This study introduces a Short-Term Load Forecasting (STLF) method that leverages multimodal data fusion combined with Bayesian Optimization (BO) to enhance prediction performance. By integrating heterogeneous information including historical load data, meteorological data, temporal characteristics, and social media sentiment, a high-dimensional feature space is constructed. Mutual information is employed to screen key variables, improving input data quality. A multi-layer LSTM (Long Short-Term Memory) architecture simulates temporal load patterns, integrated with Bayesian optimization (BO). A Gaussian Process (GP) surrogate guides efficient hyperparameter exploration via Expectation-Improvement (EI) sampling, circumventing the limitations of exhaustive grid search. Validation using real PJM grid data (2020–2022) shows significant performance gains over competitors—including ARIMA, SVR, Prophet, and standard LSTM—evidenced by lower Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) metrics. Notably, MAE decreased by 11.6%. Ablation experiments further validate that meteorological data significantly reduces prediction errors during extreme weather events (MAE decrease of 15.5%), while social media data provides effective early warning information during sudden public incidents. This method effectively enhances the accuracy and robustness of STLF, providing reliable support for real-time grid dispatch and demand response.
Liu et al. (Sun,) studied this question.