Photovoltaic (PV) system power supply is characteristically intermittent. Therefore, PV forecasting is crucial for decision makers responsible for electrical grid stability. With forecast models traditionally trained as macro-level solutions, where a single model emulates the entire PV system, there is uncertainty regarding the ability of these macro-level models to capture the low-level power output dynamics of large multi-megawatt PV systems. Instead, an aggregated inverter-level forecasting methodology is proposed to obtain an enhanced forecasting accuracy. These macro-level and inverter-level forecasting methodologies are implemented with state-of-the-art deep learning based Feedforward neural network, Long Short-Term Memory and Gated Recurrent Unit recurrent neural network models. Results are generated for a real-world scenario, with multi-step forecasts delivered 1–6 h ahead for a 75 MW rated PV system. To ensure the scalability of the proposed methodology, a unique inverter-clustering technique is presented, which reduces the effort of optimising multiple low-level forecast models. A heuristic process of systematic hyperparameter optimisation is also proposed, which serves to guide future forecasting practitioners towards unbiased model development. From the deterministic and probabilistic confidence interval evaluations, overall results demonstrate a marginal increase in forecasting accuracy from the proposed aggregated inverter-level forecasts. The best performing macro-level model obtained Mean Absolute Percentage Error (MAPE) values ranging between 1.42%–8.13% for all weather types and forecast horisons. In comparison, the equivalent inverter-level forecasts delivered MAPE values ranging from 1.27%–8.29%. Finally, it is concluded that deep learning based macro-level forecast models have a sufficient ability to capture low-level PV system behaviour.

光伏（PV）系统电源通常是间歇性的。因此，PV预测对于负责电网稳定性的决策者至关重要。传统上将预测模型训练为宏观解决方案，其中单个模型可以模拟整个光伏系统，因此这些宏观模型捕捉大型多兆瓦光伏系统的低层功率输出动态的能力存在不确定性。取而代之的是，提出了一种聚合逆变器级的预测方法，以获得更高的预测精度。这些宏级别和逆变器级别的预测方法是通过基于最新深度学习的前馈神经网络，长短期记忆和门控递归单元递归神经网络模型实现的。结果是针对实际场景生成的，额定功率为75兆瓦的光伏系统提前1-6小时进行了多步预测。为了确保所提出方法的可扩展性，提出了一种独特的逆变器群集技术，该技术减少了优化多个低层预测模型的工作量。还提出了系统超参数优化的启发式过程，该过程可指导未来的预测从业人员朝着公正的模型发展方向发展。从确定性和概率置信区间评估中，总体结果表明，与拟议的逆变器级汇总预测相比，预测精度略有提高。对于所有天气类型和预报地平线，性能最佳的宏观模型获得的平均绝对百分比误差（MAPE）值在1.42％–8.13％之间。相比下，等效逆变器级别的预测得出的MAPE值在1.27％–8.29％之间。最后，得出的结论是，基于深度学习的宏观水平预测模型具有足够的能力来捕获低层光伏系统的行为。