MCrossFormer: multi-level cross-scale transformer for photovoltaic power and lifespan prediction

Accurate prediction of photovoltaic (PV) module lifespan and power output is essential for ensuring system reliability and economic viability. However, this task remains challenging due to two main factors: the complex coupling of degradation mechanisms under varying environmental stresses, and the multi-scale temporal characteristics inherent in PV power data. To address these challenges, this study proposes an integrated approach combining a weighted power degradation coupling model with a deep learning-based forecasting framework. The research first systematically analyzes how key environmental factors-such as temperature, humidity, ultraviolet radiation, and thermal cycling-individually and interactively affect module degradation. Building on this physical understanding, we develop a neural network model capable of capturing multi-scale temporal patterns in power generation data. Besides, we propose a novel Multi-level Cross-scale Transformer (MCrossFormer) architecture to overcome the limited generalization ability of traditional PV power prediction models. It adopts three parallel encoder-decoder structures to capture the trend, periodic, and closeness characteristics, respectively. Also, in each encoder-decoder module, we design a long short-distance attention mechanism, which consists of a Short Distance Attention (SDA) module, a Long Distance Attention (LDA) module, and a Multilayer Perceptron (MLP), to dynamically identify and capture critical patterns from PV power time series data. Extensive experiments on three public benchmarks show that the proposed MCrossFormer achieves significant and consistent improvements over state-of-the-art models, underscoring its effectiveness in practical forecasting scenarios.