Multi-step ahead forecasting for electric power load using an ensemble model

Abstract

The multi-step prediction of electric power load is a crucial technology to promote power grid intelligence. Precise forecasting of short-term electric power will enhance the meticulous distribution management level of the grid and the dynamic balance between the supply and demand side of the power system. Conversely, a poor forecast of electric power load will lead to the unreasonable power supply to the grid. However, the electric power load is characterized by strong volatility and randomness, and it is a great challenge to precisely grasp the complex non-linear patterns hidden in power load sequences. For making the grid distribution more rational and efficient, a more precise electric power load prediction model needs to be explored. Thus, an ensemble model NeuralProphet-Lightgbm which combines the advanced NeuralProphet model and Lightgbm model is proposed in this paper. The experiments with different prediction horizons of 12, 24 and 48 h were conducted in this paper, and the root mean square error (RMSE), mean squared error (MSE), mean absolute percentage error (MAPE) and standard deviation (SD) are used as assessment metrics for prediction performance. In order to verify the effectiveness of the proposed model, the NeuralProphet-Quantile Regression Forest (QRF), Prophet, Autoregressive Integrated Moving Average model (ARIMA), Prophet-Lightgbm, NeuralProphet, Long Short Term Memory (LSTM) and LazyProphet models are used as benchmark models. The results of the proposed model significantly outperforms than the comparative models and is very robust as prediction horizons increasing. It is a good choice for efficiently forecasting electric power load.