A Short-Term Photovoltaic Power Prediction Method for Residential Energy Systems and Its Application in Scheduling Strategies

Abstract With the widespread adoption of distributed energy systems, accurately predicting photovoltaic (PV) power generation is essential for optimizing residential energy scheduling. However, existing methods face challenges in balancing prediction accuracy and cost-effectiveness. This study proposes a low-cost short-term PV power prediction method and applies it to residential energy systems (RES) to improve their scheduling performance. The prediction model uses a 3D convolutional neural network (3DCNN) to extract features from sequences of sky images, which are then integrated with historical power data through a DLinear-based multi-step prediction model. The proposed prediction model is further embedded into a multi-agent reinforcement learning (MARL) framework to assess the impact of short-term prediction on renewable energy scheduling. A case study is conducted on RES located in Changping District, Beijing, China. A physical model is developed, incorporating the proposed prediction model into the scheduling strategy to enable precise energy flow control and dynamic supply-demand balancing. Validation results demonstrate that the sequence model can achieve 60-minute-ahead prediction. In system-level simulations, when short-term prediction information is included as part of the state input for MARL, the long-term results show that the proposed approach reduces the operational cost of the RES and optimizes energy flow.