In the realm of meteorological forecasting, the integration of various nowcasting extrapolation methods is critical for enhancing accuracy and reliability. This paper introduces an innovative method for radar echo extrapolation called Rolling Fusion (RF), specifically designed to improve radar composite reflectivity (CREF) extrapolation. RF represents a novel synthesis of Optical Flow (OF) and Deep Learning (DL) methodologies, targeting the enhancement of nowcasting weather predictions. Central to the RF approach is RFNet, a sophisticated tool that employs a two-layer convolutional neural network. This network is optimised using Particle Swarm Optimisation (PSO), a computational methodology inspired by the social behaviour of birds and fish. PSO is particularly valuable in refining the network’s parameters to tackle the prevalent issue of CREF intensity imbalance, which can skew forecasting results. By optimising these parameters, RFNet ensures a balanced and accurate representation of various intensity levels, crucial for predicting severe weather conditions. The training process for RFNet is meticulously structured, utilising 10 steps of CREF data extrapolated from both OF and DL methods to anticipate the subsequent 10 steps. This dynamic approach not only enables high accuracy in nowcasting predictions but also enhances training efficiency by using the initially trained RFNet as a pre-trained model for further training cycles. This layered training process reduces computational demands, making the system both time-efficient and resource-efficient. Empirical results from this study reveal that RFNet effectively mitigates common drawbacks associated with deep learning predictions, specifically intensity attenuation and echo structure blurring. These enhancements allow RFNet to provide clearer and more accurate forecasts. Performance assessments across various intensity thresholds from 20 to 50 dBz demonstrate the method’s robustness. At lower thresholds, such as 20 and 30 dBz, RFNet and DL exhibit comparable performance, both of which surpass the capabilities of OF. In these scenarios, RFNet’s advanced integration of methodologies ensures superior forecasting precision. At a 40 dBz threshold, DL initially excels within the first 30 minutes of forecast duration. However, RFNet outperforms DL beyond this timeframe, highlighting its strength in extended forecasting scenarios. Notably, at the 50 dBz threshold, RFNet displays a significant performance advantage over both DL and OF, maintaining superior forecasting ability for up to 42 minutes. This capability is particularly valuable in predicting high-intensity weather events, where rapid changes necessitate agile and accurate forecasting models. Additionally, the research indicates a trend where RFNet’s extrapolation performance improves as CREF intensity surpasses 40 dBz. This improvement underscores the system’s adaptability and effectiveness in handling severe weather conditions, ultimately contributing to more reliable and actionable nowcasting weather forecasts.