A short-term time-series prediction approach for photovoltaic power generation in residential energy systems

As distributed energy systems become increasingly prevalent, residential energy systems (RES) equipped with photovoltaics (PV) face significant challenges in maintaining supply-demand balance due to power output fluctuations. This necessitates short-term PV power prediction methods that effectively balance accuracy and deployment cost. To address this issue, this paper proposes a novel short-term PV power prediction approach based on low-cost ground-based sky image sequences: the 3DCNN-DLinear model. The method leverages fisheye camera-captured sky images to extract spatiotemporal features via a three-dimensional convolutional neural network (3DCNN), and integrates a lightweight time-series model, DLinear, to enable efficient prediction. The proposed model was evaluated using real-world data collected in Changping District, Beijing, China. A comparative analysis involving six mainstream time-series models confirmed that DLinear achieved the lowest overall prediction error. Further experiments demonstrated that the 3DCNN-DLinear model reduced RMSE by 49.28%, 9.56%, and 8.82% for 30-, 60-, and 90-minute prediction tasks, respectively, compared to the baseline 3DCNN-LSTM model. Additionally, the study examined the contribution of sky image data to prediction accuracy, revealing significant improvements under varying conditions. Notably, RMSE was reduced by 40.4% and 30.5% under sunny and cloudy conditions, respectively, for the 60-minute task. Overall, the proposed method offers an effective and economically viable solution to improve the predictive performance and intelligent scheduling of RES.