Application of Statistical and Machine Learning Models in Vietnam’s Energy Consumption Demand Forecasting

Energy consumption demand forecasting plays a critical role in the planning and development of national energy security, which underpins the Vietnam’s Eighth National Power Development Plan (PDP VIII) and Vietnam’s ambitious Net-Zero 2050 commitment. However, this task becomes more difficult because the big data environment is filled with a lot of noise and highly fluctuating data. In order to deal with the problem, this paper evaluates five models: Linear Regression, Holt’s Exponential Smoothing, PSO-GM (1,1), Support Vector Regression (SVR), and Random Forest as a benchmark to conduct a rigorous comparative analysis to identify the most accurate forecasting model. The performance was evaluated by MAE, RMSE, and MAPE indexes based on Vietnam’s total primary energy demand data from 1986 to 2024. To check the accuracy of the forecasting model, this study split the data into two periods: first time for the training data (1986–2016), and second for the testing data set (2017–2024). Furthermore, a five-fold rolling-window time-series cross-validation method and Diebold–Mariano tests were employed to ensure the statistical robustness of the findings on the small-sample datasets (n = 39). The results decisively identified that the Holt’s model as the superior framework, maintaining high stability and achieving a testing MAPE of 7.19% (training MAPE of 5.52%), while the complex machine learning benchmark shows severe over-fitting. Applying this model, Vietnam’s energy demand will reach 1528.08 TWh and 1882.55 TWh in 2025 and 2030, respectively. Furthermore, this study provides empirical evidence that simpler, well-chosen statistical models can surpass complex alternatives in small-sample scenarios, offering a reliable quantitative baseline for policymakers to navigate infrastructure development and decarbonization challenges.