Chinese AI Team Achieves Breakthrough in Photovoltaic Prediction and Data Compression Research

Recent advancements in time-series prediction have been making waves in diverse fields such as data compression and extreme weather forecasting. As applications become more varied and data complexity increases, current models face significant challenges in handling heterogeneous data, modeling long-term dependencies, and capturing extreme weather fluctuations. Addressing these issues, the Artificial Intelligence team at the Computer Network Information Center of the Chinese Academy of Sciences (CAS) has developed a series of innovative algorithms and models, which have been successfully deployed in real-world systems. One of the primary challenges in photovoltaic (PV) energy is the strong weather disturbances and rapid cloud movements that affect power generation. To tackle this, the CAS team devised MCloudNet, an ultrashort-term multilayer PV power prediction framework. This model leverages high, medium, and low cloud imagery to predict optical flow trajectories, thereby improving its sensitivity and response to high-frequency power changes. MCloudNet has been implemented in several PV power stations across Hebei and Yunnan provinces, enhancing both prediction accuracy and stability in microgrid scheduling. Another critical challenge lies in the efficient modeling of potential structures within raw byte streams, which traditional compression methods often struggle with. The CAS team introduced the SEP time-series compression framework, designed specifically for lossless prediction and compression tasks involving universal byte streams. By employing semantic-enhanced patch representations and an adaptive skip mechanism, SEP boosts its ability to model the underlying structure in binary data, enabling shared memory and concurrent multitasking. Experimental results show that SEP can improve compression rates by up to 12.8% and speed by 32.5%, making it highly adaptable for scenarios like scientific data archiving. Both papers detailing these innovations have recently been accepted for presentation at the 34th International Joint Conference on Artificial Intelligence (IJCAI 2025), a top-tier conference in the field. The research was supported by the National Key Research and Development Program of China, highlighting its significance and impact. The MCloudNet framework employs a multi-layer cloud image structure to model and predict cloud movement and its effect on PV power generation. Meanwhile, the SEP framework processes multiple data streams efficiently by using enhanced patch representations and adaptive mechanisms to capture underlying patterns in binary data. These developments underscore the CAS team's commitment to pushing the boundaries of artificial intelligence in practical applications, contributing significantly to the advancement of renewable energy and data management technologies.