UC Irvine
UC Irvine Previously Published Works
Title
Developing reservoir monthly inflow forecasts using artificial intelligence and climate
phenomenon information
Permalink
https://escholarship.org/uc/item/2zp896xn
Journal
Water Resources Research, 53(4)
ISSN
0043-1397
Authors
Yang, T
Asanjan, AA
Welles, E
et al.
Publication Date
2017-04-01
DOI
10.1002/2017WR020482
Copyright Information
This work is made available under the terms of a Creative Commons Attribution License,
availalbe at https://creativecommons.org/licenses/by/4.0/
Peer reviewed
eScholarship.org Powered by the California Digital Library
University of California
RESEARCH ARTICLE
10.1002/2017WR020482
Developing reservoir monthly inflow forecasts using artificial
intelligence and climate phenomenon information
Tiantian Yang
1,2
, Ata Akbari Asanjan
1
, Edwin Welles
2
, Xiaogang Gao
1
,
Soroosh Sorooshian
1
, and Xiaomang Liu
3
1
Department of Civil and Environmental Engineering, Center for Hydrometeorology and Remote Sensing [CHRS],
University of California-Irvine, Irvine, California, USA,
2
Deltares USA Inc., Silver Spring, Maryland, USA,
3
Key Laboratory of
Water Cycle and Related Land Surface Processes, Institute of Geographic Sciences and Natural Resources Research,
Chinese Academy of Sciences, Beijing, China
Abstract Reservoirs are fundamental human-built infrastructures that collect, store, and deliver fresh surface
water in a timely manner for many purposes. Efficient reservoir operation requires policy makers and operators
to understand how reservoir inflows are changing under different hydrological and climatic conditions to enable
forecast-informed operations. Over the last decade, the uses of Artificial Intelligence and Data Mining [AI & DM]
techniques in assisting reservoir streamflow subseasonal to seasonal forecasts have been increasing. In this
study, Random Forest [RF), Artificial Neural Network (ANN), and Support Vector Regression (SVR) are employed
and compared with respect to their capabilities for predicting 1 month-ahead reservoir inflows for two headwa-
ter reservoirs in USA and China. Both current and lagged hydrological information and 17 known climate phe-
nomenon indices, i.e., PDO and ENSO, etc., are selected as predictors for simulating reservoir inflows. Results
show (1) three methods are capable of providing monthly reservoir inflows with satisfactory statistics; (2) the
results obtained by Random Forest have the best statistical performances compared with the other two meth-
ods; (3) another advantage of Random Forest algorithm is its capability of interpreting raw model inputs; (4) cli-
mate phenomenon indices are useful in assisting monthly or seasonal forecasts of reservoir inflow; and (5)
different climate conditions are autocorrelated with up to several months, and the climatic information and their
lags are cross correlated with local hydrological conditions in our case studies.
1. Introduction
Reservoirs are the vital human-built infrastructures that collect, store, and deliver fresh surface water for many
purposes in a timely manner. Efficient reservoir management is capable of providing society with resilience
against hydrological extremes, water-supply sustainability, flood protection for urban areas, and clean, renew-
able power production. Over the past century, much effort has been made by dam operators, policy makers,
and scientists to better understand reservoir operations, and develop optimal reservoir operation strategies.
According to CDWR [2014a] and [2014b], the primary focus of reservoir operations in developed regions, such
as California, is how to validate the operation strategies of existing facilities based on hydrological information,
and improve subseasonal to seasonal forecasts in mitigating changing climatic conditions at different tempo-
ral scales, i.e., real-time, subseasonal to seasonal, and single to multiple years [CDWR, 2014a, 2014b]. For exam-
ple, it is widely acknowledged that climate conditions can significantly impact water supply and many climate
phenomenon indices can be used as predictors in supporting water resources management [Pagano and
Garen, 2003; Montoya et al., 2014; Kalra et al., 2013; Guti
errez and Dracup, 2001; Garen, 1993]. The research
question for the reservoir systems in developed countries is how to utilize different types of auxiliary informa-
tion to support reservoir management and to develop forecast-informed operations for existing facilities.
However, in developing countries, such as China, besides the requirements stated above for existing reser-
voirs, many new reservoirs are under construction or being planned and a number of water diversion projects
have recently begun, i.e., the center route of the China’s South-to-North Water Diversion Project. Advanced
modeling and decision support tools, such as the AI & DM techniques, are therefore needed in developing
countries for efficient management and operation purposes.
In recent years, with the advances in computer sciences, the AI & DM techniques have become more and
more popular in the field of streamflow forecasts, reservoir operation planning and scheduling [Hejazi and
Key Points:
Artificial intelligence and data mining
(AI&DM) techniques are powerful
regression tools in developing
reservoir monthly inflow forecasts
Climate phenomenon indices have a
complex relationship with
hydrological conditions, and provide
useful information for reservoir
operations
Different AI & DM techniques have
strengths and limitations and are
suggested to use with proper
parameterization and prior
examination
Supporting Information:
Supporting Information S1
Correspondence to:
T. Yang,
tiantiay@uci.edu
Citation:
Yang, T., A. A. Asanjan, E. Welles,
X. Gao, S. Sorooshian, and
X. Liu (2017), Developing reservoir
monthly inflow forecasts using artificial
intelligence and climate phenomenon
information, Water Resour. Res., 53,
2786–2812, doi:10.1002/
2017WR020482.
Received 25 JAN 2017
Accepted 6 MAR 2017
Accepted article online 9 MAR 2017
Published online 7 APR 2017
Corrected 15 MAY 2017
V
C
2017. American Geophysical Union.
All Rights Reserved.
YANG ET AL. RESERVOIR INFLOW FORECASTS 2786
Water Resources Research
PUBLICATIONS
Cai, 2009]. Among all kinds of Artificial Intelligence & Data Mining (AI & DM) techniques, the Artificial Neural
Networks (ANN), the Decision Tree (DT), and Support Vector Machine or Regression (SVM or SVR) methods
are three of the most popular techniques in developing streamflow forecasts at different temporal scales
around the world [Schnier and Cai, 2014; Zealand et al., 1999; Yaseen et al., 2015; Cheng et al., 2015; Kumar
et al., 2013; Erdal and Karakurt, 2013; Maity et al ., 2010; Guo et al., 2011; Asefa et al., 2006].
The Artificial Neural Network is a robust, nonlinear machine learning approach, which has been extensively
applied for many classification and regression problems in various fields. In the field of streamflow and res-
ervoir inflow forecasts, Cheng et al. [2015] compared ANN and SVM in forecasting monthly inflow of the Xin-
fengjiang Reservoir in China and proved both methods have satisfactory performances. Thirumalaiah and
Deo [1998] used ANN in real-time forecasting of water levels based on upstream gauging station informa-
tion and historical records in a river system in Jagdalpur in India. Lima et al. [2016] used a simple ANN to
incorporate newly arrived meteorological data and to produce daily forecasts for two small watersheds in
British Columbia, Canada. Wu et al. [2009] used both SVR and ANN to predict the streamflow timeseries for
two river outlets located in China, and they examined the predictive skills for different lead times, including
1, 3, 6, and 12 month-ahead. Wang et al. [2006] compared different hybrid ANNs with regard to their capa-
bility of streamflow prediction at the headwater region of the Yellow River, China, at a daily scale. Linares-
Rodriguez et al. [2015] demonstrated the flexibility of ANNs on adding additional runoff indices to enhance
1 day-ahead streamflow forecast in the Northeast Guadalquivir basin in southern Spain. Jain et al. [1999]
presented the usefulness of ANNs not only in reservoir inflow prediction, but also in the optimal reservoir
scheduling and management of a diversion reservoir located in the Godavari basin and Mahanadi basin of
India. Ashaary et al. [2015] also demonstrated the application of ANN in developing the short-term forecast-
ing model for the change in the reservoir water level in the Timah Tasoh Reservoir located in Northern Pen-
insular Malaysia. A recent summary of using ANN and other AI methods in hydrological applications and
forecasting can be found in Yaseen et al. [2015].
Similarly, there are also extensive studies and applications of using DT methods to assist streamflow fore-
casts. Erdal and Karakurt [2013] compared SVR with a DT algorithm in predicting monthly streamflow in
the C¸oruh River in the Eastern Black Sea Region in Turkey and concluded that DT methods were able to
produce better results than SVR. Kumar et al. [2013] tested the performances of MLR, ANN, fuzzy logic,
and DT algorithms in predicting streamflow at an upstream reservoir in the Sutlej Basin in northern India.
They determined the DT methods performed well when compared to other methods. Galelli and Castelletti
[2013] assessed the predictive performances of multiple DT methods and ANN in forecasting streamflow
of the Marina catchment in Singapore and the Canning River in Western Australia. In addition, Galelli and
Castelletti [2013] demonstrated the DT method is superior over ANN due to its nonparametric characteris-
tics, which makes DT methods suitable for large computationally intensive problems. Wei [2012] compared
two popular DT algorithms (C5.0 and CART) in predicting reservoir releases in northern Taiwan during
typhoon events and concluded that the DT methods are skillful in discharge simulation. Cheng et al.
[2008] used a DT approach as a predictive model to determine the optimal reservoir releases before the
onset of typhoons at the Shihmen Reservoir System in Taiwan and justified its capability in assisting
streamflow prediction during flood conditions. In a more recent study, Yang et al. [2016] compared a stan-
dard DT algorithm (CART) with a Random Forest algorithm in predicting the daily reservoir discharges for
nine different river basins in California and tested the suitability of DT methods for a generalized dis-
charge simulation problem.
Besides ANN and DT methods, the Support Vector Machine (SVM) is another popular method in
streamflow forecasts. Asefa et al. [2006] demonstrated a case study in Sevier River Basin in Utah, USA,
and produced a promising streamflow prediction results at both seasonal and hourly temporal scales.
Maity et al. [2010] applied a Support Vector Regression (SVR) method to predict monthly streamflow in
the Mahanadi River Basin in the State of Orissa, India, and showed the superior performance of SVR
over autocorrelation regression. Lin et al. [2006] presented an enhanced SVR model to predict long-
term reservoir discharges from the Manwan Hydropower Reservoir in China. Guo et al. [2011] devel-
oped an adaptive SVR model to conduct a monthly streamflow prediction on the Three Gorges Area
in the Yangtze River basin in China. Guo et al. [2011] also concluded an SVR model is capable of pro-
ducing accurate predictions of streamflow and the SVR model has good generalization characteristics
in solving streamflow prediction problems.
Water Resources Research 10.1002/2017WR020482
YANG ET AL. RESERVOIR INFLOW FORECASTS 2787
As a summary of the uses of AI & DM techniques in developing streamflow forecasts as listed above, the
ANN, DT, and SVR are all proven to be powerful tools in predicting streamflow and reservoir inflows. Howev-
er, the investigation on different parameterization among those models and comprehensive comparison
study are rarely reported, which impedes the practical uses of these methods in water resources manage-
ment and planning. As communicated with the Snow Survey Office from the California Department of
Water Resources, many operational streamflow regression models are not as complex as AI & DM methods.
The decades-long experiences from many hydrologists and engineers have been instrumental in determin-
ing the regression coefficients of the operational models. The AI & DM tools provide the mechanism for
enhancing the existing hydrometerological forecasts for reservoir management, given the fact that more
and more types of data have become available (so called ‘‘Big Data Era’’) and the increasing possibilities of
taking advantages of those auxiliary information to support reservoir operations. This comparison study
aims to provide a baseline, and test the applicability and robustness of different AI & DM tools in support of
reservoir operations and hydrological forecasts. As a part of our development process, we conducted a gen-
eralized comparison of AI & DM methods, including (1) a benchmark three-layer feed-forward Artificial Neu-
ral Network, (2) the Random Forests method, and (3) a Support Vector Regression technique, which are all
commonly used approaches in the literature.
To test the robustness of different AI & DM methods, we selected two headwater reservoirs in USA and Chi-
na, and conducted a comparison experiment using different AI & DM methods with various parameteriza-
tions to simulate reservoir inflows. In USA, the Trinity Lake, also known as the Clair Engle Lake (CLE), is
selected. The CLE reservoir is one of the water supply sources in the Central Valley Project (CVP) a federally
funded water distribution project. The goal of CVP is to divert water from the northern part of California,
where water supply is relatively abundant, to the water-scarce areas in the central and southern parts of the
state for irrigation and municipal water supply purposes. The CVP is jointly operated in coordination with
the California State Water Project (SWP). SWP is operated by the California Department of Water Resources,
and directly transports water from the northern parts to the southern parts of California for residential water
uses. In China, we selected the Danjiangkou (DJK) Reservoir as another study case. The DJK reservoir is the
headwater reservoir for the central route of the China’s South-to-North Water Diversion Project which also
transports water from the water-abundant, in this case southern parts (Han River and Yangtze River), to the
water-limited northern areas of China, including the highly populated areas of Beijing and Tianjin, and
Henan and Hebei provinces.
In addition, another focus of this paper is to test and quantify the predictability of different hydrological
and climatic information for reservoir inflow forecasts. For example, in order to develop better reservoir
inflow subseasonal and seasonal forecasts, it may be possible to identify the climate phenomenon that
dominate the local hydrology, and then to incorporate climate phenomenon indicators into a given model-
ling framework. As pointed out by Burley et al. [2012] and Turner and Galelli [2016], the shifts in climate con-
ditions and streamflow should be emphasized in future studies on water resources management and
planning. According to many other studies [Montoya et al., 2014; Kalra et al., 2013; Guti
errez and Dracup,
2001; Garen, 1993; Hamlet and Lettenmaier, 1999], certain climate phenomenon variables or indices, such as
the ENSO and PDO, are potentially useful in supporting water supply planning in the western regions of
United State, and these indices are predictable with lead times up to 6 months or 1 year. However, there
are two key steps necessary to make operational use of climate indices for reservoir operations. The first
one is how to incorporate the signals of known climate phenomenon indices into subseasonal and seasonal
prediction modeling framework. Demonstrating methods to use raw climate phenomenon indices directly
in regression models of streamflow makes this potentially valuable information available to decision makers
and dam operators. The authors believe the AI & DM techniques are suitable tools to address the issue of
using climate phenomenon indices in regression models thus assisting water-supply planning and predic-
tion. The second issue is how to automatically identify and select the climate phenomena indices as predic-
tors for supporting reservoir planning and scheduling, given the fact that it is unclear to decision makers
which climate phenomenon indices are effective predictors, and, hence, representative for the regional cli-
mate variability. Many previous studies were conducted on using one or two climate phenomenon indices,
which are already known as useful predictors in a specific region. Inspired by a recent study by Yang et al.
[2016], in which many decision variables of operating reservoirs are automatically ranked and selected using
a Gini diversity index, the climate phenomenon indices can also be ranked and compared with regard to
Water Resources Research 10.1002/2017WR020482
YANG ET AL. RESERVOIR INFLOW FORECASTS 2788
their predictability for reservoir operation and seasonal forecasts. In this study, we focus on developing 1
month-ahead inflow forecasts.
In a summary, the goals of this study are (1) to apply AI & DM methods to reservoir inflow predictions and
investigate the signals of climate phenomenon indices in two headwater reservoirs located in the western
regions of the United States, and the southern part of China, respectively; (2) to compare the predictive per-
formances of three popular AI & DM methods, namely RF, ANN, and SVR, in assisting subseasonal to season-
al water-supply planning; and (3) to explore the sensitivity analysis of different AI & DM methods and
inform decision makers regarding the usefulness and capability of various regression approaches. Specifi-
cally, the sensitivity analysis of AI & DM methods includes (i) the default stopping criterion in ANN, i.e., maxi-
mum iteration, and the number of hidden nodes as one of the structural parameters; (ii) the maximum
features in developing a Random Forest, and (iii) different kernel functions and associated penalty terms in
the SVR models. The methodology and approaches employed in the current work are universally applicable
to other study cases and are not limited to the case studies and sites in the current work, which are the
source reservoirs for the U.S. federal Central Valley Project, and the China’s South-to-North Water Diversion
Project, respectively.
The organization of rest of the paper is as follows: The methodologies of ANN, DT, and SVR are summarized
in section 2; section 3 introduces the two study cites, reservoir operation data, climate phenomenon indices,
and model inputs; The results are given in section 4; section 5 provides the discussion with regard to meth-
odologies, parameterization, seasonal patterns, and the implications of using climate phenomenon indices;
Finally, the major findings, conclusions, and future works are summarized in section 6.
2. Methodology
2.1. Artificial Neural Network (ANN)
The ANN is a powerful classification and regression algorithm that is inspired by the neurological structure
of the human brain [Jain et al., 1996; Hopfield, 1988]. The concept in ANN is to interconnect input data with
output data using multiple neurons as hidden layers, which are able to extract the explicit information and
relationship between input and output data. ANN is extensively used in many fields of study, such as biolog-
ical memory [Kohonen, 2012], pattern recognition [Jain et al., 2000], image processing [Egmont-Petersen
et al., 2002; Cichocki and Amari, 2002], precipitation estimation from satellites [Hsu et al., 1997], ecological
modeling [Lek and Gu
egan, 1999], and reservoir operation [Cheng et al., 2015; Jain et al., 1999; Ashaary et al.,
2015; Shamim et al., 2016; Coulibaly et al., 2001]. In this study, a Three-Layer Feed-Forward Neural Network
(TLFFNN) is used in combination with a backpropagation learning algorithm [Werbos, 1974] (Figure 1).
A typical TLFFNN consists of an input layer, a hidden layer, and an output layer. The inputs data
~
x
x
1
; x
2
; ...; x
n
0
ðÞand output data
~
zz
1
; z
2
; ...; z
n
2
ðÞare connected by a hidden layer
~
hh
1
; h
2
; ...; h
n
1
ðÞ, where
n
0
, n
1
and n
2
represent the total
number of inputs, hidden neu-
rons, and outputs. The inputs
are connected to the hidden
layer by a transfer function (f ),
which is typically based on the
weighted sum of inputs as
shown in equation (1).
h
j
5f
X
n
0
i51
w
ij
x
i
1w
0j
(1)
where h
j
is the j-thneuroninthe
hidden layer; x
i
is the i-th input;
w
ij
represents the weight as-
signed to the i-th input in order to
calculate the j-th hidden neuron;
and j 2 1; 2; ...n
1
ðÞand i 2
1; 2; ...n
0
ðÞ. Similarly, another
Figure 1. A three-layer feed-forward neural network (TLFFNN)
Water Resources Research 10.1002/2017WR020482
YANG ET AL. RESERVOIR INFLOW FORECASTS 2789