the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 10 Jan 2024
Combining Recurrent Neural Networks with Variational Mode Decomposition and Multifractals to Predict Rainfall Time Series
Abstract. Rainfall time series prediction is essential for monitoring urban hydrological systems, but it is challenging and complex due to the extreme variability of rainfall. A hybrid deep learning model (VMD-RNN) is used in order to improve prediction performance. In this study, variational mode decomposition (VMD) is first applied to decompose the original rainfall time series into several sub-sequences according to the frequency domain, where the number of decomposed sub-sequences is determined by power spectral density (PSD) analysis. To prevent the disclosure of forthcoming data, non-training time series are sequentially appended for generating the decomposed testing samples. Following that, different recurrent neural network (RNN) variant models are used to predict individual sub-sequences and the final prediction is reconstructed by summing the prediction results of sub-sequences. These RNN-variants are long short-term memory (LSTM), gated recurrent unit (GRU), bidirectional LSTM (BiLSTM) and bidirectional GRU (BiGRU), which are optimal for sequence prediction. In addition to three common evaluation criteria, mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE), the framework of universal multifractals (UM) is also introduced to assess the performance of predictions, which enables the extreme variability of predicted rainfall time series to be characterized. The study employs two rainfall time series with daily and hourly resolutions, respectively. The results indicate that the hybrid VMD-RNN model provides a reliable one-step-ahead prediction, with better performance in predicting high and low values than the pure LSTM model without decomposition.
- Preprint
(3354 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2023-2710', Anonymous Referee #1, 26 Feb 2024
The authors have used a hybrid deep-learning model to attempt to predict rainfall. While the paper is scientifically sound and quite in-depth, I can't see it as a good article for HESS because the focus is just so strongly on the Deep-Learning infrastructure. This is already evident in the Introduction which is clearly written with an audience in mind that is up-to-date with the terminology and typical issues that come with deep-learning models, whereas there is very little attention for the real-world practical problems this model is trying to solve (the abstract mentions urban runoff issues which are never mentioned anywhere in the main body, for instance). This would be fine for a journal that focuses on that particular research area, but the typical HESS reader (or at the very least, myself) will be completely lost in the methodology section. It's way too detailed in explaining the core mathematical concepts behind the model (once again, scientifcially absolutely good work, but not for a hydrology-focused journal), whereas the section discussing the used dataset for validation purposes (section 3.1) is barely 10 lines long and doesn't contain any information about the type of data collected (is it radar, tipping bucket, time-integrated, point measurements, etc etc).
That aside, the outcomes of the study are also a bit disappointing from a practical point of view. The authors acknowledge that their chosen study area has a fairly typical rain pattern, which makes me wonder whether this means such a model can't be applied anywhere else without specifically training it for that area - which would defeat the purpose of using a model, in my opinion. Secondly, and perhaps most importantly: the authors conclude that with the used lead time (1 time step) the applicability of the model is severely limited for prediction purposes, nor can it handle the stochastic nature of rainfall variability all too well. A conclusion on my end would be then that it's not any better than just interpolating observational data...
I can only recommend to reject this paper for HESS. While I can't judge the details of the used methodology (not being an expert in the field of Deep Learning Models myself), I do believe this is a sound scientific paper and definitely interesting, but just submitted to the wrong journal. A journal which focuses on developing models such as this would be a much better fit, where the validation and application of this model in a real setting is not as important.
Citation: https://doi.org/10.5194/egusphere-2023-2710-RC1 -
AC1: 'Reply on RC1', Hai Zhou, 11 Mar 2024
Thank you for your thoughtful analysis of our manuscript. While we understand the concerns that you have raised, we believe there are compelling reasons to reconsider the suitability of the article for HESS.
- We acknowledge your concern about the manuscript’s focus on deep-learning models. In fact, deep learning is increasingly relevant in hydrological research. By doing so, we aim to bridge the gap for readers who are not familiar with deep-learning models. In the revised manuscript, we will explicitly state the pedagogical purpose of the future work in the Introduction section. We believe that by framing the deep-learning approach as a powerful tool for addressing hydrological challenges, the paper becomes more accessible to broader readers.
- The worry about the model's applicability beyond the chosen study area is understandable because the model only has to be trained once. In principle, a new dataset from different regions or time periods can be fed directly into the well-trained model without repeating the training process to obtain the prediction on the new dataset. By the way, the used dataset in this study is sourced from The POWER Project. According to the project, the precipitation data is derived from a combination of observations, satellite data, and atmospheric model simulations. We will explore transferability and potential adaptations for different datasets in Section 4 of the result analysis.
- The confusion that our work can be considered as similar to interpolation is duly noted. We will also introduce the traditional interpolation method as one of benchmark methods in the Section 4, explaining how our methodology differs from traditional interpolation methods.
- We also understand the opinion about the lack of attention to real-world practical problems. Because the study presented in this manuscript serves as a kind of pedagogical example, acting as a starting point for further research that can extend to realistic forecast and space-time nowcasting. As we described in the future works, multi-step-ahead rainfall prediction is currently under investigation, and the model combined multifractals with deep learning is being developed to analyze the variability of stochastic rainfall time series.
In light of these clarifications, we believe that the pedagogical nature of our work can contribute significantly to the hydrology community by providing a deeper understanding of the application of deep-learning models and multifractals technique in rainfall prediction. We trust that these revisions will enhance the manuscript's suitability for HESS.
Citation: https://doi.org/10.5194/egusphere-2023-2710-AC1 -
RC2: 'Comment on egusphere-2023-2710', Anonymous Referee #2, 06 Apr 2024
- Lines 47 and 48: “However, these pure variant models are…preprocessing. ” Please consider citing some previous studies to support your statement.
- The second to last and third to last paragraphs in the Introduction section should be in the Method section. They went into details about either a model or an evaluation index, rather than focusing on the context and motivation of this study.
- I suggest that the authors elaborate on their motivation and clarify the contribution of this work. Based on the current introduction, the model is not new, and the dataset is not new. It’s okay if this work is focused on applying a method to a dataset and this application has not been documented in previous research. But you will need to justify your decision with appropriate citations. For instance, why is that application important? It could be because of limitations from previous approaches or the good performance of some new approaches, and so on. You just need to justify this work by elaborating why it is important.
- Maybe I missed something, but why do you need steps 3 and 4? I suggest that the authors explain why they want to generate sub-sequences on combined sequences with both training and non-training sequences and then clip to get the non-training ones instead of generating the non-training sub-sequences using directly the non-training original sequences.
- Sub-section 3.3 open sources. The title of this sub-section is weird to me. Maybe consider using titles like Model Settings and Implementation.
- Result analysis. Since for each testing sub-sequence several RNN models were used and only the best result was kept for result aggregation, it will be really helpful to add a summary table showing the result of each RNN model on each sub-sequence. This will not only allow readers to understand how the eventual result was aggregated but will also bring insights into which model is the best, and so on.
- I feel that the Result section is not very well elaborated. So far there are only results but no discussion, which damaged the value of this study. How will readers benefit from reading this paper? To me what’s more important is the insights behind specific results. For instance, why are some models better than others? In what circumstances? What insights can I gain regarding model selection and tuning after reading this work? etc. I suggest the authors add more in-depth discussions (please also refer to my 6th comment) to improve the quality of this section.
Citation: https://doi.org/10.5194/egusphere-2023-2710-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 230 | 77 | 21 | 328 | 12 | 12 |
- HTML: 230
- PDF: 77
- XML: 21
- Total: 328
- BibTeX: 12
- EndNote: 12
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1