Electricity load forecasting using hybrid prophet–TFT model incorporating weather data

Accurate forecasting of electrical consumption is essential for controlling the integration of renewable energy sources and improving the electric grid’s performance. We introduce a new hybrid forecasting method in this study that combines Facebook Prophet’s time-series decomposition with the deep learning capabilities of a Temporal Fusion Transformer (TFT). By leveraging Prophet to model long-term trends and seasonality and using TFT to capture short-term patterns and complex multivariate relationships, the hybrid model effectively bridges statistical and neural forecasting techniques. We incorporate real-time meteorological features from Jeju Island (using data from 2012–2020) to enhance forecast adaptability to weather variations. The proposed Prophet–TFT model achieves high accuracy, outperforming each individual model and other benchmark hybrids. It significantly improves upon baseline predictions, achieving an MAE of 8.4 MW with a MAPE of about 1.43%. The inclusion of weather data, particularly during extreme temperature conditions, further enhances forecasting precision. An attention-based feature importance analysis reveals that key meteorological variables, such as Jeju’s air temperature, Seogwipo’s wind speed, and a heat-discomfort index, are among the most influential predictors. These results highlight the significant role of environmental conditions in shaping electricity consumption. The results illustrate the remarkable efficacy of the hybrid model, interpretable and flexible. It offers a practical tool for grid operators, especially in regions with significant renewable resources and weather-dependent demand, to reliably forecast loads and enhance the stability of power systems.