國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
7
1. Introduction
Empirical Mode Decomposition (EMD), first introduced by Huang et al. (1998) was developed for dealing with nonlinear and non-stationary data. The method is empirical, intuitive, direct and adaptive. By using EMD, the time-series data is decomposed into several intrinsic modes which are nearly periodic and independent with each other. The intrinsic modes may have its own physical meaning based on local characteristic time scale. For instance, if an intrinsic mode is periodic with a time scale of one month, it can be recognized as the monthly component. Similarly, an intrinsic mode with scale of three months means the seasonal component.
Generally, complex time-series data is often mixed up with many different signal sources thus difficult to understand their meaning. But, through the EMD decomposition method we can divide them into several meaningful intrinsic modes, allowing us to closely analyze their characteristics.
EMD has been successfully applied in different fields such as ocean waves (Hwang et al., 2003), earthquake engineering (RR Zhang et al., 2003), wind engineering (Li and Wu, 2007), biomedical engineering (Liang et al., 2005) and structured health monitoring (R Yan, 2006). The above applications are all related to natural science and engineering. However, in the recent years, there have been more and more applications in social sciences. For example, financial time series analysis (Huang et al., 2003b), transport geography (MC Chen, 2010), disease transmission (Cummings et al., 2004), as well as combined with artificial neural networks (ANNs) to forecast crude oil price (Lean Yu et al., 2008).
In addition to the EMD originally developed, there is an improved EMD, known as Ensemble EMD (EEMD, Wu and Huang, 2009), which was proposed to solve the mode mixing problem of EMD. In this study, we first applied the EEMD method and
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
8
then combined it with ANNs to electricity consumption forecasting and gold price forecasting.
Electricity consumption forecasting (i.e. load forecasting), is commonly classified into four categories: long-term load forecasting, medium-term load forecasting, short-term load forecasting and very short-term load forecasting. Long-term load forecasting (5, 10 and 20 years ahead) is used for system planning, scheduling construction of new generation capacity and the purchasing of generation units (Jia et al., 2001). Medium-term load forecasting (a few months to 5 years ahead) is applied to maintenance scheduling, coordination of load dispatching and the setting of prices (Jia et al., 2001). Short-term load forecasting (hourly forecasting from one day to one week ahead) is usually used for optimal generator unit commitment, fuel scheduling, hydro-thermal co-ordination, economic dispatch, generator maintenance scheduling, the buying and selling of power and security analysis. Very short term load forecasting (few minutes to an hour ahead in the future) is often used for security assessments and economic dispatching, real-time control and real-time security evaluation (Jia et al., 2001). In this study we focused on very short-term load forecasting and forecast load of one hour ahead.
Surprisingly, there has been very little research on very short-term load forecasting.
Yang et al. (2005) used a method based on the chaotic dynamics reconstruction technique and fuzzy logic theory on the load data of Shandong Heze Electric Utility, (China). Their results demonstrated that the proposed approach could calculate 15 minutes ahead load forecasting with accurate results. James W. Taylor (2008) used minute-by-minute British electricity demand observations to evaluate different forecasting methods, including ARIMA modeling and Holt-Winters' exponential smoothing method for prediction between 10 and 30 minutes ahead. Liu et al. (1996) applied the fuzzy logic method and ANNs to the previous 30 minute-by-minute
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
9
observations as input for the online forecasting process to show that the methods could outperform the simplistic non-seasonal AR model.
Unlike the various aforementioned methods, we used the EEMD-based ANN algorithm for one hour ahead forecasting. ANNs are massively parallel and robust.
They contain complicated architectures of interconnected processing elements. They can learn complex linear or non-linear input-output mappings between data sets of measurements and future demand values. Based on the features presented in the data they can be designed adaptively for learning and responding with high-speed computation. The ANNs have been applied to many areas especially when dealing with the issues of forecasting.
Hamid et al. (2004) applied ANNs to financial forecasting. Their goal was to forecast the volatility of S&P 500 Index future prices, and compare volatility forecasts from ANNs with the implied volatility from S&P 500 Index futures options using the Barone-Adesi and Whaley (BAW) model for pricing American options on futures.
Lean Yu et al. (2008) used an EMD-based neural network ensemble learning paradigm for world crude oil spot price forecasting. In Yu‘s study, West Texas Intermediate (WTI) crude oil spot price and Brent crude oil spot price were used to test the effectiveness of the method. Their results showed that the EMD-based neural network ensemble-learning model outperformed the other forecasting models in terms of criteria. Lean Yu et al. (2010) proposed an EMD-based multi-scale neural network learning paradigm to predict financial crisis events for early-warning purposes. They took the currency exchange rate series of the South Korean Won (KRW) and Thai Baht (THB) as training targets. Their tests showed that the EMD-based multi-scale neural network learning paradigm was superior to other classification methods and single-scale neural network learning paradigm when formulating currency crisis forecasting. Feng Ping et al. (2009) applied EMD-based ANNs to precipitation-runoff
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
10
forecasting. They took the annual precipitation series from 1956 to 2000 from the sub-water resource regions of upper Lanzhou, China as historical training data, and showed that the EMD could decompose the data into a multi-time scale sub-series for finding their local change rule. The results demonstrated that the EMD-based ANNs model presented higher accuracy than any other models.
The ANNs are also widely used in short-term load forecasting. Ruey-Hsun Liang and Ching-Chi Cheng (2002) used an approach based on combing ANNs with the fuzzy system and applying it to data from the Taiwan Power Company. Nahi Kandil et al. (2006) applied multi-layered feed-forward ANNs by using data from Hydro-Quebec databases for forecasting. They demonstrated ANNs‘ capabilities without using load history as an input, instead final temperatures were the only data considered in their load forecasting procedure. Mohsen Hayati and Yazdan Shirvany (2007) used Multi-Layer Perceptron (MLP), a kind of architecture of ANNs, on data from a three year time period (2004-2006) from the Illam (Middle Eastern country, west of Iran) region, while G.A. Adepoju et al. (2007) applied ANNs to the Nigeria Electric power system.
Gold has been mined since ancient times. With recent growth in production, more than a third of the world‘s gold that has ever been mined, in just the last thirty years.
The consumption of gold differs by application type: industrial, dental technology, jewelry products and inventory. Jewelry consistently accounts for over two-thirds of the gold demand, but in markets with poorly developed financial systems or markets experiencing crisis, gold is an attractive investment. The demand-supply equilibrium and inflation cause gold price to fluctuate. Gold is commonly a popular hedge instrument for investors against devaluation of the US dollar. In recent years as the value of US dollar has decreased relative to other major currencies, the price of gold
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
11
has experienced a secular increase. The dramatic rises in gold price since the start of 2009 may have resulted from investors looking to preserve their wealth.
There have been several studies on gold price analysis: Baker and Tassel (1985) used regression results to support the theoretical analysis leading to the prediction.
Akgiray, Booth, Hatem and Mustafa (1991) used the GARCH model to verify time decency of gold price. In comparison with statistical techniques, engineering- based systems, such as neural networks, make less restrictive assumptions on the underlying distribution. Mirmirani and Li (2004) used neural networks and genetic algorithm to analyze gold price. Shahriar Shafiee and ErkanTopal (2010) used long-term trend reverting jump and dip diffusion model and took monthly historical gold price data from January 1968 to December 2008, to forecast gold price for the next ten years.
Yen-Rue Chang (2011) used EEMD to decompose monthly gold price data into several IMFs to observe their important properties. Following the work of Yen-Rue Chang (2011), we used the decomposed IMFs as input factors for gold price forecasting.
This research aimed to forecast using EEMD-based ANN algorithm, more specifically; we used a back-propagation neural network (BPNN) which is a kind of ANN architecture. Our testing targets were electricity load data and gold price data.
We performed one hour ahead load forecasting and gold price forecasting for 2011.
Section 2 gives brief introduction to the basic concept of EEMD and BPNN. In section 3, we describe the subject data, and introduce our experiment strategy, and some important measures used. Experiment results and forecast performance are discussed in Section 4, along with the results of improvement by ensemble average. In Section 5, we will present our conclusion and outlook.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
12