Cambodia Mid-Term Transmission System Load Forecasting with the combination of Seasonal ARIMA and Gaussian Process Regression

Abstract

Mid-term load forecasting (MTLF) is crucial for power transmission system planning, safe operation, and maintenance. It's very significant in a developing country like Cambodia. In the past decade, many kinds of research have been developed for MLTF. However, the specific research of MLTF for Cambodia's transmission system is still insufficient to meet the optimal planning for the rapid growth of the demand. This is the reason that Cambodia would require more research work to be contributed to this field. This paper presents an attempt at a medium-term forecast (one year ahead prediction) for the transmission system load of the Electricite Du Cambodge (EDC) company in Cambodia with the monthly index of historical data from 2010 to 2019, using a hybrid model of combining the Seasonal ARIMA (SARIMA) and Gaussian Process Regression (GPR) model. The data sets were split into two sets of train (2010–2018) and test (2019) data. The proposed model procedure had two main steps which the first step was to train the SARIMA from train data sets then simulated the residual values and estimated the load values. Next, we summed the residual values to the system load values in the train data sets and used that summed up values as new variables to train the kernel-based GPR model while the estimated load from SARIMA was kept as another new variable for simulating the trained GPR model. The other variables such as dates, months of the years, relative humidity, wet-bulb temperature, and total non-working days of each month also were taken into account to train with the GPR model. Three different kernels function were tested in this research work and the modelings were conducted on Matlab 2019b. The Mean Absolute Error (MAE) and the Mean Absolute Percentage Error (MAPE) is used as the indicator to evaluate the performance of the simulated data from models to the actual data. The obtained results have shown that the MAPE of SARIMA and GPR were 3.46 percent while the proposed model performed by 1.95 percent.