由于光伏发电在电力系统中的占比越来越高，准确有效的光伏发电功率预测对于提高电力系统的安全稳定运行具有重要意义。光伏发电功率点预测在晴空条件下具有较高的预测精度，但在非晴空条件下预测误差较大并且无法反映光伏出力的不确定性，难以满足当下电力系统对光伏发电功率预测的要求。鉴于此，本文对短期光伏发电功率概率预测进行研究，以满足电网经济运行和安全分析的迫切要求。

       由于天气类型对光伏发电功率预测有较大影响，本文采用Person相关系数法筛选聚类模型的输入特征量，利用手肘法改进的K-means聚类方法筛选光伏发电功率的相似日，避免了凭经验确定相似日聚类数所存在的不足，为提高非晴空条件下的短期光伏发电功率预测精度奠定基础。

       针对非晴空条件下短期光伏发电功率点预测精度不高的问题，本文利用相关向量机核函数不受Mercer定理限制的优势，将泛化能力强的多项式核函数和学习能力强的高斯核函数相结合，建立了基于混合核函数的改进相关向量机短期光伏发电功率预测模型，然后采用鲸鱼优化算法确定模型核参数，通过自适应调整不同天气类型下预测模型的最优核参数，实现对多尺度多模态变化的气象数据和光伏发电功率之间映射关系的建立，提高了非晴空条件下短期光伏发电功率点预测的精度。

       针对短期光伏发电功率点预测无法反映光伏发电功率随机性和波动性问题，本文在上述改进相关向量机点预测模型的基础上，提出基于分位数回归和高斯核密度估计的短期光伏发电功率概率预测方法。通过分位数回归获得在不同置信水平下的功率波动区间，在此基础上采用高斯核密度估计进行短期光伏发电功率概率密度分析，实现对光伏发电功率不确定性的预测。

       采用宁夏太阳山光伏电站的实测数据进行短期光伏发电功率预测，在不同季节的晴天、雨天和多云天气类型下对比本文预测方法与现有方法的预测结果，并采用平均绝对误差、均方根误差以及预测区间覆盖率等指标来分析本文预测方法的适用性，实验结果表明，本文方法可将预测误差范围控制在1.69~6.32%。

       Due to the proportion of photovoltaic power generation in the power system is getting higher and higher. Precise and efficient projections of solar power generation are crucial for the secure and dependable functioning of the electrical grid. Given the substantial variability and unpredictability inherent in photovoltaic yield, the extant models for forecasting photovoltaic power points exhibit commendable precision when skies are clear. Nevertheless, these models are prone to significant prediction discrepancies under cloudy conditions and fail to encapsulate the inherent uncertainty of photovoltaic generation, thus struggling to fulfill the stringent forecasting demands of contemporary power grids. In view of this, it is an important task to carry out short-term photovoltaic power probability prediction to meet the economic operation and safety analysis of power grid.

        Aiming at the problem that the weather type has a great influence on the law of photovoltaic power generation, this paper uses the Person correlation coefficient method to screen the input characteristics of the prediction model, and selects the similar days of photovoltaic power generation by the improved K-means clustering method based on the elbow method. Utilizing the elbow method, the enhanced K-means clustering technique circumvents the drawback of manually specifying the number of clusters. It refines the precision in selecting analogous days for short-term photovoltaic power, thereby providing a solid basis for enhancing the forecast accuracy of short-term photovoltaic power under cloudy conditions.

        Aiming at the issue of diminished forecast precision for short-term photovoltaic power under cloudy conditions, this study capitalizes on the flexibility that the kernel function of the relevance vector machine is not constrained by Mercer's condition. It amalgamates the polynomial kernel, noted for its robust generalization, with the Gaussian kernel, renowned for its strong learning capability, to devise a short-term photovoltaic power forecasting model grounded on a composite kernel function. Subsequently, the kernel parameters of the model are fine-tuned using the whale optimization algorithm, permitting an adaptive tuning of the optimal kernel parameters for the prediction model across diverse weather conditions. This methodology institutes a correlation between multi-scale, multimodal meteorological data and photovoltaic electricity generation. It adeptly captures the sporadic fluctuations in photovoltaic power output, thereby elevating the predictive accuracy of short-term photovoltaic power estimation.

         Aiming at the problem that the short-term photovoltaic power point prediction results cannot reflect the fluctuation of photovoltaic power generation, based on the above improved relevance vector machine point prediction model, this research builds upon the aforementioned enhanced relevance vector machine point forecasting model to introduce a probabilistic method for short-term photovoltaic power prediction that incorporates both relevance vector machines and quantile regression. The point prediction value of photovoltaic power generation at multiple quantiles at different times is obtained by quantile regression, and then the power fluctuation interval at different confidence levels is obtained. On this basis, Gaussian kernel density estimation is utilized to forecast the probability distribution of short-term photovoltaic power generation, thereby offering predictions regarding the uncertainty of photovoltaic power output.

        Conduct short-term photovoltaic power generation forecasting using measured data from Ningxia Taiyangshan Photovoltaic Power Station The prediction results of the proposed prediction method and the existing method are compared under the sunny, rainy and cloudy weather types in different seasons. The mean absolute error, root mean square error, prediction interval coverage and other error evaluation indexes are used to validate the applicability of the proposed prediction method in this paper, The experimental results show that the proposed method can control the prediction error range from 1.69 % to 6.32 %.