The main purpose of the thesis was to investigate application of various computationalintelligencemethods for shorttermloadforecasting. In Chapter 3, STLF by three types of neural structures i.e. MLPNN, RNN and FLANN using back propagation training algorithm was discusses and their suitability for STLF was investigated. FLANN was modeled using Chebyshev, trigonometric, and algebraic expansion of input vector. RNN as proposed by Savaran  was modified to obtain better MAPE. Chapter 4 discusses about suitability of GA, PSO, and AIS as training algorithms for MLPNN and the corresponding suitability for STLF was investigated. GA and AIS were hybridized with back propagation to train the same neural structure and were found to be achieving better MAPE than utilizing only GA or PSO. This chapter summarizes the work reported in this thesis, specifying the limitations of the study and provides some indications for future work. Following this introduction section 5.2 lists the achievements from the work. Section 5.3 provides the limitations and section 5.4 presents indications toward future work.
applications. Loadforecasting depends on operating time horizon of load occurrence. Accordingly, loadforecasting is classified in three categories such as shorttermloadforecasting, medium termloadforecasting, and long termloadforecasting . Shorttermload forecast is applicable for load forecasts within one day to one week ahead of load occurrence, while medium termload forecast is applicable for forecasting loads in a period within one week to one month ahead of load occurrence. For each classification there are appropriate methods for loadforecasting. Here, we are going to show the impact of ANN and FL in shorttermloadforecasting. This paper attempts to compare two new techniques for shorttermloadforecasting. The paper is organized as follows: In sections 2 and 3 the mathematical models for shorttermloadforecasting with FL and ANN are proposed. In section 4 a case study to show the application of two methods is represented and compare between short-term and medium-term LF with ANN, finally in section 5 provided the conclusion.
Typically, an MPC-based EMS requires load forecasts with a prediction horizon of up to a few days, which in the literature is often referred to as short-termloadforecasting (STLF). Especially for the forecasting of the electric load of large territories, various approaches have been proposed for that purpose . In general, these approaches are divided into two categories . Classical approaches include methods such as time series models, regression models and techniques based on Kalman ﬁltering. Newer approaches apply methods from the research ﬁeld of artiﬁcial and computationalintelligence such as artiﬁcial neural networks, fuzzy inference and fuzzy-neural models, expert systems, and support vector machines (SVMs). Although there is a large volume of literature on this topic, almost no applications of STLF to the thermal and electric loads of buildings have been reported . Forrester and Wepfer , for instance, proposed a method based on multiple linear regression to provide forecasts of the energy demand of a large, commercial building. Dhar et al.  applied a Fourier series model to predict the hourly heating and cooling energy use in commercial buildings. Several researchers studied neural networks (NNs) to develop a building STLF algorithm [7, 8, 9]. Hou and Lian  studied in their work the feasibility and applicability of the SVM for the specific case of building loadforecasting. In , the performances of an autoregressive model, an autoregressive integrated moving average (ARIMA) model, a NN and a Bayesian model for the forecasting of the electric load of an air-conditioned non- residential building are examined.
Gaussian membership functions, it can approach a non-fuzzy set if the free parameter is tuned. Because of their smoothness and concise notation, Gaussian and bell membership functions are popular methods for specifying fuzzy sets. The curves have the advantage of being smooth and nonzero at all points. Season do not change frequently hence Gaussian membership function has been taken for the analysis .
error, regarding two different time horizons, is presented. Regarding the one step ahead, the results obtained, by the tested models, are all very similar to the one obtained by the autoregressive model. However the differences between models become more apparent when the forecasting horizon is increased. For the case of sixty steps ahead, the prediction error has a lower value for the hybrid SVR/Markov model when comparing to the remaining values. However the value achieved by this method is followed very closely by other models, in particular the SVR with simple Gaussian kernel (SVR-GK) and artificial neural network with wavelet decomposition (ANN-WD).
We presented a refined parametric model for shorttermloadforecasting. We began with a linear model, as in (1), that is simple and transparent, uses well-understood techniques and is easy and efficient to implement. A series of refinements improved the accuracy of our model; these included combining models from multiple weather stations, removing outliers from the historical data and special treatment of public holidays. The C++ code of our forecasting program is freely available under the GPL license and we encourage readers to experiment further with our methods.
Due to various factors such as weather changes, social activities and festival types, the load appears as a non-stationary random process in time series, but most of the factors affecting system load have regularity, to achieve effective prediction. Foundation. In order to predict the load better, people use many methods to increase the accuracy of the short-termload. The newer algorithms mainly include the neural network method, time series method, regression analysis method, support vector machine method, and fuzzy prediction method. The core problem of power loadforecasting research is how to use existing historical data to establish a predictive model to predict the load value in future time or period. Therefore, the reliability of historical data information and prediction model are the central factors that affect loadforecasting accuracy.
This paper is designed to walk through our approach to this problem in four steps. First we will reveal the necessary background knowledge so that one becomes acquainted with the problem and how we are attempting to solve this problem. We will start with the history of artificial intelligence. We then will introduce the problem along with the necessary background knowledge to fully understand it. We then take a look at other approaches that have seen real world implementations. Next, we introduce the background knowledge of our proposed solution to the problem, beginning with neural networks. We then pull from recent academic literature on this subject to further build our approach. After all of this background analysis has been done we finally detail our approach which consists of threefold, an FFNN, LSTM, and a hybrid approach. Lastly, we detail and conclude our findings.
A number of studies have developed different methods to accurately forecast electricity load in recent years. Traditional statistical loadforecastingmethods are inadequate to fully model the complex nature of electricity demand and often result in lower accuracy . Artificial Intelligence (AI) based techniques are most favourable due to their ability to tackle non-linear relationships between dependent and independent variables. Fuzzy logic , artificial neural networks , sup- port vector machines  and wavelets neural networks  are popular AI techniques for STLF. Due to the state-of- the-art success of deep neural network methods in image processing, they are naturally being adapted for general time series modelling tasks such as those for loadforecasting. Busseti et al.  predicted load demands by utilizing only time and temperature data by performing prediction using deep neural and recurrent neural networks. In fact, the outcomes behind the aforementioned study achieved an RMSE error of 530 KW/h using a three layer recurrent neural network. Khan et al.  proposed a recurrent neural network model for half an hour ahead loadforecasting whereas Agarwal et al.  developed ANN models for hour ahead load and price forecasts using the load data from ISO New England 1 ) that
A lot of researches were conducted based on AI methods , , –, but there is room for improvements. As presented in  GA and PSO are used in training multi-layer perceptron NN (MLPNN) and compared with back propagation NN (BPNN). It was found that the GA trained NN is more accurate and slower in convergence than the PSO trained NN, but both are more accurate than the BPNN. A redial basis function (RBF) NN is proposed to forecast the load without considering the price factor , then the RBF-NN forecast is adjusted with real-time price using ANFIS. One-hour- ahead forecast using ANFIS is presented by Thai Nguyen and Yuan liao . Next hour temperature, next hour dew point, day of the week, hour of the day and current day load are used as model inputs to the ANFIS model. M. Hanmandlu and B. K. Chauhan  presented a two hybrid NN models comprised of Fuzzy NN (FNN) and wavelet fuzzy NN (WFNN). Fuzzified wavelets inputs from WFNN are used in the FNN, which employed Choquet Integral through q-measure to simplify the learning process, and used reinforced learning to speed the
Short-termloadforecasting has been essential for reliable power system operation. The operational decisions in power systems such as unit commitment, economic dispatch, automatic generation control, security assessment, maintenance scheduling and energy commercialization depend on the future behaviour of the loads . Also, with the rise of deregulation and free competition of the electric power industry all over the world, loadforecasting has become more important than ever before. Load forecasts plays vital role for any energy transactions in the competitive electricity markets . In order to provide substantial amount of electric energy to grow the economy progressively, loadforecasting is required by the electrical power producers. As a result, the accuracy of the forecasts has significant effect on economy and control of power systems operations. Hence, more sophisticated forecasting tools with higher accuracy are necessary for modern power system , . Owing to the importance of short-termloadforecasting (STLF), research in this area during the past few decades has resulted in the development of numerous forecastingmethods. Most of the methods are based on time series analysis. Time series models mainly include approaches based on statistical methods ,-,. The statistical models are hard computing techniques that forecast current value of the variable by using mathematical combination of the previous values of that variable.
Short-term time horizon loadforecasting is usually used for the one-day ahead fore‐ casting and has a strong inﬂuence on the operation of electricity utilities. Many decisions depend on this type of forecast, namely scheduling of fuel purchases, scheduling of power generation, planning of energy transactions and assessment of system safety . The load forecast can be related in complex and non-linear ways with various vari‐ ables such as the past consumption pattern, the season of the year, climatic conditions and others. Several methods to model these relationships have been applied in the past such as regression, econometric, time series, decomposition, co-integration, ARIMA, artiﬁcial intelligence, fuzzy and support vector models [2, 7]. In the work of Lin et al. [8, 9], a stock exchange index (TAIEX) was used in order to better the performance of short-termloadforecasting (STLF) at times of global economic downturn.
Abstract: The forecasting of electricity demand has become one of the major research fields in Electrical Engineering. In recent years, much research has been carried out on the application of artificial intelligence techniques to the Load-Forecasting problem. Various Artificial Intelligence (AI) techniques used for loadforecasting are Expert systems, Fuzzy, Genetic Algorithm, Artificial Neural Network (ANN). This research work is an attempt to apply hybrid and integrated effort to forecast load. Regression, Fuzzy and Neural along with Genetic Algorithm will empower the analysts to strongly forecast fairly accurate load demand on hourly base.
important to select an appropriate hybrid model of ANN which gives less error in the STLF results. This includes inappropriate selection of back-propagation algorithm for ANN may cause to inaccurate estimation of STLF results. In this paper, the ANN is used to perform STLF for the next 24 hours. The performance of ANN is improved due to the significant input data provided by the sequential process of feature extractions. The first stage of feature extraction involves the transformation of raw data that is from the chronological hourly peak loads to the multiple time lags of hourly peak loads. The significant input data for ANN is finally obtained by using the principal component analysis (PCA) and this is considered as the second stage of feature extraction process. The final stage of improvement for the results of STLF is based on the stationary hourly peak loads of ANN output. The ANN can easily forecasts a stationary form of time series and this has been proven in (Mahdi et al., 2009), (Zhang, 2003), (Zhang et al., 1998), (Lachtermacher and Fuller, 1995), (Khotanzad, 1995), (El-Sharkawi, 1993). The performance of the proposed ANN in STLF is verified by using a case study of Malaysian hourly peak loads in the year 2002.
Forecasting is the process of making statements about events whose actual outcomes have not yet been observed . It is the basic facet of decision making . Loadforecasting is the projection of electrical load that will be required by a certain geographical area with the use of previous electrical load usage in the said geographical area. Electrical loadforecasting has a lot of applications. They include: Energy purchasing and generation, load switching, contract evaluation and infrastructure development. It is a very essential part of an efficient power system planning and operation. This is why it has become a major research area in the field of electrical engineering. The time period in which the forecast is carried out is fundamental to the results and use of the forecast. Short-term forecast, which is from one hour to one week, helps to provide a great saving potential for economic and secured operation of power system; medium-term forecast, which is from a week to a year, concerns with scheduling of fuel supply and maintenance operation; and long-term forecast, which is from a year upwards, is useful for planning operations.
The next step is to test the assumption of residual white noise. The results of this test showed that the models already meet the assumptions of residual white noise because p-value > 0.05. After that, we need to test the normality. the models generated from ARIMA method, although it has already been added outlier not all residual has normal distribution, This is because the data is in the leptokurtic form. Table 1 is a table of result model from ARIMA methods with added outlier by calculating MAPE, SMAPE, and RMSE of out sample’s data. The resulting models are Paiton 13.30 is (0,1,1) (0,1,1) 7 , Paiton 18.30 is (0,1,[1,2,3]) (0,1,1) 7 , Paiton 22.30 is (0,1,[1,8]) (0,1,1) 7 .
existing regression model using the same data. The NN-based model produced more accurate results in terms of forecast errors, and was robust, adaptive to weather changing conditions. According to Hong and Fan (2016), one of the most successful implementations of NN models for STLF was developed by Khotanzad et al. (1997) and sponsored by the Electric Power Research Institute (EPRI). The authors nevertheless admitted that a couple of conventional techniques were used previously with varying degrees of satisfaction, but not as accurate as would be desired. Besides, most of these models could not be used elsewhere, but at the built site. This paper investigated several types of NN architecture such as recurrent NN and radial basis NN, and came to a conclusion that there was no major advantage of these architectures over the MLP in terms of the loadforecasting problem. This NN Short-TermLoad Forecaster (ANNSTLF) constructed by the aforementioned authors, was subjected to different comparative studies using various methods as well as other NN-based models. The accuracy of the load forecasts was evaluated and expressed in terms of the MAPE. The ANNSTLF yielded very good results and induced its acceptance across Canada and USA.
Broadly, the loadforecasting techniques can be divided into two categories such as parametric or non parametric techniques. The linear regression, auto regressive moving average (ARMA), general exponential technique and stochastic time series techniques are some examples of parametric (statistical) technique. The main drawback of this technique is its capability in abrupt change of any types of environment or social changes. However, this shortcoming is overcome by applying non- parametric (artificial intelligence) based technique because of its potentiality to global search. Among these artificial intelligence based methodology, artificial neural network has emerged as one of the most prominent technique that receive much more attention of researchers. The ability to solve the complex relationships, adaptive control, image denoising, decision making under uncertainty and prediction patterns makes ANN a powerful performer than previously implemented techniques [7-11]. Hence, several variants of ANN which is generally hybridization of neural network with some learning techniques such as GA, PS0, BFO etc are proposed by several researchers. Similarly, ANN in hybridization with fuzzy logic, AIS (Artificial Immune System), LMA (Levenberg Marquardt Algorithm) etc have also shown improved performance in terms of accuracy, computational cost and time requirement. In sec 2, a brief discussion on ANN has been done. In section 3, different variants of ANN, that includes the conventional and hybrid neural network techniques that have successfully applied to STLF is described. Finally, the conclusion of the present review work has been presented in section 4.
IJEDR1504164 International Journal of Engineering Development and Research (www.ijedr.org) 938 Artificial intelligence tools using to improve the forecasting result. It is very difficult to predict the load accurately using analytical methods. A neural network is the advanced approach for accurate short-team loadforecasting. Short-team loadforecasting is using Artificial intelligence method and error minimized.
IJEDR1602012 International Journal of Engineering Development and Research (www.ijedr.org) 79 Loadforecasting is an essential tool for operation and planning of power system. It is required for unit commitment, energy transfer scheduling and load dispatch. The different types of loadforecasting  can be classified according to forecast period as: a. Short –termloadforecasting (STLF), which are usually from one hour to one month. It is important for various applications such as unit commitment, economic dispatch, energy transfer scheduling and real time control. A lot of studies have been done for using of short-termloadforecasting  with different methods. Some of these methods may be classified as follow: Regression, Kalman filtering, Box &Jenkins model, Expert system, Fuzzy inference, Neuro-fuzzy models and Chaos time series analysis. Some of these methods have main limitations such as neglecting of some forecasting attribute condition, difficulty to find functional relationship between all attribute variable and instantaneous load demand, difficulty to upgrade the set of the rules that govern at expert system and disability to adjust themselves with rapid nonlinear system –load change. The NNs can be used to solve these problems. Most of these projects using NNs considered many factors such as weather condition, holidays, weekends and special sport matches days in forecasting model, successfully. This is because of learning ability of NNs with many input factors.