MambaSolar-Forcaster: A trustworthy photovoltaic ultra-short-term power forecasting method based on normalized optimization and multi-step forecasting mechanism

Against the backdrop of the global energy structure shifting towards clean and low-carbon, the high proportion of photovoltaic grid integration poses challenges to the safe and stable operation of the power system. The intermittency and volatility of photovoltaic output make high-precision and trustworthy ultra-short-term power prediction a key link in ensuring grid safety. Although deep learning models have shown potential in this field, traditional models such as RNN and Transformer have limitations in long-range dependencies and computational efficiency, and the uncertainty of their prediction results may lead to security decision risks. Therefore, this article proposes a new photovoltaic ultra short term power prediction model called MambaSolar-Forcaster. The core innovation of this study lies in introducing the Mamba architecture, a new generation state space model, into this field for the first time. By utilizing its selective state space mechanism, it efficiently captures long-range dependencies in photovoltaic power sequences while maintaining linear computational complexity, and generates more trustworthy prediction intervals. In addition, we have designed a normalization and multi-step prediction mechanism for photovoltaic data characteristics, and constructed a trustworthy AI evaluation system that integrates point prediction and interval prediction. Then, experiments were conducted on the PVDAQ dataset, and the results showed that MambaSolar-Forcaster significantly outperformed the comparison model in terms of prediction accuracy and interval prediction reliability. Its more compact and reliable prediction interval can provide stronger support for safe scheduling decisions in the power grid. • For the first time, the Mamba architecture, a new generation state space model, has been introduced into the field of photovoltaic power prediction. This model achieves efficient capture of long sequence context while maintaining linear computational complexity through its selective state space mechanism, perfectly fitting the characteristics of long-range dependence of photovoltaic data. • A complete normalization and multi-step prediction mechanism has been designed for the special properties of photovoltaic power data, achieving significant innovation in data processing. • We have established a comprehensive evaluation system that includes point prediction accuracy and interval prediction trustworthy, providing more comprehensive credibility support for the practical application of prediction results.