||Short-term demand forecasting ,Accuracy improvement ,Impact analysis ,Artificial neural network ,Multiple linear regression ,Bayesian estimation ,Atmospheric variables ,Special days ,Base temperature ,การพยากรณ์ความต้องการระยะสั้น ,การปรับปรุงความแม่นยำ ,การวิเคราะห์ผลกระทบ ,โครงข่ายประสาทเทียม ,การถดถอยเชิงเส้นพหุคูณ ,การประมาณแบบเบย์ ,ตัวแปรบรรยากาศ ,วันพิเศษ ,อุณหภูมิฐาน
||Accurate electricity demand forecasting for a short time horizon is an important issue for day-to-day control, scheduling, operation, planning, and stability of power systems. Deterministic variables such as calendars, special events, and weather variables such as temperature are key factors that affect the forecasting accuracy. Due to the tropical climate of Thailand, the marginal impact of weather variables on electricity demand is worth analyzing. This dissertation primarily focuses on the impact of atmospheric variables such as temperature and deterministic variables such as calendar on electricity demand for Thailand and Japan.For Thailand, we present a novel approach for short-term electricity demand forecasting (STDF) by constructing different scenarios so that the model can be trained with an appropriate training dataset. We divide the whole dataset into groups, hereafter named as scenarios, on the basis of day types. Scenario 1 consists of the dataset of working days, Scenario 2 consists of the dataset of the weekend, Scenario 3 consists of the dataset of holidays, and Scenario 4 consists of the whole dataset of all day types. For each scenarios, selection of the proper variables for a model, and the length of the training dataset is conducted based on experiments. Statistical and data-driven models are constructed for forecasting purpose. The performance on forecasting accuracy measured in terms of mean absolute percentage error (MAPE) for Scenario 1 is obtained as 2.72%, 1.88%, and 1.97% for feed forward artificial neural network (FF-ANN), general least square auto-regression (GLSAR), and ordinary least square (OLS), respectively. This means the proposed simple OLS and its extension to GLSAR methods outperformed over the deep learning methodology on forecasting accuracy among four scenarios. Nevertheless, GLSAR takes higher execution time than OLS, but significantly lower than FF-ANN.Normally, the atmosphere of Thailand is hot during day hours and the general belief is that this hot temperature results in high demand for electricity during day hours. The minimum increment rate of demand approximately 50 MW per rise occurs at the morning hours (6 a.m.–7 a.m.) and evening hours (4 p.m.–6 p.m.). The low impact on morning and evening is because people are at no work condition as well as the industrial shift changing time. Throughout the day, the impact of temperature is responsible to increase by 100 MW– 200 MW per 1◦C rise in temperature. However, the result reveals that the maximum impact of temperature on electricity demand is found during night hours rather than day hours at 11 p.m. The comparative study concludes that scenario 1 gives the best forecasting accuracy for working days from simple OLS methods. However, Scenario 3 shows the worst accuracy from both the models for holidays. Therefore, it is better to choose Scenario 4 (all the data together) rather than Scenario3 (holidays data) to forecast the holidays.We also present a study on short-term load forecasting for Hokkaido, Japan. Since Hokkaido is considered a sub-tropical climate, its load shows significant dependence on many more weather variables, such as temperature, humidity, wind speed, solar radiation, and cloud cover, than Thailand load. Models are constructed on the basis of including and excluding variables and Bayesian methodology is implemented for forecasting. The meteorological dataset is collected from several weather stations of different locations. Since the study is focused on the impact analysis of weather factors, base temperature estimation and selection of the best weather station are the key issue and conducted on the basis of weighted average and piecewise linear approximation methodology. Multiple linear models including auto-regressive moving average with exogenous variable (ARMAX) are constructed and analyzed the performances. Three different models are constructed to compare their performance. The Bayesian approach is applied to estimate the weight of each variables using Gibbs sampling to approximate the estimation of the coefficients. The overall MAPE performance for a year is found as 1.72% which is 13.4% improvement by including rainfall, snowfall, solar radiation, wind speed, relative humidity, and cloud cover data.