Engineering, Technology & Applied Science Research Vol. 8, No. 2, 2018, 2818-2820 2818 www.etasr.com Shah et al.: ANN and ANFIS for Short Term Load Forecasting ANN and ANFIS for Short Term Load Forecasting S. Shah Electrical Engineering Department IITE, INDUS University Ahmedabad, India shah.sweta21@gmail.com H. N. Nagraja Electrical Engineering Department Graphic Era University Uttarakhand, India Nagraj_hp@yahoo.com Jaydeep Chakravorty Electrical Engineering Department Indus University Ahmedabad, India jaydeepchak@yahoo.co.in Abstract—Load forecasting has become one of the major areas of research in electrical engineering. Short term load forecasting (STLF) is essential for power system planning and economic load dispatch. A variety of mathematical methods has been developed for load forecasting. This paper discusses the influencing factors of STLF and an artificial intelligence (AI) based STLF model for MGVCL load. It also includes comparison of various AI models. Our main objective is to develop the best suited model for MGVCL, by critically evaluating the ways in which the AI techniques proposed are designed and tested. Keywords-load forecasting; neural network; adaptive neuro fuzzy interface system I. INTRODUCTION Electric load forecasting is the process used to forecast future electric load from the given historical load and weather information. In the last few decades, several models have been developed to forecast electric load more accurately than analytical methods. Load forecasting can be divided into three major categories [1]: 1. Long-term electric load forecasting, used to supply electric utility company management with prediction of future needs for future expansion, equipment purchases, or hiring of new staff. 2. Medium-term forecasting, used for the purpose of scheduling fuel supplies and maintenance. 3. Short-term forecasting used to supply necessary information for the system management of day-to-day operations and unit commitment for economic load dispatch. Short term load forecasting mainly aims at one hour to one week forecast. As daily load pattern is highly non linear and random, it is very difficult to obtain higher accuracy using analytical methods. Application of artificial intelligence (AI) techniques like neural networks and adaptive neuro fuzzy interface systems is an advanced approach for accurate short term load forecasting. II. ARTIFICIAL NEURAL NETWORKS A. Introduction Artificial neural networks (ANNs) have been used for many years in sectors like medical science, defense industry, robotics, electronics, economy, forecasts etc. The learning property of ANNs in solving nonlinear and complex problems is the cause of their application to forecasting problems. B. Learning Algorithm ANNs work through optimized weight values [2]. The method by which the optimized weight values are attained is called learning. In the process of learning we present to the neural network pairs of input and output data and try to teach the network how to produce the output when the corresponding input is presented. When learning is complete, the trained neural network, with the updated optimal weights, should be able to produce the output within desired accuracy. There are several learning algorithms. They can be broadly categorized into two classes: supervised and unsupervised. Supervised learning means guided learning, i.e. when the network is trained by showing the input and the desired result side by-side. This is similar to the learning experience in our childhood. As children we learn about things (input) when we see them and simultaneously are told (supervised) their names and the respective functionalities (desired result). This is unlike the unsupervised case where learning takes place from the input pattern itself. In unsupervised learning the system learns about the pattern from the data itself without a priori knowledge. This is similar to our learning experience in adulthood. For example, often in our working environment we are thrown into a project or situation which we know very little about. However, we try to familiarize with the situation as quickly as possible using our previous experiences, education, willingness and similar other factors. This adaptive mechanism is referred to as unsupervised learning. C. ANN Model Traning Process Multilayer feed forward neural network (having input layer, hidden layer and output layer) is used for STLF. The training goal was set at 0 in order to ensure zero tolerance to network computational errors [3]. The transfer function used was the tan-sigmoid in the hidden layer while a linear function was used in the output layer neurons so as not to constrain the output's values. The learning function used is the steepest gradient descent method. The Levenberg-Marquardt learning function was used as it has better learning rate compared to the other available functions in forecasting problems. The training function used was the steepest gradient descent function and in some tests the steepest gradient descent method with momentum. net art inf fuz of to app to A. lay inp ass be par fun op giv rul sum cal nor the pre AN inc Engineerin www.etasr III. ADA An adaptive twork-based f tificial neural n ference system zzy logic princ both in a sing a set of fuzzy proximate non be a universal ANFIS Archi In ANFIS ar yer has the ana put in terms sociated. This defined by su rameters. Th nctions are cal Layer 2 show erator on the ven as Oi Wi The output o le of firing po mmation of t lculated in la rmalization. TOi Wi The output oOi W1d1 In this, result e ANFIS. The emise and resu NFIS which coming signalsOi ∑ w1 ng, Technology r.com APTIVE NEURO e neuro-fuzzy fuzzy inferenc network that i m. Since it inte ciples, it has th gle framework. y if–then rules nlinear functio l estimator. itecture chitecture [4] alogous functi of its associ is also called uitable membe he foundation lled nonlinear Fig. 1. AN ws fuzzy AND input signals μpi x μqi x of this level is n ower. The ratio the firing str ayer 3. This i The throughput ; i of layer 4 is giv1 w1 aix ting parameter amount of pa ultant paramet can be give s. 1di ∑∑ y & Applied Sci O FUZZY INTER y inference s ce system (AN is based on Ta egrates both n he potential ca . Its inference s that have lea ons. Hence, A (Figure 1), ea ion. Level 1 sh iation grade membership g ership functio n parameters parameters. NFIS architecture D operations u s to generate ; i 1,2 normally know o of a rule’s f rengths of all is known as t of layer 3 cani 1,2 ven by: biy ci ; rs are ai, bi an arameters of A ters. In layer n below is ience Research RFACE SYSTEM system or ad NFIS) is a ki akagi–Sugeno neural network apturing the be system corres arning capabil ANFIS is consi ach joint in the hows that with a linguistic t grade [5] and ns with appro s associated e using an appro the output c (1) wn as the anal firing strength l the rules a firing strengt n be given by: (2) i 1,2 (3) nd ci where i=1 ANFIS is the s 5, overall outp the summatio (4) h V M aptive ind of fuzzy ks and enefits sponds lity to idered e same h each tag is it can opriate with opriate can be logous is the and is ths of 1, 2 of um of put of on of fun pos netw B. con util mat spa init ada fuz com squ mea betw whe obt for algo con esti bac gra are com whi con esti der outp laye upd A. fore and AN sho Nov sma com 0.3% imp Feb for betw Vol. 8, No. 2, 20 Shah et al.: A The ANFIS s nctions and b ssible. ANFIS work holding Learning Pro As mentione nsequent (linea lizing the learn thematical rela ace. Normally, tiated by the s aptive learning zy model and mbining gradie uares optimiza asure, usually ween actual a en either the tained. There a ANFIS. In orithm, functi nsequent para imate. In the ckward and t adient descent. learned, the o mbination of th 1 1 2 1 1 1 2 2 2 ( ) ( ) w f w w w x m w x n       ich is linear. I nsequent para imate. In the b rivatives of th tput, propagate er. In this b dated by the gr ANN Model R Table I show ecasted error. d forecasted er NN error is 0.1 ows the compa vember and D all compared mparison of er % to 0.5%. It proves the ac bruary and Ma the entire yea ween errors is 018, 2818-2820 NN and ANFIS structure is a because of t is an amalgam the benefits of ocess d above, both ar) parameters ning process, t ationships bet , as a first step system and the g process. Ba d tunes it by ent descent ba ation algorith y defined as th and desired ou predefined e are two passes the forward ional signals ameters are i e backward p the premise p . When the va overall output he consequent 2 1 1 2 1 1 1 2 2 ( ) w f w w w x n w w q      In the forward ameters are i backward pass he squared er e backward fr backward pas radient descen IV. Results ws the compar Average ANN rror is -1.346% 34% and forec arison of ANN December. It to forecasted rrors for the en t indicates tha curacy of for arch (Figure 2 ar could be sim also shown in 0 IS for Short Ter meaningful a this several c mation of fuzz f both. h the premise s of the ANFI to optimally re tween the inpu p, an approxim en improved th asically, ANFI y means of a ack propagatio hms. At each he sum of the utput, is reduc epoch numbe in the hybrid d pass of the go forward ti dentified by pass, the erro parameters ar alues of the p t (f) can be ex parameters as 2 1 1 1 1 2 2( ) f w f w w q w x m      d pass of the l dentified by s, the error sig rror with resp rom the output s, the premis nt algorithm. RESULTS rison of ANN N error for 6th % while for 7 casted error is N error and f shows that A error. Figure ntire year whic at application recasting. The 2) varies up to milarly develop n Figure 2. 2819 rm Load Foreca assignment of configurations zy logic and n e (non-linear) IS should be tu epresent the fa ut space and o mate fuzzy mo hrough an iter IS takes the i a hybrid techn on and mean h epoch, an squared differ ced. Training er or error ra learning proce e hybrid lea ill layer 4 and the least sq or rates prop re updated by premise param xpressed as a l s: 2 2 2 w f m  (5) learning algor the least sq gnals, which ar pect to each t layer to the se parameters N model error h January is 0 7th January av s -1.566%. Fig forecasted erro ANN error is e 2 also show ch varies betw of ANN for S e ANFIS erro o 0.07%. A m ped. A compa asting node s are neural ) and uned, actual output del is rative initial nique least- error rence stops ate is edure arning d the quares pagate y the meters linear rithm, quares re the node input s are r and 0.05% erage gure 2 or for very ws the ween - STLF or for model arison Engineerin www.etasr T Hour 6th Er (AN 1 0.3 2 -0. 3 0.0 4 0.0 5 0.0 6 0.3 7 0.9 8 0.8 9 0.4 10 0.2 11 -0. 12 0.7 13 -0. 14 -1. 15 1.3 16 -0. 17 1.4 18 -0. 19 -0. 20 -2. 21 1.0 22 0.2 23 0.3 24 0.2 ng, Technology r.com Fig. 2. TABLE I. A Jan rror NN) 6th J Err (Forec 390 -2.1 .986 -0.6 050 -0.6 056 -0.8 089 -0.0 324 0.17 904 -2.9 872 -3.2 422 -6.4 285 -6.1 .413 -4.7 779 -2.5 .577 -3.6 .671 -2.9 366 -2.6 .407 -0.4 429 0.83 .394 2.27 .809 -2.7 .410 -1.6 091 4.1 275 2.74 383 0.19 240 1.94 y & Applied Sci Error graphs ANN RESULTS FOR Jan ror cast) 7th Ja Erro (ANN 81 -0.04 658 0.047 695 -0.02 823 0.048 060 2.178 78 -1.51 935 0.000 281 0.019 404 0.000 42 0.000 782 1.948 515 0.036 640 0.019 938 0.018 647 -1.87 455 -3.28 39 -5.02 70 -0.03 777 0.018 692 0.036 14 -1.01 45 -0.04 92 11.67 43 0.02 ience Research R JANUARY. an or N) 7th Jan Error (Forecas 46 -1.899 7 -1.129 24 -1.335 8 -1.278 8 -2.683 2 -1.351 0 -2.145 9 -1.780 0 -3.800 0 1.042 8 -0.842 6 -2.504 9 -3.656 8 1.575 77 2.217 87 0.263 21 0.436 7 -0.405 8 -0.339 6 -2.123 8 -3.306 40 -6.583 77 -3.935 1 -2.019 h V n st) 9 9 8 0 0 2 4 6 9 6 9 [1] [2] [3] [4] [5] Vol. 8, No. 2, 20 Shah et al.: A K. Kalaitzakis, forecasting base Electric Power M. Markou, E. Load Forecasti Deregulated Ele 2008 E. Banda, K. A Neural Network pp. 108-112, Ju Z. Souzanchi-K Multi Adaptive Forecasting by on Electronics 54-57, August 1 H. P. Oak, S. J International Jo No.3, pp. 1878- 018, 2818-2820 NN and ANFIS REFER G. S. Stavrakaki ed on artificial ne Systems Research Kyriakides, M. P ing Using Multip ectricity Market I A. Folly, “Short-T k”, IEEE Lausan uly 1-5, 2007 K, H. Fanaee-T, M e Neuro Fuzzy I Using Previous D and Information 1-3, 2010 . Honade, ‘ANFI ournal of Current -1880, 2015 0 IS for Short Ter RENCES s, E. M. Anagnos eural networks pa h, Vol. 63, No. 3, Polycarpou “24-H ple MLP’, Intern Issues in South-E Term Load Forec nne Power Tech, M. Yaghoubi, M Inference System Day Features”, In Engineering, Ky IS Based Short T t Engineering an 2820 rm Load Foreca stakis, “Short-term arallel implementa , pp. 185-196, 200 Hour Ahead Short rnational Worksh Eastern Europe, p casting Using Ar Lausanne, Switze . R. Akbarzadeh- m for Short Term nternational Conf yoto, Japan, Vol. erm Load Foreca nd Technology, V asting m load ation”, 02 t Term hop on p. 1-6, rtificial erland, -T, “A m Load ference 2, pp. asting’, Vol. 5,