Load forecasting of distribution networks via modified kernel density estimation and learnable basis modalities

Load forecasting plays a critical role in power systems by enabling efficient resource allocation and power network planning. However, forecasting performance is limited by the resolution of weather recognition and data quality. To address this challenge, this paper proposes a novel load forecasting framework that integrates anomaly detection with multi-node load differences in the distribution network. Specifically, a modified kernel density estimation (MKE) method is developed to identify anomalies by introducing a weighted distance metric and kernel bandwidth update strategies. Additionally, a learnable basis modalities (LBM) approach is presented to automatically cluster and extract load profiles from various load nodes in distribution networks, with robustness to low-resolution weather inputs. To further enhance the ability to capture periodic and uncertain patterns, a bidirectional network combined with an attention mechanism is employed to strengthen the learning ability of load uncertainty. By integrating the MKE and LBM, a comprehensive load forecasting framework named the adaptive information fusion network is established. The framework is validated through extensive experiments on diverse real-world load datasets. Results demonstrated that the proposed method achieves a reduction ratio of approximately 16.12% in forecasting error and a 5.5% improvement in uncertainty estimation compared to recent approaches.