Forecasting annual electricity consumption in Vietnam using radial basis function neural network

This paper presents a machine learning (ML) model based on a radial basis function neural network (RBFNN) to forecast Vietnam's long-term annual electricity consumption for the period 2024–2030. Eight models with different combinations of socioeconomic input variables were constructed, and the mutual information and SHapley Additive exPlanations (SHAP) method were used to determine the most appropriate combination for the optimal forecasting model. The collected dataset was split into training (1990–2015) and validation (2016–2020) sets to train and validate the proposed models. To forecast electricity consumption from 2024 to 2030, gross domestic product values followed the government's scenario, and other input variables were obtained using the linear extrapolation method from 2016 to 2020. The forecasting capability of the optimal model with the extrapolated input data was tested for 2021 to 2023 by comparing the forecasted consumption with the actual consumption obtained from the Vietnam Electricity Corporation. The electricity consumption forecast in the power development plan VIII (PDP VIII) was compared and discussed. The results showed good agreement between the consumption forecasted by the RBFNN and actual consumption, whereas PDP VIII overestimated the actual values. For 2024 to 2030, the average annual growth rate of electricity consumption obtained by the RBFNN is expected to be 7.2 %, whereas PDP VIII predicts an average annual growth rate of 8.7 %. Additionally, a comparative analysis was also conducted to assess the performance of various ML models. The results show that, despite fitting the training data well, most ML models except the RBFNN failed to accurately predict electricity consumption during the validation period due to extrapolation challenges arising from the increasing trend in input variables.