An Operational Streamflow Forecasting System for a Data-Scarce Catchment in Tanzania

Abstract

This paper reports the findings of the first initiative of developing a year-round streamflow forecasting system using the HBV hydrologic model in a data-scarce Ruvu catchment in Tanzania. Considering the importance of the Ruvu catchment as the main source of water to the fast-growing mega city of Dar es Salaam, the researchers in this study made the most of the available data and their joint previous application experience of the modelling framework for the purpose of setting up a reliable operational model. In addition, the researchers adopted a phased approach of developing the streamflow forecasting system, using HBV as a hydrological model, which resulted in a simplified model structure with minimized complexity. For instance, the snow routine was removed as it is not relevant to the study area, and a few parameters were reduced to improve model efficiency. As a measure to demonstrate model performance, in addition to the Nash–Sutcliffe Efficiency (NSE) parameter used for model calibration and verification, several other error functions and graphical displays were used. The model performance values, as measured by NSE for calibration and verification periods, are 0.85 and 0.82 for Ruvu Roadbridge (1H8A), and 0.80 and 0.82 for Kidunda (1H3), respectively, and all are classified as “Very Good”. In addition, the PBIAS of less than ±5% in calibration indicates excellent water balance simulation. Furthermore, the forecast’s performance in this study is evidenced by an annual forecast R² of 0.933, with operational meteorological forecasts improving to 0.962 with “perfect” precipitation; dry season performance with R² of 0.964, demonstrating high skill in baseflow-dominated periods; and the PBIAS for forecasts of 0.866, indicating a slight systematic under-forecasting correctable by a ~15% precipitation adjustment. Although the Ruvu catchment has been characterized by this study as a data-scarce catchment, the results of the operational hydrological forecasting system vary with season and quality of forecast meteorological data, and the model is already launched for operational use. As evidenced by these study findings, the journey from data scarcity to operational forecast provision in the Ruvu catchment demonstrates that the principal barriers are fundamentally institutional and capacity-related. The authors suggest that any future forecasting initiative should put much emphasis on both the understanding of the modelling framework to be used and adequate data collection and analysis, in a synergetic manner with all relevant agencies. And it is also recommended to be vigilant regarding changes in the catchment characteristics and model performance during its life cycle, as the performance of the developed model is only valid under the condition that it was calibrated and validated.

1. Introduction

Hydrological forecasting is a translation of single deterministic or an ensemble of short, intermediate, and long lead-time meteorological forecasts into estimates of hydrological variables of interest (e.g., streamflow, river stage, snowmelt, etc.) via hydrological forecasting models at the corresponding temporal scales [1,2,3,4].

Operational hydrological models are a set of models consisting of many interacting routines handling all from validation and correction of input data to finally calculating the runoff forecast or ensemble of such, on short or long term [1,2,3,5,6]. In an operational environment, it is typically driven by inputs from a wide range of sources, including meteorological observations and forecasts, river observations, and other data types [1,2].

Hydrological forecasting is of primary importance to better inform decision-making, for instance, on flood and drought response strategies, irrigators, reservoir operations, hydropower generation, water resources planning, and environmental flow management [1,2,3,7,8,9,10,11,12]. It should be noted that hydrological forecasts possess no “intrinsic value”. They acquire value through their ability to influence the decisions made by users of the forecasts [13,14].

Hydrological models can be broadly classified into three categories: empirical models (black box), conceptual models (grey box), and physically based models (white box) [15]. Conceptual- and empirical-driven models are usually used in lumped or semi-distributed forms [1,2]. The Hydrologiska Byrans Vattenbalansavdelning (HBV) is a conceptual hydrological model extensively used in operational hydrological forecasting and water balance studies [16,17]. As with many other types of models, hydrological forecasting models are subject to many sources of uncertainty, including uncertainties in the input data, model parameters, initial conditions, and boundary conditions [1,2,18]. The model performance should be evaluated during the model calibration, and in subsequent monitoring of the operational performance, and ideally in real-time operation. For operational use, appropriate performance measures also need to be adopted for forecast verification [1,2,17].

To date, hydrological forecasting initiatives face a variety of challenges induced by the increasing trend of extreme events, changing basin climate and hydrology, and demands of a unified and versatile hydrological forecasting system operating at local to continental scales [3,6,10,18,19,20,21,22,23,24]. These challenges include (i) making the most of available data as both data-rich and data-poor countries struggle with retrieving such information, as well as quality controlling, infilling, formatting, archiving, and redistributing these data; (ii) making accurate predictions using models (i.e., modelling and forecasting); (iii) turning hydrometeorological forecasts into effective warnings; and (iv) the administration of an operational service. Furthermore, as remarked by other researchers such as [20], it is not clear how we can disentangle and reduce model structural/parameter/input uncertainty in hydrological prediction.

Some of the obstacles encountered in hydrological services in the developing countries, such as Tanzania, include inadequate modelling capacity, declining hydrometeorological monitoring networks, and inadequate data management systems; as such, they are characterized as data-poor regions [24,25,26]. Unlike developed countries (rich-data regions), many developing nations (data-poor regions) do not have hydrological models of their own, largely because of fairly low technical capacity for developing and/or simulating the models and a lack of financial resources to compute and maintain the models. This means that modellers in developing countries strive to choose an appropriate hydrological model from a variety of models developed in other regions and customize it to local catchments and conditions [27,28,29,30,31,32]. It is noteworthy that this research is conducted in a basin and region with known unreliable and scarce data issues. In addition, it is a region with limited capacity in terms of operational hydrological modelling. Limited data availability and quality hinder the accuracy and reliability of research findings and water resources management in the basin [24,27]. And insufficient capacity and expertise in hydrological modelling make it difficult to develop and implement effective water management tools and strategies [27]. Among other approaches, these issues could be addressed by embarking on collaborating with local stakeholders that can help identify data sources, build capacity, and ensure research relevance; using alternative data sources that can help supplement limited data availability; capacity building and training and capacity-building programmes for local engineers and stakeholders that can help enhance their skills and expertise in hydrological modelling and water management; and adopting flexible and adaptable methods and approaches to accommodate the limitations and uncertainties of working in data-scarce environments. As one of Tanzania’s earliest flow forecasting studies, this paper addresses challenges (i) and (ii), as discussed above and as reported by other researchers such as [19]. Therefore, the purpose of this paper is to report the findings of the first initiative of developing an operational streamflow/hydrological forecasting system using an HBV hydrological model in data-scarce catchments in Tanzania with a view to addressing the modelling capacity gap and data inadequacy, and uncertainty issues. These issues are adequately discussed in the methodology section of this paper.

2. Materials and Methods

2.1. Study Area

The study area is the Ruvu catchment, administratively located within the Wami–Ruvu Basin in Tanzania. The Ruvu catchment can be subdivided into the following three sub-catchments, namely Ngerengere, Upper Ruvu, and Lower Ruvu. (Figure 1). The Ruvu catchment area is 17,843 km², which includes parts of Morogoro and Pwani Regions, and drains into the Indian Ocean [33]. The Ruvu River originates in the Uluguru Mountains and flows eastwards and lies between latitudes 6° 05′ and 7° 45′ south and longitudes 37° 15′ and 39° 00′ east [34,35].

2.2. Characteristics of the Study Area

2.2.1. Climate

The eastern slopes of the Uluguru Mountains have a mean annual rainfall of more than 2500 mm, while the western side of the mountains receives less [36]. The Ruvu catchment has a bimodal rainfall pattern. Average monthly minimum and maximum temperatures are almost the same throughout the catchment; the coldest month is August (about 18 °C) and the hottest month is February (about 32 °C). The annual average temperature is about 26 °C [34].

2.2.2. Topography

Except for the Uluguru Mountains in the extreme west, which have an altitude of above 2000 m above mean sea level, the catchment is mostly composed of low-lying areas along the Ruvu River valley and a slightly elevated hilly area with moderate undulation, which extends from west to east around Morogoro town. Isolated rolling hills are in the middle reach of the Ruvu River. The lowermost part of the Ruvu River is meandering heavily along the floodplain and the extreme eastern edge of the catchment, where the low-lying alluvial floodplain is about 5–10 km wide.

2.3. Surface Water Availability

The Ruvu River is the main source of water supply and livelihood of the Dar es Salaam Region, the megacity of Tanzania. Over 50% of Dar es Salaam residents rely on groundwater because of the unreliable supply of water from the Ruvu River. This is due to the vulnerability of the unregulated Ruvu River to adverse impacts of droughts and floods [34]. To address the water supply problems in Dar es Salaam, the government had embarked on structural and information investments, including drilling for deep groundwater in the Kimbiji and Mpiji aquifers, upgrading of upper and Lower Ruvu water supply intakes, construction of a dam at Kidunda for water supply, and developing disaster risk management plans. The Kidunda dam will result in changes in downstream flow from 7.073 m³/s as available discharge (average drought—water discharge) as reported in [33] to the minimum discharge, to be released downstream of the dam, of 24 m³/s.

2.4. Socio-Economic Activities

The Ruvu catchment is home to different socioeconomic activities such as agriculture, which is the principal livelihood for about 20% of the rural population living in the area, and 80% of the population lives in urban areas. Outside major urban areas, 75% of total household income in the catchment is gained from agriculture. Other rural livelihood activities include irrigation practices in areas where rainfall is inadequate, pastoralism, beekeeping, and fishing activities. Major cash crops grown in the catchment include sugarcane, sisal, and cotton, while surplus food crops include maize, rice, sweet potatoes, and beans [34]. The catchment is also the major water supply source for Morogoro Municipality through Morogoro Urban Water Supply Authority (MORUWASA), but also supplies other urban towns or cities, including Dar es Salaam, Kibaha, Mlandazi, Bagamoyo, and other small towns through the Dar es Salaam Water Supply Authority (DAWASA). The Upper and Lower Ruvu water supply intakes on the river are located downstream of the flow gauging station, 1H8A. Considering the near future water scarcity as a result of climate change and urbanization impacts, DAWASA is currently constructing a Kidunda dam and reservoir in the Upper Ruvu catchment. The dam is designed to stabilize the water supply, particularly for the Dar es Salaam and Coast Regions, mitigate floods and droughts, enhance fishing and agriculture, and hydropower generation with a capacity of 20 MW.

2.5. Data and Data Analysis

In this section of the paper, besides presenting the traditional methods for data and data analysis, further data processing has been conducted in order to reflect on issue number (i) of this paper, how to make the most of the available data in the data scarce catchments with regard to data availability, sourcing, retrieving, quality controlling, infilling, formatting, archiving, and redistribution of these data. The rainfall and streamflow data records at all stations in the hydroclimatic regular monitoring network were subjected to data quality checks. This entailed checking for errors and inconsistencies in the records. Thus, data-cleaning and gap-filling techniques were applied whenever necessary. The stations finally selected for calibration of the rainfall–runoff models should also be operational still, for use during the forecast preparation by the calibrated models. This includes both precipitation stations giving input data and streamflow stations used for model updating.

2.5.1. Precipitation Data

It should be noted that, for the purpose of this research work, the authors of this paper have keenly explored various scholarly literature on the use of reanalysis precipitation data in operational hydrological forecasting in Africa and found that it is a topic of significant debate and challenge [37,38,39]. Using raw reanalysis precipitation for operational hydrological forecasting in Africa is not preferred because its inherent biases, smoothness, poor extreme event capture, and latency introduce critical errors into time-sensitive forecasts. As a result, this study decided to make the most of the available ground monitored precipitation data, as further described below. According to available information, there are fourteen (14) precipitation gauges in the regular meteorological monitoring network within the Ruvu catchment (Figure 2 and

Table 1. Metadata for precipitation gauging stations in regular monitoring network.

SN	Station	Code	Geographic Location		Validity Period		Record Length	% Missing	Thiessen Area (%)
			Lat	Long	Start	End
1	Mlali	9637051	−6.966	37.536	1956.01	2023.12	23831	4.1	1
2	Morogoro Maji	9637052	−6.818	37.660	1956.01	2023.12	24191	0.0	11
3	Kwandewa Masa (Mongwe)	9637049	−6.970	37.580	1959.12	2023.12	22171	5.2	0
4	Morning Side Farm	9637046	−6.900	37.670	1966.01	2023.12	20864	1.5	2
5	Mondo	9637045	−6.950	37.630	1970.01	2023.12	19723	2.2	1
6	Hobwe	9637047	−6.980	37.570	1971.02	2023.12	19047	1.6	0
7	Ruhungo	9637048	−6.920	37.630	1971.01	2023.12	19292	0.3	0
8	Matombo Mission	9737006	−7.080	37.770	1971.01	2023.12	17277	10.8	12
9	Kibungo juu Sec. School	9737024	−7.073	37.688	1973.01	2023.12	11466	0.0	2
10	Nghesse (Utari Bridge)	9738009	−6.992	38.291	2005.03	2023.12	5353	22.9	34
11	Morogoro Roadd Bridge	9637057	−6.863	37.620	2009.01	2023.12	5478	0.0	1
12	Mindu Dam	9638029	−6.690	38.695	2013.01	2023.12	4017	0.0	12
13	Langali Sec. at Mgeta	9737044	−7.059	37.575	2013.01	2023.12	4017	1.2	4
14	Milengwelengwe Met.	9737029	−7.435	37.633	2013.02	2023.12	3982	0.9	21

). They include both manual and automatic climatic stations. The stations are spatially distributed within the catchment. However, most of the stations are located on both the windward and leeward sides of the Uluguru mountains in the Upper Ruvu catchment.

Further analysis suggests that, with the exception of three stations (SN 9, 11, and 12), every station has missing data (Table 1), with two stations (SN 8 and 10) having missing data above 10 percent as a quality threshold used for this study. Notwithstanding, the stations have varying record lengths ranging from 3982 to 24,191 daily records, with an average of 14,336, equivalent to 11, 66, and 39 years, respectively. In addition, spatial analysis indicates that most of the rainfall gauging stations have been monitored with nonconcurrent observation periods. Consequently, available rainfall data have different lengths, periods of observation, and quality that are related to the size of missing observations. Based on Table 1 above, about three-quarters (i.e., 11 out of 14) of the stations have long rainfall records spanning between 1956 and 2023, with missing data less than 10%.

2.5.2. Selecting Rainfall Gauging Stations

Because high-quality precipitation data are essential for hydrological forecasting, the study first needed to determine which gauging stations were suitable for use as forecasting stations. The assessment of the contribution of rainfall stations to streamflow was performed using two approaches: Thiessen polygon analysis (Table 1) and a data-mining approach (machine learning models) under the Cubist tool environment, as has been proposed by other researchers [10,40,41,42]. Thiessen polygon analysis was conducted in two scenarios, in all 14 stations and 5 selected stations, to measure the consistency of results. The results for the first scenario are presented in Table 1 in the last column. The results for the second scenario indicate that five stations, Utari (46%), Milengwelengwe (22%), Matombo (15%), Mindu dam (12%), and Mondo (5%) have a higher percentage of influence in that order. You will note that the absolute values of contributions, as bracketed, of respective stations vary with the number of rainfall stations used in the analysis. In addition, the results from the two scenarios seem to somewhat correlate for some stations, whereby the three stations of Nghesse at Utari Bridge, Milengwelengwe, and Matombo Mission, consistently, appear to have higher influence and contributions to the total amount of catchment rainfall. However, Mindu and Mondo stations have less influence and contributions. In addition, this analysis suggests that the three stations of Hobwe, Ruhungo, and Kwandewa Masa (Mongwe) do not have any influence.

Such results were expected as the authors are aware that Thiessen polygons are a simplification of reality [41]. They assume that rainfall is uniform within each polygon, require that the meteorological stations are relatively dense, and the topography of the interpolated areas is similar, which may not always be the case [43]. Additionally, the accuracy of the analysis depends on the density and distribution of rainfall stations.

The cubist model was applied to the Ruvu catchment at the Morogoro Road bridge gauging station (1H8A), with 15 attributes (1 flow and 14 rainfall gauging stations) to discover which rainfall stations might be responsible for influencing runoff delivery at the outlet of the Ruvu catchment. In addition, the form of the model used is instances and rules, and the extrapolation allowed is 5%. Based on this analysis, five (5) rainfall stations—Mondo, Mindu, Matombo, Ngesse at Utari bridge, and Milengwelengwe—are the key rainfall stations contributing to runoff generation in the catchment.

It should be noted further that some rainfall stations, such as Ruvu at Morogoro Bridge and Morogoro Water Department (Morogoro Maji), that were found to have a stronger influence in contributing to catchment rainfall based on Thiessen polygon analysis, are considered redundant variables and therefore avoided. Again, such results were expected as the cubist models are designed to discover complex relationships between variables and avoid parameter redundancy in constructing relationships [40,41,42]. It is noteworthy that the cubist model analysis compares rainfall and streamflow on the same day and does not take the delay in the runoff process into consideration, which may lead to unrealistic results. The final optimization of individual weight factors for each of the selected five precipitation stations was determined during the model calibration process, as explained in Section 2.6.4.

2.5.3. Precipitation Data Quality Analysis

A more detailed data quality analysis was performed on the five selected rainfall gauging stations—Mondo, Mindu, Matombo, Milengwelengwe, and Nghesse/Utari. Although there were several small gaps in the data series, most of these could be filled by correlating with results from other stations. However, some stations have longer missed periods, as illustrated in Figure 3. After considering the correlations between stations and the overall quality of the combined areal precipitation data, the following years were deemed acceptable for analysis: 2010/11 and 2012/13–2018/19. The year 2011/12 was excluded due to questionable high precipitation recorded at Matombo. This year requires further investigation and verification before the data can be used for modelling, and was filled in by correlation to other stations. One station, Milengwelengwe, had no data before 2013/14, and data were filled in by correlation for two full years, 2010/11 and 2012/13, to have a possibility to create areal precipitation for model verification.

2.5.4. Runoff Data

In the Ruvu River Basin, a total of fifteen (15) streamflow gauging stations have been in operation during the last 15 years (Hydrological Yearbook 2010–2019 River Flows Data). Of these, 6 are Primary stations and 9 Secondary stations (Figure 2). Four (4) stations are located along the main Ruvu River reach, and the other 11 are in tributaries. Out of the 15 stations, 5 were selected for closer investigation (

Table 2. Metadata for selected streamflow gauging stations in a regular monitoring network.

Station	Code	Geographic Location		Network Density [23] **	Rating Curve Validity Period		No. of Flow Gaugings	Reliability	Remarks
		Latitude	Longitude		Start	End
Ruvu at Kidunda	1H3	−7.264	38.246	5(S)	1993	2020	19	Reliable	Recently demolished by flood waters of 29th April, 2024. Revitalization is going on.
Ruvu at Kibungo	1H5	−7.024	37.809	1(S)	2013	2021	45	Unreliable	Recalibration recommended.
Ruvu at Morogoro Road Bridge	1H8A	−6.691	38.694	7(S)	2004	2020	79	Unreliable	Underestimated recent (2023–2024) low flows and peak discharges. Recalibration recommended
Mvuha at Ngagama	1HC2	−7.200	37.838	1(S)	2007	2021	16	Unreliable	Recalibration recommended
Mgeta at Dhutumi	1HB5	−7.410	37.778	1(S)					Not published

Note: ** The network densities for streamflow were assessed based on recommended minimum densities as stipulated in [23]. For the Ruvu River catchment, coastal, interior plains, and hilly/undulating regions, a minimum drainage area of 2750 km² was considered satisfactory (S), otherwise unsatisfactory (US).

) as potential forecasting stations for the study. These stations are strategically located across the study area.

The data used for the current rating curves span from 1993 to 2021, with the number of streamflow gaugings varying from 16 to 45. In addition, note that the Ruvu catchment regular streamflow monitoring network density is satisfactory (S).

2.5.5. Selecting Runoff Gauging Stations

Given the importance of high-quality streamflow data for hydrological forecasting, it was necessary at the outset of this study to determine how many streamflow gauging stations met the criteria to serve as forecasting stations. The reliability analysis of rating curves from these gauges was employed to evaluate the accuracy and dependability of their representation of the stage–discharge relationship. In this research, the reliability of discharge rating curves was assessed using both acceptance limits for discharge measurements and the Sign Test [24].

Of the five streamflow gauging stations investigated, two stations—Ruvu at Morogoro Road Bridge and Ruvu at Kidunda—were finally recommended for model implementation. This recommendation stands despite recent modifications to their cross-section geometries, provided that these stations are appropriately upgraded.

Due to unavoidable circumstances, the earmarked streamflow gauging stations required recalibration due to modification of the cross-section geometry, emanating from socio-economic activities. In the years 2021/22, a water diversion weir or a sill was constructed just downstream of the 1H8A gauging station. Such modification made the river geometry at this site more complex and the existing rating curve ill-defined, not representing floodplain floods conveyed through culverts installed under Morogoro Road embankments.

Therefore, the poor status of the monitoring stations necessitated the initiation of new spot gauging measurements and recalibration of the rating curves for two key flow gauging stations (1H8A) and (1H3). The problem at hand, on scientific grounds, is complex, considering also the fact that a rating curve is a nonlinear equation. In addition, the number of new spot flow measurements is inadequate on statistical grounds. Therefore, traditional approaches such as Ordinary Least Squares were considered inadequate, and their uses for recalibrating the rating curves at the respective streamflow gauging stations were avoided [24,44]. In this assignment, the flow gauges that required recalibration were therefore recalibrated using the spreadsheet Solver toolkit optimization procedures as recommended by [44]. The recalibration was performed as requested and made the station suitable for monitoring during the model testing period, except for some problems at 1H8A during the flood season April–May 2025, caused by the loss of the upper part of the staff gauge.

In addition, it should be noted that the Kidunda flow gauging station (IH3) was completely damaged by floods on 29th April 2024, rendering it non-operational. All the staff gauges, including their respective anchoring/mounting posts, have been displaced and removed by flood waters and fallen/uprooted big trees. The only remaining measuring structure is a submerged and silted non-operational water levels graphical recording stilling well. Therefore, at the moment, it is noteworthy that no water level recording takes place. The chosen stations need to be upgraded in terms of infrastructure revitalization and rating curve updating. The loss of the Kidunda (1H3) station has partly been compensated for by correlation with measurements at the upstream Kidunda dam site and downstream measurements at 1H8A, but has introduced increased uncertainty in the data series, and as a consequence, it was difficult to undertake a verification study for flow forecasts at this site. The reconstruction of the 1H3 station is strongly recommended and has high priority.

The status regarding streamflow data quality and availability is outlined in Figure 3. For these stations, good data were found for the hydrological years 2011/12 to 2017/18. Less accurate but acceptable data were found for the years 2010/11 and 2018/19. Both stations were operational at the start of the project, and sufficient data for model calibration and validation were found and used.

Figure 4 highlights three (3) key sub-areas and the locations of two (2) essential flow gauging stations used for model calibration and operation: Ruvu at Kidunda (red, 1H3), Ruvu at Morogoro Road Bridge (green, 1H8A), and the Ruvu sub-catchment downstream from 1H8A to the outlet near Bagamoyo.

The Dar es Salaam water intake is near station 1H8A at Morogoro Roadbridge. The Kidunda reservoir, located close upstream to gauging station 1H3, about 80 km upstream from 1H8A, will help regulate flow in the Ruvu and is expected to be operational by the end of 2026. By then, it is also anticipated that the gauging station 1H3 will be operational again.

Table 3. Some important runoff data for the 1H3 and 1H8A catchments.

Catchment	Area, km2	Average Flow 2013–2017, m3/s	Specific Runoff L/(s × km2)	% of Total Flow at 1H8A (%)
1H3 Kidunda	6665	47.96	7.2	94
1H8A Ruvu (local)	7696	3.30	0.4	6
1H8A Ruvu (total)	14,361	51.26	3.6	100

presents key data for the catchments at 1H3 and 1H8A. The meteorological and hydrological conditions differ between these two sub-catchments; most precipitation and runoff generation occurs in the mountainous 1H3 (Kidunda) sub-catchment, whereas the remainder of the 1H8A catchment is relatively dry and contributes only a small portion to the flow at the water intake located at Morogoro Roadbridge.

The table shows that 94% of the flow at 1H8A originates from the Upper Ruvu sub-catchment (Kidunda), which covers 6665 km²—less than half of the total study catchment area measured at the gauging station of 1H8A Ruvu at Morogoro Roadbridge (14,361 km²). Specific runoff in Kidunda is 7.2 l/s/km², compared to just 0.4 l/s/km² for the remaining 7696 km².

The share of flow coming from the Kidunda sub-catchment and from the remaining (local) sub-catchment varies from year to year, during the year, and even from one rainfall event to another. Since the flow measurements at both stations are both incomplete and prone to various types of measurement errors, the computation of flow in the local catchment as the difference between flows at 1H8A and 1H3 is not feasible due to error propagation from both measurements and frequently yields negative flows. It is therefore not possible to find flow data of acceptable quality and to establish and calibrate a separate hydrological model for this catchment.

2.5.6. Potential Evapotranspiration Data

Meteorological data such as temperature, wind speed, humidity, and solar radiation are often used for calculating potential evapotranspiration, a key input for the HBV hydrological model. For the Ruvu catchment, only mean daily temperature records from 1971 to 1989 at Morogoro Meteorological station (via WRBWBB database) were sufficiently available. Since the year 2018, near-real-time weather data have been accessible from TAHMO for five (5) new stations in the catchment—Milengwelengwe, Ngerengere Utali, Matombo Mission, Langali Sec., and Kibungo Juu—with four (4) of them currently operational (

Table 4. Metadata for TAHMO automatic weather stations in Ruvu catchment as of July 2025 (Sourced from WRBWBB database).

SN	Station	Code	Geographic Location		Elevation (Masl)	Validity Period		Record Length	% Missing
			Lat	Long		Start	End
1	Kibungo Juu	TA00591	−7.070	37.690	1012	2018.12	2023.02	1515	37.8
2	Langali Sec School	TA00592	−7.059	37.575	1099	2018.12	2025.07	2416	23.6
3	Ngerengere Utali	TA00594	−6.992	38.291	115	2018.12	2025.07	2406	0.12
4	Matombo primary school	TA00792	−7.053	37.765	275	2023.02	2025.07	889	20.6
5	Milengwelengwe secondary school	TA00793	−7.437	37.634	162	2023.02	2025.07	889	23.2

).

These stations are distributed across different climatic conditions and elevation zones. The recorded climatic parameters include precipitation, temperature, wind speed, humidity, and shortwave radiation. Missing data in the datasets vary from 0.12 to 37.8 percent, with an average of 21.1 percent. The weather stations are all situated near their corresponding precipitation gauge stations (Table 1). This dataset is, unfortunately, not considered sufficient or concurrent with the streamflow data available for model setup. Due to limited data, this study decided to calculate potential evapotranspiration (PET) using the Thorntwaite method [45,46] based on monthly average temperatures from the WRWBB database and applied with smoothing for day-to-day variation (Figure 5).

Operationally, the forecasting system is based on the near-real-time hydroclimatic data from five (5) selected rainfall gauging stations and two streamflow gauging stations. All these stations require some manual processes for data reading, transfer, and model input. The programme system is open for and prepared for more automatic data input, for example, from the 5 TAHMO stations in Table 4, when the data stream from automatic stations becomes stable and reliable enough to replace manually operated sources. This is already implemented for meteorological forecast data, where direct links to YR and TMA are established and used. More complete climate data from the TAHMO stations can also be used for the implementation of improved potential evapotranspiration calculation, to replace the existing data. This can be performed easily due to the modular structure of the program system, four (4) TAHMO weather stations, and streamflow data derived from recalibrated rating curves of 1H8A and 1H3 gauging stations. In addition, TMA climatic forecast data were used for hydrological forecasting.

2.5.7. Meteorological Forecasts Data

Meteorological forecasts data on rainfall and temperature were sourced from a joint collaborative venture between NRK and Norwegian Meteorological Institute (YR: https://www.yr.no/en) (accessed on 31 March 2023) and Tanzania Meteorological Authority (TMA: https://www.meteo.go.tz/) (accessed on 12 June 2024) during model building/piloting and operational phases, respectively. It should be noted that while the model was being built, researchers had no access to local forecast data from TMA, so it was resorted to using publicly available data resources, such as the YR. Weather forecast data retrieved online from the TMA database are in Network Common Data Form (NetCDF) file format. The NC file from TMA contains various geographical coordinates (latitude and longitude), as well as time and weather-related variables (temperature and precipitation). Please note that the forecast data are presented as time series with a time step of 180 min (3 h) for 10 days at each location. The temperature data are presented in degrees centigrade and rainfall in millimetres.

The process involved a cross-matching of the five (5) selected rainfall monitoring gauging stations with the locations embedded within the NC file. The coded NC data underwent a spatial mapping procedure to pinpoint and align with the coordinates of the monitoring stations. Subsequently, the corresponding weather station values associated with these stations were extracted. The correct formatting was applied to ensure the data is presented clearly. It is noteworthy that data from YR are read directly and do not require reformatting.

2.5.8. Final Data Selection for the Calibration and Verification Periods

To set up, calibrate the model, and then verify the goodness of calibration, it is necessary to have some concurrent years with good data (precipitation, temperature, and streamflow) for the calibration, and some other years with equally good data for verification. The verification must be performed using independent data (years), not those used in the calibration. For the HBV-model (and probably also other similar models), it has been found that up to 5 years of data is enough for calibration, and similar for verification.

This strategy may be problematic if the total number of years with good data available is limited, which is also the case here. If we have only, for example, 7 years of data, the question is, should the number of years used for calibration be reduced to have enough years for verification? Recently, modelling experts, for example, [47,48] have argued that in such cases, the most important thing is to prioritize data for calibration and take what is possible to have for verification.

In this research work, this is clearly an important question. Both precipitation and streamflow data are limited, and it is a problem to find 5 + 5 years with high concurrent quality data. Figure 3 illustrates the problem.

The initial data analyses helped to identify 5 precipitation stations for use in the calibration and operation of the HBV models. These stations are Mondo, Mindu, Matombo, Milengwelengwe, and Nghesse/Utari. The individual weight of each station in the computation of mean (Areal) precipitation in each catchment was initially guided by the Thiessen polygon and Cubist model analyses and finally fine-tuned during the model calibration process.

The conclusion from the data availability analysis was that the most promising 5-year period with both good precipitation data and good streamflow data for stations 1H8A and 1H3 is the hydrological years from 2013/14 to 2017/18. It is not possible to find another 5-year period to be used for hydrological model verification, but following the advice of [47], we decided to use three (3) remaining good hydrological years for verification (2010/11, 2012/13, and 2018/19).

2.6. The Operational Forecasting Model System—Components and Structure

In this section of the paper, besides presenting the traditional methods for hydrological forecasting modelling, further analysis has been conducted to reflect on issue number two (ii) of this paper: how to make accurate predictions using models (i.e., modelling and forecasting) in data-scarce catchments and developing countries.

2.6.1. The Ruvu–HBV Flow Forecasting System—Model Structure

As deduced based on data availability and quality analysis above, the Ruvu River flow forecasting system is designed to produce streamflow forecasts for two locations on the Ruvu River using an HBV-type rainfall–runoff model. The HBV models simulate the natural flow from the catchments at gauging stations 1H3 Kidunda and 1H8A Ruvu at Morogoro road bridge.

Since the modelling system is intended to assist in the operation of the water supply system, priority has been given to establishing a forecasting service for the inflow to the intake at 1H8A. This can be performed by establishing a hydrological model for the entire catchment upstream of 1H8A, with a total area of 14,361 km² (estimated by this study). This model setup is further called a 1-stage system (Figure 6).

In addition, a forecast service is needed for inflow to the Kidunda reservoir and its use when the reservoir becomes operational in 2026. Inflow forecasts will be needed to obtain the optimal operation of the reservoir. The catchment below 1H8A to the outlet is not included in the modelling system. The introduction of a reservoir at Kidunda leads to a heavily modified flow regime in the Ruvu River downstream of the dam, between the gauging stations 1H3 and 1H8A. The model that was established for the unregulated flow at 1H8A will not be valid anymore; a more complex model setup is needed, later called the 3-stage model system (Figure 7). It consists of the following 3 models in sequence: (i) a rainfall–runoff model for the 1H3 sub-catchment; (ii) a reservoir operation model for the Kidunda reservoir; and (iii) a river routing model for the flow along Ruvu from Kidunda to Morogoro Roadbridge, including local inflow from tributaries.

In the 3-stage model, the first stage is to compute a forecast for the 1H3 Kidunda sub-catchment, giving the inflow to the reservoir. The inflow may be stored or (partly) released, depending on the operation decided for the reservoir. The reservoir operation model is the second stage; it is designed to respond to a complex set of rules for the reservoir operation, where the action of storing or releasing water depends on inflow, water demand, and state (water level) in the reservoir. Seepage and evaporation losses are also considered. The water released from Kidunda dam will flow downstream along the Ruvu River and arrive at the intake site at Morogoro road bridge a few days later. The straight-line distance is only about 80 km, but Ruvu is meandering heavily along the floodplain, increasing the flow length and delaying the flow up to several days, typically 3–5 days. The flow released from the dam is also gradually mixed with local inflow from tributaries and from the floodplain itself. The river routing model is stage 3 in the model setup. In principle, we use a modified Muskingum model where the lateral inflow is accounted for, assuming similar flow timing as in the Kidunda forecast model and a fixed share of the inflow to Kidunda.

The 3-stage system model will only be needed after the Kidunda reservoir becomes operational, but it has already been developed and tested to be ready for use when needed. Until then, it will be sufficient to use the 1-stage model for simulating flow and generating forecasts in Ruvu at 1H8A. Both model versions have been implemented and can be operated from a common user interface, with seamless integration of data import and storage.

2.6.2. The Operational Model System—Data Collection, Transfer, and Storage

An overview of the operational Ruvu–HBV forecasting system is shown in Figure 8. The modelling system includes components for input and storage of hydrometeorological data, rainfall–runoff models for simulating and forecasting river flow, a reservoir model for the operation of Kidunda reservoir, and a routing model for river flow between Kidunda reservoir and the Morogoro Road bridge. It also includes several report generators for preparing and displaying results. All user inputs and interactions can be operated from a common user interface, which is described later.

In operational use, a hydrological forecasting system is usually driven by inputs from many sources, including meteorological observations and forecasts and streamflow observations. Implementation of the conceptual framework in Figure 8 entailed customizing the existing Excel–HBV model, calibrating, testing, and verifying the hydrological modelling system (HBV) using data from the selected flow and precipitation gauging stations in the catchment, also establishing input routines for observations linked through GSM technology (Figure 9).

2.6.3. Rainfall–Runoff Model—The HBV Model

The streamflow forecasting system for the Ruvu catchment is based on a conceptual hydrological model derived from the HBV model approach. The HBV model (Figure 10) is a widely used hydrological model and has proved to work well in climate zones ranging from tropical to arctic [49,50,51,52]. The version implemented in Ruvu is built in an Excel environment and is often referred to as Excel–HBV.

Why use the Excel–HBV model?

The HBV model, conceived over 50 years ago, has been implemented in various programming environments, including FORTRAN, C++, Python, R, Pascal, Visual Basic, Matlab, and Excel. Its simple structure, low data requirements, and strong performance across diverse catchments have made it popular for teaching, research, consulting, and operational applications.

The Excel–HBV model has its roots 35 years ago as a spreadsheet model, developed for teaching and research at NTNU [50]. This version was developed in the first successful spreadsheet programme, Lotus 123 [53], and a few years later converted to Excel. This was possibly the first HBV model with a fully interactive graphical user interface, like that introduced by HBV-Light some years later [54]. Excel–HBV soon also became popular for operational use in flow/flood forecasting, mainly by hydropower and water supply utilities, and for consulting. Excel–HBV has been under constant development and improvement since the first implementation for operational use for flood forecasting in the Ruo River in Malawi in 1992 [52]. The Ruvu–HBV model is the newest version, building on a similar system in use at the Oslo Water Supply Utility (Oslo H2O).

A hydrological model, designed for operational flow forecasting in a real-world environment, usually requires tight integration with other software systems in the organization, for data collection and transfer, data storage, report generators, etc. Forecasting models are typically used by different individuals, not just modelling experts, so the interface needs to be intuitive, user-friendly, forgiving of mistakes, and resilient when faced with missing data or errors. Repeated, “boring” tasks, like manual data entry and model updating, should be automated as much as possible, for example, by downloading forecasts and observations by Web-based applications.

The Ruvu–HBV model system consists of two rainfall–runoff models, one reservoir operation model, and a routing model, all sharing the same input data but also exchanging data from one process to the next. Excel is a widely used and very powerful platform for data analysis, visualization, and automation, and is well suited for integrating such a complex software system, with a user environment known to most technical personnel. To our knowledge, no other free, readily available HBV-model-based software existed for operational forecasting.

Large modelling systems based on the HBV-model are developed and operated by national forecast centres in many countries, as Norway (NVE), Sweden (SMHI), Slovenia, Finland, and Germany, but have not been an option for this project, considering available resources. Some of these systems could possibly be considered later if a national flow forecasting service in Tanzania is to be established. In that case, the Ruvu model will be a critical part as the main component of the national system, and hence the capacity built on various modelling processes will be directly applicable at no extra cost.

Another model option briefly considered was the very popular “HBV-Light” model [54]. This programme is known to be an excellent choice for teaching and research on hydrological modelling, but was not considered easily configured for operational use in the Ruvu system, as the integration into the total system would require access to source code and an unknown amount of programming for automated data input and storage and interaction with other forecasting system modules such as the reservoir operation model.

The complete source code for Excel–HBV has been donated to the Dar es Salaam Institute of Technology (DIT), a leading institution in this study, which has also been given the full user rights for the software, including any new implementations in Tanzania or elsewhere in Africa.

Beyond the issues that led to selecting the HBV model, other factors influencing the overall modelling approach included required expertise, available software, real-time and calibration data availability, budget constraints, and prior experience with these models [2].

HBV-model applications in Tanzania and East Africa.

The first application of the HBV model in Sub-Saharan Africa may have been an HBV model used for flood forecasting in Malawi, according to a report where Bergstrøm summarizes known applications of HBV models from 1972 up to 1992 [52]. This model was developed for the “Flood forecasting and Warning system for Lower Shire Valley in Malawi” in cooperation between the Norwegian company OCEANOR and NTNU in 1990/92. In the following years, many applications for the Excel–HBV model in East Africa can be found, many connected to two important educational programmes.

Supported by the Norwegian Program for Development, Research and Education (NUFU), the “Pangani Project” was a collaborative research initiative in water management between NTNU and the University of Dar es Salaam (UDSM), active from 1996 to 2006. Throughout this period, the HBV model was incorporated into teaching and postgraduate research, predominantly via the Excel–HBV version, serving as a resource for MSc and PhD programmes.

Another reason for Excel–HBV-model use in Africa has been its role in introducing hydrological models in NTNU’s International MSc programme in Hydropower Development. Since 1994, nearly 450 students—mainly from developing countries, including 127 from East Africa and 17 from Tanzania—have enrolled. The model has also supported many MSc and PhD projects.

Together, more than 20 MSc theses from NTNU and UDSM included the use of the HBV model as a tool for hydrological modelling, for example, in Tanzania (Pangani River), Malawi (Shire, Ruo, and Songwe rivers), Uganda (Mabuku River), Zambia (Zambezi and Kafue rivers), and Ethiopia (Awash, Gilgel-Ghibe, Omo-Ghibe, and Koga rivers).

Unfortunately, most of the HBV-related results from the MSc theses were not published in scientific journals, so it is not possible to give references to open public sources, but access is possible by direct application to NTNU.

A few publications about climate change impact on hydropower in Africa are also worth mentioning: in the Zambezi River, including Kafue and Shire [55]; in the Kwanza River in Angola [56]; and in various rivers in Tanzania and Uganda [28]. All of them used the HBV model for analysis of the change in river flow and thereby the change in hydropower potential.

In recent years, several papers have reported on the use of the “HBV-Light” model [54] in the Wami–Ruvu Basin, assessing its performance and parameter sensitivity [32], comparing it to HEC-HMS and ANN models [30], and studying the impacts of climate change [31]. Although these studies used older data, mainly from the 1970s, and other streamflow stations than this study, calibration results are comparable and discussed in Section 4.1. The findings generally were that the HBV model performed well across all three sub-catchments and could be recommended for further use. The HBV-Light model was also used in a study of various Global Precipitation Datasets (GPDs), which included 3 reanalysis data products, in the Kilombero Valley in central Tanzania [39].

Experience accumulated from the use in a variety of catchments throughout the region has led us to conclude that the HBV model can also be effectively adapted as a rainfall–runoff forecasting model in the Ruvu basin.

Given the widespread use and extensive documentation of the HBV model, only a concise overview is provided here.

HBV Model—Structure and Parameters.

The original HBV model included four (4) main storage types: Snow, soil moisture, upper zone, and lower zone. The snow storage is not needed and excluded in this implementation; precipitation enters directly into the soil moisture and passes through the three storages before exiting as runoff or evaporation (Figure 10).

The model parameters are often classified as confined parameters and free parameters. The main confined parameters are AREA, LA, and the area elevation curve (Figure 10). These parameters are determined from maps, GIS, or field surveys. The free parameters, FC, β, and LP in the soil moisture routine—KUZ, KUZ1, and UZ1 in the upper zone, and KLZ in the lower zone—are all determined by the process of model calibration. The snow routine is not needed for most African catchments, and since the programme code for snow is quite extensive, it is omitted in the Ruvu model.

In addition, as a result of implementing a conceptual framework for developing a streamflow forecasting system, as reported in [2], modelling exercise in this project took place in three (3) phases for the purposes of finding an intermediate complexity that would facilitate parameter calibration and improve the relevance of hydrologic model as attempted by others [21,23]: Phase 1—data analysis and suitability assessment of the HBV as a rainfall–runoff model in the study area; Phase 2—customizing and testing HBV as a streamflow forecasting system using meteorological forecast data sourced from Yr.no; and Phase 3—operational forecasting model integrated online with near real time hydrometeorological data and meteorological forecasts from TMA.

2.6.4. Hydrological Model Calibration and Validation

The HBV model used in this study is coded in Excel using both ordinary in-cell formulas and Macro programme code for speeding up complicated calculations. The model calibration process, however, was performed using a separate model setup, with an “optimalization core” coded in C++ integrated within an Excel interface. This combines maximum speed during optimization with easy user interaction for parameter setup and visualization of model performance. The integration with Excel gives the modeller an opportunity to study model performance in detail, for example, under special extreme climate events or episodes where model performance is poor.

The parameter optimization is guided by an error function, which measures the goodness of fit between observed and simulated flow. The error function used in the optimization was the Nash–Sutcliffe (NSE) parameter. In addition, several other goodness-of-fit parameters were also calculated (

Table 5. Error functions used for hydrological model calibration and validation [57,58].

Nash–Sutcliffe efficiency	$NSE=1-{\displaystyle \frac{\sum _{i=1}^n{\left(O_i-P_i\right)}^2}{\sum _{i=1}^n{\left(O_i-\overline{O}\right)}^2}}$	Range −∞ to 1
Pearson correlation Coefficient	$r={\displaystyle \frac{\sum _{i=1}^n\left(O_i-\overline{O}\right)(P_i-\overline{P})}{\sqrt{{\sum }_{i=1}^n{(O_i-\overline{O})}^2}\sqrt{{\sum }_{i=1}^n{(P_i-\overline{P})}^2}}}$	Range −1 to 1 Optimal value 1
Coefficient of determination	$R^2=r^2$	Range 0 to 1 Optimal value: 1
Error in average flow (water balance)	$PBIAS={\displaystyle \frac{\sum _{i=1}^n\left(O_i-P_i\right)}{\sum _{i=1}^n{O}_i}}\times 100$	Optimal value: 0
Standard deviation ratio	$RSR={\displaystyle \frac{\surd \sum _{i=1}^n{(O_i-P_i)}^2}{\surd \sum _{i=1}^n{(O_i-\overline{P})}^2}}$	Range 0 to ∞ Optimal value 0
Kling–Gupta efficiency	$KGE=1-\sqrt{{(r-1)}^2+{(\alpha -1)}^2+{(\beta -1)}^2}$ α is the variability of prediction errors β is a bias term.	Range −∞ to 1 Optimal value 1

), including bias (PBIAS) and the Kling–Gupta (KGE) parameter, as recommended by other researchers [57,58]. The optimization routine used is based on the SCE-UA (Shuffled Complex Evolution) algorithm [59]. All software code was developed and tested at NTNU. Assuming a time-series with n number of observed (O) and predicted (P) flow values, the six (6) error function parameters shown in Table 5 were computed based on recommendations in [57,58].

Model calibration was conducted for two catchments: 1H8A Ruvu at Morogoro Roadbridge (14,361 km²) and 1H3 Kidunda (6665 km²), as shown in Figure 4. For both catchments, calibration and testing were performed using a 5-year time-series (2013–2017) for calibration and three different hydrological years (2010/11, 2012/13, and 2018/19) for verification. Due to limited data availability, the longer continuous 5-year period was selected for calibration, while the remaining three years with sufficient data were used for verification.

The authors of [57,58] developed a system to translate the numerical value into a verbal classification, ranging from Unsatisfactory to Very Good. This classification scheme was also adapted and used here, and is summarized in

Table 6. Performance criteria for catchment-scale hydrological models [57,58].

Performance Rating	NSE	R2	PBIAS (%)	RSR
Unsatisfactory	≤0.50	≤0.60	≥±15	≤0.7
Satisfactory	0.50 < NSE ≤ 0.70	0.60 < R2 ≤ 0.75	±10 < PBIAS ≤ ±15	0.6 < RSR ≤ 0.7
Good	0.70 < NSE ≤ 0.80	0.75 < R2 ≤ 0.85	±5 < PBIAS ≤ ±10	0.5 < RSR ≤ 0.6
Very good	>0.80	>0.85	<±5	0 < RSR ≤ 0.5

. This classification is used both for calibration and verification. We acknowledge that using the higher-quality data (5 years) for calibration and the lower-quality data (3 years) for verification is debatable, but we believe this approach is optimal and supported by recommendations in [57,58].

2.6.5. The Operational Forecasting Model System—User Interface

As shown in Figure 8, the forecasting system consists of four main components: data input, data storage, model simulation, and reporting. Data input can be performed automatically or manually, depending on the available data platforms. The automatic data input option offers advantages for data quality and control.

The main user interface (“Dashboard”, Figure 11) guides the operator through 6 main steps that are needed to produce a forecast and the reports needed for distribution to users. The step number is marked on the six command buttons used to activate each step in the process.

The first two steps entail inputting climatic data and streamflow data either automatically or manually. For cases where the required respective data are not available, estimated values or long-term historical data values may instead be used. In step 3, the model simulates the streamflow based on observed precipitation, up to the day the forecast will be issued (“Today”). The simulated flow can be compared to the observed flow, and in case of large deviations, the HBV model may be updated to more correct model states, for further simulation, or for the generation of forecasts. This process is described in 2.6.7.

Next, a streamflow forecast is computed (step 4.1) for the next 10 days. The flow forecast will be issued based on the following:

• A meteorological forecast for precipitation (and possibly temperature);
• A hydrological model for transforming precipitation into runoff;
• The model states at the time of the forecast.

To avoid potential bias and other issues, the same types of data were used both during the model preparation phase (calibration and verification) and in the operational phase (updating and forecasting). In both cases, the same selection of rain gauges and flow gauges was used. For operational forecasts, the models were configured to use weather forecasts (rainfall) from TMA or Yr, with a lead time of 10 days.

This forecast is the best estimate of the streamflow but may still have errors. To map the uncertainty systematically (step 4.2), several alternative forecasts can be run based on different variations in the met forecasts, including the use of ensemble forecasts. The model also includes the possibility of using observed precipitation from previous years as input to see what will happen if previous extreme weather events happen again.

The impact on the Kidunda reservoir may be studied (step 5) by feeding the forecasted streamflow into the reservoir. By simulating the water balance and operation, the risk of flooding and/or emptying the reservoir can be analyzed, and the operation can be optimized.

Finally, in step 6, the results are summarized in different reports that are tailored to meet different users’ needs for information. Some examples are given later. During the forecast generation process, the most important results will be stored in statistics files and kept for later use in the model performance assessment. After some time, the study of model behaviour during a variety of different hydrological events may give indications of a possible need for recalibration or the need for new data.

As recommended [13], an optional longer lead time of 3 months was also configured to allow seasonal forecasts for informing water supply schemes and irrigation during drought periods. However, the authors note that forecast uncertainty is caused by initial hydrological or antecedent conditions (e.g., water held in storage in a catchment, in the soil, as groundwater, in surface stores, or as snow/ice) and from the skill of seasonal climate forecasts [11].

2.6.6. The Model Updating Process

The skill of streamflow forecasts depends a lot on the capability to estimate catchment initial conditions (model states) at the time the forecast is issued. The model states are issued as the amount of water in the different storage reservoirs in the model. The model reservoirs correspond to water storage in the river catchment. In the HBV model, the following reservoirs are defined (Figure 10): soil water (SM), upper zone (UZ), and lower zone (LZ).

Updating is based on using observations of streamflow to identify model state errors and correct these to improve the model during operation. Many different methods have been proposed and used in the literature [60,61,62,63,64], ranging from direct manual corrections to Kalman-filter-based methods. Here, in the Ruvu–HBV model system, a hybrid method is implemented, where the model operator can choose between purely manual methods or a semi-automatic optimization method using the powerful Solver tool in Excel [65] to determine optimal updated model states.

In principle, the contents in the natural storage in the catchment could have been measured directly in the field, but in practice, this is only possible for snow storage, which is not used in the Ruvu–HBV model. It is only possible to assess the amount of water in the underground storages (SM, UZ, and LZ) by simulation, that is, by running the model with observed precipitation up to the time of forecast (“Today”), computing each of the model states via water balance equations.

Since the computed flow is a direct function of the model states, by comparing observed and computed flow, it is possible to see if the computed states are correct and acceptable to be used as initial states for the forecast simulation. If there are deviations, there may be a need to correct and update the model states first. The updating will usually be performed at regular intervals and at least every time a forecast is prepared, to keep the model in the best possible shape (“Updated”) and ready for forecasting.

By gradually correcting the model each time when deviations between observed and simulated flows become too large, the model will always be close to the correct states in the catchment, and ready for issuing forecasts. This technique is sometimes referred to as a “nudging scheme”, where small corrections are made when needed to “nudge” the model towards states more like those in the real world. [61].

As an alternative to directly changing the model states, it is also possible to adjust precipitation data. This will lead to a corresponding change in the model states. A benefit of this method is that all model states can be updated in a physically correct way during the simulation. For large deviations, this method is preferred. For a well-calibrated model (as in this case), errors in the precipitation are usually the main source of error and the need for updating. Errors in precipitation data are not primarily caused by uncertainty in measurements per se, but mostly in the calculation of the areal average precipitation, which is the input to the model. This uncertainty is caused by the large spatial variability in rainfall and (nearly always) too few precipitation gauges. With just 5 gauging stations in a catchment with an area of > 14,000 km², it is possible to have large rainfall events falling over areas without any precipitation stations, leading to underestimated areal precipitation. On the other hand, a very local intense rainfall event at one station may lead to overestimated areal precipitation and therefore too high streamflow. Both these types of events are quite common and often lead to the need for rainfall data corrections and model updating.

A more in-depth description of all the model software can be found in the project documentation, given in 6 volumes [66,67,68,69,70,71]. These reports are distributed by DIT.

2.7. Forecast Quality Analysis—Methodology

The achievable accuracy of streamflow forecasts depends on several factors, all with inherent uncertainty:

• The precipitation forecasts;
• The hydrological model structure;
• The hydrological model calibration;
• Initial conditions in the catchment when a forecast is issued.

To analyze and identify possible weaknesses and the need for improvements, it is necessary to identify the main sources of errors among these factors, and, if possible, improve the quality of the dominant factor(s). This type of analysis has been performed for the first year of operation and presented later in Section 3 of this paper.

At the time of writing this paper, the forecasting system has been operational for one year, and a study of the model’s forecasting skills is now possible. Due to unforeseen events, some vital equipment has been damaged or destroyed since the calibration finished, limiting the amount of data available for forecast verification. The damage includes the complete destruction of station 1H3 Kidunda in April 2024, and a partial destruction of the high-water-level staff gauges for 1H8A Ruvu at Morogoro Roadbridge. The loss of 1H3 Kidunda has still not been replaced, so a verification of forecast performance for 1H3 Kidunda has not been possible, leaving only 1H8A to be studied. The damaged upper section of the staff gauge at 1H8A was repaired in June 2025, and the station is now fully operational for accurate monitoring, also at high flood levels.

Here, the loss of staff gauges for high water levels has made flow data above 100 m³/s increasingly uncertain, though correlation with other measurements has made it possible to estimate flow up to 400 m³/s and even higher. The uncertainty is, however, deemed so high that only forecasts up to ca 200 m³/s were included in the analysis, leaving, for example, the entire month of April 2025 out.

In the Ruvu–HBV model, the most important forecast results are logged for possible later analysis. Examples taken from this log during the first year of operation are used here to illustrate the model performance and try to identify possible sources of errors or poor prediction, or the need for model improvements.

The selection of forecasts for verification analysis was grouped by the three main seasons in Tanzania: dry season, short rains, and long rains. The data quality problem and the seasonal distribution of rain in the 2024/25 season resulted in the following seasonal grouping into three cases:

• Dry season (September–November, flows from 5 to 30 m³/s);
• Short Rains (December–February, flows from 10 to 120 m³/s);
• Long Rains (May–June, flows from 5 to 200 m³/s).

The highest flows were observed in April, with estimated peaks up to and possibly above 450 m³/s, but the measurement was deemed very uncertain, and it was therefore decided to exclude the data for April 2025 from the performance analysis. The analysis consists of a graphical presentation of the forecasts, compared to observed flows, and a statistical analysis quantifying the degree of correlation between forecasted and observed flow. Given that precipitation forecasts often carry significant uncertainty, they are frequently identified as a primary source of error within the flow forecasting chain. To specifically evaluate the magnitude of error attributable to precipitation forecast inaccuracies, a virtual experiment can be performed. For the same initial hydrological conditions, forecasts are re-run using the actual observed precipitation data instead of the original forecasted values. This approach, often described as a “perfect precipitation forecast” scenario, directly reveals how closely the model could match reality if meteorological forecasts were flawless. The difference between the original forecast and this idealized run quantifies the impact of precipitation forecast errors on streamflow prediction quality and highlights the potential gains achievable through improvements in meteorological forecasts. Statistical methods were finally applied to calculate numerical indices for assessing how closely flow forecasts aligned with observed flow across various seasons of the first year of forecasting.

It should be noted that the loss of some streamflow measurements at the two stations has not prevented us from making forecasts, only the possibility of verifying the precision of the forecasts. With a well-calibrated model and functional precipitation measurements, the programme system can simulate catchment response (streamflow) and model states up to the time of forecast, and present flow forecasts from there. But without flow measurements, it is not possible to verify if the initial model states are correct or to assess the quality of the forecasts. Since, without flow measurements, it will not be possible to update the model states, they could gradually deviate from correct values and thereby reduce the forecast quality. Without good streamflow measurements, the forecasting system could become less accurate but still operational to produce valuable forecasts in a critical situation. We therefore strongly recommend replacing lost equipment and, if necessary, recalibrating the station’s rating curve if river topography has been changed.

3. Results

3.1. Model Calibration Results, 1H8A Ruvu at Morogoro Roadbridge

The numerical results from both the calibration and verification periods are presented in

Table 7. 1H8A Ruvu at Morogoro Roadbridge—Result from HBV model calibration and verification.

Catchment	1H8A Ruvu at Morogoro Roadbridge		Optimal Parameter Set
Model fit Parameter	Calibration period	Verification period	Area	14,361	km2
			PKORR	0.875
$\overline{{\mathrm{Q}}_{\mathrm{o}\mathrm{b}\mathrm{s}}}$	62.106 m3/s	44.024 m3/s	FC	1743	mm
$\overline{{\mathrm{Q}}_{\mathrm{s}\mathrm{i}\mathrm{m}}}$	62.289 m3/s	42.031 m3/s	BETA	2.55
r	0.923	0.905	LP%	92%
NSE	0.850	0.820	KUZ2	0.100	1/day
R2	0.852	0.819	KUZ1	0.087	1/day
PBIAS	−0.3%	4.50%	UZ1	40	mm
RSR	0.4	0.4	PERC	0.3	mm/day
KGE	87%	84%	KLZ	0.025	1/day

, with graphical presentations shown in Figure 12, Figure 13, Figure 14 and Figure 15. On all these graphs, observed flow is shown in blue and simulated flow in red.

The table also shows the best parameter set with the highest NSE. The remaining five model-fit parameters, along with averages of observed and simulated streamflow, are also included to provide additional performance metrics for the model.

The graphical presentations in Figure 12, Figure 13, Figure 14 and Figure 15 are included to give a better visual impression of the model’s performance.

3.2. Model Calibration Results, 1H3 Kidunda

A summary of results from model calibration and verification is presented in

Table 8. 1H3 Kidunda: Result from HBV model calibration and verification.

Catchment	1H3 Kidunda		Optimal Parameter Set
Model fit Parameters	Calibration period	Verification period	Area	6665 km2
			PKORR	1.14
$\overline{Q_{obs}}$	60.823 m3/s	40.213 m3/s	FC	2200 mm
$\overline{Q_{sim}}$	60.596 m3/s	41.676 m3/s	BETA	2.00
r	0.893	0.906	LP%	80%
NSE	0.800	0.820	KUZ2	0.110 1/day
R2	0.797	0.821	KUZ1	0.100 1/day
PBIAS	0.4%	−3.60%	UZ1	20.3 mm
RSR	0.5	0.4	PERC	0.7 mm/day
KGE	86%	88%	KLZ	0.037 1/day

and Figure 16 and Figure 17. On these graphs, observed flow is shown in blue and simulated flow in red.

3.3. Operational 10-Day Flow Forecasting Results

Operational flow forecasting results are presented as tables and graphs, with the main output being a 10-day streamflow forecast for two stations: 1H3 Kidunda and 1H8A Ruvu at Morogoro Roadbridge. As an example, Figure 18 shows the forecast (red line), observed flow up to the time of forecast (blue line), normal flow (black line), and both observed and forecasted precipitation.

An example, showing the results of sensitivity analysis for precipitation volume for the forecast issued 25 June 2024, is presented in Figure 19 below.

3.4. Seasonal Flow Forecasting Results

An example of a Seasonal forecast is given in Figure 20 and Figure 21. Figure 20 shows a 3-month forecast from October to the end of December in 2025, based on model states as of the 1st of October and a 90-day precipitation forecast from TMA. The plot shows the streamflow forecast in red, together with flow percentiles 10%, 50%, and 90% based on historical flow (2009–2024) in the same period. The blue bars show the daily precipitation forecasts. The forecast is very low, below the long-term average in this period (50% percentile), and most of the time even below the 10% percentile.

Figure 21 shows a different view, focusing on the volume of streamflow during the same period. The blue bars show the actual volume of flow for individual years from 2009 to 2024, in million m³ (Mm³). To the right, the four (4) bars show this year’s forecast (red) and the 10%, 50%, and 90% percentiles for observed flow volumes for the same period. Again, we can see that this year’s forecast (29.5 Mm³) is well below the average (193.9 Mm³). Compared with individual prior years, the forecast is the lowest in the 17-year period.

3.5. Forecast Quality Analysis—Results

As highlighted in Section 2.7, the quality of forecasts is influenced by several factors, some amenable to improvement while others present more significant challenges. To determine whether it is feasible and beneficial to enhance the model system (including input data), it is essential to analyze the nature of past errors, assess whether these can be addressed through model refinement or improved input data, and, if possible, evaluate the associated costs in terms of time and resources.

Following nearly one year of system operation, we have reviewed multiple forecasts, identified principal sources of uncertainty, and documented their impacts accordingly.

This analysis will therefore concentrate on two additional factors that are crucial for forecast quality: model updating and meteorological forecasts. Figure 22 demonstrates how both elements affect the accuracy of forecasts. On the left side (A), the plot shows results of two forecasts made with a well-updated model, where the model states are simulated and adjusted so that the simulated flow (Sim) at the time of forecast (18/Jan) is very close to the observed flow (Obs).

On the right side of Figure 22B, the plot shows the results of two forecasts made with the same model, but this time the model had not been updated. Hence, the model states are too low, so the simulated flow (Sim2) at the time of forecast (18/January) is also too low compared to the observed flow (Obs).

Two different forecasts are produced for each case, one based on the actual precipitation forecast issued by TMA (AFc2), the other based on “Perfect precipitation forecast” (PFc2), which is the observed precipitation for these 10 days, the best possible forecast that could be achieved with a perfect meteorological model. The difference between these two forecasts illustrates how much the uncertainty in precipitation forecasts will impact the streamflow forecast. With an updated model as shown in Figure 22A, the difference between Obs and PFc reflects all other errors from sources like model structure or calibration. If the model is not updated as shown in Figure 22B, forecasts begin with incorrect states, leading to poorer results during most of the 10-day period.

The Ruvu–HBV model has implemented good tools for model updating and requires some time (a few minutes) from the model operators each time a forecast is prepared. Our experience from the testing period is that only a few actual corrections are needed; if the model has been updated, it will usually stay within acceptable states for several days, sometimes weeks (in baseflow-dominated periods). If the model is not updated, it may still give useful results, as shown in Figure 22B. The time needed for updating is so small that the improved forecast quality comes at a very low cost. In the detailed investigation (Section 3.5.1, Section 3.5.2 and Section 3.5.3), we have assumed a fully updated model during the entire test period.

The results are grouped by and presented for three seasons of the year: “Dry season”, “Short rains”, and “Long rains”.

Due to the loss of gauging station 1H3 Kidunda, the forecast quality analysis was only possible for station 1H8A Ruvu at Morogoro Roadbridge.

3.5.1. Case 1: Dry Season: Low Flow—September to November 2024

In the first case, the results from 8 forecasts made during the dry season were analyzed and compared to observed streamflow at station 1H8A Morogoro Roadbridge. Some main results are shown in Figure 23 and Figure 24. Each forecast is marked with a red diamond at the time the forecast was issued, and a dotted red line which shows the forecasted flow for the following 10 days. The blue line shows the observed flow, and the blue bars below the sum of predicted precipitation during the forecast period (Figure 23) and daily observed precipitation (Figure 24) during the same period.

The first, Figure 23, shows the forecasts computed from a well-updated initial state (the computed flow is close to the observed flow) and with forecasted precipitation as input. We notice that forecast numbers 1,2, 6, 7, and 8 (counted from the left) are quite close to the observed streamflow for the following 10 days. Three of the forecasts, 3, 4, and 5, have too low and falling flow, while in fact steady or even increasing flow was observed during the next 10 days.

To identify the cause of error, forecasts were recalculated using the “Perfect forecast#” (actually observed precipitation) over the 10-day period. As shown in the second figure (Figure 24), all eight forecasts would have performed well with accurate precipitation forecasts. Comparing the precipitation bars shows, the forecasted values were too low, likely the main source of error. There is no evidence of model errors.

3.5.2. Case 2: Short Rains: Medium–High Flow—December–February 2024/2025

The results (Figure 25 and Figure 26) indicate lower forecast precision compared to the first case, likely due to the dry season’s higher share of baseflow, making forecasts more dependent on initial conditions than in wetter periods dominated by stormflow.

In forecasts no. 4 (31 December) and no. 6 (20 January), predicted precipitation was less than observed, so using actual precipitation would have improved flow forecasts. However, even with observed precipitation, forecasts no. 4 and no. 7 did not match peak flows, suggesting areal precipitation may have been underestimated. The high flow deviation on 1 February may be due to both underestimated precipitation and possible errors in flow measurements, as water levels exceeded the damaged staff gauge.

The last two forecasts in February were nearly perfect, and similarly as case 1 shows that the model gives excellent results for the recession part of hydrographs, dominated by baseflow.

3.5.3. Case 3: Long Rains: Forecasts During High Flow May–June 2025

This case shows model performance during the end of the long rains in May and early June 2025, and into the start of the dry period in July. The problems with the missing staff gauge for high flows result in uncertain data quality above 100 m³/s and, unfortunately, missing data for April.

The first two forecasts (4th May and 17th May) were low due to poor precipitation predictions (Figure 27). With accurate precipitation forecasts (Figure 28), both flow forecasts would have improved, but flow uncertainty limits categorical conclusions. The other six forecasts were almost flawless, further showing the models’ strong performance in predicting low flows driven by groundwater and steady recession, with minimal impact from precipitation uncertainty.

3.5.4. Statistical Analysis

Results of statistical analysis of forecast skills are presented in

Table 9. Summary statistics for Ruvu–HBV forecast skills in hydrological year 2024/25.

Season	Dry	Short Rains	Long Rains	All year
Number# of forecast days	70	90	67	227
Met Forecast
R2	0.964	0.889	0.947	0.933
PBIAS	4.7	18.8	16.6	13.5
“Perfect” Forecast
R2	0.981	0.928	0.976	0.962
PBIAS	0.7	8.6	4.6	4.6
Improvement (r2)	0.017	0.039	0.029	0.028

. During the dry season, based on 70 forecasts (days) evaluated, the correlation coefficient between observed and predicted flow was 0.964 for actual meteorological forecasts, and 0.981 for “Perfect forecasts” that utilized observed precipitation data, resulting in an r² increase of 0.017. PBIAS was used to quantify bias by comparing average forecasted and observed flows.

Similar results were obtained for short- and long-rain periods and presented in the same Table. Better precipitation forecasts will improve flow forecasts, with r² gains between 0.017 and 0.039. Forecasts tend to underestimate precipitation (PBIAS > 0), but this can be corrected by increasing precipitation forecast values by 13.5%.

The following figures illustrate the forecast skills by plots comparing forecasted flow and observed flow for the same days. Figure 29, Figure 30 and Figure 31 compare forecasted and observed flows during the three different seasons, using both actual (Left side) and “perfect” precipitation forecasts (Right side). In the Figures, forecast-observation pairs are plotted as dots, with a dotted-trendline derived from Ordinary Least Squares (OLS) regression indicating the overall agreement.

4. Discussion

4.1. Model Calibration

Model calibration relied primarily on the Nash–Sutcliffe Efficiency (NSE) parameter, which reached values of 0.85 (1H8A) and 0.80 (1H3) during calibration and 0.82 (Both) during verification (see Table 7 and Table 8). The tables also show the optimal parameters corresponding to the highest NSE. The remaining five model-fit parameters, along with averages of observed and simulated streamflow, are also included to provide additional performance metrics for the model.

According to quality classification in Table 6, calibration results are very good across almost all efficiency parameters (NSE, R², PBIAS, and RSR). The KGE value is also notably high. The NSE value achieved both for calibration (0.85 and 0.80) and verification (0.82) are well above the limit of 0.80 needed to be classified as a “Very good” model. Similarly, the R² value of 0.852, PBIAS of −0.3% and RSR of 0.4 all classify the model calibration for 1H8A as “Very good”, and almost similar for 1H3.

The verification results remain strong despite some lower-quality input data caused by gaps in the precipitation series (see Section 2.5.8). The verification period was much drier than the calibration period, with only 67% of the average flow, which adds confidence to the model’s reliability under very different climatic conditions. Thus, the model is considered suitable for its intended operational use.

The model calibration results measured by NSE also compare favourably with other HBV-model results in the Ruvu Basin [30,31,32], where optimal NSE values typically were found in the range 0.67–0.73 in the calibration period and 0.57–0.83 in the verification period for the three (3) catchments studied. However, since both the catchments and calibration period were different from this study, a more in-depth comparison is difficult.

The HBV model manages to compute both low flows (baseflow) and flood peaks well, indicating a calibration valid over a wide range of flows. This is also supported by the flow duration curves; see Figure 14 and Figure 15.

4.2. 10-Day Forecast

As mentioned in Section 2.7, the quality of forecasts is influenced by many factors, some amenable to improvement while others present more significant challenges. To determine whether it is feasible and beneficial to enhance the model system (including more input data), it is essential to analyze the nature of past errors, assess whether these can be addressed through model refinement or improved input data, and, if possible, evaluate the associated costs in terms of time and resources.

Following nearly one year of system operation, we have reviewed and quantified the quality of forecasts, identified principal sources of uncertainty, and documented their impacts accordingly. Based on these results, it is possible to discuss whether it is feasible to improve the forecasts and where to start.

While hydrological models can always be improved through further research and better data, practical constraints often limit enhancements. The HBV model is widely used because it balances complexity with manageable data needs, especially in areas with limited and sometimes uncertain data, like in Ruvu. The strong calibration and verification results obtained here and in other studies in the region, and its long history of use with the same structure, suggest that significant improvements to its structure are challenging.

The HBV model and similar conceptual models need lengthy, high-quality time series of data for calibration, both precipitation and streamflow data. If new precipitation and streamflow stations were to be installed, gathering and preparing the necessary new input data would need to take considerable time, potentially five years or more, before the new data could be fully utilized for improved calibration. To develop a forecasting model now, we therefore find that both model structure and calibration are as good as possible, given the present state of historical and operational data sources.

The Ruvu–HBV model has implemented good tools for model updating and requires only a few minutes’ work by the model operators each time a forecast is prepared. Our experience from the testing period is that usually just a few corrections are needed, and when the model has been updated, it will usually stay within an acceptable range of states for several days, sometimes weeks, especially during baseflow-dominated periods. The time needed for updating is so small that the improved forecast quality obtained by the updating comes at a very low cost. In the detailed investigation of forecast quality, we have ensured that a fully updated model was used during the entire test period.

4.3. Seasonal Forecast

Seasonal flow forecasts will have lower accuracy than the 10-day forecasts, due to higher rainfall forecast uncertainty, but can still be useful, particularly in the dry season when groundwater storage supports longer flow response times. The HBV model’s lower zone stores water that is released very slowly, allowing sustained flow through dry months even with minimal precipitation. Updating initial storage and flow enables baseflow prediction months ahead, using occasional observed flow updates.

4.4. Ensemble Forecasts

The optimal approach for analyzing forecast uncertainty is to utilize ensemble meteorological forecasts. Once these forecasts become accessible in Tanzania, the model system can incorporate them to generate ensemble streamflow predictions. This allows statistical analysis of flow event probabilities, information useful for managing and planning future Kidunda reservoir operations. The model system is prepared for this possibility.

The key takeaway from the analysis of forecast quality is that uncertainty in areal precipitation observations and forecasts typically represent the primary source of error, especially during the rainy seasons. However, enhancing meteorological forecast accuracy may require the development of meteorological models tailored to local conditions, particularly in mountainous areas. Observed areal precipitation values can be improved by increasing the number of meteorological stations and strategically optimizing their placement. However, this approach may result in increased costs and additional time required for data collection and quality control.

We also found that it is essential to update the model as close as possible to actual states in the basin before issuing the forecast; otherwise, initial errors are likely to persist throughout much of the forecast period

5. Conclusions

This paper shares results from developing a streamflow forecasting system with an HBV model for the data-scarce Ruvu catchment in Tanzania, aiming to make the best possible use of existing data and address gaps in hydrological modelling capacity.

In case of data inadequacy and uncertainty challenge, the study has employed various approaches such as strategic data maximization, robust quality control, and optimum period selection. They entailed identifying 5 key stations from 14 rainfall stations, using a hybrid Thiessen polygon and Cubist model analysis; rigorous gap-filling, rating curve recalibration using optimization tools, and exclusion of unreliable periods; and identifying 5 high-quality years (2013–2017) for calibration and 3 years for verification. In the final analysis, data from only five rainfall and two streamflow stations proved sufficient for further modelling activities.

For the modelling capacity gap, the study has used various approaches such as appropriate model selection, phased implementation, capacity transfer, and user-centric design. They entailed adopting the user-friendly, well-documented Excel–HBV model, customized for local conditions; structuring development across three phases (data assessment, customization, operational integration); donating full source code and rights to the Dar es Salaam Institute of Technology (DIT) as a research partner higher learning institution in Tanzania, ensuring local ownership and sustainability; and provision of intuitive dashboard requiring minimal expert intervention for daily operation. The success of these interventions is quantitatively evidenced by reduced model complexity as a result of removing irrelevant model routines (snow); full operational launch with a documented, six-volume manual suite; and a system designed for use by local water authorities, not just modelling experts. In addition, the model performance, as measured by NSE for calibration and verification periods, is 0.85 and 0.82 for Ruvu Roadbridge (1H8A), and 0.80 and 0.82 for Kidunda (1H3), respectively, and all are classified as “Very Good”. Furthermore, a PBIAS of less than ±5% in calibration indicates excellent water balance simulation.

In order to ensure good operational forecast performance, the study emphasized transparent uncertainty handling, diagnosis of error sources, and model updating. It entailed conducting experiments for quantifying the primary error source of meteorological forecasts; generating 10-day and seasonal (3-month) streamflow forecasts; and maintaining forecast accuracy by hybrid manual/automatic state updating (“nudging”). The forecast’s performance in this study is evidenced by an annual forecast R² of 0.933, with operational meteorological forecasts improving to 0.962 with “perfect” precipitation; dry season performance with R² of 0.964, demonstrating high skill in baseflow-dominated periods; and PBIAS for forecasts of 0.866, indicating a slight systematic under-forecasting correctable by a ~15% precipitation adjustment.

Notwithstanding the above very good performance, the research project navigated significant practical obstacles:

• Infrastructure vulnerability: The complete destruction of the Kidunda (1H3) gauging station in April 2024 and damage to the 1H8A staff gauge highlighted the fragility of monitoring networks. Mitigation involved data correlation and temporary repairs, though this limited verification for 1H3.
• Inherent data scarcity: Non-concurrent records, missing data, and sparse spatial coverage were endemic. The response was a conservative, best-available-data approach, prioritizing quality over quantity and employing statistical infilling.
• Meteorological forecast limitations: Global and regional forecast products (Yr.no, TMA) introduced significant uncertainty, especially during rainy seasons. The study quantified this impact, providing a clear target for future improvements in local meteorological modelling.

This initiative provides a replicable blueprint for operational hydrology in data-scarce regions. It has been demonstrated that it is possible to launch a “Very Good” performing (NSE > 0.8) operational forecasting system without dense networks or long, pristine data series. In addition, the choice of Excel–HBV ensures low-cost, low-barrier maintenance and local capacity development, avoiding dependency on proprietary or overly complex systems. Furthermore, the analysis clearly shows that the largest gains in forecast accuracy will come from enhancing precipitation measurement networks and improving the resolution and accuracy of local meteorological forecasts. Investment here offers the highest return.

It may be concluded that this study successfully demonstrates the feasibility and effectiveness of developing an operational streamflow forecasting system in a data-scarce catchment typical of developing countries. By systematically addressing the dual challenges of modelling capacity gaps and data inadequacy and uncertainty, the research establishes a practical, scalable framework for hydrological forecasting in the Ruvu catchment, Tanzania—a critical water source for the megacity of Dar es Salaam in Tanzania.

This research has effectively bridged the modelling capacity and data adequacy gaps in the Ruvu catchment. By making the most of limited data through rigorous analysis, selecting and customizing an appropriate conceptual model, and embedding the system within local institutions, the study has transitioned from theoretical modelling to operational service. The deployed Ruvu–HBV forecasting system now provides actionable water inflow forecasts for Dar es Salaam’s water supply management and is pre-configured for the future Kidunda Dam operations.

This work demonstrates that with pragmatic methodology, strategic partnerships, and appropriate technology, developing countries can build and own the hydrological forecasting tools essential for water security, climate resilience, and sustainable development.

The hydrological model calibration described herein is specific to this catchment, reflecting its state and climatic conditions during the calibration and validation periods. Recalibration may become necessary if significant land-use changes occur, as has been observed in other catchments experiencing increased irrigation and water withdrawals. Since both land-use and climate change typically progress gradually, it is likely that recalibration will only be required at intervals of several years.

6. Recommendations

The successful development and operational launch of the Ruvu–HBV streamflow forecasting system provides a validated, practical framework for hydrological service development in data-scarce regions. To consolidate this achievement, ensure its sustainability, and guide similar initiatives elsewhere, the following recommendations are formulated, structured around the core challenges of data, modelling, and operations.

6.1. Institutionalize and Modernize Hydrometeorological Monitoring Networks

The vulnerability of physical infrastructure (e.g., the destruction of the Kidunda station) and inherent data scarcity remain the most critical threats to forecast reliability. We recommend the following:

• Prioritizing resilient and redundant monitoring infrastructure. Key stations like Kidunda (1H3) must be rebuilt with robust design and real-time telemetry to withstand extreme events.
• Strategic expansion of automated networks. Investments should focus on deploying cost-effective, automated weather stations (like TAHMO) to improve spatial coverage, particularly in topographically complex and data-sparse areas, to reduce areal precipitation uncertainty.
• Establishing national data stewardship protocols. This includes formalizing procedures for continuous quality control, gap-filling, rating curve maintenance, and open-data sharing to create a sustainable and high-quality data foundation for all water sector applications.

6.2. Embed Forecasting Systems Within Local Institutional Frameworks

Long-term sustainability depends on local ownership and operational capacity. We recommend the following:

• Formalizing the institutional home for the forecasting service within a relevant national or basin authority (e.g., Wami–Ruvu Basin Water Board), ensuring dedicated budgets and staffing.
• Implementing continuous capacity-building programmes. This includes advanced training for model maintenance, forecast interpretation, and system updating for engineers and technicians at local partner higher learning institutions like the Dar es Salaam Institute of Technology (DIT).
• Developing clear operational protocols that integrate forecast products into the standard decision-making workflows of water supply utilities (e.g., DAWASA) and disaster management agencies.

6.3. Prioritize Improvements in Meteorological Forecasting

Our analysis identified precipitation forecast error as the primary constraint on hydrological forecast accuracy. To address this, we considered the following:

• Foster deeper collaboration between hydrological and meteorological services (TMA) to co-develop bias-corrected, downscaled forecast products tailored for hydrological input.
• Invest in and evaluate ensemble prediction systems (EPS) to quantify and communicate forecast uncertainty, which is crucial for risk-based water management and reservoir operation.

6.4. Adopt and Promote the Phased, Pragmatic Modelling Paradigm

The study’s methodology—appropriate model selection, phased implementation, and user-centric design—proves to be highly replicable. We recommend the following:

• Promoting the use of transparent, adaptable conceptual models (like HBV) as a first step in operational forecasting in developing regions, avoiding premature adoption of overly complex systems that exceed local maintenance capacity.
• Documenting and sharing modular system blueprints, including data preparation scripts, calibration routines, and dashboard designs, to reduce start-up costs for new basins.
• Planning for adaptive model evolution. Systems should be designed from the outset, as demonstrated with the three-stage Kidunda dam model, to accommodate future changes in catchment regulation and land use.

6.5. Implement a Cycle of Continuous Performance Review and System Evolution

An operational system must learn and adapt. We recommend the following:

• Instituting a formal forecast verification routine to routinely assess performance against observations, diagnose emerging errors, and trigger necessary model updates or recalibration.
• Developing a structured feedback mechanism with forecast end-users (e.g., reservoir operators and city water managers) to ensure the system evolves to meet practical decision-making needs.
• Exploring incremental technological upgrades, such as coupling with hydraulic routing models for flood inundation mapping or integrating seasonal climate forecasts for long-term water resources planning, as local capacity grows.

6.6. Advocate for Strategic Policy and Investment

Translating a successful pilot into a permanent national service requires high-level support. We recommend the following:

• Mainstreaming operational hydrological forecasting into national water resources management policies, climate adaptation strategies, and infrastructure development plans.
• Securing dedicated, long-term funding for monitoring networks, model maintenance, and capacity development, moving from project-based to programme-based support.
• Fostering regional knowledge networks among East African countries facing similar challenges to share solutions, data, and tools, thereby increasing collective efficiency and resilience.

As evidenced by these study findings, the journey from data scarcity to operational forecast provision in the Ruvu catchment demonstrates that the principal barriers are fundamentally institutional and capacity related. The recommendations above chart a path forward that balances technical pragmatism with strategic investment. By solidifying the data foundation, empowering local institutions, and systematically reducing the largest source of forecast uncertainty—meteorological input—the promising start documented here can mature into a resilient, indispensable pillar of water security for Dar es Salaam and a transferable model for sustainable hydrological service development across the developing world.