Long-term load forecasting using grey wolf optimizer -least- squares support vector machine

ABSTRACT


INTRODUCTION
Planning in power systems is a projection of how the system should evolve over a specific time-frame. It has become more difficult to plan power systems in developing countries. Planning must be completed despite numerous vulnerabilities such as future load designs, population growth and economic growth representing developing countries, technical, economic and environmental constraints [1]. Load forecasting plays an important role in planning power systems to ensure uninterrupted and economical power generation and distribution to consumers. Precise electrical load forecasting is a major issue in the planning and management of electricity generating utilities.
The load forecast can be divided into three classifications. Short-term load forecasting (STLF) covers an hour to a week time [2]. Short term load forecasting contributes to the decision-making process, including unit commitment and process for improving security [3]. The accurate load forecasting helps enhancing the unit commitment scheduling and thus can save large amount of cost per year [4]. Medium load forecasting (MTLF) covers a week-to-year period. It is essential for fuel supply and maintenance operations to be planned [5]. Long-term load forecast (LTLF) forecast for more than one year [6]. Long term load forecasting facilitates the decision-making of an electrical utility including the purchase and generation of electricity, load switching and infrastructure development. The importance of load prediction turns out to be more noteworthy in high-growth developing countries. Understanding and forecasting of load characteristics was complex due to its dependence on a wide range of factors influencing the weather, geographical diversity, sunrise/sunset time, seasonal diversity, etc. [7]. Despite the increase in electricity demand annually, the pattern of the load profile may also change [8][9].
Analysis of long-term load forecasting has received little attention. Reference [10] presented a long term load forecast using Fuzzy Logic model. Fuzzy logic model was developed based on the weather parameters which are temperature and humidity. Two years historical load data were utilized to predict a year-ahead load demand. The value of Mean Average Percentage Error (MAPE) obtained was 6.9%. The technique is then improved in [11] using Fuzzy-Neuro. Fuzzy-Neuro is the combination of Fuzzy Logic and Artificial Neural Network (ANN). In Fuzzy-Neuro, the output of the fuzzy logic system are fed to ANN. Fuzzy logic systems used for making decision based on the rules, meanwhile, ANN is used for training and testing process. The characteristic of ANN are high learning capabilities, adaptability and generalization. By using Fuzzy-Neuro, MAPE has improved to 1.22%.
Recently, various Artificial Intelligent (AI) techniques can be implemented for forecasting. One of them is ANN. Research in ANN has received considerable attention. ANN has many advantages as compared to the conventional computational systems such as computation speed and robustness. The most profound and important characteristic of ANN is the ability to memorize which involves a large number of processors operating in parallel. Each processors consists of own knowledge and can access the data in its local memory. The ability of ANN to generalize on unseen data and learn from noisy data make it a very powerful machine learning algorithm [12]. Apart from ANN, there is another intelligent system that has been given attention recently namely Support Vector Machine (SVM). The advantages of SVM including fast convergence rate, good generalization capability and automatic determination of hidden neurons [13]. The regression problem in SVM is formulated based on convex quadratic programming problem where it use linear regressor [14]. In linear regressor, the regressor will maps the inputs into a higher dimensional feature space to minimize the cost function. The major shortcoming of SVM is the computational burden for optimization programming constrained. This weakness has been overcomed in Least-Square Support Vector Machines (LSSVM), which substitute a quadratic programming with linear equations [15][16][17][18]. The accuracy of the prediction of LSSVM depends on the hyperparameters value setting. Therefore, Grey Wolf Optimizer -Least Square Support Vector Machine(GWO-LSSVM) is proposed to solve long-term load forecasting problem. GWO is integrated with LSSVM for determining the optimal hyperparameters. GWO is inspired by hierarchy system and hunting mechanism of grey wolves in nature. GWO has been effectively implemented for solving various power system problems [19][20][21][22][23].

RESEARCH METHOD
In order to predict total load demand, a total of 1035 sets of data are collected in the area of Dayton, Ohio, US from January 2016 until May 2018. For training, 668 data sets are used and the remaining 367 data sets are used for testing. Moreover, the temperature, wind speed and humidity are the inputs, while the total output power is the output of the model. Figure 1 presents overall data used in this paper. It can be observed that the total load demand fluctuated. Therefore, an accurate prediction technique is required to ensure the electricity generated capacity meet electricity demand at all times. In this paper, GWO-LSSVM is developed to predict the total load demand in long-term duration. The efficacy of the proposed algorithm is evaluated through comparison with other techniques such as LSSVM and Ant Lion Optimizer-Least-Square Support Vector Machine (ALO-LSSVM).

Prediction measurement
Mean Absolute Percentage Error (MAPE) and coefficient of determination (R 2 ) were used for the evaluation purposes as shown in (1) and (2). Both of them act as indicators which assess the performance of suggested technique. MAPE is a measure of prediction precision of a forecasting technique in measurements.
It usually uses percentage to shows the accuracy. For MAPE, the lower the value, the better the performance. In the simulation, MAPE is utilized as objective function of GWO-LSSVM. The coefficient of determination, R 2 is a part of the output variable variance that can be predicted from the input variable. It gives a measure of how similar the outcomes are replicated by the model, in view of the extent of total variation of results clarified by the model. However, for R 2 , the closer the value to 1, the better it will be. (1) Where t=1, 2, …., t At = Actual values Ft = Predicted values/Forecasting values

Development of GWO-LSSVM
GWO is a metaheurictic-based technique stimulated by the hunting behavior and leadership hierarchy of grey wolves as depicted in Figure 2. Alpha (α) is the leader of the wolves. It is responsible to make a settlement which are relates to hunting, when and where to sleep, and etc [24]. Beta (β) is the second level which is the assistant of α in making decision. β is likewise the best applicant to replace α when α getting old or die. The lowest in grey wolves' system hierarchy is called Omega (ω). It acts as a scapegoat. ω can fulfill the entire whole grey wolves group. The third level in grey wolves' system hierarchy is Delta (δ) which must report to α and β, however they lead the ω. δ responsibility is scouting to secure and ensure the security of the group of wolves. The system hierarchy of grey wolves is an assigned characteristic used in GWO. Another amazing characteristic of wolves is the group hunting strategy. The fittest solution is considered as the α. Meanwhile, the second and third best solutions are considered as β and delta δ respectively. The rest of the individual solutions are considered as ω. During hunting, the wolves tend to encircle their prey. The encircling behaviour is represented as (3). Where ⃗ ⃗ is the coefficient vector In LSSVM, there are two tuning parameters that associated with RBF kernel, σ 2 and regularization parameter, γ. These parameters affect the LSSVM estimation model's accuracy. These parameters are optimized using GWO-LSSVM to minimize MAPE. The flowchart of GWO-LSSVM is presented in Figure 3.   Table 1. The number of search agents is set to 10. If the number of search agents is high, it will consume more time. However, small number of search agents may not give the best prediction. The position of grey wolf is then updated according to (4). α, β, and δ wolves determine the possible position of the prey. This process continue until the convergence goal is met and is taken as the best solution. In this algorithm, MAPE is set as the fitness function. MAPE value is obtained by calling the LSSVM algorithm. The position of search agents are updated according to α, β, and δ in search space as shown in (4). The process is repeated until the solution converged. The solution will converge if the different between maximum and minimum fitness reach 1x10 -7 . In this paper, the results of GWO-LSSVM is also compared to ALO-LSSVM in terms of accuracy and convergence speed. In ALO-LSSVM, ALO is utilized to determine the optimal parameters of gamma and sigma. The algorithm represents the hunting behavior of antlions in nature. There are five main steps of the algorithm including random walk of ants, constructing traps, entrapment of ants in traps built by antlions, catching ants, and re-setup the traps [25]. The number of serach agents, maximum iterations, lower and upper bound are set similar with GWO-LSSVM. Figure 4 presents the results of GWO-LSSVM for training data. The red line is defined as the estimation or predicted output, while the blue dots represented as the actual output. The optimal gamma and sigma value are 106784.335 and 0.85381 respectively. The comparison between predicted and actual data for training and testing process using GWO-LSSVM are shown in Figure 5 and Figure 6 respectively. It can be observed that the predicted results almost similar with actual data for both training and testing process. Comparison between GWO-LSSVM, ALO-LSSVM and LSSVM is tabulated in Table 2. By running GWO-LSSVM, the values of optimal parameters of γ and σ2 are 106784.335 and 0.85381 respectively. As the maximum number of iterations reaches 30, MAPE and R 2 are compared. By using LSSVM to model the total output power, the values of MAPE was 0.2570%. However, by using ALO-LSSVM, it helps in reducing the MAPE values to 0.1860%. For GWO-LSSVM, the MAPE is much better than the other two methods which is 0.1265%. For the coefficient of determination, R 2 , LSSVM gives the value of 0.9989. However, the optimal values of parameters γ and σ2 by using hybrid techniques GWO-LSSVM and ALO-LSSVM, the R 2 increased to 0.9998 and 0.9991 respectively. Thus, it can be concluded that GWO-LSSVM provide more accurate prediction for long-term load forecasting as compared to other techniques. The comparison of iterative convergence between GWO-LSSVM and ALO-LSSVM is presented in Figure 7. GWO-LSSVM converged at 7 th iteration while ALO-LSSVM converged at 12 th iteration. It can be concluded that GWO-LSSVM has a precise prediction and the speed of optimization is faster.

CONCLUSION
A hybrid prediction technique namely Grey Wolf Optimizer -Least Square Support Vector Machine (GWO-LSSVM) had been develop to forecast long-term load demand, with the objective function to minimize the error. LSSVM is one of the good techniques to solve nonlinear problems. However, the accuracy of the prediction depends on the selection of the RBF parameters. GWO has been utilized to optimized the value of these parameters. GWO mimics the hierarchy of leadership and hunting mechanism of grey wolf in nature. Four inputs were considered for the prediction; peak load demand, temperature, humidity and wind speed. After conducting the simulation for LSSVM, GWO-LSSVM and ALO-LSSVM, it can be concluded that GWO-LSSVM has better prediction accuracy and higher optimization speed to predict the long-term load forecasting as it.