Adaptive user clustering enhanced BiLSTM-attention for short-term load forecasting in smart distribution networks

Area of Science:

• Electrical Engineering
• Data Science
• Artificial Intelligence

Background:

• Multi-node load forecasting in distribution networks is complex due to user heterogeneity and computational limits.
• Existing methods struggle to balance accuracy and real-time performance.

Purpose of the Study:

• To develop an adaptive clustering enhanced forecasting framework for improved accuracy and efficiency.
• To address challenges in smart distribution network load prediction.

Main Methods:

• Proposed an adaptive clustering framework using DTW-K-Medoids for load pattern identification.
• Implemented a sliding window update mechanism and cluster-wise training strategy.
• Utilized a two-layer BiLSTM-Attention collaborative forecasting model.

Main Results:

• The proposed method significantly outperformed baseline approaches in prediction accuracy and computational efficiency.
• Ablation studies confirmed the contribution of each component.
• Scalability analysis demonstrated computational advantages for large-scale networks.

Conclusions:

• The framework offers a balance between accuracy and efficiency for day-ahead load forecasting.
• The cluster-wise training strategy meets real-time prediction demands in large-scale networks.
• Provides data support for grid optimal scheduling and demand-side management.