Photovoltaic power generation has an important place in the energy sector. In order to accurately quantify the uncertainty and fluctuation range of PV(photovoltaic) power and to improve the comprehensiveness of interval forecasts, a probabilistic prediction method for PV power intervals based on feature mining with improved TCN-BiGRU was proposed. First, the maximum information coefficient and symbolic transfer entropy causal analysis were utilized to screen the meteorological features, remove redundant information, and construct global horizontal radiation trend features, seasonal features, and weather clustering features to provide more effective information. Subsequently, the TCN-BiGRU model was improved by combining the temporal pattern attention mechanism and quantile regression methods to construct a combined model for interval prediction. Finally, the probabilistic prediction results are generated using the KDE method of empirical bandwidth selection with scatter measure semi-polar optimization. The proposed method is analyzed by real PV plant data, which verifies the high reliability and applicability of the proposed method in PV power interval probability prediction.