A Two-Stage Multistep-Ahead Electricity Load ForecastingScheme Based on LightGBM and Attention-BiLSTM

Area of Science:

• Energy Systems
• Artificial Intelligence
• Time Series Analysis

Background:

• Accurate day-ahead electricity load forecasting is crucial for efficient smart grid operation.
• High time-resolution forecasts (15-30 min) are needed, but existing methods struggle with sudden load changes and multistep predictions.
• Conventional multistep-ahead models show performance degradation with longer prediction ranges.

Purpose of the Study:

• To develop a novel two-stage forecasting model to improve multistep-ahead electricity load prediction accuracy.
• To address the limitations of single-output and conventional multistep-ahead forecasting models.
• To enhance the smart grid's ability to handle sudden load fluctuations.

Main Methods:

• A two-stage approach combining a light gradient boosting machine (LGBM) for single-output prediction and a sequence-to-sequence attention-based bidirectional long short-term memory (S2S ATT-BiLSTM) for multistep prediction.
• Stage 1: Utilized LGBM with time-series cross-validation on recent load data.
• Stage 2: Employed S2S ATT-BiLSTM with an attention mechanism for multistep forecasting, using Stage 1 output as input.

Main Results:

• The proposed two-stage model demonstrated improved performance over a single S2S ATT-BiLSTM model.
• Achieved a 3.23% improvement in mean absolute percentage error (MAPE).
• Achieved a 4.92% improvement in normalized root mean square error (NRMSE).

Conclusions:

• The novel two-stage forecasting model effectively enhances multistep-ahead electricity load prediction accuracy.
• The model shows significant improvements in key error metrics, outperforming conventional approaches.
• This strategy offers a more robust solution for smart grid energy management, particularly in handling load variability.