Water Quality, Pollution and Reservation

Abstract

An emergency multi-objective framework is developed in this study to achieve optimal reservoir operating strategy under sudden pollution injection. A large number of reservoir release and pollution injection scenarios are considered to assess a wide range of future conditions. CE-QUAL-W2 model is used to simulate pollution concentration and reservoir cleanup time (RCT) at each scenario. To minimize the computational burden of the reservoir water quality modeling, a multilayer perceptron (MLP) neural network is trained and validated against the simulated responses to various scenarios. Forcings of this surrogate model are first orthogonalized with principal component analysis (PCA) to reduce the dimensions of the problem. This surrogate model helps estimate response variables at any intermediate scenario, and can be readily coupled with a non-dominated sorting genetic algorithm-II (NSGA-II) optimization model to render Pareto optimal solutions. Finally, a multi-method decision-making procedure is applied to select the best compromise solution between all stakeholders. This study is the first attempt that proposed an optimization model in case of reservoir optimal operation using PCA tool with MLP model to utilize the most efficient scenarios besides several conflict resolution models. The results show that closer pollution injection locations to the reservoir outlets lead to lower RCT. The selected solution for the closest injection location to reservoir outlets in a 40-day simulation, resulted in an 18-day RCT. The three objective functions of this study, including not-supplied water demands, weighted combination of frequency and magnitude of pollution violation, and ratio of pollution in reservoir outlets to the total injection, are associated with values of 34.44%, 0.686 (-) and 0.808 (-), respectively, for the selected compromise solution.

Keywords: Sudden pollution injection, Principal component analysis, Non-dominated sorting genetic algorithm-II, Social choice rules, Fallback bargaining

Introduction

Due to rapid growth of industry and population many countries face immediate and emerging water resources crises (Sadegh et al. 2010), both in terms of available water resources and their quality (Zhang and Mei 2013; Monfared et al. 2017). Climate change further strains available water resources (Mallakpour et al. 2018). Development and expansion of agricultural and related chemical industries pose significant environmental hazards associated with contaminant agents that are released to the surface and groundwater resources during the processes of production, storage and transfer of food and materials used in the food chain (Wang et al. 2013). Dams are among the most important built infrastructure to supply agricultural, industrial and urban water demand, and are specifically vulnerable to contaminants entering dam lakes from various sources. It is hence necessary to monitor pollution in reservoirs and develop reservoir management strategies to deal with critical conditions such as sudden pollution injection (Li et al. 2014), originating from various sources such as overturning of chemical carriers, sudden leaks, or even terrorist attacks (Shokri et al. 2014). Such incidents can disturb water supply for different sectors with significant socioeconomic repercussions. Urgency of this problem prompted several studies for optimal reservoir management in such critical situations (e. g., Saadatpour and Afshar 2013; Shokri et al. 2014; Zhang and Xin 2017; Darmian et al. 2018). While significant strides have been made, employing Principal Component Analysis (PCA) model to decrease simulation-optimization run-time given many possible reservoir release and injection pollution scenarios has not been explored in previous studies on sudden pollution injection.

In addition, in the case of sudden pollution injection, there are different objectives that can lead to conflict between stakeholders. Traditionally, only the Nash conflict resolution model is utilized to achieve the most preferable compromise reservoir management strategy between stakeholders (Haddad et al. 2013; Yang et al. 2014). However, several more efficient conflict resolution models such as fallback bargaining (FB) and social choice rules (SCRs) have been employed in various environmental fields recently (e. g., Mahjouri and Bizhani-Manzar 2013; Ghodsi et al. 2016) attempt, proposes an optimization model with least run-time using PCA tool and FB and SCRs conflict resolution models to efficiently address reservoir optimal operation under sudden pollution injection. To simulate pollution propagation in a reservoir and analyze reservoir responses, such as cleanup time, the CE-QUAl-W2 numerical model is an appropriate tool widely used in the literature (e. g., Haddad et al. 2013; Saadatpour and Afshar 2013; Shokri et al. 2014). This model is, however, associated with a high run-time when considering different scenarios. This problem can be avoided by using artificial intelligence models to emulate reservoir response to pollution injection. For training and validation of the artificial intelligence model, the implementation of the CE-QUAL-W2 numerical model is required, which results in a large input-output dataset. The redundant information in this dataset prevents the artificial intelligence model from uncovering the actual response functions in the input-output data, and makes the surrogate model imprecise and time-consume especially when coupled with the multi-objective optimization model. PCA tool can be an appropriate tool to solve this problem. By orthogonalizing the input-output dataset and decreasing its dimension, PCA increases the precision of the emulator and reduces its run-time. The validated PCA-based multilayer perceptron (MLP) neural network (MLP-PCA), as the surrogate model, is amenable to coupling with the reservoir operation optimization model (such as NSGA-II multi-objective optimization model). Applying coupled MLP-PCA model with NSGA-II optimization model considered as the novelty of present study that have never been utilized in previous studies.

Satisfying water demands and optimal management of water quality are main goals in reservoir operation in sudden pollution injection situation (Darmian et al. 2018). Literature has explored genetic algorithm (GA) as a single-objective optimization model (Zhang and Xin 2017), non-dominated sorting genetic algorithm-II (NSGA-II) ¬¬¬¬ (Shokri et al. 2014) and multi-objective particle swarm optimization (MOPSO) (Saadatpour and Afshar 2013) to optimize these objectives. These studies point to drawbacks of using a multi-objective optimization model that is coupled with an artificial intelligence model, including the low number of considered scenarios, developing only two-objective optimization model, and high run-time of coupled artificial intelligence model with optimization model due to the large input-output dataset. While, present study tries to fill mentioned voids by using coupled MLP-PCA model with Optimization procedure which resulted in less run-time and acceptable results. In the few studies in this field that used multi-objective optimization model, only two of the following objectives were considered:

  1. minimizing not-supplied water demands,
  2.  maximizing pollution release from reservoir outlets that leads to minimizing the reservoir cleanup time (RCT), and
  3.  minimizing weighted magnitude and frequency of water quality standard violations. To avoid suboptimal strategies and to achieve superior results, we consider all three above objectives simultaneously.

In the present study, a new conflict resolution-based methodology for multi-objective reservoir emergency operation under sudden pollution injection is developed. In this framework, a certain amount of pollution with coliform index is suddenly injected to the reservoir, and using the CE-QUAL-W2 numerical model, the reservoir is simulated to determine its cleanup time. CE-QUAL-W2 is developed to simulate pollution concentration values in all reservoir outlets for different possible release and injection scenarios. Since the coupling of CE-QUAL-W2 numerical model with the NSGA-II multi-objective optimization model is computationally expensive, a PCA-based MLP neural network model is trained and validated to emulate the numerical model. Then, the validated MLP-PCA model is coupled with the NSGA-II optimization model, which yields the tradeoff surface between the three objectives defined previously. To select the best compromise solution among the Pareto optimal results, two FB methods (q-Approval fallback bargaining and unanimity fallback bargaining) and two SCRs (Borda scoring, BS, and Condorcet choice, CC) are considered.

Finally, the best of the four mentioned conflict resolution models is selected using mediation voting rule (MVR), as MVR is able to determine median solution with the highest stakeholders’ support. This paper proceeds with presenting the details of proposed methodology and case study in sections 2 and 3. The results and conclusion of the developed models for discovering the best reservoir operation policy in condition of sudden pollution injection are presented in sections 4 and 5, respectively.