计及机理机制的Stacking集成光伏发电预测

准确地光伏预测对电力调度、容量分析和机组组合至关重要。现有的数据驱动预测算法在计算速度和预测精度上有一定的提升，但未能考虑光伏发电的内在机理，存在泛化的风险。针对上述问题，提出了一种基于Stacking框架的机理模型和数据驱动结合的预测模型。其中，光伏发电机理模型将嵌入Stacking框架一层预测结构，构成基于长短期记忆神经网络(long short-term memory, LSTM)、极度梯度提升树(extreme gradient boosting, XGBoost)和机理模型的并行预测学习器。机理模型将光伏发电限制在一个合理的范围内，作为数据驱动模型的预测约束。所提出的模型能够从机理模型中提取有用的固有信息，并利用数据分析的能力提取历史数据中的非线性关系。基于安徽省某地区实际数据分析，所提模型相比传统数据驱动方法具有更高的精度。

Stacking integrated photovoltaic power forecasting considering mechanism

Accurate photovoltaic forecasting is crucial for power dispatching, capacity analysis and unit commitment. The existing data-driven prediction algorithm has a certain improvement in calculation speed and prediction accuracy, but fails to consider the internal mechanism of photovoltaic power generation. Besides, there is a risk of generalization. To solve the above problems, this paper proposes a mechanism model based on Stacking framework and a data-driven prediction model. In this paper, the photovoltaic generator model is embedded in the Stacking framework layer prediction structure to form a parallel predictive learner based on LSTM, XGBoost, and mechanism models. The mechanism model limits photovoltaic power generation to a reasonable range as a prediction constraint of the data-driven model. The intrinsic information is extracted by mechanism model and the analysis of nonlinear relationship in historical data is driven by data. Based on the actual data analysis of a certain area in Anhui Province, the proposed model has higher accuracy than the traditional data-driven method.