| 摘要 | ABSTRACT
1. Research Background and Purpose
��The increase in renewable energy generation will promote the shift from a centralized system centering on existing nuclear power and coal power generation to a decentralized system. Also, if the volatility of renewable energy becomes large, the need for more accurate forecasts of electricity demand will increase for smooth matching of demand and supply. In addition, more accurate modeling of power and energy demand forecasts will become more important in the near future if sales in the energy and electric power markets are somewhat possible, as in the developed countries.
��In this context, it is essential to develop various forecasting models. The results of forecasts depended on the existing individual models have been recorded in many cases in which the forecasting power is damaged in other periods even though they were accurately predicting in a certain period. In this study, we apply the model of forecast combination, which is a combination of various individual models for power demand, as one of the methods to solve this instability surrounding individual models.
��The point forecast estimates of a model is affected by the type of the model and the information set. The point forecast estimates contains various characteristics such as the uncertainty of the model, parameter estimation error, overfitting problem, etc. Thus, there is always the possibility that the predictive power of an individual model may decline over time, regardless of the precise predictions during a short timing of the forecast. Forecast combination is a model that literally combines the predictions of individual models. According to the forecasting literature, it has been reported that the accuracy of these forecast combinations is better and more robust than the individual models when the evaluation period is longer. In addition, since there are not many cases where the forecast combining method, which is widely used in macroeconomic variables forecasting, is applied to Korean domestic electric power demand, we introduce various forecast combining methods and analyze them empirically to see how these modeling is actually effective in Korea's electric power demand. Market participants also avoid uncertainty due to electricity demand forecasts. Therefore, information on these uncertainties should be necessary for making important decisions, such as making market entry decisions.
��The first objective of this study is to try out various forecast combination models for various individual models and to characterize them in case of Korean power demand. The second objective of this study is to construct a predictive density function. This process is necessary because it is possible to provide more information for decision making of market participants by presenting the distribution considering the uncertainty of the point forecast estimate.
2. Summary
��The main purpose of this study is to expand the point forecast estimate, which is the outlook for individual models. Also, it is to build predictive density function and provide prospect information abundantly. Existing predictions related to power can be categorized into the application of demand forecasting to forecast combinations and the study of density function building. However, both studies have not actively conducted preliminary researches both domestically and internationally. Therefore, this study is expected to contribute not only to domestic power demand research but also to overseas research by applying forecast combination and density function to domestic electric power demand forecast.
��This paper presents forecast combinations and probability density function methodology for power demand forecasting. The first methodology estimates new predictions based on predicted values from individual models as various prediction combining methods. Various typical time series models (ARIMA, ARIMAX) were used as individual models. Estimation of individual model predictions was also made by using various models of Artificial Neural Network Model and Exponential Smoothing Method. The forecast combining method is to estimate the weights of the individual prediction values. The forecast combining method includes a simple prediction combining which averages over the predicted values, a prediction combining which assigns weights according to the accuracy of the prediction values or by OLS method, respectively. In addition, we compared the predicted values of the Top 5 and Top 10 models, which were ranked with the forecasted values from the individual models, and the results of their forecast combinations.
��Various forecasting combination models were applied to the weekly, monthly and annual electricity demand data. The main results are as follows. In the case of weekly demand forecasting, the ANN model showed the most accurate forecasting performance, and among the exogenous variables, the variables such as the degree of heating and cooling were found to be very helpful. In the forecast combination, the model with the combination of five individual models has the best performance, wheras the accuracy of the other forecast combination models having to estimate the weights is relatively low. In the case of monthly power demand forecasting, various forecast combining methods were analyzed to have good predictive power. The Bates and Granger method, which is a method of estimating weights, also showed the sharp accuracy of forecasts. The Granger and Ramanathan model, which estimates weights with OLS, showed a slightly poorer accuracy than those of these models, but it outperformed the accuracy of most other individual models. In the annual forecast, it showed good accuracy comparable to that of the predictive combination model of five 10 models with good performance and the best predictive model among the individual models, and the prediction model with weighted values considering the level of information standard also showed good results.
��The forecast density function for the last data of the sample for weekly, monthly, and yearly forecasts is constructed and presented together with the measured values. In the case of these density function, two combinations are considered. One is the weight using the simple average and the other is the weighting method using the logarithm score function. The results can be summarized in three major ways. First, even if the point estimate is relatively accurate, the variance of the distribution can be large. This means that there is a large uncertainty in one prediction model and all the possibilities applied to the data, suggesting that there will be more variation in applying these individual models to the prediction in the future. Second, even if one prediction is inaccurate, the distribution can be small in variance and can be useful for forecast density combination. In addition to the other models that are likely to be underestimated, those that are highly likely to be overestimated can be used in forecasting to approximate the observed values. Third, the density function is approximated to the measured value better when the weight is estimated by the log score function. This shows that the weighting method of the logarithm score method is more effective for the density function than the simple average weighting scheme.
3. Suggestions for Further Research
��Unlike high-frequency weekly forecasts where there is no observed data other than meteorological variables, it is likely that a monthly and yearly demand with economic variables to be considered will need to be considered in a way that allows them to be considered together with various macro and micro variables. There is a need for a systematic study on how the influence of the information set on the predictive power increases as the information set increases. Adding a variable increases the information set, but there is always the possibility of overfitting in the model, and the prediction in this case is likely to be poor. In addition to considering the nonlinear relation of the artificial neural network model performed in this study, it is necessary to further studies the performance from the higher order artificial neural network model or other complex models or from various time series models with nonlinearity.
��In the case of weekly forecasting, it is necessary to study the weather variables more precisely because the prediction of the model including the cooling degree and the heating degree of the weather is most accurate. In addition, it is necessary to construct the macro variables that do not exist in the weekly or in the daily through the estimation using the interpolation method and to apply it to the forecasting of the weekly or daily power variable. |
