National-scale electricity peak load forecasting: Traditional, machine learning, or hybrid model?

• Accurate peak-load forecasts support power-system operation and generation planning, while forecast errors can impose costs from unexpected or surplus generation.The passage links rising demand to economic growth, electrification, and climate change, and describes the economic and financial consequences of forecast errors.Source: §1 · p. 3
• National-scale peak-load forecasting is presented as essential for long-term energy planning, beyond the building- or region-level focus of much prior work.In South Korea, the Korea Power Exchange uses the short-term KSLF to forecast next-day demand for power-system operation.Source: §1 · p. 3
• Time-series models can forecast acceptably under constant trends but may lose accuracy as markets change rapidly and nonlinearly; machine-learning models can process nonlinear data.The introduction also notes that hybrid models combine time-series and machine-learning methods, following Zhang.Source: §1 · p. 4
• Prior studies report that exogenous factors such as weather and calendar variables improve forecasting, helping explain the use of ARIMAX and SARIMAX over other time-series models.The introduction identifies SARIMAX as the most appropriate time-series model for peak-load forecasting among the models discussed.Source: §1 · p. 4

Introduction

• Prior hybrid studies report that ANN captures nonlinear peak-load components missed by ARIMA-based models; Zhang et al. and Nie et al., comparing ARIMA, SVM, and hybrid models, found ARIMA–SVM had greatest predictive power.Chou and Tran reported ANN outperformed SVM and other single models, while SVM improved hybrid predictive power more than ANN.Source: §1 · p. 8
• Prior studies associate peak load with weekends, holidays, and weather variables; machine learning typically outperforms time-series models, but the strongest machine-learning model varies, while hybrids often outperform single models.The cited weather variables include temperature, humidity, rainfall, and wind speed.Source: §1 · p. 9
• Previous peak-load studies mostly targeted buildings or specific regions, leaving country-level forecasting without a comprehensive comparison.The study presents its national-scale analysis as a basis for broader methodological and policy directions.Source: §1 · p. 9
• The study compares time-series, machine-learning, and hybrid models, tests whether hybrids improve on single models, and evaluates its model against KPX’s current approach.It also aims to derive electricity supply-and-demand implications regarding reserve margins and suggest improvements to Korea’s forecasting model.Source: §1 · p. 9

2. Methodology

• The study compares SARIMAX with ANN, SVR, and LSTM, as well as hybrids that combine time-series and machine-learning models.It also describes evaluation statistics and summarizes the research steps in a flowchart.Source: §2 · p. 10, Figure 1 · p. 10

2.1. SARIMAX

• SARIMAX extends seasonal ARIMA by incorporating exogenous variables, with model orders selected using correlograms or information criteria.The passage also notes that R’s auto.arima package can reduce the time required for these selection processes.Source: §2.1 · p. 11

2.2. ANN

• For ANN fitting, the study uses a NARX architecture suited to peak-load time series influenced by external variables, with a multilayer perceptron mapping inputs to outputs.The NARX architecture uses a multilayer perceptron for the mapping function.Source: §2.2 · p. 12

2.3. SVR

• SVR performs regression by finding a flat function that keeps training targets within a specified deviation, and this study tunes its hyperparameters using grid search.The tuned hyperparameters are the box constraint, epsilon margin, and kernel function.Source: §2.3 · pp. 13–14

2.4. LSTM

• The study uses basic LSTM, which addresses the RNN vanishing-gradient problem with a cell state and gates that update, maintain, or forget prior information.The described gates are the input, forget, and output gates.Source: §2.4 · pp. 15–16

2.5. Hybrid model

• The hybrid approach combines a time-series model with a machine-learning model.The passage introduces the hybrid model’s expression but does not provide its equation in the supplied text.Source: §2.5 · p. 17
• The hybrid model adds a machine-learning estimate of the residual to a time-series model’s linear estimate, defining “hybrid” by this combination.Source: §2.5 · p. 17

2.6. Evaluation of the predictive power

• Predictive power is compared using RMSE and MAPE, which capture scale-dependent and percentage-dependent error, with MSE and MAE also reported.Source: §2.6 · p. 18

3. Data and Results

• From 2014 to 2019, peak loads generally rose, with summer and winter increases and lower spring and autumn loads; weekends and holidays also had reduced loads.New Year’s Day and Thanksgiving loads were significantly lower than those on other days.Source: §3 · p. 18, Figure 5 · p. 19
• The compared models used mean temperature, mean humidity, and Saturday, Sunday, and national-holiday features as exogenous variables; among chosen variables, mean temperature was most highly correlated with peak load.Peak load was positively correlated with mean temperature in summer but negatively correlated in winter.Source: §3 · pp. 19–20

3.1. Performance comparison

• All machine-learning and hybrid models outperformed SARIMAX; LSTM led the single models, while SARIMAX-LSTM showed the greatest hybrid improvement.Hybrid models also had higher adjusted R-squared values than single models.Source: Table 2 · p. 26, §3.1 · pp. 26–27
• SARIMAX-LSTM had a smaller RMSE than LSTM, while LSTM had a smaller MAPE; the differences were marginal.The study could not determine which model was superior by simply comparing these accuracy criteria.Source: §3.1 · p. 27

3.2. The highest peak load of test data

• The LSTM model predicted the highest 2019 peak load at 89,800 MW and significantly outperformed all other models.Model performance was relatively low on days when peak loads sharply increased or decreased; the highest peak was not explained by the temperature, humidity, or holiday variables used.Source: §3.2 · p. 29
• The reported greatest result was for the SARIMAX-LSTM hybrid and single LSTM; LSTM outperformed other models on transient changes unexplained by meteorological, holiday, or weekend variables.Source: §3.2 · p. 29

3.3. Comparison with Korea’s current forecasting model

• KSLF combines a panel macroeconomic model and auxiliary time series models, updating next-day peak-load predictions to reflect current peak-load temperature volatility.Source: §3.3 · p. 29
• KSLF predictions were typically below actual values.Source: §3.3 · p. 30
• In 2019, actual peak load was 91300 MW versus KSLF’s 83440 MW prediction (RMSE 3531.0519; MAPE 4.37%); LSTM, SARIMAX-SVR, and SARIMAX-LSTM beat KSLF overall and at peak.The predicted and actual KSLF peak loads are shown in Figure 12.Source: §3.3 · p. 30, Figure 12 · p. 30

4. Conclusions

• All single and hybrid machine-learning models significantly outperformed SARIMAX in predictive power.Source: §4 · p. 31
• LSTM and SARIMAX-LSTM were best within single and hybrid models, respectively, making LSTM-based models comparatively most accurate for Korea.Source: §4 · p. 31
• LSTM and SARIMAX-LSTM showed no significant overall performance difference, although LSTM predicted sharp increases and decreases better.Source: §4 · p. 31
• LSTM and SARIMAX-LSTM outperformed Korea’s current peak-load forecasting model; the authors say including LSTM-based models could improve current forecasts.Source: §4 · p. 31
• The findings are Korea-specific; the authors say LSTM’s lead is unproven in countries with smaller annual temperature or meteorological variation.Source: §4 · pp. 31–32
• Hybrid predictive power depends on how well the time-series component fits peak-load linearity, making component-model fit a key design condition.Source: §4 · pp. 32–33
• The study omitted the latest extended models, including ANN- and SVM-extended variants, and recommends future comparisons with boosting models.Source: §4 · p. 33
• The study compared daily peak-load forecasts; the authors call for future analyses of electricity consumption and comparisons across time units and horizons.Source: §4 · pp. 33–34