An optimisation of power system load forecasting driven by deep learning model

Power system load forecasting is core to grid safety and economic operation.With diversified power consumption and high-proportion renewable energy grid connection, load sequences show strong nonlinearity, time-variability and volatility, making traditional statistical methods inadequate for high-precision forecasting.This paper proposes an LSTM model integrating attention mechanism and multi-strategy optimisation: it enhances key time-step feature weights via attention, fuses historical load, meteorological and calendar features, and adopts Bayesian optimisation + grid search hyperparameter tuning plus regularisation to suppress overfitting.Experiments on southern China regional grid data show the model outperforms ARIMA, SVM and other benchmarks in short/medium-term forecasting, with lower MAE/RMSE, stronger cross-seasonal adaptability and stability, providing a feasible path for high-precision forecasting in new power systems.