An Interpretable Photovoltaic Power Forecasting Method Based on Temporal Embedding

Photovoltaic power generation is a key driver of sustainable development, making real-time operation and maintenance of power generation equipment increasingly critical. Accurate forecasting of future power output can enhance the economic efficiency of power plants, optimize power dispatch, ensure grid stability, and prevent both power shortages and wastage, while enabling timely adjustments and maintenance. However, many remote photovoltaic power plants struggle to optimize resource utilization due to insufficient or unavailable sensor, cloud, and satellite data, often relying solely on limited historical information. This challenge is further exacerbated in regions where human intervention is costly or impractical due to accessibility issues. Traditional models, while reliable, often function as “black box” systems, lacking interpretability, making it difficult to identify the root cause when predictions deviate from actual values. To address these challenges, this study proposes a low-cost, deployable model framework that relies only on historical data and calendar features to provide long-term trend forecasting for decision-making, offering a novel solution for effective predictions under resource constraints.