Page 1 of 15

Transactions on Machine Learning and Artificial Intelligence - Vol. 10, No. 5

Publication Date: October, 25, 2022

DOI:10.14738/tmlai.105.13049. Karimanzira, D., Ritzau, L., & Emde, K. (2022) Catchment Area Multi-Streamflow Multiple Hours Ahead Forecast Based on Deep

Learning. Transactions on Machine Learning and Artificial Intelligence, 10(5). 15-29.

Services for Science and Education – United Kingdom

Catchment Area Multi-Streamflow Multiple Hours Ahead Forecast

Based on Deep Learning

Divas Karimanzira

Fraunhofer Institute of Optronics

System Technologies and Image Exploitation (IOSB)

Am Vogelherd 90, 98693 Ilmenau, Germany

Linda Ritzau

Fraunhofer Institute of Optronics

System Technologies and Image Exploitation (IOSB)

Am Vogelherd 90, 98693 Ilmenau, Germany

Katharina Emde

School 1Fraunhofer Institute of Optronic

System Technologies and Image Exploitation (IOSB)

Am Vogelherd 90, 98693 Karlsruhe, Germany

ABSTRACT

Modeling of rainfall-runoff is very critical for flood prediction studies in decision

making for disaster management. Deep learning methods have proven to be very

useful in hydrological prediction. To increase their acceptance in the hydrological

community, they must be physic-informed and show some interpretability. They

are several ways this can be achieved e.g. by learning from a fully-trained

hydrological model which assumes the availability of the hydrological model or to

use physic-informed data. In this work we developed a Graph Attention Network

(GAT) with learnable Adjacency Matrix coupled with a Bi-directional Gated

Temporal Convolutional Neural Network (2DGAT-BiLSTM). Physic-informed data

with spatial information from Digital Elevation Model and geographical data is used

to train it. Besides, precipitation, evapotranspiration and discharge, the model

utilizes the catchment area characteristic information, such as instantaneous slope,

soil type, drainage area etc. The method is compared to two different current

developments in deep learning structures for streamflow prediction, which also

utilize all the spatial and temporal information in an integrated way. One, namely

Graph Neural Rainfall-Runoff Models (GNRRM) uses timeseries prediction on each

node and a Graph Neural Network (GNN) to route the information to the target node

and another one called STA-LSTM is based on Spatial and temporal Attention

Mechanism and Long Short Term Memory (LSTM) for prediction. The different

methods were compared in their performance in predicting the flow at several

points of a pilot catchment area. With an average prediction NSE and KGE of 0.995

and 0.981, respectively for 2DGAT-BiLSTM, it could be shown that graph attention

mechanism and learning the adjacency matrix for spatial information can boost the

model performance and robustness, and bring interpretability and with the

inclusion of domain knowledge the acceptance of the models.

Page 2 of 15

16

Transactions on Machine Learning and Artificial Intelligence (TMLAI) Vol 10, Issue 5, October - 2022

Services for Science and Education – United Kingdom

Keywords: Deep learning; Streamflow prediction; Graph Attention Neural Networks,

Sequential model. Rainfall-Runoff Modelling.

INTRODUCTION

Among all extreme events such as droughts, fires, etc., flood is the most common and cause a

lot of deaths according to the UN [1,2]. Therefore, rainfall-runoff studies are very important in

several respects. Thy can be used to simulate flow peaking, changes in the water quantity and

quality and helps in decision on water resources management. Due to the nature of hydrological

processes, interconnectivity between several influencing variables, nonlinearity, etc, rainfall- runoff modelling has many challenges. Traditionally, physically-based or numerical

hydrological models are used for flood prediction, e.g. in [3], fully dynamic, diffusive, and

kinematic waves models are used to describe the surface water flow and produces very good

results in flow and water level prediction. Although the performance of the physically-based

models in terms of prediction are excellent, these models require various types of hydrological

and geomorphological observations. Their setup and operation are very time-expensive, which

makes short-term prediction required, not feasible [4]. Furthermore, developing a physically- based model requires in-depth domain knowledge and expertise on hydrological parameters

and model region, which is very challenging. Physically-based models postulates inserting

maltitudes of different types of hydrological data for learning, but in fact, it is known that there

are too many influencing factors of floods which are difficult to fully simulate them. Therefore,

Cea et al,[5] and others, suggest to use as less variables as possible to increase the

transferability of the models. For example Cea et al, [5], pointed out the correlation between

rainfall and runoff.

In recent years, data-driven and deep learning models are becoming the point of interest in

hydrological studies [6]. Mostly, LSTM (Long Short-Term Memory), TCN (Temporal

Convolutional Networks) are used as rainfall-runoff modelling is taken as a timeseries problem

[7,8,9,10].

If geospatial information such as topography and other watershed characteristics considering

the upstream-downstream relationships are integrated into deep learning, a combination of

Graph Neural Network (GNN) or Convolutional Neural Networks and LSTMs are used as in

[11],[12]. In this case, CNNs are used to extract and encode spatial information at each time

step which is then fed into an LSTM module for the time-series modeling and prediction. Due

to the nature of hydrological data that it is non-euclidean, using CNNs might not be physically

correct. The CNNs have convolution kernels with fixed size filter which assumes Euclidean data

and takes all the local neighbors of each node into consideration. In hydrology the fact that

streams or hillslopes are regionally close to one another does not mean that they are

hydrologically related. A more physically appealing solution is to use GNNs which capture the

upstream-downstream dependency graphically [13].

There are two different current developments in deep learning structures for streamflow

prediction. One from the authors [14] Graph Neural Rainfall-Runoff Models (GNRRM) assumes

raster information and uses timeseries prediction models on each grid cell and Graph Neural

Networks to aggregate grid cells with similar distance from the target node to compute its

streamflow. Through the GNN it is non-euclidean as it utilizes spatial information. The author

showed that for the sequence models in the nodes to model temporal behavior, BiGTCN