多源融合驱动的光伏电站功率日前预测

Multi-Source Fusion Driven Day-Ahead PV Power Forecasting for PV Power Plants

摘要: 针对光伏发电间歇性与波动性导致的日前预测精度不足问题，本文提出“基准构建–融合优化–空间适配”三阶段光伏电站功率日前预测体系。该体系以某6600 kW级光伏电站实测数据与数值天气预报(NWP)数据为基础：首先构建单变量长短期记忆网络(LSTM)基准模型，精准捕捉功率时序依赖；随后通过CLARANS聚类算法划分强、中、弱辐射场景，结合Bi-LSTM融合多源气象数据，增强复杂气象适应性；最后采用Co-Kriging插值法实现NWP数据空间降尺度，解决空间尺度不匹配问题。研究表明，该三阶段协同体系预测性能显著优于单一模型，均方根误差(RMSE)降至0.0007，决定系数(R²)提升至0.958，较基准模型精度累计提升30%，复杂地形区域预测误差降低20%以上。本研究通过时序–空间跨维度优化策略，有效提升了光伏功率日前预测的精准度与稳定性，验证了“分阶优化、时空协同”技术路线的可行性，为电网新能源消纳与电力系统高质量发展提供了重要技术支撑。

Abstract: To address the issue of insufficient day-ahead forecasting accuracy caused by the intermittency and variability of photovoltaic power generation, this paper proposes a three-stage day-ahead power forecasting system for photovoltaic power plants: “Benchmark Construction-Fusion Optimisation-Spatial Adaptation”. This framework utilises measured data from a 6600 kW photovoltaic power station alongside Numerical Weather Prediction (NWP) data: first, a univariate Long Short-Term Memory (LSTM) benchmark model is constructed to precisely capture temporal dependencies in power output; subsequently, the CLARANS clustering algorithm classifies scenarios into strong, moderate, and weak irradiance conditions, while Bi-LSTM integrates multi-source meteorological data to enhance adaptation to complex weather patterns; finally, Co-Kriging interpolation is employed to spatially downscale NWP data, resolving spatial scale mismatch issues. Research demonstrates that this three-stage collaborative system significantly outperforms single models in prediction performance: root mean square error (RMSE) is reduced to 0.0007, coefficient of determination (R²) increases to 0.958, cumulative accuracy improves by 30% over the baseline model, and prediction errors in complex terrain areas decrease by over 20%. Through this spatiotemporal cross-dimensional optimisation strategy, the study effectively enhanced the accuracy and stability of photovoltaic power forecasting. It validated the feasibility of the “multi-stage optimisation and spatiotemporal coordination” technical approach, providing crucial technical support for grid integration of renewable energy and the high-quality development of power systems.