Abstract

The need for, and the process of, optimizing real time reservoir operations have attracted substantial research attention. Among these is the employment of artificial neural network (ANN), singly or with supporting algorithms, for real time multi-objective reservoir operation optimization. Using Systematic Literature Review (SLR), this paper reviews 66 studies, comprising of studies that employed ANN with or without another training algorithm(s), and those that employed evolutionary algorithm (EA) of any type for real time reservoir operations optimization. From this, it highlights the necessity of using ANN and the suitability of EA as a training algorithm. This paper, from the meta-analysis of the studies reviewed, shows that, though ANN is primarily suitable for real time forecasting, the best network architecture is the real-time recurrent learning (RTRL) neural network algorithm. And, ANN supported with supervised or unsupervised learning algorithm, has better performance potential than those singly used. Also, evolutionary algorithms are presented as viable supporting training algorithms capable of extrapolating data of deeper abstraction, complex uncertainty, with consistent predictive capacities.

1. Introduction

The need to manage water resources variations, especially in reservoirs, using hydrological structures is inescapable. Reservoirs, defined as formed or modified water bodies for specific purposes ¹, always demand operations that are usually executed for the accomplishment of its primary purposes ^{2, 3}. These include water supply for domestic and industrial use, power generation, agricultural irrigation, flood control, recreation, and waste disposal, in few instances ^{1, 4}. The reservoir operations are, therefore, generally managerial and technical activities that are carried out, daily or periodically, in the handling of reservoirs, and most importantly, in achieving its purposes ^{3, 5}. These reservoir operations differ and correspond to the design purpose(s). For instance, the details of the operational activities involved such as planning, design, implementation and monitoring, and the associated data needed for decision making in the operation of hydropower generating reservoir is different from that of irrigation-aiding. Multi-objective reservoirs are, however, reservoirs designed for more than one objective ^{2, 4}. On another end, real time reservoir operations are operations whose variables (water discharge, sediment discharge, rainfall runoff, ground water flow, precipitation and water quality) are necessarily captured in real time for accurate decision making ⁶.

The need and process of optimizing real time reservoir operations have attracted substantial research attention, and numerous works have been published on both the process-driven and data-driven approaches. Process-driven optimization process, like reservoir operation rule curve ^{4, 7, 8}, deals with enhancement of the regimented approach in the long-term operation guideline and policy. The data-driven, on the other hand, optimizes the reservoir operations based on the associated data variables with the aid of computational models ^{9, 10, 11}. The data-driven approach has been the most challenging due to event uncertainties, data varieties and the velocity of its changing and update ^{12, 13}. Nonetheless, designing and implementing computational approaches to optimizing real-time reservoir operations have been largely data-driven and more advantageous over the process-driven approaches because of their abilities of strong linear mapping, uncertainty handling, and learnability ¹⁴.

Artificial Neural Network (ANN), as later shown in this work, is observed as the most prominently-used computational model in the data driven approach to optimizing reservoir operations. ANN, which has been applied in diverse fields, such as medicine, finance, management, engineering, among others ^{2, 4}, is also attracting considerable interest in reservoir operation optimization, with record of substantial studies published in this regard ^{8, 12, 15, 16}. In the same vein, several studies have been conducted on the need to enhance ANN and continuously support its workability with other algorithms ^{17, 18}, which, in our case, is evolutionary algorithm (EA).

EA, consisting of different heuristics, solves optimization tasks through the emulation of certain aspects of natural evolutions. Similar to ANN, EA uses different levels of data abstraction but always works on whole population in proffering solutions to a given task ^{19, 20}. It seamlessly learns additional problem details, and uses this for its performance improvement ²¹. EA can also be applied to a wide range of domain problems, and can optimize the target function. These abilities are the considerable factors that endeared the usage of EA as supporting algorithm in computational optimization methods ^{20, 22}.

This paper, using a Systematic Literature Review (SLR) approach, presents the necessity and viability of a novel computational model of ANN, and EA as a suitable supporting algorithm in training ANN for multi-objective framework in real time reservoir operations optimization. This work, in sum, identifies the best ANN architecture for real time forecasting, and proposes that ANN can efficiently and effectively be trained with EA for optimized real time reservoir operations. The next section of this paper describes SLR, which is the research methodology employed in this study. The third section presents the findings, which serves as the answers to this study’s research questions, and the fourth section presents the conclusion.

2. Methodology: Systematic Literature Review (SLR)

The research questions posed by this study are (a) what is the best ANN’s architecture for real time forecasting, and (b) is EA a supporting algorithm suitable for the training of ANN for a multi-objective framework of real time reservoir operation optimization? In answering these questions, SLR, being a structured literature review approach, is employed.

SLR is described as a structured method to identify, evaluate and interpret available information, in view of engaging a research topic, and answering research questions in a preliminary study ^{23, 24}. It presents a systematic way of summarizing empirical evidences for meta-analysis. By defining a review protocol and specifying the topic to be researched, SLR presents a customized approach to literature review that leads to valid and reliable findings. The three main phases in SLR, which are also adapted in this study, are (a) planning the review, (b) conducting the review and (c) reporting the review ²⁴. Figure 1 depicts the SLR research design.

The researcher outlines the necessary requirements for the collation of broad and objective information, which will be critiqued to provide insights to the research topic, or answer the research questions. In this study, Google Scholar was chosen as the indexed literature repository because it supports open access and presents vast amount of literature. The search strings used are (a) computational methods for real time reservoir operation optimization, and (b) evolutionary algorithms for real time reservoir operation optimization. These are used in identifying primary sources of information that exclusively treat the subject and suitable for meta-analysis in order to achieve an unbiased result.

From the search, a total number of 131 published articles were gathered from the order of the results presented by the Google Scholar Search Engine. The conceptual articles on (real time) reservoir operations and articles on computational methods in reservoir operations, but with no case study, modelling and simulations, were then excluded. This is to provide, most importantly, attention to articles that have implemented any computational model for reservoir operation optimization. After the sorting and filtering, a total of 66 articles were left. These articles were then categorized into two, with each article openly coded for identification and tracking. The categories are: (a) 52 articles that employed ANN, singly or supported with any other computational model, and (b) 14 articles that employed any other computational model aside ANN.

Ahmed and Naomie ²³’s Population-Intervention-Comparison-Outcome-Context (PICOC) format is used in conducting the review. The articles’ general theme, as shown by the search strings used, is described by the population. The intervention is the converging point for each of the sub-theme as related to the research questions. The comparison is the contextual comparison of the articles, as related with their respective meta-analysis. Lastly, the outcomes are the final findings drawn from the review and the conclusions reached. The PICOC review method, as it relates with this study, is presented in Table 1.

The review is reported step-by-step to highlight results in view of answering the earlier posed research questions. These findings are presented in the next section of this paper.

The articles were reviewed on one-by-one basis by identifying the overall issue attended to, the objectives and the results. Table 2 and 3 present summaries of the articles reviewed under the 2 categories as earlier mentioned.

From Table 2 presented, ANN had been applied to many different aspects of reservoir operation management to execute varying tasks, but with a unifying purpose of predictive analyses. Hydropower performance ^{5, 25, 28} among others, river sediment estimation ^{26, 37} among others, water inflow and level and flood forecasting ^{17, 18, 29, 33, 34, 35} among others, marsh land restoration ¹² and reservoir control system ^{27, 32, 36} among others are the reservoir operations identified from the review. Studies that are not task-based experimented and evaluated the performances of different ANN models ⁵⁹, optimized traditional ANN model ^{56, 57}, and compared ANN models with traditional linear models ³.

Out of the 52 studies reviewed above, 27 employed ANN with supporting algorithms in training the data, while 25 solely employed ANN. The supporting algorithms/models recorded are linear regression, Monte Carlo, particle swarm optimization (PSO), non-linear programming model (NLP), genetic algorithm, fuzzy logic, Levenberg-Marquardt, dynamic programming (DP), conjugate gradient (CG), stochastic dynamic programming (SDP), stopped training approach (STA), support vector machine (SVM), Monte Carlo statistical blockade (SB), multivariable linear regression (MLR) model, and implicit stochastic optimization (ISO). Although ANN is principally suitable for real time data-driven learning model, studies that explicitly demonstrated the capacity employed feed forward ^{17, 50}, and recurrent network ^{60, 61} as their network architectures. Notably, studies that employed recurrent network ^{60, 61} as their network architectures employed ANN for multi-objective framework and provided a better performance result. Table 3 presents the summary of articles that employed other computational model aside ANN.

Studies on optimization of reservoir operation, as shown in Table 3, are on concrete dam crack detection ⁶², optimal design of water dam ^{63, 72}, daily flow and flood forecasting ^{64, 66}, hydropower performance ^{8, 65, 67}, and reservoir control system ⁷. The computational models used by these studies are dynamic programming (DP), improved sampling stochastic DP (ISSDP), stochastic dynamic programming (SDP), genetic algorithm-simulated annealing (IGA-SA), dynamic programming successive approximation (DPSA) algorithm, particle swarm optimization (IPSO) and artificial Bee Colony (ABC). In these cases, studies on real time and/or multi-objective optimization ^{65, 69} employed PSO, GA and SA. From the summaries presented in Table 2 and 3, and the accompanying reviews, the research questions earlier posed by this study are therefore answered.

a) What is the best ANN’s network architecture for real time forecasting?

The best ANN’s network architecture for real time forecasting is the real-time recurrent learning (RTRL) neural network algorithm, which was firstly proposed by Williams and Zipser ⁷³. The comparative studies conducted by Hsu, Huang ⁶¹ (review presented in Table 2; Open tag 52) reported that RTRL performed better, among others, for multi-phase intelligent real-time reservoir operation model.

b) Is EA a supporting algorithm suitable for the training of ANN for a multi-objective framework of real time reservoir operation optimization?

EA is suitable as a supporting algorithm for the training of ANN as a multi-objective framework for real time reservoir operation optimization. This is so, according to the reviews presented in Table 2 and Table 3, based on the following. First, ANN is a computational model that works better with supporting algorithm for the training of its input data nodes for supervised and unsupervised learnings. From the review presented in Table 2, 27 out of the 52 studies presented employed supporting algorithm, and the findings reported comparatively showed that studies that employed supporting algorithms dealt with data of deeper abstraction, complex uncertain events, and produce consistent predictive capacities. Second, particle swarm optimization (PSO) and genetic algorithm which are leading EA algorithms, are recorded in 12 studies, either supporting ANN or another computational model for real-time optimization of reservoir operation. Third, studies that employed recurrent network ^{60, 61}, acclaimed as the best ANN architecture, used it for multi-objective framework. There is also record of studies ^{65, 69} (in Table 3) that employed an EA algorithm for multi-objective framework.

3. Conclusion

A multi-objective optimization framework must therefore be able to attend to more than one, if not all, of the reservoir operation objectives which can be mainly classified into hydropower performance, river sediment estimation, water inflow and level and flood forecasting, marsh land restoration, and reservoir control system. In each of these objectives, the data variables and the weighing parameters to be modelled for the respective computational purpose is different, and this would inform the choice of the computational model and any of the supporting algorithms. The choice of the computational model is basically determined by the details of the reservoir operation policy intended to be modelled.

In a real time, multi-objective optimization framework for reservoir operations, ANN has been extensively employed, with or without supporting algorithm for the training of its data for supervised and unsupervised learning. It has, however, shown that, even though ANN is primarily a robust algorithm for real time multi-objective optimization, it works better with training algorithm and EA’s usage in this regard has recorded consistent predictive quality for any of the objectives.

Conflict of Interest

The authors declare that there is no conflict of interest regarding the publication of this paper.

References