1. Introduction
Flood-related hazards are among the most devastating weather-driven natural disasters which affect the population in vulnerable areas and cause high economic losses throughout the world [
1,
2,
3]. A steadily rising world population alongside an increase in natural disasters highlights the importance of developing early-warning weather systems [
4,
5]. Such systems are not limited to providing only timely and reliable runoff forecasts to inform local communities about possible flooding, but can also be used by local authorities and businesses as proxies in water resource management and planning, water quality prediction, and economic loss mitigation. Therefore, the development of efficient runoff forecast communication strategies is of the same importance as robust runoff prediction methodology.
The vital importance of timely forecasts for many parties creates a high interest in the development of operational flood forecasting services worldwide. There are many established operational runoff forecasting systems in place regionally, nationally, and globally, which provide skillful forecasts up to two weeks in advance (e.g., see review papers [
6,
7,
8]). However, several reasons limit the efficiency and effectiveness of current systems. First, while recent advances in numerical weather prediction (NWP) allow us to get accurate weather forecasts up to a few days in advance [
9], there are still many associated uncertainties. Second, our understanding of runoff formation processes is incomplete; thus, it reflects in additional epistemic uncertainty, which is inevitable in rainfall-runoff models [
10]. Third, there is a lack of observational runoff data, which is a crucial element for successful rainfall-runoff model parameters calibration and regionalization [
11], especially for remote and uninhabited regions, or countries where observational data is not always readily available. However, despite the existing problems in data availability and natural limitations in water balance components prediction, current forecasting services are valuable to their target audience [
12].
Most operational runoff forecasting systems are based on a conventional technological stack; this stack includes a runoff formation model and short-range meteorological forecast data (deterministic or ensemble), which drives this model [
5,
7,
8]. Unfortunately, all or some of these components may be subject to restricted use, which limits the ability of a system to be reproduced or reimplemented in the regions not covered by existing services. Due to this, when developing new systems or improving existing services for runoff forecasting, more attention needs to be paid to ensuring that every system’s component remains open and freely accessible to the interested community. This ensures the reproduction of secure results can be guaranteed [
13,
14] and support for the steady development of the system by those in the community can be made [
15].
While the unanimous consensus is that effective communication of produced forecasts is essential [
16], there are still many open questions and challenges as to how to organize this efficiently and affordably [
17]. Most of the existing services share a standard way of disseminating forecasts, through interactive websites, which provide detailed information, e.g., simulated hydrographs, flood probabilities, weather forecasts, generalized outlooks [
5,
8]. As social networks have become an important daily source of information, the circulation of forecasts on networks such as Facebook and Twitter has gained popularity internationally [
18,
19]. Although the main focus is still to provide timely and reliable predictions of river runoff, particular attention should be paid to protecting users from misinformation and misinterpretation of forecasting data [
20].
While during the last few decades Russia has faced an increase in the frequency of flood-related hazards [
21], the circulation of operational runoff forecasting has remained limited. Forecasts are usually issued as news by the Central Administration of the Russian Federal Service for Hydrometeorology and Environmental Monitoring (Roshydromet) or its regional branches on official websites in textual form, mainly before the spring floods (in the middle of March) or during extreme events (e.g., for 2019 spring flood
http://www.meteorf.ru/press/releases/18773/, in Russian). Thus, runoff forecasts in Russia are scattered, mostly qualitative than quantitative, and not updated online. Some Russian cities that regularly suffer from spring floods spontaneously organize the monitoring of water levels in surrounding rivers using web-cameras and then develop websites to share ongoing water level data (e.g., for Tom and Kondoma rivers near Novokuznetsk
http://uznt42.ru/index.php?do=static&page=vsekamery, in Russian). However, these community-driven efforts are sporadic and provide only qualitative forecasts which are limited in their value to provide early flood warnings.
In this study, we developed Russia’s first open and operational online runoff forecasting system, OpenForecast (https://hydrogo.github.io/openforecast/). The present study aims to comprehensively evaluate and demonstrate the potential of OpenForecast, as well as documenting its limitations, as the operational service for short-term runoff forecasting. The primary aim of this study is to establish the interim service that would serve as a forerunner, providing a guideline for further development. The specific value of this study lies in demonstrating the utility of open data and software for making runoff forecasting workflow freely available and reproducible.
4. Conclusions
In the present paper, we describe our motivation and following production of OpenForecast—the first open-source operational runoff forecasting system in Russia—which was developed for and evaluated on two pilot basins in the European part of Russia. OpenForecast uses only open-source and freely available data and software as framework’s building blocks; thereby it follows principles of open and reproducible science, and moreover, has a potential to be implemented at a national scale, or even globally.
The initial evaluation results of OpenForecast on almost a year of continuous operational use are promising, showing good skill in flood timing prediction, i.e., date of the beginning of the flood period or date of the peak discharge. However, initial results also indicate the significant inconsistencies between simulated and observed flood volumes for both pilot basins. Therefore, at the present state of the OpenForecast development, produced runoff forecasts should not be used as a reliable reference. Revealing of the particular reasons which may cause reported errors in runoff prediction will be in the main focus of following studies. We believe that for the next update of OpenForecast operational setup, we must focus our attention on improving the underlying hydrological model as it showed limited robustness on the evaluation period. There are two most perspective directions for further model improvement. First, to re-calibrate the model against the data we obtained from CAHEM for the evaluation period. Second, to develop an error model for addressing discrepancies between simulated and observed runoff.
Additionally, we demonstrate the utility of OpenForecast for operational use in the CAHEM. Using OpenForecast runoff forecasts as a first-guess forecast, CAHEM introduces the parsimonious data assimilation technique which uses recently observed runoff data for a further forecast updating procedure. This data assimilation technique significantly improves runoff prediction efficiency, and additionally underlines and confirms the importance of observational data assimilation for getting reliable hydrological modeling results [
50,
52].
OpenForecast is running operationally and communicates runoff forecast for the next three days via the openly available website at
https://hydrogo.github.io/openforecast/. OpenForecast development and evaluation results proved the concept of the proposed computational framework, but also highlighted the necessary need for further improvements. Specifically, improvements of OpenForecast are currently underway in four directions:
Scientific development. As initial evaluation results indicated some problems in flood prediction efficiency, we want to understand the possible sources of OpenForecast errors better.
Software development and service maintenance. Under this direction, we plan to migrate from ERA-Interim to ERA-5 meteorological reanalysis data (both produced by ECMWF), from deterministic (ICON) to ensemble weather forecast product (ICON-EPS) of DWD, and from the one hydrological model (GR4J) to the family of GR models (e.g., GR5J, GR6J). We also want to document and then share OpenForecast code to engage the hydrological community in further development.
Communication of forecast. We want to promote OpenForecast to a broader audience: both specialist and non-specialist should be able to benefit from our service to make informed decisions.
OpenForecast expansion. As OpenForecast requires only historical runoff observations and watershed boundaries as input for initialization of operational forecasting routine, we want to expand our service for as many basins as satisfy these conditions disregarding their location.