Predicting River Discharge in the Niger River Basin: A Deep Learning Approach

Abstract

Across West Africa, the River Niger is a major source of freshwater. In addition, the river system also provides services such as aquaculture, transportation, and hydropower. The river network plays a critical role in the hydropolitics and hydroeconomics of the region. Therefore, River Niger is integral to the development of West Africa, hence, there is a need to ensure that the river’s ecosystem is a healthy one. In light of the changing climate and its associated threats such as droughts and floods, constant monitoring and measurements of the the river’s flow system cannot be overemphasized. This study investigates temporal variations in annual river discharge characteristics at eight stations (Koulikoro, Dioila, Kirango, Douna, Mopti, Dire, Ansongo, and Niamey) in the Niger River basin, presenting detailed quantitative findings. Analyzing discharge data of River Niger from 1950 to 1990, the minimum discharge measures (minimum and 10th percentile) exhibit a consistent decreasing trend post-1960, persisting into the 1990s at several stations. Central tendency measures (mean and 50th percentile) also consistently reduced since 1950, with near-zero median values observed in Diola and Douna. Recovery in mean discharge is evident in Ansongo after 1980. Extreme values (maximum and 90th percentile) show decreasing trends across all stations, with some locations exhibiting a slight recovery after 1980. The decreasing trend in annual minimum, mean, and maximum values has implications for water resources, affecting hydroelectric generation, fish farming, and dry season irrigation. Machine learning algorithms (MLAs) are deployed to predict the prediction of monthly river discharge, with LSTM identified as the best-performing model overall. However, model performance varies across locations, with TCN excelling in Diola but underperforming in Koulikoro. This study emphasizes the chaotic nature of time series data and external drivers limiting the long-term predictive capabilities of MLAs. Quantitative evaluation of MLA performance reveals specific strengths and weaknesses at each station. This study underscores the importance of predicting the 10th percentile of annual river discharge for water resource planning. Models exhibit diverse performance across basins, emphasizing the need for tailored approaches. Further analysis considers measures of central tendencies, predicting the 50th percentile (Q50) and mean discharge values. TCN emerges as the best model for Q50 prediction, showcasing superior performance over other models. Additionally, the study delves into predicting high and low extreme discharges, crucial for understanding potential flood events and preparing for meteorological and hydrological droughts. This study concludes by emphasizing the necessity for location-specific studies in the River Niger basin to facilitate an enhanced integrated river management system.

1. Introduction

Rivers are important components of the hydrological cycle that account for some of the richest biodiversity in the world with vibrant ecosystems. Riverine ecosystems are important because they provide certain ecosystem services such as drinking water, fish and other aquatic foods, flood protection, or spaces for recreation. These diverse and dynamic ecosystems in context drain their immediate surroundings and sometimes empty into another river, sea, or ocean [1]. For instance, the riverine ecosystem receives energy and material flux influenced by factors such as changing climate, anthropogenic pollution, global-to-local economies, and land use–land cover activities [1]. Rivers also hold the potential for improved, efficient, and climate-friendly energy generation [2]. These factors dictate the resilience of these systems in such a way that it stretches their ability to recover or initiate a new state. As a result of the pressures from these factors, river systems are facing an increased rate of extreme events, such as droughts, floods, and pollution. These activities affect biodiversity, increase temperature, and enhance thermal pollution.

As a result of the increasing pressures on river systems, various extreme and compound events are increasing [3,4,5,6,7,8] in magnitude and frequency globally. For instance, the increased global warming greatly influences the prevailing drying conditions, which invariably cause extreme hydrological events [3]. Furthermore, Trenberth and Shea [3] showed that on a global scale, a positive correlation exists between high maximum temperature and dry conditions across the continent during the warm seasons. Similarly, during the 2012–2014 drought period in California, ref. [4] found a great influence on river discharge by increased temperature. There are other variables of influence on river discharge besides rainfall variability and increased temperature. Adeyeri et al. [5] found that the increased river discharge during the period between 1971 and 2013 is approximately due to 50% rainfall variability and 50% human activities. From the foregoing, it becomes clear that there is a need for adequate monitoring of river systems for flows and fluxes. This becomes necessary to maintain the healthiness of the river system and ensure that they provide needed services. Monitoring activities, using traditional and sophisticated instruments have helped to understand riverine ecosystems much better by providing various hydrogeomorphological information and services. Globally, river monitoring is an integral part of the environment, ensuring that common concerns regarding flow, water quality, and fish habitat are monitored and well-documented; environmental standards are not violated; and that these ecosystems are as healthy as they can be [6,7]. Specifically, streamflow monitoring and measurement for various purposes such as but not limited to flood forecasting, turbidity, and phosphorus and water level carried out at (sub)-daily, weekly, and monthly scales depending on staff, finance, and the importance attached to these activities by various governments are some of the most frequently monitored aspects of river flows. Across the globe, more monitoring stations than the existing ones for streamflow are needed [9]. Sadly, monitoring stations no longer exist or partially function in most countries, especially those in the tropics. River monitoring was paramount in the early 18th century, especially in some African countries. For example, the Egyptians could forecast periods of extreme events such as hunger, abundance, and catastrophe along the Nile using Nilometers [8]. Unfortunately, since the late 19th century, most gauging stations are no longer functioning due to various reasons such as economic adjustment programs, inadequate staff, low priorities, and the level of importance attached to these activities across Africa. Hence, the continued increase in ungauged basins in most African countries, especially in West Africa [10,11], is concerning as most most gauges along River Niger are no longer functioning [10,11].

To achieve the intent and purpose of monitoring river systems, there is a need to improve gauging systems or provide validated datasets that could serve as proxies for flow measures. This becomes essential to ensure a sustainably built environment, preserve crops through irrigation, and to an extent cushion the impacts of either flood or water scarcity [8]. The economic performance of most West African countries suggests that river monitoring is a manageable issue in terms of the scale of preference. Therefore, for the foreseeable future, the number of hydrological monitoring stations will likely remain the same across most rivers in West Africa. From the foregoing, there is the need to develop and design solutions that can help in some ways provide needed streamflow data for these data-scarce regions based on testing and validation of available datasets using improved statistical algorithms [10]. Prediction of streamflow has been implemented for numerous ungauged basins globally, employing diverse techniques and algorithms [12,13,14,15,16,17]. Kratzert et al. [18] showed that machine learning algorithms, emphasizing LSTM, performed better than traditional hydrological models. In their study on the prediction of floods using machine learning algorithms (MLAs), Zehra [12], using Non-linear Autoregressive Exogenous (NARX) and Support Vector Machine (SVM) algorithms affirmed that possibilities abound in the prediction of extreme events such as floods using MLAs. Tayfur and Moramarco [13] reported that a genetic algorithm model could help predict hourly discharge hydrographs using elevation data. Furthermore, Schreider et al. [14] in their study on ungauged catchments in Northern Thailand showed that the first pass approach algorithm provides an accuracy of 13–17 percent of the relative error for the monthly time step. These studies show the potential of MLAs as an aid in the prediction of streamflow. In a study aimed at comparing MLAs for discharge prediction [16], the authors used decision tree (DT), multilayer perceptron (MLP), random forest (RF), gradient boosting (GB), RNN-LSTM, and CNN-LSTM predicted dam discharge. The study concluded that predicting the amount of discharge from a dam is possible. In another similar study, using combined learning models, the authors affirmed that combined MLAs can predict inflow through inflow learning, considering flow characteristics such as flow regimes [15]. In addition, Snieder et al. [17] in their study on resampling and ensemble techniques for flow forecast using random undersampling (RUS), random oversampling (ROS), and the synthetic minority oversampling technique for regression (SMOTER), and four ensemble techniques, randomized weights and biases, bagging, adaptive boosting (AdaBoost), and least-squares boosting (LSBoost), the authors concluded that resampling produces marginal improvements to high stage prediction accuracy, whereas ensemble methods produce more substantial improvements, with or without resampling. Other related studies include LSTM studies for river discharge in Indonesia [19], GRU prediction of river discharge in China [20], and TCN prediction of river discharge in the Yellow River, China.

These studies have shown the possibility of predicted flow through testing and validation using available flow records. The focus of this paper is to consider some characteristics of river discharge such as low and high flows. The consideration of low and high flows in rivers of West Africa has received little attention despite the global relevance of low and high flows in hydrology [21]. In their paper, Smakhtin [21] affirmed that predicting low flows would go a long way in understanding catchment processes. As considered in this study, most existing studies on flow prediction using machine learning have focused on river discharge using average flow conditions against low and high flows. Furthermore, streamflow prediction within the Niger Basin is of utmost importance based on the representation of the river as it flows through nine West African countries, playing a pivotal role in shaping regional hydropolitics, influencing the agricultural landscape, and impacting the hydroeconomies of West African member states. The significance of studying the Niger flow becomes imperative for the continuing development and stability of the region. Studies have identified the River Niger as highly dynamic and chaotic, emphasizing the complexity and variability inherent in its hydrological patterns [10,11]. Hence, highly efficient methods such as deep learning algorithms can help with short-term river flow predictions. This study aims to predict the high and low river flow across eight gauging stations within the Niger Basin using four deep-learning approaches. Identifying the most efficient algorithms will go a long way in managing water resources across the sub-region and support efficient regional decision-making.

2. Study Area

The Niger Basin, situated within equatorial West Africa [22], consists of an intricate network of river systems. It is the largest river Basin in Africa, with an area of more than 2.1 million square kilometres spanning nine countries (Benin, Burkina Faso, Cameroon, Chad, Côte d’Ivoire, Guinea, Mali, Niger, and Nigeria). It is the principal river in West Africa and the third longest on the continent; it took its source from the Guinean Highlands and stretches approximately 4200 km. The basin has diverse ecosystems, ranging from humid rainforests in the south to arid savannahs in the north. It provides home to a variety of flora, fauna, and endangered species. The Niger Basin is a region of great ecological cultural and economic importance; hence, its sustainable management is crucial for the livelihoods of millions of people who depend on its resources for the preservation of its rich biodiversity.

3. Data

The daily river discharge used in this study was obtained from the archives of the Global Runoff Data Centre (GRDC) (https://www.bafg.de/GRDC/EN/Home/homepage_node.html accessed on 1 June 2023). The data for eight stations within the Niger Basin were obtained from 1950 to 1990 (Figure 1). These stations are Koulikoro, Diola, Kirango, Douna, Mopti, Dire, Ansongo, and Niamey. The period chosen has the longest continuous data for all the locations with minimal missing data points. The statistical properties of river discharge at these locations have been reported in the literature [10]. This study considered six statistical characteristics of river discharge, including minimum discharge, quantiles (0.1, 0.5, and 0.9), mean discharge, and maximum discharge values. These properties are plotted for all the locations (Figure 2).

4. Methodology

4.1. Deep Learning

Machine learning is a branch of artificial intelligence that allows computers to make inferences and decisions without prior programming. Deep learning, a subset of machine learning, is defined as a complex dynamical model that trains any given dataset using the solution of a non-convex optimization [23]. Deep learning mimics the human brain in learning data for making predictions [24]. It employs several layers for data representation using a computational model [25]. In this study, past values were considered predictors. All algorithms were implemented using the Python package, DARTS [26]. The flow chart for the entire process is shown in Figure 3. First, the raw river discharge data statistics such as minimum and maximum value per month are estimated. The estimated statistics are then scaled and transformed for easy processing in the deep learning pipeline. The data are split into training (490 points) and test (92 points) data. The training data were used to generate the deep learning model. Based on the model, predictions were made, and compared with the test data to estimate the model errors. Deep learning algorithms have been applied successfully for several applications including speech analysis, weather forecasting, and dynamical systems prediction [27,28]. In this study, the deep learning algorithms were implemented using the Darts Python package [26].

4.2. Recurrent Neural Network (RNN)

RNNs use several successive recurrent layers to identify time relationships within a data structure and make predictions based on past input states [25]. Given a time series, $x_i$, the RNN predicts for each timestep an output of the form:

$$y=g_2(W_{ya}{a}_t+b_y)$$

(1)

where $a_t$ is the activation given by:

$$a_t=g_1(W_{aa}{a}_{t-1}+W_{ax}{x}_t+b_a)$$

(2)

where $g_i$ are activation functions and $W_{aa},W_{ya},W_{ax},b_a,$ and $b_y$ are coefficients. In this study, 50 hidden layers were used. Two problems associated with RNNs are vanishing and exploding gradients. RNNs have been used in several studies to predict river discharge [29,30,31]. A modified RNN model outperformed an artificial neural network model in river discharge prediction [32].

4.3. Gated Recurrent Unit (GRU)

The GRU attempts to solve the vanishing gradient problem in the RNN by using both the update and reset gates [33]. The update gate is defined as:

$$\begin{array}{c}\hfill f_t={\sigma }_g(W_z{x}_t+U_z{h}_{t-1}+b_z)\\ \hfill r_t={\sigma }_g(W_r{x}_t+U_r{h}_{t-1}+b_r)\end{array}$$

(3)

where $f_t$ and $r_t$ are the forget and reset gate functions, respectively, ${\sigma }_g$, $h_{t-1}$, $b_z$ are the activation function, cell state vector for time $t-1$, and $b_z$, respectively, and assigned weights are $W_z$ and $U_z$ [34]. The activation function becomes

$${\tilde{h}}_t=f(W_h{x}_t+r_t\odot U_h{h}_{t-1}+b_h)$$

(4)

where ⊙ is elementwise multiplication [35]. In this study, 50 hidden layers with batch size of 56 were employed. Prediction of river discharge has been carried out using the GRU [36].

4.4. Long Short-Term Memory (LSTM)

The LSTM tends to overcome the vanishing gradient problem commonly experienced in RNNs. It operates by using three gates (input gate, forget gate, and output gate). Given a time series x as input to the LSTM block, the current state is defined as:

$$C_t=tanh{W}_c[a_{t-1},x_t]+b_c$$

(5)

where $W_c$ and $b_c$ are the weight of the current input and block bias, respectively [27]. The input from the previous state is $a_{t-1}$. The update gate function is obtained from the expression:

$$\mathrm{\Gamma }u=\sigma (W_u[a_{t-1},x_t]+b_u)$$

(6)

where $\sigma$ is the sigmoid transfer function. The forget gate $\mathrm{\Gamma }f$ and output gate $\mathrm{\Gamma }o$ determine the recurrent cell’s eliminated information and output, respectively. The current state is obtained as:

$$a_t=\mathrm{\Gamma }o\times tanh{\tilde{C}}_t$$

(7)

where ${\tilde{C}}_t$ is the self-looped cell obtained as:

$${\tilde{C}}_t=\mathrm{\Gamma }u\times C_t+\mathrm{\Gamma }f\times {\tilde{C}}_{t-1}$$

(8)

In this study, 50 hidden layers were used. LSTM has been used to predict river discharge in Indonesia [19], South Africa [36], and India [30]. Sahoo et al. [30] reported that LSTM showed better performance in river discharge over RNNs.

4.5. Temporal Convolution Network (TCN)

A TCN has input and output of the same length with dilated and convoluted layers. Only observed inputs $x_1, x_2,\cdots , x_N$ can be used to predict the output $y_i$. In a TCN, fixed steps between adjacent filters are used rather than the whole history in dilation. To overcome the vanishing gradient problem and speed up computation time, the method also uses residual connections [37]. Given a time series $q=(q_0,\cdots , q_t)$ and filter $f:\{0,\cdots , k-1\}$, on element j of the sequence, a layer F is defined as:

$$F\left(j\right)=\left(q{\times }_{d}f\right)\left(j\right)=\sum _{i=0}^{k-1}f\left(i\right)·q_{s=d·i}$$

(9)

where $d=5$ is the dilation factor, $k=10$ is the filter size, and $s-d·i$. This study takes the values of $d,k,$ and s as 5, 10, and 5, respectively. TCNs have been shown to have longer memory retention, and are simpler and clearer than recurrent networks [38]. TCNs have been deployed to predict river discharge in different basins worldwide [39,40,41].

5. Results and Discussion

The temporal variations in the annual characteristics of river discharge at different stations are presented in Figure 2. The minimum discharge measures (minimum and 10% percentile) were found to have a decreasing trend after 1960 at all the stations. The decreasing trend persisted into the 1990s at Diola, Douna, Dire, and Niamey. Kirango and Ansongo showed spikes in minimum discharge values around the 1980s. However, locations such as Koulikoro and Mopti were found to begin an increasing trend around the 1990s. The central tendency measures of river discharge (mean and 50% percentile) also showed a decreasing trend across all the stations since 1950 (Figure 2). The median value (50% percentile) was found to decrease to near-zero values in Diola and Douna. A recovery in the mean value of river discharge could be seen in the Ansongo station after 1980. Extreme values (maximum and 90% percentile) were also considered in this study. Both indices also showed decreasing trends at all stations; however, a slight recovery could be observed after 1980 in some locations. The decreasing trend in annual minimum, mean, and maximum values of river discharge has implications for water resources in the region. The decreasing trend in minimum values suggests that water availability during the dry season will be very low. Hence, energy production through hydro-power, fish farming, and dry season irrigation-fed farming will be affected [42,43]. Broadly, the decrease in minimum discharge can be attributed to factors such as land use [44,45,46], construction of dams that regulate water flow [11,46,47], teleconnections [10,48], and drought [49,50]. Specifically, the spikes observed in Kirango and Ansongo after 1980 can be attributed to the Sahelian paradox, Sahelian recovery, and the accompanying change in land use–land cover within the region [51,52,53]. Despite the already established rainfall decrease post-1968, Sahelian stations have continued to experience increased runoff conditions promoted as a result of soil clearing and local soil crusting induced by demographic growth [54,55,56,57]. These studies and some others have shown the variability in river discharge are a result of various factors aforementioned. This could in no small way affect the services expected from the river system.

Four MLAs were deployed for the prediction of monthly river discharge time series at the locations. The predicted data were compared with the original data (Figure 4) and the associated error analysis is presented in

Table 1. Root mean square error (m${}^3$/s) from different machine learning predictions of monthly river discharge data.

Model	Koulikoro	Dioila	Kirango	Douna	Mopti	Dire	Ansongo	Niamey
RNN	1241	200	1315	193	814	421	478	454
LSTM	1044	211	953	239	517	412	412	280
TCN	1468	174	1052	245	621	445	327	291
GRU	1299	220	1171	689	948	632	674	547

. In Koulikoro, the MLAs were able to replicate the periodic nature of river discharge but not the magnitude. LSTM was the best-performing model with an RMSE of 1044 m${}^3$/s, while TCN was the worst-performing model in Koulikoro with an RMSE of 1468 m${}^3$/s. However, TCN was found to be the best-performing machine learning algorithm in predicting river discharge at Diola with an RMSE of 174 m${}^3$/s. The TCN model simulates the periodic nature of river discharge with an increasing trend. With an RMSE value of 220 m${}^3$/s, GRU was found to be the worst-performing model in Diola. In Niamey, Kirango, Dire, and Mopti, the LSTM was found to have the lowest RMSE while RNN and TCN were the best-performing models in Douna and Ansongo, respectively. In Douna, the LSTM model failed after a few steps of prediction. All models, except TCN, did not capture the low river discharge values in Ansongo. The performance of the GRU model was the worst in all locations except Kirango. Generally, the performance of LSTM was found to be the best among the four models. The low performance of these models in the long-term prediction of monthly river discharge is attributed to the chaotic nature of the time series and the influence of external drivers [10,11]. In their study, Ogunjo et al. [11] showed that the prediction of river discharge along the River Niger has a short prediction horizon due to the chaotic nature of the data. For some other studies, possibilities in the use of machine learning algorithms to predict river discharge have been investigated [12,15]. Our results differ from other studies that showed the best-performing model as multilayer perceptron when compared with DT, RF, GB, RNN-LSTM, and CNN-LSTM [15]. Invariably, the use of MLAs is effective in the prediction of flow within ungauged basins [12].

The ability to predict the annual river discharge statistics using machine learning algorithms was investigated (Figure 2, Figure 3, Figure 4, Figure 5, Figure 6, Figure 7, Figure 8, Figure 9 and Figure 10).

Table 2. Performance statistical metrics of the machine learning models at the different stations.

Statistics	Model	Koulikoro	Dioila	Kirango	Douna	Mopti	Dire	Ansongo	Niamey
Minimum	RNN	30	4	295	1	32	21	492	65
	LSTM	81	0	908	75	22	15	407	150
	TCN	61	2	106	1	35	251	410	17
	GRU	46	0.18	720	3	25	25	545	63
Q10	RNN	53	1	175	2	20	18	593	126
	LSTM	75	4	658	12	29	23	396	16
	TCN	40	7	150	6	31	234	602	29
	GRU	76	6	277	6	42	26	724	73
Q50	RNN	253	60	725	21	357	516	734	559
	LSTM	375	158	1940	18	179	1134	663	335
	TCN	187	44	642	18	218	389	403	295
	GRU	246	144	1022	33	195	425	449	272
Mean	RNN	678	31	1075	78	189	279	516	603
	LSTM	615	33	1389	97	451	258	275	314
	TCN	234	31	820	89	303	326	535	347
	GRU	123	29	866	79	289	336	1066	240
Q90	RNN	1355	87	1436	318	800	395	369	511
	LSTM	1578	72	1643	467	886	455	636	503
	TCN	1747	221	1158	260	577	495	426	496
	GRU	1619	103	1023	390	588	477	273	451
Maximum	RNN	967	252	855	1041	609	419	377	600
	LSTM	893	205	2368	492	588	438	326	474
	TCN	1898	180	3531	447	700	459	315	573
	GRU	3367	519	562	562	636	581	230	560

shows the performance statistics of the machine learning models at the different stations. As shown in Figure 2, both LSTM and TCN models underestimated the annual minimum river discharge at Koulikoro station. RNN provides better tracking of the annual minimum discharge at the location than the GRU model. Hence, RNN was found to be the best model with an RMSE value of 30 m${}^3$/s while LSTM has the worst performance with an RMSE value of 81 m${}^3$/s. At Diola, all the models could not adequately predict the annual minimum discharge due to the zero values. However, LSTM, although providing negative predictions, has the best performance with an RMSE value of 0 m${}^3$/s closely followed by the GRU model with an RMSE value of 0.18 m${}^3$/s. The RMSE values of the Kirango models were 295, 908, 106, and 720 m${}^3$/s for RNN, LSTM, TCN, and GRU, respectively. There was significant overestimation in the LSTM and GRU models after a few timesteps of prediction. The inability of TCN to adequately capture high values contributed to the high RMSE values obtained. The annual minimum river discharge was largely zero in Douna; thus, the models could not sufficiently predict the values. LSTM estimated negative values for the prediction period, while RNN and TCN estimated negative values after a few timesteps. Both RNN and TCN recorded identical performances, which is attributed to their ability to predict the trend and early timesteps of prediction. The LSTM model was found to have the best performance in predicting the annual minimum discharge at Mopti, Dire, and Ansongo, while TCN models had the best performance at Niamey.

The 10th percentile of the annual river discharge, a measure of minimum extreme, indicates low river discharge during the year. It is the minimum value below which we have extreme minimum discharge. This is an important measure whose knowledge will help in planning water resources within a region. Several factors including precipitation, land use, and teleconnection patterns can contribute to the 10th percentile of annual river discharge [10]. Figure 6 shows the 10th percentile river discharge predictions from the four machine learning algorithms. Among the machine learning algorithms, only TCN was able to capture both the trend and dynamics of the 10th percentile river discharge at Koulikoro, thus becoming the best-performing model with an RMSE value of 40 m${}^3$/s. GRU grossly underestimated the discharge at this location. The predictions by the models in Diola showed small errors in the range 1–7 m${}^3$/s. All the models, except RNN, overestimated the 10th percentile of annual river discharge at Diola, however, to different degrees. RNN had the lowest RMSE value of 1 m${}^3$/s to emerge as the best model for the location’s 10th percentile of annual river discharge. The RNN model was also found to be the best model in predicting the 10th percentile of annual river discharge at Douna, Mopti, and Dire, while the best-performing models at Kirango, Ansongo, and Niamey were TCN, LSTM, and TCN, respectively. In Dire, the RNN model was able to capture both the trend and fluctuations in the discharge data, while the TCN model predicted negative values.

From the foregoing, as observed in this study, it is not uncommon for various models to outperform each other across different basins or locations as observed in this study. Kratzert et al. [18] reported variations in the performance of different models with machine learning algorithms outperforming basic or traditional hydrological models. This is largely due to the data’s inherent strengths, weaknesses, and peculiarity. The LSTM-RNN algorithm has been found to outperform the traditional RNN and Naive Bayes algorithms in predicting low flows [30]. Although GRU has identical performance to LSTM, it was found to be better in the short-term prediction of river flow [58]. The ability to predict the minimum and 10th percentile annual discharge in a given year will help prepare the region for water scarcity and its compound effects. Farmers will be able to make irrigation plans and adjust their planting approach. Furthermore, predicting low flows aids fish spawning and catch in the basin of interest [59]. This will enable adequate mitigation plans such as pond farming and importation. Foreknowledge of the annual minimum discharge is also important for hydropower generation; it will give insight into the potentially available power during the year. In addition, it will help in planning for alternatives to river transportation during periods of low flows.

Two measures of central tendencies were considered for the annual river discharge prediction: the 50th percentile (Figure 7) and the mean value (Figure 8). The annual river discharge’s 50th percentile (Q50) represents the median value. The TCN model was the best model for Q50 (187 m${}^3$/s) prediction, while the GRU model performed best in predicting the mean value at Koulikoro. The worst performing models for both Q50 and mean value were found to be LSTM (375 m${}^3$/s) and RNN (678 m${}^3$/s). In Diola, the Q50 values showed decreasing values similar to the mean. However, the mean discharge value increased towards the end of the century. The models could not capture the decreasing trend in Q50; however, they captured the trend in mean value but not the reversal towards the end of the century. TCN was found to be the best-performing model in both Q50 with an RMSE of 44 m${}^3$/s and annual mean discharge value with an RMSE value of 31 m${}^3$/s. LSTM was the worst-performing model with RMSE values of 158 m${}^3$/s and 33 m${}^3$/s in Q50 and mean discharge, respectively. A similar result was obtained in Kirango, with TCN and LSTM as the best- and worst-performing models, respectively, for Q50 and mean discharge. TCN and LSTM have identical performance metrics (18 m${}^3$/s) for Q50 in Douna. However, the two models could only replicate the trend and not the unique patterns in the discharge. RNN and GRU have similar performance in Douna with RMSE values of 78 m${}^3$/s and 79 m${}^3$/s respectively. LSTM performed poorly in Douna with an RMSE value of 97 m${}^3$/s. In Mopti, LSTM (179 m${}^3$/s) and RNN (357 m${}^3$/s) were the best- and worst-performing models, respectively, for Q50. However, the reverse was the case in the mean discharge value with RNN showing the best metric and LSTM the worst. The TCN model was observed to have the smallest RMSE values for Q50 in Dire and Ansongo, while the LSTM model had the best performance in both stations in mean discharge values. The dynamics of both median and mean discharge values in Niamey were best captured by the GRU model. This study’s TCN model performed better than the other models in predicting Q50 values. TCN has also been better than LSTM in predicting extreme river discharge in the Yellow River [40]. The results obtained in this study were within the range 19.6–362 m${}^3$/s using GRU [20] and 8–83 m${}^3$/s using TCN [40], but higher than the range 2–6 m${}^3$/s using LSTM [19] and 0.137–0.174 m${}^3$/s using SVR [60]. It is worth noting that not much investigation has been performed regarding predicting river discharge using the TCN model. The median and mean discharge values give an insight into the river’s average discharge during the year. A low mean value suggests that the discharge might be low, while a high value might imply a higher than normal discharge. Predicting the median and mean river discharge values in different rivers will give an understanding of potentially available water resources during the year. This will enable proper planning, mitigation, adaptation, and preparation.

High river discharge extremes imply the availability of water resources. However, this causes uncertainty in water transportation and fish farming. Furthermore, high river discharge that overflows banks tends to destroy crops and that such as houses and bridges. The availability of high water volume which cannot be accommodated by a hydropower reservoir tends to be released, causing ecological damage to the surrounding environment. Machine learning prediction of the 90th percentile (Q90) and annual mean discharge are shown in Figure 9 and Figure 10, respectively. With RMSE values of 1355 m${}^3$/s and 1747 m${}^3$/s, the RNN and TCN models were the best- and worst-performing models in the prediction of Q90 in Koulikoro. However, the LSTM models gave the lowest RMSE value for the annual maximum river discharge in the same location. The GRU model was found to have the best performance for both Q90 and annual maximum discharge at Kirango and Ansongo. Similarly, TCN and RNN also showed superior performance for the two parameters in Douna and Dire, respectively. In Diola, the LSTM model outperforms the other models for Q90 but the TCN model performed best for the maximum river discharge. In Mopti, the TCN model has the best performance for Q90, but GRU has the best performance for the same quantity in Niamey. However, in both locations, the best-performing model for annual maximum river discharge was the LSTM model. Muhammad et al. [61] proposed using the GRU model for river discharge prediction due to its better performance than the LSTM model. Among the various recurrent network variants, the LSTM model has been recommended for river discharge forecasting due to its accuracy [62]. It was further confirmed that the LSTM performs better in the presence of dams and reservoirs [63,64]. Prediction of a higher than average Q90 and annual maximum river discharge suggests flooding in the given year. This will enable adequate preparation to mitigate losses and maximize the potential in the discharge.

6. Conclusions

Our study on predicting river discharge using deep learning algorithms is essential for proactive water resource planning and enhances our understanding of river behavior. Using four distinct deep learning algorithms across eight stations in the Niger Basin, we were able to identify and predict low extremes, defined as values below the 10th percentile and annual minimum. This is crucial for meteorological and hydrological drought forecast, safeguarding agricultural irrigation, aquaculture, water transportation, and hydropower generation. Additionally, our investigation into high extreme flows provides important information for anticipating and managing flood events. Extreme scenarios can lead to adverse consequences, including disease outbreaks, ecological degradation from bank overflow, and damage to vital infrastructures. The proficient performance of these algorithms yielded reliable predictions and established their suitability for the unique challenges presented by the region. Ultimately, our study not only advances our comprehension of river dynamics but also equips decision-makers with tools to address the complexities associated with low and high extreme flows in the Niger Basin.

Secondly, this study has shown the possibility of the use of deep learning algorithms in river flow (low and high) prediction within the River Niger basin. Even though this prediction cannot be made possible on a long-term basis due to the chaotic nature of the river system, deep learning algorithms can still help in short-term predictions. Across the stations studied, the study affirmed that no one-size-fits-all model can be used across the stations within the study area. The performances of various models were specific to some stations and peculiar for some flow levels. In essence, river flow prediction within the River Niger system is station-specific and should thus be studied as such. Peculiarities across each station become evident based on the station’s location in terms of distance from the coast and nearness to the Sahel.

In conclusion, this exhaustive examination of temporal variations in annual river discharge, conducted across diverse stations within the River Niger basin, furnishes invaluable insights into the dynamic comportment of this pivotal hydrological system. The temporal scope of the study spans from 1950 to 1990, elucidating discernible trends in minimum, mean, and maximum discharge values, thereby engendering implications for water resource management within the region. The observed diminishing trajectory in minimum discharge values, notably post-1960, posits a formidable challenge to water availability during the dry season, with profound repercussions for critical sectors such as hydro-power generation, fish farming, and dry season irrigation-fed agriculture. The multifaceted nature of these trends finds its roots in various contributory elements, encompassing alterations in land use, dam infrastructure development, teleconnections, and drought occurrences. This multifarious array of factors underscores the intricate responsiveness of the river system to an amalgamation of driving forces. Introducing a predictive dimension to the analysis, the deployment of deep learning algorithms (DLAs) for monthly river discharge prediction elucidates notable distinctions in performance across locations, with Long Short-Term Memory (LSTM) emerging as the preeminent model. However, it is imperative to acknowledge the location-specific variations in DLA efficacy, accentuating the necessity for bespoke methodologies tailored to the distinctive attributes of individual basins. The inherently chaotic nature of time series data, coupled with the influence of extraneous drivers, imposes limitations on the protracted predictive capabilities of DLAs—a pivotal consideration in the context of judicious water resource planning. Furthermore, this study underscores the critical significance of prognosticating the 10th percentile of annual river discharge, particularly in ameliorating the impacts of meteorological and hydrological droughts. The discerned aptitude of DLAs, particularly Temporal Convolutional Networks (TCNs), in furnishing cogent predictions for both low- and high-extreme discharge events augments their utility as indispensable tools for proactive planning and mitigation strategies. In summary, we advocate for the imperative undertaking of location-specific inquiries within the River Niger basin. Recognizing the idiosyncratic characteristics and influences operative at each station is integral to the formulation of an augmented and integrated river management paradigm. Key considerations in this regard encompass latitude, land use patterns, teleconnections, and global environmental dynamics.