To address issues such as edge distortion， blurring， and loss of realism in radar echo extrapolation using deep learning methods， a residual module， generator， and discriminator were introduced into the Convolutional Long Short-Term Memory (Conv-LSTM) framework to construct a Generative Adversarial-Residual Convolutional Long Short-Term Memory Network (GAN-rcLSTM) deep learning model. Additionally， a customized weighted loss function， which assigns different weights to radar echoes of varying intensities， was designed and integrated into GAN-rcLSTM， resulting in the Weighted Loss Function-based Generative Adversarial Residual Convolutional Long Short-Term Memory Network (Wloss-GAN-rcLSTM) model. Using a historical radar echo dataset from Shandong Province and its surrounding areas (2021-2022)， the Wloss-GAN-rcLSTM model was trained and tested for radar echo extrapolation. A spatiotemporal deep learning radar echo extrapolation model capable of 0-2 hour forecasting with 6-minute updates was established. Evaluation results indicate that， at the Critical Success Index (CSI) threshold of 45 dBZ， which is of particular interest for heavy precipitation， the Wloss-GAN-rcLSTM model outperforms the optical flow method， Predictive Recurrent Neural Network (PredRNN)， and GAN-rcLSTM by 0.12， 0.07， and 0.02， respectively. For the clarity metric， structural similarity index (SSIM)， improvements of 0.009， 0.042， and 0.11 units were observed， respectively. Case studies further demonstrate that Wloss-GAN-rcLSTM is effectively suited for forecasting mesoscale weather processes， such as squall-line systems.