The important contribution of the seasonal component to the dynamics of electricitydemand suggests that a proper description of (1) may improve the forecast performance with respect of the existing methods. In this paper we attempt to do so. Our method exploits principal components and regression to solve the dimensionality problem implied by N → ∞ and (1). We propose to extract from the data latent factors that drive most of the dynamics, and we use them in a diffusionindex (DI) forecast (see Stock and Watson, 2002). The DI model has been relatively successful to forecast macroeconomic time series, and in this paper we demonstrate that such a method also produces quite good short-run forecasts of high-frequencyelectricitydemand.
forecastingelectricity load demand is usually affected by other causal factors of data or disturbances, such as highfrequency, non-stationary, non-constant variance and mean, and multiple seasonality, which are very likely related to half-hourly, hourly, daily, and weekly periodicity, and the calendar effects, for example, holidays and weekends. Therefore, modeling such data type poses multitude of challenges and the method must satisfy the causal factor that affects forecasting process. One of the methods that eliminate the causal factor of electricitydemand data is the EMD method. Then, it had been necessary to combine the EMD with the DR method, in order to improve forecast accuracy, rather than using a single method and also to investigate the elimination of causal factor in electricitydemand data.
This study proposed hybrid exponential smoothing state space and ANN to model the series with double seasonal patterns. The exponential smoothing state space named TBATS incorporates Box- Cox transformation, ARMA error correction and trigonometric function with time varying based on Fourier. TBATS model is able to decompose and forecast time series with complex seasonal patterns. There is no guarantee that ARMA process appropriate to describe the irregular component of the series. This is probably caused by the existence of the nonlinearity pattern in the data that cannot handle by TBATS. ANN model is superior in handling nonlinear pattern. The hybrid TBATS-ANN is proposed to capture multiple seasonality, highfrequency seasonality and nonlinearity pattern of the series therefore the forecast performance can be improved. Algorithm of the proposed hybrid approach is started by decomposing the load series into level, trend, seasonal and irregular components and continuing using the decomposed components as inputs of ANN. Thus the forecast can be obtained from ANN model with the inputs are the level, trend, seasonal and irregular components of the series. Based on the results for load data of Bawen and Java Bali substations which have integer and nested seasonal pattern, the proposed hybrid model generally yields smaller MAPE and RMSE than TBATS, ANN and DSHW.
However, no consistent conclusion has been made on the superiority of specific models for forecasting peri- odic long memory series. Franses & Ooms (1997)  tried the periodic PAR model, AR model, PARFIMA model and ARFIMA model to forecast the quarterly UK inflation, but found no significant difference between these models. Those authors did find that the PARFIMA model was generally outperformed by rival models. Porter-Hudak (1990) compared the SARFIMA model and the Airline model, and found that the former outper- formed the latter . Nasr & Trabelsi (2005) tried the PARFIMA, SARFIMA, PAR, and AR models in  to forecast inflation rates in four different countries, and showed that the long memory models, the PARFIMA model and the SARFIMA model, performed better than the short memory models in terms of information crite- ria and clean residuals.
Electricitydemand forecasts for both short and long term can be conveniently produced using an MLR analysis . Utilizing the past hourly load data and temperature data, the short term electricitydemand for Sulawesi Island in Indonesia was forecasted . In the Philippines, the electric load for a grid was forecasted using a multiple linear regression analysis . The model takes past load data and future development plans as input variables. For forecasting the electricity consumption of Mexican border states’ maquiladora industries, a multiple linear regression model was used using Microsoft Excel’s regression tools . A MLR-based econometric model and univariate time series model show similar forecast results for Pakistan’s future electricity consumption . A multiple regression method when applied to India’s electricitydemand data reaps results quite comparable with a partial end-use technique . For South Africa, the use of a regression-SARIMA modeling framework revealed some important demand governing variables in the country . Based on past load data, the electricitydemandforecasting was carried out for Malaysia using a regression-based ARIMA model . When compared to other methods, a MLR sometimes may give less accurate results. For example, the electricity forecast for the agricultural and services sectors of Pakistan was carried out using a MLR in comparison to the OLS technique . This comparison shows the MLR resulting in less accurate forecasts. Similarly, the Artificial Neural Network (ANN) model outperforms the MLR model while giving a long term forecast for the electric energy consumption in Thailand . A demand forecast for Pakistan using the STAR (smooth transition autoregressive) model has been given based on an extensive time series data set of 41 years, i.e., between 1971 and 2012 .
The most important finding was that, without considering weather factors one can achieve exceptionally high accuracy in forecasting using only calendar data (MAPE = O.9%). Two day ahead forecasting with the same model gives MAPE = 1.28%, which is also a great accuracy as per current industry standards. Forecasting for three days ahead or more may be done using the model with MAPE = 2.9%, this is very good accuracy considering that this is not an autoregressive model and may be used for short-term, medium term and even long term forecasting. In terms of K-S criterion, the models do not show the same exceptional performance. However this is not due to exclusion of weather factors, the high residuals happen at approximately the same time in all years. We may recommend to address this issue in further studies.
System dynamics methodology has been applied for forecasting in various sectors. Lyneis  forecasted the jet aircraft industry and found that the system dynamics forecast perform better than stochastic forecasts. Railway cargo carrying capacity in China was forecasted by . It was found that system dynamics model generated an error of only 2.66%. Air passenger transport volume was forecasted for coming 10 years for China . In Taiwan,  forecasted air passenger demand and linked it to the passenger terminal capacity expansion decision. Malla and Kunch used simulation to predicted the diffusion of combined heat and power (CHP) system in Belgium. Dyson and Chang  forecasted the amount of solid waste generated in a city of San Antonio, USA to develop a strategy for effective management. Apart from focusing on a forecasting requirement from the model system dynamics has been successfully applied to water resources management , environmental management , deregulation in electricity industry , and shipping port terminal operations  to name a few application areas.
Sawnwood production is an important employer in many European economies. The in- dustry supplies raw material to construction industry, furniture manufacturing and pre- fabricated house production and it affects the prices and the supply of other forest sector products like pulp, paper, bioenergy, wood-based panels and veneer. Sawnwood produc- tion and consumption are diverged globally and the demand for sawnwood varies signifi- cantly depending on the economic activity of the industries which consume sawnwood. In order to balance the supply and the demand on different markets, the sawnwood pro- ducers have to make choices, to which markets they allocate their production. However, this is a complex task because the demand and supply of sawnwood may vary significantly in time and place. For example, in year 2000 the sawnwood consumption was around 30 million cubic meters both in Denmark and in Poland (Figure 1.1). By year 2015 the sawn- wood consumption had decreased to 20 million cubic meters in Denmark while in Poland the consumption had increased to 40 million cubic meters. The 10 million cubic meters’ change in yearly consumption corresponds to the average yearly production of sawnwood in Finland between years 2010 and 2015 . It is important for the sawnwood producers to detect these kind of changes, and if possible also to predict them so that they can better plan their sales efforts.
needed since it can help to make right decision in expansion planning of electricity system and has big impact to profits [3,4]. Besides, it can reduce costs of operation and maintenance, and enhance continuity level of power supply . With regards to this, one of the main methods that has been used successfully is fuzzy approach. Related to uncertainties of load characteristics and ability of fuzzy method to include human knowledge or experience in model such as in selecting input variables, and also to make smaller rule set when we have a lot of data caused fuzzy approach is suitable and interesting for load forecasting application [7,8]. Concerning its applications, paper of  proposeda fuzzy linear regression model for STLF which is composed for summer and winter seasons using weather parameters. In , the authors presented forecasting study for power system in India. In their case, fuzzy based STLF which implemented for peak, medium, and off-peak demand seasons gives better result than conventional method. Time and temperature variables are used as inputsfor the model. Meanwhile in the paper of , the authors conducted STLF study for Jordan context. A fuzzy inference model is presented with input namely last day and last week consumptions, last day and forecasted temperatures, weather, and type of day. Recently, short-term forecasting for electricitydemand in Turkey using fuzzy logic and ANFIS is presented in . Four day type models for each method are composed that is Monday, Weekday, Saturday, and Sunday models. Historical load demand, difference of temperature, and season data are inputs in their models.
Electricity supply in Indonesia shall be under the control of the state, which supplies shall be provided by the government and the regional government under the principle or regional autonomy. The implementation of electricity supply business is carried out by state-owned enterprises (BUMN), regional-owned enterprises (BUMD), private enterprises, cooperatives and self-supporting communities . Every power supply entrepreneur is obliged to hold a business license for electricity supply (IUPTL). PLN is one of the holders of IUPTL which business area covers the entire territory of the Republic of Indonesia as figure 1.
The first step in the proposed research methodology is to choose the right forecasting tool. Selection will depend on the time horizon estimates, data are available, time available, and the estimated operating costs with poor or inadequate. In this research, forecastingelectricitydemand long term being suitable for application in planning and decision making. Forecasting models are selected using available data collected for the electricity consumption of the world's energy agency, the model variable data from the Office for National Statistics. In addition, the proposed benchmarking forecastingmodel also uses data from previous studies. A reliable forecasting tool derived from the simulation must have a higher accuracy than other estimation models. The model shows a good estimate of the prediction will be elected as the development model used to forecast future demand for electricity by using economic scenario.
This study seeks to contribute to the forecasting studies in general and electricity consumption forecasting in particular. There has been a shift from simple linear forecasting tools to more complex nonlinear forecasting in the endeavor to find a more suitable tool. Accordingly, novel Artificial intelligence tools are proposed in this work. The use of artificial intelligence tech- niques such as neural networks falls within the logic of introducing complex methods that are able to deal with the non-stationary data sets. Marvin minsky divided the task of creating an intelligent machine into five main areas: Search, Pattern-Recognition, Learning, Planning, and Induction . A machine searches for solutions in a solution space but the search is often inefficient because of the vast solution space especially now with big data. It is therefore, im- portant to introduce pattern-recognition to make the search efficient by restricting the machine to use its methods only on the kind of attempts for which they are appropriate. The efficiency of the search is further improved through learning which direct Search in accord with earlier experiences. Planning helps in dividing the problem into smaller chunks that make it relatively easy for the machine to search for the solution. Induction is creating model for the machine to generalize on unseen data. Artificial intelligence has come under heavy criticism for the usage of statistical analysis. Noam Chomsky, the world renowned linguist, criticized the definition of success in AI which is defined as getting a fair approximation to a mass of chaotic unana- lyzed data . This way of studying AI, he stated, does get the kind of understanding that the sciences have always been aimed at which is to understand the underlying principles of the system. He, however, acknowledged that the statistical analysis gives much better prediction of phenomena than the physics models will ever give.
in the modality of grid usage requires a certain degree of system overhaul. In particular, one needs to ensure that DSM does not create congestions or voltage violations at any point of the distribution grid. A comprehensive approach towards the actuation of flexibility, taking into account its effects at different grid levels, is proposed in this thesis. In this work, I chose to investigate distributed control techniques for the problem of coordinating the flexibilities in the grid. Compared to centralized techniques, distributed control offers the advantage of being more scalable and allowing to preserve partially the privacy of the agents. On the other hand, distributed control presents a number of challenges, among which the need to decompose the optimization problem and distribute it among the agents, who need to solve parts of it locally, often on a hardware, which has limited computational power. To successfully apply distributed control techniques for the coordination of a set of agents represented by electrical loads or batteries, one needs to rely on the forecasts of each agents’ electrical consumption/production with relatively high time resolutions. Since both power consumption and DERs power production profiles have a daily seasonality, a 24 hours ahead planning is typically used. The high number of time steps, the frequency at which the problem must be solved, the number of agents to be coordinated (in the range of hundreds) and the limited computational power of the devices on which the distributed control problem is solved (smart meters), require a careful selection of both optimization strategy and forecasting algorithm. The efficient forecast of a high number of relatively small loads and generators is particularly challenging. Part of this thesis is specifically dedicated to the design and evaluation of forecasting techniques for production and consumption. Both the accuracy and the computational requirements of the proposed techniques
There are several possible extensions of the forecastingmodel. A possible extension is to incorporate information from the waiting list. Part of the care pathways is already known for patients who are not in treatment yet, but are admitted when a bed becomes empty. Using this information makes the probability distribution of the workload smaller. Another possible extension is to make use of a probability distribution for the discharge week of a patient, instead of using a strict transition from in treatment to discharged based on the length of the care pathway. This can be based on the realised length of stay of patients or based on an expectation of the discharge date by the multidisciplinary treatment team. It is possible to incorporate a stochastic discharge week in the forecastingmodel in the same way as Vanberkel et al. calculates the probability that patient is discharged on a certain day . Another possible extension is to make use of time dependent arrival probabilities. Table 8 gives an overview of the division of requests for care pathways over the year at the SIR care unit. It is possible that there are seasonal patterns in the arrivals of care pathways. When that is the case, it is better to use time dependent arrival probabilities. Another possible extension concerns the workload following from the care pathways where patients receive treatment for. It is possible to make use of stochastic and variable workloads following from the care pathways in the forecastingmodel, instead of deterministic and constant workloads. With stochastic workloads following from the care pathways where a patient receives treatment for, we mean that instead of a deterministic workload it is possible to use a probability distribution for the workload following from the care pathway where a patient needs treatment for, in a certain week. With variable workloads, we mean differences in the workload between the different weeks of the treatment. Furthermore, it is possible to make use of the treatment plans of current patients when these are registered, instead of their care pathways, to improve the accuracy of the prediction.
system as an application of artificial neural networks and fuzzy logic based hourly load demandforecasting with linear polynomial and exponential equation. Multivariate inputs for electrical load forecasting on hybrid neuro-fuzzy and fuzzy C- Means forecaster has been proposed by . The neuro-fuzzy approach was used with additional fuzzy C-Means clustering method before the input enters the network. An intelligent method for medium and long-term energy demandforecasting of a complicated electrical systems has been provided. The demandforecasting using time series modelling and ANFIS estimator has been developed . A clustering based genetic fuzzy expert system for electrical energy demandforecasting has been presented. A novel load forecasting approach has been developed by integration of genetic fuzzy systems and data clustering for extracting a load forecaster expert system . A new approach to short-term load forecasting in a dereg- ulated and price-sensitive environment has been presented. A real-time pricing type scenario is envisioned where energy prices could change on an hourly basis with the consumer having the ability to react to the price signal through shifting electricity usage from expensive periods to cheaper periods when possible. ANFIS modelling has been proven as a reliable forecasting method as compared to other forecasting methods. The number and type of the input variables is playing a big role in increasing or decreasing the forecasting accuracy. As a case study for this investigation, energy demand and load curve data from the Joondalup Campus of ECU is used.
Careful data cleaning is one of the most important aspects of volatility estimation from high-frequency data. This task has been given special attention in e.g. (Dacorogna, Gencay, M¨uller, Olsen and Pictet, 2001, chapter 4), Falkenberry (2001), Hansen and Lunde (2006), Brownless and Gallo (2006) and Barndorff- Nielsen, Hansen, Lunde and Shephard (2009). Interestingly, Hansen and Lunde (2006) show that tossing out a large number of observations can improve the accuracy of volatility estimators. This result may seem counter intuitive at first, but the reasoning is fairly simple. An estimator that makes optimal use of all data will typically put high weight on accurate data and be less influenced by the least accurate observations. The generalized least squares (GLS) estimator in the classical regression model is a good analogy. On the other hand, the precision of the standard least squares estimator can deteriorate when relatively noisy observations are included in the estimation. So the inclusion of poor quality observations can cause more harm than good to the least squares estimator and this is the relevant comparison to the present situation. The most commonly used data source in academic literature is the TAQ data base, NYSE (2007). Several other data sources are surveyed in Wood (2000). Regardless of the data source, a good understanding of the microstructure of the financial market from which the data originates is essential for constructing efficient data filters. Some fairly comprehensive descriptions are provided in Comerton-Forde and Rydge (2004) and (Hasbrouck, 2007, chapter 2 and appendix about U.S. equity markets). Also of interest are Andersen (2000) and Ghysels (2000) who provide valuable discussions of various aspects of analyzing highfrequency data.
Demandforecasting is an iterative process and a critical part of the supply chain that links supply to demand so that service providers have products available when and where they need them. It is essential for a firm to enable it to produce the required quantities at the right time and plan well in advance taking into view various factors of productions. Moreover, it is often critical in better planning for labour and allocation of national resources.
There are two configurations in WEKA Forecasting. They are named “Basic configuration” and “Advanced configuration”. In Basic configuration, the important parameters such as timestamp, periodicity and number of units to be forecasted are available. A timestamp is a sequence of information which identifies an event, and usually it is represented as date and time of day, sometimes accurate to a small fraction of a second. Since the used datasets are seasonal data, the time stamp is set as “Use Artificial Time Index”. This is used only when we are adjusting for trends via a real or artificial time stamp. That means that it will increment the artificial time value with time stamp. In Advanced configuration, the Base learner is available; it has configured parameters specific to the learning algorithm selected. Here the WEKA learning algorithms such as Multilayer Perceptron , Support Vector Machine , Linear Regression , and Gaussian Process  are used for implementa- tion. These algorithms are capable of predicting the numeric quantity. The input data set is available in Attri- bute-Relation File Format (ARFF).
While studies using aggregate data generally must resort to controlling for demand shocks with observable controls and/or a more limited set of fixed effects, several recent papers have attempted identify and utilize instrumental variables to more accurately modeldemand. Hughes et al. (2008) estimate a specification of nationwide monthly gasoline de- mand using crude oil production disruptions as instruments for monthly gasoline prices, but find the resulting elasticity estimates to be nearly indistinguishable from those obtained in their baseline OLS specifications. Davis and Kilian (2011) utilize monthly state-level aggre- gate gasoline consumption and average prices to estimate a first-differenced model in which changes in state gasoline tax rates serve as instrumental variables for changes in gasoline price. Their preferred IV estimate suggests a demand elasticity of −0.46 (s.e. = 0.23), while the corresponding OLS monthly state-level panel regression produces a substantially less elastic estimate of −0.19 (s.e. = 0.04) and an estimation using data aggregated to the na- tional monthly time series level produces an even smaller elasticity of −0.09 (s.e. = 0.04). 5 Davis and Kilian’s IV estimate is much closer to our disaggregated elasticity estimates, sug- gesting that both their IV approach and our disaggregated panel fixed effects approach may be overcoming the potential simultaneity bias caused by baseline demand differences over time. 6 While we don’t observe enough state tax changes during our sample to consider such an instrument, as another robustness check we do estimate an IV specification (described in Appendix C) using regional wholesale spot gasoline prices as an instrument for local retail 5 Coglianese, Davis, Kilian and Stock (2015) point out that the IV estimate of Davis and Kilian (2011) may