Kai Xu

Kai Xu

Auto-generated profile, based on publicly available information

Kai Xu


Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
13 August 2024 in Applied Energy
Verified
Journal Article
Closed access
XXXuesong Xu
Xuesong Xu
KXKai Xu
Kai Xu
ZZZiyang Zeng
Ziyang Zeng
JTJiale Tang
Jiale Tang
YHYuanxing He
Yuanxing He
Tao ZhangTao Zhang
Tao Zhang

9 April 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
GSGuangze Shi
Guangze Shi
XXxuesong xu
xuesong xu
JTJisheng Tian
Jisheng Tian
KXKai Xu
Kai Xu
MGMeiyao Guo
Meiyao Guo
The growing integration of electric vehicles (EVs) has made accurate charging load forecasting essential for the safe operation of distribution grids and the planning of charging facilities. However, charging loads are highly nonlinear and non-stationary due to random user behaviour, weather disturbances, and calendar effects. Yet, existing methods typically fail to effectively decouple linear inertial trends from nonlinear residuals, and mostly employ single time-scale feature extraction, making it difficult to fully capture multi-scale temporal feature variations. Furthermore, existing similar-day aggregation strategies lack dynamic adaptability and adequate consideration of traffic flow factors. Therefore, we propose the Parallel Attention-based Multi-source Learning (PAML) framework. It uses a linear-nonlinear decoupled architecture: an autoregressive branch captures linear inertial trends, while a nonlinear branch models nonlinear fluctuations. Within the nonlinear branch, multi-scale gated causal convolutions and LSTM are integrated to capture multi-scale temporal features. The framework also integrates PCC and MIC to construct a multi-source similar-day screening mechanism. This mechanism extracts prior information from meteorological, traffic, and historical load data, adaptively weighting each source via attention. Experiments on real-world station data demonstrate that PAML reduces MAE by 5.9% and RMSE by 6.7% compared with standard LSTM. In training-free cross-station transfer tests, Correlation improves by 4.6%, confirming the framework's applicability to newly built stations with scarce historical data.