Cloud-edge Collaborative Forecasting for Industrial Photovoltaic-storage-charging Systems

With the growing integration of renewable energy into modern power grids, photovoltaic (PV) power has emerged as a pivotal element that is frequently coupled with energy-storage systems to enhance dispatch efficiency. However, the inherent variability in PV power generation and user load present significant challenges for accurate system scheduling. A key difficulty is that the PV output is often subject to artificial power curtailment, creating a non-linear relationship with solar irradiance, whereas charging loads exhibit distinct patterns between weekdays and weekends. To address this problem, a cloud-edge collaborative prediction system is proposed to implemented in a local energy management system (EMS). This system adopts a meta-learning-based multihead regression model that integrates XGBoost, support vector regression, random forest, CatBoost, and a multilayer perceptron. For PV power prediction, historical rolling lag features are introduced to capture the curtailment patterns. For load prediction, time-based features are designed to reflect the behavioral habits of electricity consumption over time. The system is successfully deployed at a charging station in the Haier Industrial Park. A practically validated solution is demonstrated for addressing uncertainties in integrated energy systems, along with the effectiveness of edge-deployed algorithms in optimizing storage strategies and reducing operational costs.