Modeling regional peak load forecasting using dynamic narx neural network with temperature

• Main Author: Basaruddin, Faridah
• Format: Thesis
• Language: English
• Published: 2014
• Online Access: http://psasir.upm.edu.my/id/eprint/64727/1/FK%202014%20154IR.pdf; http://psasir.upm.edu.my/id/eprint/64727/

Temperature is one of the most significant weather parameters affecting load consumption. Temperature varies according to demographic region and could not be incorporated in the Malaysia load forecasting as the latter emphasized a general style of aggregate forecasting that predicts the load consumption for the whole of the country. Such load forecasting results would not be able to identify where the power load takes place and also is not helpful for power facilities construction location planning. Therefore the models are inadequate to predict the control of load in critical situations such as during drought or monsoon seasons that occur at certain time of the year or occasionally when weather is unpredicted. Hence it is of interest to implement a model that serves the above purposes as well as to improve on the supply of load. A regional peak hourly load forecasting at selected meteorological stations in Malaysia using Dynamic Narx Neural Network Model is implemented. The advanced dynamic Narx neural network model (NARXNET) without and with temperature is applied to peak hourly load forecasts at selected meteorological stations in Malaysia. The performances of both models are compared with time series, Auto Regressive Integrated Moving Average (ARIMA), ARIMA transfer function with temperature model and another neural network, Focused Time Delay (FTDNN) model in terms of parameters investigations and models’ performances. NARXNET forecasts for the week ahead peak hourly load achieved Mean Absolute Percentage Error (MAPE) ranging from 0.3422 to 0.9066 while the five hundreds -hours ahead peak hourly load at the stations with MAPE 0.0109 to 0.1733. NARXNET with temperature model forecasts for the week ahead peak hourly load produced MAPE ranging from 0.2773 to 0.6533 and the five hundreds -hours ahead peak hourly load forecasts gave MAPE ranging from 0.0248 to 0.1391. NARXNET with temperature model is able to capture the effect of temperature on the peak hourly load system at three out of five stations. ARIMA and ARIMA transfer function with temperature, for the five hundreds - hours ahead peak hourly load forecasts , however gave MAPE that ranged from 2.700 to 5.390 and 2.702 to 5.393 respectively. The effect of temperature using ARIMA transfer function is captured only at two out of five stations and the improvement in the forecast is very small. The experimental results have shown that NARXNET with temperature model due to the existence of feedback connection where the outputs are regressed to the network is capable of improving the forecasting performance through the effect of temperature. Being part of neural network, NARXNET is seen as a promising black box model in identifying a nonlinear system without/with prior knowledge. Thus Narxnet can be used for real time simulations. The simulation results proved that NARXNET, having the ability dealing with nonlinear data outperformed ARIMA and ARIMA Transfer function models. The excellence performance of NARXNET in dynamical modeling was supported by studies conducted by Lin et al.(1997), Luo and Puthusserypady (2006), Nordin (2009) and others. As both historical temperature and load data were applied to the NARXNET model, this research also considered some aspect of regression analysis involving load and weather parameters with more emphasis on temperature–load relationship. The historical peak hourly load and temperature data for a period of one year were applied to both models, NARXNET and regression. The trend of the peak hourly load consumption for a selected week and load profile on a selected day within the study period were analyzed. The analysis provides better understanding on the characteristics of Malaysian power load system. The implementation of the model was validated by comparing with other existing works. Both the validation and simulation results were similar. It can be concluded that NARXNET with temperature model performed better than other models that use temperature and thus by applying NARXNET model to predict the electricity consumption at locations that are affected by extreme changes in temperature, the problem of over production or under production of electricity that in turn influence the sustainable development of the economy could be overcome.