A Taguchi and Neural Network Based Electric Load Demand Forecaster

In this paper, we present Taguchi’s and rolling modeling methods of artificial neural network (ANN) for very-short-term electric demand forecasting (VSTEDF) from the consumers’ viewpoint. The rolling model is a metabolism technique that guarantees input data are always the most recent values. In ANN prediction, several factors that may influence the model should be well examined. Taguchi’s method was employed to optimize the parameter settings for the ANN-based electric demand-value forecaster. Our experimental result shows that the optimal settings of ANN prediction model are 3 lagged load points, 0.1 for the momentum, 5 hidden neurons and 0.1 for the learning rate. The error of forecasting is as small as 3%. That is, comparison with the results of ordinary ANN and Grey prediction, the presented Taguchi-ANN-based forecaster gives more accurate prediction for VSTEDF. INTRODUCTION Due to the lack of natural energy resources, over 95 percent of the energy consumed in Taiwan is imported from overseas, and owing to the growth of the economy and global market competition, the supply and demand of high-quality and inexpensive electric power has become an important issue to electric power plants, business owners, and government. In order to assess electricity efficiently and supply high-quality electricity to the consumers economically, electric power companies face financial and technical challenges. Tracking electric load generation at all times and knowledge of the future load is a basic requisite in the efficient operation of power-generating facility. In the past few decades, numerous researchers have presented different methods for electric load forecasting. Among these studies, many models for load forecasting have been proposed, such as models of time-series analysis, regression analysis, artificial neural networks (ANN) and Grey prediction. These models have been reported and documented demonstrating success in long-term forecasting, medium-term forecasting, short-term forecasting and very-short-term forecasting. A number of researchers have adopted ANN for load forecasting. Fung and Tummala (Fung 1993) [1] combined the economic factors such as electricity price, gross domestic product, and the weather conditions to predict the long-term load consumption for each electricity market in Hong Kong. Charytoniuk and Chen (Charytoniuk 2000) [2] moved their focus toward predicting relative changes in a daily load base. Instead of modeling relationships between load, time, weather conditions, they extrapolated the recently observed load patterns to the nearest future. Srinivasan et al. (Srinivasan 1991) [3] used a three-layer-architecture back-propagation networks model (BPN) to predict Address correspondence to this author at the Department of Mechanical and Automation Engineering, National Kaohsiung First University of Science and Technology, Kaohsiung; Taiwan; E-mail: [email protected] the daily peak load in Singapore. They conclude that using a lower learning rate for the layers gives accurate prediction. Lu et al. (Lu 1993) [4], Kiartzis (Kiartzis 1995) [5], Khotanzad (Khotanzad 1998) [6], Chow (Chow 1996) [7] and Bakirtzis (Bakirtzis 1996) [8] used different BPN models to predict hourly load, daily total load and daily peak load. They also show that there is no firm criterion for training inputs selection. However, the load demand forecasting from the viewpoint of consumers is seldom discussed. With global marketing competition, national and multinational enterprises have tried very hard to cut down on their expenditures in order to increase their competitiveness. Additionally, power companies generally have many different rate schedules and penalty policies of poor-power-factor loads for their customers. Therefore, the practice of demand-control of electric systems has drawn a lot of attention from government, manufacturers and researchers. Likewise, the electric bill is always one of the largest monthly bills in educational institutes. Yao and Ku (Yao 2003) [9] developed and installed a cost-effectively network-based and PC-based automated monitoring and control electric system (AMCES) at the National Kaohsiung First University of Science and Technology (NKFUST), Taiwan, to monitor and control the electric load on campus. AMCES (see Fig. 1) is a closed-loop demand-control system with intelligent predictor embedded and networked with a PC-based Human-Machine Interface and Data Acquisition system (HMI/DAS). AMCES provides an effective solution to electricity management. The aim of this project is to design a dynamic forecaster for VSTEDF. We present a Taguchi-rolling-model-based method for establishing the training data of ANN to overcome the deficiency of BPN on VSTEDF. This paper is organized as follows. Section 2 shows our preliminary forecasting results using ordinary BPN. Section 3 presents the development of a rolling model of BPN for VSTEDF. Section 4 presents the results and discussion. Section 5 ends with the conclusions of this study. 8 The Open Automation and Control System Journal, 2008, Volume 1 Yao et al. Fig. (1). Configuration of AMCES and its components. It is based on PC-based OMC, a PC-based visual HMI/DAS and digital power analyzers. PRELIMINARY FORECASTING BY CNVENTIONAL BPN The past research activities of ANN-based load forecasting often involved the daily peak load or total load with the factors of temperature, seasons, or day-types. Generally, a typical ANN architecture for electric load forecasting is shown in Fig. 2. It generally includes temperature, electric load and day-types as inputs for ANN training. The selection of input variables for ANN training always plays an important role in the accuracy of prediction. The model of ANN is establishing the relationships between the inputs and outputs through the training operation process. Generally, if the inputs represent less related to the predicting output, ANN may not be expected to accurately predict the outputs (Drezga 1998) [10]. However, there are no general rules for selecting input variables. Efficiently selecting appropriate variables depends on experience or preliminary tests (Swarup 2002) [11]. Fig. (2). A typical ANN architecture for electric load forecasting. It often includes the temperature, load, and the day-types as training inputs for the neural network model. In our preliminary study, we adopted ordinary BPN model and considered the electric loads and the temperatures of the last 2 minutes and 4 minutes, and the temperature of the next 2 minutes as the training inputs for the BPN. Then, using this established model to forecast and examine the electric load for the next 2 minutes (see Table 1). The forecasted electric demand values of ordinary BPN are shown in Fig. 3. The total averaged absolute error is about 14%. It concludes that the ordinary BPN model does not deal adequately with a volatile system like VSTEDF. The volatility of the usage of electricity on NKFUST’s campus is possibly due to the load characteristics of lamps, motors, and air-conditioning varying according to the class schedule and the operation of equipment. Weather related effects are also significant factors. Heat waves and humidity in the southern Taiwan are major factors of the influence of electric load in summer. Specifically, the electric load for lighting and air-conditioning is highly influenced by the temperature and the presence of clouds. However, the other uncontrolled factors of the electric load of the indoor activity are the class schedule, the attendance of students, laboratory hours, and so on. The pattern of electric load consumption is not always regular in the daily usage of the electricity on campus. From the preliminary study, the distribution plot of the load versus the temperature is shown in Fig. 4. As can be seen, it indicates that the factor of temperature is unrelated with the electric load. Therefore we considered the relationship of electric load and temperature as a relatively large time constant and ignored the temperature input in our study. But we trained our ANN model with the pattern recognition of temperature. The procedures are giving as follows. Table 1. The Description of Inputs for VSTEDF by Using the Ordinary BPN Model. Input Nodes 1 and 2 are loads of the Last 2 min. and 4 min. Input Nodes 3 and 4 are the Temperatures of the Last 2 min. and 4 min. Input node 5 is the Temperature of the Next 2 min. The Output Node is the Load Prediction Input Node Variables 1 & 2 Load (M – i ), i = 2, 4, 3 & 4 Temperature (M – i ), i = 2, 4, 5 Temperature for the next 2 minutes (M + 2) Output node Forecasted load