Abstract
In recent years, the National Climate Center has developed dynamic downscaling prediction technology based on the Climate-Weather Research and Forecasting (CWRF) regional climate model for summer precipitation prediction. However, systematic deviations persist, complicating accurate forecasting. This study utilizes a combination of Dendrite Net (DD) and Artificial Neural Networks (ANN) to analyze and correct summer precipitation data from 1996 to 2019. By comparing CWRF simulations with quality-controlled ground station observations, a correction model was established. Results indicate that the dendritic neural network algorithm significantly outperforms historical CWRF returns. Specifically, the anomaly correlation coefficient (ACC) and temporal correlation coefficient (TCC) increased by 0.1, while the mean square error (MSE) decreased by approximately 26%. These findings demonstrate that machine learning algorithms can effectively correct regional climate model biases and enhance weather forecast accuracy.
Introduction
Predicting summer precipitation is a critical component of disaster prevention and water resource management. While advanced dynamic downscaling models like the Climate-Weather Research and Forecasting (CWRF) model provide a foundation for these forecasts, they often suffer from inherent systematic deviations. The complexity of atmospheric systems makes it challenging for traditional physical models to achieve the localized precision required for effective flood planning. This research addresses the "last mile" of forecasting by employing machine learning to bridge the gap between coarse CWRF simulations and actual ground-level observations.
Methodology: The Artificial Dendritic Neural Network
The core of this research is a novel hybrid architecture: the Artificial Dendritic Neural Network. While standard Artificial Neural Networks (ANNs) are powerful for non-linear mapping, they are often criticized as "black boxes" that lack logical transparency. To address this, the researchers integrated Dendrite Net (DD) modules into the hidden layers of the network.
Unlike standard neurons, DD modules are designed to identify logical relationships—such as AND, OR, and NOT—between inputs. This "white-box" characteristic allows the model to converge more reliably toward global optimums and enhances its ability to generalize across diverse geographical regions. Furthermore, the team employed a 3×3 grid approach to incorporate spatial context from neighboring data points and used Random Forest algorithms to select the most impactful meteorological features, including wind speed, 2-meter temperature, and sea-level pressure.
Data and Regionalization
The study utilized two primary datasets spanning the summer months (June–August) from 1991 to 2020 across China: CWRF historical returns at a 30 km resolution and quality-controlled actual precipitation measurements (OBS) from ground stations. To account for China's vast geographical and climatic diversity, the country was divided into eight distinct regions.
These regions include Northwest China (NWCH), the Tibetan Plateau (TP), and the Yangtze–Huaihe River Basin (YHRB), among others. Training the model on "anomalies" (deviations from long-term averages) rather than absolute rainfall values allowed the system to focus on predictive deviations while ignoring noise from local geography.
Key Results and Discussion
The hybrid DD-ANN model consistently outperformed raw CWRF historical forecasts across all evaluated metrics. Key improvements include:
- Error Reduction: The Mean Square Error (MSE) was reduced by approximately 26%.
- Correlation Gains: Both the Anomaly Correlation Coefficient (ACC) and the Temporal Correlation Coefficient (TCC) saw improvements of roughly 0.1.
- Regional Performance: The most significant improvements were observed in the eastern "rain belt" regions, such as South China (SCH) and the Yangtze–Huaihe River Basin (YHRB).
- Trend Prediction: Overall trend anomaly (Ps) test scores indicated a superior ability to predict whether a season would be wetter or drier than the historical average.
The integration of logical DD modules allows the network to explicitly model the conditional nature of weather patterns (e.g., specific combinations of humidity and wind direction leading to rain), which standard ANNs may overlook.
Conclusion and Future Directions
While the dendritic neural network approach shows great promise in correcting seasonal precipitation forecasts, challenges remain. The model still faces difficulties in fully capturing extreme, outlier precipitation events driven by small-scale, rapid atmospheric changes. Future research will explore "layered modeling" to specifically target extreme weather patterns and incorporate broader "Big Data" sources, such as real-time satellite imagery, to further refine forecast precision.
Summary updated