4.3.1. Comparative Experiment
This comparative experiment aimed to assess the effectiveness of the proposed jump connection strategy in improving temporal prediction models. We compared widely used models including ConvLSTM, TrajGRU, PredRNN, and PredRNN-v2, specifically integrating the jump connection strategy into PredRNN-v2, henceforth known as Jump Connection PredRNN Network (JC-PredNet). The experiment used the first 10 frames of an image sequence from the MovingMNIST dataset to predict the next 10 frames. We allocated 70% of the image sequences for training, 10% for validation, and 20% for testing.
We employed a range of evaluation metrics to thoroughly assess the predictive performance of different models, including MSE, LPIPS, SSIM, and PSNR.
Table 1 presents the comparative results across these metrics, focusing on average predictions for the latter ten frames of the sequence. In the table, a lower MSE and LPIPS value (↓) indicates that the predicted image more closely aligns with the actual image, reflecting higher accuracy. Conversely, a higher SSIM and PSNR value (↑) denotes greater fidelity to the original image, signifying better image quality and structural similarity. Our results showcased that JC-PredNet exhibited the most favorable performance across the board. Notably, JC-PredNet achieved an MSE of 48.7, which is 2.7 points lower than that of PredRNN-v2, suggesting a substantial enhancement in prediction accuracy. Similarly, JC-PredNet recorded the highest SSIM at 0.895, surpassing PredRNN-v2 by 0.005 points, and an LPIPS of 0.060, which is marginally better than PredRNN-v2’s 0.066. These improvements highlight the efficacy of the jump connection mechanism in enhancing spatio-temporal prediction accuracy, especially in handling complex dynamics within the sequences. These detailed comparisons provide clear evidence of JC-PredNet’s superior capability in modeling and predicting sequences with intricate motion patterns and varying intensities, positioning it as a significant advancement over existing models like ConvLSTM, TrajGRU, and PredRNN variants.
Figure 5 displays two randomly selected examples from the test set to demonstrate their long-term prediction performance. The image from the PredRNN model, enhanced with the jump connection strategy, appears clearer. This improvement suggests that enhancing long-term memory and spatiotemporal modeling allows the model to more accurately predict future frame changes. The new model, JC-PredNet, which includes the jump connection strategy, shows enhanced prediction quality. This is particularly noticeable in the clarity of long-term predictions and trajectory accuracy.
To investigate how different jump lengths affect model performance, we studied the impact of various jump strategies.
Figure 6 compares the effects of jump lengths of 2, 3, and 4. According to the figure, the optimal performance occurs with a jump length of 2.
This experiment sought to validate the effectiveness of the temporal correlation attention mechanism in boosting model performance. Building on JC-PredNet to better capture short-term abrupt changes, we analyzed the temporal correlation attention mechanism’s performance by comparing the loss curve convergence of PredRNN-v2, illustrated in
Figure 7.
Figure 7 shows that ETCJ-PredNet, which includes the time-correlated attention mechanism, converges more quickly and with a steeper gradient than the original model.
We performed quantitative and qualitative analyses of the network’s performance using the real radar echo HKO-7 dataset. ETCJ-PredNet was compared against the ConvLSTM, TrajGRU, PredRNN, and PredRNN-v2 models. In this experiment, we input the first 10 image frames to predict the subsequent 20 frames, using radar echo images from the past hour to forecast the next two hours. The dataset distribution was 70% for training, 10% for validation, and 20% for testing, using the HKO-7 dataset. We set the model’s learning rate to 0.0001. To thoroughly assess the model’s performance, we used two popular evaluation metrics: SSIM and LPIPS. These metrics provided diverse perspectives on the model’s performance.
Figure 8 displays the performance change curves for each model over time steps. The SSIM and LPIPS comparisons show that ETCJ-PredNet consistently outperforms the other three models in radar image sequence prediction, with its superiority increasing over time.
To evaluate the accuracy of short-term precipitation forecasts, we utilized four metrics: CSI, HSS, POD, and FAR. The models evaluated included our proposed ETCJ-PredNet alongside widely used models like ConvLSTM, TrajGRU, PredRNN, and PredRNN-v2. Expanded assessments are now included across varying rainfall intensity thresholds, revealing a decline in performance with increasing thresholds (dBZ ≥ 30, 40, and 50), which can be attributed to the reduction in strong echo zones in radar images.
We conducted a detailed comparative analysis for each model at these thresholds. Our analysis shows a general trend where performance metrics such as CSI and HSS decrease as the dBZ threshold increases, highlighting the inherent challenges in accurately predicting more intense rainfall events, which are more unpredictable and less frequent. At dBZ ≥ 30, ETCJ-PredNet shows superior performance, achieving the highest CSI of 0.725 and HSS of 0.699, indicating more accurate and reliable predictions of rainfall occurrences.
Further, at higher rainfall intensities, dBZ ≥ 40 and dBZ ≥ 50, ETCJ-PredNet consistently outperforms the comparison models. For instance, at dBZ ≥ 50, ETCJ-PredNet records a CSI of 0.331 and an HSS of 0.353, significantly exceeding the metrics of PredRNN-v2 and other models. This superior performance is particularly evident in moderate to heavy rainfall conditions where the model’s innovative architecture—incorporating a temporal correlation attention mechanism and jump connection strategy—proves most beneficial.
The extended results from
Table 2,
Table 3 and
Table 4 illustrate that ETCJ-PredNet not only improves overall predictive performance but also enhances accuracy in forecasting complex meteorological phenomena and precise short-term predictions of severe weather events. This comprehensive evaluation underscores the significant advantages of ETCJ-PredNet’s unique architecture, affirming its effectiveness and utility in real-time precipitation nowcasting. These findings highlight ETCJ-PredNet as a robust solution for meteorological applications, particularly valuable in scenarios requiring precise short-term predictions of severe weather events.
As shown in
Figure 9 and
Figure 10, we selected a set of predicted images from the test set. To better reflect changes in rainfall intensity, we used the radar echo color scale to display the prediction results and conducted a detailed qualitative comparison of the prediction performance between ETCJ-PredNet and other models, including ConvLSTM, TrajGRU, PredRNN, and PredRNN-V2.
Figure 9 focuses on predictions during the precipitation growth phase, while
Figure 10 illustrates the dynamic changes during the precipitation decay phase. “Input” refers to the 10 radar echo frames received by the models, while “Ground Truth” and “Prediction” correspond to the actual and predicted radar echo images for the next 20 frames. Each row from t = 12 to t = 30 shows the predicted results from different models compared to the actual radar images at various time steps.
These comparisons clearly show that as the prediction time (t) progresses, the predicted images from each model gradually lose detail and clarity, particularly in long-term prediction scenarios. For instance, ConvLSTM and TrajGRU perform adequately in short-term predictions, but their predictions become increasingly blurry over time, with significant loss of detail. This issue is particularly evident in areas of strong radar echoes, indicating that these models struggle to capture the fine structures of intense precipitation accurately.
In contrast, ETCJ-PredNet consistently maintains higher prediction accuracy across different time steps, with predicted results that are more closely aligned with the actual radar images. Especially in high-intensity precipitation areas, ETCJ-PredNet preserves more details, not only accurately predicting the contours of the radar echoes but also precisely capturing their motion trajectories. The advantage of ETCJ-PredNet becomes more apparent in the representative cases of “precipitation growth and decay”, where its prediction of strong echoes is particularly effective, with more clearly defined precipitation boundaries.
This performance advantage is largely attributed to the architectural design of ETCJ-PredNet. Its Temporal Correlation Attention Mechanism and Jump Connection Strategy effectively mitigate the accumulation of prediction errors over time that occurs in traditional models. As a result, ETCJ-PredNet demonstrates greater stability and reliability in capturing radar echo trends. Additionally, ETCJ-PredNet consistently maintains higher image clarity, providing richer details and textures compared to other models. In strong radar echo areas, in particular, ETCJ-PredNet’s predictions are more accurate, enabling better definition of precipitation boundaries.
Through this comparative analysis, it is clear that ETCJ-PredNet excels not only in short-term precipitation forecasting but also in handling complex precipitation process changes. Whether in the growth or decay phases of precipitation, ETCJ-PredNet showcases its superior predictive capabilities, making it a more reliable tool for real-time precipitation forecasting.
4.3.2. Ablation Experiment
To validate the necessity and effectiveness of each component within the ETCJ-PredNet model, we performed a thorough ablation study, specifically examining the effects of the jump connection strategy and the temporal correlation attention mechanism.
The PredRNN model effectively handles spatiotemporal sequence data but often struggles with gradient vanishing in deep network layers. To counter this, we introduced a jump connection strategy into the PredRNN framework, creating a new model variant: JC-PredRNN. This strategy allows for direct information transfer between layers, enhancing the model’s capacity to capture long-term dependencies and improving the stability and efficiency of deep network training.
Table 5 displays the ablation results for the jump connection strategy, evaluating performance metrics like CSI, HSS, POD, and FAR, with a precipitation intensity threshold of
. The notation “w” indicates the inclusion of the jump connection strategy, while “w/o” denotes its absence. Here, “↓” signifies that lower values indicate better predictions, while “↑” implies that higher values represent better accuracy. The results show that the JC-PredRNN model, which includes the jump connection strategy, consistently outperforms the model without it across all metrics. The jump connections effectively mitigate gradient vanishing issues in deep sequence processing, thereby enhancing predictive accuracy and demonstrating superior performance in handling complex spatiotemporal data.
In our study, we compared the standard PredRNN-v2 architecture with a version enhanced by the temporal correlation attention mechanism.
Table 6 presents the ablation study results, using the same precipitation intensity threshold of
. The notation “w” indicates the inclusion of the temporal correlation attention mechanism, while “w/o” denotes its absence. The comparison clearly shows that the PredRNN-v2 model with the temporal correlation attention mechanism significantly outperforms the standard configuration.
These enhancements enable ETCJ-PredNet to perform robustly in extreme precipitation forecasting tasks, underscoring the critical roles of the jump connection strategy and temporal correlation attention mechanism in advancing spatiotemporal predictive capabilities.