE, MAPE, and maximal error of estimate) for prediction of the test set with 11 time instances are shown in Table 11. Errors obtained after applying our forecasting algorithm are again better than particular ANNs. For example, the best ANN that has 3 neurons in the hidden layer had a MAE for the entire predicted period of 34.42, while the best improvement in forecasting is achieved when MSE-weighted concurrent linear combination of three best ANN is applied. MSE in this case is 29.72, which is the improvement of about 13%.
We have now performed a seasonal ARIMA fitting process in order to get the best forecasting model. e SARIMA(2, 0, 2) (1, 1, 1) 7 model had minimal BIC value.
In forecasting, the municipal traffic flow for 11 samples from the test set with 95% confidence limits, using SAR-IMA(2, 0, 2) (1, 1, 1) 7 model, values were obtained for MAE, RMSE, MSE, and MAPE, and maximal error of estimate were 28.27, 35.13, 1234.13, 42.38, and 85.18, respectively. Corresponding performance measures in the case of linearly combined best ANN topologies and their extensions with linear regression model did not outperform the traditional     SARIMA modelling, considering different types of forecasting errors. is is systematized in Table 12. Introducing the linear regression into the suggested ANN-based forecasting model did not improve the accuracy of the prediction. ese predictions are graphically represented in Figure 15.

Discussion
Results obtained using neural networks in predicting the GNI for the Republic of Serbia have multiple qualities.
Specifically, the use of mathematical methods and models in the prediction of future economic trends in the Balkan region at the Southeast Euro zone, encompassing the Republic of Serbia, is very ungrateful. GNI is particularly vulnerable to turbulent changes and numerous of noneconomic factors to a great extent. Observing different types of forecasting performance measures in suggested ANN methodology just confirm the quality of the achieved results. e results in predicting the movement of the GNI are of extreme importance for candidates for full membership in the European Union, such as Serbia. e size of the GNI determines the obligations of membership in this integration as well as the benefits that can be granted from numerous funds. e EU budget is financed by its own system of resources whose amount is limited to 1.23% of EU GNI (for the period 2014-2020). ese funds are to be     Figure 12: Combined weights of three best networks predicting electric load consumption, corresponding to particular error types.     On the other hand, with the intention to preserve sustainable future, the importance of prediction of local traffic in large cities comes in for many reasons such as environmental and pollution monitoring; fuel usage reduction; journey planning; traffic control; urban planning; real-time route guidance; and ITS (intelligent transport system). In this analysis, ANN-based forecasting systems were developed enabling prediction of travel times, travel speeds, and traffic volumes on transportation networks using historic and real-time data.
At the end, we can also conclude that electric power load forecasting is the foundation of planning, development, and the assurance of operation efficiency and reliability of electric power systems. Because of the inherent characteristics of uncertainty, randomness, and nonlinearity, the load forecast has always been a forefront and hot issue. In the case of this forecasting task, we have achieved results using different types of forecasting performance measures in treated ANN models in order to confirm their quality.

Summary and Conclusion
In this paper, a novel methodology for increasing the predictions accuracy of different ANN-based systems has been suggested. roughout analysis of three different time series of important everyday parameters, we have introduced some efficient improvements for prediction of short time series. e proposed method has been verified on GNI forecasting at national economy level, municipal traffic volume forecasting, and suburban daily electric load consumption forecasting. ANN-based models have been trained, and the performance of the models has been analyzed by applying various performance evaluation criteria and statistical tests that included MAE, RMSE, MSE, MAPE, and maximal error of estimate. Based on their accuracy, best performing ANN topologies, considering number of the neurons in the hidden layer, have been selected and implemented into a new forecasting system that linearly combines the forecasts of most accurate individual networks. More accurate one has been assigned a greater weight value in the linear combination. e same forecasting accuracy tests have been repeated for a test set data in order to conclude which model is superior. We have concluded that results of these three case studies reveal that linear combination of three most accurate ANN forecasts could predict trend of the future changes more accurately and with more confidence and that in most cases outperforms individual ANN forecasts, ARIMA forecasts, and hybrid ANN-linear regression forecasts. Moreover, we have determined the that accuracy improvement in these three particular cases ranges from 13% in the case of the electric load prediction and up to 65% in the case of Serbian GNI prediction. Based on the analysis presented in the paper, we can anticipate that the applicability of the method can be extended to other AI forecasting and modelling methods, as well as different types of ANN topologies. Our future research would also be oriented toward further error reductions and the discussion on the smallest number of the neurons in ANN layers in order to achieve this goal.

Data Availability
Data representing the Serbian GNI time series used to support this study are obtained from World Bank National Accounts data and OECD National Accounts data files. Data sets representing daily electric load consumption are obtained from the EUNITE 2001 competition file and are cited at relevant places within the text as references [30]. e rest of the time series data representing the municipal traffic flow are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that there are no conflicts of interest regarding the publication of this paper.