摘要： 针对居民社区电动汽车充电负荷中近零负荷样本占比高、局部峰值突变等问题，提出一种基于改进贝叶斯优化的TCN-BiGRU-MHA短期预测方法。首先，融合历史负荷、气象和时间日历信息，构建多源特征输入矩阵。其次，为改善贝叶斯优化算法在超参数搜索中的初始样本分布和目标函数拟合效果，并协调探索与利用，采用Circle混沌映射生成初始样本，交替使用高斯过程与随机森林代理模型，并在探索、利用和精细搜索阶段依次采用置信下界、期望改进和改进概率采集函数，联合优化模型结构与训练超参数。采用江西省某居民社区20台，额定功率为7 kW交流充电桩的4392个小时级样本进行验证，数据覆盖2024年7月2日至12月31日，近零负荷样本占49.52%。数据经异常值处理、缺失值填补和桩级功率聚合后，按时间顺序以7：2：1划分训练集、验证集和测试集。改进贝叶斯优化算法在10维超参数空间中最多进行50次评估，连续15次评估未刷新最优值时提前停止。实验结果表明，所提模型的平均绝对误差和均方根误差分别为0.2315 kW和1.2244 kW，决定系数为0.9843；与标准贝叶斯优化得到的同类模型相比，前两项误差分别降低72.8%和19.9%；与遗传算法和麻雀搜索算法优化的同类模型相比，均方根误差分别降低55.84%和44.93%，决定系数分别提高0.0646和0.0360。在相同数据划分及预测设置下，所提模型优于Transformer、Informer及两者经改进贝叶斯优化得到的模型。消融实验支持各模块在特征提取、时序建模与时间步加权方面的互补作用。SHAP分析显示，时间日历、历史负荷和气象环境特征的贡献度分别为45.96%、29.82%和24.22%，上一小时负荷、前一天及前一周同一时刻负荷分别反映短时相关性、日周期与周周期特征。在所研究社区稀疏且峰值显著的充电负荷数据上，所提方法降低了平均绝对误差和均方根误差；面向该场景设计的预测模型与改进贝叶斯超参数优化相结合，有助于提高预测精度，可为社区有序充电控制、配电网峰谷调节和短期运行调度提供预测依据。

Abstract: Residential-community electric vehicle (EV) charging loads have many near-zero observations and abrupt local peaks. To address these characteristics, this study proposes a short-term forecasting method combining improved Bayesian optimization with a TCN-BiGRU-MHA network. First, historical load, meteorological data, and temporal and calendar information are integrated into a multivariate feature matrix. Second, to improve the distribution of initial samples and the fit to the objective function in Bayesian hyperparameter optimization while balancing exploration and exploitation, a Circle chaotic map generates the initial samples. Gaussian process and random forest surrogate models are used alternately. Lower confidence bound, expected improvement, and probability of improvement acquisition functions guide the exploration, exploitation, and refinement stages, respectively. Network structure and training hyperparameters are optimized jointly. Evaluation uses 4,392 hourly samples from 20 AC chargers rated at 7 kW each in a residential community in Jiangxi Province. The data cover 2 July to 31 December 2024, and 49.52% of samples have near-zero loads. After outlier handling, missing-value imputation, and aggregation of charger-level power, the data are divided chronologically into training, validation, and test sets at a ratio of 7:2:1. The improved Bayesian optimization algorithm searches a 10-dimensional hyperparameter space for up to 50 evaluations. It stops early if the best objective value fails to improve in 15 consecutive evaluations. Forecasting performance is assessed by mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R2). The proposed model achieves an MAE of 0.2315 kW, an RMSE of 1.2244 kW, and an R2 of 0.9843. Compared with the same network optimized by standard Bayesian optimization, MAE and RMSE decrease by 72.8% and 19.9%, respectively. Compared with the same network optimized by a genetic algorithm or the sparrow search algorithm, RMSE decreases by 55.84% and 44.93%, respectively. The corresponding R² values increase by 0.0646 and 0.0360. Under the same data split and forecasting settings, the proposed model outperforms Transformer, Informer, and their counterparts optimized by improved Bayesian optimization. Ablation experiments support the complementary roles of the modules in feature extraction, temporal modeling, and time-step weighting. SHapley Additive exPlanations (SHAP) analysis assigns feature contributions of 45.96%, 29.82%, and 24.22% to time and calendar, historical-load, and meteorological and environmental variables, respectively. The previous hour's load and loads at the same hour one day and one week earlier reflect short-term dependence and daily and weekly cycles, respectively. In the studied community, the proposed method reduces MAE and RMSE for sparse charging loads with pronounced peaks. The network design and improved Bayesian hyperparameter optimization together improve prediction accuracy, providing a basis for coordinated charging control, distribution-network peak-valley regulation, and short-term operational scheduling.