With the large-scale integration of new energy sources, the control complexity of the distribution network has substantially increased, thereby complicating the mechanisms and factors affecting power loss. Traditional power loss calculation methods have become inadequate. Addressing this challenge, we propose a novel power loss frequency-decomposition prediction model that integrates wavelet transform with BIGRU-LSTM, significantly enhancing the prediction accuracy and adaptability for power loss in distribution networks with a high proportion of new energy. Key temporal and non-temporal feature parameters were selected from electrical and environmental characteristics using grey relational analysis and the NARMA method to construct a multidimensional feature dataset. A predictive architecture combining wavelet transform with BIGRU-LSTM was developed, where historical power loss data were decomposed into frequency components to form a multidimensional input for the network prediction. The BIGRU-LSTM network facilitated frequency-decomposition prediction of the components using deterministic and probabilistic prediction channels. The results were then reconstructed using inverse wavelet transform to produce power loss prediction intervals. Simulation examples validated the effectiveness and suitability of the proposed prediction model.