Data-driven approaches to rainfall nowcasting for application in hydrological modelling
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Flash floods are amongst the most complex and destructive phenomena. An abundance of process-based and data-driven models was proposed to serve as decision support tools for flood management authorities. While various observed hydrological and meteorological characteristics were usually used as an input for flash flood modelling, it was also found that integrating rainfall forecasts could considerably enhance the models' predictive ability. This study focuses on finding reliable and efficient data-driven rainfall nowcasting models (0-2h lead time). These models could then be integrated into a short-term flash flood prediction framework to investigate the framework performance including the effect of the precipitation nowcasts on the reliability of the modelling results. It is important to note that only data from rain gauges located on the same watershed are used to predict future precipitation. Rainfall data obtained from two rain gauges installed in the Spring Creek watershed, Ontario, Canada were used in this study. The investigated watershed is highly urbanized and prone to flash floods. Investigated data spanned four years from 2013 to 2016. We tackled this data-driven modelling problem from two perspectives: (1) an algorithmic and (2) a datacentric. From the algorithmic perspective, a comparative study of three data-driven models was performed. These models included the status quo persistence model, the statistical AutoRegressive Integrated Moving Average (ARIMA) model and the deep learning Long Short-Term Memory (LSTM) model. These models were applied to each time series to predict rainfall in the respective rain gauge location (univariate modelling). Following the data-centric approach, data from both sensors were combined into one dataset to predict rainfall in each sensor location (multivariate modelling). Lagged rainfall values from the sensor at the target location and the adjacent sensor were fed into an LSTM model to predict rainfall at the target location. Models were created for each investigated year for lead times ranging from 15 minutes to 60 minutes (corresponding to the time scale of the investigated rainfall events). Data for each year were chronologically split into training and testing with a 70%:30% split ratio. Root Mean Square Error (RMSE) and Maximum Residual Error (MRE) were used as evaluation metrics. Obtained results showed that overall, according to the estimated RMSE, LSTM demonstrated a better performance for all years except the year 2015. Figure Further analysis revealed that the year 2015 had major hydrological pattern difference between the training and testing sets. MRE did not indicate major variations between the years; it was found that all the models performed approximately at the same level as the persistence model. The models failed to predict extreme values accurately. The data-centric approach, however, showed different results. According to the RMSE and MRE metrics, LSTM models trained using data from both sensors demonstrated major improvement on data from years 2014 and 2015 for both target areas. Evaluation of the model performance on data from years 2013 and 2016 gave inconsistent results. Further investigation showed that the improvement in the model predictive ability coincided with the sensors' location and the dominating wind direction in the modeled years. In general, combining data from multiple sensors when used with the LSTM model showed promising results. Further extension of input variables including meteorological data collected on the investigated watershed will be the next step of the presented study.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it