Abstract: In weather forecasting, mesoscale convective precipitation nowcasting is both crucial and challenging. High-resolution spatiotemporal precipitation data from weather radars are essential for predicting weather within 0–2 hours. Traditional radar echo extrapolation methods face limitations due to the lack of nonlinear mapping ability and the influence of local environmental parameters on the weather system. This study utilizes a recurrent neural network—PredRNN-v2—which includes a memory decoupling function and spatiotemporal long-short term memory network, specifically for nowcasting local convective weather systems. The training and test datasets are constructed from radar volumetric scanning echo intensity data obtained from the Guangzhou radar from 2010 to 2014. The model is trained using the Huber loss function, which is selected for its fast convergence and robustness. Recognizing the significant role of strong echoes in weather evolution, the study gives greater weight to strong echoes during fitting. The performances of the PredRNN-v2 model with equal weight loss functions and that with different-weight Huber loss functions are demonstrated through test validations and case studies of convective precipitation. The results of test validations show that the PredRNN-v2 model with different-weight Huber loss functions achieves a higher critical success index, higher hit rate, and lower false alarm rate for long extrapolation durations. Further, case studies show that for this model with different-weight Huber loss functions, the peak signal-to-noise ratio and structural similarity are higher under long extrapolation durations, while strong echo bias remains closer to 1 across all extrapolation durations within 2 hours. The PredRNN-v2 model with different-weight Huber loss functions can better simulate the nonlinear evolution of convective echoes, providing more reasonable and accurate precipitation location forecasts.